1 /*
2 * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2024, Alibaba Group Holding Limited. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "precompiled.hpp"
27 #include "ci/ciFlatArrayKlass.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "classfile/systemDictionary.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "gc/shared/barrierSet.hpp"
32 #include "gc/shared/c2/barrierSetC2.hpp"
33 #include "gc/shared/tlab_globals.hpp"
34 #include "memory/allocation.inline.hpp"
35 #include "memory/resourceArea.hpp"
36 #include "oops/objArrayKlass.hpp"
37 #include "opto/addnode.hpp"
38 #include "opto/arraycopynode.hpp"
39 #include "opto/cfgnode.hpp"
40 #include "opto/regalloc.hpp"
41 #include "opto/compile.hpp"
42 #include "opto/connode.hpp"
43 #include "opto/convertnode.hpp"
44 #include "opto/inlinetypenode.hpp"
45 #include "opto/loopnode.hpp"
46 #include "opto/machnode.hpp"
47 #include "opto/matcher.hpp"
48 #include "opto/memnode.hpp"
49 #include "opto/mulnode.hpp"
50 #include "opto/narrowptrnode.hpp"
51 #include "opto/phaseX.hpp"
52 #include "opto/regmask.hpp"
53 #include "opto/rootnode.hpp"
54 #include "opto/vectornode.hpp"
55 #include "utilities/align.hpp"
56 #include "utilities/copy.hpp"
57 #include "utilities/macros.hpp"
58 #include "utilities/powerOfTwo.hpp"
59 #include "utilities/vmError.hpp"
60
61 // Portions of code courtesy of Clifford Click
62
63 // Optimization - Graph Style
64
65 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
66
67 //=============================================================================
68 uint MemNode::size_of() const { return sizeof(*this); }
69
70 const TypePtr *MemNode::adr_type() const {
71 Node* adr = in(Address);
72 if (adr == nullptr) return nullptr; // node is dead
73 const TypePtr* cross_check = nullptr;
74 DEBUG_ONLY(cross_check = _adr_type);
75 return calculate_adr_type(adr->bottom_type(), cross_check);
76 }
77
78 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
79 if (adr != nullptr) {
80 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
81 return true;
82 }
83 }
84 return false;
85 }
86
87 #ifndef PRODUCT
88 void MemNode::dump_spec(outputStream *st) const {
89 if (in(Address) == nullptr) return; // node is dead
90 #ifndef ASSERT
91 // fake the missing field
92 const TypePtr* _adr_type = nullptr;
93 if (in(Address) != nullptr)
94 _adr_type = in(Address)->bottom_type()->isa_ptr();
95 #endif
96 dump_adr_type(this, _adr_type, st);
97
98 Compile* C = Compile::current();
99 if (C->alias_type(_adr_type)->is_volatile()) {
100 st->print(" Volatile!");
101 }
102 if (_unaligned_access) {
103 st->print(" unaligned");
104 }
105 if (_mismatched_access) {
106 st->print(" mismatched");
107 }
108 if (_unsafe_access) {
109 st->print(" unsafe");
110 }
111 }
112
113 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
114 st->print(" @");
115 if (adr_type == nullptr) {
116 st->print("null");
117 } else {
118 adr_type->dump_on(st);
119 Compile* C = Compile::current();
120 Compile::AliasType* atp = nullptr;
121 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
122 if (atp == nullptr)
123 st->print(", idx=?\?;");
124 else if (atp->index() == Compile::AliasIdxBot)
125 st->print(", idx=Bot;");
126 else if (atp->index() == Compile::AliasIdxTop)
127 st->print(", idx=Top;");
128 else if (atp->index() == Compile::AliasIdxRaw)
129 st->print(", idx=Raw;");
130 else {
131 ciField* field = atp->field();
132 if (field) {
133 st->print(", name=");
134 field->print_name_on(st);
135 }
136 st->print(", idx=%d;", atp->index());
137 }
138 }
139 }
140
141 extern void print_alias_types();
142
143 #endif
144
145 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
146 assert((t_oop != nullptr), "sanity");
147 bool is_instance = t_oop->is_known_instance_field();
148 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
149 (load != nullptr) && load->is_Load() &&
150 (phase->is_IterGVN() != nullptr);
151 if (!(is_instance || is_boxed_value_load))
152 return mchain; // don't try to optimize non-instance types
153 uint instance_id = t_oop->instance_id();
154 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
155 Node *prev = nullptr;
156 Node *result = mchain;
157 while (prev != result) {
158 prev = result;
159 if (result == start_mem)
160 break; // hit one of our sentinels
161 // skip over a call which does not affect this memory slice
162 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
163 Node *proj_in = result->in(0);
164 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
165 break; // hit one of our sentinels
166 } else if (proj_in->is_Call()) {
167 // ArrayCopyNodes processed here as well
168 CallNode *call = proj_in->as_Call();
169 if (!call->may_modify(t_oop, phase)) { // returns false for instances
170 result = call->in(TypeFunc::Memory);
171 }
172 } else if (proj_in->is_Initialize()) {
173 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
174 // Stop if this is the initialization for the object instance which
175 // contains this memory slice, otherwise skip over it.
176 if ((alloc == nullptr) || (alloc->_idx == instance_id)) {
177 break;
178 }
179 if (is_instance) {
180 result = proj_in->in(TypeFunc::Memory);
181 } else if (is_boxed_value_load) {
182 Node* klass = alloc->in(AllocateNode::KlassNode);
183 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
184 if (tklass->klass_is_exact() && !tklass->exact_klass()->equals(t_oop->is_instptr()->exact_klass())) {
185 result = proj_in->in(TypeFunc::Memory); // not related allocation
186 }
187 }
188 } else if (proj_in->is_MemBar()) {
189 ArrayCopyNode* ac = nullptr;
190 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
191 break;
192 }
193 result = proj_in->in(TypeFunc::Memory);
194 } else if (proj_in->is_top()) {
195 break; // dead code
196 } else {
197 assert(false, "unexpected projection");
198 }
199 } else if (result->is_ClearArray()) {
200 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
201 // Can not bypass initialization of the instance
202 // we are looking for.
203 break;
204 }
205 // Otherwise skip it (the call updated 'result' value).
206 } else if (result->is_MergeMem()) {
207 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, nullptr, tty);
208 }
209 }
210 return result;
211 }
212
213 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
214 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
215 if (t_oop == nullptr)
216 return mchain; // don't try to optimize non-oop types
217 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
218 bool is_instance = t_oop->is_known_instance_field();
219 PhaseIterGVN *igvn = phase->is_IterGVN();
220 if (is_instance && igvn != nullptr && result->is_Phi()) {
221 PhiNode *mphi = result->as_Phi();
222 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
223 const TypePtr *t = mphi->adr_type();
224 bool do_split = false;
225 // In the following cases, Load memory input can be further optimized based on
226 // its precise address type
227 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ) {
228 do_split = true;
229 } else if (t->isa_oopptr() && !t->is_oopptr()->is_known_instance()) {
230 const TypeOopPtr* mem_t =
231 t->is_oopptr()->cast_to_exactness(true)
232 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
233 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id());
234 if (t_oop->isa_aryptr()) {
235 mem_t = mem_t->is_aryptr()
236 ->cast_to_stable(t_oop->is_aryptr()->is_stable())
237 ->cast_to_size(t_oop->is_aryptr()->size())
238 ->cast_to_not_flat(t_oop->is_aryptr()->is_not_flat())
239 ->cast_to_not_null_free(t_oop->is_aryptr()->is_not_null_free())
240 ->with_offset(t_oop->is_aryptr()->offset())
241 ->is_aryptr();
242 }
243 do_split = mem_t == t_oop;
244 }
245 if (do_split) {
246 // clone the Phi with our address type
247 result = mphi->split_out_instance(t_adr, igvn);
248 } else {
249 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
250 }
251 }
252 return result;
253 }
254
255 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
256 uint alias_idx = phase->C->get_alias_index(tp);
257 Node *mem = mmem;
258 #ifdef ASSERT
259 {
260 // Check that current type is consistent with the alias index used during graph construction
261 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
262 bool consistent = adr_check == nullptr || adr_check->empty() ||
263 phase->C->must_alias(adr_check, alias_idx );
264 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
265 if( !consistent && adr_check != nullptr && !adr_check->empty() &&
266 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
267 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
268 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
269 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
270 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
271 // don't assert if it is dead code.
272 consistent = true;
273 }
274 if( !consistent ) {
275 st->print("alias_idx==%d, adr_check==", alias_idx);
276 if( adr_check == nullptr ) {
277 st->print("null");
278 } else {
279 adr_check->dump();
280 }
281 st->cr();
282 print_alias_types();
283 assert(consistent, "adr_check must match alias idx");
284 }
285 }
286 #endif
287 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
288 // means an array I have not precisely typed yet. Do not do any
289 // alias stuff with it any time soon.
290 const TypeOopPtr *toop = tp->isa_oopptr();
291 if (tp->base() != Type::AnyPtr &&
292 !(toop &&
293 toop->isa_instptr() &&
294 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
295 toop->offset() == Type::OffsetBot)) {
296 // compress paths and change unreachable cycles to TOP
297 // If not, we can update the input infinitely along a MergeMem cycle
298 // Equivalent code in PhiNode::Ideal
299 Node* m = phase->transform(mmem);
300 // If transformed to a MergeMem, get the desired slice
301 // Otherwise the returned node represents memory for every slice
302 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
303 // Update input if it is progress over what we have now
304 }
305 return mem;
306 }
307
308 //--------------------------Ideal_common---------------------------------------
309 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
310 // Unhook non-raw memories from complete (macro-expanded) initializations.
311 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
312 // If our control input is a dead region, kill all below the region
313 Node *ctl = in(MemNode::Control);
314 if (ctl && remove_dead_region(phase, can_reshape))
315 return this;
316 ctl = in(MemNode::Control);
317 // Don't bother trying to transform a dead node
318 if (ctl && ctl->is_top()) return NodeSentinel;
319
320 PhaseIterGVN *igvn = phase->is_IterGVN();
321 // Wait if control on the worklist.
322 if (ctl && can_reshape && igvn != nullptr) {
323 Node* bol = nullptr;
324 Node* cmp = nullptr;
325 if (ctl->in(0)->is_If()) {
326 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
327 bol = ctl->in(0)->in(1);
328 if (bol->is_Bool())
329 cmp = ctl->in(0)->in(1)->in(1);
330 }
331 if (igvn->_worklist.member(ctl) ||
332 (bol != nullptr && igvn->_worklist.member(bol)) ||
333 (cmp != nullptr && igvn->_worklist.member(cmp)) ) {
334 // This control path may be dead.
335 // Delay this memory node transformation until the control is processed.
336 igvn->_worklist.push(this);
337 return NodeSentinel; // caller will return null
338 }
339 }
340 // Ignore if memory is dead, or self-loop
341 Node *mem = in(MemNode::Memory);
342 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return null
343 assert(mem != this, "dead loop in MemNode::Ideal");
344
345 if (can_reshape && igvn != nullptr && igvn->_worklist.member(mem)) {
346 // This memory slice may be dead.
347 // Delay this mem node transformation until the memory is processed.
348 igvn->_worklist.push(this);
349 return NodeSentinel; // caller will return null
350 }
351
352 Node *address = in(MemNode::Address);
353 const Type *t_adr = phase->type(address);
354 if (t_adr == Type::TOP) return NodeSentinel; // caller will return null
355
356 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
357 // Unsafe off-heap access with zero address. Remove access and other control users
358 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
359 assert(ctl != nullptr, "unsafe accesses should be control dependent");
360 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
361 Node* u = ctl->fast_out(i);
362 if (u != ctl) {
363 igvn->rehash_node_delayed(u);
364 int nb = u->replace_edge(ctl, phase->C->top(), igvn);
365 --i, imax -= nb;
366 }
367 }
368 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
369 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
370 phase->C->root()->add_req(halt);
371 return this;
372 }
373
374 if (can_reshape && igvn != nullptr &&
375 (igvn->_worklist.member(address) ||
376 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
377 // The address's base and type may change when the address is processed.
378 // Delay this mem node transformation until the address is processed.
379 igvn->_worklist.push(this);
380 return NodeSentinel; // caller will return null
381 }
382
383 // Do NOT remove or optimize the next lines: ensure a new alias index
384 // is allocated for an oop pointer type before Escape Analysis.
385 // Note: C++ will not remove it since the call has side effect.
386 if (t_adr->isa_oopptr()) {
387 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
388 }
389
390 Node* base = nullptr;
391 if (address->is_AddP()) {
392 base = address->in(AddPNode::Base);
393 }
394 if (base != nullptr && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
395 !t_adr->isa_rawptr()) {
396 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
397 // Skip this node optimization if its address has TOP base.
398 return NodeSentinel; // caller will return null
399 }
400
401 // Avoid independent memory operations
402 Node* old_mem = mem;
403
404 // The code which unhooks non-raw memories from complete (macro-expanded)
405 // initializations was removed. After macro-expansion all stores caught
406 // by Initialize node became raw stores and there is no information
407 // which memory slices they modify. So it is unsafe to move any memory
408 // operation above these stores. Also in most cases hooked non-raw memories
409 // were already unhooked by using information from detect_ptr_independence()
410 // and find_previous_store().
411
412 if (mem->is_MergeMem()) {
413 MergeMemNode* mmem = mem->as_MergeMem();
414 const TypePtr *tp = t_adr->is_ptr();
415
416 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
417 }
418
419 if (mem != old_mem) {
420 set_req_X(MemNode::Memory, mem, phase);
421 if (phase->type(mem) == Type::TOP) return NodeSentinel;
422 return this;
423 }
424
425 // let the subclass continue analyzing...
426 return nullptr;
427 }
428
429 // Helper function for proving some simple control dominations.
430 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
431 // Already assumes that 'dom' is available at 'sub', and that 'sub'
432 // is not a constant (dominated by the method's StartNode).
433 // Used by MemNode::find_previous_store to prove that the
434 // control input of a memory operation predates (dominates)
435 // an allocation it wants to look past.
436 // Returns 'DomResult::Dominate' if all control inputs of 'dom'
437 // dominate 'sub', 'DomResult::NotDominate' if not,
438 // and 'DomResult::EncounteredDeadCode' if we can't decide due to
439 // dead code, but at the end of IGVN, we know the definite result
440 // once the dead code is cleaned up.
441 Node::DomResult MemNode::maybe_all_controls_dominate(Node* dom, Node* sub) {
442 if (dom == nullptr || dom->is_top() || sub == nullptr || sub->is_top()) {
443 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
444 }
445
446 // Check 'dom'. Skip Proj and CatchProj nodes.
447 dom = dom->find_exact_control(dom);
448 if (dom == nullptr || dom->is_top()) {
449 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
450 }
451
452 if (dom == sub) {
453 // For the case when, for example, 'sub' is Initialize and the original
454 // 'dom' is Proj node of the 'sub'.
455 return DomResult::NotDominate;
456 }
457
458 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) {
459 return DomResult::Dominate;
460 }
461
462 // 'dom' dominates 'sub' if its control edge and control edges
463 // of all its inputs dominate or equal to sub's control edge.
464
465 // Currently 'sub' is either Allocate, Initialize or Start nodes.
466 // Or Region for the check in LoadNode::Ideal();
467 // 'sub' should have sub->in(0) != nullptr.
468 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
469 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
470
471 // Get control edge of 'sub'.
472 Node* orig_sub = sub;
473 sub = sub->find_exact_control(sub->in(0));
474 if (sub == nullptr || sub->is_top()) {
475 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
476 }
477
478 assert(sub->is_CFG(), "expecting control");
479
480 if (sub == dom) {
481 return DomResult::Dominate;
482 }
483
484 if (sub->is_Start() || sub->is_Root()) {
485 return DomResult::NotDominate;
486 }
487
488 {
489 // Check all control edges of 'dom'.
490
491 ResourceMark rm;
492 Node_List nlist;
493 Unique_Node_List dom_list;
494
495 dom_list.push(dom);
496 bool only_dominating_controls = false;
497
498 for (uint next = 0; next < dom_list.size(); next++) {
499 Node* n = dom_list.at(next);
500 if (n == orig_sub) {
501 return DomResult::NotDominate; // One of dom's inputs dominated by sub.
502 }
503 if (!n->is_CFG() && n->pinned()) {
504 // Check only own control edge for pinned non-control nodes.
505 n = n->find_exact_control(n->in(0));
506 if (n == nullptr || n->is_top()) {
507 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
508 }
509 assert(n->is_CFG(), "expecting control");
510 dom_list.push(n);
511 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
512 only_dominating_controls = true;
513 } else if (n->is_CFG()) {
514 DomResult dom_result = n->dominates(sub, nlist);
515 if (dom_result == DomResult::Dominate) {
516 only_dominating_controls = true;
517 } else {
518 return dom_result;
519 }
520 } else {
521 // First, own control edge.
522 Node* m = n->find_exact_control(n->in(0));
523 if (m != nullptr) {
524 if (m->is_top()) {
525 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
526 }
527 dom_list.push(m);
528 }
529 // Now, the rest of edges.
530 uint cnt = n->req();
531 for (uint i = 1; i < cnt; i++) {
532 m = n->find_exact_control(n->in(i));
533 if (m == nullptr || m->is_top()) {
534 continue;
535 }
536 dom_list.push(m);
537 }
538 }
539 }
540 return only_dominating_controls ? DomResult::Dominate : DomResult::NotDominate;
541 }
542 }
543
544 //---------------------detect_ptr_independence---------------------------------
545 // Used by MemNode::find_previous_store to prove that two base
546 // pointers are never equal.
547 // The pointers are accompanied by their associated allocations,
548 // if any, which have been previously discovered by the caller.
549 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
550 Node* p2, AllocateNode* a2,
551 PhaseTransform* phase) {
552 // Attempt to prove that these two pointers cannot be aliased.
553 // They may both manifestly be allocations, and they should differ.
554 // Or, if they are not both allocations, they can be distinct constants.
555 // Otherwise, one is an allocation and the other a pre-existing value.
556 if (a1 == nullptr && a2 == nullptr) { // neither an allocation
557 return (p1 != p2) && p1->is_Con() && p2->is_Con();
558 } else if (a1 != nullptr && a2 != nullptr) { // both allocations
559 return (a1 != a2);
560 } else if (a1 != nullptr) { // one allocation a1
561 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
562 return all_controls_dominate(p2, a1);
563 } else { //(a2 != null) // one allocation a2
564 return all_controls_dominate(p1, a2);
565 }
566 return false;
567 }
568
569
570 // Find an arraycopy ac that produces the memory state represented by parameter mem.
571 // Return ac if
572 // (a) can_see_stored_value=true and ac must have set the value for this load or if
573 // (b) can_see_stored_value=false and ac could have set the value for this load or if
574 // (c) can_see_stored_value=false and ac cannot have set the value for this load.
575 // In case (c) change the parameter mem to the memory input of ac to skip it
576 // when searching stored value.
577 // Otherwise return null.
578 Node* LoadNode::find_previous_arraycopy(PhaseValues* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
579 ArrayCopyNode* ac = find_array_copy_clone(ld_alloc, mem);
580 if (ac != nullptr) {
581 Node* ld_addp = in(MemNode::Address);
582 Node* src = ac->in(ArrayCopyNode::Src);
583 const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
584
585 // This is a load from a cloned array. The corresponding arraycopy ac must
586 // have set the value for the load and we can return ac but only if the load
587 // is known to be within bounds. This is checked below.
588 if (ary_t != nullptr && ld_addp->is_AddP()) {
589 Node* ld_offs = ld_addp->in(AddPNode::Offset);
590 BasicType ary_elem = ary_t->elem()->array_element_basic_type();
591 jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
592 jlong elemsize = type2aelembytes(ary_elem);
593
594 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
595 const TypeInt* sizetype = ary_t->size();
596
597 if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
598 // The load is known to be within bounds. It receives its value from ac.
599 return ac;
600 }
601 // The load is known to be out-of-bounds.
602 }
603 // The load could be out-of-bounds. It must not be hoisted but must remain
604 // dependent on the runtime range check. This is achieved by returning null.
605 } else if (mem->is_Proj() && mem->in(0) != nullptr && mem->in(0)->is_ArrayCopy()) {
606 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
607
608 if (ac->is_arraycopy_validated() ||
609 ac->is_copyof_validated() ||
610 ac->is_copyofrange_validated()) {
611 Node* ld_addp = in(MemNode::Address);
612 if (ld_addp->is_AddP()) {
613 Node* ld_base = ld_addp->in(AddPNode::Address);
614 Node* ld_offs = ld_addp->in(AddPNode::Offset);
615
616 Node* dest = ac->in(ArrayCopyNode::Dest);
617
618 if (dest == ld_base) {
619 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
620 assert(!ld_offs_t->empty(), "dead reference should be checked already");
621 // Take into account vector or unsafe access size
622 jlong ld_size_in_bytes = (jlong)memory_size();
623 jlong offset_hi = ld_offs_t->_hi + ld_size_in_bytes - 1;
624 offset_hi = MIN2(offset_hi, (jlong)(TypeX::MAX->_hi)); // Take care for overflow in 32-bit VM
625 if (ac->modifies(ld_offs_t->_lo, (intptr_t)offset_hi, phase, can_see_stored_value)) {
626 return ac;
627 }
628 if (!can_see_stored_value) {
629 mem = ac->in(TypeFunc::Memory);
630 return ac;
631 }
632 }
633 }
634 }
635 }
636 return nullptr;
637 }
638
639 ArrayCopyNode* MemNode::find_array_copy_clone(Node* ld_alloc, Node* mem) const {
640 if (mem->is_Proj() && mem->in(0) != nullptr && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
641 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
642 if (ld_alloc != nullptr) {
643 // Check if there is an array copy for a clone
644 Node* mb = mem->in(0);
645 ArrayCopyNode* ac = nullptr;
646 if (mb->in(0) != nullptr && mb->in(0)->is_Proj() &&
647 mb->in(0)->in(0) != nullptr && mb->in(0)->in(0)->is_ArrayCopy()) {
648 ac = mb->in(0)->in(0)->as_ArrayCopy();
649 } else {
650 // Step over GC barrier when ReduceInitialCardMarks is disabled
651 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
652 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
653
654 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
655 ac = control_proj_ac->in(0)->as_ArrayCopy();
656 }
657 }
658
659 if (ac != nullptr && ac->is_clonebasic()) {
660 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest));
661 if (alloc != nullptr && alloc == ld_alloc) {
662 return ac;
663 }
664 }
665 }
666 }
667 return nullptr;
668 }
669
670 // The logic for reordering loads and stores uses four steps:
671 // (a) Walk carefully past stores and initializations which we
672 // can prove are independent of this load.
673 // (b) Observe that the next memory state makes an exact match
674 // with self (load or store), and locate the relevant store.
675 // (c) Ensure that, if we were to wire self directly to the store,
676 // the optimizer would fold it up somehow.
677 // (d) Do the rewiring, and return, depending on some other part of
678 // the optimizer to fold up the load.
679 // This routine handles steps (a) and (b). Steps (c) and (d) are
680 // specific to loads and stores, so they are handled by the callers.
681 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
682 //
683 Node* MemNode::find_previous_store(PhaseValues* phase) {
684 Node* ctrl = in(MemNode::Control);
685 Node* adr = in(MemNode::Address);
686 intptr_t offset = 0;
687 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
688 AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
689
690 if (offset == Type::OffsetBot)
691 return nullptr; // cannot unalias unless there are precise offsets
692
693 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
694 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
695
696 intptr_t size_in_bytes = memory_size();
697
698 Node* mem = in(MemNode::Memory); // start searching here...
699
700 int cnt = 50; // Cycle limiter
701 for (;;) { // While we can dance past unrelated stores...
