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