1 /*
2 * Copyright (c) 1997, 2026, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2024, Alibaba Group Holding Limited. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "classfile/javaClasses.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "gc/shared/tlab_globals.hpp"
31 #include "memory/allocation.inline.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "oops/objArrayKlass.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/arraycopynode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/compile.hpp"
38 #include "opto/connode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/loopnode.hpp"
41 #include "opto/machnode.hpp"
42 #include "opto/matcher.hpp"
43 #include "opto/memnode.hpp"
44 #include "opto/mempointer.hpp"
45 #include "opto/mulnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/phaseX.hpp"
48 #include "opto/regalloc.hpp"
49 #include "opto/regmask.hpp"
50 #include "opto/rootnode.hpp"
51 #include "opto/traceMergeStoresTag.hpp"
52 #include "opto/vectornode.hpp"
53 #include "utilities/align.hpp"
54 #include "utilities/copy.hpp"
55 #include "utilities/macros.hpp"
56 #include "utilities/powerOfTwo.hpp"
57 #include "utilities/vmError.hpp"
58
59 // Portions of code courtesy of Clifford Click
60
61 // Optimization - Graph Style
62
63 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
64
65 //=============================================================================
66 uint MemNode::size_of() const { return sizeof(*this); }
67
68 const TypePtr *MemNode::adr_type() const {
69 Node* adr = in(Address);
70 if (adr == nullptr) return nullptr; // node is dead
71 const TypePtr* cross_check = nullptr;
72 DEBUG_ONLY(cross_check = _adr_type);
73 return calculate_adr_type(adr->bottom_type(), cross_check);
74 }
75
76 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
77 if (adr != nullptr) {
78 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
79 return true;
80 }
81 }
82 return false;
83 }
84
85 #ifndef PRODUCT
86 void MemNode::dump_spec(outputStream *st) const {
87 if (in(Address) == nullptr) return; // node is dead
88 #ifndef ASSERT
89 // fake the missing field
90 const TypePtr* _adr_type = nullptr;
91 if (in(Address) != nullptr)
92 _adr_type = in(Address)->bottom_type()->isa_ptr();
93 #endif
94 dump_adr_type(_adr_type, st);
95
96 Compile* C = Compile::current();
97 if (C->alias_type(_adr_type)->is_volatile()) {
98 st->print(" Volatile!");
99 }
100 if (_unaligned_access) {
101 st->print(" unaligned");
102 }
103 if (_mismatched_access) {
104 st->print(" mismatched");
105 }
106 if (_unsafe_access) {
107 st->print(" unsafe");
108 }
109 }
110
111 void MemNode::dump_adr_type(const TypePtr* adr_type, outputStream* st) {
112 st->print(" @");
113 if (adr_type == nullptr) {
114 st->print("null");
115 } else {
116 adr_type->dump_on(st);
117 Compile* C = Compile::current();
118 Compile::AliasType* atp = nullptr;
119 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
120 if (atp == nullptr)
121 st->print(", idx=?\?;");
122 else if (atp->index() == Compile::AliasIdxBot)
123 st->print(", idx=Bot;");
124 else if (atp->index() == Compile::AliasIdxTop)
125 st->print(", idx=Top;");
126 else if (atp->index() == Compile::AliasIdxRaw)
127 st->print(", idx=Raw;");
128 else {
129 ciField* field = atp->field();
130 if (field) {
131 st->print(", name=");
132 field->print_name_on(st);
133 }
134 st->print(", idx=%d;", atp->index());
135 }
136 }
137 }
138
139 extern void print_alias_types();
140
141 #endif
142
143 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
144 assert((t_oop != nullptr), "sanity");
145 bool is_instance = t_oop->is_known_instance_field();
146 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
147 (load != nullptr) && load->is_Load() &&
148 (phase->is_IterGVN() != nullptr);
149 if (!(is_instance || is_boxed_value_load))
150 return mchain; // don't try to optimize non-instance types
151 uint instance_id = t_oop->instance_id();
152 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
153 Node *prev = nullptr;
154 Node *result = mchain;
155 while (prev != result) {
156 prev = result;
157 if (result == start_mem)
158 break; // hit one of our sentinels
159 // skip over a call which does not affect this memory slice
160 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
161 Node *proj_in = result->in(0);
162 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
163 break; // hit one of our sentinels
164 } else if (proj_in->is_Call()) {
165 // ArrayCopyNodes processed here as well
166 CallNode *call = proj_in->as_Call();
167 if (!call->may_modify(t_oop, phase)) { // returns false for instances
168 result = call->in(TypeFunc::Memory);
169 }
170 } else if (proj_in->is_Initialize()) {
171 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
172 // Stop if this is the initialization for the object instance which
173 // contains this memory slice, otherwise skip over it.
174 if ((alloc == nullptr) || (alloc->_idx == instance_id)) {
175 break;
176 }
177 if (is_instance) {
178 result = proj_in->in(TypeFunc::Memory);
179 } else if (is_boxed_value_load) {
180 Node* klass = alloc->in(AllocateNode::KlassNode);
181 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
182 if (tklass->klass_is_exact() && !tklass->exact_klass()->equals(t_oop->is_instptr()->exact_klass())) {
183 result = proj_in->in(TypeFunc::Memory); // not related allocation
184 }
185 }
186 } else if (proj_in->is_MemBar()) {
187 ArrayCopyNode* ac = nullptr;
188 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
189 break;
190 }
191 result = proj_in->in(TypeFunc::Memory);
192 } else if (proj_in->is_top()) {
193 break; // dead code
194 } else {
195 assert(false, "unexpected projection");
196 }
197 } else if (result->is_ClearArray()) {
198 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
199 // Can not bypass initialization of the instance
200 // we are looking for.
201 break;
202 }
203 // Otherwise skip it (the call updated 'result' value).
204 } else if (result->is_MergeMem()) {
205 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, nullptr, tty);
206 }
207 }
208 return result;
209 }
210
211 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
212 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
213 if (t_oop == nullptr)
214 return mchain; // don't try to optimize non-oop types
215 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
216 bool is_instance = t_oop->is_known_instance_field();
217 PhaseIterGVN *igvn = phase->is_IterGVN();
218 if (is_instance && igvn != nullptr && result->is_Phi()) {
219 PhiNode *mphi = result->as_Phi();
220 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
221 const TypePtr *t = mphi->adr_type();
222 bool do_split = false;
223 // In the following cases, Load memory input can be further optimized based on
224 // its precise address type
225 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ) {
226 do_split = true;
227 } else if (t->isa_oopptr() && !t->is_oopptr()->is_known_instance()) {
228 const TypeOopPtr* mem_t =
229 t->is_oopptr()->cast_to_exactness(true)
230 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
231 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id());
232 if (t_oop->isa_aryptr()) {
233 mem_t = mem_t->is_aryptr()
234 ->cast_to_stable(t_oop->is_aryptr()->is_stable())
235 ->cast_to_size(t_oop->is_aryptr()->size())
236 ->with_offset(t_oop->is_aryptr()->offset())
237 ->is_aryptr();
238 }
239 do_split = mem_t == t_oop;
240 }
241 if (do_split) {
242 // clone the Phi with our address type
243 result = mphi->split_out_instance(t_adr, igvn);
244 } else {
245 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
246 }
247 }
248 return result;
249 }
250
251 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
252 uint alias_idx = phase->C->get_alias_index(tp);
253 Node *mem = mmem;
254 #ifdef ASSERT
255 {
256 // Check that current type is consistent with the alias index used during graph construction
257 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
258 bool consistent = adr_check == nullptr || adr_check->empty() ||
259 phase->C->must_alias(adr_check, alias_idx );
260 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
261 if( !consistent && adr_check != nullptr && !adr_check->empty() &&
262 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
263 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
264 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
265 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
266 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
267 // don't assert if it is dead code.
268 consistent = true;
269 }
270 if( !consistent ) {
271 st->print("alias_idx==%d, adr_check==", alias_idx);
272 if( adr_check == nullptr ) {
273 st->print("null");
274 } else {
275 adr_check->dump();
276 }
277 st->cr();
278 print_alias_types();
279 assert(consistent, "adr_check must match alias idx");
280 }
281 }
282 #endif
283 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
284 // means an array I have not precisely typed yet. Do not do any
285 // alias stuff with it any time soon.
286 const TypeOopPtr *toop = tp->isa_oopptr();
287 if (tp->base() != Type::AnyPtr &&
288 !(toop &&
289 toop->isa_instptr() &&
290 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
291 toop->offset() == Type::OffsetBot)) {
292 // IGVN _delay_transform may be set to true and if that is the case and mmem
293 // is already a registered node then the validation inside transform will
294 // complain.
295 Node* m = mmem;
296 PhaseIterGVN* igvn = phase->is_IterGVN();
297 if (igvn == nullptr || !igvn->delay_transform()) {
298 // compress paths and change unreachable cycles to TOP
299 // If not, we can update the input infinitely along a MergeMem cycle
300 // Equivalent code in PhiNode::Ideal
301 m = phase->transform(mmem);
302 }
303 // If transformed to a MergeMem, get the desired slice
304 // Otherwise the returned node represents memory for every slice
305 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
306 // Update input if it is progress over what we have now
307 }
308 return mem;
309 }
310
311 //--------------------------Ideal_common---------------------------------------
312 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
313 // Unhook non-raw memories from complete (macro-expanded) initializations.
314 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
315 // If our control input is a dead region, kill all below the region
316 Node *ctl = in(MemNode::Control);
317 if (ctl && remove_dead_region(phase, can_reshape))
318 return this;
319 ctl = in(MemNode::Control);
320 // Don't bother trying to transform a dead node
321 if (ctl && ctl->is_top()) return NodeSentinel;
322
323 PhaseIterGVN *igvn = phase->is_IterGVN();
324 // Wait if control on the worklist.
325 if (ctl && can_reshape && igvn != nullptr) {
326 Node* bol = nullptr;
327 Node* cmp = nullptr;
328 if (ctl->in(0)->is_If()) {
329 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
330 bol = ctl->in(0)->in(1);
331 if (bol->is_Bool())
332 cmp = ctl->in(0)->in(1)->in(1);
333 }
334 if (igvn->_worklist.member(ctl) ||
335 (bol != nullptr && igvn->_worklist.member(bol)) ||
336 (cmp != nullptr && igvn->_worklist.member(cmp)) ) {
337 // This control path may be dead.
338 // Delay this memory node transformation until the control is processed.
339 igvn->_worklist.push(this);
340 return NodeSentinel; // caller will return null
341 }
342 }
343 // Ignore if memory is dead, or self-loop
344 Node *mem = in(MemNode::Memory);
345 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return null
346 assert(mem != this, "dead loop in MemNode::Ideal");
347
348 if (can_reshape && igvn != nullptr && igvn->_worklist.member(mem)) {
349 // This memory slice may be dead.
350 // Delay this mem node transformation until the memory is processed.
351 igvn->_worklist.push(this);
352 return NodeSentinel; // caller will return null
353 }
354
355 Node *address = in(MemNode::Address);
356 const Type *t_adr = phase->type(address);
357 if (t_adr == Type::TOP) return NodeSentinel; // caller will return null
358
359 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
360 // Unsafe off-heap access with zero address. Remove access and other control users
361 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
362 assert(ctl != nullptr, "unsafe accesses should be control dependent");
363 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
364 Node* u = ctl->fast_out(i);
365 if (u != ctl) {
366 igvn->rehash_node_delayed(u);
367 int nb = u->replace_edge(ctl, phase->C->top(), igvn);
368 --i, imax -= nb;
369 }
370 }
371 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
372 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
373 phase->C->root()->add_req(halt);
374 return this;
375 }
376
377 if (can_reshape && igvn != nullptr &&
378 (igvn->_worklist.member(address) ||
379 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
380 // The address's base and type may change when the address is processed.
381 // Delay this mem node transformation until the address is processed.
382 igvn->_worklist.push(this);
383 return NodeSentinel; // caller will return null
384 }
385
386 // Do NOT remove or optimize the next lines: ensure a new alias index
387 // is allocated for an oop pointer type before Escape Analysis.
388 // Note: C++ will not remove it since the call has side effect.
389 if (t_adr->isa_oopptr()) {
390 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
391 }
392
393 Node* base = nullptr;
394 if (address->is_AddP()) {
395 base = address->in(AddPNode::Base);
396 }
397 if (base != nullptr && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
398 !t_adr->isa_rawptr()) {
399 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
400 // Skip this node optimization if its address has TOP base.
401 return NodeSentinel; // caller will return null
402 }
403
404 // Avoid independent memory operations
405 Node* old_mem = mem;
406
407 // The code which unhooks non-raw memories from complete (macro-expanded)
408 // initializations was removed. After macro-expansion all stores caught
409 // by Initialize node became raw stores and there is no information
410 // which memory slices they modify. So it is unsafe to move any memory
411 // operation above these stores. Also in most cases hooked non-raw memories
412 // were already unhooked by using information from detect_ptr_independence()
413 // and find_previous_store().
414
415 if (mem->is_MergeMem()) {
416 MergeMemNode* mmem = mem->as_MergeMem();
417 const TypePtr *tp = t_adr->is_ptr();
418
419 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
420 }
421
422 if (mem != old_mem) {
423 set_req_X(MemNode::Memory, mem, phase);
424 if (phase->type(mem) == Type::TOP) return NodeSentinel;
425 return this;
426 }
427
428 // let the subclass continue analyzing...
429 return nullptr;
430 }
431
432 // Helper function for proving some simple control dominations.
433 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
434 // Already assumes that 'dom' is available at 'sub', and that 'sub'
435 // is not a constant (dominated by the method's StartNode).
436 // Used by MemNode::find_previous_store to prove that the
437 // control input of a memory operation predates (dominates)
438 // an allocation it wants to look past.
439 // Returns 'DomResult::Dominate' if all control inputs of 'dom'
440 // dominate 'sub', 'DomResult::NotDominate' if not,
441 // and 'DomResult::EncounteredDeadCode' if we can't decide due to
442 // dead code, but at the end of IGVN, we know the definite result
443 // once the dead code is cleaned up.
444 Node::DomResult MemNode::maybe_all_controls_dominate(Node* dom, Node* sub) {
445 if (dom == nullptr || dom->is_top() || sub == nullptr || sub->is_top()) {
446 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
447 }
448
449 // Check 'dom'. Skip Proj and CatchProj nodes.
450 dom = dom->find_exact_control(dom);
451 if (dom == nullptr || dom->is_top()) {
452 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
453 }
454
455 if (dom == sub) {
456 // For the case when, for example, 'sub' is Initialize and the original
457 // 'dom' is Proj node of the 'sub'.
458 return DomResult::NotDominate;
459 }
460
461 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) {
462 return DomResult::Dominate;
463 }
464
465 // 'dom' dominates 'sub' if its control edge and control edges
466 // of all its inputs dominate or equal to sub's control edge.
467
468 // Currently 'sub' is either Allocate, Initialize or Start nodes.
469 // Or Region for the check in LoadNode::Ideal();
470 // 'sub' should have sub->in(0) != nullptr.
471 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
472 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
473
474 // Get control edge of 'sub'.
475 Node* orig_sub = sub;
476 sub = sub->find_exact_control(sub->in(0));
477 if (sub == nullptr || sub->is_top()) {
478 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
479 }
480
481 assert(sub->is_CFG(), "expecting control");
482
483 if (sub == dom) {
484 return DomResult::Dominate;
485 }
486
487 if (sub->is_Start() || sub->is_Root()) {
488 return DomResult::NotDominate;
489 }
490
491 {
492 // Check all control edges of 'dom'.
493
494 ResourceMark rm;
495 Node_List nlist;
496 Unique_Node_List dom_list;
497
498 dom_list.push(dom);
499 bool only_dominating_controls = false;
500
501 for (uint next = 0; next < dom_list.size(); next++) {
502 Node* n = dom_list.at(next);
503 if (n == orig_sub) {
504 return DomResult::NotDominate; // One of dom's inputs dominated by sub.
505 }
506 if (!n->is_CFG() && n->pinned()) {
507 // Check only own control edge for pinned non-control nodes.
508 n = n->find_exact_control(n->in(0));
509 if (n == nullptr || n->is_top()) {
510 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
511 }
512 assert(n->is_CFG(), "expecting control");
513 dom_list.push(n);
514 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
515 only_dominating_controls = true;
516 } else if (n->is_CFG()) {
517 DomResult dom_result = n->dominates(sub, nlist);
518 if (dom_result == DomResult::Dominate) {
519 only_dominating_controls = true;
520 } else {
521 return dom_result;
522 }
523 } else {
524 // First, own control edge.
525 Node* m = n->find_exact_control(n->in(0));
526 if (m != nullptr) {
527 if (m->is_top()) {
528 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
529 }
530 dom_list.push(m);
531 }
532 // Now, the rest of edges.
533 uint cnt = n->req();
534 for (uint i = 1; i < cnt; i++) {
535 m = n->find_exact_control(n->in(i));
536 if (m == nullptr || m->is_top()) {
537 continue;
538 }
539 dom_list.push(m);
540 }
541 }
542 }
543 return only_dominating_controls ? DomResult::Dominate : DomResult::NotDominate;
544 }
545 }
546
547 //---------------------detect_ptr_independence---------------------------------
548 // Used by MemNode::find_previous_store to prove that two base
549 // pointers are never equal.
550 // The pointers are accompanied by their associated allocations,
551 // if any, which have been previously discovered by the caller.
552 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
553 Node* p2, AllocateNode* a2,
554 PhaseTransform* phase) {
555 // Attempt to prove that these two pointers cannot be aliased.
556 // They may both manifestly be allocations, and they should differ.
557 // Or, if they are not both allocations, they can be distinct constants.
558 // Otherwise, one is an allocation and the other a pre-existing value.
559 if (a1 == nullptr && a2 == nullptr) { // neither an allocation
560 return (p1 != p2) && p1->is_Con() && p2->is_Con();
561 } else if (a1 != nullptr && a2 != nullptr) { // both allocations
562 return (a1 != a2);
563 } else if (a1 != nullptr) { // one allocation a1
564 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
565 return all_controls_dominate(p2, a1);
566 } else { //(a2 != null) // one allocation a2
567 return all_controls_dominate(p1, a2);
568 }
569 return false;
570 }
571
572
573 // Find an arraycopy ac that produces the memory state represented by parameter mem.
574 // Return ac if
575 // (a) can_see_stored_value=true and ac must have set the value for this load or if
576 // (b) can_see_stored_value=false and ac could have set the value for this load or if
577 // (c) can_see_stored_value=false and ac cannot have set the value for this load.
578 // In case (c) change the parameter mem to the memory input of ac to skip it
579 // when searching stored value.
580 // Otherwise return null.
581 Node* LoadNode::find_previous_arraycopy(PhaseValues* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
582 ArrayCopyNode* ac = find_array_copy_clone(ld_alloc, mem);
583 if (ac != nullptr) {
584 Node* ld_addp = in(MemNode::Address);
585 Node* src = ac->in(ArrayCopyNode::Src);
586 const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
587
588 // This is a load from a cloned array. The corresponding arraycopy ac must
589 // have set the value for the load and we can return ac but only if the load
590 // is known to be within bounds. This is checked below.
591 if (ary_t != nullptr && ld_addp->is_AddP()) {
592 Node* ld_offs = ld_addp->in(AddPNode::Offset);
593 BasicType ary_elem = ary_t->elem()->array_element_basic_type();
594 jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
595 jlong elemsize = type2aelembytes(ary_elem);
596
597 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
598 const TypeInt* sizetype = ary_t->size();
599
600 if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
601 // The load is known to be within bounds. It receives its value from ac.
602 return ac;
603 }
604 // The load is known to be out-of-bounds.
605 }
606 // The load could be out-of-bounds. It must not be hoisted but must remain
607 // dependent on the runtime range check. This is achieved by returning null.
608 } else if (mem->is_Proj() && mem->in(0) != nullptr && mem->in(0)->is_ArrayCopy()) {
609 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
610
611 if (ac->is_arraycopy_validated() ||
612 ac->is_copyof_validated() ||
613 ac->is_copyofrange_validated()) {
614 Node* ld_addp = in(MemNode::Address);
615 if (ld_addp->is_AddP()) {
616 Node* ld_base = ld_addp->in(AddPNode::Address);
617 Node* ld_offs = ld_addp->in(AddPNode::Offset);
618
619 Node* dest = ac->in(ArrayCopyNode::Dest);
620
621 if (dest == ld_base) {
622 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
623 assert(!ld_offs_t->empty(), "dead reference should be checked already");
624 // Take into account vector or unsafe access size
625 jlong ld_size_in_bytes = (jlong)memory_size();
626 jlong offset_hi = ld_offs_t->_hi + ld_size_in_bytes - 1;
627 offset_hi = MIN2(offset_hi, (jlong)(TypeX::MAX->_hi)); // Take care for overflow in 32-bit VM
628 if (ac->modifies(ld_offs_t->_lo, (intptr_t)offset_hi, phase, can_see_stored_value)) {
629 return ac;
630 }
631 if (!can_see_stored_value) {
632 mem = ac->in(TypeFunc::Memory);
633 return ac;
634 }
635 }
636 }
637 }
638 }
639 return nullptr;
640 }
641
642 ArrayCopyNode* MemNode::find_array_copy_clone(Node* ld_alloc, Node* mem) const {
643 if (mem->is_Proj() && mem->in(0) != nullptr && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
644 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
645 if (ld_alloc != nullptr) {
646 // Check if there is an array copy for a clone
647 Node* mb = mem->in(0);
648 ArrayCopyNode* ac = nullptr;
649 if (mb->in(0) != nullptr && mb->in(0)->is_Proj() &&
650 mb->in(0)->in(0) != nullptr && mb->in(0)->in(0)->is_ArrayCopy()) {
651 ac = mb->in(0)->in(0)->as_ArrayCopy();
652 } else {
653 // Step over GC barrier when ReduceInitialCardMarks is disabled
654 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
655 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
656
657 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
658 ac = control_proj_ac->in(0)->as_ArrayCopy();
659 }
660 }
661
662 if (ac != nullptr && ac->is_clonebasic()) {
663 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest));
664 if (alloc != nullptr && alloc == ld_alloc) {
665 return ac;
666 }
667 }
668 }
669 }
670 return nullptr;
671 }
672
673 // The logic for reordering loads and stores uses four steps:
674 // (a) Walk carefully past stores and initializations which we
675 // can prove are independent of this load.
676 // (b) Observe that the next memory state makes an exact match
677 // with self (load or store), and locate the relevant store.
678 // (c) Ensure that, if we were to wire self directly to the store,
679 // the optimizer would fold it up somehow.
680 // (d) Do the rewiring, and return, depending on some other part of
681 // the optimizer to fold up the load.
682 // This routine handles steps (a) and (b). Steps (c) and (d) are
683 // specific to loads and stores, so they are handled by the callers.
684 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
685 //
686 Node* MemNode::find_previous_store(PhaseValues* phase) {
687 Node* ctrl = in(MemNode::Control);
688 Node* adr = in(MemNode::Address);
689 intptr_t offset = 0;
690 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
691 AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
692
693 if (offset == Type::OffsetBot)
694 return nullptr; // cannot unalias unless there are precise offsets
695
696 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
697 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
698
699 intptr_t size_in_bytes = memory_size();
700
701 Node* mem = in(MemNode::Memory); // start searching here...
702
703 int cnt = 50; // Cycle limiter
704 for (;;) { // While we can dance past unrelated stores...
705 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
706
707 Node* prev = mem;
708 if (mem->is_Store()) {
709 Node* st_adr = mem->in(MemNode::Address);
710 intptr_t st_offset = 0;
711 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
712 if (st_base == nullptr)
713 break; // inscrutable pointer
714
715 // For raw accesses it's not enough to prove that constant offsets don't intersect.
716 // We need the bases to be the equal in order for the offset check to make sense.
717 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
718 break;
719 }
720
721 if (st_offset != offset && st_offset != Type::OffsetBot) {
722 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
723 assert(mem->as_Store()->memory_size() <= MAX_STORE, "");
724 if (st_offset >= offset + size_in_bytes ||
725 st_offset <= offset - MAX_STORE ||
726 st_offset <= offset - mem->as_Store()->memory_size()) {
727 // Success: The offsets are provably independent.
728 // (You may ask, why not just test st_offset != offset and be done?
729 // The answer is that stores of different sizes can co-exist
730 // in the same sequence of RawMem effects. We sometimes initialize
731 // a whole 'tile' of array elements with a single jint or jlong.)
732 mem = mem->in(MemNode::Memory);
733 continue; // (a) advance through independent store memory
734 }
735 }
736 if (st_base != base &&
737 detect_ptr_independence(base, alloc,
738 st_base,
739 AllocateNode::Ideal_allocation(st_base),
740 phase)) {
741 // Success: The bases are provably independent.
742 mem = mem->in(MemNode::Memory);
743 continue; // (a) advance through independent store memory
744 }
745
746 // (b) At this point, if the bases or offsets do not agree, we lose,
747 // since we have not managed to prove 'this' and 'mem' independent.
748 if (st_base == base && st_offset == offset) {
749 return mem; // let caller handle steps (c), (d)
750 }
751
752 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
753 InitializeNode* st_init = mem->in(0)->as_Initialize();
754 AllocateNode* st_alloc = st_init->allocation();
755 if (st_alloc == nullptr) {
756 break; // something degenerated
757 }
758 bool known_identical = false;
759 bool known_independent = false;
760 if (alloc == st_alloc) {
761 known_identical = true;
762 } else if (alloc != nullptr) {
763 known_independent = true;
764 } else if (all_controls_dominate(this, st_alloc)) {
765 known_independent = true;
766 }
767
768 if (known_independent) {
769 // The bases are provably independent: Either they are
770 // manifestly distinct allocations, or else the control
771 // of this load dominates the store's allocation.
772 int alias_idx = phase->C->get_alias_index(adr_type());
773 if (alias_idx == Compile::AliasIdxRaw) {
774 mem = st_alloc->in(TypeFunc::Memory);
775 } else {
776 mem = st_init->memory(alias_idx);
777 }
778 continue; // (a) advance through independent store memory
779 }
780
781 // (b) at this point, if we are not looking at a store initializing
782 // the same allocation we are loading from, we lose.
783 if (known_identical) {
784 // From caller, can_see_stored_value will consult find_captured_store.
785 return mem; // let caller handle steps (c), (d)
786 }
787
788 } else if (find_previous_arraycopy(phase, alloc, mem, false) != nullptr) {
789 if (prev != mem) {
790 // Found an arraycopy but it doesn't affect that load
791 continue;
792 }
793 // Found an arraycopy that may affect that load
794 return mem;
795 } else if (addr_t != nullptr && addr_t->is_known_instance_field()) {
796 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
797 if (mem->is_Proj() && mem->in(0)->is_Call()) {
798 // ArrayCopyNodes processed here as well.
799 CallNode *call = mem->in(0)->as_Call();
800 if (!call->may_modify(addr_t, phase)) {
801 mem = call->in(TypeFunc::Memory);
802 continue; // (a) advance through independent call memory
803 }
804 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
805 ArrayCopyNode* ac = nullptr;
806 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
807 break;
808 }
809 mem = mem->in(0)->in(TypeFunc::Memory);
810 continue; // (a) advance through independent MemBar memory
811 } else if (mem->is_ClearArray()) {
812 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
813 // (the call updated 'mem' value)
814 continue; // (a) advance through independent allocation memory
815 } else {
816 // Can not bypass initialization of the instance
817 // we are looking for.
