1 /*
2 * Copyright (c) 1997, 2026, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2024, Alibaba Group Holding Limited. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "classfile/javaClasses.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "gc/shared/tlab_globals.hpp"
31 #include "memory/allocation.inline.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "oops/objArrayKlass.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/arraycopynode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/compile.hpp"
38 #include "opto/connode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/loopnode.hpp"
41 #include "opto/machnode.hpp"
42 #include "opto/matcher.hpp"
43 #include "opto/memnode.hpp"
44 #include "opto/mempointer.hpp"
45 #include "opto/mulnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opcodes.hpp"
48 #include "opto/phaseX.hpp"
49 #include "opto/regalloc.hpp"
50 #include "opto/regmask.hpp"
51 #include "opto/rootnode.hpp"
52 #include "opto/traceMergeStoresTag.hpp"
53 #include "opto/vectornode.hpp"
54 #include "utilities/align.hpp"
55 #include "utilities/copy.hpp"
56 #include "utilities/macros.hpp"
57 #include "utilities/powerOfTwo.hpp"
58 #include "utilities/vmError.hpp"
59
60 // Portions of code courtesy of Clifford Click
61
62 // Optimization - Graph Style
63
64 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
65
66 //=============================================================================
67 uint MemNode::size_of() const { return sizeof(*this); }
68
69 const TypePtr *MemNode::adr_type() const {
70 Node* adr = in(Address);
71 if (adr == nullptr) return nullptr; // node is dead
72 const TypePtr* cross_check = nullptr;
73 DEBUG_ONLY(cross_check = _adr_type);
74 return calculate_adr_type(adr->bottom_type(), cross_check);
75 }
76
77 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
78 if (adr != nullptr) {
79 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
80 return true;
81 }
82 }
83 return false;
84 }
85
86 #ifndef PRODUCT
87 void MemNode::dump_spec(outputStream *st) const {
88 if (in(Address) == nullptr) return; // node is dead
89 #ifndef ASSERT
90 // fake the missing field
91 const TypePtr* _adr_type = nullptr;
92 if (in(Address) != nullptr)
93 _adr_type = in(Address)->bottom_type()->isa_ptr();
94 #endif
95 dump_adr_type(_adr_type, st);
96
97 Compile* C = Compile::current();
98 if (C->alias_type(_adr_type)->is_volatile()) {
99 st->print(" Volatile!");
100 }
101 if (_unaligned_access) {
102 st->print(" unaligned");
103 }
104 if (_mismatched_access) {
105 st->print(" mismatched");
106 }
107 if (_unsafe_access) {
108 st->print(" unsafe");
109 }
110 }
111
112 void MemNode::dump_adr_type(const 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 // Trivial case: Non-overlapping values. Be careful, we can cast a raw pointer to an oop (e.g. in
557 // the allocation pattern) so joining the types only works if both are oops. join may also give
558 // an incorrect result when both pointers are nullable and the result is supposed to be
559 // TypePtr::NULL_PTR, so we exclude that case.
560 const Type* p1_type = p1->bottom_type();
561 const Type* p2_type = p2->bottom_type();
562 if (p1_type->isa_oopptr() && p2_type->isa_oopptr() &&
563 (!p1_type->maybe_null() || !p2_type->maybe_null()) &&
564 p1_type->join(p2_type)->empty()) {
565 return true;
566 }
567
568 // Attempt to prove that these two pointers cannot be aliased.
569 // They may both manifestly be allocations, and they should differ.
570 // Or, if they are not both allocations, they can be distinct constants.
571 // Otherwise, one is an allocation and the other a pre-existing value.
572 if (a1 == nullptr && a2 == nullptr) { // neither an allocation
573 return (p1 != p2) && p1->is_Con() && p2->is_Con();
574 } else if (a1 != nullptr && a2 != nullptr) { // both allocations
575 return (a1 != a2);
576 } else if (a1 != nullptr) { // one allocation a1
577 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
578 return all_controls_dominate(p2, a1);
579 } else { //(a2 != null) // one allocation a2
580 return all_controls_dominate(p1, a2);
581 }
582 return false;
583 }
584
585
586 // Find an arraycopy ac that produces the memory state represented by parameter mem.
587 // Return ac if
588 // (a) can_see_stored_value=true and ac must have set the value for this load or if
589 // (b) can_see_stored_value=false and ac could have set the value for this load or if
590 // (c) can_see_stored_value=false and ac cannot have set the value for this load.
591 // In case (c) change the parameter mem to the memory input of ac to skip it
592 // when searching stored value.
593 // Otherwise return null.
594 Node* LoadNode::find_previous_arraycopy(PhaseValues* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
595 ArrayCopyNode* ac = find_array_copy_clone(ld_alloc, mem);
596 if (ac != nullptr) {
597 Node* ld_addp = in(MemNode::Address);
598 Node* src = ac->in(ArrayCopyNode::Src);
599 const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
600
601 // This is a load from a cloned array. The corresponding arraycopy ac must
602 // have set the value for the load and we can return ac but only if the load
603 // is known to be within bounds. This is checked below.
604 if (ary_t != nullptr && ld_addp->is_AddP()) {
605 Node* ld_offs = ld_addp->in(AddPNode::Offset);
606 BasicType ary_elem = ary_t->elem()->array_element_basic_type();
607 jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
608 jlong elemsize = type2aelembytes(ary_elem);
609
610 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
611 const TypeInt* sizetype = ary_t->size();
612
613 if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
614 // The load is known to be within bounds. It receives its value from ac.
615 return ac;
616 }
617 // The load is known to be out-of-bounds.
618 }
619 // The load could be out-of-bounds. It must not be hoisted but must remain
620 // dependent on the runtime range check. This is achieved by returning null.
621 } else if (mem->is_Proj() && mem->in(0) != nullptr && mem->in(0)->is_ArrayCopy()) {
622 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
623
624 if (ac->is_arraycopy_validated() ||
625 ac->is_copyof_validated() ||
626 ac->is_copyofrange_validated()) {
627 Node* ld_addp = in(MemNode::Address);
628 if (ld_addp->is_AddP()) {
629 Node* ld_base = ld_addp->in(AddPNode::Address);
630 Node* ld_offs = ld_addp->in(AddPNode::Offset);
631
632 Node* dest = ac->in(ArrayCopyNode::Dest);
633
634 if (dest == ld_base) {
635 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
636 assert(!ld_offs_t->empty(), "dead reference should be checked already");
637 // Take into account vector or unsafe access size
638 jlong ld_size_in_bytes = (jlong)memory_size();
639 jlong offset_hi = ld_offs_t->_hi + ld_size_in_bytes - 1;
640 offset_hi = MIN2(offset_hi, (jlong)(TypeX::MAX->_hi)); // Take care for overflow in 32-bit VM
641 if (ac->modifies(ld_offs_t->_lo, (intptr_t)offset_hi, phase, can_see_stored_value)) {
642 return ac;
643 }
644 if (!can_see_stored_value) {
645 mem = ac->in(TypeFunc::Memory);
646 return ac;
647 }
648 }
649 }
650 }
651 }
652 return nullptr;
653 }
654
655 ArrayCopyNode* MemNode::find_array_copy_clone(Node* ld_alloc, Node* mem) const {
656 if (mem->is_Proj() && mem->in(0) != nullptr && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
657 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
658 if (ld_alloc != nullptr) {
659 // Check if there is an array copy for a clone
660 Node* mb = mem->in(0);
661 ArrayCopyNode* ac = nullptr;
662 if (mb->in(0) != nullptr && mb->in(0)->is_Proj() &&
663 mb->in(0)->in(0) != nullptr && mb->in(0)->in(0)->is_ArrayCopy()) {
664 ac = mb->in(0)->in(0)->as_ArrayCopy();
665 } else {
666 // Step over GC barrier when ReduceInitialCardMarks is disabled
667 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
668 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
669
670 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
671 ac = control_proj_ac->in(0)->as_ArrayCopy();
672 }
673 }
674
675 if (ac != nullptr && ac->is_clonebasic()) {
676 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest));
677 if (alloc != nullptr && alloc == ld_alloc) {
678 return ac;
679 }
680 }
681 }
682 }
683 return nullptr;
684 }
685
686 // The logic for reordering loads and stores uses four steps:
687 // (a) Walk carefully past stores and initializations which we
688 // can prove are independent of this load.
689 // (b) Observe that the next memory state makes an exact match
690 // with self (load or store), and locate the relevant store.
691 // (c) Ensure that, if we were to wire self directly to the store,
692 // the optimizer would fold it up somehow.
693 // (d) Do the rewiring, and return, depending on some other part of
694 // the optimizer to fold up the load.
695 // This routine handles steps (a) and (b). Steps (c) and (d) are
696 // specific to loads and stores, so they are handled by the callers.
697 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
698 //
699 Node* MemNode::find_previous_store(PhaseValues* phase) {
700 Node* ctrl = in(MemNode::Control);
701 Node* adr = in(MemNode::Address);
702 intptr_t offset = 0;
703 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
704 AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
705
706 const TypePtr* adr_type = this->adr_type();
707 if (adr_type == nullptr) {
708 // This means the access is dead
709 return phase->C->top();
710 } else if (adr_type->base() == TypePtr::AnyPtr) {
711 assert(adr_type->ptr() == TypePtr::Null, "MemNode should never access a wide memory");
712 // Give up, this will upset Compile::get_alias_index
713 return nullptr;
714 }
715
716 int alias_idx = phase->C->get_alias_index(adr_type);
717 assert(alias_idx != Compile::AliasIdxTop, "must not be a dead node");
718 assert(alias_idx != Compile::AliasIdxBot || !phase->C->do_aliasing(), "must not be a very wide access");
719
720 if (offset == Type::OffsetBot)
721 return nullptr; // cannot unalias unless there are precise offsets
722
723 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
724 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
725
726 intptr_t size_in_bytes = memory_size();
727
728 Node* mem = in(MemNode::Memory); // start searching here...
729
730 int cnt = 50; // Cycle limiter
731 for (;;) { // While we can dance past unrelated stores...
732 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
733
734 Node* prev = mem;
735 if (mem->is_Store()) {
736 Node* st_adr = mem->in(MemNode::Address);
737 intptr_t st_offset = 0;
738 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
739 if (st_base == nullptr) {
740 // inscrutable pointer
741 break;
742 }
743
744 // If the bases are the same and the offsets are the same, it seems that this is the exact
745 // store we are looking for, the caller will check if the type of the store matches using
746 // MemNode::can_see_stored_value
747 if (st_base == base && st_offset == offset) {
748 return mem; // (b) found the store that this access observes
749 }
750
751 // If it is provable that the memory accessed by mem does not overlap the memory accessed by
752 // this, we may walk past mem.
753 // For raw accesses, 2 accesses are independent if they have the same base and the offsets
754 // say that they do not overlap.
755 // For heap accesses, 2 accesses are independent if either the bases are provably different
756 // at runtime or the offsets say that the accesses do not overlap.
757 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
758 // Raw accesses can only be provably independent if they have the same base
759 break;
760 }
761
762 // If the offsets say that the accesses do not overlap, then it is provable that mem and this
763 // do not overlap. For example, a LoadI from Object+8 is independent from a StoreL into
764 // Object+12, no matter what the bases are.
765 if (st_offset != offset && st_offset != Type::OffsetBot) {
766 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
767 assert(mem->as_Store()->memory_size() <= MAX_STORE, "");
768 if (st_offset >= offset + size_in_bytes ||
769 st_offset <= offset - MAX_STORE ||
770 st_offset <= offset - mem->as_Store()->memory_size()) {
771 // Success: The offsets are provably independent.
772 // (You may ask, why not just test st_offset != offset and be done?
773 // The answer is that stores of different sizes can co-exist
774 // in the same sequence of RawMem effects. We sometimes initialize
775 // a whole 'tile' of array elements with a single jint or jlong.)
776 mem = mem->in(MemNode::Memory);
777 continue; // (a) advance through the independent store
778 }
779 }
780
781 // Same base and overlapping offsets, it seems provable that the accesses overlap, give up
782 if (st_base == base) {
783 break;
784 }
785
786 // Try to prove that 2 different base nodes at compile time are different values at runtime
787 bool known_independent = false;
788 if (detect_ptr_independence(base, alloc, st_base, AllocateNode::Ideal_allocation(st_base), phase)) {
789 known_independent = true;
790 }
791
792 if (known_independent) {
793 mem = mem->in(MemNode::Memory);
794 continue; // (a) advance through the independent store
795 }
796 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
797 InitializeNode* st_init = mem->in(0)->as_Initialize();
798 AllocateNode* st_alloc = st_init->allocation();
799 if (st_alloc == nullptr) {
800 break; // something degenerated
801 }
802 bool known_identical = false;
803 bool known_independent = false;
804 if (alloc == st_alloc) {
805 known_identical = true;
806 } else if (alloc != nullptr) {
807 known_independent = true;
808 } else if (all_controls_dominate(this, st_alloc)) {
809 known_independent = true;
810 }
811
812 if (known_independent) {
813 // The bases are provably independent: Either they are
814 // manifestly distinct allocations, or else the control
815 // of this load dominates the store's allocation.
816 if (alias_idx == Compile::AliasIdxRaw) {
817 mem = st_alloc->in(TypeFunc::Memory);
818 } else {
819 mem = st_init->memory(alias_idx);
820 }
821 continue; // (a) advance through independent store memory
822 }
823
824 // (b) at this point, if we are not looking at a store initializing
825 // the same allocation we are loading from, we lose.
826 if (known_identical) {
827 // From caller, can_see_stored_value will consult find_captured_store.
828 return mem; // let caller handle steps (c), (d)
829 }
830
831 } else if (find_previous_arraycopy(phase, alloc, mem, false) != nullptr) {
832 if (prev != mem) {
833 // Found an arraycopy but it doesn't affect that load
834 continue;
835 }
836 // Found an arraycopy that may affect that load
837 return mem;
838 } else if (mem->is_MergeMem()) {
839 mem = mem->as_MergeMem()->memory_at(alias_idx);
840 continue;
841 } else if (addr_t != nullptr && addr_t->is_known_instance_field()) {
842 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
843 if (mem->is_Proj() && mem->in(0)->is_Call()) {
844 // ArrayCopyNodes processed here as well.
845 CallNode *call = mem->in(0)->as_Call();
846 if (!call->may_modify(addr_t, phase)) {
847 mem = call->in(TypeFunc::Memory);
848 continue; // (a) advance through independent call memory
849 }
850 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
851 ArrayCopyNode* ac = nullptr;
852 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
853 break;
854 }
855 mem = mem->in(0)->in(TypeFunc::Memory);
856 continue; // (a) advance through independent MemBar memory
857 } else if (mem->is_ClearArray()) {
858 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
859 // (the call updated 'mem' value)
860 continue; // (a) advance through independent allocation memory
861 } else {
862 // Can not bypass initialization of the instance
863 // we are looking for.
864 return mem;
865 }
866 }
867 }
868
869 // Unless there is an explicit 'continue', we must bail out here,
870 // because 'mem' is an inscrutable memory state (e.g., a call).
871 break;
872 }
873
874 return nullptr; // bail out
875 }
876
877 //----------------------calculate_adr_type-------------------------------------
878 // Helper function. Notices when the given type of address hits top or bottom.
879 // Also, asserts a cross-check of the type against the expected address type.
880 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
881 if (t == Type::TOP) return nullptr; // does not touch memory any more?
882 #ifdef ASSERT
883 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = nullptr;
884 #endif
885 const TypePtr* tp = t->isa_ptr();
886 if (tp == nullptr) {
887 assert(cross_check == nullptr || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
888 return TypePtr::BOTTOM; // touches lots of memory
889 } else {
890 #ifdef ASSERT
891 // %%%% [phh] We don't check the alias index if cross_check is
892 // TypeRawPtr::BOTTOM. Needs to be investigated.
893 if (cross_check != nullptr &&
894 cross_check != TypePtr::BOTTOM &&
895 cross_check != TypeRawPtr::BOTTOM) {
896 // Recheck the alias index, to see if it has changed (due to a bug).
897 Compile* C = Compile::current();
898 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
899 "must stay in the original alias category");
900 // The type of the address must be contained in the adr_type,
901 // disregarding "null"-ness.
902 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
903 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
904 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
905 "real address must not escape from expected memory type");
906 }
907 #endif
908 return tp;
909 }
910 }
911
912 uint8_t MemNode::barrier_data(const Node* n) {
913 if (n->is_LoadStore()) {
914 return n->as_LoadStore()->barrier_data();
915 } else if (n->is_Mem()) {
916 return n->as_Mem()->barrier_data();
917 }
918 return 0;
919 }
920
921 //=============================================================================
922 // Should LoadNode::Ideal() attempt to remove control edges?
923 bool LoadNode::can_remove_control() const {
924 return !has_pinned_control_dependency();
925 }
926 uint LoadNode::size_of() const { return sizeof(*this); }
927 bool LoadNode::cmp(const Node &n) const {
928 LoadNode& load = (LoadNode &)n;
929 return Type::equals(_type, load._type) &&
930 _control_dependency == load._control_dependency &&
931 _mo == load._mo;
932 }
933 const Type *LoadNode::bottom_type() const { return _type; }
934 uint LoadNode::ideal_reg() const {
935 return _type->ideal_reg();
936 }
937
938 #ifndef PRODUCT
939 void LoadNode::dump_spec(outputStream *st) const {
940 MemNode::dump_spec(st);
941 if( !Verbose && !WizardMode ) {
942 // standard dump does this in Verbose and WizardMode
943 st->print(" #"); _type->dump_on(st);
944 }
945 if (in(0) != nullptr && !depends_only_on_test()) {
946 st->print(" (does not depend only on test, ");
947 if (control_dependency() == UnknownControl) {
948 st->print("unknown control");
949 } else if (control_dependency() == Pinned) {
950 st->print("pinned");
951 } else if (adr_type() == TypeRawPtr::BOTTOM) {
952 st->print("raw access");
953 } else {
954 st->print("unknown reason");
955 }
956 st->print(")");
957 }
958 }
959 #endif
960
961 #ifdef ASSERT
962 //----------------------------is_immutable_value-------------------------------
963 // Helper function to allow a raw load without control edge for some cases
964 bool LoadNode::is_immutable_value(Node* adr) {
965 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
966 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) {
967
968 jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1);
969 int offsets[] = {
970 in_bytes(JavaThread::osthread_offset()),
971 in_bytes(JavaThread::threadObj_offset()),
972 in_bytes(JavaThread::vthread_offset()),
973 in_bytes(JavaThread::scopedValueCache_offset()),
974 };
975
976 for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) {
977 if (offset == offsets[i]) {
978 return true;
979 }
980 }
981 }
982
983 return false;
984 }
985 #endif
986
987 //----------------------------LoadNode::make-----------------------------------
988 // Polymorphic factory method:
989 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo,
990 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
991 Compile* C = gvn.C;
992
993 // sanity check the alias category against the created node type
994 assert(!(adr_type->isa_oopptr() &&
995 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
996 "use LoadKlassNode instead");
997 assert(!(adr_type->isa_aryptr() &&
998 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
999 "use LoadRangeNode instead");
1000 // Check control edge of raw loads
1001 assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
1002 // oop will be recorded in oop map if load crosses safepoint
1003 rt->isa_oopptr() || is_immutable_value(adr),
1004 "raw memory operations should have control edge");
1005 LoadNode* load = nullptr;
1006 switch (bt) {
1007 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1008 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1009 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1010 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1011 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1012 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break;
1013 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
1014 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break;
1015 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
1016 case T_OBJECT:
1017 case T_NARROWOOP:
1018 #ifdef _LP64
1019 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1020 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
1021 } else
1022 #endif
1023 {
1024 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
1025 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
1026 }
1027 break;
1028 default:
1029 ShouldNotReachHere();
1030 break;
1031 }
1032 assert(load != nullptr, "LoadNode should have been created");
1033 if (unaligned) {
1034 load->set_unaligned_access();
1035 }
1036 if (mismatched) {
1037 load->set_mismatched_access();
1038 }
1039 if (unsafe) {
1040 load->set_unsafe_access();
1041 }
1042 load->set_barrier_data(barrier_data);
1043 if (load->Opcode() == Op_LoadN) {
1044 Node* ld = gvn.transform(load);
1045 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
1046 }
1047
1048 return load;
1049 }
1050
1051 //------------------------------hash-------------------------------------------
1052 uint LoadNode::hash() const {
1053 // unroll addition of interesting fields
1054 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
1055 }
1056
1057 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
1058 if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) {
1059 bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile();
1060 bool is_stable_ary = FoldStableValues &&
1061 (tp != nullptr) && (tp->isa_aryptr() != nullptr) &&
1062 tp->isa_aryptr()->is_stable();
1063
1064 return (eliminate_boxing && non_volatile) || is_stable_ary;
1065 }
1066
1067 return false;
1068 }
1069
1070 // Is the value loaded previously stored by an arraycopy? If so return
1071 // a load node that reads from the source array so we may be able to
1072 // optimize out the ArrayCopy node later.
1073 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
1074 Node* ld_adr = in(MemNode::Address);
1075 intptr_t ld_off = 0;
1076 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
1077 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
1078 if (ac != nullptr) {
1079 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
1080
1081 Node* mem = ac->in(TypeFunc::Memory);
1082 Node* ctl = ac->in(0);
1083 Node* src = ac->in(ArrayCopyNode::Src);
1084
1085 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
1086 return nullptr;
1087 }
1088
1089 // load depends on the tests that validate the arraycopy
1090 LoadNode* ld = clone_pinned();
1091 Node* addp = in(MemNode::Address)->clone();
1092 if (ac->as_ArrayCopy()->is_clonebasic()) {
1093 assert(ld_alloc != nullptr, "need an alloc");
1094 assert(addp->is_AddP(), "address must be addp");
1095 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1096 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1097 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1098 addp->set_req(AddPNode::Base, src);
1099 addp->set_req(AddPNode::Address, src);
1100 } else {
1101 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1102 ac->as_ArrayCopy()->is_copyof_validated() ||
1103 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1104 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1105 addp->set_req(AddPNode::Base, src);
1106 addp->set_req(AddPNode::Address, src);
1107
1108 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1109 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
1110 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
1111
1112 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
1113 uint shift = exact_log2(type2aelembytes(ary_elem));
1114
1115 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1116 #ifdef _LP64
1117 diff = phase->transform(new ConvI2LNode(diff));
1118 #endif
1119 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1120
1121 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1122 addp->set_req(AddPNode::Offset, offset);
1123 }
1124 addp = phase->transform(addp);
1125 #ifdef ASSERT
1126 const TypePtr* adr_type = phase->type(addp)->is_ptr();
1127 ld->_adr_type = adr_type;
1128 #endif
1129 ld->set_req(MemNode::Address, addp);
1130 ld->set_req(0, ctl);
1131 ld->set_req(MemNode::Memory, mem);
1132 return ld;
1133 }
1134 return nullptr;
1135 }
1136
1137
1138 //---------------------------can_see_stored_value------------------------------
1139 // This routine exists to make sure this set of tests is done the same
1140 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
1141 // will change the graph shape in a way which makes memory alive twice at the
1142 // same time (uses the Oracle model of aliasing), then some
1143 // LoadXNode::Identity will fold things back to the equivalence-class model
1144 // of aliasing.
1145 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const {
1146 Node* ld_adr = in(MemNode::Address);
1147 intptr_t ld_off = 0;
1148 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1149 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base);
1150 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1151 Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr;
1152 // This is more general than load from boxing objects.
1153 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1154 uint alias_idx = atp->index();
1155 Node* result = nullptr;
1156 Node* current = st;
1157 // Skip through chains of MemBarNodes checking the MergeMems for
1158 // new states for the slice of this load. Stop once any other
1159 // kind of node is encountered. Loads from final memory can skip
1160 // through any kind of MemBar but normal loads shouldn't skip
1161 // through MemBarAcquire since the could allow them to move out of
1162 // a synchronized region. It is not safe to step over MemBarCPUOrder,
1163 // because alias info above them may be inaccurate (e.g., due to
1164 // mixed/mismatched unsafe accesses).
1165 bool is_final_mem = !atp->is_rewritable();
1166 while (current->is_Proj()) {
1167 int opc = current->in(0)->Opcode();
1168 if ((is_final_mem && (opc == Op_MemBarAcquire ||
1169 opc == Op_MemBarAcquireLock ||
1170 opc == Op_LoadFence)) ||
1171 opc == Op_MemBarRelease ||
1172 opc == Op_StoreFence ||
1173 opc == Op_MemBarReleaseLock ||
1174 opc == Op_MemBarStoreStore ||
1175 opc == Op_StoreStoreFence) {
1176 Node* mem = current->in(0)->in(TypeFunc::Memory);
1177 if (mem->is_MergeMem()) {
1178 MergeMemNode* merge = mem->as_MergeMem();
1179 Node* new_st = merge->memory_at(alias_idx);
1180 if (new_st == merge->base_memory()) {
1181 // Keep searching
1182 current = new_st;
1183 continue;
1184 }
1185 // Save the new memory state for the slice and fall through
1186 // to exit.
1187 result = new_st;
1188 }
1189 }
1190 break;
1191 }
1192 if (result != nullptr) {
1193 st = result;
1194 }
1195 }
1196
1197 // Loop around twice in the case Load -> Initialize -> Store.
1198 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1199 for (int trip = 0; trip <= 1; trip++) {
1200
1201 if (st->is_Store()) {
1202 Node* st_adr = st->in(MemNode::Address);
1203 if (st_adr != ld_adr) {
1204 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1205 intptr_t st_off = 0;
1206 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1207 if (ld_base == nullptr) return nullptr;
1208 if (st_base == nullptr) return nullptr;
1209 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr;
1210 if (ld_off != st_off) return nullptr;
1211 if (ld_off == Type::OffsetBot) return nullptr;
1212 // Same base, same offset.