702 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
703
704 Node* prev = mem;
705 if (mem->is_Store()) {
706 Node* st_adr = mem->in(MemNode::Address);
707 intptr_t st_offset = 0;
708 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
709 if (st_base == nullptr)
710 break; // inscrutable pointer
711
712 // For raw accesses it's not enough to prove that constant offsets don't intersect.
713 // We need the bases to be the equal in order for the offset check to make sense.
714 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
715 break;
716 }
717
718 if (st_offset != offset && st_offset != Type::OffsetBot) {
719 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
720 assert(mem->as_Store()->memory_size() <= MAX_STORE, "");
721 if (st_offset >= offset + size_in_bytes ||
722 st_offset <= offset - MAX_STORE ||
723 st_offset <= offset - mem->as_Store()->memory_size()) {
724 // Success: The offsets are provably independent.
725 // (You may ask, why not just test st_offset != offset and be done?
726 // The answer is that stores of different sizes can co-exist
727 // in the same sequence of RawMem effects. We sometimes initialize
728 // a whole 'tile' of array elements with a single jint or jlong.)
729 mem = mem->in(MemNode::Memory);
730 continue; // (a) advance through independent store memory
731 }
732 }
733 if (st_base != base &&
734 detect_ptr_independence(base, alloc,
735 st_base,
736 AllocateNode::Ideal_allocation(st_base),
737 phase)) {
738 // Success: The bases are provably independent.
739 mem = mem->in(MemNode::Memory);
740 continue; // (a) advance through independent store memory
741 }
742
743 // (b) At this point, if the bases or offsets do not agree, we lose,
744 // since we have not managed to prove 'this' and 'mem' independent.
745 if (st_base == base && st_offset == offset) {
746 return mem; // let caller handle steps (c), (d)
747 }
748
749 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
750 InitializeNode* st_init = mem->in(0)->as_Initialize();
751 AllocateNode* st_alloc = st_init->allocation();
752 if (st_alloc == nullptr) {
753 break; // something degenerated
754 }
755 bool known_identical = false;
756 bool known_independent = false;
757 if (alloc == st_alloc) {
758 known_identical = true;
759 } else if (alloc != nullptr) {
760 known_independent = true;
761 } else if (all_controls_dominate(this, st_alloc)) {
762 known_independent = true;
763 }
764
765 if (known_independent) {
766 // The bases are provably independent: Either they are
767 // manifestly distinct allocations, or else the control
768 // of this load dominates the store's allocation.
769 int alias_idx = phase->C->get_alias_index(adr_type());
770 if (alias_idx == Compile::AliasIdxRaw) {
771 mem = st_alloc->in(TypeFunc::Memory);
772 } else {
773 mem = st_init->memory(alias_idx);
774 }
775 continue; // (a) advance through independent store memory
776 }
777
778 // (b) at this point, if we are not looking at a store initializing
779 // the same allocation we are loading from, we lose.
780 if (known_identical) {
781 // From caller, can_see_stored_value will consult find_captured_store.
782 return mem; // let caller handle steps (c), (d)
783 }
784
785 } else if (find_previous_arraycopy(phase, alloc, mem, false) != nullptr) {
786 if (prev != mem) {
787 // Found an arraycopy but it doesn't affect that load
788 continue;
789 }
790 // Found an arraycopy that may affect that load
791 return mem;
792 } else if (addr_t != nullptr && addr_t->is_known_instance_field()) {
793 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
794 if (mem->is_Proj() && mem->in(0)->is_Call()) {
795 // ArrayCopyNodes processed here as well.
796 CallNode *call = mem->in(0)->as_Call();
797 if (!call->may_modify(addr_t, phase)) {
798 mem = call->in(TypeFunc::Memory);
799 continue; // (a) advance through independent call memory
800 }
801 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
802 ArrayCopyNode* ac = nullptr;
803 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
804 break;
805 }
806 mem = mem->in(0)->in(TypeFunc::Memory);
807 continue; // (a) advance through independent MemBar memory
808 } else if (mem->is_ClearArray()) {
809 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
810 // (the call updated 'mem' value)
811 continue; // (a) advance through independent allocation memory
812 } else {
813 // Can not bypass initialization of the instance
814 // we are looking for.
815 return mem;
816 }
817 } else if (mem->is_MergeMem()) {
818 int alias_idx = phase->C->get_alias_index(adr_type());
819 mem = mem->as_MergeMem()->memory_at(alias_idx);
820 continue; // (a) advance through independent MergeMem memory
821 }
822 }
823
824 // Unless there is an explicit 'continue', we must bail out here,
825 // because 'mem' is an inscrutable memory state (e.g., a call).
826 break;
827 }
828
829 return nullptr; // bail out
830 }
831
832 //----------------------calculate_adr_type-------------------------------------
833 // Helper function. Notices when the given type of address hits top or bottom.
834 // Also, asserts a cross-check of the type against the expected address type.
835 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
836 if (t == Type::TOP) return nullptr; // does not touch memory any more?
837 #ifdef ASSERT
838 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = nullptr;
839 #endif
840 const TypePtr* tp = t->isa_ptr();
841 if (tp == nullptr) {
842 assert(cross_check == nullptr || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
843 return TypePtr::BOTTOM; // touches lots of memory
844 } else {
845 #ifdef ASSERT
846 // %%%% [phh] We don't check the alias index if cross_check is
847 // TypeRawPtr::BOTTOM. Needs to be investigated.
848 if (cross_check != nullptr &&
849 cross_check != TypePtr::BOTTOM &&
850 cross_check != TypeRawPtr::BOTTOM) {
851 // Recheck the alias index, to see if it has changed (due to a bug).
852 Compile* C = Compile::current();
853 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
854 "must stay in the original alias category");
855 // The type of the address must be contained in the adr_type,
856 // disregarding "null"-ness.
857 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
858 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
859 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
860 "real address must not escape from expected memory type");
861 }
862 #endif
863 return tp;
864 }
865 }
866
867 uint8_t MemNode::barrier_data(const Node* n) {
868 if (n->is_LoadStore()) {
869 return n->as_LoadStore()->barrier_data();
870 } else if (n->is_Mem()) {
871 return n->as_Mem()->barrier_data();
872 }
873 return 0;
874 }
875
876 //=============================================================================
877 // Should LoadNode::Ideal() attempt to remove control edges?
878 bool LoadNode::can_remove_control() const {
879 return !has_pinned_control_dependency();
880 }
881 uint LoadNode::size_of() const { return sizeof(*this); }
882 bool LoadNode::cmp(const Node &n) const {
883 LoadNode& load = (LoadNode &)n;
884 return Type::equals(_type, load._type) &&
885 _control_dependency == load._control_dependency &&
886 _mo == load._mo;
887 }
888 const Type *LoadNode::bottom_type() const { return _type; }
889 uint LoadNode::ideal_reg() const {
890 return _type->ideal_reg();
891 }
892
893 #ifndef PRODUCT
894 void LoadNode::dump_spec(outputStream *st) const {
895 MemNode::dump_spec(st);
896 if( !Verbose && !WizardMode ) {
897 // standard dump does this in Verbose and WizardMode
898 st->print(" #"); _type->dump_on(st);
899 }
900 if (!depends_only_on_test()) {
901 st->print(" (does not depend only on test, ");
902 if (control_dependency() == UnknownControl) {
903 st->print("unknown control");
904 } else if (control_dependency() == Pinned) {
905 st->print("pinned");
906 } else if (adr_type() == TypeRawPtr::BOTTOM) {
907 st->print("raw access");
908 } else {
909 st->print("unknown reason");
910 }
911 st->print(")");
912 }
913 }
914 #endif
915
916 #ifdef ASSERT
917 //----------------------------is_immutable_value-------------------------------
918 // Helper function to allow a raw load without control edge for some cases
919 bool LoadNode::is_immutable_value(Node* adr) {
920 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
921 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) {
922
923 jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1);
924 int offsets[] = {
925 in_bytes(JavaThread::osthread_offset()),
926 in_bytes(JavaThread::threadObj_offset()),
927 in_bytes(JavaThread::vthread_offset()),
928 in_bytes(JavaThread::scopedValueCache_offset()),
929 };
930
931 for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) {
932 if (offset == offsets[i]) {
933 return true;
934 }
935 }
936 }
937
938 return false;
939 }
940 #endif
941
942 //----------------------------LoadNode::make-----------------------------------
943 // Polymorphic factory method:
944 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo,
945 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
946 Compile* C = gvn.C;
947
948 // sanity check the alias category against the created node type
949 assert(!(adr_type->isa_oopptr() &&
950 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
951 "use LoadKlassNode instead");
952 assert(!(adr_type->isa_aryptr() &&
953 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
954 "use LoadRangeNode instead");
955 // Check control edge of raw loads
956 assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
957 // oop will be recorded in oop map if load crosses safepoint
958 rt->isa_oopptr() || is_immutable_value(adr),
959 "raw memory operations should have control edge");
960 LoadNode* load = nullptr;
961 switch (bt) {
962 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
963 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
964 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
965 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
966 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
967 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break;
968 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
969 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break;
970 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
971 case T_OBJECT:
972 case T_NARROWOOP:
973 #ifdef _LP64
974 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
975 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
976 } else
977 #endif
978 {
979 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
980 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
981 }
982 break;
983 default:
984 ShouldNotReachHere();
985 break;
986 }
987 assert(load != nullptr, "LoadNode should have been created");
988 if (unaligned) {
989 load->set_unaligned_access();
990 }
991 if (mismatched) {
992 load->set_mismatched_access();
993 }
994 if (unsafe) {
995 load->set_unsafe_access();
996 }
997 load->set_barrier_data(barrier_data);
998 if (load->Opcode() == Op_LoadN) {
999 Node* ld = gvn.transform(load);
1000 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
1001 }
1002
1003 return load;
1004 }
1005
1006 //------------------------------hash-------------------------------------------
1007 uint LoadNode::hash() const {
1008 // unroll addition of interesting fields
1009 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
1010 }
1011
1012 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
1013 if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) {
1014 bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile();
1015 bool is_stable_ary = FoldStableValues &&
1016 (tp != nullptr) && (tp->isa_aryptr() != nullptr) &&
1017 tp->isa_aryptr()->is_stable();
1018
1019 return (eliminate_boxing && non_volatile) || is_stable_ary || tp->is_inlinetypeptr();
1020 }
1021
1022 return false;
1023 }
1024
1025 LoadNode* LoadNode::pin_array_access_node() const {
1026 const TypePtr* adr_type = this->adr_type();
1027 if (adr_type != nullptr && adr_type->isa_aryptr()) {
1028 return clone_pinned();
1029 }
1030 return nullptr;
1031 }
1032
1033 // Is the value loaded previously stored by an arraycopy? If so return
1034 // a load node that reads from the source array so we may be able to
1035 // optimize out the ArrayCopy node later.
1036 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
1037 Node* ld_adr = in(MemNode::Address);
1038 intptr_t ld_off = 0;
1039 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
1040 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
1041 if (ac != nullptr) {
1042 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
1043
1044 Node* mem = ac->in(TypeFunc::Memory);
1045 Node* ctl = ac->in(0);
1046 Node* src = ac->in(ArrayCopyNode::Src);
1047
1048 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
1049 return nullptr;
1050 }
1051
1052 // load depends on the tests that validate the arraycopy
1053 LoadNode* ld = clone_pinned();
1054 Node* addp = in(MemNode::Address)->clone();
1055 if (ac->as_ArrayCopy()->is_clonebasic()) {
1056 assert(ld_alloc != nullptr, "need an alloc");
1057 assert(addp->is_AddP(), "address must be addp");
1058 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1059 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1060 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1061 addp->set_req(AddPNode::Base, src);
1062 addp->set_req(AddPNode::Address, src);
1063 } else {
1064 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1065 ac->as_ArrayCopy()->is_copyof_validated() ||
1066 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1067 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1068 addp->set_req(AddPNode::Base, src);
1069 addp->set_req(AddPNode::Address, src);
1070
1071 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1072 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
1073 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
1074
1075 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
1076 uint shift = ary_t->is_flat() ? ary_t->flat_log_elem_size() : exact_log2(type2aelembytes(ary_elem));
1077
1078 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1079 #ifdef _LP64
1080 diff = phase->transform(new ConvI2LNode(diff));
1081 #endif
1082 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1083
1084 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1085 addp->set_req(AddPNode::Offset, offset);
1086 }
1087 addp = phase->transform(addp);
1088 #ifdef ASSERT
1089 const TypePtr* adr_type = phase->type(addp)->is_ptr();
1090 ld->_adr_type = adr_type;
1091 #endif
1092 ld->set_req(MemNode::Address, addp);
1093 ld->set_req(0, ctl);
1094 ld->set_req(MemNode::Memory, mem);
1095 return ld;
1096 }
1097 return nullptr;
1098 }
1099
1100
1101 //---------------------------can_see_stored_value------------------------------
1102 // This routine exists to make sure this set of tests is done the same
1103 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
1104 // will change the graph shape in a way which makes memory alive twice at the
1105 // same time (uses the Oracle model of aliasing), then some
1106 // LoadXNode::Identity will fold things back to the equivalence-class model
1107 // of aliasing.
1108 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const {
1109 Node* ld_adr = in(MemNode::Address);
1110 intptr_t ld_off = 0;
1111 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1112 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base);
1113 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1114 Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr;
1115 // This is more general than load from boxing objects.
1116 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1117 uint alias_idx = atp->index();
1118 Node* result = nullptr;
1119 Node* current = st;
1120 // Skip through chains of MemBarNodes checking the MergeMems for
1121 // new states for the slice of this load. Stop once any other
1122 // kind of node is encountered. Loads from final memory can skip
1123 // through any kind of MemBar but normal loads shouldn't skip
1124 // through MemBarAcquire since the could allow them to move out of
1125 // a synchronized region. It is not safe to step over MemBarCPUOrder,
1126 // because alias info above them may be inaccurate (e.g., due to
1127 // mixed/mismatched unsafe accesses).
1128 bool is_final_mem = !atp->is_rewritable();
1129 while (current->is_Proj()) {
1130 int opc = current->in(0)->Opcode();
1131 if ((is_final_mem && (opc == Op_MemBarAcquire ||
1132 opc == Op_MemBarAcquireLock ||
1133 opc == Op_LoadFence)) ||
1134 opc == Op_MemBarRelease ||
1135 opc == Op_StoreFence ||
1136 opc == Op_MemBarReleaseLock ||
1137 opc == Op_MemBarStoreStore ||
1138 opc == Op_StoreStoreFence) {
1139 Node* mem = current->in(0)->in(TypeFunc::Memory);
1140 if (mem->is_MergeMem()) {
1141 MergeMemNode* merge = mem->as_MergeMem();
1142 Node* new_st = merge->memory_at(alias_idx);
1143 if (new_st == merge->base_memory()) {
1144 // Keep searching
1145 current = new_st;
1146 continue;
1147 }
1148 // Save the new memory state for the slice and fall through
1149 // to exit.
1150 result = new_st;
1151 }
1152 }
1153 break;
1154 }
1155 if (result != nullptr) {
1156 st = result;
1157 }
1158 }
1159
1160 // Loop around twice in the case Load -> Initialize -> Store.
1161 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1162 for (int trip = 0; trip <= 1; trip++) {
1163
1164 if (st->is_Store()) {
1165 Node* st_adr = st->in(MemNode::Address);
1166 if (st_adr != ld_adr) {
1167 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1168 intptr_t st_off = 0;
1169 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1170 if (ld_base == nullptr) return nullptr;
1171 if (st_base == nullptr) return nullptr;
1172 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr;
1173 if (ld_off != st_off) return nullptr;
1174 if (ld_off == Type::OffsetBot) return nullptr;
1175 // Same base, same offset.
1176 // Possible improvement for arrays: check index value instead of absolute offset.
1177
1178 // At this point we have proven something like this setup:
1179 // B = << base >>
1180 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1181 // S = StoreQ(AddP( B , #Off), V)
1182 // (Actually, we haven't yet proven the Q's are the same.)
1183 // In other words, we are loading from a casted version of
1184 // the same pointer-and-offset that we stored to.
1185 // Casted version may carry a dependency and it is respected.
1186 // Thus, we are able to replace L by V.
1187 }
1188 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1189 if (store_Opcode() != st->Opcode()) {
1190 return nullptr;
1191 }
1192 // LoadVector/StoreVector needs additional check to ensure the types match.
1193 if (st->is_StoreVector()) {
1194 const TypeVect* in_vt = st->as_StoreVector()->vect_type();
1195 const TypeVect* out_vt = is_Load() ? as_LoadVector()->vect_type() : as_StoreVector()->vect_type();
1196 if (in_vt != out_vt) {
1197 return nullptr;
1198 }
1199 }
1200 return st->in(MemNode::ValueIn);
1201 }
1202
1203 // A load from a freshly-created object always returns zero.
1204 // (This can happen after LoadNode::Ideal resets the load's memory input
1205 // to find_captured_store, which returned InitializeNode::zero_memory.)
1206 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1207 (st->in(0) == ld_alloc) &&
1208 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1209 // return a zero value for the load's basic type
1210 // (This is one of the few places where a generic PhaseTransform
1211 // can create new nodes. Think of it as lazily manifesting
1212 // virtually pre-existing constants.)
1213 Node* default_value = ld_alloc->in(AllocateNode::DefaultValue);
1214 if (default_value != nullptr) {
1215 return default_value;
1216 }
1217 assert(ld_alloc->in(AllocateNode::RawDefaultValue) == nullptr, "default value may not be null");
1218 if (memory_type() != T_VOID) {
1219 if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) {
1220 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1221 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1222 // by the ArrayCopyNode.
1223 return phase->zerocon(memory_type());
1224 }
1225 } else {
1226 // TODO: materialize all-zero vector constant
1227 assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1228 }
1229 }
1230
1231 // A load from an initialization barrier can match a captured store.
1232 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1233 InitializeNode* init = st->in(0)->as_Initialize();
1234 AllocateNode* alloc = init->allocation();
1235 if ((alloc != nullptr) && (alloc == ld_alloc)) {
1236 // examine a captured store value
1237 st = init->find_captured_store(ld_off, memory_size(), phase);
1238 if (st != nullptr) {
1239 continue; // take one more trip around
1240 }
1241 }
1242 }
1243
1244 // Load boxed value from result of valueOf() call is input parameter.
1245 if (this->is_Load() && ld_adr->is_AddP() &&
1246 (tp != nullptr) && tp->is_ptr_to_boxed_value()) {
1247 intptr_t ignore = 0;
1248 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1249 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1250 base = bs->step_over_gc_barrier(base);
1251 if (base != nullptr && base->is_Proj() &&
1252 base->as_Proj()->_con == TypeFunc::Parms &&
1253 base->in(0)->is_CallStaticJava() &&
1254 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1255 return base->in(0)->in(TypeFunc::Parms);
1256 }
1257 }
1258
1259 break;
1260 }
1261
1262 return nullptr;
1263 }
1264
1265 //----------------------is_instance_field_load_with_local_phi------------------
1266 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1267 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1268 in(Address)->is_AddP() ) {
1269 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1270 // Only instances and boxed values.
1271 if( t_oop != nullptr &&
1272 (t_oop->is_ptr_to_boxed_value() ||
1273 t_oop->is_known_instance_field()) &&
1274 t_oop->offset() != Type::OffsetBot &&
1275 t_oop->offset() != Type::OffsetTop) {
1276 return true;
1277 }
1278 }
1279 return false;
1280 }
1281
1282 //------------------------------Identity---------------------------------------
1283 // Loads are identity if previous store is to same address
1284 Node* LoadNode::Identity(PhaseGVN* phase) {
1285 // Loading from an InlineType? The InlineType has the values of
1286 // all fields as input. Look for the field with matching offset.
1287 Node* addr = in(Address);
1288 intptr_t offset;
1289 Node* base = AddPNode::Ideal_base_and_offset(addr, phase, offset);
1290 if (base != nullptr && base->is_InlineType() && offset > oopDesc::klass_offset_in_bytes()) {
1291 Node* value = base->as_InlineType()->field_value_by_offset((int)offset, true);
1292 if (value != nullptr) {
1293 if (Opcode() == Op_LoadN) {
1294 // Encode oop value if we are loading a narrow oop
1295 assert(!phase->type(value)->isa_narrowoop(), "should already be decoded");
1296 value = phase->transform(new EncodePNode(value, bottom_type()));
1297 }
1298 return value;
1299 }
1300 }
1301
1302 // If the previous store-maker is the right kind of Store, and the store is
1303 // to the same address, then we are equal to the value stored.
1304 Node* mem = in(Memory);
1305 Node* value = can_see_stored_value(mem, phase);
1306 if( value ) {
1307 // byte, short & char stores truncate naturally.
1308 // A load has to load the truncated value which requires
1309 // some sort of masking operation and that requires an
1310 // Ideal call instead of an Identity call.
1311 if (memory_size() < BytesPerInt) {
1312 // If the input to the store does not fit with the load's result type,
1313 // it must be truncated via an Ideal call.
1314 if (!phase->type(value)->higher_equal(phase->type(this)))
1315 return this;
1316 }
1317 // (This works even when value is a Con, but LoadNode::Value
1318 // usually runs first, producing the singleton type of the Con.)
1319 if (!has_pinned_control_dependency() || value->is_Con()) {
1320 return value;
1321 } else {
1322 return this;
1323 }
1324 }
1325
1326 if (has_pinned_control_dependency()) {
1327 return this;
1328 }
1329 // Search for an existing data phi which was generated before for the same
1330 // instance's field to avoid infinite generation of phis in a loop.
1331 Node *region = mem->in(0);
1332 if (is_instance_field_load_with_local_phi(region)) {
1333 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1334 int this_index = phase->C->get_alias_index(addr_t);
1335 int this_offset = addr_t->offset();
1336 int this_iid = addr_t->instance_id();
1337 if (!addr_t->is_known_instance() &&
1338 addr_t->is_ptr_to_boxed_value()) {
1339 // Use _idx of address base (could be Phi node) for boxed values.
1340 intptr_t ignore = 0;
1341 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1342 if (base == nullptr) {
1343 return this;
1344 }
1345 this_iid = base->_idx;
1346 }
1347 const Type* this_type = bottom_type();
1348 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1349 Node* phi = region->fast_out(i);
1350 if (phi->is_Phi() && phi != mem &&
1351 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1352 return phi;
1353 }
1354 }
1355 }
1356
1357 return this;
1358 }
1359
1360 // Construct an equivalent unsigned load.
1361 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1362 BasicType bt = T_ILLEGAL;
1363 const Type* rt = nullptr;
1364 switch (Opcode()) {
1365 case Op_LoadUB: return this;
1366 case Op_LoadUS: return this;
1367 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1368 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1369 default:
1370 assert(false, "no unsigned variant: %s", Name());
1371 return nullptr;
1372 }
1373 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1374 raw_adr_type(), rt, bt, _mo, _control_dependency,
1375 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1376 }
1377
1378 // Construct an equivalent signed load.