818 return mem;
819 }
820 } else if (mem->is_MergeMem()) {
821 int alias_idx = phase->C->get_alias_index(adr_type());
822 mem = mem->as_MergeMem()->memory_at(alias_idx);
823 continue; // (a) advance through independent MergeMem memory
824 }
825 }
826
827 // Unless there is an explicit 'continue', we must bail out here,
828 // because 'mem' is an inscrutable memory state (e.g., a call).
829 break;
830 }
831
832 return nullptr; // bail out
833 }
834
835 //----------------------calculate_adr_type-------------------------------------
836 // Helper function. Notices when the given type of address hits top or bottom.
837 // Also, asserts a cross-check of the type against the expected address type.
838 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
839 if (t == Type::TOP) return nullptr; // does not touch memory any more?
840 #ifdef ASSERT
841 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = nullptr;
842 #endif
843 const TypePtr* tp = t->isa_ptr();
844 if (tp == nullptr) {
845 assert(cross_check == nullptr || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
846 return TypePtr::BOTTOM; // touches lots of memory
847 } else {
848 #ifdef ASSERT
849 // %%%% [phh] We don't check the alias index if cross_check is
850 // TypeRawPtr::BOTTOM. Needs to be investigated.
851 if (cross_check != nullptr &&
852 cross_check != TypePtr::BOTTOM &&
853 cross_check != TypeRawPtr::BOTTOM) {
854 // Recheck the alias index, to see if it has changed (due to a bug).
855 Compile* C = Compile::current();
856 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
857 "must stay in the original alias category");
858 // The type of the address must be contained in the adr_type,
859 // disregarding "null"-ness.
860 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
861 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
862 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
863 "real address must not escape from expected memory type");
864 }
865 #endif
866 return tp;
867 }
868 }
869
870 uint8_t MemNode::barrier_data(const Node* n) {
871 if (n->is_LoadStore()) {
872 return n->as_LoadStore()->barrier_data();
873 } else if (n->is_Mem()) {
874 return n->as_Mem()->barrier_data();
875 }
876 return 0;
877 }
878
879 //=============================================================================
880 // Should LoadNode::Ideal() attempt to remove control edges?
881 bool LoadNode::can_remove_control() const {
882 return !has_pinned_control_dependency();
883 }
884 uint LoadNode::size_of() const { return sizeof(*this); }
885 bool LoadNode::cmp(const Node &n) const {
886 LoadNode& load = (LoadNode &)n;
887 return Type::equals(_type, load._type) &&
888 _control_dependency == load._control_dependency &&
889 _mo == load._mo;
890 }
891 const Type *LoadNode::bottom_type() const { return _type; }
892 uint LoadNode::ideal_reg() const {
893 return _type->ideal_reg();
894 }
895
896 #ifndef PRODUCT
897 void LoadNode::dump_spec(outputStream *st) const {
898 MemNode::dump_spec(st);
899 if( !Verbose && !WizardMode ) {
900 // standard dump does this in Verbose and WizardMode
901 st->print(" #"); _type->dump_on(st);
902 }
903 if (in(0) != nullptr && !depends_only_on_test()) {
904 st->print(" (does not depend only on test, ");
905 if (control_dependency() == UnknownControl) {
906 st->print("unknown control");
907 } else if (control_dependency() == Pinned) {
908 st->print("pinned");
909 } else if (adr_type() == TypeRawPtr::BOTTOM) {
910 st->print("raw access");
911 } else {
912 st->print("unknown reason");
913 }
914 st->print(")");
915 }
916 }
917 #endif
918
919 #ifdef ASSERT
920 //----------------------------is_immutable_value-------------------------------
921 // Helper function to allow a raw load without control edge for some cases
922 bool LoadNode::is_immutable_value(Node* adr) {
923 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
924 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) {
925
926 jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1);
927 int offsets[] = {
928 in_bytes(JavaThread::osthread_offset()),
929 in_bytes(JavaThread::threadObj_offset()),
930 in_bytes(JavaThread::vthread_offset()),
931 in_bytes(JavaThread::scopedValueCache_offset()),
932 };
933
934 for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) {
935 if (offset == offsets[i]) {
936 return true;
937 }
938 }
939 }
940
941 return false;
942 }
943 #endif
944
945 //----------------------------LoadNode::make-----------------------------------
946 // Polymorphic factory method:
947 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo,
948 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
949 Compile* C = gvn.C;
950
951 // sanity check the alias category against the created node type
952 assert(!(adr_type->isa_oopptr() &&
953 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
954 "use LoadKlassNode instead");
955 assert(!(adr_type->isa_aryptr() &&
956 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
957 "use LoadRangeNode instead");
958 // Check control edge of raw loads
959 assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
960 // oop will be recorded in oop map if load crosses safepoint
961 rt->isa_oopptr() || is_immutable_value(adr),
962 "raw memory operations should have control edge");
963 LoadNode* load = nullptr;
964 switch (bt) {
965 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
966 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
967 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
968 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
969 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
970 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break;
971 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
972 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break;
973 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
974 case T_OBJECT:
975 case T_NARROWOOP:
976 #ifdef _LP64
977 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
978 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
979 } else
980 #endif
981 {
982 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
983 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
984 }
985 break;
986 default:
987 ShouldNotReachHere();
988 break;
989 }
990 assert(load != nullptr, "LoadNode should have been created");
991 if (unaligned) {
992 load->set_unaligned_access();
993 }
994 if (mismatched) {
995 load->set_mismatched_access();
996 }
997 if (unsafe) {
998 load->set_unsafe_access();
999 }
1000 load->set_barrier_data(barrier_data);
1001 if (load->Opcode() == Op_LoadN) {
1002 Node* ld = gvn.transform(load);
1003 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
1004 }
1005
1006 return load;
1007 }
1008
1009 //------------------------------hash-------------------------------------------
1010 uint LoadNode::hash() const {
1011 // unroll addition of interesting fields
1012 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
1013 }
1014
1015 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
1016 if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) {
1017 bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile();
1018 bool is_stable_ary = FoldStableValues &&
1019 (tp != nullptr) && (tp->isa_aryptr() != nullptr) &&
1020 tp->isa_aryptr()->is_stable();
1021
1022 return (eliminate_boxing && non_volatile) || is_stable_ary;
1023 }
1024
1025 return false;
1026 }
1027
1028 // Is the value loaded previously stored by an arraycopy? If so return
1029 // a load node that reads from the source array so we may be able to
1030 // optimize out the ArrayCopy node later.
1031 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
1032 Node* ld_adr = in(MemNode::Address);
1033 intptr_t ld_off = 0;
1034 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
1035 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
1036 if (ac != nullptr) {
1037 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
1038
1039 Node* mem = ac->in(TypeFunc::Memory);
1040 Node* ctl = ac->in(0);
1041 Node* src = ac->in(ArrayCopyNode::Src);
1042
1043 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
1044 return nullptr;
1045 }
1046
1047 // load depends on the tests that validate the arraycopy
1048 LoadNode* ld = clone_pinned();
1049 Node* addp = in(MemNode::Address)->clone();
1050 if (ac->as_ArrayCopy()->is_clonebasic()) {
1051 assert(ld_alloc != nullptr, "need an alloc");
1052 assert(addp->is_AddP(), "address must be addp");
1053 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1054 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1055 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1056 addp->set_req(AddPNode::Base, src);
1057 addp->set_req(AddPNode::Address, src);
1058 } else {
1059 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1060 ac->as_ArrayCopy()->is_copyof_validated() ||
1061 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1062 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1063 addp->set_req(AddPNode::Base, src);
1064 addp->set_req(AddPNode::Address, src);
1065
1066 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1067 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
1068 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
1069
1070 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
1071 uint shift = exact_log2(type2aelembytes(ary_elem));
1072
1073 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1074 #ifdef _LP64
1075 diff = phase->transform(new ConvI2LNode(diff));
1076 #endif
1077 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1078
1079 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1080 addp->set_req(AddPNode::Offset, offset);
1081 }
1082 addp = phase->transform(addp);
1083 #ifdef ASSERT
1084 const TypePtr* adr_type = phase->type(addp)->is_ptr();
1085 ld->_adr_type = adr_type;
1086 #endif
1087 ld->set_req(MemNode::Address, addp);
1088 ld->set_req(0, ctl);
1089 ld->set_req(MemNode::Memory, mem);
1090 return ld;
1091 }
1092 return nullptr;
1093 }
1094
1095
1096 //---------------------------can_see_stored_value------------------------------
1097 // This routine exists to make sure this set of tests is done the same
1098 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
1099 // will change the graph shape in a way which makes memory alive twice at the
1100 // same time (uses the Oracle model of aliasing), then some
1101 // LoadXNode::Identity will fold things back to the equivalence-class model
1102 // of aliasing.
1103 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const {
1104 Node* ld_adr = in(MemNode::Address);
1105 intptr_t ld_off = 0;
1106 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1107 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base);
1108 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1109 Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr;
1110 // This is more general than load from boxing objects.
1111 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1112 uint alias_idx = atp->index();
1113 Node* result = nullptr;
1114 Node* current = st;
1115 // Skip through chains of MemBarNodes checking the MergeMems for
1116 // new states for the slice of this load. Stop once any other
1117 // kind of node is encountered. Loads from final memory can skip
1118 // through any kind of MemBar but normal loads shouldn't skip
1119 // through MemBarAcquire since the could allow them to move out of
1120 // a synchronized region. It is not safe to step over MemBarCPUOrder,
1121 // because alias info above them may be inaccurate (e.g., due to
1122 // mixed/mismatched unsafe accesses).
1123 bool is_final_mem = !atp->is_rewritable();
1124 while (current->is_Proj()) {
1125 int opc = current->in(0)->Opcode();
1126 if ((is_final_mem && (opc == Op_MemBarAcquire ||
1127 opc == Op_MemBarAcquireLock ||
1128 opc == Op_LoadFence)) ||
1129 opc == Op_MemBarRelease ||
1130 opc == Op_StoreFence ||
1131 opc == Op_MemBarReleaseLock ||
1132 opc == Op_MemBarStoreStore ||
1133 opc == Op_StoreStoreFence) {
1134 Node* mem = current->in(0)->in(TypeFunc::Memory);
1135 if (mem->is_MergeMem()) {
1136 MergeMemNode* merge = mem->as_MergeMem();
1137 Node* new_st = merge->memory_at(alias_idx);
1138 if (new_st == merge->base_memory()) {
1139 // Keep searching
1140 current = new_st;
1141 continue;
1142 }
1143 // Save the new memory state for the slice and fall through
1144 // to exit.
1145 result = new_st;
1146 }
1147 }
1148 break;
1149 }
1150 if (result != nullptr) {
1151 st = result;
1152 }
1153 }
1154
1155 // Loop around twice in the case Load -> Initialize -> Store.
1156 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1157 for (int trip = 0; trip <= 1; trip++) {
1158
1159 if (st->is_Store()) {
1160 Node* st_adr = st->in(MemNode::Address);
1161 if (st_adr != ld_adr) {
1162 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1163 intptr_t st_off = 0;
1164 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1165 if (ld_base == nullptr) return nullptr;
1166 if (st_base == nullptr) return nullptr;
1167 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr;
1168 if (ld_off != st_off) return nullptr;
1169 if (ld_off == Type::OffsetBot) return nullptr;
1170 // Same base, same offset.
1171 // Possible improvement for arrays: check index value instead of absolute offset.
1172
1173 // At this point we have proven something like this setup:
1174 // B = << base >>
1175 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1176 // S = StoreQ(AddP( B , #Off), V)
1177 // (Actually, we haven't yet proven the Q's are the same.)
1178 // In other words, we are loading from a casted version of
1179 // the same pointer-and-offset that we stored to.
1180 // Casted version may carry a dependency and it is respected.
1181 // Thus, we are able to replace L by V.
1182 }
1183 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1184 if (store_Opcode() != st->Opcode()) {
1185 return nullptr;
1186 }
1187 // LoadVector/StoreVector needs additional check to ensure the types match.
1188 if (st->is_StoreVector()) {
1189 const TypeVect* in_vt = st->as_StoreVector()->vect_type();
1190 const TypeVect* out_vt = as_LoadVector()->vect_type();
1191 if (in_vt != out_vt) {
1192 return nullptr;
1193 }
1194 }
1195 return st->in(MemNode::ValueIn);
1196 }
1197
1198 // A load from a freshly-created object always returns zero.
1199 // (This can happen after LoadNode::Ideal resets the load's memory input
1200 // to find_captured_store, which returned InitializeNode::zero_memory.)
1201 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1202 (st->in(0) == ld_alloc) &&
1203 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1204 // return a zero value for the load's basic type
1205 // (This is one of the few places where a generic PhaseTransform
1206 // can create new nodes. Think of it as lazily manifesting
1207 // virtually pre-existing constants.)
1208 if (value_basic_type() != T_VOID) {
1209 if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) {
1210 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1211 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1212 // by the ArrayCopyNode.
1213 return phase->zerocon(value_basic_type());
1214 }
1215 } else {
1216 // TODO: materialize all-zero vector constant
1217 assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1218 }
1219 }
1220
1221 // A load from an initialization barrier can match a captured store.
1222 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1223 InitializeNode* init = st->in(0)->as_Initialize();
1224 AllocateNode* alloc = init->allocation();
1225 if ((alloc != nullptr) && (alloc == ld_alloc)) {
1226 // examine a captured store value
1227 st = init->find_captured_store(ld_off, memory_size(), phase);
1228 if (st != nullptr) {
1229 continue; // take one more trip around
1230 }
1231 }
1232 }
1233
1234 // Load boxed value from result of valueOf() call is input parameter.
1235 if (this->is_Load() && ld_adr->is_AddP() &&
1236 (tp != nullptr) && tp->is_ptr_to_boxed_value()) {
1237 intptr_t ignore = 0;
1238 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1239 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1240 base = bs->step_over_gc_barrier(base);
1241 if (base != nullptr && base->is_Proj() &&
1242 base->as_Proj()->_con == TypeFunc::Parms &&
1243 base->in(0)->is_CallStaticJava() &&
1244 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1245 return base->in(0)->in(TypeFunc::Parms);
1246 }
1247 }
1248
1249 break;
1250 }
1251
1252 return nullptr;
1253 }
1254
1255 //----------------------is_instance_field_load_with_local_phi------------------
1256 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1257 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1258 in(Address)->is_AddP() ) {
1259 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1260 // Only instances and boxed values.
1261 if( t_oop != nullptr &&
1262 (t_oop->is_ptr_to_boxed_value() ||
1263 t_oop->is_known_instance_field()) &&
1264 t_oop->offset() != Type::OffsetBot &&
1265 t_oop->offset() != Type::OffsetTop) {
1266 return true;
1267 }
1268 }
1269 return false;
1270 }
1271
1272 //------------------------------Identity---------------------------------------
1273 // Loads are identity if previous store is to same address
1274 Node* LoadNode::Identity(PhaseGVN* phase) {
1275 // If the previous store-maker is the right kind of Store, and the store is
1276 // to the same address, then we are equal to the value stored.
1277 Node* mem = in(Memory);
1278 Node* value = can_see_stored_value(mem, phase);
1279 if( value ) {
1280 // byte, short & char stores truncate naturally.
1281 // A load has to load the truncated value which requires
1282 // some sort of masking operation and that requires an
1283 // Ideal call instead of an Identity call.
1284 if (memory_size() < BytesPerInt) {
1285 // If the input to the store does not fit with the load's result type,
1286 // it must be truncated via an Ideal call.
1287 if (!phase->type(value)->higher_equal(phase->type(this)))
1288 return this;
1289 }
1290 // (This works even when value is a Con, but LoadNode::Value
1291 // usually runs first, producing the singleton type of the Con.)
1292 if (!has_pinned_control_dependency() || value->is_Con()) {
1293 return value;
1294 } else {
1295 return this;
1296 }
1297 }
1298
1299 if (has_pinned_control_dependency()) {
1300 return this;
1301 }
1302 // Search for an existing data phi which was generated before for the same
1303 // instance's field to avoid infinite generation of phis in a loop.
1304 Node *region = mem->in(0);
1305 if (is_instance_field_load_with_local_phi(region)) {
1306 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1307 int this_index = phase->C->get_alias_index(addr_t);
1308 int this_offset = addr_t->offset();
1309 int this_iid = addr_t->instance_id();
1310 if (!addr_t->is_known_instance() &&
1311 addr_t->is_ptr_to_boxed_value()) {
1312 // Use _idx of address base (could be Phi node) for boxed values.
1313 intptr_t ignore = 0;
1314 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1315 if (base == nullptr) {
1316 return this;
1317 }
1318 this_iid = base->_idx;
1319 }
1320 const Type* this_type = bottom_type();
1321 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1322 Node* phi = region->fast_out(i);
1323 if (phi->is_Phi() && phi != mem &&
1324 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1325 return phi;
1326 }
1327 }
1328 }
1329
1330 return this;
1331 }
1332
1333 // Construct an equivalent unsigned load.
1334 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1335 BasicType bt = T_ILLEGAL;
1336 const Type* rt = nullptr;
1337 switch (Opcode()) {
1338 case Op_LoadUB: return this;
1339 case Op_LoadUS: return this;
1340 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1341 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1342 default:
1343 assert(false, "no unsigned variant: %s", Name());
1344 return nullptr;
1345 }
1346 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1347 raw_adr_type(), rt, bt, _mo, _control_dependency,
1348 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1349 }
1350
1351 // Construct an equivalent signed load.
1352 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1353 BasicType bt = T_ILLEGAL;
1354 const Type* rt = nullptr;
1355 switch (Opcode()) {
1356 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1357 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1358 case Op_LoadB: // fall through
1359 case Op_LoadS: // fall through
1360 case Op_LoadI: // fall through
1361 case Op_LoadL: return this;
1362 default:
1363 assert(false, "no signed variant: %s", Name());
1364 return nullptr;
1365 }
1366 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1367 raw_adr_type(), rt, bt, _mo, _control_dependency,
1368 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1369 }
1370
1371 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1372 BasicType bt = rt->basic_type();
1373 switch (Opcode()) {
1374 case Op_LoadI: return (bt == T_FLOAT);
1375 case Op_LoadL: return (bt == T_DOUBLE);
1376 case Op_LoadF: return (bt == T_INT);
1377 case Op_LoadD: return (bt == T_LONG);
1378
1379 default: return false;
1380 }
1381 }
1382
1383 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1384 BasicType bt = rt->basic_type();
1385 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1386 bool is_mismatched = is_mismatched_access();
1387 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1388 if (raw_type == nullptr) {
1389 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1390 }
1391 const int op = Opcode();
1392 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) ||
1393 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access());
1394 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1395 raw_adr_type(), rt, bt, _mo, _control_dependency,
1396 require_atomic_access, is_unaligned_access(), is_mismatched);
1397 }
1398
1399 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1400 BasicType bt = vt->basic_type();
1401 switch (Opcode()) {
1402 case Op_StoreI: return (bt == T_FLOAT);
1403 case Op_StoreL: return (bt == T_DOUBLE);
1404 case Op_StoreF: return (bt == T_INT);
1405 case Op_StoreD: return (bt == T_LONG);
1406
1407 default: return false;
1408 }
1409 }
1410
1411 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1412 BasicType bt = vt->basic_type();
1413 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1414 const int op = Opcode();
1415 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) ||
1416 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access());
1417 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1418 raw_adr_type(), val, bt, _mo, require_atomic_access);
1419
1420 bool is_mismatched = is_mismatched_access();
1421 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1422 if (raw_type == nullptr) {
1423 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1424 }
1425 if (is_mismatched) {
1426 st->set_mismatched_access();
1427 }
1428 return st;
1429 }
1430
1431 // We're loading from an object which has autobox behaviour.
1432 // If this object is result of a valueOf call we'll have a phi
1433 // merging a newly allocated object and a load from the cache.
1434 // We want to replace this load with the original incoming
1435 // argument to the valueOf call.
1436 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1437 assert(igvn->C->eliminate_boxing(), "sanity");
1438 intptr_t ignore = 0;
1439 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1440 if ((base == nullptr) || base->is_Phi()) {
1441 // Push the loads from the phi that comes from valueOf up
1442 // through it to allow elimination of the loads and the recovery
1443 // of the original value. It is done in split_through_phi().
1444 return nullptr;
1445 } else if (base->is_Load() ||
1446 (base->is_DecodeN() && base->in(1)->is_Load())) {
1447 // Eliminate the load of boxed value for integer types from the cache
1448 // array by deriving the value from the index into the array.
1449 // Capture the offset of the load and then reverse the computation.
1450
1451 // Get LoadN node which loads a boxing object from 'cache' array.
1452 if (base->is_DecodeN()) {
1453 base = base->in(1);
1454 }
1455 if (!base->in(Address)->is_AddP()) {
1456 return nullptr; // Complex address
1457 }
1458 AddPNode* address = base->in(Address)->as_AddP();
1459 Node* cache_base = address->in(AddPNode::Base);
1460 if ((cache_base != nullptr) && cache_base->is_DecodeN()) {
1461 // Get ConP node which is static 'cache' field.
1462 cache_base = cache_base->in(1);
1463 }
1464 if ((cache_base != nullptr) && cache_base->is_Con()) {
1465 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1466 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
1467 Node* elements[4];
1468 int shift = exact_log2(type2aelembytes(T_OBJECT));
1469 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1470 if (count > 0 && elements[0]->is_Con() &&
1471 (count == 1 ||
1472 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1473 elements[1]->in(2) == igvn->intcon(shift)))) {
1474 ciObjArray* array = base_type->const_oop()->as_obj_array();
1475 // Fetch the box object cache[0] at the base of the array and get its value
1476 ciInstance* box = array->obj_at(0)->as_instance();
1477 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1478 assert(ik->is_box_klass(), "sanity");
1479 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1480 if (ik->nof_nonstatic_fields() == 1) {
1481 // This should be true nonstatic_field_at requires calling
1482 // nof_nonstatic_fields so check it anyway
1483 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1484 BasicType bt = c.basic_type();
1485 // Only integer types have boxing cache.
1486 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1487 bt == T_BYTE || bt == T_SHORT ||
1488 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1489 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1490 if (cache_low != (int)cache_low) {
1491 return nullptr; // should not happen since cache is array indexed by value
1492 }
1493 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1494 if (offset != (int)offset) {
1495 return nullptr; // should not happen since cache is array indexed by value
1496 }
1497 // Add up all the offsets making of the address of the load
1498 Node* result = elements[0];
1499 for (int i = 1; i < count; i++) {
1500 result = igvn->transform(new AddXNode(result, elements[i]));
1501 }
1502 // Remove the constant offset from the address and then
1503 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1504 // remove the scaling of the offset to recover the original index.
1505 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1506 // Peel the shift off directly but wrap it in a dummy node
1507 // since Ideal can't return existing nodes
1508 igvn->_worklist.push(result); // remove dead node later
1509 result = new RShiftXNode(result->in(1), igvn->intcon(0));
1510 } else if (result->is_Add() && result->in(2)->is_Con() &&
1511 result->in(1)->Opcode() == Op_LShiftX &&
1512 result->in(1)->in(2) == igvn->intcon(shift)) {
1513 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1514 // but for boxing cache access we know that X<<Z will not overflow
1515 // (there is range check) so we do this optimizatrion by hand here.
1516 igvn->_worklist.push(result); // remove dead node later
1517 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1518 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1519 } else {
1520 result = new RShiftXNode(result, igvn->intcon(shift));
1521 }
1522 #ifdef _LP64
1523 if (bt != T_LONG) {
1524 result = new ConvL2INode(igvn->transform(result));
1525 }
1526 #else
1527 if (bt == T_LONG) {
1528 result = new ConvI2LNode(igvn->transform(result));
1529 }
1530 #endif
1531 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1532 // Need to preserve unboxing load type if it is unsigned.
1533 switch(this->Opcode()) {
1534 case Op_LoadUB:
1535 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1536 break;
1537 case Op_LoadUS:
1538 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1539 break;
1540 }
1541 return result;
1542 }
1543 }
1544 }
1545 }
1546 }
1547 return nullptr;
1548 }
1549
1550 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1551 Node* region = phi->in(0);
1552 if (region == nullptr) {
1553 return false; // Wait stable graph
1554 }
1555 uint cnt = phi->req();
1556 for (uint i = 1; i < cnt; i++) {
1557 Node* rc = region->in(i);
1558 if (rc == nullptr || phase->type(rc) == Type::TOP)
1559 return false; // Wait stable graph
1560 Node* in = phi->in(i);
1561 if (in == nullptr || phase->type(in) == Type::TOP)
1562 return false; // Wait stable graph
1563 }
1564 return true;
1565 }
1566
1567 //------------------------------split_through_phi------------------------------
1568 // Check whether a call to 'split_through_phi' would split this load through the
1569 // Phi *base*. This method is essentially a copy of the validations performed
1570 // by 'split_through_phi'. The first use of this method was in EA code as part
1571 // of simplification of allocation merges.
1572 // Some differences from original method (split_through_phi):
1573 // - If base->is_CastPP(): base = base->in(1)
1574 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) {
1575 Node* mem = in(Memory);
1576 Node* address = in(Address);
1577 intptr_t ignore = 0;
1578 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1579
1580 if (base->is_CastPP()) {
1581 base = base->in(1);
1582 }
1583
1584 if (req() > 3 || base == nullptr || !base->is_Phi()) {
1585 return false;
1586 }
1587
1588 if (!mem->is_Phi()) {
1589 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1590 return false;
1591 }
1592 } else if (base->in(0) != mem->in(0)) {
1593 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1594 return false;
1595 }
1596 }
1597
1598 return true;
1599 }
1600
1601 //------------------------------split_through_phi------------------------------
1602 // Split instance or boxed field load through Phi.
1603 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) {
1604 if (req() > 3) {
1605 assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs");
1606 // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account
1607 return nullptr;
1608 }
1609 Node* mem = in(Memory);
1610 Node* address = in(Address);
1611 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1612
1613 assert((t_oop != nullptr) &&
1614 (ignore_missing_instance_id ||
1615 t_oop->is_known_instance_field() ||
1616 t_oop->is_ptr_to_boxed_value()), "invalid conditions");
1617
1618 Compile* C = phase->C;
1619 intptr_t ignore = 0;
1620 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1621 bool base_is_phi = (base != nullptr) && base->is_Phi();
1622 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1623 (base != nullptr) && (base == address->in(AddPNode::Base)) &&
1624 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1625
1626 if (!((mem->is_Phi() || base_is_phi) &&
1627 (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) {
1628 return nullptr; // Neither memory or base are Phi
1629 }
1630
1631 if (mem->is_Phi()) {
1632 if (!stable_phi(mem->as_Phi(), phase)) {
1633 return nullptr; // Wait stable graph
1634 }
1635 uint cnt = mem->req();
1636 // Check for loop invariant memory.
1637 if (cnt == 3) {
1638 for (uint i = 1; i < cnt; i++) {
1639 Node* in = mem->in(i);
1640 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1641 if (m == mem) {
1642 if (i == 1) {
1643 // if the first edge was a loop, check second edge too.