1213 // Possible improvement for arrays: check index value instead of absolute offset.
1214
1215 // At this point we have proven something like this setup:
1216 // B = << base >>
1217 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1218 // S = StoreQ(AddP( B , #Off), V)
1219 // (Actually, we haven't yet proven the Q's are the same.)
1220 // In other words, we are loading from a casted version of
1221 // the same pointer-and-offset that we stored to.
1222 // Casted version may carry a dependency and it is respected.
1223 // Thus, we are able to replace L by V.
1224 }
1225 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1226 if (store_Opcode() != st->Opcode()) {
1227 return nullptr;
1228 }
1229 // LoadVector/StoreVector needs additional check to ensure the types match.
1230 if (st->is_StoreVector()) {
1231 if ((Opcode() != Op_LoadVector && Opcode() != Op_StoreVector) || st->Opcode() != Op_StoreVector) {
1232 // Some kind of masked access or gather/scatter
1233 return nullptr;
1234 }
1235
1236 const TypeVect* in_vt = st->as_StoreVector()->vect_type();
1237 const TypeVect* out_vt = is_Load() ? as_LoadVector()->vect_type() : as_StoreVector()->vect_type();
1238 if (in_vt != out_vt) {
1239 return nullptr;
1240 }
1241 }
1242 return st->in(MemNode::ValueIn);
1243 }
1244
1245 // A load from a freshly-created object always returns zero.
1246 // (This can happen after LoadNode::Ideal resets the load's memory input
1247 // to find_captured_store, which returned InitializeNode::zero_memory.)
1248 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1249 (st->in(0) == ld_alloc) &&
1250 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1251 // return a zero value for the load's basic type
1252 // (This is one of the few places where a generic PhaseTransform
1253 // can create new nodes. Think of it as lazily manifesting
1254 // virtually pre-existing constants.)
1255 if (value_basic_type() != T_VOID) {
1256 if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) {
1257 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1258 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1259 // by the ArrayCopyNode.
1260 return phase->zerocon(value_basic_type());
1261 }
1262 } else {
1263 // TODO: materialize all-zero vector constant
1264 assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1265 }
1266 }
1267
1268 // A load from an initialization barrier can match a captured store.
1269 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1270 InitializeNode* init = st->in(0)->as_Initialize();
1271 AllocateNode* alloc = init->allocation();
1272 if ((alloc != nullptr) && (alloc == ld_alloc)) {
1273 // examine a captured store value
1274 st = init->find_captured_store(ld_off, memory_size(), phase);
1275 if (st != nullptr) {
1276 continue; // take one more trip around
1277 }
1278 }
1279 }
1280
1281 // Load boxed value from result of valueOf() call is input parameter.
1282 if (this->is_Load() && ld_adr->is_AddP() &&
1283 (tp != nullptr) && tp->is_ptr_to_boxed_value()) {
1284 intptr_t ignore = 0;
1285 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1286 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1287 base = bs->step_over_gc_barrier(base);
1288 if (base != nullptr && base->is_Proj() &&
1289 base->as_Proj()->_con == TypeFunc::Parms &&
1290 base->in(0)->is_CallStaticJava() &&
1291 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1292 return base->in(0)->in(TypeFunc::Parms);
1293 }
1294 }
1295
1296 break;
1297 }
1298
1299 return nullptr;
1300 }
1301
1302 //----------------------is_instance_field_load_with_local_phi------------------
1303 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1304 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1305 in(Address)->is_AddP() ) {
1306 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1307 // Only instances and boxed values.
1308 if( t_oop != nullptr &&
1309 (t_oop->is_ptr_to_boxed_value() ||
1310 t_oop->is_known_instance_field()) &&
1311 t_oop->offset() != Type::OffsetBot &&
1312 t_oop->offset() != Type::OffsetTop) {
1313 return true;
1314 }
1315 }
1316 return false;
1317 }
1318
1319 //------------------------------Identity---------------------------------------
1320 // Loads are identity if previous store is to same address
1321 Node* LoadNode::Identity(PhaseGVN* phase) {
1322 // If the previous store-maker is the right kind of Store, and the store is
1323 // to the same address, then we are equal to the value stored.
1324 Node* mem = in(Memory);
1325 Node* value = can_see_stored_value(mem, phase);
1326 if( value ) {
1327 // byte, short & char stores truncate naturally.
1328 // A load has to load the truncated value which requires
1329 // some sort of masking operation and that requires an
1330 // Ideal call instead of an Identity call.
1331 if (memory_size() < BytesPerInt) {
1332 // If the input to the store does not fit with the load's result type,
1333 // it must be truncated via an Ideal call.
1334 if (!phase->type(value)->higher_equal(phase->type(this)))
1335 return this;
1336 }
1337 // (This works even when value is a Con, but LoadNode::Value
1338 // usually runs first, producing the singleton type of the Con.)
1339 if (!has_pinned_control_dependency() || value->is_Con()) {
1340 return value;
1341 } else {
1342 return this;
1343 }
1344 }
1345
1346 if (has_pinned_control_dependency()) {
1347 return this;
1348 }
1349 // Search for an existing data phi which was generated before for the same
1350 // instance's field to avoid infinite generation of phis in a loop.
1351 Node *region = mem->in(0);
1352 if (is_instance_field_load_with_local_phi(region)) {
1353 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1354 int this_index = phase->C->get_alias_index(addr_t);
1355 int this_offset = addr_t->offset();
1356 int this_iid = addr_t->instance_id();
1357 if (!addr_t->is_known_instance() &&
1358 addr_t->is_ptr_to_boxed_value()) {
1359 // Use _idx of address base (could be Phi node) for boxed values.
1360 intptr_t ignore = 0;
1361 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1362 if (base == nullptr) {
1363 return this;
1364 }
1365 this_iid = base->_idx;
1366 }
1367 const Type* this_type = bottom_type();
1368 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1369 Node* phi = region->fast_out(i);
1370 if (phi->is_Phi() && phi != mem &&
1371 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1372 return phi;
1373 }
1374 }
1375 }
1376
1377 return this;
1378 }
1379
1380 // Construct an equivalent unsigned load.
1381 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1382 BasicType bt = T_ILLEGAL;
1383 const Type* rt = nullptr;
1384 switch (Opcode()) {
1385 case Op_LoadUB: return this;
1386 case Op_LoadUS: return this;
1387 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1388 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1389 default:
1390 assert(false, "no unsigned variant: %s", Name());
1391 return nullptr;
1392 }
1393 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1394 raw_adr_type(), rt, bt, _mo, _control_dependency,
1395 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1396 }
1397
1398 // Construct an equivalent signed load.
1399 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1400 BasicType bt = T_ILLEGAL;
1401 const Type* rt = nullptr;
1402 switch (Opcode()) {
1403 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1404 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1405 case Op_LoadB: // fall through
1406 case Op_LoadS: // fall through
1407 case Op_LoadI: // fall through
1408 case Op_LoadL: return this;
1409 default:
1410 assert(false, "no signed variant: %s", Name());
1411 return nullptr;
1412 }
1413 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1414 raw_adr_type(), rt, bt, _mo, _control_dependency,
1415 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1416 }
1417
1418 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1419 BasicType bt = rt->basic_type();
1420 switch (Opcode()) {
1421 case Op_LoadI: return (bt == T_FLOAT);
1422 case Op_LoadL: return (bt == T_DOUBLE);
1423 case Op_LoadF: return (bt == T_INT);
1424 case Op_LoadD: return (bt == T_LONG);
1425
1426 default: return false;
1427 }
1428 }
1429
1430 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1431 BasicType bt = rt->basic_type();
1432 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1433 bool is_mismatched = is_mismatched_access();
1434 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1435 if (raw_type == nullptr) {
1436 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1437 }
1438 const int op = Opcode();
1439 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) ||
1440 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access());
1441 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1442 raw_adr_type(), rt, bt, _mo, _control_dependency,
1443 require_atomic_access, is_unaligned_access(), is_mismatched);
1444 }
1445
1446 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1447 BasicType bt = vt->basic_type();
1448 switch (Opcode()) {
1449 case Op_StoreI: return (bt == T_FLOAT);
1450 case Op_StoreL: return (bt == T_DOUBLE);
1451 case Op_StoreF: return (bt == T_INT);
1452 case Op_StoreD: return (bt == T_LONG);
1453
1454 default: return false;
1455 }
1456 }
1457
1458 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1459 BasicType bt = vt->basic_type();
1460 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1461 const int op = Opcode();
1462 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) ||
1463 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access());
1464 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1465 raw_adr_type(), val, bt, _mo, require_atomic_access);
1466
1467 bool is_mismatched = is_mismatched_access();
1468 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1469 if (raw_type == nullptr) {
1470 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1471 }
1472 if (is_mismatched) {
1473 st->set_mismatched_access();
1474 }
1475 return st;
1476 }
1477
1478 // We're loading from an object which has autobox behaviour.
1479 // If this object is result of a valueOf call we'll have a phi
1480 // merging a newly allocated object and a load from the cache.
1481 // We want to replace this load with the original incoming
1482 // argument to the valueOf call.
1483 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1484 assert(igvn->C->eliminate_boxing(), "sanity");
1485 intptr_t ignore = 0;
1486 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1487 if ((base == nullptr) || base->is_Phi()) {
1488 // Push the loads from the phi that comes from valueOf up
1489 // through it to allow elimination of the loads and the recovery
1490 // of the original value. It is done in split_through_phi().
1491 return nullptr;
1492 } else if (base->is_Load() ||
1493 (base->is_DecodeN() && base->in(1)->is_Load())) {
1494 // Eliminate the load of boxed value for integer types from the cache
1495 // array by deriving the value from the index into the array.
1496 // Capture the offset of the load and then reverse the computation.
1497
1498 // Get LoadN node which loads a boxing object from 'cache' array.
1499 if (base->is_DecodeN()) {
1500 base = base->in(1);
1501 }
1502 if (!base->in(Address)->is_AddP()) {
1503 return nullptr; // Complex address
1504 }
1505 AddPNode* address = base->in(Address)->as_AddP();
1506 Node* cache_base = address->in(AddPNode::Base);
1507 if ((cache_base != nullptr) && cache_base->is_DecodeN()) {
1508 // Get ConP node which is static 'cache' field.
1509 cache_base = cache_base->in(1);
1510 }
1511 if ((cache_base != nullptr) && cache_base->is_Con()) {
1512 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1513 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
1514 Node* elements[4];
1515 int shift = exact_log2(type2aelembytes(T_OBJECT));
1516 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1517 if (count > 0 && elements[0]->is_Con() &&
1518 (count == 1 ||
1519 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1520 elements[1]->in(2) == igvn->intcon(shift)))) {
1521 ciObjArray* array = base_type->const_oop()->as_obj_array();
1522 // Fetch the box object cache[0] at the base of the array and get its value
1523 ciInstance* box = array->obj_at(0)->as_instance();
1524 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1525 assert(ik->is_box_klass(), "sanity");
1526 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1527 if (ik->nof_nonstatic_fields() == 1) {
1528 // This should be true nonstatic_field_at requires calling
1529 // nof_nonstatic_fields so check it anyway
1530 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1531 BasicType bt = c.basic_type();
1532 // Only integer types have boxing cache.
1533 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1534 bt == T_BYTE || bt == T_SHORT ||
1535 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1536 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1537 if (cache_low != (int)cache_low) {
1538 return nullptr; // should not happen since cache is array indexed by value
1539 }
1540 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1541 if (offset != (int)offset) {
1542 return nullptr; // should not happen since cache is array indexed by value
1543 }
1544 // Add up all the offsets making of the address of the load
1545 Node* result = elements[0];
1546 for (int i = 1; i < count; i++) {
1547 result = igvn->transform(new AddXNode(result, elements[i]));
1548 }
1549 // Remove the constant offset from the address and then
1550 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1551 // remove the scaling of the offset to recover the original index.
1552 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1553 // Peel the shift off directly but wrap it in a dummy node
1554 // since Ideal can't return existing nodes
1555 igvn->_worklist.push(result); // remove dead node later
1556 result = new RShiftXNode(result->in(1), igvn->intcon(0));
1557 } else if (result->is_Add() && result->in(2)->is_Con() &&
1558 result->in(1)->Opcode() == Op_LShiftX &&
1559 result->in(1)->in(2) == igvn->intcon(shift)) {
1560 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1561 // but for boxing cache access we know that X<<Z will not overflow
1562 // (there is range check) so we do this optimizatrion by hand here.
1563 igvn->_worklist.push(result); // remove dead node later
1564 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1565 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1566 } else {
1567 result = new RShiftXNode(result, igvn->intcon(shift));
1568 }
1569 #ifdef _LP64
1570 if (bt != T_LONG) {
1571 result = new ConvL2INode(igvn->transform(result));
1572 }
1573 #else
1574 if (bt == T_LONG) {
1575 result = new ConvI2LNode(igvn->transform(result));
1576 }
1577 #endif
1578 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1579 // Need to preserve unboxing load type if it is unsigned.
1580 switch(this->Opcode()) {
1581 case Op_LoadUB:
1582 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1583 break;
1584 case Op_LoadUS:
1585 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1586 break;
1587 }
1588 return result;
1589 }
1590 }
1591 }
1592 }
1593 }
1594 return nullptr;
1595 }
1596
1597 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1598 Node* region = phi->in(0);
1599 if (region == nullptr) {
1600 return false; // Wait stable graph
1601 }
1602 uint cnt = phi->req();
1603 for (uint i = 1; i < cnt; i++) {
1604 Node* rc = region->in(i);
1605 if (rc == nullptr || phase->type(rc) == Type::TOP)
1606 return false; // Wait stable graph
1607 Node* in = phi->in(i);
1608 if (in == nullptr || phase->type(in) == Type::TOP)
1609 return false; // Wait stable graph
1610 }
1611 return true;
1612 }
1613
1614 //------------------------------split_through_phi------------------------------
1615 // Check whether a call to 'split_through_phi' would split this load through the
1616 // Phi *base*. This method is essentially a copy of the validations performed
1617 // by 'split_through_phi'. The first use of this method was in EA code as part
1618 // of simplification of allocation merges.
1619 // Some differences from original method (split_through_phi):
1620 // - If base->is_CastPP(): base = base->in(1)
1621 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) {
1622 Node* mem = in(Memory);
1623 Node* address = in(Address);
1624 intptr_t ignore = 0;
1625 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1626
1627 if (base->is_CastPP()) {
1628 base = base->in(1);
1629 }
1630
1631 if (req() > 3 || base == nullptr || !base->is_Phi()) {
1632 return false;
1633 }
1634
1635 if (!mem->is_Phi()) {
1636 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1637 return false;
1638 }
1639 } else if (base->in(0) != mem->in(0)) {
1640 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1641 return false;
1642 }
1643 }
1644
1645 return true;
1646 }
1647
1648 //------------------------------split_through_phi------------------------------
1649 // Split instance or boxed field load through Phi.
1650 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) {
1651 if (req() > 3) {
1652 assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs");
1653 // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account
1654 return nullptr;
1655 }
1656 Node* mem = in(Memory);
1657 Node* address = in(Address);
1658 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1659
1660 assert((t_oop != nullptr) &&
1661 (ignore_missing_instance_id ||
1662 t_oop->is_known_instance_field() ||
1663 t_oop->is_ptr_to_boxed_value()), "invalid conditions");
1664
1665 Compile* C = phase->C;
1666 intptr_t ignore = 0;
1667 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1668 bool base_is_phi = (base != nullptr) && base->is_Phi();
1669 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1670 (base != nullptr) && (base == address->in(AddPNode::Base)) &&
1671 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1672
1673 if (!((mem->is_Phi() || base_is_phi) &&
1674 (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) {
1675 return nullptr; // Neither memory or base are Phi
1676 }
1677
1678 if (mem->is_Phi()) {
1679 if (!stable_phi(mem->as_Phi(), phase)) {
1680 return nullptr; // Wait stable graph
1681 }
1682 uint cnt = mem->req();
1683 // Check for loop invariant memory.
1684 if (cnt == 3) {
1685 for (uint i = 1; i < cnt; i++) {
1686 Node* in = mem->in(i);
1687 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1688 if (m == mem) {
1689 if (i == 1) {
1690 // if the first edge was a loop, check second edge too.
1691 // If both are replaceable - we are in an infinite loop
1692 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1693 if (n == mem) {
1694 break;
1695 }
1696 }
1697 set_req(Memory, mem->in(cnt - i));
1698 return this; // made change
1699 }
1700 }
1701 }
1702 }
1703 if (base_is_phi) {
1704 if (!stable_phi(base->as_Phi(), phase)) {
1705 return nullptr; // Wait stable graph
1706 }
1707 uint cnt = base->req();
1708 // Check for loop invariant memory.
1709 if (cnt == 3) {
1710 for (uint i = 1; i < cnt; i++) {
1711 if (base->in(i) == base) {
1712 return nullptr; // Wait stable graph
1713 }
1714 }
1715 }
1716 }
1717
1718 // Split through Phi (see original code in loopopts.cpp).
1719 assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type");
1720
1721 // Do nothing here if Identity will find a value
1722 // (to avoid infinite chain of value phis generation).
1723 if (this != Identity(phase)) {
1724 return nullptr;
1725 }
1726
1727 // Select Region to split through.
1728 Node* region;
1729 DomResult dom_result = DomResult::Dominate;
1730 if (!base_is_phi) {
1731 assert(mem->is_Phi(), "sanity");
1732 region = mem->in(0);
1733 // Skip if the region dominates some control edge of the address.
1734 // We will check `dom_result` later.
1735 dom_result = MemNode::maybe_all_controls_dominate(address, region);
1736 } else if (!mem->is_Phi()) {
1737 assert(base_is_phi, "sanity");
1738 region = base->in(0);
1739 // Skip if the region dominates some control edge of the memory.
1740 // We will check `dom_result` later.
1741 dom_result = MemNode::maybe_all_controls_dominate(mem, region);
1742 } else if (base->in(0) != mem->in(0)) {
1743 assert(base_is_phi && mem->is_Phi(), "sanity");
1744 dom_result = MemNode::maybe_all_controls_dominate(mem, base->in(0));
1745 if (dom_result == DomResult::Dominate) {
1746 region = base->in(0);
1747 } else {
1748 dom_result = MemNode::maybe_all_controls_dominate(address, mem->in(0));
1749 if (dom_result == DomResult::Dominate) {
1750 region = mem->in(0);
1751 }
1752 // Otherwise we encountered a complex graph.
1753 }
1754 } else {
1755 assert(base->in(0) == mem->in(0), "sanity");
1756 region = mem->in(0);
1757 }
1758
1759 PhaseIterGVN* igvn = phase->is_IterGVN();
1760 if (dom_result != DomResult::Dominate) {
1761 if (dom_result == DomResult::EncounteredDeadCode) {
1762 // There is some dead code which eventually will be removed in IGVN.
1763 // Once this is the case, we get an unambiguous dominance result.
1764 // Push the node to the worklist again until the dead code is removed.
1765 igvn->_worklist.push(this);
1766 }
1767 return nullptr;
1768 }
1769
1770 Node* phi = nullptr;
1771 const Type* this_type = this->bottom_type();
1772 if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) {
1773 int this_index = C->get_alias_index(t_oop);
1774 int this_offset = t_oop->offset();
1775 int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx;
1776 phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset);
1777 } else if (ignore_missing_instance_id) {
1778 phi = new PhiNode(region, this_type, nullptr, mem->_idx);
1779 } else {
1780 return nullptr;
1781 }
1782
1783 for (uint i = 1; i < region->req(); i++) {
1784 Node* x;
1785 Node* the_clone = nullptr;
1786 Node* in = region->in(i);
1787 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1788 in != nullptr && in->is_OuterStripMinedLoop()) {
1789 // No node should go in the outer strip mined loop
1790 in = in->in(LoopNode::EntryControl);
1791 }
1792 if (in == nullptr || in == C->top()) {
1793 x = C->top(); // Dead path? Use a dead data op
1794 } else {
1795 x = this->clone(); // Else clone up the data op
1796 the_clone = x; // Remember for possible deletion.
1797 // Alter data node to use pre-phi inputs
1798 if (this->in(0) == region) {
1799 x->set_req(0, in);
1800 } else {
1801 x->set_req(0, nullptr);
1802 }
1803 if (mem->is_Phi() && (mem->in(0) == region)) {
1804 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1805 }
1806 if (address->is_Phi() && address->in(0) == region) {
1807 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1808 }
1809 if (base_is_phi && (base->in(0) == region)) {
1810 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1811 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1812 x->set_req(Address, adr_x);
1813 }
1814 }
1815 // Check for a 'win' on some paths
1816 const Type *t = x->Value(igvn);
1817
1818 bool singleton = t->singleton();
1819
1820 // See comments in PhaseIdealLoop::split_thru_phi().
1821 if (singleton && t == Type::TOP) {
1822 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1823 }
1824
1825 if (singleton) {
1826 x = igvn->makecon(t);
1827 } else {
1828 // We now call Identity to try to simplify the cloned node.
1829 // Note that some Identity methods call phase->type(this).
1830 // Make sure that the type array is big enough for
1831 // our new node, even though we may throw the node away.
1832 // (This tweaking with igvn only works because x is a new node.)
1833 igvn->set_type(x, t);
1834 // If x is a TypeNode, capture any more-precise type permanently into Node
1835 // otherwise it will be not updated during igvn->transform since
1836 // igvn->type(x) is set to x->Value() already.
1837 x->raise_bottom_type(t);
1838 Node* y = x->Identity(igvn);
1839 if (y != x) {
1840 x = y;
1841 } else {
1842 y = igvn->hash_find_insert(x);
1843 if (y) {
1844 x = y;
1845 } else {
1846 // Else x is a new node we are keeping
1847 // We do not need register_new_node_with_optimizer
1848 // because set_type has already been called.
1849 igvn->_worklist.push(x);
1850 }
1851 }
1852 }
1853 if (x != the_clone && the_clone != nullptr) {
1854 igvn->remove_dead_node(the_clone);
1855 }
1856 phi->set_req(i, x);
1857 }
1858 // Record Phi
1859 igvn->register_new_node_with_optimizer(phi);
1860 return phi;
1861 }
1862
1863 AllocateNode* LoadNode::is_new_object_mark_load() const {
1864 if (Opcode() == Op_LoadX) {
1865 Node* address = in(MemNode::Address);
1866 AllocateNode* alloc = AllocateNode::Ideal_allocation(address);
1867 Node* mem = in(MemNode::Memory);
1868 if (alloc != nullptr && mem->is_Proj() &&
1869 mem->in(0) != nullptr &&
1870 mem->in(0) == alloc->initialization() &&
1871 alloc->initialization()->proj_out_or_null(0) != nullptr) {
1872 return alloc;
1873 }
1874 }
1875 return nullptr;
1876 }
1877
1878
1879 //------------------------------Ideal------------------------------------------
1880 // If the load is from Field memory and the pointer is non-null, it might be possible to
1881 // zero out the control input.
1882 // If the offset is constant and the base is an object allocation,
1883 // try to hook me up to the exact initializing store.
1884 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1885 if (has_pinned_control_dependency()) {
1886 return nullptr;
1887 }
1888 Node* p = MemNode::Ideal_common(phase, can_reshape);
1889 if (p) return (p == NodeSentinel) ? nullptr : p;
1890
1891 Node* ctrl = in(MemNode::Control);
1892 Node* address = in(MemNode::Address);
1893
1894 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1895 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1896
1897 // Skip up past a SafePoint control. Cannot do this for Stores because
1898 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1899 if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint &&
1900 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1901 !addr_mark &&
1902 (depends_only_on_test() || has_unknown_control_dependency())) {
1903 ctrl = ctrl->in(0);
1904 set_req(MemNode::Control,ctrl);
1905 return this;
1906 }
1907
1908 intptr_t ignore = 0;
1909 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1910 if (base != nullptr
1911 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1912 // Check for useless control edge in some common special cases
1913 if (in(MemNode::Control) != nullptr
1914 && can_remove_control()
1915 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1916 && all_controls_dominate(base, phase->C->start())) {
1917 // A method-invariant, non-null address (constant or 'this' argument).
1918 set_req(MemNode::Control, nullptr);
1919 return this;
1920 }
1921 }
1922
1923 Node* mem = in(MemNode::Memory);
1924 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1925
1926 if (can_reshape && (addr_t != nullptr)) {
1927 // try to optimize our memory input
1928 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1929 if (opt_mem != mem) {
1930 set_req_X(MemNode::Memory, opt_mem, phase);
1931 if (phase->type( opt_mem ) == Type::TOP) return nullptr;
1932 return this;
1933 }
1934 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1935 if ((t_oop != nullptr) &&
1936 (t_oop->is_known_instance_field() ||
1937 t_oop->is_ptr_to_boxed_value())) {
1938 PhaseIterGVN *igvn = phase->is_IterGVN();
1939 assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true");
1940 if (igvn->_worklist.member(opt_mem)) {
1941 // Delay this transformation until memory Phi is processed.
1942 igvn->_worklist.push(this);
1943 return nullptr;
1944 }
1945 // Split instance field load through Phi.