1379 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1380 BasicType bt = T_ILLEGAL;
1381 const Type* rt = nullptr;
1382 switch (Opcode()) {
1383 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1384 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1385 case Op_LoadB: // fall through
1386 case Op_LoadS: // fall through
1387 case Op_LoadI: // fall through
1388 case Op_LoadL: return this;
1389 default:
1390 assert(false, "no signed 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 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1399 BasicType bt = rt->basic_type();
1400 switch (Opcode()) {
1401 case Op_LoadI: return (bt == T_FLOAT);
1402 case Op_LoadL: return (bt == T_DOUBLE);
1403 case Op_LoadF: return (bt == T_INT);
1404 case Op_LoadD: return (bt == T_LONG);
1405
1406 default: return false;
1407 }
1408 }
1409
1410 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1411 BasicType bt = rt->basic_type();
1412 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1413 bool is_mismatched = is_mismatched_access();
1414 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1415 if (raw_type == nullptr) {
1416 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1417 }
1418 const int op = Opcode();
1419 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) ||
1420 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access());
1421 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1422 raw_adr_type(), rt, bt, _mo, _control_dependency,
1423 require_atomic_access, is_unaligned_access(), is_mismatched);
1424 }
1425
1426 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1427 BasicType bt = vt->basic_type();
1428 switch (Opcode()) {
1429 case Op_StoreI: return (bt == T_FLOAT);
1430 case Op_StoreL: return (bt == T_DOUBLE);
1431 case Op_StoreF: return (bt == T_INT);
1432 case Op_StoreD: return (bt == T_LONG);
1433
1434 default: return false;
1435 }
1436 }
1437
1438 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1439 BasicType bt = vt->basic_type();
1440 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1441 const int op = Opcode();
1442 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) ||
1443 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access());
1444 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1445 raw_adr_type(), val, bt, _mo, require_atomic_access);
1446
1447 bool is_mismatched = is_mismatched_access();
1448 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1449 if (raw_type == nullptr) {
1450 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1451 }
1452 if (is_mismatched) {
1453 st->set_mismatched_access();
1454 }
1455 return st;
1456 }
1457
1458 // We're loading from an object which has autobox behaviour.
1459 // If this object is result of a valueOf call we'll have a phi
1460 // merging a newly allocated object and a load from the cache.
1461 // We want to replace this load with the original incoming
1462 // argument to the valueOf call.
1463 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1464 assert(igvn->C->eliminate_boxing(), "sanity");
1465 intptr_t ignore = 0;
1466 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1467 if ((base == nullptr) || base->is_Phi()) {
1468 // Push the loads from the phi that comes from valueOf up
1469 // through it to allow elimination of the loads and the recovery
1470 // of the original value. It is done in split_through_phi().
1471 return nullptr;
1472 } else if (base->is_Load() ||
1473 (base->is_DecodeN() && base->in(1)->is_Load())) {
1474 // Eliminate the load of boxed value for integer types from the cache
1475 // array by deriving the value from the index into the array.
1476 // Capture the offset of the load and then reverse the computation.
1477
1478 // Get LoadN node which loads a boxing object from 'cache' array.
1479 if (base->is_DecodeN()) {
1480 base = base->in(1);
1481 }
1482 if (!base->in(Address)->is_AddP()) {
1483 return nullptr; // Complex address
1484 }
1485 AddPNode* address = base->in(Address)->as_AddP();
1486 Node* cache_base = address->in(AddPNode::Base);
1487 if ((cache_base != nullptr) && cache_base->is_DecodeN()) {
1488 // Get ConP node which is static 'cache' field.
1489 cache_base = cache_base->in(1);
1490 }
1491 if ((cache_base != nullptr) && cache_base->is_Con()) {
1492 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1493 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
1494 Node* elements[4];
1495 int shift = exact_log2(type2aelembytes(T_OBJECT));
1496 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1497 if (count > 0 && elements[0]->is_Con() &&
1498 (count == 1 ||
1499 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1500 elements[1]->in(2) == igvn->intcon(shift)))) {
1501 ciObjArray* array = base_type->const_oop()->as_obj_array();
1502 // Fetch the box object cache[0] at the base of the array and get its value
1503 ciInstance* box = array->obj_at(0)->as_instance();
1504 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1505 assert(ik->is_box_klass(), "sanity");
1506 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1507 if (ik->nof_nonstatic_fields() == 1) {
1508 // This should be true nonstatic_field_at requires calling
1509 // nof_nonstatic_fields so check it anyway
1510 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1511 BasicType bt = c.basic_type();
1512 // Only integer types have boxing cache.
1513 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1514 bt == T_BYTE || bt == T_SHORT ||
1515 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1516 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1517 if (cache_low != (int)cache_low) {
1518 return nullptr; // should not happen since cache is array indexed by value
1519 }
1520 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1521 if (offset != (int)offset) {
1522 return nullptr; // should not happen since cache is array indexed by value
1523 }
1524 // Add up all the offsets making of the address of the load
1525 Node* result = elements[0];
1526 for (int i = 1; i < count; i++) {
1527 result = igvn->transform(new AddXNode(result, elements[i]));
1528 }
1529 // Remove the constant offset from the address and then
1530 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1531 // remove the scaling of the offset to recover the original index.
1532 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1533 // Peel the shift off directly but wrap it in a dummy node
1534 // since Ideal can't return existing nodes
1535 igvn->_worklist.push(result); // remove dead node later
1536 result = new RShiftXNode(result->in(1), igvn->intcon(0));
1537 } else if (result->is_Add() && result->in(2)->is_Con() &&
1538 result->in(1)->Opcode() == Op_LShiftX &&
1539 result->in(1)->in(2) == igvn->intcon(shift)) {
1540 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1541 // but for boxing cache access we know that X<<Z will not overflow
1542 // (there is range check) so we do this optimizatrion by hand here.
1543 igvn->_worklist.push(result); // remove dead node later
1544 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1545 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1546 } else {
1547 result = new RShiftXNode(result, igvn->intcon(shift));
1548 }
1549 #ifdef _LP64
1550 if (bt != T_LONG) {
1551 result = new ConvL2INode(igvn->transform(result));
1552 }
1553 #else
1554 if (bt == T_LONG) {
1555 result = new ConvI2LNode(igvn->transform(result));
1556 }
1557 #endif
1558 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1559 // Need to preserve unboxing load type if it is unsigned.
1560 switch(this->Opcode()) {
1561 case Op_LoadUB:
1562 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1563 break;
1564 case Op_LoadUS:
1565 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1566 break;
1567 }
1568 return result;
1569 }
1570 }
1571 }
1572 }
1573 }
1574 return nullptr;
1575 }
1576
1577 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1578 Node* region = phi->in(0);
1579 if (region == nullptr) {
1580 return false; // Wait stable graph
1581 }
1582 uint cnt = phi->req();
1583 for (uint i = 1; i < cnt; i++) {
1584 Node* rc = region->in(i);
1585 if (rc == nullptr || phase->type(rc) == Type::TOP)
1586 return false; // Wait stable graph
1587 Node* in = phi->in(i);
1588 if (in == nullptr || phase->type(in) == Type::TOP)
1589 return false; // Wait stable graph
1590 }
1591 return true;
1592 }
1593
1594 //------------------------------split_through_phi------------------------------
1595 // Check whether a call to 'split_through_phi' would split this load through the
1596 // Phi *base*. This method is essentially a copy of the validations performed
1597 // by 'split_through_phi'. The first use of this method was in EA code as part
1598 // of simplification of allocation merges.
1599 // Some differences from original method (split_through_phi):
1600 // - If base->is_CastPP(): base = base->in(1)
1601 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) {
1602 Node* mem = in(Memory);
1603 Node* address = in(Address);
1604 intptr_t ignore = 0;
1605 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1606
1607 if (base->is_CastPP()) {
1608 base = base->in(1);
1609 }
1610
1611 if (req() > 3 || base == nullptr || !base->is_Phi()) {
1612 return false;
1613 }
1614
1615 if (!mem->is_Phi()) {
1616 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1617 return false;
1618 }
1619 } else if (base->in(0) != mem->in(0)) {
1620 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1621 return false;
1622 }
1623 }
1624
1625 return true;
1626 }
1627
1628 //------------------------------split_through_phi------------------------------
1629 // Split instance or boxed field load through Phi.
1630 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) {
1631 if (req() > 3) {
1632 assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs");
1633 // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account
1634 return nullptr;
1635 }
1636 Node* mem = in(Memory);
1637 Node* address = in(Address);
1638 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1639
1640 assert((t_oop != nullptr) &&
1641 (ignore_missing_instance_id ||
1642 t_oop->is_known_instance_field() ||
1643 t_oop->is_ptr_to_boxed_value()), "invalid conditions");
1644
1645 Compile* C = phase->C;
1646 intptr_t ignore = 0;
1647 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1648 bool base_is_phi = (base != nullptr) && base->is_Phi();
1649 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1650 (base != nullptr) && (base == address->in(AddPNode::Base)) &&
1651 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1652
1653 if (!((mem->is_Phi() || base_is_phi) &&
1654 (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) {
1655 return nullptr; // Neither memory or base are Phi
1656 }
1657
1658 if (mem->is_Phi()) {
1659 if (!stable_phi(mem->as_Phi(), phase)) {
1660 return nullptr; // Wait stable graph
1661 }
1662 uint cnt = mem->req();
1663 // Check for loop invariant memory.
1664 if (cnt == 3) {
1665 for (uint i = 1; i < cnt; i++) {
1666 Node* in = mem->in(i);
1667 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1668 if (m == mem) {
1669 if (i == 1) {
1670 // if the first edge was a loop, check second edge too.
1671 // If both are replaceable - we are in an infinite loop
1672 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1673 if (n == mem) {
1674 break;
1675 }
1676 }
1677 set_req(Memory, mem->in(cnt - i));
1678 return this; // made change
1679 }
1680 }
1681 }
1682 }
1683 if (base_is_phi) {
1684 if (!stable_phi(base->as_Phi(), phase)) {
1685 return nullptr; // Wait stable graph
1686 }
1687 uint cnt = base->req();
1688 // Check for loop invariant memory.
1689 if (cnt == 3) {
1690 for (uint i = 1; i < cnt; i++) {
1691 if (base->in(i) == base) {
1692 return nullptr; // Wait stable graph
1693 }
1694 }
1695 }
1696 }
1697
1698 // Split through Phi (see original code in loopopts.cpp).
1699 assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type");
1700
1701 // Do nothing here if Identity will find a value
1702 // (to avoid infinite chain of value phis generation).
1703 if (this != Identity(phase)) {
1704 return nullptr;
1705 }
1706
1707 // Select Region to split through.
1708 Node* region;
1709 DomResult dom_result = DomResult::Dominate;
1710 if (!base_is_phi) {
1711 assert(mem->is_Phi(), "sanity");
1712 region = mem->in(0);
1713 // Skip if the region dominates some control edge of the address.
1714 // We will check `dom_result` later.
1715 dom_result = MemNode::maybe_all_controls_dominate(address, region);
1716 } else if (!mem->is_Phi()) {
1717 assert(base_is_phi, "sanity");
1718 region = base->in(0);
1719 // Skip if the region dominates some control edge of the memory.
1720 // We will check `dom_result` later.
1721 dom_result = MemNode::maybe_all_controls_dominate(mem, region);
1722 } else if (base->in(0) != mem->in(0)) {
1723 assert(base_is_phi && mem->is_Phi(), "sanity");
1724 dom_result = MemNode::maybe_all_controls_dominate(mem, base->in(0));
1725 if (dom_result == DomResult::Dominate) {
1726 region = base->in(0);
1727 } else {
1728 dom_result = MemNode::maybe_all_controls_dominate(address, mem->in(0));
1729 if (dom_result == DomResult::Dominate) {
1730 region = mem->in(0);
1731 }
1732 // Otherwise we encountered a complex graph.
1733 }
1734 } else {
1735 assert(base->in(0) == mem->in(0), "sanity");
1736 region = mem->in(0);
1737 }
1738
1739 PhaseIterGVN* igvn = phase->is_IterGVN();
1740 if (dom_result != DomResult::Dominate) {
1741 if (dom_result == DomResult::EncounteredDeadCode) {
1742 // There is some dead code which eventually will be removed in IGVN.
1743 // Once this is the case, we get an unambiguous dominance result.
1744 // Push the node to the worklist again until the dead code is removed.
1745 igvn->_worklist.push(this);
1746 }
1747 return nullptr;
1748 }
1749
1750 Node* phi = nullptr;
1751 const Type* this_type = this->bottom_type();
1752 if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) {
1753 int this_index = C->get_alias_index(t_oop);
1754 int this_offset = t_oop->offset();
1755 int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx;
1756 phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset);
1757 } else if (ignore_missing_instance_id) {
1758 phi = new PhiNode(region, this_type, nullptr, mem->_idx);
1759 } else {
1760 return nullptr;
1761 }
1762
1763 for (uint i = 1; i < region->req(); i++) {
1764 Node* x;
1765 Node* the_clone = nullptr;
1766 Node* in = region->in(i);
1767 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1768 in != nullptr && in->is_OuterStripMinedLoop()) {
1769 // No node should go in the outer strip mined loop
1770 in = in->in(LoopNode::EntryControl);
1771 }
1772 if (in == nullptr || in == C->top()) {
1773 x = C->top(); // Dead path? Use a dead data op
1774 } else {
1775 x = this->clone(); // Else clone up the data op
1776 the_clone = x; // Remember for possible deletion.
1777 // Alter data node to use pre-phi inputs
1778 if (this->in(0) == region) {
1779 x->set_req(0, in);
1780 } else {
1781 x->set_req(0, nullptr);
1782 }
1783 if (mem->is_Phi() && (mem->in(0) == region)) {
1784 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1785 }
1786 if (address->is_Phi() && address->in(0) == region) {
1787 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1788 }
1789 if (base_is_phi && (base->in(0) == region)) {
1790 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1791 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1792 x->set_req(Address, adr_x);
1793 }
1794 }
1795 // Check for a 'win' on some paths
1796 const Type *t = x->Value(igvn);
1797
1798 bool singleton = t->singleton();
1799
1800 // See comments in PhaseIdealLoop::split_thru_phi().
1801 if (singleton && t == Type::TOP) {
1802 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1803 }
1804
1805 if (singleton) {
1806 x = igvn->makecon(t);
1807 } else {
1808 // We now call Identity to try to simplify the cloned node.
1809 // Note that some Identity methods call phase->type(this).
1810 // Make sure that the type array is big enough for
1811 // our new node, even though we may throw the node away.
1812 // (This tweaking with igvn only works because x is a new node.)
1813 igvn->set_type(x, t);
1814 // If x is a TypeNode, capture any more-precise type permanently into Node
1815 // otherwise it will be not updated during igvn->transform since
1816 // igvn->type(x) is set to x->Value() already.
1817 x->raise_bottom_type(t);
1818 Node* y = x->Identity(igvn);
1819 if (y != x) {
1820 x = y;
1821 } else {
1822 y = igvn->hash_find_insert(x);
1823 if (y) {
1824 x = y;
1825 } else {
1826 // Else x is a new node we are keeping
1827 // We do not need register_new_node_with_optimizer
1828 // because set_type has already been called.
1829 igvn->_worklist.push(x);
1830 }
1831 }
1832 }
1833 if (x != the_clone && the_clone != nullptr) {
1834 igvn->remove_dead_node(the_clone);
1835 }
1836 phi->set_req(i, x);
1837 }
1838 // Record Phi
1839 igvn->register_new_node_with_optimizer(phi);
1840 return phi;
1841 }
1842
1843 AllocateNode* LoadNode::is_new_object_mark_load() const {
1844 if (Opcode() == Op_LoadX) {
1845 Node* address = in(MemNode::Address);
1846 AllocateNode* alloc = AllocateNode::Ideal_allocation(address);
1847 Node* mem = in(MemNode::Memory);
1848 if (alloc != nullptr && mem->is_Proj() &&
1849 mem->in(0) != nullptr &&
1850 mem->in(0) == alloc->initialization() &&
1851 alloc->initialization()->proj_out_or_null(0) != nullptr) {
1852 return alloc;
1853 }
1854 }
1855 return nullptr;
1856 }
1857
1858
1859 //------------------------------Ideal------------------------------------------
1860 // If the load is from Field memory and the pointer is non-null, it might be possible to
1861 // zero out the control input.
1862 // If the offset is constant and the base is an object allocation,
1863 // try to hook me up to the exact initializing store.
1864 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1865 if (has_pinned_control_dependency()) {
1866 return nullptr;
1867 }
1868 Node* p = MemNode::Ideal_common(phase, can_reshape);
1869 if (p) return (p == NodeSentinel) ? nullptr : p;
1870
1871 Node* ctrl = in(MemNode::Control);
1872 Node* address = in(MemNode::Address);
1873 bool progress = false;
1874
1875 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1876 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1877
1878 // Skip up past a SafePoint control. Cannot do this for Stores because
1879 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1880 if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint &&
1881 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1882 !addr_mark &&
1883 (depends_only_on_test() || has_unknown_control_dependency())) {
1884 ctrl = ctrl->in(0);
1885 set_req(MemNode::Control,ctrl);
1886 progress = true;
1887 }
1888
1889 intptr_t ignore = 0;
1890 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1891 if (base != nullptr
1892 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1893 // Check for useless control edge in some common special cases
1894 if (in(MemNode::Control) != nullptr
1895 && can_remove_control()
1896 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1897 && all_controls_dominate(base, phase->C->start())) {
1898 // A method-invariant, non-null address (constant or 'this' argument).
1899 set_req(MemNode::Control, nullptr);
1900 progress = true;
1901 }
1902 }
1903
1904 Node* mem = in(MemNode::Memory);
1905 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1906
1907 if (can_reshape && (addr_t != nullptr)) {
1908 // try to optimize our memory input
1909 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1910 if (opt_mem != mem) {
1911 set_req_X(MemNode::Memory, opt_mem, phase);
1912 if (phase->type( opt_mem ) == Type::TOP) return nullptr;
1913 return this;
1914 }
1915 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1916 if ((t_oop != nullptr) &&
1917 (t_oop->is_known_instance_field() ||
1918 t_oop->is_ptr_to_boxed_value())) {
1919 PhaseIterGVN *igvn = phase->is_IterGVN();
1920 assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true");
1921 if (igvn->_worklist.member(opt_mem)) {
1922 // Delay this transformation until memory Phi is processed.
1923 igvn->_worklist.push(this);
1924 return nullptr;
1925 }
1926 // Split instance field load through Phi.
1927 Node* result = split_through_phi(phase);
1928 if (result != nullptr) return result;
1929
1930 if (t_oop->is_ptr_to_boxed_value()) {
1931 Node* result = eliminate_autobox(igvn);
1932 if (result != nullptr) return result;
1933 }
1934 }
1935 }
1936
1937 // Is there a dominating load that loads the same value? Leave
1938 // anything that is not a load of a field/array element (like
1939 // barriers etc.) alone
1940 if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) {
1941 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1942 Node *use = mem->fast_out(i);
1943 if (use != this &&
1944 use->Opcode() == Opcode() &&
1945 use->in(0) != nullptr &&
1946 use->in(0) != in(0) &&
1947 use->in(Address) == in(Address)) {
1948 Node* ctl = in(0);
1949 for (int i = 0; i < 10 && ctl != nullptr; i++) {
1950 ctl = IfNode::up_one_dom(ctl);
1951 if (ctl == use->in(0)) {
1952 set_req(0, use->in(0));
1953 return this;
1954 }
1955 }
1956 }
1957 }
1958 }
1959
1960 // Check for prior store with a different base or offset; make Load
1961 // independent. Skip through any number of them. Bail out if the stores
1962 // are in an endless dead cycle and report no progress. This is a key
1963 // transform for Reflection. However, if after skipping through the Stores
1964 // we can't then fold up against a prior store do NOT do the transform as
1965 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1966 // array memory alive twice: once for the hoisted Load and again after the
1967 // bypassed Store. This situation only works if EVERYBODY who does
1968 // anti-dependence work knows how to bypass. I.e. we need all
1969 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1970 // the alias index stuff. So instead, peek through Stores and IFF we can
1971 // fold up, do so.
1972 Node* prev_mem = find_previous_store(phase);
1973 if (prev_mem != nullptr) {
1974 Node* value = can_see_arraycopy_value(prev_mem, phase);
1975 if (value != nullptr) {
1976 return value;
1977 }
1978 }
1979 // Steps (a), (b): Walk past independent stores to find an exact match.
1980 if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) {
1981 // (c) See if we can fold up on the spot, but don't fold up here.
1982 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1983 // just return a prior value, which is done by Identity calls.
1984 if (can_see_stored_value(prev_mem, phase)) {
1985 // Make ready for step (d):
1986 set_req_X(MemNode::Memory, prev_mem, phase);
1987 return this;
1988 }
1989 }
1990
1991 return progress ? this : nullptr;
1992 }
1993
1994 // Helper to recognize certain Klass fields which are invariant across
1995 // some group of array types (e.g., int[] or all T[] where T < Object).
1996 const Type*
1997 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1998 ciKlass* klass) const {
1999 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
2000 // The field is Klass::_modifier_flags. Return its (constant) value.
2001 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
2002 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
2003 return TypeInt::make(klass->modifier_flags());
2004 }
2005 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
2006 // The field is Klass::_access_flags. Return its (constant) value.
2007 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
2008 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
2009 return TypeInt::make(klass->access_flags());
2010 }
2011 if (tkls->offset() == in_bytes(Klass::misc_flags_offset())) {
2012 // The field is Klass::_misc_flags. Return its (constant) value.
2013 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
2014 assert(this->Opcode() == Op_LoadUB, "must load an unsigned byte from _misc_flags");
2015 return TypeInt::make(klass->misc_flags());
2016 }
2017 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
2018 // The field is Klass::_layout_helper. Return its constant value if known.
2019 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
2020 return TypeInt::make(klass->layout_helper());
2021 }
2022
2023 // No match.
2024 return nullptr;
2025 }
2026
2027 //------------------------------Value-----------------------------------------
2028 const Type* LoadNode::Value(PhaseGVN* phase) const {
2029 // Either input is TOP ==> the result is TOP
2030 Node* mem = in(MemNode::Memory);
2031 const Type *t1 = phase->type(mem);
2032 if (t1 == Type::TOP) return Type::TOP;
2033 Node* adr = in(MemNode::Address);
2034 const TypePtr* tp = phase->type(adr)->isa_ptr();
2035 if (tp == nullptr || tp->empty()) return Type::TOP;
2036 int off = tp->offset();
2037 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
2038 Compile* C = phase->C;
2039
2040 // Try to guess loaded type from pointer type
2041 if (tp->isa_aryptr()) {
2042 const TypeAryPtr* ary = tp->is_aryptr();
2043 const Type* t = ary->elem();
2044
2045 // Determine whether the reference is beyond the header or not, by comparing
2046 // the offset against the offset of the start of the array's data.
2047 // Different array types begin at slightly different offsets (12 vs. 16).
2048 // We choose T_BYTE as an example base type that is least restrictive
2049 // as to alignment, which will therefore produce the smallest
2050 // possible base offset.
2051 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2052 const bool off_beyond_header = (off >= min_base_off);
2053
2054 // Try to constant-fold a stable array element.
2055 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
2056 // Make sure the reference is not into the header and the offset is constant
2057 ciObject* aobj = ary->const_oop();
2058 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
2059 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
2060 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
2061 stable_dimension,
2062 memory_type(), is_unsigned());
2063 if (con_type != nullptr) {
2064 return con_type;
2065 }
2066 }
2067 }
2068
2069 // Don't do this for integer types. There is only potential profit if
2070 // the element type t is lower than _type; that is, for int types, if _type is
2071 // more restrictive than t. This only happens here if one is short and the other
2072 // char (both 16 bits), and in those cases we've made an intentional decision
2073 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2074 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2075 //
2076 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2077 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
2078 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2079 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
2080 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2081 // In fact, that could have been the original type of p1, and p1 could have
2082 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2083 // expression (LShiftL quux 3) independently optimized to the constant 8.