1644 // If both are replaceable - we are in an infinite loop
1645 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1646 if (n == mem) {
1647 break;
1648 }
1649 }
1650 set_req(Memory, mem->in(cnt - i));
1651 return this; // made change
1652 }
1653 }
1654 }
1655 }
1656 if (base_is_phi) {
1657 if (!stable_phi(base->as_Phi(), phase)) {
1658 return nullptr; // Wait stable graph
1659 }
1660 uint cnt = base->req();
1661 // Check for loop invariant memory.
1662 if (cnt == 3) {
1663 for (uint i = 1; i < cnt; i++) {
1664 if (base->in(i) == base) {
1665 return nullptr; // Wait stable graph
1666 }
1667 }
1668 }
1669 }
1670
1671 // Split through Phi (see original code in loopopts.cpp).
1672 assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type");
1673
1674 // Do nothing here if Identity will find a value
1675 // (to avoid infinite chain of value phis generation).
1676 if (this != Identity(phase)) {
1677 return nullptr;
1678 }
1679
1680 // Select Region to split through.
1681 Node* region;
1682 DomResult dom_result = DomResult::Dominate;
1683 if (!base_is_phi) {
1684 assert(mem->is_Phi(), "sanity");
1685 region = mem->in(0);
1686 // Skip if the region dominates some control edge of the address.
1687 // We will check `dom_result` later.
1688 dom_result = MemNode::maybe_all_controls_dominate(address, region);
1689 } else if (!mem->is_Phi()) {
1690 assert(base_is_phi, "sanity");
1691 region = base->in(0);
1692 // Skip if the region dominates some control edge of the memory.
1693 // We will check `dom_result` later.
1694 dom_result = MemNode::maybe_all_controls_dominate(mem, region);
1695 } else if (base->in(0) != mem->in(0)) {
1696 assert(base_is_phi && mem->is_Phi(), "sanity");
1697 dom_result = MemNode::maybe_all_controls_dominate(mem, base->in(0));
1698 if (dom_result == DomResult::Dominate) {
1699 region = base->in(0);
1700 } else {
1701 dom_result = MemNode::maybe_all_controls_dominate(address, mem->in(0));
1702 if (dom_result == DomResult::Dominate) {
1703 region = mem->in(0);
1704 }
1705 // Otherwise we encountered a complex graph.
1706 }
1707 } else {
1708 assert(base->in(0) == mem->in(0), "sanity");
1709 region = mem->in(0);
1710 }
1711
1712 PhaseIterGVN* igvn = phase->is_IterGVN();
1713 if (dom_result != DomResult::Dominate) {
1714 if (dom_result == DomResult::EncounteredDeadCode) {
1715 // There is some dead code which eventually will be removed in IGVN.
1716 // Once this is the case, we get an unambiguous dominance result.
1717 // Push the node to the worklist again until the dead code is removed.
1718 igvn->_worklist.push(this);
1719 }
1720 return nullptr;
1721 }
1722
1723 Node* phi = nullptr;
1724 const Type* this_type = this->bottom_type();
1725 if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) {
1726 int this_index = C->get_alias_index(t_oop);
1727 int this_offset = t_oop->offset();
1728 int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx;
1729 phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset);
1730 } else if (ignore_missing_instance_id) {
1731 phi = new PhiNode(region, this_type, nullptr, mem->_idx);
1732 } else {
1733 return nullptr;
1734 }
1735
1736 for (uint i = 1; i < region->req(); i++) {
1737 Node* x;
1738 Node* the_clone = nullptr;
1739 Node* in = region->in(i);
1740 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1741 in != nullptr && in->is_OuterStripMinedLoop()) {
1742 // No node should go in the outer strip mined loop
1743 in = in->in(LoopNode::EntryControl);
1744 }
1745 if (in == nullptr || in == C->top()) {
1746 x = C->top(); // Dead path? Use a dead data op
1747 } else {
1748 x = this->clone(); // Else clone up the data op
1749 the_clone = x; // Remember for possible deletion.
1750 // Alter data node to use pre-phi inputs
1751 if (this->in(0) == region) {
1752 x->set_req(0, in);
1753 } else {
1754 x->set_req(0, nullptr);
1755 }
1756 if (mem->is_Phi() && (mem->in(0) == region)) {
1757 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1758 }
1759 if (address->is_Phi() && address->in(0) == region) {
1760 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1761 }
1762 if (base_is_phi && (base->in(0) == region)) {
1763 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1764 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1765 x->set_req(Address, adr_x);
1766 }
1767 }
1768 // Check for a 'win' on some paths
1769 const Type *t = x->Value(igvn);
1770
1771 bool singleton = t->singleton();
1772
1773 // See comments in PhaseIdealLoop::split_thru_phi().
1774 if (singleton && t == Type::TOP) {
1775 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1776 }
1777
1778 if (singleton) {
1779 x = igvn->makecon(t);
1780 } else {
1781 // We now call Identity to try to simplify the cloned node.
1782 // Note that some Identity methods call phase->type(this).
1783 // Make sure that the type array is big enough for
1784 // our new node, even though we may throw the node away.
1785 // (This tweaking with igvn only works because x is a new node.)
1786 igvn->set_type(x, t);
1787 // If x is a TypeNode, capture any more-precise type permanently into Node
1788 // otherwise it will be not updated during igvn->transform since
1789 // igvn->type(x) is set to x->Value() already.
1790 x->raise_bottom_type(t);
1791 Node* y = x->Identity(igvn);
1792 if (y != x) {
1793 x = y;
1794 } else {
1795 y = igvn->hash_find_insert(x);
1796 if (y) {
1797 x = y;
1798 } else {
1799 // Else x is a new node we are keeping
1800 // We do not need register_new_node_with_optimizer
1801 // because set_type has already been called.
1802 igvn->_worklist.push(x);
1803 }
1804 }
1805 }
1806 if (x != the_clone && the_clone != nullptr) {
1807 igvn->remove_dead_node(the_clone);
1808 }
1809 phi->set_req(i, x);
1810 }
1811 // Record Phi
1812 igvn->register_new_node_with_optimizer(phi);
1813 return phi;
1814 }
1815
1816 AllocateNode* LoadNode::is_new_object_mark_load() const {
1817 if (Opcode() == Op_LoadX) {
1818 Node* address = in(MemNode::Address);
1819 AllocateNode* alloc = AllocateNode::Ideal_allocation(address);
1820 Node* mem = in(MemNode::Memory);
1821 if (alloc != nullptr && mem->is_Proj() &&
1822 mem->in(0) != nullptr &&
1823 mem->in(0) == alloc->initialization() &&
1824 alloc->initialization()->proj_out_or_null(0) != nullptr) {
1825 return alloc;
1826 }
1827 }
1828 return nullptr;
1829 }
1830
1831
1832 //------------------------------Ideal------------------------------------------
1833 // If the load is from Field memory and the pointer is non-null, it might be possible to
1834 // zero out the control input.
1835 // If the offset is constant and the base is an object allocation,
1836 // try to hook me up to the exact initializing store.
1837 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1838 if (has_pinned_control_dependency()) {
1839 return nullptr;
1840 }
1841 Node* p = MemNode::Ideal_common(phase, can_reshape);
1842 if (p) return (p == NodeSentinel) ? nullptr : p;
1843
1844 Node* ctrl = in(MemNode::Control);
1845 Node* address = in(MemNode::Address);
1846 bool progress = false;
1847
1848 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1849 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1850
1851 // Skip up past a SafePoint control. Cannot do this for Stores because
1852 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1853 if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint &&
1854 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1855 !addr_mark &&
1856 (depends_only_on_test() || has_unknown_control_dependency())) {
1857 ctrl = ctrl->in(0);
1858 set_req(MemNode::Control,ctrl);
1859 progress = true;
1860 }
1861
1862 intptr_t ignore = 0;
1863 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1864 if (base != nullptr
1865 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1866 // Check for useless control edge in some common special cases
1867 if (in(MemNode::Control) != nullptr
1868 && can_remove_control()
1869 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1870 && all_controls_dominate(base, phase->C->start())) {
1871 // A method-invariant, non-null address (constant or 'this' argument).
1872 set_req(MemNode::Control, nullptr);
1873 progress = true;
1874 }
1875 }
1876
1877 Node* mem = in(MemNode::Memory);
1878 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1879
1880 if (can_reshape && (addr_t != nullptr)) {
1881 // try to optimize our memory input
1882 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1883 if (opt_mem != mem) {
1884 set_req_X(MemNode::Memory, opt_mem, phase);
1885 if (phase->type( opt_mem ) == Type::TOP) return nullptr;
1886 return this;
1887 }
1888 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1889 if ((t_oop != nullptr) &&
1890 (t_oop->is_known_instance_field() ||
1891 t_oop->is_ptr_to_boxed_value())) {
1892 PhaseIterGVN *igvn = phase->is_IterGVN();
1893 assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true");
1894 if (igvn->_worklist.member(opt_mem)) {
1895 // Delay this transformation until memory Phi is processed.
1896 igvn->_worklist.push(this);
1897 return nullptr;
1898 }
1899 // Split instance field load through Phi.
1900 Node* result = split_through_phi(phase);
1901 if (result != nullptr) return result;
1902
1903 if (t_oop->is_ptr_to_boxed_value()) {
1904 Node* result = eliminate_autobox(igvn);
1905 if (result != nullptr) return result;
1906 }
1907 }
1908 }
1909
1910 // Is there a dominating load that loads the same value? Leave
1911 // anything that is not a load of a field/array element (like
1912 // barriers etc.) alone
1913 if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) {
1914 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1915 Node *use = mem->fast_out(i);
1916 if (use != this &&
1917 use->Opcode() == Opcode() &&
1918 use->in(0) != nullptr &&
1919 use->in(0) != in(0) &&
1920 use->in(Address) == in(Address)) {
1921 Node* ctl = in(0);
1922 for (int i = 0; i < 10 && ctl != nullptr; i++) {
1923 ctl = IfNode::up_one_dom(ctl);
1924 if (ctl == use->in(0)) {
1925 set_req(0, use->in(0));
1926 return this;
1927 }
1928 }
1929 }
1930 }
1931 }
1932
1933 // Check for prior store with a different base or offset; make Load
1934 // independent. Skip through any number of them. Bail out if the stores
1935 // are in an endless dead cycle and report no progress. This is a key
1936 // transform for Reflection. However, if after skipping through the Stores
1937 // we can't then fold up against a prior store do NOT do the transform as
1938 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1939 // array memory alive twice: once for the hoisted Load and again after the
1940 // bypassed Store. This situation only works if EVERYBODY who does
1941 // anti-dependence work knows how to bypass. I.e. we need all
1942 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1943 // the alias index stuff. So instead, peek through Stores and IFF we can
1944 // fold up, do so.
1945 Node* prev_mem = find_previous_store(phase);
1946 if (prev_mem != nullptr) {
1947 Node* value = can_see_arraycopy_value(prev_mem, phase);
1948 if (value != nullptr) {
1949 return value;
1950 }
1951 }
1952 // Steps (a), (b): Walk past independent stores to find an exact match.
1953 if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) {
1954 // (c) See if we can fold up on the spot, but don't fold up here.
1955 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1956 // just return a prior value, which is done by Identity calls.
1957 if (can_see_stored_value(prev_mem, phase)) {
1958 // Make ready for step (d):
1959 set_req_X(MemNode::Memory, prev_mem, phase);
1960 return this;
1961 }
1962 }
1963
1964 return progress ? this : nullptr;
1965 }
1966
1967 // Helper to recognize certain Klass fields which are invariant across
1968 // some group of array types (e.g., int[] or all T[] where T < Object).
1969 const Type*
1970 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1971 ciKlass* klass) const {
1972 assert(!UseCompactObjectHeaders || tkls->offset() != in_bytes(Klass::prototype_header_offset()),
1973 "must not happen");
1974
1975 if (tkls->isa_instklassptr() && tkls->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
1976 // The field is InstanceKlass::_access_flags. Return its (constant) value.
1977 assert(Opcode() == Op_LoadUS, "must load an unsigned short from _access_flags");
1978 ciInstanceKlass* iklass = tkls->is_instklassptr()->instance_klass();
1979 return TypeInt::make(iklass->access_flags());
1980 }
1981 if (tkls->offset() == in_bytes(Klass::misc_flags_offset())) {
1982 // The field is Klass::_misc_flags. Return its (constant) value.
1983 assert(Opcode() == Op_LoadUB, "must load an unsigned byte from _misc_flags");
1984 return TypeInt::make(klass->misc_flags());
1985 }
1986 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1987 // The field is Klass::_layout_helper. Return its constant value if known.
1988 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1989 return TypeInt::make(klass->layout_helper());
1990 }
1991
1992 // No match.
1993 return nullptr;
1994 }
1995
1996 //------------------------------Value-----------------------------------------
1997 const Type* LoadNode::Value(PhaseGVN* phase) const {
1998 // Either input is TOP ==> the result is TOP
1999 Node* mem = in(MemNode::Memory);
2000 const Type *t1 = phase->type(mem);
2001 if (t1 == Type::TOP) return Type::TOP;
2002 Node* adr = in(MemNode::Address);
2003 const TypePtr* tp = phase->type(adr)->isa_ptr();
2004 if (tp == nullptr || tp->empty()) return Type::TOP;
2005 int off = tp->offset();
2006 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
2007 Compile* C = phase->C;
2008
2009 // If load can see a previous constant store, use that.
2010 Node* value = can_see_stored_value(mem, phase);
2011 if (value != nullptr && value->is_Con()) {
2012 assert(value->bottom_type()->higher_equal(_type), "sanity");
2013 return value->bottom_type();
2014 }
2015
2016 // Try to guess loaded type from pointer type
2017 if (tp->isa_aryptr()) {
2018 const TypeAryPtr* ary = tp->is_aryptr();
2019 const Type* t = ary->elem();
2020
2021 // Determine whether the reference is beyond the header or not, by comparing
2022 // the offset against the offset of the start of the array's data.
2023 // Different array types begin at slightly different offsets (12 vs. 16).
2024 // We choose T_BYTE as an example base type that is least restrictive
2025 // as to alignment, which will therefore produce the smallest
2026 // possible base offset.
2027 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2028 const bool off_beyond_header = (off >= min_base_off);
2029
2030 // Try to constant-fold a stable array element.
2031 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
2032 // Make sure the reference is not into the header and the offset is constant
2033 ciObject* aobj = ary->const_oop();
2034 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
2035 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
2036 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
2037 stable_dimension,
2038 value_basic_type(), is_unsigned());
2039 if (con_type != nullptr) {
2040 return con_type;
2041 }
2042 }
2043 }
2044
2045 // Don't do this for integer types. There is only potential profit if
2046 // the element type t is lower than _type; that is, for int types, if _type is
2047 // more restrictive than t. This only happens here if one is short and the other
2048 // char (both 16 bits), and in those cases we've made an intentional decision
2049 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2050 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2051 //
2052 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2053 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
2054 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2055 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
2056 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2057 // In fact, that could have been the original type of p1, and p1 could have
2058 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2059 // expression (LShiftL quux 3) independently optimized to the constant 8.
2060 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2061 && (_type->isa_vect() == nullptr)
2062 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2063 // t might actually be lower than _type, if _type is a unique
2064 // concrete subclass of abstract class t.
2065 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
2066 const Type* jt = t->join_speculative(_type);
2067 // In any case, do not allow the join, per se, to empty out the type.
2068 if (jt->empty() && !t->empty()) {
2069 // This can happen if a interface-typed array narrows to a class type.
2070 jt = _type;
2071 }
2072 #ifdef ASSERT
2073 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2074 // The pointers in the autobox arrays are always non-null
2075 Node* base = adr->in(AddPNode::Base);
2076 if ((base != nullptr) && base->is_DecodeN()) {
2077 // Get LoadN node which loads IntegerCache.cache field
2078 base = base->in(1);
2079 }
2080 if ((base != nullptr) && base->is_Con()) {
2081 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2082 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2083 // It could be narrow oop
2084 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2085 }
2086 }
2087 }
2088 #endif
2089 return jt;
2090 }
2091 }
2092 } else if (tp->base() == Type::InstPtr) {
2093 assert( off != Type::OffsetBot ||
2094 // arrays can be cast to Objects
2095 !tp->isa_instptr() ||
2096 tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
2097 // unsafe field access may not have a constant offset
2098 C->has_unsafe_access(),
2099 "Field accesses must be precise" );
2100 // For oop loads, we expect the _type to be precise.
2101
2102 // Optimize loads from constant fields.
2103 const TypeInstPtr* tinst = tp->is_instptr();
2104 ciObject* const_oop = tinst->const_oop();
2105 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2106 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), value_basic_type());
2107 if (con_type != nullptr) {
2108 return con_type;
2109 }
2110 }
2111 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2112 assert(off != Type::OffsetBot ||
2113 !tp->isa_instklassptr() ||
2114 // arrays can be cast to Objects
2115 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2116 // also allow array-loading from the primary supertype
2117 // array during subtype checks
2118 Opcode() == Op_LoadKlass,
2119 "Field accesses must be precise");
2120 // For klass/static loads, we expect the _type to be precise
2121 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2122 /* With mirrors being an indirect in the Klass*
2123 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2124 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2125 *
2126 * So check the type and klass of the node before the LoadP.
2127 */
2128 Node* adr2 = adr->in(MemNode::Address);
2129 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2130 if (tkls != nullptr && !StressReflectiveCode) {
2131 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2132 ciKlass* klass = tkls->exact_klass();
2133 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2134 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2135 return TypeInstPtr::make(klass->java_mirror());
2136 }
2137 }
2138 }
2139
2140 const TypeKlassPtr *tkls = tp->isa_klassptr();
2141 if (tkls != nullptr) {
2142 if (tkls->is_loaded() && tkls->klass_is_exact()) {
2143 ciKlass* klass = tkls->exact_klass();
2144 // We are loading a field from a Klass metaobject whose identity
2145 // is known at compile time (the type is "exact" or "precise").
2146 // Check for fields we know are maintained as constants by the VM.
2147 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2148 // The field is Klass::_super_check_offset. Return its (constant) value.
2149 // (Folds up type checking code.)
2150 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2151 return TypeInt::make(klass->super_check_offset());
2152 }
2153 if (UseCompactObjectHeaders) {
2154 if (tkls->offset() == in_bytes(Klass::prototype_header_offset())) {
2155 // The field is Klass::_prototype_header. Return its (constant) value.
2156 assert(this->Opcode() == Op_LoadX, "must load a proper type from _prototype_header");
2157 return TypeX::make(klass->prototype_header());
2158 }
2159 }
2160 // Compute index into primary_supers array
2161 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2162 // Check for overflowing; use unsigned compare to handle the negative case.
2163 if( depth < ciKlass::primary_super_limit() ) {
2164 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2165 // (Folds up type checking code.)
2166 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2167 ciKlass *ss = klass->super_of_depth(depth);
2168 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2169 }
2170 const Type* aift = load_array_final_field(tkls, klass);
2171 if (aift != nullptr) return aift;
2172 }
2173
2174 // We can still check if we are loading from the primary_supers array at a
2175 // shallow enough depth. Even though the klass is not exact, entries less
2176 // than or equal to its super depth are correct.
2177 if (tkls->is_loaded()) {
2178 ciKlass* klass = nullptr;
2179 if (tkls->isa_instklassptr()) {
2180 klass = tkls->is_instklassptr()->instance_klass();
2181 } else {
2182 int dims;
2183 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims);
2184 if (inner->isa_instklassptr()) {
2185 klass = inner->is_instklassptr()->instance_klass();
2186 klass = ciObjArrayKlass::make(klass, dims);
2187 }
2188 }
2189 if (klass != nullptr) {
2190 // Compute index into primary_supers array
2191 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2192 // Check for overflowing; use unsigned compare to handle the negative case.
2193 if (depth < ciKlass::primary_super_limit() &&
2194 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case
2195 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2196 // (Folds up type checking code.)
2197 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2198 ciKlass *ss = klass->super_of_depth(depth);
2199 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2200 }
2201 }
2202 }
2203
2204 // If the type is enough to determine that the thing is not an array,
2205 // we can give the layout_helper a positive interval type.
2206 // This will help short-circuit some reflective code.
2207 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) &&
2208 tkls->isa_instklassptr() && // not directly typed as an array
2209 !tkls->is_instklassptr()->might_be_an_array() // not the supertype of all T[] (java.lang.Object) or has an interface that is not Serializable or Cloneable
2210 ) {
2211 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2212 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2213 // The key property of this type is that it folds up tests
2214 // for array-ness, since it proves that the layout_helper is positive.
2215 // Thus, a generic value like the basic object layout helper works fine.
2216 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2217 }
2218 }
2219
2220 // If we are loading from a freshly-allocated object/array, produce a zero.
2221 // Things to check:
2222 // 1. Load is beyond the header: headers are not guaranteed to be zero
2223 // 2. Load is not vectorized: vectors have no zero constant
2224 // 3. Load has no matching store, i.e. the input is the initial memory state
2225 const TypeOopPtr* tinst = tp->isa_oopptr();
2226 bool is_not_header = (tinst != nullptr) && tinst->is_known_instance_field();
2227 bool is_not_vect = (_type->isa_vect() == nullptr);
2228 if (is_not_header && is_not_vect) {
2229 Node* mem = in(MemNode::Memory);
2230 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2231 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2232 return Type::get_zero_type(_type->basic_type());
2233 }
2234 }
2235
2236 if (!UseCompactObjectHeaders) {
2237 Node* alloc = is_new_object_mark_load();
2238 if (alloc != nullptr) {
2239 return TypeX::make(markWord::prototype().value());
2240 }
2241 }
2242
2243 return _type;
2244 }
2245
2246 //------------------------------match_edge-------------------------------------
2247 // Do we Match on this edge index or not? Match only the address.
2248 uint LoadNode::match_edge(uint idx) const {
2249 return idx == MemNode::Address;
2250 }
2251
2252 //--------------------------LoadBNode::Ideal--------------------------------------
2253 //
2254 // If the previous store is to the same address as this load,
2255 // and the value stored was larger than a byte, replace this load
2256 // with the value stored truncated to a byte. If no truncation is
2257 // needed, the replacement is done in LoadNode::Identity().
2258 //
2259 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2260 Node* mem = in(MemNode::Memory);
2261 Node* value = can_see_stored_value(mem,phase);
2262 if (value != nullptr) {
2263 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2264 if (narrow != value) {
2265 return narrow;
2266 }
2267 }
2268 // Identity call will handle the case where truncation is not needed.
2269 return LoadNode::Ideal(phase, can_reshape);
2270 }
2271
2272 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2273 Node* mem = in(MemNode::Memory);
2274 Node* value = can_see_stored_value(mem,phase);
2275 if (value != nullptr && value->is_Con() &&
2276 !value->bottom_type()->higher_equal(_type)) {
2277 // If the input to the store does not fit with the load's result type,
2278 // it must be truncated. We can't delay until Ideal call since
2279 // a singleton Value is needed for split_thru_phi optimization.
2280 int con = value->get_int();
2281 return TypeInt::make((con << 24) >> 24);
2282 }
2283 return LoadNode::Value(phase);
2284 }
2285
2286 //--------------------------LoadUBNode::Ideal-------------------------------------
2287 //
2288 // If the previous store is to the same address as this load,
2289 // and the value stored was larger than a byte, replace this load
2290 // with the value stored truncated to a byte. If no truncation is
2291 // needed, the replacement is done in LoadNode::Identity().
2292 //
2293 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2294 Node* mem = in(MemNode::Memory);
2295 Node* value = can_see_stored_value(mem, phase);
2296 if (value != nullptr) {
2297 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2298 if (narrow != value) {
2299 return narrow;
2300 }
2301 }
2302 // Identity call will handle the case where truncation is not needed.
2303 return LoadNode::Ideal(phase, can_reshape);
2304 }
2305
2306 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2307 Node* mem = in(MemNode::Memory);
2308 Node* value = can_see_stored_value(mem,phase);
2309 if (value != nullptr && value->is_Con() &&
2310 !value->bottom_type()->higher_equal(_type)) {
2311 // If the input to the store does not fit with the load's result type,
2312 // it must be truncated. We can't delay until Ideal call since
2313 // a singleton Value is needed for split_thru_phi optimization.
2314 int con = value->get_int();
2315 return TypeInt::make(con & 0xFF);
2316 }
2317 return LoadNode::Value(phase);
2318 }
2319
2320 //--------------------------LoadUSNode::Ideal-------------------------------------
2321 //
2322 // If the previous store is to the same address as this load,
2323 // and the value stored was larger than a char, replace this load
2324 // with the value stored truncated to a char. If no truncation is
2325 // needed, the replacement is done in LoadNode::Identity().
2326 //
2327 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2328 Node* mem = in(MemNode::Memory);
2329 Node* value = can_see_stored_value(mem,phase);
2330 if (value != nullptr) {
2331 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2332 if (narrow != value) {
2333 return narrow;
2334 }
2335 }
2336 // Identity call will handle the case where truncation is not needed.
2337 return LoadNode::Ideal(phase, can_reshape);
2338 }
2339
2340 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2341 Node* mem = in(MemNode::Memory);
2342 Node* value = can_see_stored_value(mem,phase);
2343 if (value != nullptr && value->is_Con() &&
2344 !value->bottom_type()->higher_equal(_type)) {
2345 // If the input to the store does not fit with the load's result type,
2346 // it must be truncated. We can't delay until Ideal call since
2347 // a singleton Value is needed for split_thru_phi optimization.
2348 int con = value->get_int();
2349 return TypeInt::make(con & 0xFFFF);
2350 }
2351 return LoadNode::Value(phase);
2352 }
2353
2354 //--------------------------LoadSNode::Ideal--------------------------------------
2355 //
2356 // If the previous store is to the same address as this load,
2357 // and the value stored was larger than a short, replace this load
2358 // with the value stored truncated to a short. If no truncation is
2359 // needed, the replacement is done in LoadNode::Identity().
2360 //
2361 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2362 Node* mem = in(MemNode::Memory);
2363 Node* value = can_see_stored_value(mem,phase);
2364 if (value != nullptr) {
2365 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2366 if (narrow != value) {
2367 return narrow;
2368 }
2369 }
2370 // Identity call will handle the case where truncation is not needed.
2371 return LoadNode::Ideal(phase, can_reshape);
2372 }
2373
2374 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2375 Node* mem = in(MemNode::Memory);
2376 Node* value = can_see_stored_value(mem,phase);
2377 if (value != nullptr && value->is_Con() &&
2378 !value->bottom_type()->higher_equal(_type)) {
2379 // If the input to the store does not fit with the load's result type,
2380 // it must be truncated. We can't delay until Ideal call since
2381 // a singleton Value is needed for split_thru_phi optimization.