1946 Node* result = split_through_phi(phase);
1947 if (result != nullptr) return result;
1948
1949 if (t_oop->is_ptr_to_boxed_value()) {
1950 Node* result = eliminate_autobox(igvn);
1951 if (result != nullptr) return result;
1952 }
1953 }
1954 }
1955
1956 // Is there a dominating load that loads the same value? Leave
1957 // anything that is not a load of a field/array element (like
1958 // barriers etc.) alone
1959 if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) {
1960 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1961 Node *use = mem->fast_out(i);
1962 if (use != this &&
1963 use->Opcode() == Opcode() &&
1964 use->in(0) != nullptr &&
1965 use->in(0) != in(0) &&
1966 use->in(Address) == in(Address)) {
1967 Node* ctl = in(0);
1968 for (int i = 0; i < 10 && ctl != nullptr; i++) {
1969 ctl = IfNode::up_one_dom(ctl);
1970 if (ctl == use->in(0)) {
1971 set_req(0, use->in(0));
1972 return this;
1973 }
1974 }
1975 }
1976 }
1977 }
1978
1979 // Check for prior store with a different base or offset; make Load
1980 // independent. Skip through any number of them. Bail out if the stores
1981 // are in an endless dead cycle and report no progress. This is a key
1982 // transform for Reflection. However, if after skipping through the Stores
1983 // we can't then fold up against a prior store do NOT do the transform as
1984 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1985 // array memory alive twice: once for the hoisted Load and again after the
1986 // bypassed Store. This situation only works if EVERYBODY who does
1987 // anti-dependence work knows how to bypass. I.e. we need all
1988 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1989 // the alias index stuff. So instead, peek through Stores and IFF we can
1990 // fold up, do so.
1991 Node* prev_mem = find_previous_store(phase);
1992 if (prev_mem != nullptr && prev_mem->is_top()) {
1993 // find_previous_store returns top when the access is dead
1994 return prev_mem;
1995 }
1996 if (prev_mem != nullptr) {
1997 Node* value = can_see_arraycopy_value(prev_mem, phase);
1998 if (value != nullptr) {
1999 return value;
2000 }
2001 }
2002 // Steps (a), (b): Walk past independent stores to find an exact match.
2003 if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) {
2004 // (c) See if we can fold up on the spot, but don't fold up here.
2005 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
2006 // just return a prior value, which is done by Identity calls.
2007 if (can_see_stored_value(prev_mem, phase)) {
2008 // Make ready for step (d):
2009 set_req_X(MemNode::Memory, prev_mem, phase);
2010 return this;
2011 }
2012 }
2013
2014 if (!can_reshape) {
2015 phase->record_for_igvn(this);
2016 }
2017
2018 return nullptr;
2019 }
2020
2021 // Helper to recognize certain Klass fields which are invariant across
2022 // some group of array types (e.g., int[] or all T[] where T < Object).
2023 const Type*
2024 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
2025 ciKlass* klass) const {
2026 assert(!UseCompactObjectHeaders || tkls->offset() != in_bytes(Klass::prototype_header_offset()),
2027 "must not happen");
2028
2029 if (tkls->isa_instklassptr() && tkls->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
2030 // The field is InstanceKlass::_access_flags. Return its (constant) value.
2031 assert(Opcode() == Op_LoadUS, "must load an unsigned short from _access_flags");
2032 ciInstanceKlass* iklass = tkls->is_instklassptr()->instance_klass();
2033 return TypeInt::make(iklass->access_flags());
2034 }
2035 if (tkls->offset() == in_bytes(Klass::misc_flags_offset())) {
2036 // The field is Klass::_misc_flags. Return its (constant) value.
2037 assert(Opcode() == Op_LoadUB, "must load an unsigned byte from _misc_flags");
2038 return TypeInt::make(klass->misc_flags());
2039 }
2040 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
2041 // The field is Klass::_layout_helper. Return its constant value if known.
2042 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2043 return TypeInt::make(klass->layout_helper());
2044 }
2045
2046 // No match.
2047 return nullptr;
2048 }
2049
2050 //------------------------------Value-----------------------------------------
2051 const Type* LoadNode::Value(PhaseGVN* phase) const {
2052 // Either input is TOP ==> the result is TOP
2053 Node* mem = in(MemNode::Memory);
2054 const Type *t1 = phase->type(mem);
2055 if (t1 == Type::TOP) return Type::TOP;
2056 Node* adr = in(MemNode::Address);
2057 const TypePtr* tp = phase->type(adr)->isa_ptr();
2058 if (tp == nullptr || tp->empty()) return Type::TOP;
2059 int off = tp->offset();
2060 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
2061 Compile* C = phase->C;
2062
2063 // If load can see a previous constant store, use that.
2064 Node* value = can_see_stored_value(mem, phase);
2065 if (value != nullptr && value->is_Con()) {
2066 assert(value->bottom_type()->higher_equal(_type), "sanity");
2067 return value->bottom_type();
2068 }
2069
2070 // Try to guess loaded type from pointer type
2071 if (tp->isa_aryptr()) {
2072 const TypeAryPtr* ary = tp->is_aryptr();
2073 const Type* t = ary->elem();
2074
2075 // Determine whether the reference is beyond the header or not, by comparing
2076 // the offset against the offset of the start of the array's data.
2077 // Different array types begin at slightly different offsets (12 vs. 16).
2078 // We choose T_BYTE as an example base type that is least restrictive
2079 // as to alignment, which will therefore produce the smallest
2080 // possible base offset.
2081 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2082 const bool off_beyond_header = (off >= min_base_off);
2083
2084 // Try to constant-fold a stable array element.
2085 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
2086 // Make sure the reference is not into the header and the offset is constant
2087 ciObject* aobj = ary->const_oop();
2088 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
2089 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
2090 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
2091 stable_dimension,
2092 value_basic_type(), is_unsigned());
2093 if (con_type != nullptr) {
2094 return con_type;
2095 }
2096 }
2097 }
2098
2099 // Don't do this for integer types. There is only potential profit if
2100 // the element type t is lower than _type; that is, for int types, if _type is
2101 // more restrictive than t. This only happens here if one is short and the other
2102 // char (both 16 bits), and in those cases we've made an intentional decision
2103 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2104 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2105 //
2106 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2107 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
2108 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2109 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
2110 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2111 // In fact, that could have been the original type of p1, and p1 could have
2112 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2113 // expression (LShiftL quux 3) independently optimized to the constant 8.
2114 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2115 && (_type->isa_vect() == nullptr)
2116 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2117 // t might actually be lower than _type, if _type is a unique
2118 // concrete subclass of abstract class t.
2119 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
2120 const Type* jt = t->join_speculative(_type);
2121 // In any case, do not allow the join, per se, to empty out the type.
2122 if (jt->empty() && !t->empty()) {
2123 // This can happen if a interface-typed array narrows to a class type.
2124 jt = _type;
2125 }
2126 #ifdef ASSERT
2127 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2128 // The pointers in the autobox arrays are always non-null
2129 Node* base = adr->in(AddPNode::Base);
2130 if ((base != nullptr) && base->is_DecodeN()) {
2131 // Get LoadN node which loads IntegerCache.cache field
2132 base = base->in(1);
2133 }
2134 if ((base != nullptr) && base->is_Con()) {
2135 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2136 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2137 // It could be narrow oop
2138 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2139 }
2140 }
2141 }
2142 #endif
2143 return jt;
2144 }
2145 }
2146 } else if (tp->base() == Type::InstPtr) {
2147 assert( off != Type::OffsetBot ||
2148 // arrays can be cast to Objects
2149 !tp->isa_instptr() ||
2150 tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
2151 // unsafe field access may not have a constant offset
2152 C->has_unsafe_access(),
2153 "Field accesses must be precise" );
2154 // For oop loads, we expect the _type to be precise.
2155
2156 // Optimize loads from constant fields.
2157 const TypeInstPtr* tinst = tp->is_instptr();
2158 ciObject* const_oop = tinst->const_oop();
2159 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2160 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), value_basic_type());
2161 if (con_type != nullptr) {
2162 return con_type;
2163 }
2164 }
2165 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2166 assert(off != Type::OffsetBot ||
2167 !tp->isa_instklassptr() ||
2168 // arrays can be cast to Objects
2169 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2170 // also allow array-loading from the primary supertype
2171 // array during subtype checks
2172 Opcode() == Op_LoadKlass,
2173 "Field accesses must be precise");
2174 // For klass/static loads, we expect the _type to be precise
2175 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2176 /* With mirrors being an indirect in the Klass*
2177 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2178 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2179 *
2180 * So check the type and klass of the node before the LoadP.
2181 */
2182 Node* adr2 = adr->in(MemNode::Address);
2183 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2184 if (tkls != nullptr && !StressReflectiveCode) {
2185 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2186 ciKlass* klass = tkls->exact_klass();
2187 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2188 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2189 return TypeInstPtr::make(klass->java_mirror());
2190 }
2191 }
2192 }
2193
2194 const TypeKlassPtr *tkls = tp->isa_klassptr();
2195 if (tkls != nullptr) {
2196 if (tkls->is_loaded() && tkls->klass_is_exact()) {
2197 ciKlass* klass = tkls->exact_klass();
2198 // We are loading a field from a Klass metaobject whose identity
2199 // is known at compile time (the type is "exact" or "precise").
2200 // Check for fields we know are maintained as constants by the VM.
2201 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2202 // The field is Klass::_super_check_offset. Return its (constant) value.
2203 // (Folds up type checking code.)
2204 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2205 return TypeInt::make(klass->super_check_offset());
2206 }
2207 if (UseCompactObjectHeaders) {
2208 if (tkls->offset() == in_bytes(Klass::prototype_header_offset())) {
2209 // The field is Klass::_prototype_header. Return its (constant) value.
2210 assert(this->Opcode() == Op_LoadX, "must load a proper type from _prototype_header");
2211 return TypeX::make(klass->prototype_header());
2212 }
2213 }
2214 // Compute index into primary_supers array
2215 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2216 // Check for overflowing; use unsigned compare to handle the negative case.
2217 if( depth < ciKlass::primary_super_limit() ) {
2218 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2219 // (Folds up type checking code.)
2220 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2221 ciKlass *ss = klass->super_of_depth(depth);
2222 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2223 }
2224 const Type* aift = load_array_final_field(tkls, klass);
2225 if (aift != nullptr) return aift;
2226 }
2227
2228 // We can still check if we are loading from the primary_supers array at a
2229 // shallow enough depth. Even though the klass is not exact, entries less
2230 // than or equal to its super depth are correct.
2231 if (tkls->is_loaded()) {
2232 ciKlass* klass = nullptr;
2233 if (tkls->isa_instklassptr()) {
2234 klass = tkls->is_instklassptr()->instance_klass();
2235 } else {
2236 int dims;
2237 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims);
2238 if (inner->isa_instklassptr()) {
2239 klass = inner->is_instklassptr()->instance_klass();
2240 klass = ciObjArrayKlass::make(klass, dims);
2241 }
2242 }
2243 if (klass != nullptr) {
2244 // Compute index into primary_supers array
2245 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2246 // Check for overflowing; use unsigned compare to handle the negative case.
2247 if (depth < ciKlass::primary_super_limit() &&
2248 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case
2249 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2250 // (Folds up type checking code.)
2251 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2252 ciKlass *ss = klass->super_of_depth(depth);
2253 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2254 }
2255 }
2256 }
2257
2258 // If the type is enough to determine that the thing is not an array,
2259 // we can give the layout_helper a positive interval type.
2260 // This will help short-circuit some reflective code.
2261 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) &&
2262 tkls->isa_instklassptr() && // not directly typed as an array
2263 !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
2264 ) {
2265 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2266 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2267 // The key property of this type is that it folds up tests
2268 // for array-ness, since it proves that the layout_helper is positive.
2269 // Thus, a generic value like the basic object layout helper works fine.
2270 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2271 }
2272 }
2273
2274 // If we are loading from a freshly-allocated object/array, produce a zero.
2275 // Things to check:
2276 // 1. Load is beyond the header: headers are not guaranteed to be zero
2277 // 2. Load is not vectorized: vectors have no zero constant
2278 // 3. Load has no matching store, i.e. the input is the initial memory state
2279 const TypeOopPtr* tinst = tp->isa_oopptr();
2280 bool is_not_header = (tinst != nullptr) && tinst->is_known_instance_field();
2281 bool is_not_vect = (_type->isa_vect() == nullptr);
2282 if (is_not_header && is_not_vect) {
2283 Node* mem = in(MemNode::Memory);
2284 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2285 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2286 return Type::get_zero_type(_type->basic_type());
2287 }
2288 }
2289
2290 if (!UseCompactObjectHeaders) {
2291 Node* alloc = is_new_object_mark_load();
2292 if (alloc != nullptr) {
2293 return TypeX::make(markWord::prototype().value());
2294 }
2295 }
2296
2297 return _type;
2298 }
2299
2300 //------------------------------match_edge-------------------------------------
2301 // Do we Match on this edge index or not? Match only the address.
2302 uint LoadNode::match_edge(uint idx) const {
2303 return idx == MemNode::Address;
2304 }
2305
2306 //--------------------------LoadBNode::Ideal--------------------------------------
2307 //
2308 // If the previous store is to the same address as this load,
2309 // and the value stored was larger than a byte, replace this load
2310 // with the value stored truncated to a byte. If no truncation is
2311 // needed, the replacement is done in LoadNode::Identity().
2312 //
2313 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2314 Node* mem = in(MemNode::Memory);
2315 Node* value = can_see_stored_value(mem,phase);
2316 if (value != nullptr) {
2317 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2318 if (narrow != value) {
2319 return narrow;
2320 }
2321 }
2322 // Identity call will handle the case where truncation is not needed.
2323 return LoadNode::Ideal(phase, can_reshape);
2324 }
2325
2326 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2327 Node* mem = in(MemNode::Memory);
2328 Node* value = can_see_stored_value(mem,phase);
2329 if (value != nullptr && value->is_Con() &&
2330 !value->bottom_type()->higher_equal(_type)) {
2331 // If the input to the store does not fit with the load's result type,
2332 // it must be truncated. We can't delay until Ideal call since
2333 // a singleton Value is needed for split_thru_phi optimization.
2334 int con = value->get_int();
2335 return TypeInt::make((con << 24) >> 24);
2336 }
2337 return LoadNode::Value(phase);
2338 }
2339
2340 //--------------------------LoadUBNode::Ideal-------------------------------------
2341 //
2342 // If the previous store is to the same address as this load,
2343 // and the value stored was larger than a byte, replace this load
2344 // with the value stored truncated to a byte. If no truncation is
2345 // needed, the replacement is done in LoadNode::Identity().
2346 //
2347 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2348 Node* mem = in(MemNode::Memory);
2349 Node* value = can_see_stored_value(mem, phase);
2350 if (value != nullptr) {
2351 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2352 if (narrow != value) {
2353 return narrow;
2354 }
2355 }
2356 // Identity call will handle the case where truncation is not needed.
2357 return LoadNode::Ideal(phase, can_reshape);
2358 }
2359
2360 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2361 Node* mem = in(MemNode::Memory);
2362 Node* value = can_see_stored_value(mem,phase);
2363 if (value != nullptr && value->is_Con() &&
2364 !value->bottom_type()->higher_equal(_type)) {
2365 // If the input to the store does not fit with the load's result type,
2366 // it must be truncated. We can't delay until Ideal call since
2367 // a singleton Value is needed for split_thru_phi optimization.
2368 int con = value->get_int();
2369 return TypeInt::make(con & 0xFF);
2370 }
2371 return LoadNode::Value(phase);
2372 }
2373
2374 //--------------------------LoadUSNode::Ideal-------------------------------------
2375 //
2376 // If the previous store is to the same address as this load,
2377 // and the value stored was larger than a char, replace this load
2378 // with the value stored truncated to a char. If no truncation is
2379 // needed, the replacement is done in LoadNode::Identity().
2380 //
2381 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2382 Node* mem = in(MemNode::Memory);
2383 Node* value = can_see_stored_value(mem,phase);
2384 if (value != nullptr) {
2385 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2386 if (narrow != value) {
2387 return narrow;
2388 }
2389 }
2390 // Identity call will handle the case where truncation is not needed.
2391 return LoadNode::Ideal(phase, can_reshape);
2392 }
2393
2394 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2395 Node* mem = in(MemNode::Memory);
2396 Node* value = can_see_stored_value(mem,phase);
2397 if (value != nullptr && value->is_Con() &&
2398 !value->bottom_type()->higher_equal(_type)) {
2399 // If the input to the store does not fit with the load's result type,
2400 // it must be truncated. We can't delay until Ideal call since
2401 // a singleton Value is needed for split_thru_phi optimization.
2402 int con = value->get_int();
2403 return TypeInt::make(con & 0xFFFF);
2404 }
2405 return LoadNode::Value(phase);
2406 }
2407
2408 //--------------------------LoadSNode::Ideal--------------------------------------
2409 //
2410 // If the previous store is to the same address as this load,
2411 // and the value stored was larger than a short, replace this load
2412 // with the value stored truncated to a short. If no truncation is
2413 // needed, the replacement is done in LoadNode::Identity().
2414 //
2415 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2416 Node* mem = in(MemNode::Memory);
2417 Node* value = can_see_stored_value(mem,phase);
2418 if (value != nullptr) {
2419 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2420 if (narrow != value) {
2421 return narrow;
2422 }
2423 }
2424 // Identity call will handle the case where truncation is not needed.
2425 return LoadNode::Ideal(phase, can_reshape);
2426 }
2427
2428 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2429 Node* mem = in(MemNode::Memory);
2430 Node* value = can_see_stored_value(mem,phase);
2431 if (value != nullptr && value->is_Con() &&
2432 !value->bottom_type()->higher_equal(_type)) {
2433 // If the input to the store does not fit with the load's result type,
2434 // it must be truncated. We can't delay until Ideal call since
2435 // a singleton Value is needed for split_thru_phi optimization.
2436 int con = value->get_int();
2437 return TypeInt::make((con << 16) >> 16);
2438 }
2439 return LoadNode::Value(phase);
2440 }
2441
2442 //=============================================================================
2443 //----------------------------LoadKlassNode::make------------------------------
2444 // Polymorphic factory method:
2445 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2446 // sanity check the alias category against the created node type
2447 const TypePtr* adr_type = adr->bottom_type()->isa_ptr();
2448 assert(adr_type != nullptr, "expecting TypeKlassPtr");
2449 #ifdef _LP64
2450 if (adr_type->is_ptr_to_narrowklass()) {
2451 assert(UseCompressedClassPointers, "no compressed klasses");
2452 Node* load_klass = gvn.transform(new LoadNKlassNode(mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2453 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2454 }
2455 #endif
2456 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2457 return new LoadKlassNode(mem, adr, at, tk, MemNode::unordered);
2458 }
2459
2460 //------------------------------Value------------------------------------------
2461 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2462 return klass_value_common(phase);
2463 }
2464
2465 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2466 // Either input is TOP ==> the result is TOP
2467 const Type *t1 = phase->type( in(MemNode::Memory) );
2468 if (t1 == Type::TOP) return Type::TOP;
2469 Node *adr = in(MemNode::Address);
2470 const Type *t2 = phase->type( adr );
2471 if (t2 == Type::TOP) return Type::TOP;
2472 const TypePtr *tp = t2->is_ptr();
2473 if (TypePtr::above_centerline(tp->ptr()) ||
2474 tp->ptr() == TypePtr::Null) return Type::TOP;
2475
2476 // Return a more precise klass, if possible
2477 const TypeInstPtr *tinst = tp->isa_instptr();
2478 if (tinst != nullptr) {
2479 ciInstanceKlass* ik = tinst->instance_klass();
2480 int offset = tinst->offset();
2481 if (ik == phase->C->env()->Class_klass()
2482 && (offset == java_lang_Class::klass_offset() ||
2483 offset == java_lang_Class::array_klass_offset())) {
2484 // We are loading a special hidden field from a Class mirror object,
2485 // the field which points to the VM's Klass metaobject.
2486 ciType* t = tinst->java_mirror_type();
2487 // java_mirror_type returns non-null for compile-time Class constants.
2488 if (t != nullptr) {
2489 // constant oop => constant klass
2490 if (offset == java_lang_Class::array_klass_offset()) {
2491 if (t->is_void()) {
2492 // We cannot create a void array. Since void is a primitive type return null
2493 // klass. Users of this result need to do a null check on the returned klass.
2494 return TypePtr::NULL_PTR;
2495 }
2496 return TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);
2497 }
2498 if (!t->is_klass()) {
2499 // a primitive Class (e.g., int.class) has null for a klass field
2500 return TypePtr::NULL_PTR;
2501 }
2502 // Fold up the load of the hidden field
2503 return TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);
2504 }
2505 // non-constant mirror, so we can't tell what's going on
2506 }
2507 if (!tinst->is_loaded())
2508 return _type; // Bail out if not loaded
2509 if (offset == oopDesc::klass_offset_in_bytes()) {
2510 return tinst->as_klass_type(true);
2511 }
2512 }
2513
2514 // Check for loading klass from an array
2515 const TypeAryPtr *tary = tp->isa_aryptr();
2516 if (tary != nullptr &&
2517 tary->offset() == oopDesc::klass_offset_in_bytes()) {
2518 return tary->as_klass_type(true);
2519 }
2520
2521 // Check for loading klass from an array klass
2522 const TypeKlassPtr *tkls = tp->isa_klassptr();
2523 if (tkls != nullptr && !StressReflectiveCode) {
2524 if (!tkls->is_loaded())
2525 return _type; // Bail out if not loaded
2526 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2527 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2528 // // Always returning precise element type is incorrect,
2529 // // e.g., element type could be object and array may contain strings
2530 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2531
2532 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2533 // according to the element type's subclassing.
2534 return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2535 }
2536 if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() &&
2537 tkls->offset() == in_bytes(Klass::super_offset())) {
2538 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super();
2539 // The field is Klass::_super. Return its (constant) value.
2540 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2541 return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR;
2542 }
2543 }
2544
2545 if (tkls != nullptr && !UseSecondarySupersCache
2546 && tkls->offset() == in_bytes(Klass::secondary_super_cache_offset())) {
2547 // Treat Klass::_secondary_super_cache as a constant when the cache is disabled.
2548 return TypePtr::NULL_PTR;
2549 }
2550
2551 // Bailout case
2552 return LoadNode::Value(phase);
2553 }
2554
2555 //------------------------------Identity---------------------------------------
2556 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2557 // Also feed through the klass in Allocate(...klass...)._klass.
2558 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2559 return klass_identity_common(phase);
2560 }
2561
2562 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2563 Node* x = LoadNode::Identity(phase);
2564 if (x != this) return x;
2565
2566 // Take apart the address into an oop and offset.
2567 // Return 'this' if we cannot.
2568 Node* adr = in(MemNode::Address);
2569 intptr_t offset = 0;
2570 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2571 if (base == nullptr) return this;
2572 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2573 if (toop == nullptr) return this;
2574
2575 // Step over potential GC barrier for OopHandle resolve
2576 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2577 if (bs->is_gc_barrier_node(base)) {
2578 base = bs->step_over_gc_barrier(base);
2579 }
2580
2581 // We can fetch the klass directly through an AllocateNode.
2582 // This works even if the klass is not constant (clone or newArray).
2583 if (offset == oopDesc::klass_offset_in_bytes()) {
2584 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2585 if (allocated_klass != nullptr) {
2586 return allocated_klass;
2587 }
2588 }
2589
2590 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2591 // See inline_native_Class_query for occurrences of these patterns.
2592 // Java Example: x.getClass().isAssignableFrom(y)
2593 //
2594 // This improves reflective code, often making the Class
2595 // mirror go completely dead. (Current exception: Class
2596 // mirrors may appear in debug info, but we could clean them out by
2597 // introducing a new debug info operator for Klass.java_mirror).
2598
2599 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass()
2600 && offset == java_lang_Class::klass_offset()) {
2601 if (base->is_Load()) {
2602 Node* base2 = base->in(MemNode::Address);
2603 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2604 Node* adr2 = base2->in(MemNode::Address);
2605 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2606 if (tkls != nullptr && !tkls->empty()
2607 && (tkls->isa_instklassptr() || tkls->isa_aryklassptr())
2608 && adr2->is_AddP()
2609 ) {
2610 int mirror_field = in_bytes(Klass::java_mirror_offset());
2611 if (tkls->offset() == mirror_field) {
2612 #ifdef ASSERT
2613 const TypeKlassPtr* tkls2 = phase->type(adr2->in(AddPNode::Address))->is_klassptr();
2614 assert(tkls2->offset() == 0, "not a load of java_mirror");
2615 #endif
2616 assert(adr2->in(AddPNode::Base)->is_top(), "not an off heap load");
2617 assert(adr2->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(Klass::java_mirror_offset()), "incorrect offset");
2618 return adr2->in(AddPNode::Address);
2619 }
2620 }
2621 }
2622 }
2623 }
2624
2625 return this;
2626 }
2627
2628 LoadNode* LoadNode::clone_pinned() const {
2629 LoadNode* ld = clone()->as_Load();
2630 ld->_control_dependency = UnknownControl;
2631 return ld;
2632 }
2633
2634 // Pin a LoadNode if it carries a dependency on its control input. There are cases when the node
2635 // does not actually have any dependency on its control input. For example, if we have a LoadNode
2636 // being used only outside a loop but it must be scheduled inside the loop, we can clone the node
2637 // for each of its use so that all the clones can be scheduled outside the loop. Then, to prevent
2638 // the clones from being GVN-ed again, we add a control input for each of them at the loop exit. In
2639 // those case, since there is not a dependency between the node and its control input, we do not
2640 // need to pin it.
2641 LoadNode* LoadNode::pin_node_under_control_impl() const {
2642 const TypePtr* adr_type = this->adr_type();
2643 if (adr_type != nullptr && adr_type->isa_aryptr()) {
2644 // Only array accesses have dependencies on their control input
2645 return clone_pinned();
2646 }
2647 return nullptr;
2648 }
2649
2650 //------------------------------Value------------------------------------------
2651 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2652 const Type *t = klass_value_common(phase);
2653 if (t == Type::TOP)
2654 return t;
2655
2656 return t->make_narrowklass();
2657 }
2658
2659 //------------------------------Identity---------------------------------------
2660 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2661 // Also feed through the klass in Allocate(...klass...)._klass.