2084 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2085 && (_type->isa_vect() == nullptr)
2086 && !ary->is_flat()
2087 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2088 // t might actually be lower than _type, if _type is a unique
2089 // concrete subclass of abstract class t.
2090 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
2091 const Type* jt = t->join_speculative(_type);
2092 // In any case, do not allow the join, per se, to empty out the type.
2093 if (jt->empty() && !t->empty()) {
2094 // This can happen if a interface-typed array narrows to a class type.
2095 jt = _type;
2096 }
2097 #ifdef ASSERT
2098 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2099 // The pointers in the autobox arrays are always non-null
2100 Node* base = adr->in(AddPNode::Base);
2101 if ((base != nullptr) && base->is_DecodeN()) {
2102 // Get LoadN node which loads IntegerCache.cache field
2103 base = base->in(1);
2104 }
2105 if ((base != nullptr) && base->is_Con()) {
2106 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2107 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2108 // It could be narrow oop
2109 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2110 }
2111 }
2112 }
2113 #endif
2114 return jt;
2115 }
2116 }
2117 } else if (tp->base() == Type::InstPtr) {
2118 assert( off != Type::OffsetBot ||
2119 // arrays can be cast to Objects
2120 !tp->isa_instptr() ||
2121 tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
2122 // Default value load
2123 tp->is_instptr()->instance_klass() == ciEnv::current()->Class_klass() ||
2124 // unsafe field access may not have a constant offset
2125 C->has_unsafe_access(),
2126 "Field accesses must be precise" );
2127 // For oop loads, we expect the _type to be precise.
2128
2129 const TypeInstPtr* tinst = tp->is_instptr();
2130 BasicType bt = memory_type();
2131
2132 // Optimize loads from constant fields.
2133 ciObject* const_oop = tinst->const_oop();
2134 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2135 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt);
2136 if (con_type != nullptr) {
2137 return con_type;
2138 }
2139 }
2140 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2141 assert(off != Type::OffsetBot ||
2142 !tp->isa_instklassptr() ||
2143 // arrays can be cast to Objects
2144 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2145 // also allow array-loading from the primary supertype
2146 // array during subtype checks
2147 Opcode() == Op_LoadKlass,
2148 "Field accesses must be precise");
2149 // For klass/static loads, we expect the _type to be precise
2150 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2151 /* With mirrors being an indirect in the Klass*
2152 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2153 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2154 *
2155 * So check the type and klass of the node before the LoadP.
2156 */
2157 Node* adr2 = adr->in(MemNode::Address);
2158 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2159 if (tkls != nullptr && !StressReflectiveCode) {
2160 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2161 ciKlass* klass = tkls->exact_klass();
2162 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2163 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2164 return TypeInstPtr::make(klass->java_mirror());
2165 }
2166 }
2167 }
2168
2169 const TypeKlassPtr *tkls = tp->isa_klassptr();
2170 if (tkls != nullptr) {
2171 if (tkls->is_loaded() && tkls->klass_is_exact()) {
2172 ciKlass* klass = tkls->exact_klass();
2173 // We are loading a field from a Klass metaobject whose identity
2174 // is known at compile time (the type is "exact" or "precise").
2175 // Check for fields we know are maintained as constants by the VM.
2176 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2177 // The field is Klass::_super_check_offset. Return its (constant) value.
2178 // (Folds up type checking code.)
2179 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2180 return TypeInt::make(klass->super_check_offset());
2181 }
2182 // Compute index into primary_supers array
2183 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2184 // Check for overflowing; use unsigned compare to handle the negative case.
2185 if( depth < ciKlass::primary_super_limit() ) {
2186 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2187 // (Folds up type checking code.)
2188 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2189 ciKlass *ss = klass->super_of_depth(depth);
2190 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2191 }
2192 const Type* aift = load_array_final_field(tkls, klass);
2193 if (aift != nullptr) return aift;
2194 }
2195
2196 // We can still check if we are loading from the primary_supers array at a
2197 // shallow enough depth. Even though the klass is not exact, entries less
2198 // than or equal to its super depth are correct.
2199 if (tkls->is_loaded()) {
2200 ciKlass* klass = nullptr;
2201 if (tkls->isa_instklassptr()) {
2202 klass = tkls->is_instklassptr()->instance_klass();
2203 } else {
2204 int dims;
2205 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims);
2206 if (inner->isa_instklassptr()) {
2207 klass = inner->is_instklassptr()->instance_klass();
2208 klass = ciObjArrayKlass::make(klass, dims);
2209 }
2210 }
2211 if (klass != nullptr) {
2212 // Compute index into primary_supers array
2213 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2214 // Check for overflowing; use unsigned compare to handle the negative case.
2215 if (depth < ciKlass::primary_super_limit() &&
2216 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case
2217 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2218 // (Folds up type checking code.)
2219 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2220 ciKlass *ss = klass->super_of_depth(depth);
2221 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2222 }
2223 }
2224 }
2225
2226 // If the type is enough to determine that the thing is not an array,
2227 // we can give the layout_helper a positive interval type.
2228 // This will help short-circuit some reflective code.
2229 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) &&
2230 tkls->isa_instklassptr() && // not directly typed as an array
2231 !tkls->is_instklassptr()->instance_klass()->is_java_lang_Object() // not the supertype of all T[] and specifically not Serializable & Cloneable
2232 ) {
2233 // Note: When interfaces are reliable, we can narrow the interface
2234 // test to (klass != Serializable && klass != Cloneable).
2235 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2236 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2237 // The key property of this type is that it folds up tests
2238 // for array-ness, since it proves that the layout_helper is positive.
2239 // Thus, a generic value like the basic object layout helper works fine.
2240 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2241 }
2242 }
2243
2244 // If we are loading from a freshly-allocated object, produce a zero,
2245 // if the load is provably beyond the header of the object.
2246 // (Also allow a variable load from a fresh array to produce zero.)
2247 const TypeOopPtr *tinst = tp->isa_oopptr();
2248 bool is_instance = (tinst != nullptr) && tinst->is_known_instance_field();
2249 bool is_boxed_value = (tinst != nullptr) && tinst->is_ptr_to_boxed_value();
2250 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
2251 Node* value = can_see_stored_value(mem,phase);
2252 if (value != nullptr && value->is_Con()) {
2253 assert(value->bottom_type()->higher_equal(_type),"sanity");
2254 return value->bottom_type();
2255 }
2256 }
2257
2258 bool is_vect = (_type->isa_vect() != nullptr);
2259 if (is_instance && !is_vect) {
2260 // If we have an instance type and our memory input is the
2261 // programs's initial memory state, there is no matching store,
2262 // so just return a zero of the appropriate type -
2263 // except if it is vectorized - then we have no zero constant.
2264 Node *mem = in(MemNode::Memory);
2265 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2266 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2267 return Type::get_zero_type(_type->basic_type());
2268 }
2269 }
2270 Node* alloc = is_new_object_mark_load();
2271 if (alloc != nullptr) {
2272 if (EnableValhalla) {
2273 // The mark word may contain property bits (inline, flat, null-free)
2274 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2275 const TypeKlassPtr* tkls = phase->type(klass_node)->isa_klassptr();
2276 if (tkls != nullptr && tkls->is_loaded() && tkls->klass_is_exact()) {
2277 return TypeX::make(tkls->exact_klass()->prototype_header().value());
2278 }
2279 } else {
2280 return TypeX::make(markWord::prototype().value());
2281 }
2282 }
2283
2284 return _type;
2285 }
2286
2287 //------------------------------match_edge-------------------------------------
2288 // Do we Match on this edge index or not? Match only the address.
2289 uint LoadNode::match_edge(uint idx) const {
2290 return idx == MemNode::Address;
2291 }
2292
2293 //--------------------------LoadBNode::Ideal--------------------------------------
2294 //
2295 // If the previous store is to the same address as this load,
2296 // and the value stored was larger than a byte, replace this load
2297 // with the value stored truncated to a byte. If no truncation is
2298 // needed, the replacement is done in LoadNode::Identity().
2299 //
2300 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2301 Node* mem = in(MemNode::Memory);
2302 Node* value = can_see_stored_value(mem,phase);
2303 if (value != nullptr) {
2304 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2305 if (narrow != value) {
2306 return narrow;
2307 }
2308 }
2309 // Identity call will handle the case where truncation is not needed.
2310 return LoadNode::Ideal(phase, can_reshape);
2311 }
2312
2313 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2314 Node* mem = in(MemNode::Memory);
2315 Node* value = can_see_stored_value(mem,phase);
2316 if (value != nullptr && value->is_Con() &&
2317 !value->bottom_type()->higher_equal(_type)) {
2318 // If the input to the store does not fit with the load's result type,
2319 // it must be truncated. We can't delay until Ideal call since
2320 // a singleton Value is needed for split_thru_phi optimization.
2321 int con = value->get_int();
2322 return TypeInt::make((con << 24) >> 24);
2323 }
2324 return LoadNode::Value(phase);
2325 }
2326
2327 //--------------------------LoadUBNode::Ideal-------------------------------------
2328 //
2329 // If the previous store is to the same address as this load,
2330 // and the value stored was larger than a byte, replace this load
2331 // with the value stored truncated to a byte. If no truncation is
2332 // needed, the replacement is done in LoadNode::Identity().
2333 //
2334 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2335 Node* mem = in(MemNode::Memory);
2336 Node* value = can_see_stored_value(mem, phase);
2337 if (value != nullptr) {
2338 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2339 if (narrow != value) {
2340 return narrow;
2341 }
2342 }
2343 // Identity call will handle the case where truncation is not needed.
2344 return LoadNode::Ideal(phase, can_reshape);
2345 }
2346
2347 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2348 Node* mem = in(MemNode::Memory);
2349 Node* value = can_see_stored_value(mem,phase);
2350 if (value != nullptr && value->is_Con() &&
2351 !value->bottom_type()->higher_equal(_type)) {
2352 // If the input to the store does not fit with the load's result type,
2353 // it must be truncated. We can't delay until Ideal call since
2354 // a singleton Value is needed for split_thru_phi optimization.
2355 int con = value->get_int();
2356 return TypeInt::make(con & 0xFF);
2357 }
2358 return LoadNode::Value(phase);
2359 }
2360
2361 //--------------------------LoadUSNode::Ideal-------------------------------------
2362 //
2363 // If the previous store is to the same address as this load,
2364 // and the value stored was larger than a char, replace this load
2365 // with the value stored truncated to a char. If no truncation is
2366 // needed, the replacement is done in LoadNode::Identity().
2367 //
2368 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2369 Node* mem = in(MemNode::Memory);
2370 Node* value = can_see_stored_value(mem,phase);
2371 if (value != nullptr) {
2372 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2373 if (narrow != value) {
2374 return narrow;
2375 }
2376 }
2377 // Identity call will handle the case where truncation is not needed.
2378 return LoadNode::Ideal(phase, can_reshape);
2379 }
2380
2381 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2382 Node* mem = in(MemNode::Memory);
2383 Node* value = can_see_stored_value(mem,phase);
2384 if (value != nullptr && value->is_Con() &&
2385 !value->bottom_type()->higher_equal(_type)) {
2386 // If the input to the store does not fit with the load's result type,
2387 // it must be truncated. We can't delay until Ideal call since
2388 // a singleton Value is needed for split_thru_phi optimization.
2389 int con = value->get_int();
2390 return TypeInt::make(con & 0xFFFF);
2391 }
2392 return LoadNode::Value(phase);
2393 }
2394
2395 //--------------------------LoadSNode::Ideal--------------------------------------
2396 //
2397 // If the previous store is to the same address as this load,
2398 // and the value stored was larger than a short, replace this load
2399 // with the value stored truncated to a short. If no truncation is
2400 // needed, the replacement is done in LoadNode::Identity().
2401 //
2402 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2403 Node* mem = in(MemNode::Memory);
2404 Node* value = can_see_stored_value(mem,phase);
2405 if (value != nullptr) {
2406 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2407 if (narrow != value) {
2408 return narrow;
2409 }
2410 }
2411 // Identity call will handle the case where truncation is not needed.
2412 return LoadNode::Ideal(phase, can_reshape);
2413 }
2414
2415 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2416 Node* mem = in(MemNode::Memory);
2417 Node* value = can_see_stored_value(mem,phase);
2418 if (value != nullptr && value->is_Con() &&
2419 !value->bottom_type()->higher_equal(_type)) {
2420 // If the input to the store does not fit with the load's result type,
2421 // it must be truncated. We can't delay until Ideal call since
2422 // a singleton Value is needed for split_thru_phi optimization.
2423 int con = value->get_int();
2424 return TypeInt::make((con << 16) >> 16);
2425 }
2426 return LoadNode::Value(phase);
2427 }
2428
2429 //=============================================================================
2430 //----------------------------LoadKlassNode::make------------------------------
2431 // Polymorphic factory method:
2432 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at,
2433 const TypeKlassPtr* tk) {
2434 // sanity check the alias category against the created node type
2435 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2436 assert(adr_type != nullptr, "expecting TypeKlassPtr");
2437 #ifdef _LP64
2438 if (adr_type->is_ptr_to_narrowklass()) {
2439 assert(UseCompressedClassPointers, "no compressed klasses");
2440 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2441 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2442 }
2443 #endif
2444 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2445 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2446 }
2447
2448 //------------------------------Value------------------------------------------
2449 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2450 return klass_value_common(phase);
2451 }
2452
2453 // In most cases, LoadKlassNode does not have the control input set. If the control
2454 // input is set, it must not be removed (by LoadNode::Ideal()).
2455 bool LoadKlassNode::can_remove_control() const {
2456 return false;
2457 }
2458
2459 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2460 // Either input is TOP ==> the result is TOP
2461 const Type *t1 = phase->type( in(MemNode::Memory) );
2462 if (t1 == Type::TOP) return Type::TOP;
2463 Node *adr = in(MemNode::Address);
2464 const Type *t2 = phase->type( adr );
2465 if (t2 == Type::TOP) return Type::TOP;
2466 const TypePtr *tp = t2->is_ptr();
2467 if (TypePtr::above_centerline(tp->ptr()) ||
2468 tp->ptr() == TypePtr::Null) return Type::TOP;
2469
2470 // Return a more precise klass, if possible
2471 const TypeInstPtr *tinst = tp->isa_instptr();
2472 if (tinst != nullptr) {
2473 ciInstanceKlass* ik = tinst->instance_klass();
2474 int offset = tinst->offset();
2475 if (ik == phase->C->env()->Class_klass()
2476 && (offset == java_lang_Class::klass_offset() ||
2477 offset == java_lang_Class::array_klass_offset())) {
2478 // We are loading a special hidden field from a Class mirror object,
2479 // the field which points to the VM's Klass metaobject.
2480 bool is_null_free_array = false;
2481 ciType* t = tinst->java_mirror_type(&is_null_free_array);
2482 // java_mirror_type returns non-null for compile-time Class constants.
2483 if (t != nullptr) {
2484 // constant oop => constant klass
2485 if (offset == java_lang_Class::array_klass_offset()) {
2486 if (t->is_void()) {
2487 // We cannot create a void array. Since void is a primitive type return null
2488 // klass. Users of this result need to do a null check on the returned klass.
2489 return TypePtr::NULL_PTR;
2490 }
2491 const TypeKlassPtr* tklass = TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);
2492 if (is_null_free_array) {
2493 tklass = tklass->is_aryklassptr()->cast_to_null_free();
2494 }
2495 return tklass;
2496 }
2497 if (!t->is_klass()) {
2498 // a primitive Class (e.g., int.class) has null for a klass field
2499 return TypePtr::NULL_PTR;
2500 }
2501 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2502 const TypeKlassPtr* tklass = TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);
2503 if (is_null_free_array) {
2504 tklass = tklass->is_aryklassptr()->cast_to_null_free();
2505 }
2506 return tklass;
2507 }
2508 // non-constant mirror, so we can't tell what's going on
2509 }
2510 if (!tinst->is_loaded())
2511 return _type; // Bail out if not loaded
2512 if (offset == oopDesc::klass_offset_in_bytes()) {
2513 return tinst->as_klass_type(true);
2514 }
2515 }
2516
2517 // Check for loading klass from an array
2518 const TypeAryPtr* tary = tp->isa_aryptr();
2519 if (tary != nullptr &&
2520 tary->offset() == oopDesc::klass_offset_in_bytes()) {
2521 return tary->as_klass_type(true);
2522 }
2523
2524 // Check for loading klass from an array klass
2525 const TypeKlassPtr *tkls = tp->isa_klassptr();
2526 if (tkls != nullptr && !StressReflectiveCode) {
2527 if (!tkls->is_loaded())
2528 return _type; // Bail out if not loaded
2529 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2530 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2531 // // Always returning precise element type is incorrect,
2532 // // e.g., element type could be object and array may contain strings
2533 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2534
2535 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2536 // according to the element type's subclassing.
2537 return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2538 }
2539 if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() &&
2540 tkls->offset() == in_bytes(Klass::super_offset())) {
2541 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super();
2542 // The field is Klass::_super. Return its (constant) value.
2543 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2544 return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR;
2545 }
2546 }
2547
2548 if (tkls != nullptr && !UseSecondarySupersCache
2549 && tkls->offset() == in_bytes(Klass::secondary_super_cache_offset())) {
2550 // Treat Klass::_secondary_super_cache as a constant when the cache is disabled.
2551 return TypePtr::NULL_PTR;
2552 }
2553
2554 // Bailout case
2555 return LoadNode::Value(phase);
2556 }
2557
2558 //------------------------------Identity---------------------------------------
2559 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2560 // Also feed through the klass in Allocate(...klass...)._klass.
2561 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2562 return klass_identity_common(phase);
2563 }
2564
2565 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2566 Node* x = LoadNode::Identity(phase);
2567 if (x != this) return x;
2568
2569 // Take apart the address into an oop and offset.
2570 // Return 'this' if we cannot.
2571 Node* adr = in(MemNode::Address);
2572 intptr_t offset = 0;
2573 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2574 if (base == nullptr) return this;
2575 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2576 if (toop == nullptr) return this;
2577
2578 // Step over potential GC barrier for OopHandle resolve
2579 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2580 if (bs->is_gc_barrier_node(base)) {
2581 base = bs->step_over_gc_barrier(base);
2582 }
2583
2584 // We can fetch the klass directly through an AllocateNode.
2585 // This works even if the klass is not constant (clone or newArray).
2586 if (offset == oopDesc::klass_offset_in_bytes()) {
2587 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2588 if (allocated_klass != nullptr) {
2589 return allocated_klass;
2590 }
2591 }
2592
2593 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2594 // See inline_native_Class_query for occurrences of these patterns.
2595 // Java Example: x.getClass().isAssignableFrom(y)
2596 //
2597 // This improves reflective code, often making the Class
2598 // mirror go completely dead. (Current exception: Class
2599 // mirrors may appear in debug info, but we could clean them out by
2600 // introducing a new debug info operator for Klass.java_mirror).
2601
2602 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass()
2603 && offset == java_lang_Class::klass_offset()) {
2604 if (base->is_Load()) {
2605 Node* base2 = base->in(MemNode::Address);
2606 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2607 Node* adr2 = base2->in(MemNode::Address);
2608 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2609 if (tkls != nullptr && !tkls->empty()
2610 && (tkls->isa_instklassptr() || tkls->isa_aryklassptr())
2611 && adr2->is_AddP()
2612 ) {
2613 int mirror_field = in_bytes(Klass::java_mirror_offset());
2614 if (tkls->offset() == mirror_field) {
2615 return adr2->in(AddPNode::Base);
2616 }
2617 }
2618 }
2619 }
2620 }
2621
2622 return this;
2623 }
2624
2625 LoadNode* LoadNode::clone_pinned() const {
2626 LoadNode* ld = clone()->as_Load();
2627 ld->_control_dependency = UnknownControl;
2628 return ld;
2629 }
2630
2631
2632 //------------------------------Value------------------------------------------
2633 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2634 const Type *t = klass_value_common(phase);
2635 if (t == Type::TOP)
2636 return t;
2637
2638 return t->make_narrowklass();
2639 }
2640
2641 //------------------------------Identity---------------------------------------
2642 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2643 // Also feed through the klass in Allocate(...klass...)._klass.
2644 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2645 Node *x = klass_identity_common(phase);
2646
2647 const Type *t = phase->type( x );
2648 if( t == Type::TOP ) return x;
2649 if( t->isa_narrowklass()) return x;
2650 assert (!t->isa_narrowoop(), "no narrow oop here");
2651
2652 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2653 }
2654
2655 //------------------------------Value-----------------------------------------
2656 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2657 // Either input is TOP ==> the result is TOP
2658 const Type *t1 = phase->type( in(MemNode::Memory) );
2659 if( t1 == Type::TOP ) return Type::TOP;
2660 Node *adr = in(MemNode::Address);
2661 const Type *t2 = phase->type( adr );
2662 if( t2 == Type::TOP ) return Type::TOP;
2663 const TypePtr *tp = t2->is_ptr();
2664 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2665 const TypeAryPtr *tap = tp->isa_aryptr();
2666 if( !tap ) return _type;
2667 return tap->size();
2668 }
2669
2670 //-------------------------------Ideal---------------------------------------
2671 // Feed through the length in AllocateArray(...length...)._length.
2672 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2673 Node* p = MemNode::Ideal_common(phase, can_reshape);
2674 if (p) return (p == NodeSentinel) ? nullptr : p;
2675
2676 // Take apart the address into an oop and offset.
2677 // Return 'this' if we cannot.
2678 Node* adr = in(MemNode::Address);
2679 intptr_t offset = 0;
2680 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2681 if (base == nullptr) return nullptr;
2682 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2683 if (tary == nullptr) return nullptr;
2684
2685 // We can fetch the length directly through an AllocateArrayNode.
2686 // This works even if the length is not constant (clone or newArray).
2687 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2688 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2689 if (alloc != nullptr) {
2690 Node* allocated_length = alloc->Ideal_length();
2691 Node* len = alloc->make_ideal_length(tary, phase);
2692 if (allocated_length != len) {
2693 // New CastII improves on this.
2694 return len;
2695 }
2696 }
2697 }
2698
2699 return nullptr;
2700 }
2701
2702 //------------------------------Identity---------------------------------------
2703 // Feed through the length in AllocateArray(...length...)._length.
2704 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2705 Node* x = LoadINode::Identity(phase);
2706 if (x != this) return x;
2707
2708 // Take apart the address into an oop and offset.
2709 // Return 'this' if we cannot.
2710 Node* adr = in(MemNode::Address);
2711 intptr_t offset = 0;
2712 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2713 if (base == nullptr) return this;
2714 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2715 if (tary == nullptr) return this;
2716
2717 // We can fetch the length directly through an AllocateArrayNode.
2718 // This works even if the length is not constant (clone or newArray).
2719 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2720 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2721 if (alloc != nullptr) {
2722 Node* allocated_length = alloc->Ideal_length();
2723 // Do not allow make_ideal_length to allocate a CastII node.
2724 Node* len = alloc->make_ideal_length(tary, phase, false);
2725 if (allocated_length == len) {
2726 // Return allocated_length only if it would not be improved by a CastII.