2382 int con = value->get_int();
2383 return TypeInt::make((con << 16) >> 16);
2384 }
2385 return LoadNode::Value(phase);
2386 }
2387
2388 //=============================================================================
2389 //----------------------------LoadKlassNode::make------------------------------
2390 // Polymorphic factory method:
2391 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2392 // sanity check the alias category against the created node type
2393 const TypePtr* adr_type = adr->bottom_type()->isa_ptr();
2394 assert(adr_type != nullptr, "expecting TypeKlassPtr");
2395 #ifdef _LP64
2396 if (adr_type->is_ptr_to_narrowklass()) {
2397 assert(UseCompressedClassPointers, "no compressed klasses");
2398 Node* load_klass = gvn.transform(new LoadNKlassNode(mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2399 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2400 }
2401 #endif
2402 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2403 return new LoadKlassNode(mem, adr, at, tk, MemNode::unordered);
2404 }
2405
2406 //------------------------------Value------------------------------------------
2407 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2408 return klass_value_common(phase);
2409 }
2410
2411 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2412 // Either input is TOP ==> the result is TOP
2413 const Type *t1 = phase->type( in(MemNode::Memory) );
2414 if (t1 == Type::TOP) return Type::TOP;
2415 Node *adr = in(MemNode::Address);
2416 const Type *t2 = phase->type( adr );
2417 if (t2 == Type::TOP) return Type::TOP;
2418 const TypePtr *tp = t2->is_ptr();
2419 if (TypePtr::above_centerline(tp->ptr()) ||
2420 tp->ptr() == TypePtr::Null) return Type::TOP;
2421
2422 // Return a more precise klass, if possible
2423 const TypeInstPtr *tinst = tp->isa_instptr();
2424 if (tinst != nullptr) {
2425 ciInstanceKlass* ik = tinst->instance_klass();
2426 int offset = tinst->offset();
2427 if (ik == phase->C->env()->Class_klass()
2428 && (offset == java_lang_Class::klass_offset() ||
2429 offset == java_lang_Class::array_klass_offset())) {
2430 // We are loading a special hidden field from a Class mirror object,
2431 // the field which points to the VM's Klass metaobject.
2432 ciType* t = tinst->java_mirror_type();
2433 // java_mirror_type returns non-null for compile-time Class constants.
2434 if (t != nullptr) {
2435 // constant oop => constant klass
2436 if (offset == java_lang_Class::array_klass_offset()) {
2437 if (t->is_void()) {
2438 // We cannot create a void array. Since void is a primitive type return null
2439 // klass. Users of this result need to do a null check on the returned klass.
2440 return TypePtr::NULL_PTR;
2441 }
2442 return TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);
2443 }
2444 if (!t->is_klass()) {
2445 // a primitive Class (e.g., int.class) has null for a klass field
2446 return TypePtr::NULL_PTR;
2447 }
2448 // Fold up the load of the hidden field
2449 return TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);
2450 }
2451 // non-constant mirror, so we can't tell what's going on
2452 }
2453 if (!tinst->is_loaded())
2454 return _type; // Bail out if not loaded
2455 if (offset == oopDesc::klass_offset_in_bytes()) {
2456 return tinst->as_klass_type(true);
2457 }
2458 }
2459
2460 // Check for loading klass from an array
2461 const TypeAryPtr *tary = tp->isa_aryptr();
2462 if (tary != nullptr &&
2463 tary->offset() == oopDesc::klass_offset_in_bytes()) {
2464 return tary->as_klass_type(true);
2465 }
2466
2467 // Check for loading klass from an array klass
2468 const TypeKlassPtr *tkls = tp->isa_klassptr();
2469 if (tkls != nullptr && !StressReflectiveCode) {
2470 if (!tkls->is_loaded())
2471 return _type; // Bail out if not loaded
2472 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2473 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2474 // // Always returning precise element type is incorrect,
2475 // // e.g., element type could be object and array may contain strings
2476 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2477
2478 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2479 // according to the element type's subclassing.
2480 return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2481 }
2482 if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() &&
2483 tkls->offset() == in_bytes(Klass::super_offset())) {
2484 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super();
2485 // The field is Klass::_super. Return its (constant) value.
2486 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2487 return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR;
2488 }
2489 }
2490
2491 if (tkls != nullptr && !UseSecondarySupersCache
2492 && tkls->offset() == in_bytes(Klass::secondary_super_cache_offset())) {
2493 // Treat Klass::_secondary_super_cache as a constant when the cache is disabled.
2494 return TypePtr::NULL_PTR;
2495 }
2496
2497 // Bailout case
2498 return LoadNode::Value(phase);
2499 }
2500
2501 //------------------------------Identity---------------------------------------
2502 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2503 // Also feed through the klass in Allocate(...klass...)._klass.
2504 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2505 return klass_identity_common(phase);
2506 }
2507
2508 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2509 Node* x = LoadNode::Identity(phase);
2510 if (x != this) return x;
2511
2512 // Take apart the address into an oop and offset.
2513 // Return 'this' if we cannot.
2514 Node* adr = in(MemNode::Address);
2515 intptr_t offset = 0;
2516 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2517 if (base == nullptr) return this;
2518 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2519 if (toop == nullptr) return this;
2520
2521 // Step over potential GC barrier for OopHandle resolve
2522 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2523 if (bs->is_gc_barrier_node(base)) {
2524 base = bs->step_over_gc_barrier(base);
2525 }
2526
2527 // We can fetch the klass directly through an AllocateNode.
2528 // This works even if the klass is not constant (clone or newArray).
2529 if (offset == oopDesc::klass_offset_in_bytes()) {
2530 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2531 if (allocated_klass != nullptr) {
2532 return allocated_klass;
2533 }
2534 }
2535
2536 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2537 // See inline_native_Class_query for occurrences of these patterns.
2538 // Java Example: x.getClass().isAssignableFrom(y)
2539 //
2540 // This improves reflective code, often making the Class
2541 // mirror go completely dead. (Current exception: Class
2542 // mirrors may appear in debug info, but we could clean them out by
2543 // introducing a new debug info operator for Klass.java_mirror).
2544
2545 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass()
2546 && offset == java_lang_Class::klass_offset()) {
2547 if (base->is_Load()) {
2548 Node* base2 = base->in(MemNode::Address);
2549 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2550 Node* adr2 = base2->in(MemNode::Address);
2551 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2552 if (tkls != nullptr && !tkls->empty()
2553 && (tkls->isa_instklassptr() || tkls->isa_aryklassptr())
2554 && adr2->is_AddP()
2555 ) {
2556 int mirror_field = in_bytes(Klass::java_mirror_offset());
2557 if (tkls->offset() == mirror_field) {
2558 #ifdef ASSERT
2559 const TypeKlassPtr* tkls2 = phase->type(adr2->in(AddPNode::Address))->is_klassptr();
2560 assert(tkls2->offset() == 0, "not a load of java_mirror");
2561 #endif
2562 assert(adr2->in(AddPNode::Base)->is_top(), "not an off heap load");
2563 assert(adr2->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(Klass::java_mirror_offset()), "incorrect offset");
2564 return adr2->in(AddPNode::Address);
2565 }
2566 }
2567 }
2568 }
2569 }
2570
2571 return this;
2572 }
2573
2574 LoadNode* LoadNode::clone_pinned() const {
2575 LoadNode* ld = clone()->as_Load();
2576 ld->_control_dependency = UnknownControl;
2577 return ld;
2578 }
2579
2580 // Pin a LoadNode if it carries a dependency on its control input. There are cases when the node
2581 // does not actually have any dependency on its control input. For example, if we have a LoadNode
2582 // being used only outside a loop but it must be scheduled inside the loop, we can clone the node
2583 // for each of its use so that all the clones can be scheduled outside the loop. Then, to prevent
2584 // the clones from being GVN-ed again, we add a control input for each of them at the loop exit. In
2585 // those case, since there is not a dependency between the node and its control input, we do not
2586 // need to pin it.
2587 LoadNode* LoadNode::pin_node_under_control_impl() const {
2588 const TypePtr* adr_type = this->adr_type();
2589 if (adr_type != nullptr && adr_type->isa_aryptr()) {
2590 // Only array accesses have dependencies on their control input
2591 return clone_pinned();
2592 }
2593 return nullptr;
2594 }
2595
2596 //------------------------------Value------------------------------------------
2597 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2598 const Type *t = klass_value_common(phase);
2599 if (t == Type::TOP)
2600 return t;
2601
2602 return t->make_narrowklass();
2603 }
2604
2605 //------------------------------Identity---------------------------------------
2606 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2607 // Also feed through the klass in Allocate(...klass...)._klass.
2608 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2609 Node *x = klass_identity_common(phase);
2610
2611 const Type *t = phase->type( x );
2612 if( t == Type::TOP ) return x;
2613 if( t->isa_narrowklass()) return x;
2614 assert (!t->isa_narrowoop(), "no narrow oop here");
2615
2616 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2617 }
2618
2619 //------------------------------Value-----------------------------------------
2620 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2621 // Either input is TOP ==> the result is TOP
2622 const Type *t1 = phase->type( in(MemNode::Memory) );
2623 if( t1 == Type::TOP ) return Type::TOP;
2624 Node *adr = in(MemNode::Address);
2625 const Type *t2 = phase->type( adr );
2626 if( t2 == Type::TOP ) return Type::TOP;
2627 const TypePtr *tp = t2->is_ptr();
2628 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2629 const TypeAryPtr *tap = tp->isa_aryptr();
2630 if( !tap ) return _type;
2631 return tap->size();
2632 }
2633
2634 //-------------------------------Ideal---------------------------------------
2635 // Feed through the length in AllocateArray(...length...)._length.
2636 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2637 Node* p = MemNode::Ideal_common(phase, can_reshape);
2638 if (p) return (p == NodeSentinel) ? nullptr : p;
2639
2640 // Take apart the address into an oop and offset.
2641 // Return 'this' if we cannot.
2642 Node* adr = in(MemNode::Address);
2643 intptr_t offset = 0;
2644 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2645 if (base == nullptr) return nullptr;
2646 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2647 if (tary == nullptr) return nullptr;
2648
2649 // We can fetch the length directly through an AllocateArrayNode.
2650 // This works even if the length is not constant (clone or newArray).
2651 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2652 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2653 if (alloc != nullptr) {
2654 Node* allocated_length = alloc->Ideal_length();
2655 Node* len = alloc->make_ideal_length(tary, phase);
2656 if (allocated_length != len) {
2657 // New CastII improves on this.
2658 return len;
2659 }
2660 }
2661 }
2662
2663 return nullptr;
2664 }
2665
2666 //------------------------------Identity---------------------------------------
2667 // Feed through the length in AllocateArray(...length...)._length.
2668 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2669 Node* x = LoadINode::Identity(phase);
2670 if (x != this) return x;
2671
2672 // Take apart the address into an oop and offset.
2673 // Return 'this' if we cannot.
2674 Node* adr = in(MemNode::Address);
2675 intptr_t offset = 0;
2676 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2677 if (base == nullptr) return this;
2678 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2679 if (tary == nullptr) return this;
2680
2681 // We can fetch the length directly through an AllocateArrayNode.
2682 // This works even if the length is not constant (clone or newArray).
2683 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2684 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2685 if (alloc != nullptr) {
2686 Node* allocated_length = alloc->Ideal_length();
2687 // Do not allow make_ideal_length to allocate a CastII node.
2688 Node* len = alloc->make_ideal_length(tary, phase, false);
2689 if (allocated_length == len) {
2690 // Return allocated_length only if it would not be improved by a CastII.
2691 return allocated_length;
2692 }
2693 }
2694 }
2695
2696 return this;
2697
2698 }
2699
2700 //=============================================================================
2701 //---------------------------StoreNode::make-----------------------------------
2702 // Polymorphic factory method:
2703 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) {
2704 assert((mo == unordered || mo == release), "unexpected");
2705 Compile* C = gvn.C;
2706 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2707 ctl != nullptr, "raw memory operations should have control edge");
2708
2709 switch (bt) {
2710 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2711 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2712 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2713 case T_CHAR:
2714 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2715 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2716 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2717 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2718 case T_METADATA:
2719 case T_ADDRESS:
2720 case T_OBJECT:
2721 #ifdef _LP64
2722 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2723 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2724 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2725 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2726 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2727 adr->bottom_type()->isa_rawptr())) {
2728 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2729 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2730 }
2731 #endif
2732 {
2733 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2734 }
2735 default:
2736 ShouldNotReachHere();
2737 return (StoreNode*)nullptr;
2738 }
2739 }
2740
2741 //--------------------------bottom_type----------------------------------------
2742 const Type *StoreNode::bottom_type() const {
2743 return Type::MEMORY;
2744 }
2745
2746 //------------------------------hash-------------------------------------------
2747 uint StoreNode::hash() const {
2748 // unroll addition of interesting fields
2749 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2750
2751 // Since they are not commoned, do not hash them:
2752 return NO_HASH;
2753 }
2754
2755 // Link together multiple stores (B/S/C/I) into a longer one.
2756 //
2757 // Example: _store = StoreB[i+3]
2758 //
2759 // RangeCheck[i+0] RangeCheck[i+0]
2760 // StoreB[i+0]
2761 // RangeCheck[i+3] RangeCheck[i+3]
2762 // StoreB[i+1] --> pass: fail:
2763 // StoreB[i+2] StoreI[i+0] StoreB[i+0]
2764 // StoreB[i+3]
2765 //
2766 // The 4 StoreB are merged into a single StoreI node. We have to be careful with RangeCheck[i+1]: before
2767 // the optimization, if this RangeCheck[i+1] fails, then we execute only StoreB[i+0], and then trap. After
2768 // the optimization, the new StoreI[i+0] is on the passing path of RangeCheck[i+3], and StoreB[i+0] on the
2769 // failing path.
2770 //
2771 // Note: For normal array stores, every store at first has a RangeCheck. But they can be removed with:
2772 // - RCE (RangeCheck Elimination): the RangeChecks in the loop are hoisted out and before the loop,
2773 // and possibly no RangeChecks remain between the stores.
2774 // - RangeCheck smearing: the earlier RangeChecks are adjusted such that they cover later RangeChecks,
2775 // and those later RangeChecks can be removed. Example:
2776 //
2777 // RangeCheck[i+0] RangeCheck[i+0] <- before first store
2778 // StoreB[i+0] StoreB[i+0] <- first store
2779 // RangeCheck[i+1] --> smeared --> RangeCheck[i+3] <- only RC between first and last store
2780 // StoreB[i+1] StoreB[i+1] <- second store
2781 // RangeCheck[i+2] --> removed
2782 // StoreB[i+2] StoreB[i+2]
2783 // RangeCheck[i+3] --> removed
2784 // StoreB[i+3] StoreB[i+3] <- last store
2785 //
2786 // Thus, it is a common pattern that between the first and last store in a chain
2787 // of adjacent stores there remains exactly one RangeCheck, located between the
2788 // first and the second store (e.g. RangeCheck[i+3]).
2789 //
2790 class MergePrimitiveStores : public StackObj {
2791 private:
2792 PhaseGVN* const _phase;
2793 StoreNode* const _store;
2794 // State machine with initial state Unknown
2795 // Allowed transitions:
2796 // Unknown -> Const
2797 // Unknown -> Platform
2798 // Unknown -> Reverse
2799 // Unknown -> NotAdjacent
2800 // Const -> Const
2801 // Const -> NotAdjacent
2802 // Platform -> Platform
2803 // Platform -> NotAdjacent
2804 // Reverse -> Reverse
2805 // Reverse -> NotAdjacent
2806 // NotAdjacent -> NotAdjacent
2807 enum ValueOrder : uint8_t {
2808 Unknown, // Initial state
2809 Const, // Input values are const
2810 Platform, // Platform order
2811 Reverse, // Reverse platform order
2812 NotAdjacent // Not adjacent
2813 };
2814 ValueOrder _value_order;
2815
2816 NOT_PRODUCT( const CHeapBitMap &_trace_tags; )
2817
2818 public:
2819 MergePrimitiveStores(PhaseGVN* phase, StoreNode* store) :
2820 _phase(phase), _store(store), _value_order(ValueOrder::Unknown)
2821 NOT_PRODUCT( COMMA _trace_tags(Compile::current()->directive()->trace_merge_stores_tags()) )
2822 {}
2823
2824 StoreNode* run();
2825
2826 private:
2827 bool is_compatible_store(const StoreNode* other_store) const;
2828 bool is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const;
2829 bool is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const;
2830 static bool is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase);
2831 enum CFGStatus { SuccessNoRangeCheck, SuccessWithRangeCheck, Failure };
2832 static CFGStatus cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store);
2833
2834 class Status {
2835 private:
2836 StoreNode* _found_store;
2837 bool _found_range_check;
2838
2839 Status(StoreNode* found_store, bool found_range_check)
2840 : _found_store(found_store), _found_range_check(found_range_check) {}
2841
2842 public:
2843 StoreNode* found_store() const { return _found_store; }
2844 bool found_range_check() const { return _found_range_check; }
2845 static Status make_failure() { return Status(nullptr, false); }
2846
2847 static Status make(StoreNode* found_store, const CFGStatus cfg_status) {
2848 if (cfg_status == CFGStatus::Failure) {
2849 return Status::make_failure();
2850 }
2851 return Status(found_store, cfg_status == CFGStatus::SuccessWithRangeCheck);
2852 }
2853
2854 #ifndef PRODUCT
2855 void print_on(outputStream* st) const {
2856 if (_found_store == nullptr) {
2857 st->print_cr("None");
2858 } else {
2859 st->print_cr("Found[%d %s, %s]", _found_store->_idx, _found_store->Name(),
2860 _found_range_check ? "RC" : "no-RC");
2861 }
2862 }
2863 #endif
2864 };
2865
2866 enum ValueOrder find_adjacent_input_value_order(const Node* n1, const Node* n2, const int memory_size) const;
2867 Status find_adjacent_use_store(const StoreNode* def_store) const;
2868 Status find_adjacent_def_store(const StoreNode* use_store) const;
2869 Status find_use_store(const StoreNode* def_store) const;
2870 Status find_def_store(const StoreNode* use_store) const;
2871 Status find_use_store_unidirectional(const StoreNode* def_store) const;
2872 Status find_def_store_unidirectional(const StoreNode* use_store) const;
2873
2874 void collect_merge_list(Node_List& merge_list) const;
2875 Node* make_merged_input_value(const Node_List& merge_list);
2876 StoreNode* make_merged_store(const Node_List& merge_list, Node* merged_input_value);
2877
2878 #ifndef PRODUCT
2879 // Access to TraceMergeStores tags
2880 bool is_trace(TraceMergeStores::Tag tag) const {
2881 return _trace_tags.at(tag);
2882 }
2883
2884 bool is_trace_basic() const {
2885 return is_trace(TraceMergeStores::Tag::BASIC);
2886 }
2887
2888 bool is_trace_pointer_parsing() const {
2889 return is_trace(TraceMergeStores::Tag::POINTER_PARSING);
2890 }
2891
2892 bool is_trace_pointer_aliasing() const {
2893 return is_trace(TraceMergeStores::Tag::POINTER_ALIASING);
2894 }
2895
2896 bool is_trace_pointer_adjacency() const {
2897 return is_trace(TraceMergeStores::Tag::POINTER_ADJACENCY);
2898 }
2899
2900 bool is_trace_success() const {
2901 return is_trace(TraceMergeStores::Tag::SUCCESS);
2902 }
2903 #endif
2904
2905 NOT_PRODUCT( void trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const; )
2906 };
2907
2908 StoreNode* MergePrimitiveStores::run() {
2909 // Check for B/S/C/I
2910 int opc = _store->Opcode();
2911 if (opc != Op_StoreB && opc != Op_StoreC && opc != Op_StoreI) {
2912 return nullptr;
2913 }
2914
2915 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] MergePrimitiveStores::run: "); _store->dump(); })
2916
2917 // The _store must be the "last" store in a chain. If we find a use we could merge with
2918 // then that use or a store further down is the "last" store.
2919 Status status_use = find_adjacent_use_store(_store);
2920 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect no use: "); status_use.print_on(tty); })
2921 if (status_use.found_store() != nullptr) {
2922 return nullptr;
2923 }
2924
2925 // Check if we can merge with at least one def, so that we have at least 2 stores to merge.
2926 Status status_def = find_adjacent_def_store(_store);
2927 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect def: "); status_def.print_on(tty); })
2928 Node* def_store = status_def.found_store();
2929 if (def_store == nullptr) {
2930 return nullptr;
2931 }
2932
2933 // Initialize value order
2934 _value_order = find_adjacent_input_value_order(def_store->in(MemNode::ValueIn),
2935 _store->in(MemNode::ValueIn),
2936 _store->memory_size());
2937 assert(_value_order != ValueOrder::NotAdjacent && _value_order != ValueOrder::Unknown, "Order should be checked");
2938
2939 ResourceMark rm;
2940 Node_List merge_list;
2941 collect_merge_list(merge_list);
2942
2943 Node* merged_input_value = make_merged_input_value(merge_list);
2944 if (merged_input_value == nullptr) { return nullptr; }
2945
2946 StoreNode* merged_store = make_merged_store(merge_list, merged_input_value);
2947
2948 NOT_PRODUCT( if (is_trace_success()) { trace(merge_list, merged_input_value, merged_store); } )
2949
2950 return merged_store;
2951 }
2952
2953 // Check compatibility between _store and other_store.
2954 bool MergePrimitiveStores::is_compatible_store(const StoreNode* other_store) const {
2955 int opc = _store->Opcode();
2956 assert(opc == Op_StoreB || opc == Op_StoreC || opc == Op_StoreI, "precondition");
2957
2958 if (other_store == nullptr ||
2959 _store->Opcode() != other_store->Opcode()) {
2960 return false;
2961 }
2962
2963 return true;
2964 }
2965
2966 bool MergePrimitiveStores::is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const {
2967 if (!is_adjacent_input_pair(def_store->in(MemNode::ValueIn),
2968 use_store->in(MemNode::ValueIn),
2969 def_store->memory_size())) {
2970 return false;
2971 }
2972
2973 ResourceMark rm;
2974 #ifndef PRODUCT
2975 const TraceMemPointer trace(is_trace_pointer_parsing(),
2976 is_trace_pointer_aliasing(),
2977 is_trace_pointer_adjacency(),
2978 true);
2979 #endif
2980 const MemPointer pointer_use(use_store NOT_PRODUCT(COMMA trace));
2981 const MemPointer pointer_def(def_store NOT_PRODUCT(COMMA trace));
2982 return pointer_def.is_adjacent_to_and_before(pointer_use);
2983 }
2984
2985 // Check input values n1 and n2 can be merged and return the value order
2986 MergePrimitiveStores::ValueOrder MergePrimitiveStores::find_adjacent_input_value_order(const Node* n1, const Node* n2,
2987 const int memory_size) const {
2988 // Pattern: [n1 = ConI, n2 = ConI]
2989 if (n1->Opcode() == Op_ConI && n2->Opcode() == Op_ConI) {
2990 return ValueOrder::Const;
2991 }
2992
2993 Node const *base_n2;
2994 jint shift_n2;
2995 if (!is_con_RShift(n2, base_n2, shift_n2, _phase)) {
2996 return ValueOrder::NotAdjacent;
2997 }
2998 Node const *base_n1;
2999 jint shift_n1;
3000 if (!is_con_RShift(n1, base_n1, shift_n1, _phase)) {
3001 return ValueOrder::NotAdjacent;
3002 }
3003
3004 int bits_per_store = memory_size * 8;
3005 if (base_n1 != base_n2 ||
3006 abs(shift_n1 - shift_n2) != bits_per_store ||
3007 shift_n1 % bits_per_store != 0) {
3008 // Values are not adjacent
3009 return ValueOrder::NotAdjacent;
3010 }
3011
3012 // Detect value order
3013 #ifdef VM_LITTLE_ENDIAN
3014 return shift_n1 < shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3015 : ValueOrder::Reverse; // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3016 #else
3017 return shift_n1 > shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3018 : ValueOrder::Reverse; // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3019 #endif
3020 }
3021
3022 bool MergePrimitiveStores::is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const {
3023 ValueOrder input_value_order = find_adjacent_input_value_order(n1, n2, memory_size);
3024
3025 switch (input_value_order) {
3026 case ValueOrder::NotAdjacent:
3027 return false;
3028 case ValueOrder::Reverse:
3029 if (memory_size != 1 ||
3030 !Matcher::match_rule_supported(Op_ReverseBytesS) ||
3031 !Matcher::match_rule_supported(Op_ReverseBytesI) ||
3032 !Matcher::match_rule_supported(Op_ReverseBytesL)) {
3033 // ReverseBytes are not supported by platform
3034 return false;
3035 }
3036 // fall-through.
3037 case ValueOrder::Const:
3038 case ValueOrder::Platform:
3039 if (_value_order == ValueOrder::Unknown) {
3040 // Initial state is Unknown, and we find a valid input value order
3041 return true;
3042 }
3043 // The value order can not be changed
3044 return _value_order == input_value_order;
3045 case ValueOrder::Unknown:
3046 default:
3047 ShouldNotReachHere();
3048 }
3049 return false;
3050 }
3051
3052 // Detect pattern: n = base_out >> shift_out
3053 bool MergePrimitiveStores::is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase) {
3054 assert(n != nullptr, "precondition");
3055
3056 int opc = n->Opcode();
3057 if (opc == Op_ConvL2I) {
3058 n = n->in(1);
3059 opc = n->Opcode();
3060 }
3061
3062 if ((opc == Op_RShiftI ||
3063 opc == Op_RShiftL ||
3064 opc == Op_URShiftI ||
3065 opc == Op_URShiftL) &&
3066 n->in(2)->is_ConI()) {
3067 base_out = n->in(1);
3068 shift_out = n->in(2)->get_int();
3069 // The shift must be positive:
3070 return shift_out >= 0;
3071 }
3072
3073 if (phase->type(n)->isa_int() != nullptr ||
3074 phase->type(n)->isa_long() != nullptr) {
3075 // (base >> 0)
3076 base_out = n;
3077 shift_out = 0;
3078 return true;
3079 }
3080 return false;
3081 }
3082
3083 // Check if there is nothing between the two stores, except optionally a RangeCheck leading to an uncommon trap.
3084 MergePrimitiveStores::CFGStatus MergePrimitiveStores::cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store) {
3085 assert(use_store->in(MemNode::Memory) == def_store, "use-def relationship");
3086
3087 Node* ctrl_use = use_store->in(MemNode::Control);
3088 Node* ctrl_def = def_store->in(MemNode::Control);
3089 if (ctrl_use == nullptr || ctrl_def == nullptr) {
3090 return CFGStatus::Failure;
3091 }
3092
3093 if (ctrl_use == ctrl_def) {
3094 // Same ctrl -> no RangeCheck in between.
3095 // Check: use_store must be the only use of def_store.
3096 if (def_store->outcnt() > 1) {
3097 return CFGStatus::Failure;
3098 }
3099 return CFGStatus::SuccessNoRangeCheck;
3100 }
3101
3102 // Different ctrl -> could have RangeCheck in between.
3103 // Check: 1. def_store only has these uses: use_store and MergeMem for uncommon trap, and
3104 // 2. ctrl separated by RangeCheck.
3105 if (def_store->outcnt() != 2) {
3106 return CFGStatus::Failure; // Cannot have exactly these uses: use_store and MergeMem for uncommon trap.
3107 }
3108 int use_store_out_idx = def_store->raw_out(0) == use_store ? 0 : 1;
3109 Node* merge_mem = def_store->raw_out(1 - use_store_out_idx)->isa_MergeMem();
3110 if (merge_mem == nullptr ||
3111 merge_mem->outcnt() != 1) {
3112 return CFGStatus::Failure; // Does not have MergeMem for uncommon trap.