2662 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2663 Node *x = klass_identity_common(phase);
2664
2665 const Type *t = phase->type( x );
2666 if( t == Type::TOP ) return x;
2667 if( t->isa_narrowklass()) return x;
2668 assert (!t->isa_narrowoop(), "no narrow oop here");
2669
2670 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2671 }
2672
2673 //------------------------------Value-----------------------------------------
2674 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2675 // Either input is TOP ==> the result is TOP
2676 const Type *t1 = phase->type( in(MemNode::Memory) );
2677 if( t1 == Type::TOP ) return Type::TOP;
2678 Node *adr = in(MemNode::Address);
2679 const Type *t2 = phase->type( adr );
2680 if( t2 == Type::TOP ) return Type::TOP;
2681 const TypePtr *tp = t2->is_ptr();
2682 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2683 const TypeAryPtr *tap = tp->isa_aryptr();
2684 if( !tap ) return _type;
2685 return tap->size();
2686 }
2687
2688 //-------------------------------Ideal---------------------------------------
2689 // Feed through the length in AllocateArray(...length...)._length.
2690 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2691 Node* p = MemNode::Ideal_common(phase, can_reshape);
2692 if (p) return (p == NodeSentinel) ? nullptr : p;
2693
2694 // Take apart the address into an oop and offset.
2695 // Return 'this' if we cannot.
2696 Node* adr = in(MemNode::Address);
2697 intptr_t offset = 0;
2698 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2699 if (base == nullptr) return nullptr;
2700 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2701 if (tary == nullptr) return nullptr;
2702
2703 // We can fetch the length directly through an AllocateArrayNode.
2704 // This works even if the length is not constant (clone or newArray).
2705 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2706 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2707 if (alloc != nullptr) {
2708 Node* allocated_length = alloc->Ideal_length();
2709 Node* len = alloc->make_ideal_length(tary, phase);
2710 if (allocated_length != len) {
2711 // New CastII improves on this.
2712 return len;
2713 }
2714 }
2715 }
2716
2717 return nullptr;
2718 }
2719
2720 //------------------------------Identity---------------------------------------
2721 // Feed through the length in AllocateArray(...length...)._length.
2722 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2723 Node* x = LoadINode::Identity(phase);
2724 if (x != this) return x;
2725
2726 // Take apart the address into an oop and offset.
2727 // Return 'this' if we cannot.
2728 Node* adr = in(MemNode::Address);
2729 intptr_t offset = 0;
2730 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2731 if (base == nullptr) return this;
2732 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2733 if (tary == nullptr) return this;
2734
2735 // We can fetch the length directly through an AllocateArrayNode.
2736 // This works even if the length is not constant (clone or newArray).
2737 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2738 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2739 if (alloc != nullptr) {
2740 Node* allocated_length = alloc->Ideal_length();
2741 // Do not allow make_ideal_length to allocate a CastII node.
2742 Node* len = alloc->make_ideal_length(tary, phase, false);
2743 if (allocated_length == len) {
2744 // Return allocated_length only if it would not be improved by a CastII.
2745 return allocated_length;
2746 }
2747 }
2748 }
2749
2750 return this;
2751
2752 }
2753
2754 //=============================================================================
2755 //---------------------------StoreNode::make-----------------------------------
2756 // Polymorphic factory method:
2757 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) {
2758 assert((mo == unordered || mo == release), "unexpected");
2759 Compile* C = gvn.C;
2760 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2761 ctl != nullptr, "raw memory operations should have control edge");
2762
2763 switch (bt) {
2764 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2765 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2766 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2767 case T_CHAR:
2768 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2769 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2770 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2771 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2772 case T_METADATA:
2773 case T_ADDRESS:
2774 case T_OBJECT:
2775 #ifdef _LP64
2776 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2777 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2778 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2779 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2780 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2781 adr->bottom_type()->isa_rawptr())) {
2782 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2783 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2784 }
2785 #endif
2786 {
2787 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2788 }
2789 default:
2790 ShouldNotReachHere();
2791 return (StoreNode*)nullptr;
2792 }
2793 }
2794
2795 //--------------------------bottom_type----------------------------------------
2796 const Type *StoreNode::bottom_type() const {
2797 return Type::MEMORY;
2798 }
2799
2800 //------------------------------hash-------------------------------------------
2801 uint StoreNode::hash() const {
2802 // unroll addition of interesting fields
2803 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2804
2805 // Since they are not commoned, do not hash them:
2806 return NO_HASH;
2807 }
2808
2809 // Link together multiple stores (B/S/C/I) into a longer one.
2810 //
2811 // Example: _store = StoreB[i+3]
2812 //
2813 // RangeCheck[i+0] RangeCheck[i+0]
2814 // StoreB[i+0]
2815 // RangeCheck[i+3] RangeCheck[i+3]
2816 // StoreB[i+1] --> pass: fail:
2817 // StoreB[i+2] StoreI[i+0] StoreB[i+0]
2818 // StoreB[i+3]
2819 //
2820 // The 4 StoreB are merged into a single StoreI node. We have to be careful with RangeCheck[i+1]: before
2821 // the optimization, if this RangeCheck[i+1] fails, then we execute only StoreB[i+0], and then trap. After
2822 // the optimization, the new StoreI[i+0] is on the passing path of RangeCheck[i+3], and StoreB[i+0] on the
2823 // failing path.
2824 //
2825 // Note: For normal array stores, every store at first has a RangeCheck. But they can be removed with:
2826 // - RCE (RangeCheck Elimination): the RangeChecks in the loop are hoisted out and before the loop,
2827 // and possibly no RangeChecks remain between the stores.
2828 // - RangeCheck smearing: the earlier RangeChecks are adjusted such that they cover later RangeChecks,
2829 // and those later RangeChecks can be removed. Example:
2830 //
2831 // RangeCheck[i+0] RangeCheck[i+0] <- before first store
2832 // StoreB[i+0] StoreB[i+0] <- first store
2833 // RangeCheck[i+1] --> smeared --> RangeCheck[i+3] <- only RC between first and last store
2834 // StoreB[i+1] StoreB[i+1] <- second store
2835 // RangeCheck[i+2] --> removed
2836 // StoreB[i+2] StoreB[i+2]
2837 // RangeCheck[i+3] --> removed
2838 // StoreB[i+3] StoreB[i+3] <- last store
2839 //
2840 // Thus, it is a common pattern that between the first and last store in a chain
2841 // of adjacent stores there remains exactly one RangeCheck, located between the
2842 // first and the second store (e.g. RangeCheck[i+3]).
2843 //
2844 class MergePrimitiveStores : public StackObj {
2845 private:
2846 PhaseGVN* const _phase;
2847 StoreNode* const _store;
2848 // State machine with initial state Unknown
2849 // Allowed transitions:
2850 // Unknown -> Const
2851 // Unknown -> Platform
2852 // Unknown -> Reverse
2853 // Unknown -> NotAdjacent
2854 // Const -> Const
2855 // Const -> NotAdjacent
2856 // Platform -> Platform
2857 // Platform -> NotAdjacent
2858 // Reverse -> Reverse
2859 // Reverse -> NotAdjacent
2860 // NotAdjacent -> NotAdjacent
2861 enum ValueOrder : uint8_t {
2862 Unknown, // Initial state
2863 Const, // Input values are const
2864 Platform, // Platform order
2865 Reverse, // Reverse platform order
2866 NotAdjacent // Not adjacent
2867 };
2868 ValueOrder _value_order;
2869
2870 NOT_PRODUCT( const CHeapBitMap &_trace_tags; )
2871
2872 public:
2873 MergePrimitiveStores(PhaseGVN* phase, StoreNode* store) :
2874 _phase(phase), _store(store), _value_order(ValueOrder::Unknown)
2875 NOT_PRODUCT( COMMA _trace_tags(Compile::current()->directive()->trace_merge_stores_tags()) )
2876 {}
2877
2878 StoreNode* run();
2879
2880 private:
2881 bool is_compatible_store(const StoreNode* other_store) const;
2882 bool is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const;
2883 bool is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const;
2884 static bool is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase);
2885 enum CFGStatus { SuccessNoRangeCheck, SuccessWithRangeCheck, Failure };
2886 static CFGStatus cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store);
2887
2888 class Status {
2889 private:
2890 StoreNode* _found_store;
2891 bool _found_range_check;
2892
2893 Status(StoreNode* found_store, bool found_range_check)
2894 : _found_store(found_store), _found_range_check(found_range_check) {}
2895
2896 public:
2897 StoreNode* found_store() const { return _found_store; }
2898 bool found_range_check() const { return _found_range_check; }
2899 static Status make_failure() { return Status(nullptr, false); }
2900
2901 static Status make(StoreNode* found_store, const CFGStatus cfg_status) {
2902 if (cfg_status == CFGStatus::Failure) {
2903 return Status::make_failure();
2904 }
2905 return Status(found_store, cfg_status == CFGStatus::SuccessWithRangeCheck);
2906 }
2907
2908 #ifndef PRODUCT
2909 void print_on(outputStream* st) const {
2910 if (_found_store == nullptr) {
2911 st->print_cr("None");
2912 } else {
2913 st->print_cr("Found[%d %s, %s]", _found_store->_idx, _found_store->Name(),
2914 _found_range_check ? "RC" : "no-RC");
2915 }
2916 }
2917 #endif
2918 };
2919
2920 enum ValueOrder find_adjacent_input_value_order(const Node* n1, const Node* n2, const int memory_size) const;
2921 Status find_adjacent_use_store(const StoreNode* def_store) const;
2922 Status find_adjacent_def_store(const StoreNode* use_store) const;
2923 Status find_use_store(const StoreNode* def_store) const;
2924 Status find_def_store(const StoreNode* use_store) const;
2925 Status find_use_store_unidirectional(const StoreNode* def_store) const;
2926 Status find_def_store_unidirectional(const StoreNode* use_store) const;
2927
2928 void collect_merge_list(Node_List& merge_list) const;
2929 Node* make_merged_input_value(const Node_List& merge_list);
2930 StoreNode* make_merged_store(const Node_List& merge_list, Node* merged_input_value);
2931
2932 #ifndef PRODUCT
2933 // Access to TraceMergeStores tags
2934 bool is_trace(TraceMergeStores::Tag tag) const {
2935 return _trace_tags.at(tag);
2936 }
2937
2938 bool is_trace_basic() const {
2939 return is_trace(TraceMergeStores::Tag::BASIC);
2940 }
2941
2942 bool is_trace_pointer_parsing() const {
2943 return is_trace(TraceMergeStores::Tag::POINTER_PARSING);
2944 }
2945
2946 bool is_trace_pointer_aliasing() const {
2947 return is_trace(TraceMergeStores::Tag::POINTER_ALIASING);
2948 }
2949
2950 bool is_trace_pointer_adjacency() const {
2951 return is_trace(TraceMergeStores::Tag::POINTER_ADJACENCY);
2952 }
2953
2954 bool is_trace_success() const {
2955 return is_trace(TraceMergeStores::Tag::SUCCESS);
2956 }
2957 #endif
2958
2959 NOT_PRODUCT( void trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const; )
2960 };
2961
2962 StoreNode* MergePrimitiveStores::run() {
2963 // Check for B/S/C/I
2964 int opc = _store->Opcode();
2965 if (opc != Op_StoreB && opc != Op_StoreC && opc != Op_StoreI) {
2966 return nullptr;
2967 }
2968
2969 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] MergePrimitiveStores::run: "); _store->dump(); })
2970
2971 // The _store must be the "last" store in a chain. If we find a use we could merge with
2972 // then that use or a store further down is the "last" store.
2973 Status status_use = find_adjacent_use_store(_store);
2974 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect no use: "); status_use.print_on(tty); })
2975 if (status_use.found_store() != nullptr) {
2976 return nullptr;
2977 }
2978
2979 // Check if we can merge with at least one def, so that we have at least 2 stores to merge.
2980 Status status_def = find_adjacent_def_store(_store);
2981 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect def: "); status_def.print_on(tty); })
2982 Node* def_store = status_def.found_store();
2983 if (def_store == nullptr) {
2984 return nullptr;
2985 }
2986
2987 // Initialize value order
2988 _value_order = find_adjacent_input_value_order(def_store->in(MemNode::ValueIn),
2989 _store->in(MemNode::ValueIn),
2990 _store->memory_size());
2991 assert(_value_order != ValueOrder::NotAdjacent && _value_order != ValueOrder::Unknown, "Order should be checked");
2992
2993 ResourceMark rm;
2994 Node_List merge_list;
2995 collect_merge_list(merge_list);
2996
2997 Node* merged_input_value = make_merged_input_value(merge_list);
2998 if (merged_input_value == nullptr) { return nullptr; }
2999
3000 StoreNode* merged_store = make_merged_store(merge_list, merged_input_value);
3001
3002 NOT_PRODUCT( if (is_trace_success()) { trace(merge_list, merged_input_value, merged_store); } )
3003
3004 return merged_store;
3005 }
3006
3007 // Check compatibility between _store and other_store.
3008 bool MergePrimitiveStores::is_compatible_store(const StoreNode* other_store) const {
3009 int opc = _store->Opcode();
3010 assert(opc == Op_StoreB || opc == Op_StoreC || opc == Op_StoreI, "precondition");
3011
3012 if (other_store == nullptr ||
3013 _store->Opcode() != other_store->Opcode()) {
3014 return false;
3015 }
3016
3017 return true;
3018 }
3019
3020 bool MergePrimitiveStores::is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const {
3021 if (!is_adjacent_input_pair(def_store->in(MemNode::ValueIn),
3022 use_store->in(MemNode::ValueIn),
3023 def_store->memory_size())) {
3024 return false;
3025 }
3026
3027 ResourceMark rm;
3028 #ifndef PRODUCT
3029 const TraceMemPointer trace(is_trace_pointer_parsing(),
3030 is_trace_pointer_aliasing(),
3031 is_trace_pointer_adjacency(),
3032 true);
3033 #endif
3034 const MemPointer pointer_use(use_store NOT_PRODUCT(COMMA trace));
3035 const MemPointer pointer_def(def_store NOT_PRODUCT(COMMA trace));
3036 return pointer_def.is_adjacent_to_and_before(pointer_use);
3037 }
3038
3039 // Check input values n1 and n2 can be merged and return the value order
3040 MergePrimitiveStores::ValueOrder MergePrimitiveStores::find_adjacent_input_value_order(const Node* n1, const Node* n2,
3041 const int memory_size) const {
3042 // Pattern: [n1 = ConI, n2 = ConI]
3043 if (n1->Opcode() == Op_ConI && n2->Opcode() == Op_ConI) {
3044 return ValueOrder::Const;
3045 }
3046
3047 Node const *base_n2;
3048 jint shift_n2;
3049 if (!is_con_RShift(n2, base_n2, shift_n2, _phase)) {
3050 return ValueOrder::NotAdjacent;
3051 }
3052 Node const *base_n1;
3053 jint shift_n1;
3054 if (!is_con_RShift(n1, base_n1, shift_n1, _phase)) {
3055 return ValueOrder::NotAdjacent;
3056 }
3057
3058 int bits_per_store = memory_size * 8;
3059 if (base_n1 != base_n2 ||
3060 abs(shift_n1 - shift_n2) != bits_per_store ||
3061 shift_n1 % bits_per_store != 0) {
3062 // Values are not adjacent
3063 return ValueOrder::NotAdjacent;
3064 }
3065
3066 // Detect value order
3067 #ifdef VM_LITTLE_ENDIAN
3068 return shift_n1 < shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3069 : ValueOrder::Reverse; // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3070 #else
3071 return shift_n1 > shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3072 : ValueOrder::Reverse; // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3073 #endif
3074 }
3075
3076 bool MergePrimitiveStores::is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const {
3077 ValueOrder input_value_order = find_adjacent_input_value_order(n1, n2, memory_size);
3078
3079 switch (input_value_order) {
3080 case ValueOrder::NotAdjacent:
3081 return false;
3082 case ValueOrder::Reverse:
3083 if (memory_size != 1 ||
3084 !Matcher::match_rule_supported(Op_ReverseBytesS) ||
3085 !Matcher::match_rule_supported(Op_ReverseBytesI) ||
3086 !Matcher::match_rule_supported(Op_ReverseBytesL)) {
3087 // ReverseBytes are not supported by platform
3088 return false;
3089 }
3090 // fall-through.
3091 case ValueOrder::Const:
3092 case ValueOrder::Platform:
3093 if (_value_order == ValueOrder::Unknown) {
3094 // Initial state is Unknown, and we find a valid input value order
3095 return true;
3096 }
3097 // The value order can not be changed
3098 return _value_order == input_value_order;
3099 case ValueOrder::Unknown:
3100 default:
3101 ShouldNotReachHere();
3102 }
3103 return false;
3104 }
3105
3106 // Detect pattern: n = base_out >> shift_out
3107 bool MergePrimitiveStores::is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase) {
3108 assert(n != nullptr, "precondition");
3109
3110 int opc = n->Opcode();
3111 if (opc == Op_ConvL2I) {
3112 n = n->in(1);
3113 opc = n->Opcode();
3114 }
3115
3116 if ((opc == Op_RShiftI ||
3117 opc == Op_RShiftL ||
3118 opc == Op_URShiftI ||
3119 opc == Op_URShiftL) &&
3120 n->in(2)->is_ConI()) {
3121 base_out = n->in(1);
3122 shift_out = n->in(2)->get_int();
3123 // The shift must be positive:
3124 return shift_out >= 0;
3125 }
3126
3127 if (phase->type(n)->isa_int() != nullptr ||
3128 phase->type(n)->isa_long() != nullptr) {
3129 // (base >> 0)
3130 base_out = n;
3131 shift_out = 0;
3132 return true;
3133 }
3134 return false;
3135 }
3136
3137 // Check if there is nothing between the two stores, except optionally a RangeCheck leading to an uncommon trap.
3138 MergePrimitiveStores::CFGStatus MergePrimitiveStores::cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store) {
3139 assert(use_store->in(MemNode::Memory) == def_store, "use-def relationship");
3140
3141 Node* ctrl_use = use_store->in(MemNode::Control);
3142 Node* ctrl_def = def_store->in(MemNode::Control);
3143 if (ctrl_use == nullptr || ctrl_def == nullptr) {
3144 return CFGStatus::Failure;
3145 }
3146
3147 if (ctrl_use == ctrl_def) {
3148 // Same ctrl -> no RangeCheck in between.
3149 // Check: use_store must be the only use of def_store.
3150 if (def_store->outcnt() > 1) {
3151 return CFGStatus::Failure;
3152 }
3153 return CFGStatus::SuccessNoRangeCheck;
3154 }
3155
3156 // Different ctrl -> could have RangeCheck in between.
3157 // Check: 1. def_store only has these uses: use_store and MergeMem for uncommon trap, and
3158 // 2. ctrl separated by RangeCheck.
3159 if (def_store->outcnt() != 2) {
3160 return CFGStatus::Failure; // Cannot have exactly these uses: use_store and MergeMem for uncommon trap.
3161 }
3162 int use_store_out_idx = def_store->raw_out(0) == use_store ? 0 : 1;
3163 Node* merge_mem = def_store->raw_out(1 - use_store_out_idx)->isa_MergeMem();
3164 if (merge_mem == nullptr ||
3165 merge_mem->outcnt() != 1) {
3166 return CFGStatus::Failure; // Does not have MergeMem for uncommon trap.
3167 }
3168 if (!ctrl_use->is_IfProj() ||
3169 !ctrl_use->in(0)->is_RangeCheck() ||
3170 ctrl_use->in(0)->outcnt() != 2) {
3171 return CFGStatus::Failure; // Not RangeCheck.
3172 }
3173 IfProjNode* other_proj = ctrl_use->as_IfProj()->other_if_proj();
3174 Node* trap = other_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check);
3175 if (trap != merge_mem->unique_out() ||
3176 ctrl_use->in(0)->in(0) != ctrl_def) {
3177 return CFGStatus::Failure; // Not RangeCheck with merge_mem leading to uncommon trap.
3178 }
3179
3180 return CFGStatus::SuccessWithRangeCheck;
3181 }
3182
3183 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_use_store(const StoreNode* def_store) const {
3184 Status status_use = find_use_store(def_store);
3185 StoreNode* use_store = status_use.found_store();
3186 if (use_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3187 return Status::make_failure();
3188 }
3189 return status_use;
3190 }
3191
3192 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_def_store(const StoreNode* use_store) const {
3193 Status status_def = find_def_store(use_store);
3194 StoreNode* def_store = status_def.found_store();
3195 if (def_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3196 return Status::make_failure();
3197 }
3198 return status_def;
3199 }
3200
3201 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store(const StoreNode* def_store) const {
3202 Status status_use = find_use_store_unidirectional(def_store);
3203
3204 #ifdef ASSERT
3205 StoreNode* use_store = status_use.found_store();
3206 if (use_store != nullptr) {
3207 Status status_def = find_def_store_unidirectional(use_store);
3208 assert(status_def.found_store() == def_store &&
3209 status_def.found_range_check() == status_use.found_range_check(),
3210 "find_use_store and find_def_store must be symmetric");
3211 }
3212 #endif
3213
3214 return status_use;
3215 }
3216
3217 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store(const StoreNode* use_store) const {
3218 Status status_def = find_def_store_unidirectional(use_store);
3219
3220 #ifdef ASSERT
3221 StoreNode* def_store = status_def.found_store();
3222 if (def_store != nullptr) {
3223 Status status_use = find_use_store_unidirectional(def_store);
3224 assert(status_use.found_store() == use_store &&
3225 status_use.found_range_check() == status_def.found_range_check(),
3226 "find_use_store and find_def_store must be symmetric");
3227 }
3228 #endif
3229
3230 return status_def;
3231 }
3232
3233 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store_unidirectional(const StoreNode* def_store) const {
3234 assert(is_compatible_store(def_store), "precondition: must be compatible with _store");
3235
3236 for (DUIterator_Fast imax, i = def_store->fast_outs(imax); i < imax; i++) {
3237 StoreNode* use_store = def_store->fast_out(i)->isa_Store();
3238 if (is_compatible_store(use_store)) {
3239 return Status::make(use_store, cfg_status_for_pair(use_store, def_store));
3240 }
3241 }
3242
3243 return Status::make_failure();
3244 }
3245
3246 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store_unidirectional(const StoreNode* use_store) const {
3247 assert(is_compatible_store(use_store), "precondition: must be compatible with _store");
3248
3249 StoreNode* def_store = use_store->in(MemNode::Memory)->isa_Store();
3250 if (!is_compatible_store(def_store)) {
3251 return Status::make_failure();
3252 }
3253
3254 return Status::make(def_store, cfg_status_for_pair(use_store, def_store));
3255 }
3256
3257 void MergePrimitiveStores::collect_merge_list(Node_List& merge_list) const {
3258 // The merged store can be at most 8 bytes.
3259 const uint merge_list_max_size = 8 / _store->memory_size();
3260 assert(merge_list_max_size >= 2 &&
3261 merge_list_max_size <= 8 &&
3262 is_power_of_2(merge_list_max_size),
3263 "must be 2, 4 or 8");
3264
3265 // Traverse up the chain of adjacent def stores.
3266 StoreNode* current = _store;
3267 merge_list.push(current);
3268 while (current != nullptr && merge_list.size() < merge_list_max_size) {
3269 Status status = find_adjacent_def_store(current);
3270 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] find def: "); status.print_on(tty); })
3271
3272 current = status.found_store();
3273 if (current != nullptr) {
3274 merge_list.push(current);
3275
3276 // We can have at most one RangeCheck.
3277 if (status.found_range_check()) {
3278 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found RangeCheck, stop traversal."); })
3279 break;
3280 }
3281 }
3282 }
3283
3284 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found:"); merge_list.dump(); })
3285
3286 // Truncate the merge_list to a power of 2.
3287 const uint pow2size = round_down_power_of_2(merge_list.size());
3288 assert(pow2size >= 2, "must be merging at least 2 stores");
3289 while (merge_list.size() > pow2size) { merge_list.pop(); }
3290
3291 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] truncated:"); merge_list.dump(); })
3292 }
3293
3294 // Merge the input values of the smaller stores to a single larger input value.