2727 return allocated_length;
2728 }
2729 }
2730 }
2731
2732 return this;
2733
2734 }
2735
2736 //=============================================================================
2737 //---------------------------StoreNode::make-----------------------------------
2738 // Polymorphic factory method:
2739 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) {
2740 assert((mo == unordered || mo == release), "unexpected");
2741 Compile* C = gvn.C;
2742 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2743 ctl != nullptr, "raw memory operations should have control edge");
2744
2745 switch (bt) {
2746 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2747 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2748 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2749 case T_CHAR:
2750 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2751 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2752 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2753 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2754 case T_METADATA:
2755 case T_ADDRESS:
2756 case T_OBJECT:
2757 #ifdef _LP64
2758 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2759 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2760 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2761 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2762 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2763 adr->bottom_type()->isa_rawptr())) {
2764 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2765 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2766 }
2767 #endif
2768 {
2769 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2770 }
2771 default:
2772 ShouldNotReachHere();
2773 return (StoreNode*)nullptr;
2774 }
2775 }
2776
2777 //--------------------------bottom_type----------------------------------------
2778 const Type *StoreNode::bottom_type() const {
2779 return Type::MEMORY;
2780 }
2781
2782 //------------------------------hash-------------------------------------------
2783 uint StoreNode::hash() const {
2784 // unroll addition of interesting fields
2785 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2786
2787 // Since they are not commoned, do not hash them:
2788 return NO_HASH;
2789 }
2790
2791 // Class to parse array pointers, and determine if they are adjacent. We parse the form:
2792 //
2793 // pointer = base
2794 // + constant_offset
2795 // + LShiftL( ConvI2L(int_offset + int_con), int_offset_shift)
2796 // + sum(other_offsets)
2797 //
2798 //
2799 // Note: we accumulate all constant offsets into constant_offset, even the int constant behind
2800 // the "LShiftL(ConvI2L(...))" pattern. We convert "ConvI2L(int_offset + int_con)" to
2801 // "ConvI2L(int_offset) + int_con", which is only safe if we can assume that either all
2802 // compared addresses have an overflow for "int_offset + int_con" or none.
2803 // For loads and stores on arrays, we know that if one overflows and the other not, then
2804 // the two addresses lay almost max_int indices apart, but the maximal array size is
2805 // only about half of that. Therefore, the RangeCheck on at least one of them must have
2806 // failed.
2807 //
2808 // constant_offset += LShiftL( ConvI2L(int_con), int_offset_shift)
2809 //
2810 // pointer = base
2811 // + constant_offset
2812 // + LShiftL( ConvI2L(int_offset), int_offset_shift)
2813 // + sum(other_offsets)
2814 //
2815 class ArrayPointer {
2816 private:
2817 const Node* _pointer; // The final pointer to the position in the array
2818 const Node* _base; // Base address of the array
2819 const jlong _constant_offset; // Sum of collected constant offsets
2820 const Node* _int_offset; // (optional) Offset behind LShiftL and ConvI2L
2821 const GrowableArray<Node*>* _other_offsets; // List of other AddP offsets
2822 const jint _int_offset_shift; // (optional) Shift value for int_offset
2823 const bool _is_valid; // The parsing succeeded
2824
2825 ArrayPointer(const bool is_valid,
2826 const Node* pointer,
2827 const Node* base,
2828 const jlong constant_offset,
2829 const Node* int_offset,
2830 const jint int_offset_shift,
2831 const GrowableArray<Node*>* other_offsets) :
2832 _pointer(pointer),
2833 _base(base),
2834 _constant_offset(constant_offset),
2835 _int_offset(int_offset),
2836 _other_offsets(other_offsets),
2837 _int_offset_shift(int_offset_shift),
2838 _is_valid(is_valid)
2839 {
2840 assert(_pointer != nullptr, "must always have pointer");
2841 assert(is_valid == (_base != nullptr), "have base exactly if valid");
2842 assert(is_valid == (_other_offsets != nullptr), "have other_offsets exactly if valid");
2843 }
2844
2845 static ArrayPointer make_invalid(const Node* pointer) {
2846 return ArrayPointer(false, pointer, nullptr, 0, nullptr, 0, nullptr);
2847 }
2848
2849 static bool parse_int_offset(Node* offset, Node*& int_offset, jint& int_offset_shift) {
2850 // offset = LShiftL( ConvI2L(int_offset), int_offset_shift)
2851 if (offset->Opcode() == Op_LShiftL &&
2852 offset->in(1)->Opcode() == Op_ConvI2L &&
2853 offset->in(2)->Opcode() == Op_ConI) {
2854 int_offset = offset->in(1)->in(1); // LShiftL -> ConvI2L -> int_offset
2855 int_offset_shift = offset->in(2)->get_int(); // LShiftL -> int_offset_shift
2856 return true;
2857 }
2858
2859 // offset = ConvI2L(int_offset) = LShiftL( ConvI2L(int_offset), 0)
2860 if (offset->Opcode() == Op_ConvI2L) {
2861 int_offset = offset->in(1);
2862 int_offset_shift = 0;
2863 return true;
2864 }
2865
2866 // parse failed
2867 return false;
2868 }
2869
2870 public:
2871 // Parse the structure above the pointer
2872 static ArrayPointer make(PhaseGVN* phase, const Node* pointer) {
2873 assert(phase->type(pointer)->isa_aryptr() != nullptr, "must be array pointer");
2874 if (!pointer->is_AddP()) { return ArrayPointer::make_invalid(pointer); }
2875
2876 const Node* base = pointer->in(AddPNode::Base);
2877 if (base == nullptr) { return ArrayPointer::make_invalid(pointer); }
2878
2879 const int search_depth = 5;
2880 Node* offsets[search_depth];
2881 int count = pointer->as_AddP()->unpack_offsets(offsets, search_depth);
2882
2883 // We expect at least a constant each
2884 if (count <= 0) { return ArrayPointer::make_invalid(pointer); }
2885
2886 // We extract the form:
2887 //
2888 // pointer = base
2889 // + constant_offset
2890 // + LShiftL( ConvI2L(int_offset + int_con), int_offset_shift)
2891 // + sum(other_offsets)
2892 //
2893 jlong constant_offset = 0;
2894 Node* int_offset = nullptr;
2895 jint int_offset_shift = 0;
2896 GrowableArray<Node*>* other_offsets = new GrowableArray<Node*>(count);
2897
2898 for (int i = 0; i < count; i++) {
2899 Node* offset = offsets[i];
2900 if (offset->Opcode() == Op_ConI) {
2901 // Constant int offset
2902 constant_offset += offset->get_int();
2903 } else if (offset->Opcode() == Op_ConL) {
2904 // Constant long offset
2905 constant_offset += offset->get_long();
2906 } else if(int_offset == nullptr && parse_int_offset(offset, int_offset, int_offset_shift)) {
2907 // LShiftL( ConvI2L(int_offset), int_offset_shift)
2908 int_offset = int_offset->uncast();
2909 if (int_offset->Opcode() == Op_AddI && int_offset->in(2)->Opcode() == Op_ConI) {
2910 // LShiftL( ConvI2L(int_offset + int_con), int_offset_shift)
2911 constant_offset += ((jlong)int_offset->in(2)->get_int()) << int_offset_shift;
2912 int_offset = int_offset->in(1);
2913 }
2914 } else {
2915 // All others
2916 other_offsets->append(offset);
2917 }
2918 }
2919
2920 return ArrayPointer(true, pointer, base, constant_offset, int_offset, int_offset_shift, other_offsets);
2921 }
2922
2923 bool is_adjacent_to_and_before(const ArrayPointer& other, const jlong data_size) const {
2924 if (!_is_valid || !other._is_valid) { return false; }
2925
2926 // Offset adjacent?
2927 if (this->_constant_offset + data_size != other._constant_offset) { return false; }
2928
2929 // All other components identical?
2930 if (this->_base != other._base ||
2931 this->_int_offset != other._int_offset ||
2932 this->_int_offset_shift != other._int_offset_shift ||
2933 this->_other_offsets->length() != other._other_offsets->length()) {
2934 return false;
2935 }
2936
2937 for (int i = 0; i < this->_other_offsets->length(); i++) {
2938 Node* o1 = this->_other_offsets->at(i);
2939 Node* o2 = other._other_offsets->at(i);
2940 if (o1 != o2) { return false; }
2941 }
2942
2943 return true;
2944 }
2945
2946 #ifndef PRODUCT
2947 void dump() {
2948 if (!_is_valid) {
2949 tty->print("ArrayPointer[%d %s, invalid]", _pointer->_idx, _pointer->Name());
2950 return;
2951 }
2952 tty->print("ArrayPointer[%d %s, base[%d %s] + %lld",
2953 _pointer->_idx, _pointer->Name(),
2954 _base->_idx, _base->Name(),
2955 (long long)_constant_offset);
2956 if (_int_offset != nullptr) {
2957 tty->print(" + I2L[%d %s] << %d",
2958 _int_offset->_idx, _int_offset->Name(), _int_offset_shift);
2959 }
2960 for (int i = 0; i < _other_offsets->length(); i++) {
2961 Node* n = _other_offsets->at(i);
2962 tty->print(" + [%d %s]", n->_idx, n->Name());
2963 }
2964 tty->print_cr("]");
2965 }
2966 #endif
2967 };
2968
2969 // Link together multiple stores (B/S/C/I) into a longer one.
2970 //
2971 // Example: _store = StoreB[i+3]
2972 //
2973 // RangeCheck[i+0] RangeCheck[i+0]
2974 // StoreB[i+0]
2975 // RangeCheck[i+3] RangeCheck[i+3]
2976 // StoreB[i+1] --> pass: fail:
2977 // StoreB[i+2] StoreI[i+0] StoreB[i+0]
2978 // StoreB[i+3]
2979 //
2980 // The 4 StoreB are merged into a single StoreI node. We have to be careful with RangeCheck[i+1]: before
2981 // the optimization, if this RangeCheck[i+1] fails, then we execute only StoreB[i+0], and then trap. After
2982 // the optimization, the new StoreI[i+0] is on the passing path of RangeCheck[i+3], and StoreB[i+0] on the
2983 // failing path.
2984 //
2985 // Note: For normal array stores, every store at first has a RangeCheck. But they can be removed with:
2986 // - RCE (RangeCheck Elimination): the RangeChecks in the loop are hoisted out and before the loop,
2987 // and possibly no RangeChecks remain between the stores.
2988 // - RangeCheck smearing: the earlier RangeChecks are adjusted such that they cover later RangeChecks,
2989 // and those later RangeChecks can be removed. Example:
2990 //
2991 // RangeCheck[i+0] RangeCheck[i+0] <- before first store
2992 // StoreB[i+0] StoreB[i+0] <- first store
2993 // RangeCheck[i+1] --> smeared --> RangeCheck[i+3] <- only RC between first and last store
2994 // StoreB[i+1] StoreB[i+1] <- second store
2995 // RangeCheck[i+2] --> removed
2996 // StoreB[i+2] StoreB[i+2]
2997 // RangeCheck[i+3] --> removed
2998 // StoreB[i+3] StoreB[i+3] <- last store
2999 //
3000 // Thus, it is a common pattern that between the first and last store in a chain
3001 // of adjacent stores there remains exactly one RangeCheck, located between the
3002 // first and the second store (e.g. RangeCheck[i+3]).
3003 //
3004 class MergePrimitiveArrayStores : public StackObj {
3005 private:
3006 PhaseGVN* _phase;
3007 StoreNode* _store;
3008
3009 public:
3010 MergePrimitiveArrayStores(PhaseGVN* phase, StoreNode* store) : _phase(phase), _store(store) {}
3011
3012 StoreNode* run();
3013
3014 private:
3015 bool is_compatible_store(const StoreNode* other_store) const;
3016 bool is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const;
3017 bool is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const;
3018 static bool is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out);
3019 enum CFGStatus { SuccessNoRangeCheck, SuccessWithRangeCheck, Failure };
3020 static CFGStatus cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store);
3021
3022 class Status {
3023 private:
3024 StoreNode* _found_store;
3025 bool _found_range_check;
3026
3027 Status(StoreNode* found_store, bool found_range_check)
3028 : _found_store(found_store), _found_range_check(found_range_check) {}
3029
3030 public:
3031 StoreNode* found_store() const { return _found_store; }
3032 bool found_range_check() const { return _found_range_check; }
3033 static Status make_failure() { return Status(nullptr, false); }
3034
3035 static Status make(StoreNode* found_store, const CFGStatus cfg_status) {
3036 if (cfg_status == CFGStatus::Failure) {
3037 return Status::make_failure();
3038 }
3039 return Status(found_store, cfg_status == CFGStatus::SuccessWithRangeCheck);
3040 }
3041 };
3042
3043 Status find_adjacent_use_store(const StoreNode* def_store) const;
3044 Status find_adjacent_def_store(const StoreNode* use_store) const;
3045 Status find_use_store(const StoreNode* def_store) const;
3046 Status find_def_store(const StoreNode* use_store) const;
3047 Status find_use_store_unidirectional(const StoreNode* def_store) const;
3048 Status find_def_store_unidirectional(const StoreNode* use_store) const;
3049
3050 void collect_merge_list(Node_List& merge_list) const;
3051 Node* make_merged_input_value(const Node_List& merge_list);
3052 StoreNode* make_merged_store(const Node_List& merge_list, Node* merged_input_value);
3053
3054 DEBUG_ONLY( void trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const; )
3055 };
3056
3057 StoreNode* MergePrimitiveArrayStores::run() {
3058 // Check for B/S/C/I
3059 int opc = _store->Opcode();
3060 if (opc != Op_StoreB && opc != Op_StoreC && opc != Op_StoreI) {
3061 return nullptr;
3062 }
3063
3064 // Only merge stores on arrays, and the stores must have the same size as the elements.
3065 const TypePtr* ptr_t = _store->adr_type();
3066 if (ptr_t == nullptr) {
3067 return nullptr;
3068 }
3069 const TypeAryPtr* aryptr_t = ptr_t->isa_aryptr();
3070 if (aryptr_t == nullptr) {
3071 return nullptr;
3072 }
3073 BasicType bt = aryptr_t->elem()->array_element_basic_type();
3074 if (!is_java_primitive(bt) ||
3075 type2aelembytes(bt) != _store->memory_size()) {
3076 return nullptr;
3077 }
3078 if (_store->is_unsafe_access()) {
3079 return nullptr;
3080 }
3081
3082 // The _store must be the "last" store in a chain. If we find a use we could merge with
3083 // then that use or a store further down is the "last" store.
3084 Status status_use = find_adjacent_use_store(_store);
3085 if (status_use.found_store() != nullptr) {
3086 return nullptr;
3087 }
3088
3089 // Check if we can merge with at least one def, so that we have at least 2 stores to merge.
3090 Status status_def = find_adjacent_def_store(_store);
3091 if (status_def.found_store() == nullptr) {
3092 return nullptr;
3093 }
3094
3095 ResourceMark rm;
3096 Node_List merge_list;
3097 collect_merge_list(merge_list);
3098
3099 Node* merged_input_value = make_merged_input_value(merge_list);
3100 if (merged_input_value == nullptr) { return nullptr; }
3101
3102 StoreNode* merged_store = make_merged_store(merge_list, merged_input_value);
3103
3104 DEBUG_ONLY( if(TraceMergeStores) { trace(merge_list, merged_input_value, merged_store); } )
3105
3106 return merged_store;
3107 }
3108
3109 // Check compatibility between _store and other_store.
3110 bool MergePrimitiveArrayStores::is_compatible_store(const StoreNode* other_store) const {
3111 int opc = _store->Opcode();
3112 assert(opc == Op_StoreB || opc == Op_StoreC || opc == Op_StoreI, "precondition");
3113 assert(_store->adr_type()->isa_aryptr() != nullptr, "must be array store");
3114 assert(!_store->is_unsafe_access(), "no unsafe accesses");
3115
3116 if (other_store == nullptr ||
3117 _store->Opcode() != other_store->Opcode() ||
3118 other_store->adr_type() == nullptr ||
3119 other_store->adr_type()->isa_aryptr() == nullptr ||
3120 other_store->is_unsafe_access()) {
3121 return false;
3122 }
3123
3124 // Check that the size of the stores, and the array elements are all the same.
3125 const TypeAryPtr* aryptr_t1 = _store->adr_type()->is_aryptr();
3126 const TypeAryPtr* aryptr_t2 = other_store->adr_type()->is_aryptr();
3127 BasicType aryptr_bt1 = aryptr_t1->elem()->array_element_basic_type();
3128 BasicType aryptr_bt2 = aryptr_t2->elem()->array_element_basic_type();
3129 if (!is_java_primitive(aryptr_bt1) || !is_java_primitive(aryptr_bt2)) {
3130 return false;
3131 }
3132 int size1 = type2aelembytes(aryptr_bt1);
3133 int size2 = type2aelembytes(aryptr_bt2);
3134 if (size1 != size2 ||
3135 size1 != _store->memory_size() ||
3136 _store->memory_size() != other_store->memory_size()) {
3137 return false;
3138 }
3139 return true;
3140 }
3141
3142 bool MergePrimitiveArrayStores::is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const {
3143 if (!is_adjacent_input_pair(def_store->in(MemNode::ValueIn),
3144 use_store->in(MemNode::ValueIn),
3145 def_store->memory_size())) {
3146 return false;
3147 }
3148
3149 ResourceMark rm;
3150 ArrayPointer array_pointer_use = ArrayPointer::make(_phase, use_store->in(MemNode::Address));
3151 ArrayPointer array_pointer_def = ArrayPointer::make(_phase, def_store->in(MemNode::Address));
3152 if (!array_pointer_def.is_adjacent_to_and_before(array_pointer_use, use_store->memory_size())) {
3153 return false;
3154 }
3155
3156 return true;
3157 }
3158
3159 bool MergePrimitiveArrayStores::is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const {
3160 // Pattern: [n1 = ConI, n2 = ConI]
3161 if (n1->Opcode() == Op_ConI) {
3162 return n2->Opcode() == Op_ConI;
3163 }
3164
3165 // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3166 #ifndef VM_LITTLE_ENDIAN
3167 // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3168 // Swapping n1 with n2 gives same pattern as on little endian platforms.
3169 swap(n1, n2);
3170 #endif // !VM_LITTLE_ENDIAN
3171 Node const* base_n2;
3172 jint shift_n2;
3173 if (!is_con_RShift(n2, base_n2, shift_n2)) {
3174 return false;
3175 }
3176 if (n1->Opcode() == Op_ConvL2I) {
3177 // look through
3178 n1 = n1->in(1);
3179 }
3180 Node const* base_n1;
3181 jint shift_n1;
3182 if (n1 == base_n2) {
3183 // n1 = base = base >> 0
3184 base_n1 = n1;
3185 shift_n1 = 0;
3186 } else if (!is_con_RShift(n1, base_n1, shift_n1)) {
3187 return false;
3188 }
3189 int bits_per_store = memory_size * 8;
3190 if (base_n1 != base_n2 ||
3191 shift_n1 + bits_per_store != shift_n2 ||
3192 shift_n1 % bits_per_store != 0) {
3193 return false;
3194 }
3195
3196 // both load from same value with correct shift
3197 return true;
3198 }
3199
3200 // Detect pattern: n = base_out >> shift_out
3201 bool MergePrimitiveArrayStores::is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out) {
3202 assert(n != nullptr, "precondition");
3203
3204 int opc = n->Opcode();
3205 if (opc == Op_ConvL2I) {
3206 n = n->in(1);
3207 opc = n->Opcode();
3208 }
3209
3210 if ((opc == Op_RShiftI ||
3211 opc == Op_RShiftL ||
3212 opc == Op_URShiftI ||
3213 opc == Op_URShiftL) &&
3214 n->in(2)->is_ConI()) {
3215 base_out = n->in(1);
3216 shift_out = n->in(2)->get_int();
3217 // The shift must be positive:
3218 return shift_out >= 0;
3219 }
3220 return false;
3221 }
3222
3223 // Check if there is nothing between the two stores, except optionally a RangeCheck leading to an uncommon trap.
3224 MergePrimitiveArrayStores::CFGStatus MergePrimitiveArrayStores::cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store) {
3225 assert(use_store->in(MemNode::Memory) == def_store, "use-def relationship");
3226
3227 Node* ctrl_use = use_store->in(MemNode::Control);
3228 Node* ctrl_def = def_store->in(MemNode::Control);
3229 if (ctrl_use == nullptr || ctrl_def == nullptr) {
3230 return CFGStatus::Failure;
3231 }
3232
3233 if (ctrl_use == ctrl_def) {
3234 // Same ctrl -> no RangeCheck in between.
3235 // Check: use_store must be the only use of def_store.
3236 if (def_store->outcnt() > 1) {
3237 return CFGStatus::Failure;
3238 }
3239 return CFGStatus::SuccessNoRangeCheck;
3240 }
3241
3242 // Different ctrl -> could have RangeCheck in between.
3243 // Check: 1. def_store only has these uses: use_store and MergeMem for uncommon trap, and
3244 // 2. ctrl separated by RangeCheck.
3245 if (def_store->outcnt() != 2) {
3246 return CFGStatus::Failure; // Cannot have exactly these uses: use_store and MergeMem for uncommon trap.
3247 }
3248 int use_store_out_idx = def_store->raw_out(0) == use_store ? 0 : 1;
3249 Node* merge_mem = def_store->raw_out(1 - use_store_out_idx)->isa_MergeMem();
3250 if (merge_mem == nullptr ||
3251 merge_mem->outcnt() != 1) {
3252 return CFGStatus::Failure; // Does not have MergeMem for uncommon trap.
3253 }
3254 if (!ctrl_use->is_IfProj() ||
3255 !ctrl_use->in(0)->is_RangeCheck() ||
3256 ctrl_use->in(0)->outcnt() != 2) {
3257 return CFGStatus::Failure; // Not RangeCheck.
3258 }
3259 ProjNode* other_proj = ctrl_use->as_IfProj()->other_if_proj();
3260 Node* trap = other_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check);
3261 if (trap != merge_mem->unique_out() ||
3262 ctrl_use->in(0)->in(0) != ctrl_def) {
3263 return CFGStatus::Failure; // Not RangeCheck with merge_mem leading to uncommon trap.