3113 }
3114 if (!ctrl_use->is_IfProj() ||
3115 !ctrl_use->in(0)->is_RangeCheck() ||
3116 ctrl_use->in(0)->outcnt() != 2) {
3117 return CFGStatus::Failure; // Not RangeCheck.
3118 }
3119 IfProjNode* other_proj = ctrl_use->as_IfProj()->other_if_proj();
3120 Node* trap = other_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check);
3121 if (trap != merge_mem->unique_out() ||
3122 ctrl_use->in(0)->in(0) != ctrl_def) {
3123 return CFGStatus::Failure; // Not RangeCheck with merge_mem leading to uncommon trap.
3124 }
3125
3126 return CFGStatus::SuccessWithRangeCheck;
3127 }
3128
3129 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_use_store(const StoreNode* def_store) const {
3130 Status status_use = find_use_store(def_store);
3131 StoreNode* use_store = status_use.found_store();
3132 if (use_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3133 return Status::make_failure();
3134 }
3135 return status_use;
3136 }
3137
3138 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_def_store(const StoreNode* use_store) const {
3139 Status status_def = find_def_store(use_store);
3140 StoreNode* def_store = status_def.found_store();
3141 if (def_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3142 return Status::make_failure();
3143 }
3144 return status_def;
3145 }
3146
3147 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store(const StoreNode* def_store) const {
3148 Status status_use = find_use_store_unidirectional(def_store);
3149
3150 #ifdef ASSERT
3151 StoreNode* use_store = status_use.found_store();
3152 if (use_store != nullptr) {
3153 Status status_def = find_def_store_unidirectional(use_store);
3154 assert(status_def.found_store() == def_store &&
3155 status_def.found_range_check() == status_use.found_range_check(),
3156 "find_use_store and find_def_store must be symmetric");
3157 }
3158 #endif
3159
3160 return status_use;
3161 }
3162
3163 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store(const StoreNode* use_store) const {
3164 Status status_def = find_def_store_unidirectional(use_store);
3165
3166 #ifdef ASSERT
3167 StoreNode* def_store = status_def.found_store();
3168 if (def_store != nullptr) {
3169 Status status_use = find_use_store_unidirectional(def_store);
3170 assert(status_use.found_store() == use_store &&
3171 status_use.found_range_check() == status_def.found_range_check(),
3172 "find_use_store and find_def_store must be symmetric");
3173 }
3174 #endif
3175
3176 return status_def;
3177 }
3178
3179 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store_unidirectional(const StoreNode* def_store) const {
3180 assert(is_compatible_store(def_store), "precondition: must be compatible with _store");
3181
3182 for (DUIterator_Fast imax, i = def_store->fast_outs(imax); i < imax; i++) {
3183 StoreNode* use_store = def_store->fast_out(i)->isa_Store();
3184 if (is_compatible_store(use_store)) {
3185 return Status::make(use_store, cfg_status_for_pair(use_store, def_store));
3186 }
3187 }
3188
3189 return Status::make_failure();
3190 }
3191
3192 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store_unidirectional(const StoreNode* use_store) const {
3193 assert(is_compatible_store(use_store), "precondition: must be compatible with _store");
3194
3195 StoreNode* def_store = use_store->in(MemNode::Memory)->isa_Store();
3196 if (!is_compatible_store(def_store)) {
3197 return Status::make_failure();
3198 }
3199
3200 return Status::make(def_store, cfg_status_for_pair(use_store, def_store));
3201 }
3202
3203 void MergePrimitiveStores::collect_merge_list(Node_List& merge_list) const {
3204 // The merged store can be at most 8 bytes.
3205 const uint merge_list_max_size = 8 / _store->memory_size();
3206 assert(merge_list_max_size >= 2 &&
3207 merge_list_max_size <= 8 &&
3208 is_power_of_2(merge_list_max_size),
3209 "must be 2, 4 or 8");
3210
3211 // Traverse up the chain of adjacent def stores.
3212 StoreNode* current = _store;
3213 merge_list.push(current);
3214 while (current != nullptr && merge_list.size() < merge_list_max_size) {
3215 Status status = find_adjacent_def_store(current);
3216 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] find def: "); status.print_on(tty); })
3217
3218 current = status.found_store();
3219 if (current != nullptr) {
3220 merge_list.push(current);
3221
3222 // We can have at most one RangeCheck.
3223 if (status.found_range_check()) {
3224 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found RangeCheck, stop traversal."); })
3225 break;
3226 }
3227 }
3228 }
3229
3230 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found:"); merge_list.dump(); })
3231
3232 // Truncate the merge_list to a power of 2.
3233 const uint pow2size = round_down_power_of_2(merge_list.size());
3234 assert(pow2size >= 2, "must be merging at least 2 stores");
3235 while (merge_list.size() > pow2size) { merge_list.pop(); }
3236
3237 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] truncated:"); merge_list.dump(); })
3238 }
3239
3240 // Merge the input values of the smaller stores to a single larger input value.
3241 Node* MergePrimitiveStores::make_merged_input_value(const Node_List& merge_list) {
3242 int new_memory_size = _store->memory_size() * merge_list.size();
3243 Node* first = merge_list.at(merge_list.size()-1);
3244 Node* merged_input_value = nullptr;
3245 if (_store->in(MemNode::ValueIn)->Opcode() == Op_ConI) {
3246 assert(_value_order == ValueOrder::Const, "must match");
3247 // Pattern: [ConI, ConI, ...] -> new constant
3248 jlong con = 0;
3249 jlong bits_per_store = _store->memory_size() * 8;
3250 jlong mask = (((jlong)1) << bits_per_store) - 1;
3251 for (uint i = 0; i < merge_list.size(); i++) {
3252 jlong con_i = merge_list.at(i)->in(MemNode::ValueIn)->get_int();
3253 #ifdef VM_LITTLE_ENDIAN
3254 con = con << bits_per_store;
3255 con = con | (mask & con_i);
3256 #else // VM_LITTLE_ENDIAN
3257 con_i = (mask & con_i) << (i * bits_per_store);
3258 con = con | con_i;
3259 #endif // VM_LITTLE_ENDIAN
3260 }
3261 merged_input_value = _phase->longcon(con);
3262 } else {
3263 assert(_value_order == ValueOrder::Platform || _value_order == ValueOrder::Reverse, "must match");
3264 // Pattern: [base >> 24, base >> 16, base >> 8, base] -> base
3265 // | |
3266 // _store first
3267 //
3268 Node* hi = _store->in(MemNode::ValueIn);
3269 Node* lo = first->in(MemNode::ValueIn);
3270 #ifndef VM_LITTLE_ENDIAN
3271 // `_store` and `first` are swapped in the diagram above
3272 swap(hi, lo);
3273 #endif // !VM_LITTLE_ENDIAN
3274 if (_value_order == ValueOrder::Reverse) {
3275 swap(hi, lo);
3276 }
3277 Node const* hi_base;
3278 jint hi_shift;
3279 merged_input_value = lo;
3280 bool is_true = is_con_RShift(hi, hi_base, hi_shift, _phase);
3281 assert(is_true, "must detect con RShift");
3282 if (merged_input_value != hi_base && merged_input_value->Opcode() == Op_ConvL2I) {
3283 // look through
3284 merged_input_value = merged_input_value->in(1);
3285 }
3286 if (merged_input_value != hi_base) {
3287 // merged_input_value is not the base
3288 return nullptr;
3289 }
3290 }
3291
3292 if (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size <= 4) {
3293 // Example:
3294 //
3295 // long base = ...;
3296 // a[0] = (byte)(base >> 0);
3297 // a[1] = (byte)(base >> 8);
3298 //
3299 merged_input_value = _phase->transform(new ConvL2INode(merged_input_value));
3300 }
3301
3302 assert((_phase->type(merged_input_value)->isa_int() != nullptr && new_memory_size <= 4) ||
3303 (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size == 8),
3304 "merged_input_value is either int or long, and new_memory_size is small enough");
3305
3306 if (_value_order == ValueOrder::Reverse) {
3307 assert(_store->memory_size() == 1, "only implemented for bytes");
3308 if (new_memory_size == 8) {
3309 merged_input_value = _phase->transform(new ReverseBytesLNode(merged_input_value));
3310 } else if (new_memory_size == 4) {
3311 merged_input_value = _phase->transform(new ReverseBytesINode(merged_input_value));
3312 } else {
3313 assert(new_memory_size == 2, "sanity check");
3314 merged_input_value = _phase->transform(new ReverseBytesSNode(merged_input_value));
3315 }
3316 }
3317 return merged_input_value;
3318 }
3319
3320 // //
3321 // first_ctrl first_mem first_adr first_ctrl first_mem first_adr //
3322 // | | | | | | //
3323 // | | | | +---------------+ | //
3324 // | | | | | | | //
3325 // | | +---------+ | | +---------------+ //
3326 // | | | | | | | | //
3327 // +--------------+ | | v1 +------------------------------+ | | v1 //
3328 // | | | | | | | | | | | | //
3329 // RangeCheck first_store RangeCheck | | first_store //
3330 // | | | | | | | //
3331 // last_ctrl | +----> unc_trap last_ctrl | | +----> unc_trap //
3332 // | | ===> | | | //
3333 // +--------------+ | a2 v2 | | | //
3334 // | | | | | | | | //
3335 // | second_store | | | //
3336 // | | | | | [v1 v2 ... vn] //
3337 // ... ... | | | | //
3338 // | | | | | v //
3339 // +--------------+ | an vn +--------------+ | | merged_input_value //
3340 // | | | | | | | | //
3341 // last_store (= _store) merged_store //
3342 // //
3343 StoreNode* MergePrimitiveStores::make_merged_store(const Node_List& merge_list, Node* merged_input_value) {
3344 Node* first_store = merge_list.at(merge_list.size()-1);
3345 Node* last_ctrl = _store->in(MemNode::Control); // after (optional) RangeCheck
3346 Node* first_mem = first_store->in(MemNode::Memory);
3347 Node* first_adr = first_store->in(MemNode::Address);
3348
3349 const TypePtr* new_adr_type = _store->adr_type();
3350
3351 int new_memory_size = _store->memory_size() * merge_list.size();
3352 BasicType bt = T_ILLEGAL;
3353 switch (new_memory_size) {
3354 case 2: bt = T_SHORT; break;
3355 case 4: bt = T_INT; break;
3356 case 8: bt = T_LONG; break;
3357 }
3358
3359 StoreNode* merged_store = StoreNode::make(*_phase, last_ctrl, first_mem, first_adr,
3360 new_adr_type, merged_input_value, bt, MemNode::unordered);
3361
3362 // Marking the store mismatched is sufficient to prevent reordering, since array stores
3363 // are all on the same slice. Hence, we need no barriers.
3364 merged_store->set_mismatched_access();
3365
3366 // Constants above may now also be be packed -> put candidate on worklist
3367 _phase->is_IterGVN()->_worklist.push(first_mem);
3368
3369 return merged_store;
3370 }
3371
3372 #ifndef PRODUCT
3373 void MergePrimitiveStores::trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const {
3374 stringStream ss;
3375 ss.print_cr("[TraceMergeStores]: Replace");
3376 for (int i = (int)merge_list.size() - 1; i >= 0; i--) {
3377 merge_list.at(i)->dump("\n", false, &ss);
3378 }
3379 ss.print_cr("[TraceMergeStores]: with");
3380 merged_input_value->dump("\n", false, &ss);
3381 merged_store->dump("\n", false, &ss);
3382 tty->print("%s", ss.as_string());
3383 }
3384 #endif
3385
3386 //------------------------------Ideal------------------------------------------
3387 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
3388 // When a store immediately follows a relevant allocation/initialization,
3389 // try to capture it into the initialization, or hoist it above.
3390 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3391 Node* p = MemNode::Ideal_common(phase, can_reshape);
3392 if (p) return (p == NodeSentinel) ? nullptr : p;
3393
3394 Node* mem = in(MemNode::Memory);
3395 Node* address = in(MemNode::Address);
3396 Node* value = in(MemNode::ValueIn);
3397 // Back-to-back stores to same address? Fold em up. Generally
3398 // unsafe if I have intervening uses.
3399 {
3400 Node* st = mem;
3401 // If Store 'st' has more than one use, we cannot fold 'st' away.
3402 // For example, 'st' might be the final state at a conditional
3403 // return. Or, 'st' might be used by some node which is live at
3404 // the same time 'st' is live, which might be unschedulable. So,
3405 // require exactly ONE user until such time as we clone 'mem' for
3406 // each of 'mem's uses (thus making the exactly-1-user-rule hold
3407 // true).
3408 while (st->is_Store() && st->outcnt() == 1) {
3409 // Looking at a dead closed cycle of memory?
3410 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
3411 assert(Opcode() == st->Opcode() ||
3412 st->Opcode() == Op_StoreVector ||
3413 Opcode() == Op_StoreVector ||
3414 st->Opcode() == Op_StoreVectorScatter ||
3415 Opcode() == Op_StoreVectorScatter ||
3416 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
3417 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
3418 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
3419 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
3420 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
3421
3422 if (st->in(MemNode::Address)->eqv_uncast(address) &&
3423 st->as_Store()->memory_size() <= this->memory_size()) {
3424 Node* use = st->raw_out(0);
3425 if (phase->is_IterGVN()) {
3426 phase->is_IterGVN()->rehash_node_delayed(use);
3427 }
3428 // It's OK to do this in the parser, since DU info is always accurate,
3429 // and the parser always refers to nodes via SafePointNode maps.
3430 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
3431 return this;
3432 }
3433 st = st->in(MemNode::Memory);
3434 }
3435 }
3436
3437
3438 // Capture an unaliased, unconditional, simple store into an initializer.
3439 // Or, if it is independent of the allocation, hoist it above the allocation.
3440 if (ReduceFieldZeroing && /*can_reshape &&*/
3441 mem->is_Proj() && mem->in(0)->is_Initialize()) {
3442 InitializeNode* init = mem->in(0)->as_Initialize();
3443 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
3444 if (offset > 0) {
3445 Node* moved = init->capture_store(this, offset, phase, can_reshape);
3446 // If the InitializeNode captured me, it made a raw copy of me,
3447 // and I need to disappear.
3448 if (moved != nullptr) {
3449 // %%% hack to ensure that Ideal returns a new node:
3450 mem = MergeMemNode::make(mem);
3451 return mem; // fold me away
3452 }
3453 }
3454 }
3455
3456 // Fold reinterpret cast into memory operation:
3457 // StoreX mem (MoveY2X v) => StoreY mem v
3458 if (value->is_Move()) {
3459 const Type* vt = value->in(1)->bottom_type();
3460 if (has_reinterpret_variant(vt)) {
3461 if (phase->C->post_loop_opts_phase()) {
3462 return convert_to_reinterpret_store(*phase, value->in(1), vt);
3463 } else {
3464 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
3465 }
3466 }
3467 }
3468
3469 if (MergeStores && UseUnalignedAccesses) {
3470 if (phase->C->merge_stores_phase()) {
3471 MergePrimitiveStores merge(phase, this);
3472 Node* progress = merge.run();
3473 if (progress != nullptr) { return progress; }
3474 } else {
3475 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
3476 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
3477 // the stores, and we merge the wrong sequence of stores.
3478 // Example:
3479 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
3480 // Apply MergeStores:
3481 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
3482 // Remove more RangeChecks:
3483 // StoreI [ StoreL ] StoreI
3484 // But now it would have been better to do this instead:
3485 // [ StoreL ] [ StoreL ]
3486 phase->C->record_for_merge_stores_igvn(this);
3487 }
3488 }
3489
3490 return nullptr; // No further progress
3491 }
3492
3493 //------------------------------Value-----------------------------------------
3494 const Type* StoreNode::Value(PhaseGVN* phase) const {
3495 // Either input is TOP ==> the result is TOP
3496 const Type *t1 = phase->type( in(MemNode::Memory) );
3497 if( t1 == Type::TOP ) return Type::TOP;
3498 const Type *t2 = phase->type( in(MemNode::Address) );
3499 if( t2 == Type::TOP ) return Type::TOP;
3500 const Type *t3 = phase->type( in(MemNode::ValueIn) );
3501 if( t3 == Type::TOP ) return Type::TOP;
3502 return Type::MEMORY;
3503 }
3504
3505 //------------------------------Identity---------------------------------------
3506 // Remove redundant stores:
3507 // Store(m, p, Load(m, p)) changes to m.
3508 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
3509 Node* StoreNode::Identity(PhaseGVN* phase) {
3510 Node* mem = in(MemNode::Memory);
3511 Node* adr = in(MemNode::Address);
3512 Node* val = in(MemNode::ValueIn);
3513
3514 Node* result = this;
3515
3516 // Load then Store? Then the Store is useless
3517 if (val->is_Load() &&
3518 val->in(MemNode::Address)->eqv_uncast(adr) &&
3519 val->in(MemNode::Memory )->eqv_uncast(mem) &&
3520 val->as_Load()->store_Opcode() == Opcode()) {
3521 // Ensure vector type is the same
3522 if (!is_StoreVector() || (mem->is_LoadVector() && as_StoreVector()->vect_type() == mem->as_LoadVector()->vect_type())) {
3523 result = mem;
3524 }
3525 }
3526
3527 // Two stores in a row of the same value?
3528 if (result == this &&
3529 mem->is_Store() &&
3530 mem->in(MemNode::Address)->eqv_uncast(adr) &&
3531 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
3532 mem->Opcode() == Opcode()) {
3533 if (!is_StoreVector()) {
3534 result = mem;
3535 } else {
3536 const StoreVectorNode* store_vector = as_StoreVector();
3537 const StoreVectorNode* mem_vector = mem->as_StoreVector();
3538 const Node* store_indices = store_vector->indices();
3539 const Node* mem_indices = mem_vector->indices();
3540 const Node* store_mask = store_vector->mask();
3541 const Node* mem_mask = mem_vector->mask();
3542 // Ensure types, indices, and masks match
3543 if (store_vector->vect_type() == mem_vector->vect_type() &&
3544 ((store_indices == nullptr) == (mem_indices == nullptr) &&
3545 (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) &&
3546 ((store_mask == nullptr) == (mem_mask == nullptr) &&
3547 (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) {
3548 result = mem;
3549 }
3550 }
3551 }
3552
3553 // Store of zero anywhere into a freshly-allocated object?
3554 // Then the store is useless.
3555 // (It must already have been captured by the InitializeNode.)
3556 if (result == this &&
3557 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
3558 // a newly allocated object is already all-zeroes everywhere
3559 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
3560 result = mem;
3561 }
3562
3563 if (result == this) {
3564 // the store may also apply to zero-bits in an earlier object
3565 Node* prev_mem = find_previous_store(phase);
3566 // Steps (a), (b): Walk past independent stores to find an exact match.
3567 if (prev_mem != nullptr) {
3568 Node* prev_val = can_see_stored_value(prev_mem, phase);
3569 if (prev_val != nullptr && prev_val == val) {
3570 // prev_val and val might differ by a cast; it would be good
3571 // to keep the more informative of the two.
3572 result = mem;
3573 }
3574 }
3575 }
3576 }
3577
3578 PhaseIterGVN* igvn = phase->is_IterGVN();
3579 if (result != this && igvn != nullptr) {
3580 MemBarNode* trailing = trailing_membar();
3581 if (trailing != nullptr) {
3582 #ifdef ASSERT
3583 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
3584 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects");
3585 #endif
3586 trailing->remove(igvn);
3587 }
3588 }
3589
3590 return result;
3591 }
3592
3593 //------------------------------match_edge-------------------------------------
3594 // Do we Match on this edge index or not? Match only memory & value
3595 uint StoreNode::match_edge(uint idx) const {
3596 return idx == MemNode::Address || idx == MemNode::ValueIn;
3597 }
3598
3599 //------------------------------cmp--------------------------------------------
3600 // Do not common stores up together. They generally have to be split
3601 // back up anyways, so do not bother.
3602 bool StoreNode::cmp( const Node &n ) const {
3603 return (&n == this); // Always fail except on self
3604 }
3605
3606 //------------------------------Ideal_masked_input-----------------------------
3607 // Check for a useless mask before a partial-word store
3608 // (StoreB ... (AndI valIn conIa) )
3609 // If (conIa & mask == mask) this simplifies to
3610 // (StoreB ... (valIn) )
3611 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
3612 Node *val = in(MemNode::ValueIn);
3613 if( val->Opcode() == Op_AndI ) {
3614 const TypeInt *t = phase->type( val->in(2) )->isa_int();
3615 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
3616 set_req_X(MemNode::ValueIn, val->in(1), phase);
3617 return this;
3618 }
3619 }
3620 return nullptr;
3621 }
3622
3623
3624 //------------------------------Ideal_sign_extended_input----------------------
3625 // Check for useless sign-extension before a partial-word store
3626 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
3627 // If (conIL == conIR && conIR <= num_rejected_bits) this simplifies to
3628 // (StoreB ... (v))
3629 // If (conIL > conIR) under some conditions, it can be simplified into
3630 // (StoreB ... (LShiftI _ v (conIL - conIR)))
3631 // This case happens when the value of the store was itself a left shift, that
3632 // gets merged into the inner left shift of the sign-extension. For instance,
3633 // if we have
3634 // array_of_shorts[0] = (short)(v << 2)
3635 // We get a structure such as:
3636 // (StoreB ... (RShiftI _ (LShiftI _ (LShiftI _ v 2) 16) 16))
3637 // that is simplified into
3638 // (StoreB ... (RShiftI _ (LShiftI _ v 18) 16)).
3639 // It is thus useful to handle cases where conIL > conIR. But this simplification
3640 // does not always hold. Let's see in which cases it's valid.
3641 //
3642 // Let's assume we have the following 32 bits integer v:
3643 // +----------------------------------+
3644 // | v[0..31] |
3645 // +----------------------------------+
3646 // 31 0
3647 // that will be stuffed in 8 bits byte after a shift left and a shift right of
3648 // potentially different magnitudes.
3649 // We denote num_rejected_bits the number of bits of the discarded part. In this
3650 // case, num_rejected_bits == 24.
3651 //
3652 // Statement (proved further below in case analysis):
3653 // Given:
3654 // - 0 <= conIL < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
3655 // - 0 <= conIR < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
3656 // - conIL >= conIR
3657 // - num_rejected_bits >= conIR
3658 // Then this form:
3659 // (RShiftI _ (LShiftI _ v conIL) conIR)
3660 // can be replaced with this form:
3661 // (LShiftI _ v (conIL-conIR))
3662 //
3663 // Note: We only have to show that the non-rejected lowest bits (8 bits for byte)
3664 // have to be correct, as the higher bits are rejected / truncated by the store.
3665 //
3666 // The hypotheses
3667 // 0 <= conIL < BitsPerJavaInteger
3668 // 0 <= conIR < BitsPerJavaInteger
3669 // are ensured by maskShiftAmount (called from ::Ideal of shift nodes). Indeed,
3670 // (v << 31) << 2 must be simplified into 0, not into v << 33 (which is equivalent
3671 // to v << 1).
3672 //
3673 //
3674 // If you don't like case analysis, jump after the conclusion.
3675 // ### Case 1 : conIL == conIR
3676 // ###### Case 1.1: conIL == conIR == num_rejected_bits
3677 // If we do the shift left then right by 24 bits, we get:
3678 // after: v << 24
3679 // +---------+------------------------+
3680 // | v[0..7] | 0 |
3681 // +---------+------------------------+
3682 // 31 24 23 0
3683 // after: (v << 24) >> 24
3684 // +------------------------+---------+
3685 // | sign bit | v[0..7] |
3686 // +------------------------+---------+
3687 // 31 8 7 0
3688 // The non-rejected bits (bits kept by the store, that is the 8 lower bits of the
3689 // result) are the same before and after, so, indeed, simplifying is correct.
3690
3691 // ###### Case 1.2: conIL == conIR < num_rejected_bits
3692 // If we do the shift left then right by 22 bits, we get:
3693 // after: v << 22
3694 // +---------+------------------------+
3695 // | v[0..9] | 0 |
3696 // +---------+------------------------+
3697 // 31 22 21 0
3698 // after: (v << 22) >> 22
3699 // +------------------------+---------+
3700 // | sign bit | v[0..9] |
3701 // +------------------------+---------+
3702 // 31 10 9 0
3703 // The non-rejected bits are the 8 lower bits of v. The bits 8 and 9 of v are still
3704 // present in (v << 22) >> 22 but will be dropped by the store. The simplification is
3705 // still correct.
3706
3707 // ###### But! Case 1.3: conIL == conIR > num_rejected_bits
3708 // If we do the shift left then right by 26 bits, we get:
3709 // after: v << 26
3710 // +---------+------------------------+
3711 // | v[0..5] | 0 |
3712 // +---------+------------------------+
3713 // 31 26 25 0
3714 // after: (v << 26) >> 26
3715 // +------------------------+---------+
3716 // | sign bit | v[0..5] |
3717 // +------------------------+---------+
3718 // 31 6 5 0
3719 // The non-rejected bits are made of
3720 // - 0-5 => the bits 0 to 5 of v
3721 // - 6-7 => the sign bit of v[0..5] (that is v[5])
3722 // Simplifying this as v is not correct.
3723 // The condition conIR <= num_rejected_bits is indeed necessary in Case 1
3724 //
3725 // ### Case 2: conIL > conIR
3726 // ###### Case 2.1: num_rejected_bits == conIR
3727 // We take conIL == 26 for this example.
3728 // after: v << 26
3729 // +---------+------------------------+
3730 // | v[0..5] | 0 |
3731 // +---------+------------------------+
3732 // 31 26 25 0
3733 // after: (v << 26) >> 24
3734 // +------------------+---------+-----+
3735 // | sign bit | v[0..5] | 0 |
3736 // +------------------+---------+-----+
3737 // 31 8 7 2 1 0
3738 // The non-rejected bits are the 8 lower ones of (v << conIL - conIR).
3739 // The bits 6 and 7 of v have been thrown away after the shift left.
3740 // The simplification is still correct.
3741 //
3742 // ###### Case 2.2: num_rejected_bits > conIR.
3743 // Let's say conIL == 26 and conIR == 22.
3744 // after: v << 26
3745 // +---------+------------------------+
3746 // | v[0..5] | 0 |
3747 // +---------+------------------------+
3748 // 31 26 25 0
3749 // after: (v << 26) >> 22
3750 // +------------------+---------+-----+
3751 // | sign bit | v[0..5] | 0 |
3752 // +------------------+---------+-----+
3753 // 31 10 9 4 3 0
3754 // The bits non-rejected by the store are exactly the 8 lower ones of (v << (conIL - conIR)):
3755 // - 0-3 => 0
3756 // - 4-7 => bits 0 to 3 of v
3757 // The simplification is still correct.
3758 // The bits 4 and 5 of v are still present in (v << (conIL - conIR)) but they don't
3759 // matter as they are not in the 8 lower bits: they will be cut out by the store.
3760 //
3761 // ###### But! Case 2.3: num_rejected_bits < conIR.
3762 // Let's see that this case is not as easy to simplify.
3763 // Let's say conIL == 28 and conIR == 26.