3295 Node* MergePrimitiveStores::make_merged_input_value(const Node_List& merge_list) {
3296 int new_memory_size = _store->memory_size() * merge_list.size();
3297 Node* first = merge_list.at(merge_list.size()-1);
3298 Node* merged_input_value = nullptr;
3299 if (_store->in(MemNode::ValueIn)->Opcode() == Op_ConI) {
3300 assert(_value_order == ValueOrder::Const, "must match");
3301 // Pattern: [ConI, ConI, ...] -> new constant
3302 jlong con = 0;
3303 jlong bits_per_store = _store->memory_size() * 8;
3304 jlong mask = (((jlong)1) << bits_per_store) - 1;
3305 for (uint i = 0; i < merge_list.size(); i++) {
3306 jlong con_i = merge_list.at(i)->in(MemNode::ValueIn)->get_int();
3307 #ifdef VM_LITTLE_ENDIAN
3308 con = con << bits_per_store;
3309 con = con | (mask & con_i);
3310 #else // VM_LITTLE_ENDIAN
3311 con_i = (mask & con_i) << (i * bits_per_store);
3312 con = con | con_i;
3313 #endif // VM_LITTLE_ENDIAN
3314 }
3315 merged_input_value = _phase->longcon(con);
3316 } else {
3317 assert(_value_order == ValueOrder::Platform || _value_order == ValueOrder::Reverse, "must match");
3318 // Pattern: [base >> 24, base >> 16, base >> 8, base] -> base
3319 // | |
3320 // _store first
3321 //
3322 Node* hi = _store->in(MemNode::ValueIn);
3323 Node* lo = first->in(MemNode::ValueIn);
3324 #ifndef VM_LITTLE_ENDIAN
3325 // `_store` and `first` are swapped in the diagram above
3326 swap(hi, lo);
3327 #endif // !VM_LITTLE_ENDIAN
3328 if (_value_order == ValueOrder::Reverse) {
3329 swap(hi, lo);
3330 }
3331 Node const* hi_base;
3332 jint hi_shift;
3333 merged_input_value = lo;
3334 bool is_true = is_con_RShift(hi, hi_base, hi_shift, _phase);
3335 assert(is_true, "must detect con RShift");
3336 if (merged_input_value != hi_base && merged_input_value->Opcode() == Op_ConvL2I) {
3337 // look through
3338 merged_input_value = merged_input_value->in(1);
3339 }
3340 if (merged_input_value != hi_base) {
3341 // merged_input_value is not the base
3342 return nullptr;
3343 }
3344 }
3345
3346 if (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size <= 4) {
3347 // Example:
3348 //
3349 // long base = ...;
3350 // a[0] = (byte)(base >> 0);
3351 // a[1] = (byte)(base >> 8);
3352 //
3353 merged_input_value = _phase->transform(new ConvL2INode(merged_input_value));
3354 }
3355
3356 assert((_phase->type(merged_input_value)->isa_int() != nullptr && new_memory_size <= 4) ||
3357 (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size == 8),
3358 "merged_input_value is either int or long, and new_memory_size is small enough");
3359
3360 if (_value_order == ValueOrder::Reverse) {
3361 assert(_store->memory_size() == 1, "only implemented for bytes");
3362 if (new_memory_size == 8) {
3363 merged_input_value = _phase->transform(new ReverseBytesLNode(merged_input_value));
3364 } else if (new_memory_size == 4) {
3365 merged_input_value = _phase->transform(new ReverseBytesINode(merged_input_value));
3366 } else {
3367 assert(new_memory_size == 2, "sanity check");
3368 merged_input_value = _phase->transform(new ReverseBytesSNode(merged_input_value));
3369 }
3370 }
3371 return merged_input_value;
3372 }
3373
3374 // //
3375 // first_ctrl first_mem first_adr first_ctrl first_mem first_adr //
3376 // | | | | | | //
3377 // | | | | +---------------+ | //
3378 // | | | | | | | //
3379 // | | +---------+ | | +---------------+ //
3380 // | | | | | | | | //
3381 // +--------------+ | | v1 +------------------------------+ | | v1 //
3382 // | | | | | | | | | | | | //
3383 // RangeCheck first_store RangeCheck | | first_store //
3384 // | | | | | | | //
3385 // last_ctrl | +----> unc_trap last_ctrl | | +----> unc_trap //
3386 // | | ===> | | | //
3387 // +--------------+ | a2 v2 | | | //
3388 // | | | | | | | | //
3389 // | second_store | | | //
3390 // | | | | | [v1 v2 ... vn] //
3391 // ... ... | | | | //
3392 // | | | | | v //
3393 // +--------------+ | an vn +--------------+ | | merged_input_value //
3394 // | | | | | | | | //
3395 // last_store (= _store) merged_store //
3396 // //
3397 StoreNode* MergePrimitiveStores::make_merged_store(const Node_List& merge_list, Node* merged_input_value) {
3398 Node* first_store = merge_list.at(merge_list.size()-1);
3399 Node* last_ctrl = _store->in(MemNode::Control); // after (optional) RangeCheck
3400 Node* first_mem = first_store->in(MemNode::Memory);
3401 Node* first_adr = first_store->in(MemNode::Address);
3402
3403 const TypePtr* new_adr_type = _store->adr_type();
3404
3405 int new_memory_size = _store->memory_size() * merge_list.size();
3406 BasicType bt = T_ILLEGAL;
3407 switch (new_memory_size) {
3408 case 2: bt = T_SHORT; break;
3409 case 4: bt = T_INT; break;
3410 case 8: bt = T_LONG; break;
3411 }
3412
3413 StoreNode* merged_store = StoreNode::make(*_phase, last_ctrl, first_mem, first_adr,
3414 new_adr_type, merged_input_value, bt, MemNode::unordered);
3415
3416 // Marking the store mismatched is sufficient to prevent reordering, since array stores
3417 // are all on the same slice. Hence, we need no barriers.
3418 merged_store->set_mismatched_access();
3419
3420 // Constants above may now also be be packed -> put candidate on worklist
3421 _phase->is_IterGVN()->_worklist.push(first_mem);
3422
3423 return merged_store;
3424 }
3425
3426 #ifndef PRODUCT
3427 void MergePrimitiveStores::trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const {
3428 stringStream ss;
3429 ss.print_cr("[TraceMergeStores]: Replace");
3430 for (int i = (int)merge_list.size() - 1; i >= 0; i--) {
3431 merge_list.at(i)->dump("\n", false, &ss);
3432 }
3433 ss.print_cr("[TraceMergeStores]: with");
3434 merged_input_value->dump("\n", false, &ss);
3435 merged_store->dump("\n", false, &ss);
3436 tty->print("%s", ss.as_string());
3437 }
3438 #endif
3439
3440 //------------------------------Ideal------------------------------------------
3441 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
3442 // When a store immediately follows a relevant allocation/initialization,
3443 // try to capture it into the initialization, or hoist it above.
3444 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3445 Node* p = MemNode::Ideal_common(phase, can_reshape);
3446 if (p) return (p == NodeSentinel) ? nullptr : p;
3447
3448 Node* mem = in(MemNode::Memory);
3449 Node* address = in(MemNode::Address);
3450 Node* value = in(MemNode::ValueIn);
3451 // Back-to-back stores to same address? Fold em up. Generally
3452 // unsafe if I have intervening uses.
3453 {
3454 Node* st = mem;
3455 // If Store 'st' has more than one use, we cannot fold 'st' away.
3456 // For example, 'st' might be the final state at a conditional
3457 // return. Or, 'st' might be used by some node which is live at
3458 // the same time 'st' is live, which might be unschedulable. So,
3459 // require exactly ONE user until such time as we clone 'mem' for
3460 // each of 'mem's uses (thus making the exactly-1-user-rule hold
3461 // true).
3462 while (st->is_Store() && st->outcnt() == 1) {
3463 // Looking at a dead closed cycle of memory?
3464 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
3465 assert(Opcode() == st->Opcode() ||
3466 st->Opcode() == Op_StoreVector ||
3467 Opcode() == Op_StoreVector ||
3468 st->Opcode() == Op_StoreVectorScatter ||
3469 Opcode() == Op_StoreVectorScatter ||
3470 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
3471 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
3472 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
3473 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
3474 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
3475
3476 if (st->in(MemNode::Address)->eqv_uncast(address) &&
3477 st->as_Store()->memory_size() <= this->memory_size()) {
3478 Node* use = st->raw_out(0);
3479 if (phase->is_IterGVN()) {
3480 phase->is_IterGVN()->rehash_node_delayed(use);
3481 }
3482 // It's OK to do this in the parser, since DU info is always accurate,
3483 // and the parser always refers to nodes via SafePointNode maps.
3484 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
3485 return this;
3486 }
3487 st = st->in(MemNode::Memory);
3488 }
3489 }
3490
3491
3492 // Capture an unaliased, unconditional, simple store into an initializer.
3493 // Or, if it is independent of the allocation, hoist it above the allocation.
3494 if (ReduceFieldZeroing && /*can_reshape &&*/
3495 mem->is_Proj() && mem->in(0)->is_Initialize()) {
3496 InitializeNode* init = mem->in(0)->as_Initialize();
3497 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
3498 if (offset > 0) {
3499 Node* moved = init->capture_store(this, offset, phase, can_reshape);
3500 // If the InitializeNode captured me, it made a raw copy of me,
3501 // and I need to disappear.
3502 if (moved != nullptr) {
3503 // %%% hack to ensure that Ideal returns a new node:
3504 mem = MergeMemNode::make(mem);
3505 return mem; // fold me away
3506 }
3507 }
3508 }
3509
3510 // Fold reinterpret cast into memory operation:
3511 // StoreX mem (MoveY2X v) => StoreY mem v
3512 if (value->is_Move()) {
3513 const Type* vt = value->in(1)->bottom_type();
3514 if (has_reinterpret_variant(vt)) {
3515 if (phase->C->post_loop_opts_phase()) {
3516 return convert_to_reinterpret_store(*phase, value->in(1), vt);
3517 } else {
3518 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
3519 }
3520 }
3521 }
3522
3523 if (MergeStores && UseUnalignedAccesses) {
3524 if (phase->C->merge_stores_phase()) {
3525 MergePrimitiveStores merge(phase, this);
3526 Node* progress = merge.run();
3527 if (progress != nullptr) { return progress; }
3528 } else {
3529 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
3530 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
3531 // the stores, and we merge the wrong sequence of stores.
3532 // Example:
3533 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
3534 // Apply MergeStores:
3535 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
3536 // Remove more RangeChecks:
3537 // StoreI [ StoreL ] StoreI
3538 // But now it would have been better to do this instead:
3539 // [ StoreL ] [ StoreL ]
3540 phase->C->record_for_merge_stores_igvn(this);
3541 }
3542 }
3543
3544 return nullptr; // No further progress
3545 }
3546
3547 //------------------------------Value-----------------------------------------
3548 const Type* StoreNode::Value(PhaseGVN* phase) const {
3549 // Either input is TOP ==> the result is TOP
3550 const Type *t1 = phase->type( in(MemNode::Memory) );
3551 if( t1 == Type::TOP ) return Type::TOP;
3552 const Type *t2 = phase->type( in(MemNode::Address) );
3553 if( t2 == Type::TOP ) return Type::TOP;
3554 const Type *t3 = phase->type( in(MemNode::ValueIn) );
3555 if( t3 == Type::TOP ) return Type::TOP;
3556 return Type::MEMORY;
3557 }
3558
3559 //------------------------------Identity---------------------------------------
3560 // Remove redundant stores:
3561 // Store(m, p, Load(m, p)) changes to m.
3562 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
3563 Node* StoreNode::Identity(PhaseGVN* phase) {
3564 Node* mem = in(MemNode::Memory);
3565 Node* adr = in(MemNode::Address);
3566 Node* val = in(MemNode::ValueIn);
3567
3568 Node* result = this;
3569
3570 // Load then Store? Then the Store is useless
3571 if (val->is_Load() &&
3572 val->in(MemNode::Address)->eqv_uncast(adr) &&
3573 val->in(MemNode::Memory )->eqv_uncast(mem) &&
3574 val->as_Load()->store_Opcode() == Opcode()) {
3575 if (!is_StoreVector()) {
3576 result = mem;
3577 } else if (Opcode() == Op_StoreVector && val->Opcode() == Op_LoadVector &&
3578 as_StoreVector()->vect_type() == val->as_LoadVector()->vect_type()) {
3579 // Ensure both are not masked accesses or gathers/scatters and vector types are the same
3580 result = mem;
3581 }
3582 }
3583
3584 // Two stores in a row of the same value?
3585 if (result == this &&
3586 mem->is_Store() &&
3587 mem->in(MemNode::Address)->eqv_uncast(adr) &&
3588 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
3589 mem->Opcode() == Opcode()) {
3590 if (!is_StoreVector()) {
3591 result = mem;
3592 } else {
3593 const StoreVectorNode* store_vector = as_StoreVector();
3594 const StoreVectorNode* mem_vector = mem->as_StoreVector();
3595 const Node* store_indices = store_vector->indices();
3596 const Node* mem_indices = mem_vector->indices();
3597 const Node* store_mask = store_vector->mask();
3598 const Node* mem_mask = mem_vector->mask();
3599 // Ensure types, indices, and masks match
3600 if (store_vector->vect_type() == mem_vector->vect_type() &&
3601 ((store_indices == nullptr) == (mem_indices == nullptr) &&
3602 (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) &&
3603 ((store_mask == nullptr) == (mem_mask == nullptr) &&
3604 (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) {
3605 result = mem;
3606 }
3607 }
3608 }
3609
3610 // Store of zero anywhere into a freshly-allocated object?
3611 // Then the store is useless.
3612 // (It must already have been captured by the InitializeNode.)
3613 if (result == this &&
3614 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
3615 // a newly allocated object is already all-zeroes everywhere
3616 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
3617 result = mem;
3618 }
3619
3620 if (result == this) {
3621 // the store may also apply to zero-bits in an earlier object
3622 Node* prev_mem = find_previous_store(phase);
3623 // Steps (a), (b): Walk past independent stores to find an exact match.
3624 if (prev_mem != nullptr) {
3625 if (prev_mem->is_top()) {
3626 // find_previous_store returns top when the access is dead
3627 return prev_mem;
3628 }
3629 Node* prev_val = can_see_stored_value(prev_mem, phase);
3630 if (prev_val != nullptr && prev_val == val) {
3631 // prev_val and val might differ by a cast; it would be good
3632 // to keep the more informative of the two.
3633 result = mem;
3634 }
3635 }
3636 }
3637 }
3638
3639 PhaseIterGVN* igvn = phase->is_IterGVN();
3640 if (result != this && igvn != nullptr) {
3641 MemBarNode* trailing = trailing_membar();
3642 if (trailing != nullptr) {
3643 #ifdef ASSERT
3644 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
3645 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects");
3646 #endif
3647 trailing->remove(igvn);
3648 }
3649 }
3650
3651 return result;
3652 }
3653
3654 //------------------------------match_edge-------------------------------------
3655 // Do we Match on this edge index or not? Match only memory & value
3656 uint StoreNode::match_edge(uint idx) const {
3657 return idx == MemNode::Address || idx == MemNode::ValueIn;
3658 }
3659
3660 //------------------------------cmp--------------------------------------------
3661 // Do not common stores up together. They generally have to be split
3662 // back up anyways, so do not bother.
3663 bool StoreNode::cmp( const Node &n ) const {
3664 return (&n == this); // Always fail except on self
3665 }
3666
3667 //------------------------------Ideal_masked_input-----------------------------
3668 // Check for a useless mask before a partial-word store
3669 // (StoreB ... (AndI valIn conIa) )
3670 // If (conIa & mask == mask) this simplifies to
3671 // (StoreB ... (valIn) )
3672 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
3673 Node *val = in(MemNode::ValueIn);
3674 if( val->Opcode() == Op_AndI ) {
3675 const TypeInt *t = phase->type( val->in(2) )->isa_int();
3676 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
3677 set_req_X(MemNode::ValueIn, val->in(1), phase);
3678 return this;
3679 }
3680 }
3681 return nullptr;
3682 }
3683
3684
3685 //------------------------------Ideal_sign_extended_input----------------------
3686 // Check for useless sign-extension before a partial-word store
3687 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
3688 // If (conIL == conIR && conIR <= num_rejected_bits) this simplifies to
3689 // (StoreB ... (v))
3690 // If (conIL > conIR) under some conditions, it can be simplified into
3691 // (StoreB ... (LShiftI _ v (conIL - conIR)))
3692 // This case happens when the value of the store was itself a left shift, that
3693 // gets merged into the inner left shift of the sign-extension. For instance,
3694 // if we have
3695 // array_of_shorts[0] = (short)(v << 2)
3696 // We get a structure such as:
3697 // (StoreB ... (RShiftI _ (LShiftI _ (LShiftI _ v 2) 16) 16))
3698 // that is simplified into
3699 // (StoreB ... (RShiftI _ (LShiftI _ v 18) 16)).
3700 // It is thus useful to handle cases where conIL > conIR. But this simplification
3701 // does not always hold. Let's see in which cases it's valid.
3702 //
3703 // Let's assume we have the following 32 bits integer v:
3704 // +----------------------------------+
3705 // | v[0..31] |
3706 // +----------------------------------+
3707 // 31 0
3708 // that will be stuffed in 8 bits byte after a shift left and a shift right of
3709 // potentially different magnitudes.
3710 // We denote num_rejected_bits the number of bits of the discarded part. In this
3711 // case, num_rejected_bits == 24.
3712 //
3713 // Statement (proved further below in case analysis):
3714 // Given:
3715 // - 0 <= conIL < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
3716 // - 0 <= conIR < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
3717 // - conIL >= conIR
3718 // - num_rejected_bits >= conIR
3719 // Then this form:
3720 // (RShiftI _ (LShiftI _ v conIL) conIR)
3721 // can be replaced with this form:
3722 // (LShiftI _ v (conIL-conIR))
3723 //
3724 // Note: We only have to show that the non-rejected lowest bits (8 bits for byte)
3725 // have to be correct, as the higher bits are rejected / truncated by the store.
3726 //
3727 // The hypotheses
3728 // 0 <= conIL < BitsPerJavaInteger
3729 // 0 <= conIR < BitsPerJavaInteger
3730 // are ensured by maskShiftAmount (called from ::Ideal of shift nodes). Indeed,
3731 // (v << 31) << 2 must be simplified into 0, not into v << 33 (which is equivalent
3732 // to v << 1).
3733 //
3734 //
3735 // If you don't like case analysis, jump after the conclusion.
3736 // ### Case 1 : conIL == conIR
3737 // ###### Case 1.1: conIL == conIR == num_rejected_bits
3738 // If we do the shift left then right by 24 bits, we get:
3739 // after: v << 24
3740 // +---------+------------------------+
3741 // | v[0..7] | 0 |
3742 // +---------+------------------------+
3743 // 31 24 23 0
3744 // after: (v << 24) >> 24
3745 // +------------------------+---------+
3746 // | sign bit | v[0..7] |
3747 // +------------------------+---------+
3748 // 31 8 7 0
3749 // The non-rejected bits (bits kept by the store, that is the 8 lower bits of the
3750 // result) are the same before and after, so, indeed, simplifying is correct.
3751
3752 // ###### Case 1.2: conIL == conIR < num_rejected_bits
3753 // If we do the shift left then right by 22 bits, we get:
3754 // after: v << 22
3755 // +---------+------------------------+
3756 // | v[0..9] | 0 |
3757 // +---------+------------------------+
3758 // 31 22 21 0
3759 // after: (v << 22) >> 22
3760 // +------------------------+---------+
3761 // | sign bit | v[0..9] |
3762 // +------------------------+---------+
3763 // 31 10 9 0
3764 // The non-rejected bits are the 8 lower bits of v. The bits 8 and 9 of v are still
3765 // present in (v << 22) >> 22 but will be dropped by the store. The simplification is
3766 // still correct.
3767
3768 // ###### But! Case 1.3: conIL == conIR > num_rejected_bits
3769 // If we do the shift left then right by 26 bits, we get:
3770 // after: v << 26
3771 // +---------+------------------------+
3772 // | v[0..5] | 0 |
3773 // +---------+------------------------+
3774 // 31 26 25 0
3775 // after: (v << 26) >> 26
3776 // +------------------------+---------+
3777 // | sign bit | v[0..5] |
3778 // +------------------------+---------+
3779 // 31 6 5 0
3780 // The non-rejected bits are made of
3781 // - 0-5 => the bits 0 to 5 of v
3782 // - 6-7 => the sign bit of v[0..5] (that is v[5])
3783 // Simplifying this as v is not correct.
3784 // The condition conIR <= num_rejected_bits is indeed necessary in Case 1
3785 //
3786 // ### Case 2: conIL > conIR
3787 // ###### Case 2.1: num_rejected_bits == conIR
3788 // We take conIL == 26 for this example.
3789 // after: v << 26
3790 // +---------+------------------------+
3791 // | v[0..5] | 0 |
3792 // +---------+------------------------+
3793 // 31 26 25 0
3794 // after: (v << 26) >> 24
3795 // +------------------+---------+-----+
3796 // | sign bit | v[0..5] | 0 |
3797 // +------------------+---------+-----+
3798 // 31 8 7 2 1 0
3799 // The non-rejected bits are the 8 lower ones of (v << conIL - conIR).
3800 // The bits 6 and 7 of v have been thrown away after the shift left.
3801 // The simplification is still correct.
3802 //
3803 // ###### Case 2.2: num_rejected_bits > conIR.
3804 // Let's say conIL == 26 and conIR == 22.
3805 // after: v << 26
3806 // +---------+------------------------+
3807 // | v[0..5] | 0 |
3808 // +---------+------------------------+
3809 // 31 26 25 0
3810 // after: (v << 26) >> 22
3811 // +------------------+---------+-----+
3812 // | sign bit | v[0..5] | 0 |
3813 // +------------------+---------+-----+
3814 // 31 10 9 4 3 0
3815 // The bits non-rejected by the store are exactly the 8 lower ones of (v << (conIL - conIR)):
3816 // - 0-3 => 0
3817 // - 4-7 => bits 0 to 3 of v
3818 // The simplification is still correct.
3819 // The bits 4 and 5 of v are still present in (v << (conIL - conIR)) but they don't
3820 // matter as they are not in the 8 lower bits: they will be cut out by the store.
3821 //
3822 // ###### But! Case 2.3: num_rejected_bits < conIR.
3823 // Let's see that this case is not as easy to simplify.
3824 // Let's say conIL == 28 and conIR == 26.
3825 // after: v << 28
3826 // +---------+------------------------+
3827 // | v[0..3] | 0 |
3828 // +---------+------------------------+
3829 // 31 28 27 0
3830 // after: (v << 28) >> 26
3831 // +------------------+---------+-----+
3832 // | sign bit | v[0..3] | 0 |
3833 // +------------------+---------+-----+
3834 // 31 6 5 2 1 0
3835 // The non-rejected bits are made of
3836 // - 0-1 => 0
3837 // - 2-5 => the bits 0 to 3 of v
3838 // - 6-7 => the sign bit of v[0..3] (that is v[3])
3839 // Simplifying this as (v << 2) is not correct.
3840 // The condition conIR <= num_rejected_bits is indeed necessary in Case 2.
3841 //
3842 // ### Conclusion:
3843 // Our hypotheses are indeed sufficient:
3844 // - 0 <= conIL < BitsPerJavaInteger
3845 // - 0 <= conIR < BitsPerJavaInteger
3846 // - conIL >= conIR
3847 // - num_rejected_bits >= conIR
3848 //
3849 // ### A rationale without case analysis:
3850 // After the shift left, conIL upper bits of v are discarded and conIL lower bit
3851 // zeroes are added. After the shift right, conIR lower bits of the previous result
3852 // are discarded. If conIL >= conIR, we discard only the zeroes we made up during
3853 // the shift left, but if conIL < conIR, then we discard also lower bits of v. But
3854 // the point of the simplification is to get an expression of the form
3855 // (v << (conIL - conIR)). This expression discard only higher bits of v, thus the
3856 // simplification is not correct if conIL < conIR.
3857 //
3858 // Moreover, after the shift right, the higher bit of (v << conIL) is repeated on the
3859 // conIR higher bits of ((v << conIL) >> conIR), it's the sign-extension. If
3860 // conIR > num_rejected_bits, then at least one artificial copy of this sign bit will
3861 // be in the window of the store. Thus ((v << conIL) >> conIR) is not equivalent to
3862 // (v << (conIL-conIR)) if conIR > num_rejected_bits.
3863 //
3864 // We do not treat the case conIL < conIR here since the point of this function is
3865 // to skip sign-extensions (that is conIL == conIR == num_rejected_bits). The need
3866 // of treating conIL > conIR comes from the cases where the sign-extended value is
3867 // also left-shift expression. Computing the sign-extension of a right-shift expression
3868 // doesn't yield a situation such as
3869 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
3870 // where conIL < conIR.
3871 Node* StoreNode::Ideal_sign_extended_input(PhaseGVN* phase, int num_rejected_bits) {
3872 Node* shr = in(MemNode::ValueIn);
3873 if (shr->Opcode() == Op_RShiftI) {
3874 const TypeInt* conIR = phase->type(shr->in(2))->isa_int();
3875 if (conIR != nullptr && conIR->is_con() && conIR->get_con() >= 0 && conIR->get_con() < BitsPerJavaInteger && conIR->get_con() <= num_rejected_bits) {
3876 Node* shl = shr->in(1);
3877 if (shl->Opcode() == Op_LShiftI) {
3878 const TypeInt* conIL = phase->type(shl->in(2))->isa_int();
3879 if (conIL != nullptr && conIL->is_con() && conIL->get_con() >= 0 && conIL->get_con() < BitsPerJavaInteger) {
3880 if (conIL->get_con() == conIR->get_con()) {
3881 set_req_X(MemNode::ValueIn, shl->in(1), phase);
3882 return this;
3883 }
3884 if (conIL->get_con() > conIR->get_con()) {
3885 Node* new_shl = phase->transform(new LShiftINode(shl->in(1), phase->intcon(conIL->get_con() - conIR->get_con())));
3886 set_req_X(MemNode::ValueIn, new_shl, phase);
3887 return this;
3888 }
3889 }
3890 }
3891 }
3892 }
3893 return nullptr;
3894 }
3895
3896 //------------------------------value_never_loaded-----------------------------------
3897 // Determine whether there are any possible loads of the value stored.
3898 // For simplicity, we actually check if there are any loads from the
3899 // address stored to, not just for loads of the value stored by this node.
3900 //
3901 bool StoreNode::value_never_loaded(PhaseValues* phase) const {
3902 Node *adr = in(Address);
3903 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
3904 if (adr_oop == nullptr)
3905 return false;
3906 if (!adr_oop->is_known_instance_field())
3907 return false; // if not a distinct instance, there may be aliases of the address
3908 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
3909 Node *use = adr->fast_out(i);
3910 if (use->is_Load() || use->is_LoadStore()) {
3911 return false;
3912 }
3913 }
3914 return true;
3915 }
3916
3917 MemBarNode* StoreNode::trailing_membar() const {
3918 if (is_release()) {
3919 MemBarNode* trailing_mb = nullptr;
3920 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3921 Node* u = fast_out(i);
3922 if (u->is_MemBar()) {
3923 if (u->as_MemBar()->trailing_store()) {
3924 assert(u->Opcode() == Op_MemBarVolatile, "");
3925 assert(trailing_mb == nullptr, "only one");
3926 trailing_mb = u->as_MemBar();
3927 #ifdef ASSERT
3928 Node* leading = u->as_MemBar()->leading_membar();
3929 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3930 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
3931 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
3932 #endif
3933 } else {
3934 assert(u->as_MemBar()->standalone(), "");
3935 }
3936 }
3937 }
3938 return trailing_mb;
3939 }
3940 return nullptr;
3941 }
3942
3943
3944 //=============================================================================
3945 //------------------------------Ideal------------------------------------------
3946 // If the store is from an AND mask that leaves the low bits untouched, then
3947 // we can skip the AND operation. If the store is from a sign-extension
3948 // (a left shift, then right shift) we can skip both.