3264 }
3265
3266 return CFGStatus::SuccessWithRangeCheck;
3267 }
3268
3269 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_adjacent_use_store(const StoreNode* def_store) const {
3270 Status status_use = find_use_store(def_store);
3271 StoreNode* use_store = status_use.found_store();
3272 if (use_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3273 return Status::make_failure();
3274 }
3275 return status_use;
3276 }
3277
3278 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_adjacent_def_store(const StoreNode* use_store) const {
3279 Status status_def = find_def_store(use_store);
3280 StoreNode* def_store = status_def.found_store();
3281 if (def_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3282 return Status::make_failure();
3283 }
3284 return status_def;
3285 }
3286
3287 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_use_store(const StoreNode* def_store) const {
3288 Status status_use = find_use_store_unidirectional(def_store);
3289
3290 #ifdef ASSERT
3291 StoreNode* use_store = status_use.found_store();
3292 if (use_store != nullptr) {
3293 Status status_def = find_def_store_unidirectional(use_store);
3294 assert(status_def.found_store() == def_store &&
3295 status_def.found_range_check() == status_use.found_range_check(),
3296 "find_use_store and find_def_store must be symmetric");
3297 }
3298 #endif
3299
3300 return status_use;
3301 }
3302
3303 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_def_store(const StoreNode* use_store) const {
3304 Status status_def = find_def_store_unidirectional(use_store);
3305
3306 #ifdef ASSERT
3307 StoreNode* def_store = status_def.found_store();
3308 if (def_store != nullptr) {
3309 Status status_use = find_use_store_unidirectional(def_store);
3310 assert(status_use.found_store() == use_store &&
3311 status_use.found_range_check() == status_def.found_range_check(),
3312 "find_use_store and find_def_store must be symmetric");
3313 }
3314 #endif
3315
3316 return status_def;
3317 }
3318
3319 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_use_store_unidirectional(const StoreNode* def_store) const {
3320 assert(is_compatible_store(def_store), "precondition: must be compatible with _store");
3321
3322 for (DUIterator_Fast imax, i = def_store->fast_outs(imax); i < imax; i++) {
3323 StoreNode* use_store = def_store->fast_out(i)->isa_Store();
3324 if (is_compatible_store(use_store)) {
3325 return Status::make(use_store, cfg_status_for_pair(use_store, def_store));
3326 }
3327 }
3328
3329 return Status::make_failure();
3330 }
3331
3332 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_def_store_unidirectional(const StoreNode* use_store) const {
3333 assert(is_compatible_store(use_store), "precondition: must be compatible with _store");
3334
3335 StoreNode* def_store = use_store->in(MemNode::Memory)->isa_Store();
3336 if (!is_compatible_store(def_store)) {
3337 return Status::make_failure();
3338 }
3339
3340 return Status::make(def_store, cfg_status_for_pair(use_store, def_store));
3341 }
3342
3343 void MergePrimitiveArrayStores::collect_merge_list(Node_List& merge_list) const {
3344 // The merged store can be at most 8 bytes.
3345 const uint merge_list_max_size = 8 / _store->memory_size();
3346 assert(merge_list_max_size >= 2 &&
3347 merge_list_max_size <= 8 &&
3348 is_power_of_2(merge_list_max_size),
3349 "must be 2, 4 or 8");
3350
3351 // Traverse up the chain of adjacent def stores.
3352 StoreNode* current = _store;
3353 merge_list.push(current);
3354 while (current != nullptr && merge_list.size() < merge_list_max_size) {
3355 Status status = find_adjacent_def_store(current);
3356 current = status.found_store();
3357 if (current != nullptr) {
3358 merge_list.push(current);
3359
3360 // We can have at most one RangeCheck.
3361 if (status.found_range_check()) {
3362 break;
3363 }
3364 }
3365 }
3366
3367 // Truncate the merge_list to a power of 2.
3368 const uint pow2size = round_down_power_of_2(merge_list.size());
3369 assert(pow2size >= 2, "must be merging at least 2 stores");
3370 while (merge_list.size() > pow2size) { merge_list.pop(); }
3371 }
3372
3373 // Merge the input values of the smaller stores to a single larger input value.
3374 Node* MergePrimitiveArrayStores::make_merged_input_value(const Node_List& merge_list) {
3375 int new_memory_size = _store->memory_size() * merge_list.size();
3376 Node* first = merge_list.at(merge_list.size()-1);
3377 Node* merged_input_value = nullptr;
3378 if (_store->in(MemNode::ValueIn)->Opcode() == Op_ConI) {
3379 // Pattern: [ConI, ConI, ...] -> new constant
3380 jlong con = 0;
3381 jlong bits_per_store = _store->memory_size() * 8;
3382 jlong mask = (((jlong)1) << bits_per_store) - 1;
3383 for (uint i = 0; i < merge_list.size(); i++) {
3384 jlong con_i = merge_list.at(i)->in(MemNode::ValueIn)->get_int();
3385 #ifdef VM_LITTLE_ENDIAN
3386 con = con << bits_per_store;
3387 con = con | (mask & con_i);
3388 #else // VM_LITTLE_ENDIAN
3389 con_i = (mask & con_i) << (i * bits_per_store);
3390 con = con | con_i;
3391 #endif // VM_LITTLE_ENDIAN
3392 }
3393 merged_input_value = _phase->longcon(con);
3394 } else {
3395 // Pattern: [base >> 24, base >> 16, base >> 8, base] -> base
3396 // | |
3397 // _store first
3398 //
3399 Node* hi = _store->in(MemNode::ValueIn);
3400 Node* lo = first->in(MemNode::ValueIn);
3401 #ifndef VM_LITTLE_ENDIAN
3402 // `_store` and `first` are swapped in the diagram above
3403 swap(hi, lo);
3404 #endif // !VM_LITTLE_ENDIAN
3405 Node const* hi_base;
3406 jint hi_shift;
3407 merged_input_value = lo;
3408 bool is_true = is_con_RShift(hi, hi_base, hi_shift);
3409 assert(is_true, "must detect con RShift");
3410 if (merged_input_value != hi_base && merged_input_value->Opcode() == Op_ConvL2I) {
3411 // look through
3412 merged_input_value = merged_input_value->in(1);
3413 }
3414 if (merged_input_value != hi_base) {
3415 // merged_input_value is not the base
3416 return nullptr;
3417 }
3418 }
3419
3420 if (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size <= 4) {
3421 // Example:
3422 //
3423 // long base = ...;
3424 // a[0] = (byte)(base >> 0);
3425 // a[1] = (byte)(base >> 8);
3426 //
3427 merged_input_value = _phase->transform(new ConvL2INode(merged_input_value));
3428 }
3429
3430 assert((_phase->type(merged_input_value)->isa_int() != nullptr && new_memory_size <= 4) ||
3431 (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size == 8),
3432 "merged_input_value is either int or long, and new_memory_size is small enough");
3433
3434 return merged_input_value;
3435 }
3436
3437 // //
3438 // first_ctrl first_mem first_adr first_ctrl first_mem first_adr //
3439 // | | | | | | //
3440 // | | | | +---------------+ | //
3441 // | | | | | | | //
3442 // | | +---------+ | | +---------------+ //
3443 // | | | | | | | | //
3444 // +--------------+ | | v1 +------------------------------+ | | v1 //
3445 // | | | | | | | | | | | | //
3446 // RangeCheck first_store RangeCheck | | first_store //
3447 // | | | | | | | //
3448 // last_ctrl | +----> unc_trap last_ctrl | | +----> unc_trap //
3449 // | | ===> | | | //
3450 // +--------------+ | a2 v2 | | | //
3451 // | | | | | | | | //
3452 // | second_store | | | //
3453 // | | | | | [v1 v2 ... vn] //
3454 // ... ... | | | | //
3455 // | | | | | v //
3456 // +--------------+ | an vn +--------------+ | | merged_input_value //
3457 // | | | | | | | | //
3458 // last_store (= _store) merged_store //
3459 // //
3460 StoreNode* MergePrimitiveArrayStores::make_merged_store(const Node_List& merge_list, Node* merged_input_value) {
3461 Node* first_store = merge_list.at(merge_list.size()-1);
3462 Node* last_ctrl = _store->in(MemNode::Control); // after (optional) RangeCheck
3463 Node* first_mem = first_store->in(MemNode::Memory);
3464 Node* first_adr = first_store->in(MemNode::Address);
3465
3466 const TypePtr* new_adr_type = _store->adr_type();
3467
3468 int new_memory_size = _store->memory_size() * merge_list.size();
3469 BasicType bt = T_ILLEGAL;
3470 switch (new_memory_size) {
3471 case 2: bt = T_SHORT; break;
3472 case 4: bt = T_INT; break;
3473 case 8: bt = T_LONG; break;
3474 }
3475
3476 StoreNode* merged_store = StoreNode::make(*_phase, last_ctrl, first_mem, first_adr,
3477 new_adr_type, merged_input_value, bt, MemNode::unordered);
3478
3479 // Marking the store mismatched is sufficient to prevent reordering, since array stores
3480 // are all on the same slice. Hence, we need no barriers.
3481 merged_store->set_mismatched_access();
3482
3483 // Constants above may now also be be packed -> put candidate on worklist
3484 _phase->is_IterGVN()->_worklist.push(first_mem);
3485
3486 return merged_store;
3487 }
3488
3489 #ifdef ASSERT
3490 void MergePrimitiveArrayStores::trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const {
3491 stringStream ss;
3492 ss.print_cr("[TraceMergeStores]: Replace");
3493 for (int i = (int)merge_list.size() - 1; i >= 0; i--) {
3494 merge_list.at(i)->dump("\n", false, &ss);
3495 }
3496 ss.print_cr("[TraceMergeStores]: with");
3497 merged_input_value->dump("\n", false, &ss);
3498 merged_store->dump("\n", false, &ss);
3499 tty->print("%s", ss.as_string());
3500 }
3501 #endif
3502
3503 //------------------------------Ideal------------------------------------------
3504 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
3505 // When a store immediately follows a relevant allocation/initialization,
3506 // try to capture it into the initialization, or hoist it above.
3507 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3508 Node* p = MemNode::Ideal_common(phase, can_reshape);
3509 if (p) return (p == NodeSentinel) ? nullptr : p;
3510
3511 Node* mem = in(MemNode::Memory);
3512 Node* address = in(MemNode::Address);
3513 Node* value = in(MemNode::ValueIn);
3514 // Back-to-back stores to same address? Fold em up. Generally
3515 // unsafe if I have intervening uses...
3516 if (phase->C->get_adr_type(phase->C->get_alias_index(adr_type())) != TypeAryPtr::INLINES) {
3517 Node* st = mem;
3518 // If Store 'st' has more than one use, we cannot fold 'st' away.
3519 // For example, 'st' might be the final state at a conditional
3520 // return. Or, 'st' might be used by some node which is live at
3521 // the same time 'st' is live, which might be unschedulable. So,
3522 // require exactly ONE user until such time as we clone 'mem' for
3523 // each of 'mem's uses (thus making the exactly-1-user-rule hold
3524 // true).
3525 while (st->is_Store() && st->outcnt() == 1) {
3526 // Looking at a dead closed cycle of memory?
3527 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
3528 assert(Opcode() == st->Opcode() ||
3529 st->Opcode() == Op_StoreVector ||
3530 Opcode() == Op_StoreVector ||
3531 st->Opcode() == Op_StoreVectorScatter ||
3532 Opcode() == Op_StoreVectorScatter ||
3533 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
3534 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
3535 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
3536 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) ||
3537 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
3538 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
3539
3540 if (st->in(MemNode::Address)->eqv_uncast(address) &&
3541 st->as_Store()->memory_size() <= this->memory_size()) {
3542 Node* use = st->raw_out(0);
3543 if (phase->is_IterGVN()) {
3544 phase->is_IterGVN()->rehash_node_delayed(use);
3545 }
3546 // It's OK to do this in the parser, since DU info is always accurate,
3547 // and the parser always refers to nodes via SafePointNode maps.
3548 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
3549 return this;
3550 }
3551 st = st->in(MemNode::Memory);
3552 }
3553 }
3554
3555
3556 // Capture an unaliased, unconditional, simple store into an initializer.
3557 // Or, if it is independent of the allocation, hoist it above the allocation.
3558 if (ReduceFieldZeroing && /*can_reshape &&*/
3559 mem->is_Proj() && mem->in(0)->is_Initialize()) {
3560 InitializeNode* init = mem->in(0)->as_Initialize();
3561 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
3562 if (offset > 0) {
3563 Node* moved = init->capture_store(this, offset, phase, can_reshape);
3564 // If the InitializeNode captured me, it made a raw copy of me,
3565 // and I need to disappear.
3566 if (moved != nullptr) {
3567 // %%% hack to ensure that Ideal returns a new node:
3568 mem = MergeMemNode::make(mem);
3569 return mem; // fold me away
3570 }
3571 }
3572 }
3573
3574 // Fold reinterpret cast into memory operation:
3575 // StoreX mem (MoveY2X v) => StoreY mem v
3576 if (value->is_Move()) {
3577 const Type* vt = value->in(1)->bottom_type();
3578 if (has_reinterpret_variant(vt)) {
3579 if (phase->C->post_loop_opts_phase()) {
3580 return convert_to_reinterpret_store(*phase, value->in(1), vt);
3581 } else {
3582 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
3583 }
3584 }
3585 }
3586
3587 if (MergeStores && UseUnalignedAccesses) {
3588 if (phase->C->post_loop_opts_phase()) {
3589 MergePrimitiveArrayStores merge(phase, this);
3590 Node* progress = merge.run();
3591 if (progress != nullptr) { return progress; }
3592 } else {
3593 phase->C->record_for_post_loop_opts_igvn(this);
3594 }
3595 }
3596
3597 return nullptr; // No further progress
3598 }
3599
3600 //------------------------------Value-----------------------------------------
3601 const Type* StoreNode::Value(PhaseGVN* phase) const {
3602 // Either input is TOP ==> the result is TOP
3603 const Type *t1 = phase->type( in(MemNode::Memory) );
3604 if( t1 == Type::TOP ) return Type::TOP;
3605 const Type *t2 = phase->type( in(MemNode::Address) );
3606 if( t2 == Type::TOP ) return Type::TOP;
3607 const Type *t3 = phase->type( in(MemNode::ValueIn) );
3608 if( t3 == Type::TOP ) return Type::TOP;
3609 return Type::MEMORY;
3610 }
3611
3612 //------------------------------Identity---------------------------------------
3613 // Remove redundant stores:
3614 // Store(m, p, Load(m, p)) changes to m.
3615 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
3616 Node* StoreNode::Identity(PhaseGVN* phase) {
3617 Node* mem = in(MemNode::Memory);
3618 Node* adr = in(MemNode::Address);
3619 Node* val = in(MemNode::ValueIn);
3620
3621 Node* result = this;
3622
3623 // Load then Store? Then the Store is useless
3624 if (val->is_Load() &&
3625 val->in(MemNode::Address)->eqv_uncast(adr) &&
3626 val->in(MemNode::Memory )->eqv_uncast(mem) &&
3627 val->as_Load()->store_Opcode() == Opcode()) {
3628 // Ensure vector type is the same
3629 if (!is_StoreVector() || (mem->is_LoadVector() && as_StoreVector()->vect_type() == mem->as_LoadVector()->vect_type())) {
3630 result = mem;
3631 }
3632 }
3633
3634 // Two stores in a row of the same value?
3635 if (result == this &&
3636 mem->is_Store() &&
3637 mem->in(MemNode::Address)->eqv_uncast(adr) &&
3638 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
3639 mem->Opcode() == Opcode()) {
3640 if (!is_StoreVector()) {
3641 result = mem;
3642 } else {
3643 const StoreVectorNode* store_vector = as_StoreVector();
3644 const StoreVectorNode* mem_vector = mem->as_StoreVector();
3645 const Node* store_indices = store_vector->indices();
3646 const Node* mem_indices = mem_vector->indices();
3647 const Node* store_mask = store_vector->mask();
3648 const Node* mem_mask = mem_vector->mask();
3649 // Ensure types, indices, and masks match
3650 if (store_vector->vect_type() == mem_vector->vect_type() &&
3651 ((store_indices == nullptr) == (mem_indices == nullptr) &&
3652 (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) &&
3653 ((store_mask == nullptr) == (mem_mask == nullptr) &&
3654 (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) {
3655 result = mem;
3656 }
3657 }
3658 }
3659
3660 // Store of zero anywhere into a freshly-allocated object?
3661 // Then the store is useless.
3662 // (It must already have been captured by the InitializeNode.)
3663 if (result == this && ReduceFieldZeroing) {
3664 // a newly allocated object is already all-zeroes everywhere
3665 if (mem->is_Proj() && mem->in(0)->is_Allocate() &&
3666 (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == val)) {
3667 result = mem;
3668 }
3669
3670 if (result == this && phase->type(val)->is_zero_type()) {
3671 // the store may also apply to zero-bits in an earlier object
3672 Node* prev_mem = find_previous_store(phase);
3673 // Steps (a), (b): Walk past independent stores to find an exact match.
3674 if (prev_mem != nullptr) {
3675 Node* prev_val = can_see_stored_value(prev_mem, phase);
3676 if (prev_val != nullptr && prev_val == val) {
3677 // prev_val and val might differ by a cast; it would be good
3678 // to keep the more informative of the two.
3679 result = mem;
3680 }
3681 }
3682 }
3683 }
3684
3685 PhaseIterGVN* igvn = phase->is_IterGVN();
3686 if (result != this && igvn != nullptr) {
3687 MemBarNode* trailing = trailing_membar();
3688 if (trailing != nullptr) {
3689 #ifdef ASSERT
3690 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
3691 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects");
3692 #endif
3693 trailing->remove(igvn);
3694 }
3695 }
3696
3697 return result;
3698 }
3699
3700 //------------------------------match_edge-------------------------------------
3701 // Do we Match on this edge index or not? Match only memory & value
3702 uint StoreNode::match_edge(uint idx) const {
3703 return idx == MemNode::Address || idx == MemNode::ValueIn;
3704 }
3705
3706 //------------------------------cmp--------------------------------------------
3707 // Do not common stores up together. They generally have to be split
3708 // back up anyways, so do not bother.
3709 bool StoreNode::cmp( const Node &n ) const {
3710 return (&n == this); // Always fail except on self
3711 }
3712
3713 //------------------------------Ideal_masked_input-----------------------------
3714 // Check for a useless mask before a partial-word store
3715 // (StoreB ... (AndI valIn conIa) )
3716 // If (conIa & mask == mask) this simplifies to
3717 // (StoreB ... (valIn) )
3718 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
3719 Node *val = in(MemNode::ValueIn);
3720 if( val->Opcode() == Op_AndI ) {
3721 const TypeInt *t = phase->type( val->in(2) )->isa_int();
3722 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
3723 set_req_X(MemNode::ValueIn, val->in(1), phase);
3724 return this;
3725 }
3726 }
3727 return nullptr;
3728 }
3729
3730
3731 //------------------------------Ideal_sign_extended_input----------------------
3732 // Check for useless sign-extension before a partial-word store
3733 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
3734 // If (conIL == conIR && conIR <= num_bits) this simplifies to
3735 // (StoreB ... (valIn) )
3736 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
3737 Node *val = in(MemNode::ValueIn);
3738 if( val->Opcode() == Op_RShiftI ) {
3739 const TypeInt *t = phase->type( val->in(2) )->isa_int();
3740 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
3741 Node *shl = val->in(1);
3742 if( shl->Opcode() == Op_LShiftI ) {
3743 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
3744 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
3745 set_req_X(MemNode::ValueIn, shl->in(1), phase);
3746 return this;
3747 }
3748 }
3749 }
3750 }
3751 return nullptr;
3752 }
3753
3754 //------------------------------value_never_loaded-----------------------------------
3755 // Determine whether there are any possible loads of the value stored.
3756 // For simplicity, we actually check if there are any loads from the
3757 // address stored to, not just for loads of the value stored by this node.
3758 //
3759 bool StoreNode::value_never_loaded(PhaseValues* phase) const {
3760 Node *adr = in(Address);
3761 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
3762 if (adr_oop == nullptr)
3763 return false;
3764 if (!adr_oop->is_known_instance_field())
3765 return false; // if not a distinct instance, there may be aliases of the address
3766 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
3767 Node *use = adr->fast_out(i);
3768 if (use->is_Load() || use->is_LoadStore()) {
3769 return false;
3770 }
3771 }
3772 return true;
3773 }
3774
3775 MemBarNode* StoreNode::trailing_membar() const {
3776 if (is_release()) {
3777 MemBarNode* trailing_mb = nullptr;
3778 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3779 Node* u = fast_out(i);
3780 if (u->is_MemBar()) {
3781 if (u->as_MemBar()->trailing_store()) {
3782 assert(u->Opcode() == Op_MemBarVolatile, "");
3783 assert(trailing_mb == nullptr, "only one");
3784 trailing_mb = u->as_MemBar();
3785 #ifdef ASSERT
3786 Node* leading = u->as_MemBar()->leading_membar();
3787 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3788 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
3789 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
3790 #endif
3791 } else {
3792 assert(u->as_MemBar()->standalone(), "");
3793 }
3794 }
3795 }
3796 return trailing_mb;
3797 }
3798 return nullptr;
3799 }
3800
3801
3802 //=============================================================================
3803 //------------------------------Ideal------------------------------------------
3804 // If the store is from an AND mask that leaves the low bits untouched, then
3805 // we can skip the AND operation. If the store is from a sign-extension
3806 // (a left shift, then right shift) we can skip both.
3807 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
3808 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
3809 if( progress != nullptr ) return progress;
3810
3811 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
3812 if( progress != nullptr ) return progress;
3813
3814 // Finally check the default case
3815 return StoreNode::Ideal(phase, can_reshape);
3816 }
3817
3818 //=============================================================================
3819 //------------------------------Ideal------------------------------------------
3820 // If the store is from an AND mask that leaves the low bits untouched, then
3821 // we can skip the AND operation
3822 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
3823 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
3824 if( progress != nullptr ) return progress;
3825
3826 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
3827 if( progress != nullptr ) return progress;
3828
3829 // Finally check the default case
3830 return StoreNode::Ideal(phase, can_reshape);
3831 }
3832
3833 //=============================================================================
3834 //----------------------------------SCMemProjNode------------------------------
3835 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
3836 {
3837 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
3838 return Type::TOP;
3839 }
3840 return bottom_type();
3841 }
3842
3843 //=============================================================================
3844 //----------------------------------LoadStoreNode------------------------------
3845 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
3846 : Node(required),
3847 _type(rt),
3848 _adr_type(at),
3849 _barrier_data(0)
3850 {
3851 init_req(MemNode::Control, c );
3852 init_req(MemNode::Memory , mem);
3853 init_req(MemNode::Address, adr);
3854 init_req(MemNode::ValueIn, val);
3855 init_class_id(Class_LoadStore);
3856 }
3857
3858 //------------------------------Value-----------------------------------------
3859 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
3860 // Either input is TOP ==> the result is TOP
3861 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
3862 return Type::TOP;
3863 }
3864 const Type* t = phase->type(in(MemNode::Memory));
3865 if (t == Type::TOP) {
3866 return Type::TOP;
3867 }
3868 t = phase->type(in(MemNode::Address));
3869 if (t == Type::TOP) {
3870 return Type::TOP;
3871 }
3872 t = phase->type(in(MemNode::ValueIn));
3873 if (t == Type::TOP) {
3874 return Type::TOP;
3875 }
3876 return bottom_type();
3877 }
3878
3879 uint LoadStoreNode::ideal_reg() const {
3880 return _type->ideal_reg();
3881 }
3882
3883 // This method conservatively checks if the result of a LoadStoreNode is
3884 // used, that is, if it returns true, then it is definitely the case that
3885 // the result of the node is not needed.