3764 // after: v << 28
3765 // +---------+------------------------+
3766 // | v[0..3] | 0 |
3767 // +---------+------------------------+
3768 // 31 28 27 0
3769 // after: (v << 28) >> 26
3770 // +------------------+---------+-----+
3771 // | sign bit | v[0..3] | 0 |
3772 // +------------------+---------+-----+
3773 // 31 6 5 2 1 0
3774 // The non-rejected bits are made of
3775 // - 0-1 => 0
3776 // - 2-5 => the bits 0 to 3 of v
3777 // - 6-7 => the sign bit of v[0..3] (that is v[3])
3778 // Simplifying this as (v << 2) is not correct.
3779 // The condition conIR <= num_rejected_bits is indeed necessary in Case 2.
3780 //
3781 // ### Conclusion:
3782 // Our hypotheses are indeed sufficient:
3783 // - 0 <= conIL < BitsPerJavaInteger
3784 // - 0 <= conIR < BitsPerJavaInteger
3785 // - conIL >= conIR
3786 // - num_rejected_bits >= conIR
3787 //
3788 // ### A rationale without case analysis:
3789 // After the shift left, conIL upper bits of v are discarded and conIL lower bit
3790 // zeroes are added. After the shift right, conIR lower bits of the previous result
3791 // are discarded. If conIL >= conIR, we discard only the zeroes we made up during
3792 // the shift left, but if conIL < conIR, then we discard also lower bits of v. But
3793 // the point of the simplification is to get an expression of the form
3794 // (v << (conIL - conIR)). This expression discard only higher bits of v, thus the
3795 // simplification is not correct if conIL < conIR.
3796 //
3797 // Moreover, after the shift right, the higher bit of (v << conIL) is repeated on the
3798 // conIR higher bits of ((v << conIL) >> conIR), it's the sign-extension. If
3799 // conIR > num_rejected_bits, then at least one artificial copy of this sign bit will
3800 // be in the window of the store. Thus ((v << conIL) >> conIR) is not equivalent to
3801 // (v << (conIL-conIR)) if conIR > num_rejected_bits.
3802 //
3803 // We do not treat the case conIL < conIR here since the point of this function is
3804 // to skip sign-extensions (that is conIL == conIR == num_rejected_bits). The need
3805 // of treating conIL > conIR comes from the cases where the sign-extended value is
3806 // also left-shift expression. Computing the sign-extension of a right-shift expression
3807 // doesn't yield a situation such as
3808 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
3809 // where conIL < conIR.
3810 Node* StoreNode::Ideal_sign_extended_input(PhaseGVN* phase, int num_rejected_bits) {
3811 Node* shr = in(MemNode::ValueIn);
3812 if (shr->Opcode() == Op_RShiftI) {
3813 const TypeInt* conIR = phase->type(shr->in(2))->isa_int();
3814 if (conIR != nullptr && conIR->is_con() && conIR->get_con() >= 0 && conIR->get_con() < BitsPerJavaInteger && conIR->get_con() <= num_rejected_bits) {
3815 Node* shl = shr->in(1);
3816 if (shl->Opcode() == Op_LShiftI) {
3817 const TypeInt* conIL = phase->type(shl->in(2))->isa_int();
3818 if (conIL != nullptr && conIL->is_con() && conIL->get_con() >= 0 && conIL->get_con() < BitsPerJavaInteger) {
3819 if (conIL->get_con() == conIR->get_con()) {
3820 set_req_X(MemNode::ValueIn, shl->in(1), phase);
3821 return this;
3822 }
3823 if (conIL->get_con() > conIR->get_con()) {
3824 Node* new_shl = phase->transform(new LShiftINode(shl->in(1), phase->intcon(conIL->get_con() - conIR->get_con())));
3825 set_req_X(MemNode::ValueIn, new_shl, phase);
3826 return this;
3827 }
3828 }
3829 }
3830 }
3831 }
3832 return nullptr;
3833 }
3834
3835 //------------------------------value_never_loaded-----------------------------------
3836 // Determine whether there are any possible loads of the value stored.
3837 // For simplicity, we actually check if there are any loads from the
3838 // address stored to, not just for loads of the value stored by this node.
3839 //
3840 bool StoreNode::value_never_loaded(PhaseValues* phase) const {
3841 Node *adr = in(Address);
3842 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
3843 if (adr_oop == nullptr)
3844 return false;
3845 if (!adr_oop->is_known_instance_field())
3846 return false; // if not a distinct instance, there may be aliases of the address
3847 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
3848 Node *use = adr->fast_out(i);
3849 if (use->is_Load() || use->is_LoadStore()) {
3850 return false;
3851 }
3852 }
3853 return true;
3854 }
3855
3856 MemBarNode* StoreNode::trailing_membar() const {
3857 if (is_release()) {
3858 MemBarNode* trailing_mb = nullptr;
3859 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3860 Node* u = fast_out(i);
3861 if (u->is_MemBar()) {
3862 if (u->as_MemBar()->trailing_store()) {
3863 assert(u->Opcode() == Op_MemBarVolatile, "");
3864 assert(trailing_mb == nullptr, "only one");
3865 trailing_mb = u->as_MemBar();
3866 #ifdef ASSERT
3867 Node* leading = u->as_MemBar()->leading_membar();
3868 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3869 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
3870 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
3871 #endif
3872 } else {
3873 assert(u->as_MemBar()->standalone(), "");
3874 }
3875 }
3876 }
3877 return trailing_mb;
3878 }
3879 return nullptr;
3880 }
3881
3882
3883 //=============================================================================
3884 //------------------------------Ideal------------------------------------------
3885 // If the store is from an AND mask that leaves the low bits untouched, then
3886 // we can skip the AND operation. If the store is from a sign-extension
3887 // (a left shift, then right shift) we can skip both.
3888 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
3889 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
3890 if( progress != nullptr ) return progress;
3891
3892 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
3893 if( progress != nullptr ) return progress;
3894
3895 // Finally check the default case
3896 return StoreNode::Ideal(phase, can_reshape);
3897 }
3898
3899 //=============================================================================
3900 //------------------------------Ideal------------------------------------------
3901 // If the store is from an AND mask that leaves the low bits untouched, then
3902 // we can skip the AND operation
3903 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
3904 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
3905 if( progress != nullptr ) return progress;
3906
3907 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
3908 if( progress != nullptr ) return progress;
3909
3910 // Finally check the default case
3911 return StoreNode::Ideal(phase, can_reshape);
3912 }
3913
3914 //=============================================================================
3915 //----------------------------------SCMemProjNode------------------------------
3916 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
3917 {
3918 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
3919 return Type::TOP;
3920 }
3921 return bottom_type();
3922 }
3923
3924 //=============================================================================
3925 //----------------------------------LoadStoreNode------------------------------
3926 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
3927 : Node(required),
3928 _type(rt),
3929 _barrier_data(0)
3930 {
3931 init_req(MemNode::Control, c );
3932 init_req(MemNode::Memory , mem);
3933 init_req(MemNode::Address, adr);
3934 init_req(MemNode::ValueIn, val);
3935 init_class_id(Class_LoadStore);
3936 DEBUG_ONLY(_adr_type = at; adr_type();)
3937 }
3938
3939 //------------------------------Value-----------------------------------------
3940 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
3941 // Either input is TOP ==> the result is TOP
3942 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
3943 return Type::TOP;
3944 }
3945 const Type* t = phase->type(in(MemNode::Memory));
3946 if (t == Type::TOP) {
3947 return Type::TOP;
3948 }
3949 t = phase->type(in(MemNode::Address));
3950 if (t == Type::TOP) {
3951 return Type::TOP;
3952 }
3953 t = phase->type(in(MemNode::ValueIn));
3954 if (t == Type::TOP) {
3955 return Type::TOP;
3956 }
3957 return bottom_type();
3958 }
3959
3960 const TypePtr* LoadStoreNode::adr_type() const {
3961 const TypePtr* cross_check = DEBUG_ONLY(_adr_type) NOT_DEBUG(nullptr);
3962 return MemNode::calculate_adr_type(in(MemNode::Address)->bottom_type(), cross_check);
3963 }
3964
3965 uint LoadStoreNode::ideal_reg() const {
3966 return _type->ideal_reg();
3967 }
3968
3969 // This method conservatively checks if the result of a LoadStoreNode is
3970 // used, that is, if it returns true, then it is definitely the case that
3971 // the result of the node is not needed.
3972 // For example, GetAndAdd can be matched into a lock_add instead of a
3973 // lock_xadd if the result of LoadStoreNode::result_not_used() is true
3974 bool LoadStoreNode::result_not_used() const {
3975 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3976 Node *x = fast_out(i);
3977 if (x->Opcode() == Op_SCMemProj) {
3978 continue;
3979 }
3980 if (x->bottom_type() == TypeTuple::MEMBAR &&
3981 !x->is_Call() &&
3982 x->Opcode() != Op_Blackhole) {
3983 continue;
3984 }
3985 return false;
3986 }
3987 return true;
3988 }
3989
3990 MemBarNode* LoadStoreNode::trailing_membar() const {
3991 MemBarNode* trailing = nullptr;
3992 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3993 Node* u = fast_out(i);
3994 if (u->is_MemBar()) {
3995 if (u->as_MemBar()->trailing_load_store()) {
3996 assert(u->Opcode() == Op_MemBarAcquire, "");
3997 assert(trailing == nullptr, "only one");
3998 trailing = u->as_MemBar();
3999 #ifdef ASSERT
4000 Node* leading = trailing->leading_membar();
4001 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
4002 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
4003 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
4004 #endif
4005 } else {
4006 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
4007 }
4008 }
4009 }
4010
4011 return trailing;
4012 }
4013
4014 uint LoadStoreNode::size_of() const { return sizeof(*this); }
4015
4016 //=============================================================================
4017 //----------------------------------LoadStoreConditionalNode--------------------
4018 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) {
4019 init_req(ExpectedIn, ex );
4020 }
4021
4022 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
4023 // Either input is TOP ==> the result is TOP
4024 const Type* t = phase->type(in(ExpectedIn));
4025 if (t == Type::TOP) {
4026 return Type::TOP;
4027 }
4028 return LoadStoreNode::Value(phase);
4029 }
4030
4031 //=============================================================================
4032 //-------------------------------adr_type--------------------------------------
4033 const TypePtr* ClearArrayNode::adr_type() const {
4034 Node *adr = in(3);
4035 if (adr == nullptr) return nullptr; // node is dead
4036 return MemNode::calculate_adr_type(adr->bottom_type());
4037 }
4038
4039 //------------------------------match_edge-------------------------------------
4040 // Do we Match on this edge index or not? Do not match memory
4041 uint ClearArrayNode::match_edge(uint idx) const {
4042 return idx > 1;
4043 }
4044
4045 //------------------------------Identity---------------------------------------
4046 // Clearing a zero length array does nothing
4047 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
4048 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
4049 }
4050
4051 //------------------------------Idealize---------------------------------------
4052 // Clearing a short array is faster with stores
4053 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4054 // Already know this is a large node, do not try to ideal it
4055 if (_is_large) return nullptr;
4056
4057 const int unit = BytesPerLong;
4058 const TypeX* t = phase->type(in(2))->isa_intptr_t();
4059 if (!t) return nullptr;
4060 if (!t->is_con()) return nullptr;
4061 intptr_t raw_count = t->get_con();
4062 intptr_t size = raw_count;
4063 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
4064 // Clearing nothing uses the Identity call.
4065 // Negative clears are possible on dead ClearArrays
4066 // (see jck test stmt114.stmt11402.val).
4067 if (size <= 0 || size % unit != 0) return nullptr;
4068 intptr_t count = size / unit;
4069 // Length too long; communicate this to matchers and assemblers.
4070 // Assemblers are responsible to produce fast hardware clears for it.
4071 if (size > InitArrayShortSize) {
4072 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
4073 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
4074 return nullptr;
4075 }
4076 if (!IdealizeClearArrayNode) return nullptr;
4077 Node *mem = in(1);
4078 if( phase->type(mem)==Type::TOP ) return nullptr;
4079 Node *adr = in(3);
4080 const Type* at = phase->type(adr);
4081 if( at==Type::TOP ) return nullptr;
4082 const TypePtr* atp = at->isa_ptr();
4083 // adjust atp to be the correct array element address type
4084 if (atp == nullptr) atp = TypePtr::BOTTOM;
4085 else atp = atp->add_offset(Type::OffsetBot);
4086 // Get base for derived pointer purposes
4087 if( adr->Opcode() != Op_AddP ) Unimplemented();
4088 Node *base = adr->in(1);
4089
4090 Node *zero = phase->makecon(TypeLong::ZERO);
4091 Node *off = phase->MakeConX(BytesPerLong);
4092 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
4093 count--;
4094 while( count-- ) {
4095 mem = phase->transform(mem);
4096 adr = phase->transform(new AddPNode(base,adr,off));
4097 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
4098 }
4099 return mem;
4100 }
4101
4102 //----------------------------step_through----------------------------------
4103 // Return allocation input memory edge if it is different instance
4104 // or itself if it is the one we are looking for.
4105 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
4106 Node* n = *np;
4107 assert(n->is_ClearArray(), "sanity");
4108 intptr_t offset;
4109 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
4110 // This method is called only before Allocate nodes are expanded
4111 // during macro nodes expansion. Before that ClearArray nodes are
4112 // only generated in PhaseMacroExpand::generate_arraycopy() (before
4113 // Allocate nodes are expanded) which follows allocations.
4114 assert(alloc != nullptr, "should have allocation");
4115 if (alloc->_idx == instance_id) {
4116 // Can not bypass initialization of the instance we are looking for.
4117 return false;
4118 }
4119 // Otherwise skip it.
4120 InitializeNode* init = alloc->initialization();
4121 if (init != nullptr)
4122 *np = init->in(TypeFunc::Memory);
4123 else
4124 *np = alloc->in(TypeFunc::Memory);
4125 return true;
4126 }
4127
4128 Node* ClearArrayNode::make_address(Node* dest, Node* offset, bool raw_base, PhaseGVN* phase) {
4129 Node* base = dest;
4130 if (raw_base) {
4131 // May be called as part of the initialization of a just allocated object
4132 base = phase->C->top();
4133 }
4134 return phase->transform(new AddPNode(base, dest, offset));
4135 }
4136
4137 //----------------------------clear_memory-------------------------------------
4138 // Generate code to initialize object storage to zero.
4139 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4140 intptr_t start_offset,
4141 Node* end_offset,
4142 bool raw_base,
4143 PhaseGVN* phase) {
4144 intptr_t offset = start_offset;
4145
4146 int unit = BytesPerLong;
4147 if ((offset % unit) != 0) {
4148 Node* adr = make_address(dest, phase->MakeConX(offset), raw_base, phase);
4149 const TypePtr* atp = TypeRawPtr::BOTTOM;
4150 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4151 mem = phase->transform(mem);
4152 offset += BytesPerInt;
4153 }
4154 assert((offset % unit) == 0, "");
4155
4156 // Initialize the remaining stuff, if any, with a ClearArray.
4157 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, raw_base, phase);
4158 }
4159
4160 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4161 Node* start_offset,
4162 Node* end_offset,
4163 bool raw_base,
4164 PhaseGVN* phase) {
4165 if (start_offset == end_offset) {
4166 // nothing to do
4167 return mem;
4168 }
4169
4170 int unit = BytesPerLong;
4171 Node* zbase = start_offset;
4172 Node* zend = end_offset;
4173
4174 // Scale to the unit required by the CPU:
4175 if (!Matcher::init_array_count_is_in_bytes) {
4176 Node* shift = phase->intcon(exact_log2(unit));
4177 zbase = phase->transform(new URShiftXNode(zbase, shift) );
4178 zend = phase->transform(new URShiftXNode(zend, shift) );
4179 }
4180
4181 // Bulk clear double-words
4182 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
4183 Node* adr = make_address(dest, start_offset, raw_base, phase);
4184 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
4185 return phase->transform(mem);
4186 }
4187
4188 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4189 intptr_t start_offset,
4190 intptr_t end_offset,
4191 bool raw_base,
4192 PhaseGVN* phase) {
4193 if (start_offset == end_offset) {
4194 // nothing to do
4195 return mem;
4196 }
4197
4198 assert((end_offset % BytesPerInt) == 0, "odd end offset");
4199 intptr_t done_offset = end_offset;
4200 if ((done_offset % BytesPerLong) != 0) {
4201 done_offset -= BytesPerInt;
4202 }
4203 if (done_offset > start_offset) {
4204 mem = clear_memory(ctl, mem, dest,
4205 start_offset, phase->MakeConX(done_offset), raw_base, phase);
4206 }
4207 if (done_offset < end_offset) { // emit the final 32-bit store
4208 Node* adr = make_address(dest, phase->MakeConX(done_offset), raw_base, phase);
4209 const TypePtr* atp = TypeRawPtr::BOTTOM;
4210 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4211 mem = phase->transform(mem);
4212 done_offset += BytesPerInt;
4213 }
4214 assert(done_offset == end_offset, "");
4215 return mem;
4216 }
4217
4218 //=============================================================================
4219 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
4220 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
4221 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
4222 #ifdef ASSERT
4223 , _pair_idx(0)
4224 #endif
4225 {
4226 init_class_id(Class_MemBar);
4227 Node* top = C->top();
4228 init_req(TypeFunc::I_O,top);
4229 init_req(TypeFunc::FramePtr,top);
4230 init_req(TypeFunc::ReturnAdr,top);
4231 if (precedent != nullptr)
4232 init_req(TypeFunc::Parms, precedent);
4233 }
4234
4235 //------------------------------cmp--------------------------------------------
4236 uint MemBarNode::hash() const { return NO_HASH; }
4237 bool MemBarNode::cmp( const Node &n ) const {
4238 return (&n == this); // Always fail except on self
4239 }
4240
4241 //------------------------------make-------------------------------------------
4242 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
4243 switch (opcode) {
4244 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
4245 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
4246 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
4247 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
4248 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
4249 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn);
4250 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
4251 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
4252 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
4253 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
4254 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
4255 case Op_Initialize: return new InitializeNode(C, atp, pn);
4256 default: ShouldNotReachHere(); return nullptr;
4257 }
4258 }
4259
4260 void MemBarNode::remove(PhaseIterGVN *igvn) {
4261 assert(outcnt() > 0 && (outcnt() <= 2 || Opcode() == Op_Initialize), "Only one or two out edges allowed");
4262 if (trailing_store() || trailing_load_store()) {
4263 MemBarNode* leading = leading_membar();
4264 if (leading != nullptr) {
4265 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
4266 leading->remove(igvn);
4267 }
4268 }
4269 if (proj_out_or_null(TypeFunc::Control) != nullptr) {
4270 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
4271 }
4272 if (is_Initialize()) {
4273 as_Initialize()->replace_mem_projs_by(in(TypeFunc::Memory), igvn);
4274 } else {
4275 if (proj_out_or_null(TypeFunc::Memory) != nullptr) {
4276 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
4277 }
4278 }
4279 }
4280
4281 //------------------------------Ideal------------------------------------------
4282 // Return a node which is more "ideal" than the current node. Strip out
4283 // control copies
4284 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4285 if (remove_dead_region(phase, can_reshape)) return this;
4286 // Don't bother trying to transform a dead node
4287 if (in(0) && in(0)->is_top()) {
4288 return nullptr;
4289 }
4290
4291 bool progress = false;
4292 // Eliminate volatile MemBars for scalar replaced objects.
4293 if (can_reshape && req() == (Precedent+1)) {
4294 bool eliminate = false;
4295 int opc = Opcode();
4296 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
4297 // Volatile field loads and stores.
4298 Node* my_mem = in(MemBarNode::Precedent);
4299 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
4300 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
4301 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
4302 // replace this Precedent (decodeN) with the Load instead.
4303 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
4304 Node* load_node = my_mem->in(1);
4305 set_req(MemBarNode::Precedent, load_node);
4306 phase->is_IterGVN()->_worklist.push(my_mem);
4307 my_mem = load_node;
4308 } else {
4309 assert(my_mem->unique_out() == this, "sanity");
4310 assert(!trailing_load_store(), "load store node can't be eliminated");
4311 del_req(Precedent);
4312 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
4313 my_mem = nullptr;
4314 }
4315 progress = true;
4316 }
4317 if (my_mem != nullptr && my_mem->is_Mem()) {
4318 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
4319 // Check for scalar replaced object reference.
4320 if( t_oop != nullptr && t_oop->is_known_instance_field() &&
4321 t_oop->offset() != Type::OffsetBot &&
4322 t_oop->offset() != Type::OffsetTop) {
4323 eliminate = true;
4324 }
4325 }
4326 } else if (opc == Op_MemBarRelease || (UseStoreStoreForCtor && opc == Op_MemBarStoreStore)) {
4327 // Final field stores.
4328 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent));
4329 if ((alloc != nullptr) && alloc->is_Allocate() &&
4330 alloc->as_Allocate()->does_not_escape_thread()) {
4331 // The allocated object does not escape.
4332 eliminate = true;
4333 }
4334 }
4335 if (eliminate) {
4336 // Replace MemBar projections by its inputs.
4337 PhaseIterGVN* igvn = phase->is_IterGVN();
4338 remove(igvn);
4339 // Must return either the original node (now dead) or a new node
4340 // (Do not return a top here, since that would break the uniqueness of top.)
4341 return new ConINode(TypeInt::ZERO);
4342 }
4343 }
4344 return progress ? this : nullptr;
4345 }
4346
4347 //------------------------------Value------------------------------------------
4348 const Type* MemBarNode::Value(PhaseGVN* phase) const {
4349 if( !in(0) ) return Type::TOP;
4350 if( phase->type(in(0)) == Type::TOP )
4351 return Type::TOP;
4352 return TypeTuple::MEMBAR;
4353 }
4354
4355 //------------------------------match------------------------------------------
4356 // Construct projections for memory.
4357 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
4358 switch (proj->_con) {
4359 case TypeFunc::Control:
4360 case TypeFunc::Memory:
4361 return new MachProjNode(this, proj->_con, RegMask::EMPTY, MachProjNode::unmatched_proj);
4362 }
4363 ShouldNotReachHere();
4364 return nullptr;
4365 }
4366
4367 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4368 trailing->_kind = TrailingStore;
4369 leading->_kind = LeadingStore;
4370 #ifdef ASSERT
4371 trailing->_pair_idx = leading->_idx;
4372 leading->_pair_idx = leading->_idx;
4373 #endif
4374 }
4375
4376 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4377 trailing->_kind = TrailingLoadStore;
4378 leading->_kind = LeadingLoadStore;
4379 #ifdef ASSERT
4380 trailing->_pair_idx = leading->_idx;
4381 leading->_pair_idx = leading->_idx;
4382 #endif
4383 }
4384
4385 MemBarNode* MemBarNode::trailing_membar() const {
4386 ResourceMark rm;
4387 Node* trailing = (Node*)this;
4388 VectorSet seen;
4389 Node_Stack multis(0);
4390 do {
4391 Node* c = trailing;
4392 uint i = 0;
4393 do {
4394 trailing = nullptr;
4395 for (; i < c->outcnt(); i++) {
4396 Node* next = c->raw_out(i);
4397 if (next != c && next->is_CFG()) {
4398 if (c->is_MultiBranch()) {
4399 if (multis.node() == c) {
4400 multis.set_index(i+1);
4401 } else {
4402 multis.push(c, i+1);
4403 }
4404 }
4405 trailing = next;
4406 break;
4407 }
4408 }
4409 if (trailing != nullptr && !seen.test_set(trailing->_idx)) {
4410 break;
4411 }
4412 while (multis.size() > 0) {
4413 c = multis.node();
4414 i = multis.index();
4415 if (i < c->req()) {
4416 break;
4417 }
4418 multis.pop();
4419 }
4420 } while (multis.size() > 0);
4421 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
4422
4423 MemBarNode* mb = trailing->as_MemBar();
4424 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
4425 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
4426 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
4427 return mb;
4428 }
4429
4430 MemBarNode* MemBarNode::leading_membar() const {
4431 ResourceMark rm;
4432 VectorSet seen;
4433 Node_Stack regions(0);
4434 Node* leading = in(0);
4435 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
4436 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) {
4437 leading = nullptr;
4438 while (regions.size() > 0 && leading == nullptr) {
4439 Node* r = regions.node();
4440 uint i = regions.index();
4441 if (i < r->req()) {
4442 leading = r->in(i);
4443 regions.set_index(i+1);
4444 } else {
4445 regions.pop();
4446 }
4447 }
4448 if (leading == nullptr) {
4449 assert(regions.size() == 0, "all paths should have been tried");
4450 return nullptr;
4451 }
4452 }
4453 if (leading->is_Region()) {
4454 regions.push(leading, 2);
4455 leading = leading->in(1);
4456 } else {
4457 leading = leading->in(0);
4458 }
4459 }
4460 #ifdef ASSERT
4461 Unique_Node_List wq;
4462 wq.push((Node*)this);
4463 uint found = 0;
4464 for (uint i = 0; i < wq.size(); i++) {
4465 Node* n = wq.at(i);
4466 if (n->is_Region()) {
4467 for (uint j = 1; j < n->req(); j++) {
4468 Node* in = n->in(j);
4469 if (in != nullptr && !in->is_top()) {
4470 wq.push(in);
4471 }
4472 }
4473 } else {
4474 if (n->is_MemBar() && n->as_MemBar()->leading()) {
4475 assert(n == leading, "consistency check failed");
4476 found++;
4477 } else {
4478 Node* in = n->in(0);
4479 if (in != nullptr && !in->is_top()) {
4480 wq.push(in);
4481 }
4482 }
4483 }
4484 }
4485 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed");
4486 #endif
4487 if (leading == nullptr) {
4488 return nullptr;
4489 }
4490 MemBarNode* mb = leading->as_MemBar();
4491 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
4492 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
4493 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
4494 return mb;
4495 }
4496
4497
4498 //===========================InitializeNode====================================
4499 // SUMMARY:
4500 // This node acts as a memory barrier on raw memory, after some raw stores.
4501 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
4502 // The Initialize can 'capture' suitably constrained stores as raw inits.
4503 // It can coalesce related raw stores into larger units (called 'tiles').
4504 // It can avoid zeroing new storage for memory units which have raw inits.
4505 // At macro-expansion, it is marked 'complete', and does not optimize further.
4506 //
4507 // EXAMPLE:
4508 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
4509 // ctl = incoming control; mem* = incoming memory
4510 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
4511 // First allocate uninitialized memory and fill in the header:
4512 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
4513 // ctl := alloc.Control; mem* := alloc.Memory*
4514 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
4515 // Then initialize to zero the non-header parts of the raw memory block:
4516 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
4517 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
4518 // After the initialize node executes, the object is ready for service:
4519 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
4520 // Suppose its body is immediately initialized as {1,2}:
4521 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4522 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4523 // mem.SLICE(#short[*]) := store2
4524 //
4525 // DETAILS:
4526 // An InitializeNode collects and isolates object initialization after
4527 // an AllocateNode and before the next possible safepoint. As a
4528 // memory barrier (MemBarNode), it keeps critical stores from drifting
4529 // down past any safepoint or any publication of the allocation.
4530 // Before this barrier, a newly-allocated object may have uninitialized bits.
4531 // After this barrier, it may be treated as a real oop, and GC is allowed.
4532 //
4533 // The semantics of the InitializeNode include an implicit zeroing of
4534 // the new object from object header to the end of the object.
4535 // (The object header and end are determined by the AllocateNode.)
4536 //
4537 // Certain stores may be added as direct inputs to the InitializeNode.