3949 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
3950 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
3951 if( progress != nullptr ) return progress;
3952
3953 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
3954 if( progress != nullptr ) return progress;
3955
3956 // Finally check the default case
3957 return StoreNode::Ideal(phase, can_reshape);
3958 }
3959
3960 //=============================================================================
3961 //------------------------------Ideal------------------------------------------
3962 // If the store is from an AND mask that leaves the low bits untouched, then
3963 // we can skip the AND operation
3964 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
3965 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
3966 if( progress != nullptr ) return progress;
3967
3968 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
3969 if( progress != nullptr ) return progress;
3970
3971 // Finally check the default case
3972 return StoreNode::Ideal(phase, can_reshape);
3973 }
3974
3975 //=============================================================================
3976 //----------------------------------SCMemProjNode------------------------------
3977 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
3978 {
3979 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
3980 return Type::TOP;
3981 }
3982 return bottom_type();
3983 }
3984
3985 //=============================================================================
3986 //----------------------------------LoadStoreNode------------------------------
3987 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
3988 : Node(required),
3989 _type(rt),
3990 _barrier_data(0)
3991 {
3992 init_req(MemNode::Control, c );
3993 init_req(MemNode::Memory , mem);
3994 init_req(MemNode::Address, adr);
3995 init_req(MemNode::ValueIn, val);
3996 init_class_id(Class_LoadStore);
3997 DEBUG_ONLY(_adr_type = at; adr_type();)
3998 }
3999
4000 //------------------------------Value-----------------------------------------
4001 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
4002 // Either input is TOP ==> the result is TOP
4003 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
4004 return Type::TOP;
4005 }
4006 const Type* t = phase->type(in(MemNode::Memory));
4007 if (t == Type::TOP) {
4008 return Type::TOP;
4009 }
4010 t = phase->type(in(MemNode::Address));
4011 if (t == Type::TOP) {
4012 return Type::TOP;
4013 }
4014 t = phase->type(in(MemNode::ValueIn));
4015 if (t == Type::TOP) {
4016 return Type::TOP;
4017 }
4018 return bottom_type();
4019 }
4020
4021 const TypePtr* LoadStoreNode::adr_type() const {
4022 const TypePtr* cross_check = DEBUG_ONLY(_adr_type) NOT_DEBUG(nullptr);
4023 return MemNode::calculate_adr_type(in(MemNode::Address)->bottom_type(), cross_check);
4024 }
4025
4026 uint LoadStoreNode::ideal_reg() const {
4027 return _type->ideal_reg();
4028 }
4029
4030 // This method conservatively checks if the result of a LoadStoreNode is
4031 // used, that is, if it returns true, then it is definitely the case that
4032 // the result of the node is not needed.
4033 // For example, GetAndAdd can be matched into a lock_add instead of a
4034 // lock_xadd if the result of LoadStoreNode::result_not_used() is true
4035 bool LoadStoreNode::result_not_used() const {
4036 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4037 Node *x = fast_out(i);
4038 if (x->Opcode() == Op_SCMemProj) {
4039 continue;
4040 }
4041 if (x->bottom_type() == TypeTuple::MEMBAR &&
4042 !x->is_Call() &&
4043 x->Opcode() != Op_Blackhole) {
4044 continue;
4045 }
4046 return false;
4047 }
4048 return true;
4049 }
4050
4051 MemBarNode* LoadStoreNode::trailing_membar() const {
4052 MemBarNode* trailing = nullptr;
4053 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4054 Node* u = fast_out(i);
4055 if (u->is_MemBar()) {
4056 if (u->as_MemBar()->trailing_load_store()) {
4057 assert(u->Opcode() == Op_MemBarAcquire, "");
4058 assert(trailing == nullptr, "only one");
4059 trailing = u->as_MemBar();
4060 #ifdef ASSERT
4061 Node* leading = trailing->leading_membar();
4062 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
4063 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
4064 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
4065 #endif
4066 } else {
4067 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
4068 }
4069 }
4070 }
4071
4072 return trailing;
4073 }
4074
4075 uint LoadStoreNode::size_of() const { return sizeof(*this); }
4076
4077 //=============================================================================
4078 //----------------------------------LoadStoreConditionalNode--------------------
4079 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) {
4080 init_req(ExpectedIn, ex );
4081 }
4082
4083 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
4084 // Either input is TOP ==> the result is TOP
4085 const Type* t = phase->type(in(ExpectedIn));
4086 if (t == Type::TOP) {
4087 return Type::TOP;
4088 }
4089 return LoadStoreNode::Value(phase);
4090 }
4091
4092 //=============================================================================
4093 //-------------------------------adr_type--------------------------------------
4094 const TypePtr* ClearArrayNode::adr_type() const {
4095 Node *adr = in(3);
4096 if (adr == nullptr) return nullptr; // node is dead
4097 return MemNode::calculate_adr_type(adr->bottom_type());
4098 }
4099
4100 //------------------------------match_edge-------------------------------------
4101 // Do we Match on this edge index or not? Do not match memory
4102 uint ClearArrayNode::match_edge(uint idx) const {
4103 return idx > 1;
4104 }
4105
4106 //------------------------------Identity---------------------------------------
4107 // Clearing a zero length array does nothing
4108 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
4109 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
4110 }
4111
4112 //------------------------------Idealize---------------------------------------
4113 // Clearing a short array is faster with stores
4114 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4115 // Already know this is a large node, do not try to ideal it
4116 if (_is_large) return nullptr;
4117
4118 const int unit = BytesPerLong;
4119 const TypeX* t = phase->type(in(2))->isa_intptr_t();
4120 if (!t) return nullptr;
4121 if (!t->is_con()) return nullptr;
4122 intptr_t raw_count = t->get_con();
4123 intptr_t size = raw_count;
4124 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
4125 // Clearing nothing uses the Identity call.
4126 // Negative clears are possible on dead ClearArrays
4127 // (see jck test stmt114.stmt11402.val).
4128 if (size <= 0 || size % unit != 0) return nullptr;
4129 intptr_t count = size / unit;
4130 // Length too long; communicate this to matchers and assemblers.
4131 // Assemblers are responsible to produce fast hardware clears for it.
4132 if (size > InitArrayShortSize) {
4133 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
4134 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
4135 return nullptr;
4136 }
4137 if (!IdealizeClearArrayNode) return nullptr;
4138 Node *mem = in(1);
4139 if( phase->type(mem)==Type::TOP ) return nullptr;
4140 Node *adr = in(3);
4141 const Type* at = phase->type(adr);
4142 if( at==Type::TOP ) return nullptr;
4143 const TypePtr* atp = at->isa_ptr();
4144 // adjust atp to be the correct array element address type
4145 if (atp == nullptr) atp = TypePtr::BOTTOM;
4146 else atp = atp->add_offset(Type::OffsetBot);
4147 // Get base for derived pointer purposes
4148 if( adr->Opcode() != Op_AddP ) Unimplemented();
4149 Node *base = adr->in(1);
4150
4151 Node *zero = phase->makecon(TypeLong::ZERO);
4152 Node *off = phase->MakeConX(BytesPerLong);
4153 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
4154 count--;
4155 while( count-- ) {
4156 mem = phase->transform(mem);
4157 adr = phase->transform(new AddPNode(base,adr,off));
4158 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
4159 }
4160 return mem;
4161 }
4162
4163 //----------------------------step_through----------------------------------
4164 // Return allocation input memory edge if it is different instance
4165 // or itself if it is the one we are looking for.
4166 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
4167 Node* n = *np;
4168 assert(n->is_ClearArray(), "sanity");
4169 intptr_t offset;
4170 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
4171 // This method is called only before Allocate nodes are expanded
4172 // during macro nodes expansion. Before that ClearArray nodes are
4173 // only generated in PhaseMacroExpand::generate_arraycopy() (before
4174 // Allocate nodes are expanded) which follows allocations.
4175 assert(alloc != nullptr, "should have allocation");
4176 if (alloc->_idx == instance_id) {
4177 // Can not bypass initialization of the instance we are looking for.
4178 return false;
4179 }
4180 // Otherwise skip it.
4181 InitializeNode* init = alloc->initialization();
4182 if (init != nullptr)
4183 *np = init->in(TypeFunc::Memory);
4184 else
4185 *np = alloc->in(TypeFunc::Memory);
4186 return true;
4187 }
4188
4189 Node* ClearArrayNode::make_address(Node* dest, Node* offset, bool raw_base, PhaseGVN* phase) {
4190 Node* base = dest;
4191 if (raw_base) {
4192 // May be called as part of the initialization of a just allocated object
4193 base = phase->C->top();
4194 }
4195 return phase->transform(new AddPNode(base, dest, offset));
4196 }
4197
4198 //----------------------------clear_memory-------------------------------------
4199 // Generate code to initialize object storage to zero.
4200 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4201 intptr_t start_offset,
4202 Node* end_offset,
4203 bool raw_base,
4204 PhaseGVN* phase) {
4205 intptr_t offset = start_offset;
4206
4207 int unit = BytesPerLong;
4208 if ((offset % unit) != 0) {
4209 Node* adr = make_address(dest, phase->MakeConX(offset), raw_base, phase);
4210 const TypePtr* atp = TypeRawPtr::BOTTOM;
4211 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4212 mem = phase->transform(mem);
4213 offset += BytesPerInt;
4214 }
4215 assert((offset % unit) == 0, "");
4216
4217 // Initialize the remaining stuff, if any, with a ClearArray.
4218 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, raw_base, phase);
4219 }
4220
4221 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4222 Node* start_offset,
4223 Node* end_offset,
4224 bool raw_base,
4225 PhaseGVN* phase) {
4226 if (start_offset == end_offset) {
4227 // nothing to do
4228 return mem;
4229 }
4230
4231 int unit = BytesPerLong;
4232 Node* zbase = start_offset;
4233 Node* zend = end_offset;
4234
4235 // Scale to the unit required by the CPU:
4236 if (!Matcher::init_array_count_is_in_bytes) {
4237 Node* shift = phase->intcon(exact_log2(unit));
4238 zbase = phase->transform(new URShiftXNode(zbase, shift) );
4239 zend = phase->transform(new URShiftXNode(zend, shift) );
4240 }
4241
4242 // Bulk clear double-words
4243 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
4244 Node* adr = make_address(dest, start_offset, raw_base, phase);
4245 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
4246 return phase->transform(mem);
4247 }
4248
4249 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4250 intptr_t start_offset,
4251 intptr_t end_offset,
4252 bool raw_base,
4253 PhaseGVN* phase) {
4254 if (start_offset == end_offset) {
4255 // nothing to do
4256 return mem;
4257 }
4258
4259 assert((end_offset % BytesPerInt) == 0, "odd end offset");
4260 intptr_t done_offset = end_offset;
4261 if ((done_offset % BytesPerLong) != 0) {
4262 done_offset -= BytesPerInt;
4263 }
4264 if (done_offset > start_offset) {
4265 mem = clear_memory(ctl, mem, dest,
4266 start_offset, phase->MakeConX(done_offset), raw_base, phase);
4267 }
4268 if (done_offset < end_offset) { // emit the final 32-bit store
4269 Node* adr = make_address(dest, phase->MakeConX(done_offset), raw_base, phase);
4270 const TypePtr* atp = TypeRawPtr::BOTTOM;
4271 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4272 mem = phase->transform(mem);
4273 done_offset += BytesPerInt;
4274 }
4275 assert(done_offset == end_offset, "");
4276 return mem;
4277 }
4278
4279 //=============================================================================
4280 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
4281 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
4282 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
4283 #ifdef ASSERT
4284 , _pair_idx(0)
4285 #endif
4286 {
4287 init_class_id(Class_MemBar);
4288 Node* top = C->top();
4289 init_req(TypeFunc::I_O,top);
4290 init_req(TypeFunc::FramePtr,top);
4291 init_req(TypeFunc::ReturnAdr,top);
4292 if (precedent != nullptr)
4293 init_req(TypeFunc::Parms, precedent);
4294 }
4295
4296 //------------------------------cmp--------------------------------------------
4297 uint MemBarNode::hash() const { return NO_HASH; }
4298 bool MemBarNode::cmp( const Node &n ) const {
4299 return (&n == this); // Always fail except on self
4300 }
4301
4302 //------------------------------make-------------------------------------------
4303 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
4304 switch (opcode) {
4305 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
4306 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
4307 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
4308 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
4309 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
4310 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn);
4311 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
4312 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
4313 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
4314 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
4315 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
4316 case Op_Initialize: return new InitializeNode(C, atp, pn);
4317 default: ShouldNotReachHere(); return nullptr;
4318 }
4319 }
4320
4321 void MemBarNode::remove(PhaseIterGVN *igvn) {
4322 assert(outcnt() > 0 && (outcnt() <= 2 || Opcode() == Op_Initialize), "Only one or two out edges allowed");
4323 if (trailing_store() || trailing_load_store()) {
4324 MemBarNode* leading = leading_membar();
4325 if (leading != nullptr) {
4326 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
4327 leading->remove(igvn);
4328 }
4329 }
4330 if (proj_out_or_null(TypeFunc::Control) != nullptr) {
4331 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
4332 }
4333 if (is_Initialize()) {
4334 as_Initialize()->replace_mem_projs_by(in(TypeFunc::Memory), igvn);
4335 } else {
4336 if (proj_out_or_null(TypeFunc::Memory) != nullptr) {
4337 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
4338 }
4339 }
4340 }
4341
4342 //------------------------------Ideal------------------------------------------
4343 // Return a node which is more "ideal" than the current node. Strip out
4344 // control copies
4345 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4346 if (remove_dead_region(phase, can_reshape)) return this;
4347 // Don't bother trying to transform a dead node
4348 if (in(0) && in(0)->is_top()) {
4349 return nullptr;
4350 }
4351
4352 bool progress = false;
4353 // Eliminate volatile MemBars for scalar replaced objects.
4354 if (can_reshape && req() == (Precedent+1)) {
4355 bool eliminate = false;
4356 int opc = Opcode();
4357 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
4358 // Volatile field loads and stores.
4359 Node* my_mem = in(MemBarNode::Precedent);
4360 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
4361 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
4362 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
4363 // replace this Precedent (decodeN) with the Load instead.
4364 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
4365 Node* load_node = my_mem->in(1);
4366 set_req(MemBarNode::Precedent, load_node);
4367 phase->is_IterGVN()->_worklist.push(my_mem);
4368 my_mem = load_node;
4369 } else {
4370 assert(my_mem->unique_out() == this, "sanity");
4371 assert(!trailing_load_store(), "load store node can't be eliminated");
4372 del_req(Precedent);
4373 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
4374 my_mem = nullptr;
4375 }
4376 progress = true;
4377 }
4378 if (my_mem != nullptr && my_mem->is_Mem()) {
4379 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
4380 // Check for scalar replaced object reference.
4381 if( t_oop != nullptr && t_oop->is_known_instance_field() &&
4382 t_oop->offset() != Type::OffsetBot &&
4383 t_oop->offset() != Type::OffsetTop) {
4384 eliminate = true;
4385 }
4386 }
4387 } else if (opc == Op_MemBarRelease || (UseStoreStoreForCtor && opc == Op_MemBarStoreStore)) {
4388 // Final field stores.
4389 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent));
4390 if ((alloc != nullptr) && alloc->is_Allocate() &&
4391 alloc->as_Allocate()->does_not_escape_thread()) {
4392 // The allocated object does not escape.
4393 eliminate = true;
4394 }
4395 }
4396 if (eliminate) {
4397 // Replace MemBar projections by its inputs.
4398 PhaseIterGVN* igvn = phase->is_IterGVN();
4399 remove(igvn);
4400 // Must return either the original node (now dead) or a new node
4401 // (Do not return a top here, since that would break the uniqueness of top.)
4402 return new ConINode(TypeInt::ZERO);
4403 }
4404 }
4405 return progress ? this : nullptr;
4406 }
4407
4408 //------------------------------Value------------------------------------------
4409 const Type* MemBarNode::Value(PhaseGVN* phase) const {
4410 if( !in(0) ) return Type::TOP;
4411 if( phase->type(in(0)) == Type::TOP )
4412 return Type::TOP;
4413 return TypeTuple::MEMBAR;
4414 }
4415
4416 //------------------------------match------------------------------------------
4417 // Construct projections for memory.
4418 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
4419 switch (proj->_con) {
4420 case TypeFunc::Control:
4421 case TypeFunc::Memory:
4422 return new MachProjNode(this, proj->_con, RegMask::EMPTY, MachProjNode::unmatched_proj);
4423 }
4424 ShouldNotReachHere();
4425 return nullptr;
4426 }
4427
4428 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4429 trailing->_kind = TrailingStore;
4430 leading->_kind = LeadingStore;
4431 #ifdef ASSERT
4432 trailing->_pair_idx = leading->_idx;
4433 leading->_pair_idx = leading->_idx;
4434 #endif
4435 }
4436
4437 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4438 trailing->_kind = TrailingLoadStore;
4439 leading->_kind = LeadingLoadStore;
4440 #ifdef ASSERT
4441 trailing->_pair_idx = leading->_idx;
4442 leading->_pair_idx = leading->_idx;
4443 #endif
4444 }
4445
4446 MemBarNode* MemBarNode::trailing_membar() const {
4447 ResourceMark rm;
4448 Node* trailing = (Node*)this;
4449 VectorSet seen;
4450 Node_Stack multis(0);
4451 do {
4452 Node* c = trailing;
4453 uint i = 0;
4454 do {
4455 trailing = nullptr;
4456 for (; i < c->outcnt(); i++) {
4457 Node* next = c->raw_out(i);
4458 if (next != c && next->is_CFG()) {
4459 if (c->is_MultiBranch()) {
4460 if (multis.node() == c) {
4461 multis.set_index(i+1);
4462 } else {
4463 multis.push(c, i+1);
4464 }
4465 }
4466 trailing = next;
4467 break;
4468 }
4469 }
4470 if (trailing != nullptr && !seen.test_set(trailing->_idx)) {
4471 break;
4472 }
4473 while (multis.size() > 0) {
4474 c = multis.node();
4475 i = multis.index();
4476 if (i < c->req()) {
4477 break;
4478 }
4479 multis.pop();
4480 }
4481 } while (multis.size() > 0);
4482 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
4483
4484 MemBarNode* mb = trailing->as_MemBar();
4485 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
4486 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
4487 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
4488 return mb;
4489 }
4490
4491 MemBarNode* MemBarNode::leading_membar() const {
4492 ResourceMark rm;
4493 VectorSet seen;
4494 Node_Stack regions(0);
4495 Node* leading = in(0);
4496 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
4497 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) {
4498 leading = nullptr;
4499 while (regions.size() > 0 && leading == nullptr) {
4500 Node* r = regions.node();
4501 uint i = regions.index();
4502 if (i < r->req()) {
4503 leading = r->in(i);
4504 regions.set_index(i+1);
4505 } else {
4506 regions.pop();
4507 }
4508 }
4509 if (leading == nullptr) {
4510 assert(regions.size() == 0, "all paths should have been tried");
4511 return nullptr;
4512 }
4513 }
4514 if (leading->is_Region()) {
4515 regions.push(leading, 2);
4516 leading = leading->in(1);
4517 } else {
4518 leading = leading->in(0);
4519 }
4520 }
4521 #ifdef ASSERT
4522 Unique_Node_List wq;
4523 wq.push((Node*)this);
4524 uint found = 0;
4525 for (uint i = 0; i < wq.size(); i++) {
4526 Node* n = wq.at(i);
4527 if (n->is_Region()) {
4528 for (uint j = 1; j < n->req(); j++) {
4529 Node* in = n->in(j);
4530 if (in != nullptr && !in->is_top()) {
4531 wq.push(in);
4532 }
4533 }
4534 } else {
4535 if (n->is_MemBar() && n->as_MemBar()->leading()) {
4536 assert(n == leading, "consistency check failed");
4537 found++;
4538 } else {
4539 Node* in = n->in(0);
4540 if (in != nullptr && !in->is_top()) {
4541 wq.push(in);
4542 }
4543 }
4544 }
4545 }
4546 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed");
4547 #endif
4548 if (leading == nullptr) {
4549 return nullptr;
4550 }
4551 MemBarNode* mb = leading->as_MemBar();
4552 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
4553 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
4554 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
4555 return mb;
4556 }
4557
4558
4559 //===========================InitializeNode====================================
4560 // SUMMARY:
4561 // This node acts as a memory barrier on raw memory, after some raw stores.
4562 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
4563 // The Initialize can 'capture' suitably constrained stores as raw inits.
4564 // It can coalesce related raw stores into larger units (called 'tiles').
4565 // It can avoid zeroing new storage for memory units which have raw inits.
4566 // At macro-expansion, it is marked 'complete', and does not optimize further.
4567 //
4568 // EXAMPLE:
4569 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
4570 // ctl = incoming control; mem* = incoming memory
4571 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
4572 // First allocate uninitialized memory and fill in the header:
4573 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
4574 // ctl := alloc.Control; mem* := alloc.Memory*
4575 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
4576 // Then initialize to zero the non-header parts of the raw memory block:
4577 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
4578 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
4579 // After the initialize node executes, the object is ready for service:
4580 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
4581 // Suppose its body is immediately initialized as {1,2}:
4582 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4583 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4584 // mem.SLICE(#short[*]) := store2
4585 //
4586 // DETAILS:
4587 // An InitializeNode collects and isolates object initialization after
4588 // an AllocateNode and before the next possible safepoint. As a
4589 // memory barrier (MemBarNode), it keeps critical stores from drifting
4590 // down past any safepoint or any publication of the allocation.
4591 // Before this barrier, a newly-allocated object may have uninitialized bits.
4592 // After this barrier, it may be treated as a real oop, and GC is allowed.
4593 //
4594 // The semantics of the InitializeNode include an implicit zeroing of
4595 // the new object from object header to the end of the object.
4596 // (The object header and end are determined by the AllocateNode.)
4597 //
4598 // Certain stores may be added as direct inputs to the InitializeNode.
4599 // These stores must update raw memory, and they must be to addresses
4600 // derived from the raw address produced by AllocateNode, and with
4601 // a constant offset. They must be ordered by increasing offset.
4602 // The first one is at in(RawStores), the last at in(req()-1).
4603 // Unlike most memory operations, they are not linked in a chain,
4604 // but are displayed in parallel as users of the rawmem output of
4605 // the allocation.
4606 //
4607 // (See comments in InitializeNode::capture_store, which continue
4608 // the example given above.)
4609 //
4610 // When the associated Allocate is macro-expanded, the InitializeNode
4611 // may be rewritten to optimize collected stores. A ClearArrayNode
4612 // may also be created at that point to represent any required zeroing.
4613 // The InitializeNode is then marked 'complete', prohibiting further
4614 // capturing of nearby memory operations.
4615 //
4616 // During macro-expansion, all captured initializations which store
4617 // constant values of 32 bits or smaller are coalesced (if advantageous)
4618 // into larger 'tiles' 32 or 64 bits. This allows an object to be
4619 // initialized in fewer memory operations. Memory words which are
4620 // covered by neither tiles nor non-constant stores are pre-zeroed
4621 // by explicit stores of zero. (The code shape happens to do all
4622 // zeroing first, then all other stores, with both sequences occurring
4623 // in order of ascending offsets.)
4624 //
4625 // Alternatively, code may be inserted between an AllocateNode and its
4626 // InitializeNode, to perform arbitrary initialization of the new object.
4627 // E.g., the object copying intrinsics insert complex data transfers here.
4628 // The initialization must then be marked as 'complete' disable the
4629 // built-in zeroing semantics and the collection of initializing stores.
4630 //
4631 // While an InitializeNode is incomplete, reads from the memory state
4632 // produced by it are optimizable if they match the control edge and
4633 // new oop address associated with the allocation/initialization.
4634 // They return a stored value (if the offset matches) or else zero.
4635 // A write to the memory state, if it matches control and address,
4636 // and if it is to a constant offset, may be 'captured' by the
4637 // InitializeNode. It is cloned as a raw memory operation and rewired
4638 // inside the initialization, to the raw oop produced by the allocation.
4639 // Operations on addresses which are provably distinct (e.g., to
4640 // other AllocateNodes) are allowed to bypass the initialization.
4641 //
4642 // The effect of all this is to consolidate object initialization
4643 // (both arrays and non-arrays, both piecewise and bulk) into a
4644 // single location, where it can be optimized as a unit.
4645 //
4646 // Only stores with an offset less than TrackedInitializationLimit words
4647 // will be considered for capture by an InitializeNode. This puts a
4648 // reasonable limit on the complexity of optimized initializations.
4649
4650 //---------------------------InitializeNode------------------------------------
4651 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
4652 : MemBarNode(C, adr_type, rawoop),
4653 _is_complete(Incomplete), _does_not_escape(false)
4654 {
4655 init_class_id(Class_Initialize);
4656
4657 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
4658 assert(in(RawAddress) == rawoop, "proper init");
4659 // Note: allocation() can be null, for secondary initialization barriers
4660 }
4661
4662 // Since this node is not matched, it will be processed by the
4663 // register allocator. Declare that there are no constraints
4664 // on the allocation of the RawAddress edge.