3886 // For example, GetAndAdd can be matched into a lock_add instead of a
3887 // lock_xadd if the result of LoadStoreNode::result_not_used() is true
3888 bool LoadStoreNode::result_not_used() const {
3889 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3890 Node *x = fast_out(i);
3891 if (x->Opcode() == Op_SCMemProj) {
3892 continue;
3893 }
3894 if (x->bottom_type() == TypeTuple::MEMBAR &&
3895 !x->is_Call() &&
3896 x->Opcode() != Op_Blackhole) {
3897 continue;
3898 }
3899 return false;
3900 }
3901 return true;
3902 }
3903
3904 MemBarNode* LoadStoreNode::trailing_membar() const {
3905 MemBarNode* trailing = nullptr;
3906 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3907 Node* u = fast_out(i);
3908 if (u->is_MemBar()) {
3909 if (u->as_MemBar()->trailing_load_store()) {
3910 assert(u->Opcode() == Op_MemBarAcquire, "");
3911 assert(trailing == nullptr, "only one");
3912 trailing = u->as_MemBar();
3913 #ifdef ASSERT
3914 Node* leading = trailing->leading_membar();
3915 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3916 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
3917 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
3918 #endif
3919 } else {
3920 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
3921 }
3922 }
3923 }
3924
3925 return trailing;
3926 }
3927
3928 uint LoadStoreNode::size_of() const { return sizeof(*this); }
3929
3930 //=============================================================================
3931 //----------------------------------LoadStoreConditionalNode--------------------
3932 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) {
3933 init_req(ExpectedIn, ex );
3934 }
3935
3936 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
3937 // Either input is TOP ==> the result is TOP
3938 const Type* t = phase->type(in(ExpectedIn));
3939 if (t == Type::TOP) {
3940 return Type::TOP;
3941 }
3942 return LoadStoreNode::Value(phase);
3943 }
3944
3945 //=============================================================================
3946 //-------------------------------adr_type--------------------------------------
3947 const TypePtr* ClearArrayNode::adr_type() const {
3948 Node *adr = in(3);
3949 if (adr == nullptr) return nullptr; // node is dead
3950 return MemNode::calculate_adr_type(adr->bottom_type());
3951 }
3952
3953 //------------------------------match_edge-------------------------------------
3954 // Do we Match on this edge index or not? Do not match memory
3955 uint ClearArrayNode::match_edge(uint idx) const {
3956 return idx > 1;
3957 }
3958
3959 //------------------------------Identity---------------------------------------
3960 // Clearing a zero length array does nothing
3961 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
3962 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
3963 }
3964
3965 //------------------------------Idealize---------------------------------------
3966 // Clearing a short array is faster with stores
3967 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3968 // Already know this is a large node, do not try to ideal it
3969 if (_is_large) return nullptr;
3970
3971 const int unit = BytesPerLong;
3972 const TypeX* t = phase->type(in(2))->isa_intptr_t();
3973 if (!t) return nullptr;
3974 if (!t->is_con()) return nullptr;
3975 intptr_t raw_count = t->get_con();
3976 intptr_t size = raw_count;
3977 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
3978 // Clearing nothing uses the Identity call.
3979 // Negative clears are possible on dead ClearArrays
3980 // (see jck test stmt114.stmt11402.val).
3981 if (size <= 0 || size % unit != 0) return nullptr;
3982 intptr_t count = size / unit;
3983 // Length too long; communicate this to matchers and assemblers.
3984 // Assemblers are responsible to produce fast hardware clears for it.
3985 if (size > InitArrayShortSize) {
3986 return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true);
3987 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
3988 return nullptr;
3989 }
3990 if (!IdealizeClearArrayNode) return nullptr;
3991 Node *mem = in(1);
3992 if( phase->type(mem)==Type::TOP ) return nullptr;
3993 Node *adr = in(3);
3994 const Type* at = phase->type(adr);
3995 if( at==Type::TOP ) return nullptr;
3996 const TypePtr* atp = at->isa_ptr();
3997 // adjust atp to be the correct array element address type
3998 if (atp == nullptr) atp = TypePtr::BOTTOM;
3999 else atp = atp->add_offset(Type::OffsetBot);
4000 // Get base for derived pointer purposes
4001 if( adr->Opcode() != Op_AddP ) Unimplemented();
4002 Node *base = adr->in(1);
4003
4004 Node *val = in(4);
4005 Node *off = phase->MakeConX(BytesPerLong);
4006 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
4007 count--;
4008 while( count-- ) {
4009 mem = phase->transform(mem);
4010 adr = phase->transform(new AddPNode(base,adr,off));
4011 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
4012 }
4013 return mem;
4014 }
4015
4016 //----------------------------step_through----------------------------------
4017 // Return allocation input memory edge if it is different instance
4018 // or itself if it is the one we are looking for.
4019 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
4020 Node* n = *np;
4021 assert(n->is_ClearArray(), "sanity");
4022 intptr_t offset;
4023 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
4024 // This method is called only before Allocate nodes are expanded
4025 // during macro nodes expansion. Before that ClearArray nodes are
4026 // only generated in PhaseMacroExpand::generate_arraycopy() (before
4027 // Allocate nodes are expanded) which follows allocations.
4028 assert(alloc != nullptr, "should have allocation");
4029 if (alloc->_idx == instance_id) {
4030 // Can not bypass initialization of the instance we are looking for.
4031 return false;
4032 }
4033 // Otherwise skip it.
4034 InitializeNode* init = alloc->initialization();
4035 if (init != nullptr)
4036 *np = init->in(TypeFunc::Memory);
4037 else
4038 *np = alloc->in(TypeFunc::Memory);
4039 return true;
4040 }
4041
4042 //----------------------------clear_memory-------------------------------------
4043 // Generate code to initialize object storage to zero.
4044 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4045 Node* val,
4046 Node* raw_val,
4047 intptr_t start_offset,
4048 Node* end_offset,
4049 PhaseGVN* phase) {
4050 intptr_t offset = start_offset;
4051
4052 int unit = BytesPerLong;
4053 if ((offset % unit) != 0) {
4054 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
4055 adr = phase->transform(adr);
4056 const TypePtr* atp = TypeRawPtr::BOTTOM;
4057 if (val != nullptr) {
4058 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
4059 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
4060 } else {
4061 assert(raw_val == nullptr, "val may not be null");
4062 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4063 }
4064 mem = phase->transform(mem);
4065 offset += BytesPerInt;
4066 }
4067 assert((offset % unit) == 0, "");
4068
4069 // Initialize the remaining stuff, if any, with a ClearArray.
4070 return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, phase);
4071 }
4072
4073 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4074 Node* raw_val,
4075 Node* start_offset,
4076 Node* end_offset,
4077 PhaseGVN* phase) {
4078 if (start_offset == end_offset) {
4079 // nothing to do
4080 return mem;
4081 }
4082
4083 int unit = BytesPerLong;
4084 Node* zbase = start_offset;
4085 Node* zend = end_offset;
4086
4087 // Scale to the unit required by the CPU:
4088 if (!Matcher::init_array_count_is_in_bytes) {
4089 Node* shift = phase->intcon(exact_log2(unit));
4090 zbase = phase->transform(new URShiftXNode(zbase, shift) );
4091 zend = phase->transform(new URShiftXNode(zend, shift) );
4092 }
4093
4094 // Bulk clear double-words
4095 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
4096 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
4097 if (raw_val == nullptr) {
4098 raw_val = phase->MakeConX(0);
4099 }
4100 mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false);
4101 return phase->transform(mem);
4102 }
4103
4104 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4105 Node* val,
4106 Node* raw_val,
4107 intptr_t start_offset,
4108 intptr_t end_offset,
4109 PhaseGVN* phase) {
4110 if (start_offset == end_offset) {
4111 // nothing to do
4112 return mem;
4113 }
4114
4115 assert((end_offset % BytesPerInt) == 0, "odd end offset");
4116 intptr_t done_offset = end_offset;
4117 if ((done_offset % BytesPerLong) != 0) {
4118 done_offset -= BytesPerInt;
4119 }
4120 if (done_offset > start_offset) {
4121 mem = clear_memory(ctl, mem, dest, val, raw_val,
4122 start_offset, phase->MakeConX(done_offset), phase);
4123 }
4124 if (done_offset < end_offset) { // emit the final 32-bit store
4125 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
4126 adr = phase->transform(adr);
4127 const TypePtr* atp = TypeRawPtr::BOTTOM;
4128 if (val != nullptr) {
4129 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
4130 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
4131 } else {
4132 assert(raw_val == nullptr, "val may not be null");
4133 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4134 }
4135 mem = phase->transform(mem);
4136 done_offset += BytesPerInt;
4137 }
4138 assert(done_offset == end_offset, "");
4139 return mem;
4140 }
4141
4142 //=============================================================================
4143 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
4144 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
4145 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
4146 #ifdef ASSERT
4147 , _pair_idx(0)
4148 #endif
4149 {
4150 init_class_id(Class_MemBar);
4151 Node* top = C->top();
4152 init_req(TypeFunc::I_O,top);
4153 init_req(TypeFunc::FramePtr,top);
4154 init_req(TypeFunc::ReturnAdr,top);
4155 if (precedent != nullptr)
4156 init_req(TypeFunc::Parms, precedent);
4157 }
4158
4159 //------------------------------cmp--------------------------------------------
4160 uint MemBarNode::hash() const { return NO_HASH; }
4161 bool MemBarNode::cmp( const Node &n ) const {
4162 return (&n == this); // Always fail except on self
4163 }
4164
4165 //------------------------------make-------------------------------------------
4166 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
4167 switch (opcode) {
4168 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
4169 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
4170 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
4171 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
4172 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
4173 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn);
4174 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
4175 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
4176 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
4177 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
4178 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
4179 case Op_Initialize: return new InitializeNode(C, atp, pn);
4180 default: ShouldNotReachHere(); return nullptr;
4181 }
4182 }
4183
4184 void MemBarNode::remove(PhaseIterGVN *igvn) {
4185 if (outcnt() != 2) {
4186 assert(Opcode() == Op_Initialize, "Only seen when there are no use of init memory");
4187 assert(outcnt() == 1, "Only control then");
4188 }
4189 if (trailing_store() || trailing_load_store()) {
4190 MemBarNode* leading = leading_membar();
4191 if (leading != nullptr) {
4192 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
4193 leading->remove(igvn);
4194 }
4195 }
4196 if (proj_out_or_null(TypeFunc::Memory) != nullptr) {
4197 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
4198 }
4199 if (proj_out_or_null(TypeFunc::Control) != nullptr) {
4200 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
4201 }
4202 }
4203
4204 //------------------------------Ideal------------------------------------------
4205 // Return a node which is more "ideal" than the current node. Strip out
4206 // control copies
4207 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4208 if (remove_dead_region(phase, can_reshape)) return this;
4209 // Don't bother trying to transform a dead node
4210 if (in(0) && in(0)->is_top()) {
4211 return nullptr;
4212 }
4213
4214 bool progress = false;
4215 // Eliminate volatile MemBars for scalar replaced objects.
4216 if (can_reshape && req() == (Precedent+1)) {
4217 bool eliminate = false;
4218 int opc = Opcode();
4219 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
4220 // Volatile field loads and stores.
4221 Node* my_mem = in(MemBarNode::Precedent);
4222 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
4223 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
4224 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
4225 // replace this Precedent (decodeN) with the Load instead.
4226 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
4227 Node* load_node = my_mem->in(1);
4228 set_req(MemBarNode::Precedent, load_node);
4229 phase->is_IterGVN()->_worklist.push(my_mem);
4230 my_mem = load_node;
4231 } else {
4232 assert(my_mem->unique_out() == this, "sanity");
4233 assert(!trailing_load_store(), "load store node can't be eliminated");
4234 del_req(Precedent);
4235 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
4236 my_mem = nullptr;
4237 }
4238 progress = true;
4239 }
4240 if (my_mem != nullptr && my_mem->is_Mem()) {
4241 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
4242 // Check for scalar replaced object reference.
4243 if( t_oop != nullptr && t_oop->is_known_instance_field() &&
4244 t_oop->offset() != Type::OffsetBot &&
4245 t_oop->offset() != Type::OffsetTop) {
4246 eliminate = true;
4247 }
4248 }
4249 } else if (opc == Op_MemBarRelease || (UseStoreStoreForCtor && opc == Op_MemBarStoreStore)) {
4250 // Final field stores.
4251 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent));
4252 if ((alloc != nullptr) && alloc->is_Allocate() &&
4253 alloc->as_Allocate()->does_not_escape_thread()) {
4254 // The allocated object does not escape.
4255 eliminate = true;
4256 }
4257 }
4258 if (eliminate) {
4259 // Replace MemBar projections by its inputs.
4260 PhaseIterGVN* igvn = phase->is_IterGVN();
4261 remove(igvn);
4262 // Must return either the original node (now dead) or a new node
4263 // (Do not return a top here, since that would break the uniqueness of top.)
4264 return new ConINode(TypeInt::ZERO);
4265 }
4266 }
4267 return progress ? this : nullptr;
4268 }
4269
4270 //------------------------------Value------------------------------------------
4271 const Type* MemBarNode::Value(PhaseGVN* phase) const {
4272 if( !in(0) ) return Type::TOP;
4273 if( phase->type(in(0)) == Type::TOP )
4274 return Type::TOP;
4275 return TypeTuple::MEMBAR;
4276 }
4277
4278 //------------------------------match------------------------------------------
4279 // Construct projections for memory.
4280 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) {
4281 switch (proj->_con) {
4282 case TypeFunc::Control:
4283 case TypeFunc::Memory:
4284 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
4285 }
4286 ShouldNotReachHere();
4287 return nullptr;
4288 }
4289
4290 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4291 trailing->_kind = TrailingStore;
4292 leading->_kind = LeadingStore;
4293 #ifdef ASSERT
4294 trailing->_pair_idx = leading->_idx;
4295 leading->_pair_idx = leading->_idx;
4296 #endif
4297 }
4298
4299 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4300 trailing->_kind = TrailingLoadStore;
4301 leading->_kind = LeadingLoadStore;
4302 #ifdef ASSERT
4303 trailing->_pair_idx = leading->_idx;
4304 leading->_pair_idx = leading->_idx;
4305 #endif
4306 }
4307
4308 MemBarNode* MemBarNode::trailing_membar() const {
4309 ResourceMark rm;
4310 Node* trailing = (Node*)this;
4311 VectorSet seen;
4312 Node_Stack multis(0);
4313 do {
4314 Node* c = trailing;
4315 uint i = 0;
4316 do {
4317 trailing = nullptr;
4318 for (; i < c->outcnt(); i++) {
4319 Node* next = c->raw_out(i);
4320 if (next != c && next->is_CFG()) {
4321 if (c->is_MultiBranch()) {
4322 if (multis.node() == c) {
4323 multis.set_index(i+1);
4324 } else {
4325 multis.push(c, i+1);
4326 }
4327 }
4328 trailing = next;
4329 break;
4330 }
4331 }
4332 if (trailing != nullptr && !seen.test_set(trailing->_idx)) {
4333 break;
4334 }
4335 while (multis.size() > 0) {
4336 c = multis.node();
4337 i = multis.index();
4338 if (i < c->req()) {
4339 break;
4340 }
4341 multis.pop();
4342 }
4343 } while (multis.size() > 0);
4344 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
4345
4346 MemBarNode* mb = trailing->as_MemBar();
4347 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
4348 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
4349 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
4350 return mb;
4351 }
4352
4353 MemBarNode* MemBarNode::leading_membar() const {
4354 ResourceMark rm;
4355 VectorSet seen;
4356 Node_Stack regions(0);
4357 Node* leading = in(0);
4358 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
4359 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) {
4360 leading = nullptr;
4361 while (regions.size() > 0 && leading == nullptr) {
4362 Node* r = regions.node();
4363 uint i = regions.index();
4364 if (i < r->req()) {
4365 leading = r->in(i);
4366 regions.set_index(i+1);
4367 } else {
4368 regions.pop();
4369 }
4370 }
4371 if (leading == nullptr) {
4372 assert(regions.size() == 0, "all paths should have been tried");
4373 return nullptr;
4374 }
4375 }
4376 if (leading->is_Region()) {
4377 regions.push(leading, 2);
4378 leading = leading->in(1);
4379 } else {
4380 leading = leading->in(0);
4381 }
4382 }
4383 #ifdef ASSERT
4384 Unique_Node_List wq;
4385 wq.push((Node*)this);
4386 uint found = 0;
4387 for (uint i = 0; i < wq.size(); i++) {
4388 Node* n = wq.at(i);
4389 if (n->is_Region()) {
4390 for (uint j = 1; j < n->req(); j++) {
4391 Node* in = n->in(j);
4392 if (in != nullptr && !in->is_top()) {
4393 wq.push(in);
4394 }
4395 }
4396 } else {
4397 if (n->is_MemBar() && n->as_MemBar()->leading()) {
4398 assert(n == leading, "consistency check failed");
4399 found++;
4400 } else {
4401 Node* in = n->in(0);
4402 if (in != nullptr && !in->is_top()) {
4403 wq.push(in);
4404 }
4405 }
4406 }
4407 }
4408 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed");
4409 #endif
4410 if (leading == nullptr) {
4411 return nullptr;
4412 }
4413 MemBarNode* mb = leading->as_MemBar();
4414 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
4415 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
4416 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
4417 return mb;
4418 }
4419
4420
4421 //===========================InitializeNode====================================
4422 // SUMMARY:
4423 // This node acts as a memory barrier on raw memory, after some raw stores.
4424 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
4425 // The Initialize can 'capture' suitably constrained stores as raw inits.
4426 // It can coalesce related raw stores into larger units (called 'tiles').
4427 // It can avoid zeroing new storage for memory units which have raw inits.
4428 // At macro-expansion, it is marked 'complete', and does not optimize further.
4429 //
4430 // EXAMPLE:
4431 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
4432 // ctl = incoming control; mem* = incoming memory
4433 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
4434 // First allocate uninitialized memory and fill in the header:
4435 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
4436 // ctl := alloc.Control; mem* := alloc.Memory*
4437 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
4438 // Then initialize to zero the non-header parts of the raw memory block:
4439 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
4440 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
4441 // After the initialize node executes, the object is ready for service:
4442 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
4443 // Suppose its body is immediately initialized as {1,2}:
4444 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4445 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4446 // mem.SLICE(#short[*]) := store2
4447 //
4448 // DETAILS:
4449 // An InitializeNode collects and isolates object initialization after
4450 // an AllocateNode and before the next possible safepoint. As a
4451 // memory barrier (MemBarNode), it keeps critical stores from drifting
4452 // down past any safepoint or any publication of the allocation.
4453 // Before this barrier, a newly-allocated object may have uninitialized bits.
4454 // After this barrier, it may be treated as a real oop, and GC is allowed.
4455 //
4456 // The semantics of the InitializeNode include an implicit zeroing of
4457 // the new object from object header to the end of the object.
4458 // (The object header and end are determined by the AllocateNode.)
4459 //
4460 // Certain stores may be added as direct inputs to the InitializeNode.
4461 // These stores must update raw memory, and they must be to addresses
4462 // derived from the raw address produced by AllocateNode, and with
4463 // a constant offset. They must be ordered by increasing offset.
4464 // The first one is at in(RawStores), the last at in(req()-1).
4465 // Unlike most memory operations, they are not linked in a chain,
4466 // but are displayed in parallel as users of the rawmem output of
4467 // the allocation.
4468 //
4469 // (See comments in InitializeNode::capture_store, which continue
4470 // the example given above.)
4471 //
4472 // When the associated Allocate is macro-expanded, the InitializeNode
4473 // may be rewritten to optimize collected stores. A ClearArrayNode
4474 // may also be created at that point to represent any required zeroing.
4475 // The InitializeNode is then marked 'complete', prohibiting further
4476 // capturing of nearby memory operations.
4477 //
4478 // During macro-expansion, all captured initializations which store
4479 // constant values of 32 bits or smaller are coalesced (if advantageous)
4480 // into larger 'tiles' 32 or 64 bits. This allows an object to be
4481 // initialized in fewer memory operations. Memory words which are
4482 // covered by neither tiles nor non-constant stores are pre-zeroed
4483 // by explicit stores of zero. (The code shape happens to do all
4484 // zeroing first, then all other stores, with both sequences occurring
4485 // in order of ascending offsets.)
4486 //
4487 // Alternatively, code may be inserted between an AllocateNode and its
4488 // InitializeNode, to perform arbitrary initialization of the new object.
4489 // E.g., the object copying intrinsics insert complex data transfers here.
4490 // The initialization must then be marked as 'complete' disable the
4491 // built-in zeroing semantics and the collection of initializing stores.
4492 //
4493 // While an InitializeNode is incomplete, reads from the memory state
4494 // produced by it are optimizable if they match the control edge and
4495 // new oop address associated with the allocation/initialization.
4496 // They return a stored value (if the offset matches) or else zero.
4497 // A write to the memory state, if it matches control and address,
4498 // and if it is to a constant offset, may be 'captured' by the
4499 // InitializeNode. It is cloned as a raw memory operation and rewired
4500 // inside the initialization, to the raw oop produced by the allocation.
4501 // Operations on addresses which are provably distinct (e.g., to
4502 // other AllocateNodes) are allowed to bypass the initialization.
4503 //
4504 // The effect of all this is to consolidate object initialization
4505 // (both arrays and non-arrays, both piecewise and bulk) into a
4506 // single location, where it can be optimized as a unit.
4507 //
4508 // Only stores with an offset less than TrackedInitializationLimit words
4509 // will be considered for capture by an InitializeNode. This puts a
4510 // reasonable limit on the complexity of optimized initializations.
4511
4512 //---------------------------InitializeNode------------------------------------
4513 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
4514 : MemBarNode(C, adr_type, rawoop),
4515 _is_complete(Incomplete), _does_not_escape(false)
4516 {
4517 init_class_id(Class_Initialize);
4518
4519 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
4520 assert(in(RawAddress) == rawoop, "proper init");
4521 // Note: allocation() can be null, for secondary initialization barriers
4522 }
4523
4524 // Since this node is not matched, it will be processed by the
4525 // register allocator. Declare that there are no constraints
4526 // on the allocation of the RawAddress edge.
4527 const RegMask &InitializeNode::in_RegMask(uint idx) const {
4528 // This edge should be set to top, by the set_complete. But be conservative.
4529 if (idx == InitializeNode::RawAddress)
4530 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
4531 return RegMask::Empty;
4532 }
4533
4534 Node* InitializeNode::memory(uint alias_idx) {
4535 Node* mem = in(Memory);
4536 if (mem->is_MergeMem()) {
4537 return mem->as_MergeMem()->memory_at(alias_idx);
4538 } else {
4539 // incoming raw memory is not split
4540 return mem;
4541 }
4542 }
4543
4544 bool InitializeNode::is_non_zero() {
4545 if (is_complete()) return false;
4546 remove_extra_zeroes();
4547 return (req() > RawStores);
4548 }
4549
4550 void InitializeNode::set_complete(PhaseGVN* phase) {
4551 assert(!is_complete(), "caller responsibility");
4552 _is_complete = Complete;
4553
4554 // After this node is complete, it contains a bunch of
4555 // raw-memory initializations. There is no need for
4556 // it to have anything to do with non-raw memory effects.
4557 // Therefore, tell all non-raw users to re-optimize themselves,
4558 // after skipping the memory effects of this initialization.
4559 PhaseIterGVN* igvn = phase->is_IterGVN();
4560 if (igvn) igvn->add_users_to_worklist(this);
4561 }
4562
4563 // convenience function
4564 // return false if the init contains any stores already
4565 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
4566 InitializeNode* init = initialization();
4567 if (init == nullptr || init->is_complete()) {
4568 return false;
4569 }
4570 init->remove_extra_zeroes();
4571 // for now, if this allocation has already collected any inits, bail:
4572 if (init->is_non_zero()) return false;
4573 init->set_complete(phase);
4574 return true;
4575 }
4576
4577 void InitializeNode::remove_extra_zeroes() {
4578 if (req() == RawStores) return;
4579 Node* zmem = zero_memory();
4580 uint fill = RawStores;
4581 for (uint i = fill; i < req(); i++) {
4582 Node* n = in(i);
4583 if (n->is_top() || n == zmem) continue; // skip
4584 if (fill < i) set_req(fill, n); // compact
4585 ++fill;
4586 }
4587 // delete any empty spaces created:
4588 while (fill < req()) {
4589 del_req(fill);
4590 }
4591 }
4592
4593 // Helper for remembering which stores go with which offsets.