4538 // These stores must update raw memory, and they must be to addresses
4539 // derived from the raw address produced by AllocateNode, and with
4540 // a constant offset. They must be ordered by increasing offset.
4541 // The first one is at in(RawStores), the last at in(req()-1).
4542 // Unlike most memory operations, they are not linked in a chain,
4543 // but are displayed in parallel as users of the rawmem output of
4544 // the allocation.
4545 //
4546 // (See comments in InitializeNode::capture_store, which continue
4547 // the example given above.)
4548 //
4549 // When the associated Allocate is macro-expanded, the InitializeNode
4550 // may be rewritten to optimize collected stores. A ClearArrayNode
4551 // may also be created at that point to represent any required zeroing.
4552 // The InitializeNode is then marked 'complete', prohibiting further
4553 // capturing of nearby memory operations.
4554 //
4555 // During macro-expansion, all captured initializations which store
4556 // constant values of 32 bits or smaller are coalesced (if advantageous)
4557 // into larger 'tiles' 32 or 64 bits. This allows an object to be
4558 // initialized in fewer memory operations. Memory words which are
4559 // covered by neither tiles nor non-constant stores are pre-zeroed
4560 // by explicit stores of zero. (The code shape happens to do all
4561 // zeroing first, then all other stores, with both sequences occurring
4562 // in order of ascending offsets.)
4563 //
4564 // Alternatively, code may be inserted between an AllocateNode and its
4565 // InitializeNode, to perform arbitrary initialization of the new object.
4566 // E.g., the object copying intrinsics insert complex data transfers here.
4567 // The initialization must then be marked as 'complete' disable the
4568 // built-in zeroing semantics and the collection of initializing stores.
4569 //
4570 // While an InitializeNode is incomplete, reads from the memory state
4571 // produced by it are optimizable if they match the control edge and
4572 // new oop address associated with the allocation/initialization.
4573 // They return a stored value (if the offset matches) or else zero.
4574 // A write to the memory state, if it matches control and address,
4575 // and if it is to a constant offset, may be 'captured' by the
4576 // InitializeNode. It is cloned as a raw memory operation and rewired
4577 // inside the initialization, to the raw oop produced by the allocation.
4578 // Operations on addresses which are provably distinct (e.g., to
4579 // other AllocateNodes) are allowed to bypass the initialization.
4580 //
4581 // The effect of all this is to consolidate object initialization
4582 // (both arrays and non-arrays, both piecewise and bulk) into a
4583 // single location, where it can be optimized as a unit.
4584 //
4585 // Only stores with an offset less than TrackedInitializationLimit words
4586 // will be considered for capture by an InitializeNode. This puts a
4587 // reasonable limit on the complexity of optimized initializations.
4588
4589 //---------------------------InitializeNode------------------------------------
4590 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
4591 : MemBarNode(C, adr_type, rawoop),
4592 _is_complete(Incomplete), _does_not_escape(false)
4593 {
4594 init_class_id(Class_Initialize);
4595
4596 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
4597 assert(in(RawAddress) == rawoop, "proper init");
4598 // Note: allocation() can be null, for secondary initialization barriers
4599 }
4600
4601 // Since this node is not matched, it will be processed by the
4602 // register allocator. Declare that there are no constraints
4603 // on the allocation of the RawAddress edge.
4604 const RegMask &InitializeNode::in_RegMask(uint idx) const {
4605 // This edge should be set to top, by the set_complete. But be conservative.
4606 if (idx == InitializeNode::RawAddress)
4607 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
4608 return RegMask::EMPTY;
4609 }
4610
4611 Node* InitializeNode::memory(uint alias_idx) {
4612 Node* mem = in(Memory);
4613 if (mem->is_MergeMem()) {
4614 return mem->as_MergeMem()->memory_at(alias_idx);
4615 } else {
4616 // incoming raw memory is not split
4617 return mem;
4618 }
4619 }
4620
4621 bool InitializeNode::is_non_zero() {
4622 if (is_complete()) return false;
4623 remove_extra_zeroes();
4624 return (req() > RawStores);
4625 }
4626
4627 void InitializeNode::set_complete(PhaseGVN* phase) {
4628 assert(!is_complete(), "caller responsibility");
4629 _is_complete = Complete;
4630
4631 // After this node is complete, it contains a bunch of
4632 // raw-memory initializations. There is no need for
4633 // it to have anything to do with non-raw memory effects.
4634 // Therefore, tell all non-raw users to re-optimize themselves,
4635 // after skipping the memory effects of this initialization.
4636 PhaseIterGVN* igvn = phase->is_IterGVN();
4637 if (igvn) igvn->add_users_to_worklist(this);
4638 }
4639
4640 // convenience function
4641 // return false if the init contains any stores already
4642 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
4643 InitializeNode* init = initialization();
4644 if (init == nullptr || init->is_complete()) return false;
4645 init->remove_extra_zeroes();
4646 // for now, if this allocation has already collected any inits, bail:
4647 if (init->is_non_zero()) return false;
4648 init->set_complete(phase);
4649 return true;
4650 }
4651
4652 void InitializeNode::remove_extra_zeroes() {
4653 if (req() == RawStores) return;
4654 Node* zmem = zero_memory();
4655 uint fill = RawStores;
4656 for (uint i = fill; i < req(); i++) {
4657 Node* n = in(i);
4658 if (n->is_top() || n == zmem) continue; // skip
4659 if (fill < i) set_req(fill, n); // compact
4660 ++fill;
4661 }
4662 // delete any empty spaces created:
4663 while (fill < req()) {
4664 del_req(fill);
4665 }
4666 }
4667
4668 // Helper for remembering which stores go with which offsets.
4669 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) {
4670 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
4671 intptr_t offset = -1;
4672 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
4673 phase, offset);
4674 if (base == nullptr) return -1; // something is dead,
4675 if (offset < 0) return -1; // dead, dead
4676 return offset;
4677 }
4678
4679 // Helper for proving that an initialization expression is
4680 // "simple enough" to be folded into an object initialization.
4681 // Attempts to prove that a store's initial value 'n' can be captured
4682 // within the initialization without creating a vicious cycle, such as:
4683 // { Foo p = new Foo(); p.next = p; }
4684 // True for constants and parameters and small combinations thereof.
4685 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
4686 ResourceMark rm;
4687 Unique_Node_List worklist;
4688 worklist.push(value);
4689
4690 uint complexity_limit = 20;
4691 for (uint j = 0; j < worklist.size(); j++) {
4692 if (j >= complexity_limit) {
4693 return false; // Bail out if processed too many nodes
4694 }
4695
4696 Node* n = worklist.at(j);
4697 if (n == nullptr) continue; // (can this really happen?)
4698 if (n->is_Proj()) n = n->in(0);
4699 if (n == this) return false; // found a cycle
4700 if (n->is_Con()) continue;
4701 if (n->is_Start()) continue; // params, etc., are OK
4702 if (n->is_Root()) continue; // even better
4703
4704 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
4705 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
4706 continue;
4707 }
4708
4709 Node* ctl = n->in(0);
4710 if (ctl != nullptr && !ctl->is_top()) {
4711 if (ctl->is_Proj()) ctl = ctl->in(0);
4712 if (ctl == this) return false;
4713
4714 // If we already know that the enclosing memory op is pinned right after
4715 // the init, then any control flow that the store has picked up
4716 // must have preceded the init, or else be equal to the init.
4717 // Even after loop optimizations (which might change control edges)
4718 // a store is never pinned *before* the availability of its inputs.
4719 if (!MemNode::all_controls_dominate(n, this)) {
4720 return false; // failed to prove a good control
4721 }
4722 }
4723
4724 // Check data edges for possible dependencies on 'this'.
4725 for (uint i = 1; i < n->req(); i++) {
4726 Node* m = n->in(i);
4727 if (m == nullptr || m == n || m->is_top()) continue;
4728
4729 // Only process data inputs once
4730 worklist.push(m);
4731 }
4732 }
4733
4734 return true;
4735 }
4736
4737 // Here are all the checks a Store must pass before it can be moved into
4738 // an initialization. Returns zero if a check fails.
4739 // On success, returns the (constant) offset to which the store applies,
4740 // within the initialized memory.
4741 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
4742 const int FAIL = 0;
4743 if (st->req() != MemNode::ValueIn + 1)
4744 return FAIL; // an inscrutable StoreNode (card mark?)
4745 Node* ctl = st->in(MemNode::Control);
4746 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this))
4747 return FAIL; // must be unconditional after the initialization
4748 Node* mem = st->in(MemNode::Memory);
4749 if (!(mem->is_Proj() && mem->in(0) == this))
4750 return FAIL; // must not be preceded by other stores
4751 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4752 if ((st->Opcode() == Op_StoreP || st->Opcode() == Op_StoreN) &&
4753 !bs->can_initialize_object(st)) {
4754 return FAIL;
4755 }
4756 Node* adr = st->in(MemNode::Address);
4757 intptr_t offset;
4758 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
4759 if (alloc == nullptr)
4760 return FAIL; // inscrutable address
4761 if (alloc != allocation())
4762 return FAIL; // wrong allocation! (store needs to float up)
4763 int size_in_bytes = st->memory_size();
4764 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
4765 return FAIL; // mismatched access
4766 }
4767 Node* val = st->in(MemNode::ValueIn);
4768
4769 if (!detect_init_independence(val, phase))
4770 return FAIL; // stored value must be 'simple enough'
4771
4772 // The Store can be captured only if nothing after the allocation
4773 // and before the Store is using the memory location that the store
4774 // overwrites.
4775 bool failed = false;
4776 // If is_complete_with_arraycopy() is true the shape of the graph is
4777 // well defined and is safe so no need for extra checks.
4778 if (!is_complete_with_arraycopy()) {
4779 // We are going to look at each use of the memory state following
4780 // the allocation to make sure nothing reads the memory that the
4781 // Store writes.
4782 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
4783 int alias_idx = phase->C->get_alias_index(t_adr);
4784 ResourceMark rm;
4785 Unique_Node_List mems;
4786 mems.push(mem);
4787 Node* unique_merge = nullptr;
4788 for (uint next = 0; next < mems.size(); ++next) {
4789 Node *m = mems.at(next);
4790 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
4791 Node *n = m->fast_out(j);
4792 if (n->outcnt() == 0) {
4793 continue;
4794 }
4795 if (n == st) {
4796 continue;
4797 } else if (n->in(0) != nullptr && n->in(0) != ctl) {
4798 // If the control of this use is different from the control
4799 // of the Store which is right after the InitializeNode then
4800 // this node cannot be between the InitializeNode and the
4801 // Store.
4802 continue;
4803 } else if (n->is_MergeMem()) {
4804 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
4805 // We can hit a MergeMemNode (that will likely go away
4806 // later) that is a direct use of the memory state
4807 // following the InitializeNode on the same slice as the
4808 // store node that we'd like to capture. We need to check
4809 // the uses of the MergeMemNode.
4810 mems.push(n);
4811 }
4812 } else if (n->is_Mem()) {
4813 Node* other_adr = n->in(MemNode::Address);
4814 if (other_adr == adr) {
4815 failed = true;
4816 break;
4817 } else {
4818 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
4819 if (other_t_adr != nullptr) {
4820 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
4821 if (other_alias_idx == alias_idx) {
4822 // A load from the same memory slice as the store right
4823 // after the InitializeNode. We check the control of the
4824 // object/array that is loaded from. If it's the same as
4825 // the store control then we cannot capture the store.
4826 assert(!n->is_Store(), "2 stores to same slice on same control?");
4827 Node* base = other_adr;
4828 assert(base->is_AddP(), "should be addp but is %s", base->Name());
4829 base = base->in(AddPNode::Base);
4830 if (base != nullptr) {
4831 base = base->uncast();
4832 if (base->is_Proj() && base->in(0) == alloc) {
4833 failed = true;
4834 break;
4835 }
4836 }
4837 }
4838 }
4839 }
4840 } else {
4841 failed = true;
4842 break;
4843 }
4844 }
4845 }
4846 }
4847 if (failed) {
4848 if (!can_reshape) {
4849 // We decided we couldn't capture the store during parsing. We
4850 // should try again during the next IGVN once the graph is
4851 // cleaner.
4852 phase->C->record_for_igvn(st);
4853 }
4854 return FAIL;
4855 }
4856
4857 return offset; // success
4858 }
4859
4860 // Find the captured store in(i) which corresponds to the range
4861 // [start..start+size) in the initialized object.
4862 // If there is one, return its index i. If there isn't, return the
4863 // negative of the index where it should be inserted.
4864 // Return 0 if the queried range overlaps an initialization boundary
4865 // or if dead code is encountered.
4866 // If size_in_bytes is zero, do not bother with overlap checks.
4867 int InitializeNode::captured_store_insertion_point(intptr_t start,
4868 int size_in_bytes,
4869 PhaseValues* phase) {
4870 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
4871
4872 if (is_complete())
4873 return FAIL; // arraycopy got here first; punt
4874
4875 assert(allocation() != nullptr, "must be present");
4876
4877 // no negatives, no header fields:
4878 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
4879
4880 // after a certain size, we bail out on tracking all the stores:
4881 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4882 if (start >= ti_limit) return FAIL;
4883
4884 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
4885 if (i >= limit) return -(int)i; // not found; here is where to put it
4886
4887 Node* st = in(i);
4888 intptr_t st_off = get_store_offset(st, phase);
4889 if (st_off < 0) {
4890 if (st != zero_memory()) {
4891 return FAIL; // bail out if there is dead garbage
4892 }
4893 } else if (st_off > start) {
4894 // ...we are done, since stores are ordered
4895 if (st_off < start + size_in_bytes) {
4896 return FAIL; // the next store overlaps
4897 }
4898 return -(int)i; // not found; here is where to put it
4899 } else if (st_off < start) {
4900 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
4901 if (size_in_bytes != 0 &&
4902 start < st_off + MAX_STORE &&
4903 start < st_off + st->as_Store()->memory_size()) {
4904 return FAIL; // the previous store overlaps
4905 }
4906 } else {
4907 if (size_in_bytes != 0 &&
4908 st->as_Store()->memory_size() != size_in_bytes) {
4909 return FAIL; // mismatched store size
4910 }
4911 return i;
4912 }
4913
4914 ++i;
4915 }
4916 }
4917
4918 // Look for a captured store which initializes at the offset 'start'
4919 // with the given size. If there is no such store, and no other
4920 // initialization interferes, then return zero_memory (the memory
4921 // projection of the AllocateNode).
4922 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
4923 PhaseValues* phase) {
4924 assert(stores_are_sane(phase), "");
4925 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4926 if (i == 0) {
4927 return nullptr; // something is dead
4928 } else if (i < 0) {
4929 return zero_memory(); // just primordial zero bits here
4930 } else {
4931 Node* st = in(i); // here is the store at this position
4932 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
4933 return st;
4934 }
4935 }
4936
4937 // Create, as a raw pointer, an address within my new object at 'offset'.
4938 Node* InitializeNode::make_raw_address(intptr_t offset,
4939 PhaseGVN* phase) {
4940 Node* addr = in(RawAddress);
4941 if (offset != 0) {
4942 Compile* C = phase->C;
4943 addr = phase->transform( new AddPNode(C->top(), addr,
4944 phase->MakeConX(offset)) );
4945 }
4946 return addr;
4947 }
4948
4949 // Clone the given store, converting it into a raw store
4950 // initializing a field or element of my new object.
4951 // Caller is responsible for retiring the original store,
4952 // with subsume_node or the like.
4953 //
4954 // From the example above InitializeNode::InitializeNode,
4955 // here are the old stores to be captured:
4956 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4957 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4958 //
4959 // Here is the changed code; note the extra edges on init:
4960 // alloc = (Allocate ...)
4961 // rawoop = alloc.RawAddress
4962 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
4963 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
4964 // init = (Initialize alloc.Control alloc.Memory rawoop
4965 // rawstore1 rawstore2)
4966 //
4967 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
4968 PhaseGVN* phase, bool can_reshape) {
4969 assert(stores_are_sane(phase), "");
4970
4971 if (start < 0) return nullptr;
4972 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
4973
4974 Compile* C = phase->C;
4975 int size_in_bytes = st->memory_size();
4976 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4977 if (i == 0) return nullptr; // bail out
4978 Node* prev_mem = nullptr; // raw memory for the captured store
4979 if (i > 0) {
4980 prev_mem = in(i); // there is a pre-existing store under this one
4981 set_req(i, C->top()); // temporarily disconnect it
4982 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
4983 } else {
4984 i = -i; // no pre-existing store
4985 prev_mem = zero_memory(); // a slice of the newly allocated object
4986 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
4987 set_req(--i, C->top()); // reuse this edge; it has been folded away
4988 else
4989 ins_req(i, C->top()); // build a new edge
4990 }
4991 Node* new_st = st->clone();
4992 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4993 new_st->set_req(MemNode::Control, in(Control));
4994 new_st->set_req(MemNode::Memory, prev_mem);
4995 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
4996 bs->eliminate_gc_barrier_data(new_st);
4997 new_st = phase->transform(new_st);
4998
4999 // At this point, new_st might have swallowed a pre-existing store
5000 // at the same offset, or perhaps new_st might have disappeared,
5001 // if it redundantly stored the same value (or zero to fresh memory).
5002
5003 // In any case, wire it in:
5004 PhaseIterGVN* igvn = phase->is_IterGVN();
5005 if (igvn) {
5006 igvn->rehash_node_delayed(this);
5007 }
5008 set_req(i, new_st);
5009
5010 // The caller may now kill the old guy.
5011 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
5012 assert(check_st == new_st || check_st == nullptr, "must be findable");
5013 assert(!is_complete(), "");
5014 return new_st;
5015 }
5016
5017 static bool store_constant(jlong* tiles, int num_tiles,
5018 intptr_t st_off, int st_size,
5019 jlong con) {
5020 if ((st_off & (st_size-1)) != 0)
5021 return false; // strange store offset (assume size==2**N)
5022 address addr = (address)tiles + st_off;
5023 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
5024 switch (st_size) {
5025 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
5026 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
5027 case sizeof(jint): *(jint*) addr = (jint) con; break;
5028 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
5029 default: return false; // strange store size (detect size!=2**N here)
5030 }
5031 return true; // return success to caller
5032 }
5033
5034 // Coalesce subword constants into int constants and possibly
5035 // into long constants. The goal, if the CPU permits,
5036 // is to initialize the object with a small number of 64-bit tiles.
5037 // Also, convert floating-point constants to bit patterns.
5038 // Non-constants are not relevant to this pass.
5039 //
5040 // In terms of the running example on InitializeNode::InitializeNode
5041 // and InitializeNode::capture_store, here is the transformation
5042 // of rawstore1 and rawstore2 into rawstore12:
5043 // alloc = (Allocate ...)
5044 // rawoop = alloc.RawAddress
5045 // tile12 = 0x00010002
5046 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
5047 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
5048 //
5049 void
5050 InitializeNode::coalesce_subword_stores(intptr_t header_size,
5051 Node* size_in_bytes,
5052 PhaseGVN* phase) {
5053 Compile* C = phase->C;
5054
5055 assert(stores_are_sane(phase), "");
5056 // Note: After this pass, they are not completely sane,
5057 // since there may be some overlaps.
5058
5059 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
5060
5061 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
5062 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
5063 size_limit = MIN2(size_limit, ti_limit);
5064 size_limit = align_up(size_limit, BytesPerLong);
5065 int num_tiles = size_limit / BytesPerLong;
5066
5067 // allocate space for the tile map:
5068 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
5069 jlong tiles_buf[small_len];
5070 Node* nodes_buf[small_len];
5071 jlong inits_buf[small_len];
5072 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
5073 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5074 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
5075 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
5076 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
5077 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5078 // tiles: exact bitwise model of all primitive constants
5079 // nodes: last constant-storing node subsumed into the tiles model
5080 // inits: which bytes (in each tile) are touched by any initializations
5081
5082 //// Pass A: Fill in the tile model with any relevant stores.
5083
5084 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
5085 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
5086 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
5087 Node* zmem = zero_memory(); // initially zero memory state
5088 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5089 Node* st = in(i);
5090 intptr_t st_off = get_store_offset(st, phase);
5091
5092 // Figure out the store's offset and constant value:
5093 if (st_off < header_size) continue; //skip (ignore header)
5094 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
5095 int st_size = st->as_Store()->memory_size();
5096 if (st_off + st_size > size_limit) break;
5097
5098 // Record which bytes are touched, whether by constant or not.
5099 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
5100 continue; // skip (strange store size)
5101
5102 const Type* val = phase->type(st->in(MemNode::ValueIn));
5103 if (!val->singleton()) continue; //skip (non-con store)
5104 BasicType type = val->basic_type();
5105
5106 jlong con = 0;
5107 switch (type) {
5108 case T_INT: con = val->is_int()->get_con(); break;
5109 case T_LONG: con = val->is_long()->get_con(); break;
5110 case T_FLOAT: con = jint_cast(val->getf()); break;
5111 case T_DOUBLE: con = jlong_cast(val->getd()); break;
5112 default: continue; //skip (odd store type)
5113 }
5114
5115 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
5116 st->Opcode() == Op_StoreL) {
5117 continue; // This StoreL is already optimal.
5118 }
5119
5120 // Store down the constant.
5121 store_constant(tiles, num_tiles, st_off, st_size, con);
5122
5123 intptr_t j = st_off >> LogBytesPerLong;
5124
5125 if (type == T_INT && st_size == BytesPerInt
5126 && (st_off & BytesPerInt) == BytesPerInt) {
5127 jlong lcon = tiles[j];
5128 if (!Matcher::isSimpleConstant64(lcon) &&
5129 st->Opcode() == Op_StoreI) {
5130 // This StoreI is already optimal by itself.
5131 jint* intcon = (jint*) &tiles[j];
5132 intcon[1] = 0; // undo the store_constant()
5133
5134 // If the previous store is also optimal by itself, back up and
5135 // undo the action of the previous loop iteration... if we can.
5136 // But if we can't, just let the previous half take care of itself.
5137 st = nodes[j];
5138 st_off -= BytesPerInt;
5139 con = intcon[0];
5140 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) {
5141 assert(st_off >= header_size, "still ignoring header");
5142 assert(get_store_offset(st, phase) == st_off, "must be");
5143 assert(in(i-1) == zmem, "must be");
5144 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
5145 assert(con == tcon->is_int()->get_con(), "must be");
5146 // Undo the effects of the previous loop trip, which swallowed st:
5147 intcon[0] = 0; // undo store_constant()
5148 set_req(i-1, st); // undo set_req(i, zmem)
5149 nodes[j] = nullptr; // undo nodes[j] = st
5150 --old_subword; // undo ++old_subword
5151 }
5152 continue; // This StoreI is already optimal.
5153 }
5154 }
5155
5156 // This store is not needed.
5157 set_req(i, zmem);
5158 nodes[j] = st; // record for the moment
5159 if (st_size < BytesPerLong) // something has changed
5160 ++old_subword; // includes int/float, but who's counting...
5161 else ++old_long;
5162 }
5163
5164 if ((old_subword + old_long) == 0)
5165 return; // nothing more to do
5166
5167 //// Pass B: Convert any non-zero tiles into optimal constant stores.
5168 // Be sure to insert them before overlapping non-constant stores.
5169 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
5170 for (int j = 0; j < num_tiles; j++) {
5171 jlong con = tiles[j];
5172 jlong init = inits[j];
5173 if (con == 0) continue;
5174 jint con0, con1; // split the constant, address-wise
5175 jint init0, init1; // split the init map, address-wise
5176 { union { jlong con; jint intcon[2]; } u;
5177 u.con = con;
5178 con0 = u.intcon[0];
5179 con1 = u.intcon[1];
5180 u.con = init;
5181 init0 = u.intcon[0];
5182 init1 = u.intcon[1];
5183 }
5184
5185 Node* old = nodes[j];
5186 assert(old != nullptr, "need the prior store");
5187 intptr_t offset = (j * BytesPerLong);
5188
5189 bool split = !Matcher::isSimpleConstant64(con);
5190
5191 if (offset < header_size) {
5192 assert(offset + BytesPerInt >= header_size, "second int counts");
5193 assert(*(jint*)&tiles[j] == 0, "junk in header");
5194 split = true; // only the second word counts
5195 // Example: int a[] = { 42 ... }
5196 } else if (con0 == 0 && init0 == -1) {
5197 split = true; // first word is covered by full inits
5198 // Example: int a[] = { ... foo(), 42 ... }
5199 } else if (con1 == 0 && init1 == -1) {
5200 split = true; // second word is covered by full inits
5201 // Example: int a[] = { ... 42, foo() ... }
5202 }
5203
5204 // Here's a case where init0 is neither 0 nor -1:
5205 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
5206 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
5207 // In this case the tile is not split; it is (jlong)42.
5208 // The big tile is stored down, and then the foo() value is inserted.
5209 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
5210
5211 Node* ctl = old->in(MemNode::Control);
5212 Node* adr = make_raw_address(offset, phase);
5213 const TypePtr* atp = TypeRawPtr::BOTTOM;
5214
5215 // One or two coalesced stores to plop down.
5216 Node* st[2];
5217 intptr_t off[2];
5218 int nst = 0;
5219 if (!split) {
5220 ++new_long;
5221 off[nst] = offset;
5222 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5223 phase->longcon(con), T_LONG, MemNode::unordered);
5224 } else {
5225 // Omit either if it is a zero.
5226 if (con0 != 0) {
5227 ++new_int;
5228 off[nst] = offset;
5229 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5230 phase->intcon(con0), T_INT, MemNode::unordered);
5231 }
5232 if (con1 != 0) {
5233 ++new_int;
5234 offset += BytesPerInt;
5235 adr = make_raw_address(offset, phase);
5236 off[nst] = offset;
5237 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5238 phase->intcon(con1), T_INT, MemNode::unordered);
5239 }
5240 }
5241
5242 // Insert second store first, then the first before the second.
5243 // Insert each one just before any overlapping non-constant stores.
5244 while (nst > 0) {
5245 Node* st1 = st[--nst];
5246 C->copy_node_notes_to(st1, old);
5247 st1 = phase->transform(st1);
5248 offset = off[nst];
5249 assert(offset >= header_size, "do not smash header");
5250 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
5251 guarantee(ins_idx != 0, "must re-insert constant store");
5252 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
5253 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
5254 set_req(--ins_idx, st1);
5255 else
5256 ins_req(ins_idx, st1);
5257 }
5258 }
5259
5260 if (PrintCompilation && WizardMode)
5261 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
5262 old_subword, old_long, new_int, new_long);
5263 if (C->log() != nullptr)
5264 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
5265 old_subword, old_long, new_int, new_long);
5266
5267 // Clean up any remaining occurrences of zmem:
5268 remove_extra_zeroes();
5269 }
5270
5271 // Explore forward from in(start) to find the first fully initialized
5272 // word, and return its offset. Skip groups of subword stores which
5273 // together initialize full words. If in(start) is itself part of a
5274 // fully initialized word, return the offset of in(start). If there
5275 // are no following full-word stores, or if something is fishy, return
5276 // a negative value.