4665 const RegMask &InitializeNode::in_RegMask(uint idx) const {
4666 // This edge should be set to top, by the set_complete. But be conservative.
4667 if (idx == InitializeNode::RawAddress)
4668 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
4669 return RegMask::EMPTY;
4670 }
4671
4672 Node* InitializeNode::memory(uint alias_idx) {
4673 Node* mem = in(Memory);
4674 if (mem->is_MergeMem()) {
4675 return mem->as_MergeMem()->memory_at(alias_idx);
4676 } else {
4677 // incoming raw memory is not split
4678 return mem;
4679 }
4680 }
4681
4682 bool InitializeNode::is_non_zero() {
4683 if (is_complete()) return false;
4684 remove_extra_zeroes();
4685 return (req() > RawStores);
4686 }
4687
4688 void InitializeNode::set_complete(PhaseGVN* phase) {
4689 assert(!is_complete(), "caller responsibility");
4690 _is_complete = Complete;
4691
4692 // After this node is complete, it contains a bunch of
4693 // raw-memory initializations. There is no need for
4694 // it to have anything to do with non-raw memory effects.
4695 // Therefore, tell all non-raw users to re-optimize themselves,
4696 // after skipping the memory effects of this initialization.
4697 PhaseIterGVN* igvn = phase->is_IterGVN();
4698 if (igvn) igvn->add_users_to_worklist(this);
4699 }
4700
4701 // convenience function
4702 // return false if the init contains any stores already
4703 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
4704 InitializeNode* init = initialization();
4705 if (init == nullptr || init->is_complete()) return false;
4706 init->remove_extra_zeroes();
4707 // for now, if this allocation has already collected any inits, bail:
4708 if (init->is_non_zero()) return false;
4709 init->set_complete(phase);
4710 return true;
4711 }
4712
4713 void InitializeNode::remove_extra_zeroes() {
4714 if (req() == RawStores) return;
4715 Node* zmem = zero_memory();
4716 uint fill = RawStores;
4717 for (uint i = fill; i < req(); i++) {
4718 Node* n = in(i);
4719 if (n->is_top() || n == zmem) continue; // skip
4720 if (fill < i) set_req(fill, n); // compact
4721 ++fill;
4722 }
4723 // delete any empty spaces created:
4724 while (fill < req()) {
4725 del_req(fill);
4726 }
4727 }
4728
4729 // Helper for remembering which stores go with which offsets.
4730 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) {
4731 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
4732 intptr_t offset = -1;
4733 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
4734 phase, offset);
4735 if (base == nullptr) return -1; // something is dead,
4736 if (offset < 0) return -1; // dead, dead
4737 return offset;
4738 }
4739
4740 // Helper for proving that an initialization expression is
4741 // "simple enough" to be folded into an object initialization.
4742 // Attempts to prove that a store's initial value 'n' can be captured
4743 // within the initialization without creating a vicious cycle, such as:
4744 // { Foo p = new Foo(); p.next = p; }
4745 // True for constants and parameters and small combinations thereof.
4746 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
4747 ResourceMark rm;
4748 Unique_Node_List worklist;
4749 worklist.push(value);
4750
4751 uint complexity_limit = 20;
4752 for (uint j = 0; j < worklist.size(); j++) {
4753 if (j >= complexity_limit) {
4754 return false; // Bail out if processed too many nodes
4755 }
4756
4757 Node* n = worklist.at(j);
4758 if (n == nullptr) continue; // (can this really happen?)
4759 if (n->is_Proj()) n = n->in(0);
4760 if (n == this) return false; // found a cycle
4761 if (n->is_Con()) continue;
4762 if (n->is_Start()) continue; // params, etc., are OK
4763 if (n->is_Root()) continue; // even better
4764
4765 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
4766 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
4767 continue;
4768 }
4769
4770 Node* ctl = n->in(0);
4771 if (ctl != nullptr && !ctl->is_top()) {
4772 if (ctl->is_Proj()) ctl = ctl->in(0);
4773 if (ctl == this) return false;
4774
4775 // If we already know that the enclosing memory op is pinned right after
4776 // the init, then any control flow that the store has picked up
4777 // must have preceded the init, or else be equal to the init.
4778 // Even after loop optimizations (which might change control edges)
4779 // a store is never pinned *before* the availability of its inputs.
4780 if (!MemNode::all_controls_dominate(n, this)) {
4781 return false; // failed to prove a good control
4782 }
4783 }
4784
4785 // Check data edges for possible dependencies on 'this'.
4786 for (uint i = 1; i < n->req(); i++) {
4787 Node* m = n->in(i);
4788 if (m == nullptr || m == n || m->is_top()) continue;
4789
4790 // Only process data inputs once
4791 worklist.push(m);
4792 }
4793 }
4794
4795 return true;
4796 }
4797
4798 // Here are all the checks a Store must pass before it can be moved into
4799 // an initialization. Returns zero if a check fails.
4800 // On success, returns the (constant) offset to which the store applies,
4801 // within the initialized memory.
4802 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
4803 const int FAIL = 0;
4804 if (st->req() != MemNode::ValueIn + 1)
4805 return FAIL; // an inscrutable StoreNode (card mark?)
4806 Node* ctl = st->in(MemNode::Control);
4807 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this))
4808 return FAIL; // must be unconditional after the initialization
4809 Node* mem = st->in(MemNode::Memory);
4810 if (!(mem->is_Proj() && mem->in(0) == this))
4811 return FAIL; // must not be preceded by other stores
4812 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4813 if ((st->Opcode() == Op_StoreP || st->Opcode() == Op_StoreN) &&
4814 !bs->can_initialize_object(st)) {
4815 return FAIL;
4816 }
4817 Node* adr = st->in(MemNode::Address);
4818 intptr_t offset;
4819 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
4820 if (alloc == nullptr)
4821 return FAIL; // inscrutable address
4822 if (alloc != allocation())
4823 return FAIL; // wrong allocation! (store needs to float up)
4824 int size_in_bytes = st->memory_size();
4825 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
4826 return FAIL; // mismatched access
4827 }
4828 Node* val = st->in(MemNode::ValueIn);
4829
4830 if (!detect_init_independence(val, phase))
4831 return FAIL; // stored value must be 'simple enough'
4832
4833 // The Store can be captured only if nothing after the allocation
4834 // and before the Store is using the memory location that the store
4835 // overwrites.
4836 bool failed = false;
4837 // If is_complete_with_arraycopy() is true the shape of the graph is
4838 // well defined and is safe so no need for extra checks.
4839 if (!is_complete_with_arraycopy()) {
4840 // We are going to look at each use of the memory state following
4841 // the allocation to make sure nothing reads the memory that the
4842 // Store writes.
4843 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
4844 int alias_idx = phase->C->get_alias_index(t_adr);
4845 ResourceMark rm;
4846 Unique_Node_List mems;
4847 mems.push(mem);
4848 Node* unique_merge = nullptr;
4849 for (uint next = 0; next < mems.size(); ++next) {
4850 Node *m = mems.at(next);
4851 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
4852 Node *n = m->fast_out(j);
4853 if (n->outcnt() == 0) {
4854 continue;
4855 }
4856 if (n == st) {
4857 continue;
4858 } else if (n->in(0) != nullptr && n->in(0) != ctl) {
4859 // If the control of this use is different from the control
4860 // of the Store which is right after the InitializeNode then
4861 // this node cannot be between the InitializeNode and the
4862 // Store.
4863 continue;
4864 } else if (n->is_MergeMem()) {
4865 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
4866 // We can hit a MergeMemNode (that will likely go away
4867 // later) that is a direct use of the memory state
4868 // following the InitializeNode on the same slice as the
4869 // store node that we'd like to capture. We need to check
4870 // the uses of the MergeMemNode.
4871 mems.push(n);
4872 }
4873 } else if (n->is_Mem()) {
4874 Node* other_adr = n->in(MemNode::Address);
4875 if (other_adr == adr) {
4876 failed = true;
4877 break;
4878 } else {
4879 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
4880 if (other_t_adr != nullptr) {
4881 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
4882 if (other_alias_idx == alias_idx) {
4883 // A load from the same memory slice as the store right
4884 // after the InitializeNode. We check the control of the
4885 // object/array that is loaded from. If it's the same as
4886 // the store control then we cannot capture the store.
4887 assert(!n->is_Store(), "2 stores to same slice on same control?");
4888 Node* base = other_adr;
4889 assert(base->is_AddP(), "should be addp but is %s", base->Name());
4890 base = base->in(AddPNode::Base);
4891 if (base != nullptr) {
4892 base = base->uncast();
4893 if (base->is_Proj() && base->in(0) == alloc) {
4894 failed = true;
4895 break;
4896 }
4897 }
4898 }
4899 }
4900 }
4901 } else {
4902 failed = true;
4903 break;
4904 }
4905 }
4906 }
4907 }
4908 if (failed) {
4909 if (!can_reshape) {
4910 // We decided we couldn't capture the store during parsing. We
4911 // should try again during the next IGVN once the graph is
4912 // cleaner.
4913 phase->C->record_for_igvn(st);
4914 }
4915 return FAIL;
4916 }
4917
4918 return offset; // success
4919 }
4920
4921 // Find the captured store in(i) which corresponds to the range
4922 // [start..start+size) in the initialized object.
4923 // If there is one, return its index i. If there isn't, return the
4924 // negative of the index where it should be inserted.
4925 // Return 0 if the queried range overlaps an initialization boundary
4926 // or if dead code is encountered.
4927 // If size_in_bytes is zero, do not bother with overlap checks.
4928 int InitializeNode::captured_store_insertion_point(intptr_t start,
4929 int size_in_bytes,
4930 PhaseValues* phase) {
4931 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
4932
4933 if (is_complete())
4934 return FAIL; // arraycopy got here first; punt
4935
4936 assert(allocation() != nullptr, "must be present");
4937
4938 // no negatives, no header fields:
4939 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
4940
4941 // after a certain size, we bail out on tracking all the stores:
4942 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4943 if (start >= ti_limit) return FAIL;
4944
4945 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
4946 if (i >= limit) return -(int)i; // not found; here is where to put it
4947
4948 Node* st = in(i);
4949 intptr_t st_off = get_store_offset(st, phase);
4950 if (st_off < 0) {
4951 if (st != zero_memory()) {
4952 return FAIL; // bail out if there is dead garbage
4953 }
4954 } else if (st_off > start) {
4955 // ...we are done, since stores are ordered
4956 if (st_off < start + size_in_bytes) {
4957 return FAIL; // the next store overlaps
4958 }
4959 return -(int)i; // not found; here is where to put it
4960 } else if (st_off < start) {
4961 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
4962 if (size_in_bytes != 0 &&
4963 start < st_off + MAX_STORE &&
4964 start < st_off + st->as_Store()->memory_size()) {
4965 return FAIL; // the previous store overlaps
4966 }
4967 } else {
4968 if (size_in_bytes != 0 &&
4969 st->as_Store()->memory_size() != size_in_bytes) {
4970 return FAIL; // mismatched store size
4971 }
4972 return i;
4973 }
4974
4975 ++i;
4976 }
4977 }
4978
4979 // Look for a captured store which initializes at the offset 'start'
4980 // with the given size. If there is no such store, and no other
4981 // initialization interferes, then return zero_memory (the memory
4982 // projection of the AllocateNode).
4983 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
4984 PhaseValues* phase) {
4985 assert(stores_are_sane(phase), "");
4986 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4987 if (i == 0) {
4988 return nullptr; // something is dead
4989 } else if (i < 0) {
4990 return zero_memory(); // just primordial zero bits here
4991 } else {
4992 Node* st = in(i); // here is the store at this position
4993 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
4994 return st;
4995 }
4996 }
4997
4998 // Create, as a raw pointer, an address within my new object at 'offset'.
4999 Node* InitializeNode::make_raw_address(intptr_t offset,
5000 PhaseGVN* phase) {
5001 Node* addr = in(RawAddress);
5002 if (offset != 0) {
5003 Compile* C = phase->C;
5004 addr = phase->transform( new AddPNode(C->top(), addr,
5005 phase->MakeConX(offset)) );
5006 }
5007 return addr;
5008 }
5009
5010 // Clone the given store, converting it into a raw store
5011 // initializing a field or element of my new object.
5012 // Caller is responsible for retiring the original store,
5013 // with subsume_node or the like.
5014 //
5015 // From the example above InitializeNode::InitializeNode,
5016 // here are the old stores to be captured:
5017 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
5018 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
5019 //
5020 // Here is the changed code; note the extra edges on init:
5021 // alloc = (Allocate ...)
5022 // rawoop = alloc.RawAddress
5023 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
5024 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
5025 // init = (Initialize alloc.Control alloc.Memory rawoop
5026 // rawstore1 rawstore2)
5027 //
5028 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
5029 PhaseGVN* phase, bool can_reshape) {
5030 assert(stores_are_sane(phase), "");
5031
5032 if (start < 0) return nullptr;
5033 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
5034
5035 Compile* C = phase->C;
5036 int size_in_bytes = st->memory_size();
5037 int i = captured_store_insertion_point(start, size_in_bytes, phase);
5038 if (i == 0) return nullptr; // bail out
5039 Node* prev_mem = nullptr; // raw memory for the captured store
5040 if (i > 0) {
5041 prev_mem = in(i); // there is a pre-existing store under this one
5042 set_req(i, C->top()); // temporarily disconnect it
5043 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
5044 } else {
5045 i = -i; // no pre-existing store
5046 prev_mem = zero_memory(); // a slice of the newly allocated object
5047 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
5048 set_req(--i, C->top()); // reuse this edge; it has been folded away
5049 else
5050 ins_req(i, C->top()); // build a new edge
5051 }
5052 Node* new_st = st->clone();
5053 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
5054 new_st->set_req(MemNode::Control, in(Control));
5055 new_st->set_req(MemNode::Memory, prev_mem);
5056 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
5057 bs->eliminate_gc_barrier_data(new_st);
5058 new_st = phase->transform(new_st);
5059
5060 // At this point, new_st might have swallowed a pre-existing store
5061 // at the same offset, or perhaps new_st might have disappeared,
5062 // if it redundantly stored the same value (or zero to fresh memory).
5063
5064 // In any case, wire it in:
5065 PhaseIterGVN* igvn = phase->is_IterGVN();
5066 if (igvn) {
5067 igvn->rehash_node_delayed(this);
5068 }
5069 set_req(i, new_st);
5070
5071 // The caller may now kill the old guy.
5072 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
5073 assert(check_st == new_st || check_st == nullptr, "must be findable");
5074 assert(!is_complete(), "");
5075 return new_st;
5076 }
5077
5078 static bool store_constant(jlong* tiles, int num_tiles,
5079 intptr_t st_off, int st_size,
5080 jlong con) {
5081 if ((st_off & (st_size-1)) != 0)
5082 return false; // strange store offset (assume size==2**N)
5083 address addr = (address)tiles + st_off;
5084 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
5085 switch (st_size) {
5086 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
5087 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
5088 case sizeof(jint): *(jint*) addr = (jint) con; break;
5089 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
5090 default: return false; // strange store size (detect size!=2**N here)
5091 }
5092 return true; // return success to caller
5093 }
5094
5095 // Coalesce subword constants into int constants and possibly
5096 // into long constants. The goal, if the CPU permits,
5097 // is to initialize the object with a small number of 64-bit tiles.
5098 // Also, convert floating-point constants to bit patterns.
5099 // Non-constants are not relevant to this pass.
5100 //
5101 // In terms of the running example on InitializeNode::InitializeNode
5102 // and InitializeNode::capture_store, here is the transformation
5103 // of rawstore1 and rawstore2 into rawstore12:
5104 // alloc = (Allocate ...)
5105 // rawoop = alloc.RawAddress
5106 // tile12 = 0x00010002
5107 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
5108 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
5109 //
5110 void
5111 InitializeNode::coalesce_subword_stores(intptr_t header_size,
5112 Node* size_in_bytes,
5113 PhaseGVN* phase) {
5114 Compile* C = phase->C;
5115
5116 assert(stores_are_sane(phase), "");
5117 // Note: After this pass, they are not completely sane,
5118 // since there may be some overlaps.
5119
5120 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
5121
5122 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
5123 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
5124 size_limit = MIN2(size_limit, ti_limit);
5125 size_limit = align_up(size_limit, BytesPerLong);
5126 int num_tiles = size_limit / BytesPerLong;
5127
5128 // allocate space for the tile map:
5129 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
5130 jlong tiles_buf[small_len];
5131 Node* nodes_buf[small_len];
5132 jlong inits_buf[small_len];
5133 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
5134 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5135 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
5136 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
5137 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
5138 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5139 // tiles: exact bitwise model of all primitive constants
5140 // nodes: last constant-storing node subsumed into the tiles model
5141 // inits: which bytes (in each tile) are touched by any initializations
5142
5143 //// Pass A: Fill in the tile model with any relevant stores.
5144
5145 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
5146 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
5147 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
5148 Node* zmem = zero_memory(); // initially zero memory state
5149 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5150 Node* st = in(i);
5151 intptr_t st_off = get_store_offset(st, phase);
5152
5153 // Figure out the store's offset and constant value:
5154 if (st_off < header_size) continue; //skip (ignore header)
5155 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
5156 int st_size = st->as_Store()->memory_size();
5157 if (st_off + st_size > size_limit) break;
5158
5159 // Record which bytes are touched, whether by constant or not.
5160 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
5161 continue; // skip (strange store size)
5162
5163 const Type* val = phase->type(st->in(MemNode::ValueIn));
5164 if (!val->singleton()) continue; //skip (non-con store)
5165 BasicType type = val->basic_type();
5166
5167 jlong con = 0;
5168 switch (type) {
5169 case T_INT: con = val->is_int()->get_con(); break;
5170 case T_LONG: con = val->is_long()->get_con(); break;
5171 case T_FLOAT: con = jint_cast(val->getf()); break;
5172 case T_DOUBLE: con = jlong_cast(val->getd()); break;
5173 default: continue; //skip (odd store type)
5174 }
5175
5176 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
5177 st->Opcode() == Op_StoreL) {
5178 continue; // This StoreL is already optimal.
5179 }
5180
5181 // Store down the constant.
5182 store_constant(tiles, num_tiles, st_off, st_size, con);
5183
5184 intptr_t j = st_off >> LogBytesPerLong;
5185
5186 if (type == T_INT && st_size == BytesPerInt
5187 && (st_off & BytesPerInt) == BytesPerInt) {
5188 jlong lcon = tiles[j];
5189 if (!Matcher::isSimpleConstant64(lcon) &&
5190 st->Opcode() == Op_StoreI) {
5191 // This StoreI is already optimal by itself.
5192 jint* intcon = (jint*) &tiles[j];
5193 intcon[1] = 0; // undo the store_constant()
5194
5195 // If the previous store is also optimal by itself, back up and
5196 // undo the action of the previous loop iteration... if we can.
5197 // But if we can't, just let the previous half take care of itself.
5198 st = nodes[j];
5199 st_off -= BytesPerInt;
5200 con = intcon[0];
5201 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) {
5202 assert(st_off >= header_size, "still ignoring header");
5203 assert(get_store_offset(st, phase) == st_off, "must be");
5204 assert(in(i-1) == zmem, "must be");
5205 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
5206 assert(con == tcon->is_int()->get_con(), "must be");
5207 // Undo the effects of the previous loop trip, which swallowed st:
5208 intcon[0] = 0; // undo store_constant()
5209 set_req(i-1, st); // undo set_req(i, zmem)
5210 nodes[j] = nullptr; // undo nodes[j] = st
5211 --old_subword; // undo ++old_subword
5212 }
5213 continue; // This StoreI is already optimal.
5214 }
5215 }
5216
5217 // This store is not needed.
5218 set_req(i, zmem);
5219 nodes[j] = st; // record for the moment
5220 if (st_size < BytesPerLong) // something has changed
5221 ++old_subword; // includes int/float, but who's counting...
5222 else ++old_long;
5223 }
5224
5225 if ((old_subword + old_long) == 0)
5226 return; // nothing more to do
5227
5228 //// Pass B: Convert any non-zero tiles into optimal constant stores.
5229 // Be sure to insert them before overlapping non-constant stores.
5230 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
5231 for (int j = 0; j < num_tiles; j++) {
5232 jlong con = tiles[j];
5233 jlong init = inits[j];
5234 if (con == 0) continue;
5235 jint con0, con1; // split the constant, address-wise
5236 jint init0, init1; // split the init map, address-wise
5237 { union { jlong con; jint intcon[2]; } u;
5238 u.con = con;
5239 con0 = u.intcon[0];
5240 con1 = u.intcon[1];
5241 u.con = init;
5242 init0 = u.intcon[0];
5243 init1 = u.intcon[1];
5244 }
5245
5246 Node* old = nodes[j];
5247 assert(old != nullptr, "need the prior store");
5248 intptr_t offset = (j * BytesPerLong);
5249
5250 bool split = !Matcher::isSimpleConstant64(con);
5251
5252 if (offset < header_size) {
5253 assert(offset + BytesPerInt >= header_size, "second int counts");
5254 assert(*(jint*)&tiles[j] == 0, "junk in header");
5255 split = true; // only the second word counts
5256 // Example: int a[] = { 42 ... }
5257 } else if (con0 == 0 && init0 == -1) {
5258 split = true; // first word is covered by full inits
5259 // Example: int a[] = { ... foo(), 42 ... }
5260 } else if (con1 == 0 && init1 == -1) {
5261 split = true; // second word is covered by full inits
5262 // Example: int a[] = { ... 42, foo() ... }
5263 }
5264
5265 // Here's a case where init0 is neither 0 nor -1:
5266 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
5267 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
5268 // In this case the tile is not split; it is (jlong)42.
5269 // The big tile is stored down, and then the foo() value is inserted.
5270 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
5271
5272 Node* ctl = old->in(MemNode::Control);
5273 Node* adr = make_raw_address(offset, phase);
5274 const TypePtr* atp = TypeRawPtr::BOTTOM;
5275
5276 // One or two coalesced stores to plop down.
5277 Node* st[2];
5278 intptr_t off[2];
5279 int nst = 0;
5280 if (!split) {
5281 ++new_long;
5282 off[nst] = offset;
5283 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5284 phase->longcon(con), T_LONG, MemNode::unordered);
5285 } else {
5286 // Omit either if it is a zero.
5287 if (con0 != 0) {
5288 ++new_int;
5289 off[nst] = offset;
5290 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5291 phase->intcon(con0), T_INT, MemNode::unordered);
5292 }
5293 if (con1 != 0) {
5294 ++new_int;
5295 offset += BytesPerInt;
5296 adr = make_raw_address(offset, phase);
5297 off[nst] = offset;
5298 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5299 phase->intcon(con1), T_INT, MemNode::unordered);
5300 }
5301 }
5302
5303 // Insert second store first, then the first before the second.
5304 // Insert each one just before any overlapping non-constant stores.
5305 while (nst > 0) {
5306 Node* st1 = st[--nst];
5307 C->copy_node_notes_to(st1, old);
5308 st1 = phase->transform(st1);
5309 offset = off[nst];
5310 assert(offset >= header_size, "do not smash header");
5311 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
5312 guarantee(ins_idx != 0, "must re-insert constant store");
5313 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
5314 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
5315 set_req(--ins_idx, st1);
5316 else
5317 ins_req(ins_idx, st1);
5318 }
5319 }
5320
5321 if (PrintCompilation && WizardMode)
5322 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
5323 old_subword, old_long, new_int, new_long);
5324 if (C->log() != nullptr)
5325 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
5326 old_subword, old_long, new_int, new_long);
5327
5328 // Clean up any remaining occurrences of zmem:
5329 remove_extra_zeroes();
5330 }
5331
5332 // Explore forward from in(start) to find the first fully initialized
5333 // word, and return its offset. Skip groups of subword stores which
5334 // together initialize full words. If in(start) is itself part of a
5335 // fully initialized word, return the offset of in(start). If there
5336 // are no following full-word stores, or if something is fishy, return
5337 // a negative value.
5338 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
5339 int int_map = 0;
5340 intptr_t int_map_off = 0;
5341 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
5342
5343 for (uint i = start, limit = req(); i < limit; i++) {
5344 Node* st = in(i);
5345
5346 intptr_t st_off = get_store_offset(st, phase);
5347 if (st_off < 0) break; // return conservative answer
5348
5349 int st_size = st->as_Store()->memory_size();
5350 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
5351 return st_off; // we found a complete word init
5352 }
5353
5354 // update the map:
5355
5356 intptr_t this_int_off = align_down(st_off, BytesPerInt);
5357 if (this_int_off != int_map_off) {
5358 // reset the map:
5359 int_map = 0;
5360 int_map_off = this_int_off;
5361 }
5362
5363 int subword_off = st_off - this_int_off;
5364 int_map |= right_n_bits(st_size) << subword_off;
5365 if ((int_map & FULL_MAP) == FULL_MAP) {
5366 return this_int_off; // we found a complete word init
5367 }
5368
5369 // Did this store hit or cross the word boundary?
5370 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
5371 if (next_int_off == this_int_off + BytesPerInt) {
5372 // We passed the current int, without fully initializing it.
5373 int_map_off = next_int_off;
5374 int_map >>= BytesPerInt;
5375 } else if (next_int_off > this_int_off + BytesPerInt) {
5376 // We passed the current and next int.
5377 return this_int_off + BytesPerInt;
5378 }
5379 }
5380
5381 return -1;
5382 }
5383
5384
5385 // Called when the associated AllocateNode is expanded into CFG.
5386 // At this point, we may perform additional optimizations.
5387 // Linearize the stores by ascending offset, to make memory
5388 // activity as coherent as possible.
5389 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
5390 intptr_t header_size,
5391 Node* size_in_bytes,
5392 PhaseIterGVN* phase) {
5393 assert(!is_complete(), "not already complete");
5394 assert(stores_are_sane(phase), "");
5395 assert(allocation() != nullptr, "must be present");
5396
5397 remove_extra_zeroes();
5398
5399 if (ReduceFieldZeroing || ReduceBulkZeroing)
5400 // reduce instruction count for common initialization patterns
5401 coalesce_subword_stores(header_size, size_in_bytes, phase);
5402
5403 Node* zmem = zero_memory(); // initially zero memory state
5404 Node* inits = zmem; // accumulating a linearized chain of inits
5405 #ifdef ASSERT
5406 intptr_t first_offset = allocation()->minimum_header_size();
5407 intptr_t last_init_off = first_offset; // previous init offset
5408 intptr_t last_init_end = first_offset; // previous init offset+size
5409 intptr_t last_tile_end = first_offset; // previous tile offset+size
5410 #endif
5411 intptr_t zeroes_done = header_size;
5412
5413 bool do_zeroing = true; // we might give up if inits are very sparse
5414 int big_init_gaps = 0; // how many large gaps have we seen?
5415
5416 if (UseTLAB && ZeroTLAB) do_zeroing = false;
5417 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
5418
5419 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5420 Node* st = in(i);
5421 intptr_t st_off = get_store_offset(st, phase);
5422 if (st_off < 0)
5423 break; // unknown junk in the inits
5424 if (st->in(MemNode::Memory) != zmem)
5425 break; // complicated store chains somehow in list
5426
5427 int st_size = st->as_Store()->memory_size();
5428 intptr_t next_init_off = st_off + st_size;
5429
5430 if (do_zeroing && zeroes_done < next_init_off) {
5431 // See if this store needs a zero before it or under it.
5432 intptr_t zeroes_needed = st_off;
5433
5434 if (st_size < BytesPerInt) {
5435 // Look for subword stores which only partially initialize words.
5436 // If we find some, we must lay down some word-level zeroes first,
5437 // underneath the subword stores.
5438 //
5439 // Examples:
5440 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
5441 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
5442 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
5443 //
5444 // Note: coalesce_subword_stores may have already done this,
5445 // if it was prompted by constant non-zero subword initializers.
5446 // But this case can still arise with non-constant stores.
5447
5448 intptr_t next_full_store = find_next_fullword_store(i, phase);
5449
5450 // In the examples above:
5451 // in(i) p q r s x y z
5452 // st_off 12 13 14 15 12 13 14
5453 // st_size 1 1 1 1 1 1 1
5454 // next_full_s. 12 16 16 16 16 16 16
5455 // z's_done 12 16 16 16 12 16 12
5456 // z's_needed 12 16 16 16 16 16 16
5457 // zsize 0 0 0 0 4 0 4
5458 if (next_full_store < 0) {
5459 // Conservative tack: Zero to end of current word.
5460 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
5461 } else {
5462 // Zero to beginning of next fully initialized word.
5463 // Or, don't zero at all, if we are already in that word.
5464 assert(next_full_store >= zeroes_needed, "must go forward");
5465 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
5466 zeroes_needed = next_full_store;
5467 }
5468 }
5469
5470 if (zeroes_needed > zeroes_done) {
5471 intptr_t zsize = zeroes_needed - zeroes_done;
5472 // Do some incremental zeroing on rawmem, in parallel with inits.
5473 zeroes_done = align_down(zeroes_done, BytesPerInt);
5474 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5475 zeroes_done, zeroes_needed,
5476 true,
5477 phase);
5478 zeroes_done = zeroes_needed;
5479 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
5480 do_zeroing = false; // leave the hole, next time
5481 }
5482 }
5483
5484 // Collect the store and move on:
5485 phase->replace_input_of(st, MemNode::Memory, inits);
5486 inits = st; // put it on the linearized chain
5487 set_req(i, zmem); // unhook from previous position
5488
5489 if (zeroes_done == st_off)
5490 zeroes_done = next_init_off;
5491
5492 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
5493
5494 #ifdef ASSERT
5495 // Various order invariants. Weaker than stores_are_sane because
5496 // a large constant tile can be filled in by smaller non-constant stores.
5497 assert(st_off >= last_init_off, "inits do not reverse");
5498 last_init_off = st_off;
5499 const Type* val = nullptr;
5500 if (st_size >= BytesPerInt &&
5501 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
5502 (int)val->basic_type() < (int)T_OBJECT) {
5503 assert(st_off >= last_tile_end, "tiles do not overlap");
5504 assert(st_off >= last_init_end, "tiles do not overwrite inits");
5505 last_tile_end = MAX2(last_tile_end, next_init_off);
5506 } else {
5507 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
5508 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
5509 assert(st_off >= last_init_end, "inits do not overlap");
5510 last_init_end = next_init_off; // it's a non-tile
5511 }
5512 #endif //ASSERT
5513 }
5514
5515 remove_extra_zeroes(); // clear out all the zmems left over
5516 add_req(inits);
5517
5518 if (!(UseTLAB && ZeroTLAB)) {
5519 // If anything remains to be zeroed, zero it all now.
5520 zeroes_done = align_down(zeroes_done, BytesPerInt);
5521 // if it is the last unused 4 bytes of an instance, forget about it
5522 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
5523 if (zeroes_done + BytesPerLong >= size_limit) {
5524 AllocateNode* alloc = allocation();
5525 assert(alloc != nullptr, "must be present");
5526 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
5527 Node* klass_node = alloc->in(AllocateNode::KlassNode);
5528 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
5529 if (zeroes_done == k->layout_helper())
5530 zeroes_done = size_limit;
5531 }
5532 }
5533 if (zeroes_done < size_limit) {
5534 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5535 zeroes_done, size_in_bytes, true, phase);
5536 }
5537 }
5538
5539 set_complete(phase);
5540 return rawmem;
5541 }
5542
5543 void InitializeNode::replace_mem_projs_by(Node* mem, Compile* C) {
5544 auto replace_proj = [&](ProjNode* proj) {
5545 C->gvn_replace_by(proj, mem);
5546 return CONTINUE;
5547 };
5548 apply_to_projs(replace_proj, TypeFunc::Memory);
5549 }
5550
5551 void InitializeNode::replace_mem_projs_by(Node* mem, PhaseIterGVN* igvn) {
5552 DUIterator_Fast imax, i = fast_outs(imax);
5553 auto replace_proj = [&](ProjNode* proj) {
5554 igvn->replace_node(proj, mem);
5555 --i; --imax;
5556 return CONTINUE;
5557 };
5558 apply_to_projs(imax, i, replace_proj, TypeFunc::Memory);
5559 }
5560
5561 bool InitializeNode::already_has_narrow_mem_proj_with_adr_type(const TypePtr* adr_type) const {
5562 auto find_proj = [&](ProjNode* proj) {
5563 if (proj->adr_type() == adr_type) {
5564 return BREAK_AND_RETURN_CURRENT_PROJ;
5565 }
5566 return CONTINUE;
5567 };
5568 DUIterator_Fast imax, i = fast_outs(imax);
5569 return apply_to_narrow_mem_projs_any_iterator(UsesIteratorFast(imax, i, this), find_proj) != nullptr;
5570 }
5571
5572 MachProjNode* InitializeNode::mem_mach_proj() const {
5573 auto find_proj = [](ProjNode* proj) {
5574 if (proj->is_MachProj()) {
5575 return BREAK_AND_RETURN_CURRENT_PROJ;
5576 }
5577 return CONTINUE;
5578 };
5579 ProjNode* proj = apply_to_projs(find_proj, TypeFunc::Memory);
5580 if (proj == nullptr) {
5581 return nullptr;
5582 }
5583 return proj->as_MachProj();
5584 }
5585
5586 #ifdef ASSERT
5587 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
5588 if (is_complete())
5589 return true; // stores could be anything at this point
5590 assert(allocation() != nullptr, "must be present");
5591 intptr_t last_off = allocation()->minimum_header_size();
5592 for (uint i = InitializeNode::RawStores; i < req(); i++) {
5593 Node* st = in(i);
5594 intptr_t st_off = get_store_offset(st, phase);
5595 if (st_off < 0) continue; // ignore dead garbage
5596 if (last_off > st_off) {
5597 tty->print_cr("*** bad store offset at %d: %zd > %zd", i, last_off, st_off);
5598 this->dump(2);
5599 assert(false, "ascending store offsets");
5600 return false;
5601 }
5602 last_off = st_off + st->as_Store()->memory_size();
5603 }
5604 return true;
5605 }
5606 #endif //ASSERT
5607
5608
5609
5610
5611 //============================MergeMemNode=====================================
5612 //
5613 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
5614 // contributing store or call operations. Each contributor provides the memory
5615 // state for a particular "alias type" (see Compile::alias_type). For example,
5616 // if a MergeMem has an input X for alias category #6, then any memory reference
5617 // to alias category #6 may use X as its memory state input, as an exact equivalent
5618 // to using the MergeMem as a whole.
5619 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
5620 //
5621 // (Here, the <N> notation gives the index of the relevant adr_type.)
5622 //
5623 // In one special case (and more cases in the future), alias categories overlap.
5624 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
5625 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
5626 // it is exactly equivalent to that state W:
5627 // MergeMem(<Bot>: W) <==> W
5628 //
5629 // Usually, the merge has more than one input. In that case, where inputs
5630 // overlap (i.e., one is Bot), the narrower alias type determines the memory
5631 // state for that type, and the wider alias type (Bot) fills in everywhere else:
5632 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
5633 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
5634 //
5635 // A merge can take a "wide" memory state as one of its narrow inputs.
5636 // This simply means that the merge observes out only the relevant parts of
5637 // the wide input. That is, wide memory states arriving at narrow merge inputs
5638 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
5639 //
5640 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
5641 // and that memory slices "leak through":
5642 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
5643 //
5644 // But, in such a cascade, repeated memory slices can "block the leak":
5645 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
5646 //
5647 // In the last example, Y is not part of the combined memory state of the
5648 // outermost MergeMem. The system must, of course, prevent unschedulable
5649 // memory states from arising, so you can be sure that the state Y is somehow
5650 // a precursor to state Y'.
5651 //
5652 //
5653 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
5654 // of each MergeMemNode array are exactly the numerical alias indexes, including
5655 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
5656 // Compile::alias_type (and kin) produce and manage these indexes.
5657 //
5658 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
5659 // (Note that this provides quick access to the top node inside MergeMem methods,
5660 // without the need to reach out via TLS to Compile::current.)
5661 //
5662 // As a consequence of what was just described, a MergeMem that represents a full
5663 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
5664 // containing all alias categories.
5665 //
5666 // MergeMem nodes never (?) have control inputs, so in(0) is null.
5667 //
5668 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
5669 // a memory state for the alias type <N>, or else the top node, meaning that
5670 // there is no particular input for that alias type. Note that the length of
5671 // a MergeMem is variable, and may be extended at any time to accommodate new
5672 // memory states at larger alias indexes. When merges grow, they are of course
5673 // filled with "top" in the unused in() positions.
5674 //
5675 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
5676 // (Top was chosen because it works smoothly with passes like GCM.)
5677 //
5678 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
5679 // the type of random VM bits like TLS references.) Since it is always the
5680 // first non-Bot memory slice, some low-level loops use it to initialize an
5681 // index variable: for (i = AliasIdxRaw; i < req(); i++).
5682 //
5683 //
5684 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
5685 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
5686 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
5687 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
5688 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
5689 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
5690 //
5691 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
5692 // really that different from the other memory inputs. An abbreviation called
5693 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
5694 //
5695 //
5696 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
5697 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
5698 // that "emerges though" the base memory will be marked as excluding the alias types
5699 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
5700 //
5701 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
5702 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
5703 //
5704 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
5705 // (It is currently unimplemented.) As you can see, the resulting merge is
5706 // actually a disjoint union of memory states, rather than an overlay.
5707 //
5708
5709 //------------------------------MergeMemNode-----------------------------------
5710 Node* MergeMemNode::make_empty_memory() {
5711 Node* empty_memory = (Node*) Compile::current()->top();
5712 assert(empty_memory->is_top(), "correct sentinel identity");
5713 return empty_memory;
5714 }
5715
5716 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
5717 init_class_id(Class_MergeMem);
5718 // all inputs are nullified in Node::Node(int)
5719 // set_input(0, nullptr); // no control input
5720
5721 // Initialize the edges uniformly to top, for starters.
5722 Node* empty_mem = make_empty_memory();
5723 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
5724 init_req(i,empty_mem);
5725 }
5726 assert(empty_memory() == empty_mem, "");
5727
5728 if( new_base != nullptr && new_base->is_MergeMem() ) {
5729 MergeMemNode* mdef = new_base->as_MergeMem();
5730 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
5731 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
5732 mms.set_memory(mms.memory2());
5733 }
5734 assert(base_memory() == mdef->base_memory(), "");
5735 } else {
5736 set_base_memory(new_base);
5737 }
5738 }
5739
5740 // Make a new, untransformed MergeMem with the same base as 'mem'.
5741 // If mem is itself a MergeMem, populate the result with the same edges.
5742 MergeMemNode* MergeMemNode::make(Node* mem) {
5743 return new MergeMemNode(mem);
5744 }
5745
5746 //------------------------------cmp--------------------------------------------
5747 uint MergeMemNode::hash() const { return NO_HASH; }
5748 bool MergeMemNode::cmp( const Node &n ) const {
5749 return (&n == this); // Always fail except on self
5750 }
5751
5752 //------------------------------Identity---------------------------------------
5753 Node* MergeMemNode::Identity(PhaseGVN* phase) {
5754 // Identity if this merge point does not record any interesting memory
5755 // disambiguations.
5756 Node* base_mem = base_memory();
5757 Node* empty_mem = empty_memory();
5758 if (base_mem != empty_mem) { // Memory path is not dead?
5759 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5760 Node* mem = in(i);
5761 if (mem != empty_mem && mem != base_mem) {
5762 return this; // Many memory splits; no change
5763 }
5764 }
5765 }
5766 return base_mem; // No memory splits; ID on the one true input
5767 }
5768
5769 //------------------------------Ideal------------------------------------------
5770 // This method is invoked recursively on chains of MergeMem nodes
5771 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
5772 // Remove chain'd MergeMems
5773 //
5774 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
5775 // relative to the "in(Bot)". Since we are patching both at the same time,
5776 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
5777 // but rewrite each "in(i)" relative to the new "in(Bot)".
5778 Node *progress = nullptr;
5779
5780
5781 Node* old_base = base_memory();
5782 Node* empty_mem = empty_memory();
5783 if (old_base == empty_mem)
5784 return nullptr; // Dead memory path.
5785
5786 MergeMemNode* old_mbase;
5787 if (old_base != nullptr && old_base->is_MergeMem())
5788 old_mbase = old_base->as_MergeMem();
5789 else
5790 old_mbase = nullptr;
5791 Node* new_base = old_base;
5792
5793 // simplify stacked MergeMems in base memory
5794 if (old_mbase) new_base = old_mbase->base_memory();
5795
5796 // the base memory might contribute new slices beyond my req()
5797 if (old_mbase) grow_to_match(old_mbase);
5798
5799 // Note: We do not call verify_sparse on entry, because inputs
5800 // can normalize to the base_memory via subsume_node or similar
5801 // mechanisms. This method repairs that damage.
5802
5803 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
5804
5805 // Look at each slice.
5806 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5807 Node* old_in = in(i);
5808 // calculate the old memory value
5809 Node* old_mem = old_in;
5810 if (old_mem == empty_mem) old_mem = old_base;
5811 assert(old_mem == memory_at(i), "");
5812
5813 // maybe update (reslice) the old memory value
5814
5815 // simplify stacked MergeMems
5816 Node* new_mem = old_mem;
5817 MergeMemNode* old_mmem;
5818 if (old_mem != nullptr && old_mem->is_MergeMem())
5819 old_mmem = old_mem->as_MergeMem();
5820 else
5821 old_mmem = nullptr;
5822 if (old_mmem == this) {
5823 // This can happen if loops break up and safepoints disappear.
5824 // A merge of BotPtr (default) with a RawPtr memory derived from a
5825 // safepoint can be rewritten to a merge of the same BotPtr with
5826 // the BotPtr phi coming into the loop. If that phi disappears
5827 // also, we can end up with a self-loop of the mergemem.
5828 // In general, if loops degenerate and memory effects disappear,
5829 // a mergemem can be left looking at itself. This simply means
5830 // that the mergemem's default should be used, since there is
5831 // no longer any apparent effect on this slice.
5832 // Note: If a memory slice is a MergeMem cycle, it is unreachable
5833 // from start. Update the input to TOP.
5834 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
5835 }
5836 else if (old_mmem != nullptr) {
5837 new_mem = old_mmem->memory_at(i);
5838 }
5839 // else preceding memory was not a MergeMem
5840
5841 // maybe store down a new value
5842 Node* new_in = new_mem;
5843 if (new_in == new_base) new_in = empty_mem;
5844
5845 if (new_in != old_in) {
5846 // Warning: Do not combine this "if" with the previous "if"
5847 // A memory slice might have be be rewritten even if it is semantically
5848 // unchanged, if the base_memory value has changed.
5849 set_req_X(i, new_in, phase);
5850 progress = this; // Report progress
5851 }
5852 }
5853
5854 if (new_base != old_base) {
5855 set_req_X(Compile::AliasIdxBot, new_base, phase);
5856 // Don't use set_base_memory(new_base), because we need to update du.
5857 assert(base_memory() == new_base, "");
5858 progress = this;
5859 }
5860
5861 if( base_memory() == this ) {
5862 // a self cycle indicates this memory path is dead
5863 set_req(Compile::AliasIdxBot, empty_mem);
5864 }
5865
5866 // Resolve external cycles by calling Ideal on a MergeMem base_memory
5867 // Recursion must occur after the self cycle check above
5868 if( base_memory()->is_MergeMem() ) {
5869 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
5870 Node *m = phase->transform(new_mbase); // Rollup any cycles
5871 if( m != nullptr &&
5872 (m->is_top() ||
5873 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
5874 // propagate rollup of dead cycle to self
5875 set_req(Compile::AliasIdxBot, empty_mem);
5876 }
5877 }
5878
5879 if( base_memory() == empty_mem ) {
5880 progress = this;
5881 // Cut inputs during Parse phase only.
5882 // During Optimize phase a dead MergeMem node will be subsumed by Top.
5883 if( !can_reshape ) {
5884 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5885 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
5886 }
5887 }
5888 }
5889
5890 if( !progress && base_memory()->is_Phi() && can_reshape ) {
5891 // Check if PhiNode::Ideal's "Split phis through memory merges"
5892 // transform should be attempted. Look for this->phi->this cycle.
5893 uint merge_width = req();
5894 if (merge_width > Compile::AliasIdxRaw) {
5895 PhiNode* phi = base_memory()->as_Phi();
5896 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
5897 if (phi->in(i) == this) {
5898 phase->is_IterGVN()->_worklist.push(phi);
5899 break;
5900 }
5901 }
5902 }
5903 }
5904
5905 assert(progress || verify_sparse(), "please, no dups of base");
5906 return progress;
5907 }
5908
5909 //-------------------------set_base_memory-------------------------------------
5910 void MergeMemNode::set_base_memory(Node *new_base) {
5911 Node* empty_mem = empty_memory();
5912 set_req(Compile::AliasIdxBot, new_base);
5913 assert(memory_at(req()) == new_base, "must set default memory");
5914 // Clear out other occurrences of new_base:
5915 if (new_base != empty_mem) {
5916 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5917 if (in(i) == new_base) set_req(i, empty_mem);
5918 }
5919 }
5920 }
5921
5922 //------------------------------out_RegMask------------------------------------
5923 const RegMask &MergeMemNode::out_RegMask() const {
5924 return RegMask::EMPTY;
5925 }
5926
5927 //------------------------------dump_spec--------------------------------------
5928 #ifndef PRODUCT
5929 void MergeMemNode::dump_spec(outputStream *st) const {
5930 st->print(" {");
5931 Node* base_mem = base_memory();
5932 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
5933 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem;
5934 if (mem == base_mem) { st->print(" -"); continue; }
5935 st->print( " N%d:", mem->_idx );
5936 Compile::current()->get_adr_type(i)->dump_on(st);
5937 }
5938 st->print(" }");
5939 }
5940 #endif // !PRODUCT
5941
5942
5943 #ifdef ASSERT
5944 static bool might_be_same(Node* a, Node* b) {
5945 if (a == b) return true;
5946 if (!(a->is_Phi() || b->is_Phi())) return false;
5947 // phis shift around during optimization
5948 return true; // pretty stupid...
5949 }
5950
5951 // verify a narrow slice (either incoming or outgoing)
5952 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
5953 if (!VerifyAliases) return; // don't bother to verify unless requested
5954 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
5955 if (Node::in_dump()) return; // muzzle asserts when printing
5956 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
5957 assert(n != nullptr, "");
5958 // Elide intervening MergeMem's
5959 while (n->is_MergeMem()) {
5960 n = n->as_MergeMem()->memory_at(alias_idx);
5961 }
5962 Compile* C = Compile::current();
5963 const TypePtr* n_adr_type = n->adr_type();
5964 if (n == m->empty_memory()) {
5965 // Implicit copy of base_memory()
5966 } else if (n_adr_type != TypePtr::BOTTOM) {
5967 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type");
5968 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
5969 } else {
5970 // A few places like make_runtime_call "know" that VM calls are narrow,
5971 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
5972 bool expected_wide_mem = false;
5973 if (n == m->base_memory()) {
5974 expected_wide_mem = true;
5975 } else if (alias_idx == Compile::AliasIdxRaw ||
5976 n == m->memory_at(Compile::AliasIdxRaw)) {
5977 expected_wide_mem = true;
5978 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
5979 // memory can "leak through" calls on channels that
5980 // are write-once. Allow this also.
5981 expected_wide_mem = true;
5982 }
5983 assert(expected_wide_mem, "expected narrow slice replacement");
5984 }
5985 }
5986 #else // !ASSERT
5987 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
5988 #endif
5989
5990
5991 //-----------------------------memory_at---------------------------------------
5992 Node* MergeMemNode::memory_at(uint alias_idx) const {
5993 assert(alias_idx >= Compile::AliasIdxRaw ||
5994 (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()),
5995 "must avoid base_memory and AliasIdxTop");
5996
5997 // Otherwise, it is a narrow slice.
5998 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
5999 if (is_empty_memory(n)) {
6000 // the array is sparse; empty slots are the "top" node
6001 n = base_memory();
6002 assert(Node::in_dump()
6003 || n == nullptr || n->bottom_type() == Type::TOP
6004 || n->adr_type() == nullptr // address is TOP
6005 || n->adr_type() == TypePtr::BOTTOM
6006 || n->adr_type() == TypeRawPtr::BOTTOM
6007 || n->is_NarrowMemProj()
6008 || !Compile::current()->do_aliasing(),
6009 "must be a wide memory");
6010 // do_aliasing == false if we are organizing the memory states manually.
6011 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
6012 } else {
6013 // make sure the stored slice is sane
6014 #ifdef ASSERT
6015 if (VMError::is_error_reported() || Node::in_dump()) {
6016 } else if (might_be_same(n, base_memory())) {
6017 // Give it a pass: It is a mostly harmless repetition of the base.
6018 // This can arise normally from node subsumption during optimization.
6019 } else {
6020 verify_memory_slice(this, alias_idx, n);
6021 }
6022 #endif
6023 }
6024 return n;
6025 }
6026
6027 //---------------------------set_memory_at-------------------------------------
6028 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
6029 verify_memory_slice(this, alias_idx, n);
6030 Node* empty_mem = empty_memory();
6031 if (n == base_memory()) n = empty_mem; // collapse default
6032 uint need_req = alias_idx+1;
6033 if (req() < need_req) {
6034 if (n == empty_mem) return; // already the default, so do not grow me
6035 // grow the sparse array
6036 do {
6037 add_req(empty_mem);
6038 } while (req() < need_req);
6039 }
6040 set_req( alias_idx, n );
6041 }
6042
6043
6044
6045 //--------------------------iteration_setup------------------------------------
6046 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
6047 if (other != nullptr) {
6048 grow_to_match(other);
6049 // invariant: the finite support of mm2 is within mm->req()
6050 #ifdef ASSERT
6051 for (uint i = req(); i < other->req(); i++) {
6052 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
6053 }
6054 #endif
6055 }
6056 // Replace spurious copies of base_memory by top.
6057 Node* base_mem = base_memory();
6058 if (base_mem != nullptr && !base_mem->is_top()) {
6059 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
6060 if (in(i) == base_mem)
6061 set_req(i, empty_memory());
6062 }
6063 }
6064 }
6065
6066 //---------------------------grow_to_match-------------------------------------
6067 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
6068 Node* empty_mem = empty_memory();
6069 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
6070 // look for the finite support of the other memory
6071 for (uint i = other->req(); --i >= req(); ) {
6072 if (other->in(i) != empty_mem) {
6073 uint new_len = i+1;
6074 while (req() < new_len) add_req(empty_mem);
6075 break;
6076 }
6077 }
6078 }
6079
6080 //---------------------------verify_sparse-------------------------------------
6081 #ifndef PRODUCT
6082 bool MergeMemNode::verify_sparse() const {
6083 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
6084 Node* base_mem = base_memory();
6085 // The following can happen in degenerate cases, since empty==top.
6086 if (is_empty_memory(base_mem)) return true;
6087 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6088 assert(in(i) != nullptr, "sane slice");
6089 if (in(i) == base_mem) return false; // should have been the sentinel value!
6090 }
6091 return true;
6092 }
6093
6094 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
6095 Node* n;
6096 n = mm->in(idx);
6097 if (mem == n) return true; // might be empty_memory()
6098 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
6099 if (mem == n) return true;
6100 return false;
6101 }
6102 #endif // !PRODUCT