5277 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
5278 int int_map = 0;
5279 intptr_t int_map_off = 0;
5280 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
5281
5282 for (uint i = start, limit = req(); i < limit; i++) {
5283 Node* st = in(i);
5284
5285 intptr_t st_off = get_store_offset(st, phase);
5286 if (st_off < 0) break; // return conservative answer
5287
5288 int st_size = st->as_Store()->memory_size();
5289 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
5290 return st_off; // we found a complete word init
5291 }
5292
5293 // update the map:
5294
5295 intptr_t this_int_off = align_down(st_off, BytesPerInt);
5296 if (this_int_off != int_map_off) {
5297 // reset the map:
5298 int_map = 0;
5299 int_map_off = this_int_off;
5300 }
5301
5302 int subword_off = st_off - this_int_off;
5303 int_map |= right_n_bits(st_size) << subword_off;
5304 if ((int_map & FULL_MAP) == FULL_MAP) {
5305 return this_int_off; // we found a complete word init
5306 }
5307
5308 // Did this store hit or cross the word boundary?
5309 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
5310 if (next_int_off == this_int_off + BytesPerInt) {
5311 // We passed the current int, without fully initializing it.
5312 int_map_off = next_int_off;
5313 int_map >>= BytesPerInt;
5314 } else if (next_int_off > this_int_off + BytesPerInt) {
5315 // We passed the current and next int.
5316 return this_int_off + BytesPerInt;
5317 }
5318 }
5319
5320 return -1;
5321 }
5322
5323
5324 // Called when the associated AllocateNode is expanded into CFG.
5325 // At this point, we may perform additional optimizations.
5326 // Linearize the stores by ascending offset, to make memory
5327 // activity as coherent as possible.
5328 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
5329 intptr_t header_size,
5330 Node* size_in_bytes,
5331 PhaseIterGVN* phase) {
5332 assert(!is_complete(), "not already complete");
5333 assert(stores_are_sane(phase), "");
5334 assert(allocation() != nullptr, "must be present");
5335
5336 remove_extra_zeroes();
5337
5338 if (ReduceFieldZeroing || ReduceBulkZeroing)
5339 // reduce instruction count for common initialization patterns
5340 coalesce_subword_stores(header_size, size_in_bytes, phase);
5341
5342 Node* zmem = zero_memory(); // initially zero memory state
5343 Node* inits = zmem; // accumulating a linearized chain of inits
5344 #ifdef ASSERT
5345 intptr_t first_offset = allocation()->minimum_header_size();
5346 intptr_t last_init_off = first_offset; // previous init offset
5347 intptr_t last_init_end = first_offset; // previous init offset+size
5348 intptr_t last_tile_end = first_offset; // previous tile offset+size
5349 #endif
5350 intptr_t zeroes_done = header_size;
5351
5352 bool do_zeroing = true; // we might give up if inits are very sparse
5353 int big_init_gaps = 0; // how many large gaps have we seen?
5354
5355 if (UseTLAB && ZeroTLAB) do_zeroing = false;
5356 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
5357
5358 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5359 Node* st = in(i);
5360 intptr_t st_off = get_store_offset(st, phase);
5361 if (st_off < 0)
5362 break; // unknown junk in the inits
5363 if (st->in(MemNode::Memory) != zmem)
5364 break; // complicated store chains somehow in list
5365
5366 int st_size = st->as_Store()->memory_size();
5367 intptr_t next_init_off = st_off + st_size;
5368
5369 if (do_zeroing && zeroes_done < next_init_off) {
5370 // See if this store needs a zero before it or under it.
5371 intptr_t zeroes_needed = st_off;
5372
5373 if (st_size < BytesPerInt) {
5374 // Look for subword stores which only partially initialize words.
5375 // If we find some, we must lay down some word-level zeroes first,
5376 // underneath the subword stores.
5377 //
5378 // Examples:
5379 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
5380 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
5381 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
5382 //
5383 // Note: coalesce_subword_stores may have already done this,
5384 // if it was prompted by constant non-zero subword initializers.
5385 // But this case can still arise with non-constant stores.
5386
5387 intptr_t next_full_store = find_next_fullword_store(i, phase);
5388
5389 // In the examples above:
5390 // in(i) p q r s x y z
5391 // st_off 12 13 14 15 12 13 14
5392 // st_size 1 1 1 1 1 1 1
5393 // next_full_s. 12 16 16 16 16 16 16
5394 // z's_done 12 16 16 16 12 16 12
5395 // z's_needed 12 16 16 16 16 16 16
5396 // zsize 0 0 0 0 4 0 4
5397 if (next_full_store < 0) {
5398 // Conservative tack: Zero to end of current word.
5399 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
5400 } else {
5401 // Zero to beginning of next fully initialized word.
5402 // Or, don't zero at all, if we are already in that word.
5403 assert(next_full_store >= zeroes_needed, "must go forward");
5404 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
5405 zeroes_needed = next_full_store;
5406 }
5407 }
5408
5409 if (zeroes_needed > zeroes_done) {
5410 intptr_t zsize = zeroes_needed - zeroes_done;
5411 // Do some incremental zeroing on rawmem, in parallel with inits.
5412 zeroes_done = align_down(zeroes_done, BytesPerInt);
5413 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5414 zeroes_done, zeroes_needed,
5415 true,
5416 phase);
5417 zeroes_done = zeroes_needed;
5418 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
5419 do_zeroing = false; // leave the hole, next time
5420 }
5421 }
5422
5423 // Collect the store and move on:
5424 phase->replace_input_of(st, MemNode::Memory, inits);
5425 inits = st; // put it on the linearized chain
5426 set_req(i, zmem); // unhook from previous position
5427
5428 if (zeroes_done == st_off)
5429 zeroes_done = next_init_off;
5430
5431 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
5432
5433 #ifdef ASSERT
5434 // Various order invariants. Weaker than stores_are_sane because
5435 // a large constant tile can be filled in by smaller non-constant stores.
5436 assert(st_off >= last_init_off, "inits do not reverse");
5437 last_init_off = st_off;
5438 const Type* val = nullptr;
5439 if (st_size >= BytesPerInt &&
5440 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
5441 (int)val->basic_type() < (int)T_OBJECT) {
5442 assert(st_off >= last_tile_end, "tiles do not overlap");
5443 assert(st_off >= last_init_end, "tiles do not overwrite inits");
5444 last_tile_end = MAX2(last_tile_end, next_init_off);
5445 } else {
5446 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
5447 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
5448 assert(st_off >= last_init_end, "inits do not overlap");
5449 last_init_end = next_init_off; // it's a non-tile
5450 }
5451 #endif //ASSERT
5452 }
5453
5454 remove_extra_zeroes(); // clear out all the zmems left over
5455 add_req(inits);
5456
5457 if (!(UseTLAB && ZeroTLAB)) {
5458 // If anything remains to be zeroed, zero it all now.
5459 zeroes_done = align_down(zeroes_done, BytesPerInt);
5460 // if it is the last unused 4 bytes of an instance, forget about it
5461 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
5462 if (zeroes_done + BytesPerLong >= size_limit) {
5463 AllocateNode* alloc = allocation();
5464 assert(alloc != nullptr, "must be present");
5465 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
5466 Node* klass_node = alloc->in(AllocateNode::KlassNode);
5467 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
5468 if (zeroes_done == k->layout_helper())
5469 zeroes_done = size_limit;
5470 }
5471 }
5472 if (zeroes_done < size_limit) {
5473 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5474 zeroes_done, size_in_bytes, true, phase);
5475 }
5476 }
5477
5478 set_complete(phase);
5479 return rawmem;
5480 }
5481
5482 void InitializeNode::replace_mem_projs_by(Node* mem, Compile* C) {
5483 auto replace_proj = [&](ProjNode* proj) {
5484 C->gvn_replace_by(proj, mem);
5485 return CONTINUE;
5486 };
5487 apply_to_projs(replace_proj, TypeFunc::Memory);
5488 }
5489
5490 void InitializeNode::replace_mem_projs_by(Node* mem, PhaseIterGVN* igvn) {
5491 DUIterator_Fast imax, i = fast_outs(imax);
5492 auto replace_proj = [&](ProjNode* proj) {
5493 igvn->replace_node(proj, mem);
5494 --i; --imax;
5495 return CONTINUE;
5496 };
5497 apply_to_projs(imax, i, replace_proj, TypeFunc::Memory);
5498 }
5499
5500 bool InitializeNode::already_has_narrow_mem_proj_with_adr_type(const TypePtr* adr_type) const {
5501 auto find_proj = [&](ProjNode* proj) {
5502 if (proj->adr_type() == adr_type) {
5503 return BREAK_AND_RETURN_CURRENT_PROJ;
5504 }
5505 return CONTINUE;
5506 };
5507 DUIterator_Fast imax, i = fast_outs(imax);
5508 return apply_to_narrow_mem_projs_any_iterator(UsesIteratorFast(imax, i, this), find_proj) != nullptr;
5509 }
5510
5511 MachProjNode* InitializeNode::mem_mach_proj() const {
5512 auto find_proj = [](ProjNode* proj) {
5513 if (proj->is_MachProj()) {
5514 return BREAK_AND_RETURN_CURRENT_PROJ;
5515 }
5516 return CONTINUE;
5517 };
5518 ProjNode* proj = apply_to_projs(find_proj, TypeFunc::Memory);
5519 if (proj == nullptr) {
5520 return nullptr;
5521 }
5522 return proj->as_MachProj();
5523 }
5524
5525 #ifdef ASSERT
5526 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
5527 if (is_complete())
5528 return true; // stores could be anything at this point
5529 assert(allocation() != nullptr, "must be present");
5530 intptr_t last_off = allocation()->minimum_header_size();
5531 for (uint i = InitializeNode::RawStores; i < req(); i++) {
5532 Node* st = in(i);
5533 intptr_t st_off = get_store_offset(st, phase);
5534 if (st_off < 0) continue; // ignore dead garbage
5535 if (last_off > st_off) {
5536 tty->print_cr("*** bad store offset at %d: %zd > %zd", i, last_off, st_off);
5537 this->dump(2);
5538 assert(false, "ascending store offsets");
5539 return false;
5540 }
5541 last_off = st_off + st->as_Store()->memory_size();
5542 }
5543 return true;
5544 }
5545 #endif //ASSERT
5546
5547
5548
5549
5550 //============================MergeMemNode=====================================
5551 //
5552 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
5553 // contributing store or call operations. Each contributor provides the memory
5554 // state for a particular "alias type" (see Compile::alias_type). For example,
5555 // if a MergeMem has an input X for alias category #6, then any memory reference
5556 // to alias category #6 may use X as its memory state input, as an exact equivalent
5557 // to using the MergeMem as a whole.
5558 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
5559 //
5560 // (Here, the <N> notation gives the index of the relevant adr_type.)
5561 //
5562 // In one special case (and more cases in the future), alias categories overlap.
5563 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
5564 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
5565 // it is exactly equivalent to that state W:
5566 // MergeMem(<Bot>: W) <==> W
5567 //
5568 // Usually, the merge has more than one input. In that case, where inputs
5569 // overlap (i.e., one is Bot), the narrower alias type determines the memory
5570 // state for that type, and the wider alias type (Bot) fills in everywhere else:
5571 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
5572 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
5573 //
5574 // A merge can take a "wide" memory state as one of its narrow inputs.
5575 // This simply means that the merge observes out only the relevant parts of
5576 // the wide input. That is, wide memory states arriving at narrow merge inputs
5577 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
5578 //
5579 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
5580 // and that memory slices "leak through":
5581 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
5582 //
5583 // But, in such a cascade, repeated memory slices can "block the leak":
5584 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
5585 //
5586 // In the last example, Y is not part of the combined memory state of the
5587 // outermost MergeMem. The system must, of course, prevent unschedulable
5588 // memory states from arising, so you can be sure that the state Y is somehow
5589 // a precursor to state Y'.
5590 //
5591 //
5592 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
5593 // of each MergeMemNode array are exactly the numerical alias indexes, including
5594 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
5595 // Compile::alias_type (and kin) produce and manage these indexes.
5596 //
5597 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
5598 // (Note that this provides quick access to the top node inside MergeMem methods,
5599 // without the need to reach out via TLS to Compile::current.)
5600 //
5601 // As a consequence of what was just described, a MergeMem that represents a full
5602 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
5603 // containing all alias categories.
5604 //
5605 // MergeMem nodes never (?) have control inputs, so in(0) is null.
5606 //
5607 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
5608 // a memory state for the alias type <N>, or else the top node, meaning that
5609 // there is no particular input for that alias type. Note that the length of
5610 // a MergeMem is variable, and may be extended at any time to accommodate new
5611 // memory states at larger alias indexes. When merges grow, they are of course
5612 // filled with "top" in the unused in() positions.
5613 //
5614 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
5615 // (Top was chosen because it works smoothly with passes like GCM.)
5616 //
5617 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
5618 // the type of random VM bits like TLS references.) Since it is always the
5619 // first non-Bot memory slice, some low-level loops use it to initialize an
5620 // index variable: for (i = AliasIdxRaw; i < req(); i++).
5621 //
5622 //
5623 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
5624 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
5625 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
5626 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
5627 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
5628 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
5629 //
5630 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
5631 // really that different from the other memory inputs. An abbreviation called
5632 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
5633 //
5634 //
5635 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
5636 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
5637 // that "emerges though" the base memory will be marked as excluding the alias types
5638 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
5639 //
5640 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
5641 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
5642 //
5643 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
5644 // (It is currently unimplemented.) As you can see, the resulting merge is
5645 // actually a disjoint union of memory states, rather than an overlay.
5646 //
5647
5648 //------------------------------MergeMemNode-----------------------------------
5649 Node* MergeMemNode::make_empty_memory() {
5650 Node* empty_memory = (Node*) Compile::current()->top();
5651 assert(empty_memory->is_top(), "correct sentinel identity");
5652 return empty_memory;
5653 }
5654
5655 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
5656 init_class_id(Class_MergeMem);
5657 // all inputs are nullified in Node::Node(int)
5658 // set_input(0, nullptr); // no control input
5659
5660 // Initialize the edges uniformly to top, for starters.
5661 Node* empty_mem = make_empty_memory();
5662 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
5663 init_req(i,empty_mem);
5664 }
5665 assert(empty_memory() == empty_mem, "");
5666
5667 if( new_base != nullptr && new_base->is_MergeMem() ) {
5668 MergeMemNode* mdef = new_base->as_MergeMem();
5669 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
5670 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
5671 mms.set_memory(mms.memory2());
5672 }
5673 assert(base_memory() == mdef->base_memory(), "");
5674 } else {
5675 set_base_memory(new_base);
5676 }
5677 }
5678
5679 // Make a new, untransformed MergeMem with the same base as 'mem'.
5680 // If mem is itself a MergeMem, populate the result with the same edges.
5681 MergeMemNode* MergeMemNode::make(Node* mem) {
5682 return new MergeMemNode(mem);
5683 }
5684
5685 //------------------------------cmp--------------------------------------------
5686 uint MergeMemNode::hash() const { return NO_HASH; }
5687 bool MergeMemNode::cmp( const Node &n ) const {
5688 return (&n == this); // Always fail except on self
5689 }
5690
5691 //------------------------------Identity---------------------------------------
5692 Node* MergeMemNode::Identity(PhaseGVN* phase) {
5693 // Identity if this merge point does not record any interesting memory
5694 // disambiguations.
5695 Node* base_mem = base_memory();
5696 Node* empty_mem = empty_memory();
5697 if (base_mem != empty_mem) { // Memory path is not dead?
5698 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5699 Node* mem = in(i);
5700 if (mem != empty_mem && mem != base_mem) {
5701 return this; // Many memory splits; no change
5702 }
5703 }
5704 }
5705 return base_mem; // No memory splits; ID on the one true input
5706 }
5707
5708 //------------------------------Ideal------------------------------------------
5709 // This method is invoked recursively on chains of MergeMem nodes
5710 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
5711 // Remove chain'd MergeMems
5712 //
5713 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
5714 // relative to the "in(Bot)". Since we are patching both at the same time,
5715 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
5716 // but rewrite each "in(i)" relative to the new "in(Bot)".
5717 Node *progress = nullptr;
5718
5719
5720 Node* old_base = base_memory();
5721 Node* empty_mem = empty_memory();
5722 if (old_base == empty_mem)
5723 return nullptr; // Dead memory path.
5724
5725 MergeMemNode* old_mbase;
5726 if (old_base != nullptr && old_base->is_MergeMem())
5727 old_mbase = old_base->as_MergeMem();
5728 else
5729 old_mbase = nullptr;
5730 Node* new_base = old_base;
5731
5732 // simplify stacked MergeMems in base memory
5733 if (old_mbase) new_base = old_mbase->base_memory();
5734
5735 // the base memory might contribute new slices beyond my req()
5736 if (old_mbase) grow_to_match(old_mbase);
5737
5738 // Note: We do not call verify_sparse on entry, because inputs
5739 // can normalize to the base_memory via subsume_node or similar
5740 // mechanisms. This method repairs that damage.
5741
5742 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
5743
5744 // Look at each slice.
5745 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5746 Node* old_in = in(i);
5747 // calculate the old memory value
5748 Node* old_mem = old_in;
5749 if (old_mem == empty_mem) old_mem = old_base;
5750 assert(old_mem == memory_at(i), "");
5751
5752 // maybe update (reslice) the old memory value
5753
5754 // simplify stacked MergeMems
5755 Node* new_mem = old_mem;
5756 MergeMemNode* old_mmem;
5757 if (old_mem != nullptr && old_mem->is_MergeMem())
5758 old_mmem = old_mem->as_MergeMem();
5759 else
5760 old_mmem = nullptr;
5761 if (old_mmem == this) {
5762 // This can happen if loops break up and safepoints disappear.
5763 // A merge of BotPtr (default) with a RawPtr memory derived from a
5764 // safepoint can be rewritten to a merge of the same BotPtr with
5765 // the BotPtr phi coming into the loop. If that phi disappears
5766 // also, we can end up with a self-loop of the mergemem.
5767 // In general, if loops degenerate and memory effects disappear,
5768 // a mergemem can be left looking at itself. This simply means
5769 // that the mergemem's default should be used, since there is
5770 // no longer any apparent effect on this slice.
5771 // Note: If a memory slice is a MergeMem cycle, it is unreachable
5772 // from start. Update the input to TOP.
5773 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
5774 }
5775 else if (old_mmem != nullptr) {
5776 new_mem = old_mmem->memory_at(i);
5777 }
5778 // else preceding memory was not a MergeMem
5779
5780 // maybe store down a new value
5781 Node* new_in = new_mem;
5782 if (new_in == new_base) new_in = empty_mem;
5783
5784 if (new_in != old_in) {
5785 // Warning: Do not combine this "if" with the previous "if"
5786 // A memory slice might have be be rewritten even if it is semantically
5787 // unchanged, if the base_memory value has changed.
5788 set_req_X(i, new_in, phase);
5789 progress = this; // Report progress
5790 }
5791 }
5792
5793 if (new_base != old_base) {
5794 set_req_X(Compile::AliasIdxBot, new_base, phase);
5795 // Don't use set_base_memory(new_base), because we need to update du.
5796 assert(base_memory() == new_base, "");
5797 progress = this;
5798 }
5799
5800 if( base_memory() == this ) {
5801 // a self cycle indicates this memory path is dead
5802 set_req(Compile::AliasIdxBot, empty_mem);
5803 }
5804
5805 // Resolve external cycles by calling Ideal on a MergeMem base_memory
5806 // Recursion must occur after the self cycle check above
5807 if( base_memory()->is_MergeMem() ) {
5808 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
5809 Node *m = phase->transform(new_mbase); // Rollup any cycles
5810 if( m != nullptr &&
5811 (m->is_top() ||
5812 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
5813 // propagate rollup of dead cycle to self
5814 set_req(Compile::AliasIdxBot, empty_mem);
5815 }
5816 }
5817
5818 if( base_memory() == empty_mem ) {
5819 progress = this;
5820 // Cut inputs during Parse phase only.
5821 // During Optimize phase a dead MergeMem node will be subsumed by Top.
5822 if( !can_reshape ) {
5823 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5824 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
5825 }
5826 }
5827 }
5828
5829 if( !progress && base_memory()->is_Phi() && can_reshape ) {
5830 // Check if PhiNode::Ideal's "Split phis through memory merges"
5831 // transform should be attempted. Look for this->phi->this cycle.
5832 uint merge_width = req();
5833 if (merge_width > Compile::AliasIdxRaw) {
5834 PhiNode* phi = base_memory()->as_Phi();
5835 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
5836 if (phi->in(i) == this) {
5837 phase->is_IterGVN()->_worklist.push(phi);
5838 break;
5839 }
5840 }
5841 }
5842 }
5843
5844 assert(progress || verify_sparse(), "please, no dups of base");
5845 return progress;
5846 }
5847
5848 //-------------------------set_base_memory-------------------------------------
5849 void MergeMemNode::set_base_memory(Node *new_base) {
5850 Node* empty_mem = empty_memory();
5851 set_req(Compile::AliasIdxBot, new_base);
5852 assert(memory_at(req()) == new_base, "must set default memory");
5853 // Clear out other occurrences of new_base:
5854 if (new_base != empty_mem) {
5855 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5856 if (in(i) == new_base) set_req(i, empty_mem);
5857 }
5858 }
5859 }
5860
5861 //------------------------------out_RegMask------------------------------------
5862 const RegMask &MergeMemNode::out_RegMask() const {
5863 return RegMask::EMPTY;
5864 }
5865
5866 //------------------------------dump_spec--------------------------------------
5867 #ifndef PRODUCT
5868 void MergeMemNode::dump_spec(outputStream *st) const {
5869 st->print(" {");
5870 Node* base_mem = base_memory();
5871 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
5872 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem;
5873 if (mem == base_mem) { st->print(" -"); continue; }
5874 st->print( " N%d:", mem->_idx );
5875 Compile::current()->get_adr_type(i)->dump_on(st);
5876 }
5877 st->print(" }");
5878 }
5879 #endif // !PRODUCT
5880
5881
5882 #ifdef ASSERT
5883 static bool might_be_same(Node* a, Node* b) {
5884 if (a == b) return true;
5885 if (!(a->is_Phi() || b->is_Phi())) return false;
5886 // phis shift around during optimization
5887 return true; // pretty stupid...
5888 }
5889
5890 // verify a narrow slice (either incoming or outgoing)
5891 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
5892 if (!VerifyAliases) return; // don't bother to verify unless requested
5893 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
5894 if (Node::in_dump()) return; // muzzle asserts when printing
5895 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
5896 assert(n != nullptr, "");
5897 // Elide intervening MergeMem's
5898 while (n->is_MergeMem()) {
5899 n = n->as_MergeMem()->memory_at(alias_idx);
5900 }
5901 Compile* C = Compile::current();
5902 const TypePtr* n_adr_type = n->adr_type();
5903 if (n == m->empty_memory()) {
5904 // Implicit copy of base_memory()
5905 } else if (n_adr_type != TypePtr::BOTTOM) {
5906 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type");
5907 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
5908 } else {
5909 // A few places like make_runtime_call "know" that VM calls are narrow,
5910 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
5911 bool expected_wide_mem = false;
5912 if (n == m->base_memory()) {
5913 expected_wide_mem = true;
5914 } else if (alias_idx == Compile::AliasIdxRaw ||
5915 n == m->memory_at(Compile::AliasIdxRaw)) {
5916 expected_wide_mem = true;
5917 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
5918 // memory can "leak through" calls on channels that
5919 // are write-once. Allow this also.
5920 expected_wide_mem = true;
5921 }
5922 assert(expected_wide_mem, "expected narrow slice replacement");
5923 }
5924 }
5925 #else // !ASSERT
5926 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
5927 #endif
5928
5929
5930 //-----------------------------memory_at---------------------------------------
5931 Node* MergeMemNode::memory_at(uint alias_idx) const {
5932 assert(alias_idx >= Compile::AliasIdxRaw ||
5933 (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()),
5934 "must avoid base_memory and AliasIdxTop");
5935
5936 // Otherwise, it is a narrow slice.
5937 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
5938 if (is_empty_memory(n)) {
5939 // the array is sparse; empty slots are the "top" node
5940 n = base_memory();
5941 assert(Node::in_dump()
5942 || n == nullptr || n->bottom_type() == Type::TOP
5943 || n->adr_type() == nullptr // address is TOP
5944 || n->adr_type() == TypePtr::BOTTOM
5945 || n->adr_type() == TypeRawPtr::BOTTOM
5946 || n->is_NarrowMemProj()
5947 || !Compile::current()->do_aliasing(),
5948 "must be a wide memory");
5949 // do_aliasing == false if we are organizing the memory states manually.
5950 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
5951 } else {
5952 // make sure the stored slice is sane
5953 #ifdef ASSERT
5954 if (VMError::is_error_reported() || Node::in_dump()) {
5955 } else if (might_be_same(n, base_memory())) {
5956 // Give it a pass: It is a mostly harmless repetition of the base.
5957 // This can arise normally from node subsumption during optimization.
5958 } else {
5959 verify_memory_slice(this, alias_idx, n);
5960 }
5961 #endif
5962 }
5963 return n;
5964 }
5965
5966 //---------------------------set_memory_at-------------------------------------
5967 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
5968 verify_memory_slice(this, alias_idx, n);
5969 Node* empty_mem = empty_memory();
5970 if (n == base_memory()) n = empty_mem; // collapse default
5971 uint need_req = alias_idx+1;
5972 if (req() < need_req) {
5973 if (n == empty_mem) return; // already the default, so do not grow me
5974 // grow the sparse array
5975 do {
5976 add_req(empty_mem);
5977 } while (req() < need_req);
5978 }
5979 set_req( alias_idx, n );
5980 }
5981
5982
5983
5984 //--------------------------iteration_setup------------------------------------
5985 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
5986 if (other != nullptr) {
5987 grow_to_match(other);
5988 // invariant: the finite support of mm2 is within mm->req()
5989 #ifdef ASSERT
5990 for (uint i = req(); i < other->req(); i++) {
5991 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
5992 }
5993 #endif
5994 }
5995 // Replace spurious copies of base_memory by top.
5996 Node* base_mem = base_memory();
5997 if (base_mem != nullptr && !base_mem->is_top()) {
5998 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
5999 if (in(i) == base_mem)
6000 set_req(i, empty_memory());
6001 }
6002 }
6003 }
6004
6005 //---------------------------grow_to_match-------------------------------------
6006 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
6007 Node* empty_mem = empty_memory();
6008 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
6009 // look for the finite support of the other memory
6010 for (uint i = other->req(); --i >= req(); ) {
6011 if (other->in(i) != empty_mem) {
6012 uint new_len = i+1;
6013 while (req() < new_len) add_req(empty_mem);
6014 break;
6015 }
6016 }
6017 }
6018
6019 //---------------------------verify_sparse-------------------------------------
6020 #ifndef PRODUCT
6021 bool MergeMemNode::verify_sparse() const {
6022 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
6023 Node* base_mem = base_memory();
6024 // The following can happen in degenerate cases, since empty==top.
6025 if (is_empty_memory(base_mem)) return true;
6026 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6027 assert(in(i) != nullptr, "sane slice");
6028 if (in(i) == base_mem) return false; // should have been the sentinel value!
6029 }
6030 return true;
6031 }
6032
6033 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
6034 Node* n;
6035 n = mm->in(idx);
6036 if (mem == n) return true; // might be empty_memory()
6037 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
6038 if (mem == n) return true;
6039 return false;
6040 }
6041 #endif // !PRODUCT