4594 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) {
4595 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
4596 intptr_t offset = -1;
4597 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
4598 phase, offset);
4599 if (base == nullptr) return -1; // something is dead,
4600 if (offset < 0) return -1; // dead, dead
4601 return offset;
4602 }
4603
4604 // Helper for proving that an initialization expression is
4605 // "simple enough" to be folded into an object initialization.
4606 // Attempts to prove that a store's initial value 'n' can be captured
4607 // within the initialization without creating a vicious cycle, such as:
4608 // { Foo p = new Foo(); p.next = p; }
4609 // True for constants and parameters and small combinations thereof.
4610 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
4611 ResourceMark rm;
4612 Unique_Node_List worklist;
4613 worklist.push(value);
4614
4615 uint complexity_limit = 20;
4616 for (uint j = 0; j < worklist.size(); j++) {
4617 if (j >= complexity_limit) {
4618 return false; // Bail out if processed too many nodes
4619 }
4620
4621 Node* n = worklist.at(j);
4622 if (n == nullptr) continue; // (can this really happen?)
4623 if (n->is_Proj()) n = n->in(0);
4624 if (n == this) return false; // found a cycle
4625 if (n->is_Con()) continue;
4626 if (n->is_Start()) continue; // params, etc., are OK
4627 if (n->is_Root()) continue; // even better
4628
4629 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
4630 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
4631 continue;
4632 }
4633
4634 Node* ctl = n->in(0);
4635 if (ctl != nullptr && !ctl->is_top()) {
4636 if (ctl->is_Proj()) ctl = ctl->in(0);
4637 if (ctl == this) return false;
4638
4639 // If we already know that the enclosing memory op is pinned right after
4640 // the init, then any control flow that the store has picked up
4641 // must have preceded the init, or else be equal to the init.
4642 // Even after loop optimizations (which might change control edges)
4643 // a store is never pinned *before* the availability of its inputs.
4644 if (!MemNode::all_controls_dominate(n, this)) {
4645 return false; // failed to prove a good control
4646 }
4647 }
4648
4649 // Check data edges for possible dependencies on 'this'.
4650 for (uint i = 1; i < n->req(); i++) {
4651 Node* m = n->in(i);
4652 if (m == nullptr || m == n || m->is_top()) continue;
4653
4654 // Only process data inputs once
4655 worklist.push(m);
4656 }
4657 }
4658
4659 return true;
4660 }
4661
4662 // Here are all the checks a Store must pass before it can be moved into
4663 // an initialization. Returns zero if a check fails.
4664 // On success, returns the (constant) offset to which the store applies,
4665 // within the initialized memory.
4666 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
4667 const int FAIL = 0;
4668 if (st->req() != MemNode::ValueIn + 1)
4669 return FAIL; // an inscrutable StoreNode (card mark?)
4670 Node* ctl = st->in(MemNode::Control);
4671 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this))
4672 return FAIL; // must be unconditional after the initialization
4673 Node* mem = st->in(MemNode::Memory);
4674 if (!(mem->is_Proj() && mem->in(0) == this))
4675 return FAIL; // must not be preceded by other stores
4676 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4677 if ((st->Opcode() == Op_StoreP || st->Opcode() == Op_StoreN) &&
4678 !bs->can_initialize_object(st)) {
4679 return FAIL;
4680 }
4681 Node* adr = st->in(MemNode::Address);
4682 intptr_t offset;
4683 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
4684 if (alloc == nullptr)
4685 return FAIL; // inscrutable address
4686 if (alloc != allocation())
4687 return FAIL; // wrong allocation! (store needs to float up)
4688 int size_in_bytes = st->memory_size();
4689 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
4690 return FAIL; // mismatched access
4691 }
4692 Node* val = st->in(MemNode::ValueIn);
4693
4694 if (!detect_init_independence(val, phase))
4695 return FAIL; // stored value must be 'simple enough'
4696
4697 // The Store can be captured only if nothing after the allocation
4698 // and before the Store is using the memory location that the store
4699 // overwrites.
4700 bool failed = false;
4701 // If is_complete_with_arraycopy() is true the shape of the graph is
4702 // well defined and is safe so no need for extra checks.
4703 if (!is_complete_with_arraycopy()) {
4704 // We are going to look at each use of the memory state following
4705 // the allocation to make sure nothing reads the memory that the
4706 // Store writes.
4707 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
4708 int alias_idx = phase->C->get_alias_index(t_adr);
4709 ResourceMark rm;
4710 Unique_Node_List mems;
4711 mems.push(mem);
4712 Node* unique_merge = nullptr;
4713 for (uint next = 0; next < mems.size(); ++next) {
4714 Node *m = mems.at(next);
4715 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
4716 Node *n = m->fast_out(j);
4717 if (n->outcnt() == 0) {
4718 continue;
4719 }
4720 if (n == st) {
4721 continue;
4722 } else if (n->in(0) != nullptr && n->in(0) != ctl) {
4723 // If the control of this use is different from the control
4724 // of the Store which is right after the InitializeNode then
4725 // this node cannot be between the InitializeNode and the
4726 // Store.
4727 continue;
4728 } else if (n->is_MergeMem()) {
4729 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
4730 // We can hit a MergeMemNode (that will likely go away
4731 // later) that is a direct use of the memory state
4732 // following the InitializeNode on the same slice as the
4733 // store node that we'd like to capture. We need to check
4734 // the uses of the MergeMemNode.
4735 mems.push(n);
4736 }
4737 } else if (n->is_Mem()) {
4738 Node* other_adr = n->in(MemNode::Address);
4739 if (other_adr == adr) {
4740 failed = true;
4741 break;
4742 } else {
4743 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
4744 if (other_t_adr != nullptr) {
4745 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
4746 if (other_alias_idx == alias_idx) {
4747 // A load from the same memory slice as the store right
4748 // after the InitializeNode. We check the control of the
4749 // object/array that is loaded from. If it's the same as
4750 // the store control then we cannot capture the store.
4751 assert(!n->is_Store(), "2 stores to same slice on same control?");
4752 Node* base = other_adr;
4753 if (base->is_Phi()) {
4754 // In rare case, base may be a PhiNode and it may read
4755 // the same memory slice between InitializeNode and store.
4756 failed = true;
4757 break;
4758 }
4759 assert(base->is_AddP(), "should be addp but is %s", base->Name());
4760 base = base->in(AddPNode::Base);
4761 if (base != nullptr) {
4762 base = base->uncast();
4763 if (base->is_Proj() && base->in(0) == alloc) {
4764 failed = true;
4765 break;
4766 }
4767 }
4768 }
4769 }
4770 }
4771 } else {
4772 failed = true;
4773 break;
4774 }
4775 }
4776 }
4777 }
4778 if (failed) {
4779 if (!can_reshape) {
4780 // We decided we couldn't capture the store during parsing. We
4781 // should try again during the next IGVN once the graph is
4782 // cleaner.
4783 phase->C->record_for_igvn(st);
4784 }
4785 return FAIL;
4786 }
4787
4788 return offset; // success
4789 }
4790
4791 // Find the captured store in(i) which corresponds to the range
4792 // [start..start+size) in the initialized object.
4793 // If there is one, return its index i. If there isn't, return the
4794 // negative of the index where it should be inserted.
4795 // Return 0 if the queried range overlaps an initialization boundary
4796 // or if dead code is encountered.
4797 // If size_in_bytes is zero, do not bother with overlap checks.
4798 int InitializeNode::captured_store_insertion_point(intptr_t start,
4799 int size_in_bytes,
4800 PhaseValues* phase) {
4801 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
4802
4803 if (is_complete())
4804 return FAIL; // arraycopy got here first; punt
4805
4806 assert(allocation() != nullptr, "must be present");
4807
4808 // no negatives, no header fields:
4809 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
4810
4811 // after a certain size, we bail out on tracking all the stores:
4812 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4813 if (start >= ti_limit) return FAIL;
4814
4815 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
4816 if (i >= limit) return -(int)i; // not found; here is where to put it
4817
4818 Node* st = in(i);
4819 intptr_t st_off = get_store_offset(st, phase);
4820 if (st_off < 0) {
4821 if (st != zero_memory()) {
4822 return FAIL; // bail out if there is dead garbage
4823 }
4824 } else if (st_off > start) {
4825 // ...we are done, since stores are ordered
4826 if (st_off < start + size_in_bytes) {
4827 return FAIL; // the next store overlaps
4828 }
4829 return -(int)i; // not found; here is where to put it
4830 } else if (st_off < start) {
4831 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
4832 if (size_in_bytes != 0 &&
4833 start < st_off + MAX_STORE &&
4834 start < st_off + st->as_Store()->memory_size()) {
4835 return FAIL; // the previous store overlaps
4836 }
4837 } else {
4838 if (size_in_bytes != 0 &&
4839 st->as_Store()->memory_size() != size_in_bytes) {
4840 return FAIL; // mismatched store size
4841 }
4842 return i;
4843 }
4844
4845 ++i;
4846 }
4847 }
4848
4849 // Look for a captured store which initializes at the offset 'start'
4850 // with the given size. If there is no such store, and no other
4851 // initialization interferes, then return zero_memory (the memory
4852 // projection of the AllocateNode).
4853 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
4854 PhaseValues* phase) {
4855 assert(stores_are_sane(phase), "");
4856 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4857 if (i == 0) {
4858 return nullptr; // something is dead
4859 } else if (i < 0) {
4860 return zero_memory(); // just primordial zero bits here
4861 } else {
4862 Node* st = in(i); // here is the store at this position
4863 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
4864 return st;
4865 }
4866 }
4867
4868 // Create, as a raw pointer, an address within my new object at 'offset'.
4869 Node* InitializeNode::make_raw_address(intptr_t offset,
4870 PhaseGVN* phase) {
4871 Node* addr = in(RawAddress);
4872 if (offset != 0) {
4873 Compile* C = phase->C;
4874 addr = phase->transform( new AddPNode(C->top(), addr,
4875 phase->MakeConX(offset)) );
4876 }
4877 return addr;
4878 }
4879
4880 // Clone the given store, converting it into a raw store
4881 // initializing a field or element of my new object.
4882 // Caller is responsible for retiring the original store,
4883 // with subsume_node or the like.
4884 //
4885 // From the example above InitializeNode::InitializeNode,
4886 // here are the old stores to be captured:
4887 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4888 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4889 //
4890 // Here is the changed code; note the extra edges on init:
4891 // alloc = (Allocate ...)
4892 // rawoop = alloc.RawAddress
4893 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
4894 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
4895 // init = (Initialize alloc.Control alloc.Memory rawoop
4896 // rawstore1 rawstore2)
4897 //
4898 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
4899 PhaseGVN* phase, bool can_reshape) {
4900 assert(stores_are_sane(phase), "");
4901
4902 if (start < 0) return nullptr;
4903 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
4904
4905 Compile* C = phase->C;
4906 int size_in_bytes = st->memory_size();
4907 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4908 if (i == 0) return nullptr; // bail out
4909 Node* prev_mem = nullptr; // raw memory for the captured store
4910 if (i > 0) {
4911 prev_mem = in(i); // there is a pre-existing store under this one
4912 set_req(i, C->top()); // temporarily disconnect it
4913 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
4914 } else {
4915 i = -i; // no pre-existing store
4916 prev_mem = zero_memory(); // a slice of the newly allocated object
4917 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
4918 set_req(--i, C->top()); // reuse this edge; it has been folded away
4919 else
4920 ins_req(i, C->top()); // build a new edge
4921 }
4922 Node* new_st = st->clone();
4923 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4924 new_st->set_req(MemNode::Control, in(Control));
4925 new_st->set_req(MemNode::Memory, prev_mem);
4926 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
4927 bs->eliminate_gc_barrier_data(new_st);
4928 new_st = phase->transform(new_st);
4929
4930 // At this point, new_st might have swallowed a pre-existing store
4931 // at the same offset, or perhaps new_st might have disappeared,
4932 // if it redundantly stored the same value (or zero to fresh memory).
4933
4934 // In any case, wire it in:
4935 PhaseIterGVN* igvn = phase->is_IterGVN();
4936 if (igvn) {
4937 igvn->rehash_node_delayed(this);
4938 }
4939 set_req(i, new_st);
4940
4941 // The caller may now kill the old guy.
4942 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
4943 assert(check_st == new_st || check_st == nullptr, "must be findable");
4944 assert(!is_complete(), "");
4945 return new_st;
4946 }
4947
4948 static bool store_constant(jlong* tiles, int num_tiles,
4949 intptr_t st_off, int st_size,
4950 jlong con) {
4951 if ((st_off & (st_size-1)) != 0)
4952 return false; // strange store offset (assume size==2**N)
4953 address addr = (address)tiles + st_off;
4954 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
4955 switch (st_size) {
4956 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
4957 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
4958 case sizeof(jint): *(jint*) addr = (jint) con; break;
4959 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
4960 default: return false; // strange store size (detect size!=2**N here)
4961 }
4962 return true; // return success to caller
4963 }
4964
4965 // Coalesce subword constants into int constants and possibly
4966 // into long constants. The goal, if the CPU permits,
4967 // is to initialize the object with a small number of 64-bit tiles.
4968 // Also, convert floating-point constants to bit patterns.
4969 // Non-constants are not relevant to this pass.
4970 //
4971 // In terms of the running example on InitializeNode::InitializeNode
4972 // and InitializeNode::capture_store, here is the transformation
4973 // of rawstore1 and rawstore2 into rawstore12:
4974 // alloc = (Allocate ...)
4975 // rawoop = alloc.RawAddress
4976 // tile12 = 0x00010002
4977 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
4978 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
4979 //
4980 void
4981 InitializeNode::coalesce_subword_stores(intptr_t header_size,
4982 Node* size_in_bytes,
4983 PhaseGVN* phase) {
4984 Compile* C = phase->C;
4985
4986 assert(stores_are_sane(phase), "");
4987 // Note: After this pass, they are not completely sane,
4988 // since there may be some overlaps.
4989
4990 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
4991
4992 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4993 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
4994 size_limit = MIN2(size_limit, ti_limit);
4995 size_limit = align_up(size_limit, BytesPerLong);
4996 int num_tiles = size_limit / BytesPerLong;
4997
4998 // allocate space for the tile map:
4999 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
5000 jlong tiles_buf[small_len];
5001 Node* nodes_buf[small_len];
5002 jlong inits_buf[small_len];
5003 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
5004 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5005 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
5006 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
5007 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
5008 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5009 // tiles: exact bitwise model of all primitive constants
5010 // nodes: last constant-storing node subsumed into the tiles model
5011 // inits: which bytes (in each tile) are touched by any initializations
5012
5013 //// Pass A: Fill in the tile model with any relevant stores.
5014
5015 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
5016 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
5017 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
5018 Node* zmem = zero_memory(); // initially zero memory state
5019 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5020 Node* st = in(i);
5021 intptr_t st_off = get_store_offset(st, phase);
5022
5023 // Figure out the store's offset and constant value:
5024 if (st_off < header_size) continue; //skip (ignore header)
5025 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
5026 int st_size = st->as_Store()->memory_size();
5027 if (st_off + st_size > size_limit) break;
5028
5029 // Record which bytes are touched, whether by constant or not.
5030 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
5031 continue; // skip (strange store size)
5032
5033 const Type* val = phase->type(st->in(MemNode::ValueIn));
5034 if (!val->singleton()) continue; //skip (non-con store)
5035 BasicType type = val->basic_type();
5036
5037 jlong con = 0;
5038 switch (type) {
5039 case T_INT: con = val->is_int()->get_con(); break;
5040 case T_LONG: con = val->is_long()->get_con(); break;
5041 case T_FLOAT: con = jint_cast(val->getf()); break;
5042 case T_DOUBLE: con = jlong_cast(val->getd()); break;
5043 default: continue; //skip (odd store type)
5044 }
5045
5046 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
5047 st->Opcode() == Op_StoreL) {
5048 continue; // This StoreL is already optimal.
5049 }
5050
5051 // Store down the constant.
5052 store_constant(tiles, num_tiles, st_off, st_size, con);
5053
5054 intptr_t j = st_off >> LogBytesPerLong;
5055
5056 if (type == T_INT && st_size == BytesPerInt
5057 && (st_off & BytesPerInt) == BytesPerInt) {
5058 jlong lcon = tiles[j];
5059 if (!Matcher::isSimpleConstant64(lcon) &&
5060 st->Opcode() == Op_StoreI) {
5061 // This StoreI is already optimal by itself.
5062 jint* intcon = (jint*) &tiles[j];
5063 intcon[1] = 0; // undo the store_constant()
5064
5065 // If the previous store is also optimal by itself, back up and
5066 // undo the action of the previous loop iteration... if we can.
5067 // But if we can't, just let the previous half take care of itself.
5068 st = nodes[j];
5069 st_off -= BytesPerInt;
5070 con = intcon[0];
5071 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) {
5072 assert(st_off >= header_size, "still ignoring header");
5073 assert(get_store_offset(st, phase) == st_off, "must be");
5074 assert(in(i-1) == zmem, "must be");
5075 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
5076 assert(con == tcon->is_int()->get_con(), "must be");
5077 // Undo the effects of the previous loop trip, which swallowed st:
5078 intcon[0] = 0; // undo store_constant()
5079 set_req(i-1, st); // undo set_req(i, zmem)
5080 nodes[j] = nullptr; // undo nodes[j] = st
5081 --old_subword; // undo ++old_subword
5082 }
5083 continue; // This StoreI is already optimal.
5084 }
5085 }
5086
5087 // This store is not needed.
5088 set_req(i, zmem);
5089 nodes[j] = st; // record for the moment
5090 if (st_size < BytesPerLong) // something has changed
5091 ++old_subword; // includes int/float, but who's counting...
5092 else ++old_long;
5093 }
5094
5095 if ((old_subword + old_long) == 0)
5096 return; // nothing more to do
5097
5098 //// Pass B: Convert any non-zero tiles into optimal constant stores.
5099 // Be sure to insert them before overlapping non-constant stores.
5100 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
5101 for (int j = 0; j < num_tiles; j++) {
5102 jlong con = tiles[j];
5103 jlong init = inits[j];
5104 if (con == 0) continue;
5105 jint con0, con1; // split the constant, address-wise
5106 jint init0, init1; // split the init map, address-wise
5107 { union { jlong con; jint intcon[2]; } u;
5108 u.con = con;
5109 con0 = u.intcon[0];
5110 con1 = u.intcon[1];
5111 u.con = init;
5112 init0 = u.intcon[0];
5113 init1 = u.intcon[1];
5114 }
5115
5116 Node* old = nodes[j];
5117 assert(old != nullptr, "need the prior store");
5118 intptr_t offset = (j * BytesPerLong);
5119
5120 bool split = !Matcher::isSimpleConstant64(con);
5121
5122 if (offset < header_size) {
5123 assert(offset + BytesPerInt >= header_size, "second int counts");
5124 assert(*(jint*)&tiles[j] == 0, "junk in header");
5125 split = true; // only the second word counts
5126 // Example: int a[] = { 42 ... }
5127 } else if (con0 == 0 && init0 == -1) {
5128 split = true; // first word is covered by full inits
5129 // Example: int a[] = { ... foo(), 42 ... }
5130 } else if (con1 == 0 && init1 == -1) {
5131 split = true; // second word is covered by full inits
5132 // Example: int a[] = { ... 42, foo() ... }
5133 }
5134
5135 // Here's a case where init0 is neither 0 nor -1:
5136 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
5137 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
5138 // In this case the tile is not split; it is (jlong)42.
5139 // The big tile is stored down, and then the foo() value is inserted.
5140 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
5141
5142 Node* ctl = old->in(MemNode::Control);
5143 Node* adr = make_raw_address(offset, phase);
5144 const TypePtr* atp = TypeRawPtr::BOTTOM;
5145
5146 // One or two coalesced stores to plop down.
5147 Node* st[2];
5148 intptr_t off[2];
5149 int nst = 0;
5150 if (!split) {
5151 ++new_long;
5152 off[nst] = offset;
5153 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5154 phase->longcon(con), T_LONG, MemNode::unordered);
5155 } else {
5156 // Omit either if it is a zero.
5157 if (con0 != 0) {
5158 ++new_int;
5159 off[nst] = offset;
5160 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5161 phase->intcon(con0), T_INT, MemNode::unordered);
5162 }
5163 if (con1 != 0) {
5164 ++new_int;
5165 offset += BytesPerInt;
5166 adr = make_raw_address(offset, phase);
5167 off[nst] = offset;
5168 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5169 phase->intcon(con1), T_INT, MemNode::unordered);
5170 }
5171 }
5172
5173 // Insert second store first, then the first before the second.
5174 // Insert each one just before any overlapping non-constant stores.
5175 while (nst > 0) {
5176 Node* st1 = st[--nst];
5177 C->copy_node_notes_to(st1, old);
5178 st1 = phase->transform(st1);
5179 offset = off[nst];
5180 assert(offset >= header_size, "do not smash header");
5181 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
5182 guarantee(ins_idx != 0, "must re-insert constant store");
5183 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
5184 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
5185 set_req(--ins_idx, st1);
5186 else
5187 ins_req(ins_idx, st1);
5188 }
5189 }
5190
5191 if (PrintCompilation && WizardMode)
5192 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
5193 old_subword, old_long, new_int, new_long);
5194 if (C->log() != nullptr)
5195 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
5196 old_subword, old_long, new_int, new_long);
5197
5198 // Clean up any remaining occurrences of zmem:
5199 remove_extra_zeroes();
5200 }
5201
5202 // Explore forward from in(start) to find the first fully initialized
5203 // word, and return its offset. Skip groups of subword stores which
5204 // together initialize full words. If in(start) is itself part of a
5205 // fully initialized word, return the offset of in(start). If there
5206 // are no following full-word stores, or if something is fishy, return
5207 // a negative value.
5208 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
5209 int int_map = 0;
5210 intptr_t int_map_off = 0;
5211 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
5212
5213 for (uint i = start, limit = req(); i < limit; i++) {
5214 Node* st = in(i);
5215
5216 intptr_t st_off = get_store_offset(st, phase);
5217 if (st_off < 0) break; // return conservative answer
5218
5219 int st_size = st->as_Store()->memory_size();
5220 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
5221 return st_off; // we found a complete word init
5222 }
5223
5224 // update the map:
5225
5226 intptr_t this_int_off = align_down(st_off, BytesPerInt);
5227 if (this_int_off != int_map_off) {
5228 // reset the map:
5229 int_map = 0;
5230 int_map_off = this_int_off;
5231 }
5232
5233 int subword_off = st_off - this_int_off;
5234 int_map |= right_n_bits(st_size) << subword_off;
5235 if ((int_map & FULL_MAP) == FULL_MAP) {
5236 return this_int_off; // we found a complete word init
5237 }
5238
5239 // Did this store hit or cross the word boundary?
5240 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
5241 if (next_int_off == this_int_off + BytesPerInt) {
5242 // We passed the current int, without fully initializing it.
5243 int_map_off = next_int_off;
5244 int_map >>= BytesPerInt;
5245 } else if (next_int_off > this_int_off + BytesPerInt) {
5246 // We passed the current and next int.
5247 return this_int_off + BytesPerInt;
5248 }
5249 }
5250
5251 return -1;
5252 }
5253
5254
5255 // Called when the associated AllocateNode is expanded into CFG.
5256 // At this point, we may perform additional optimizations.
5257 // Linearize the stores by ascending offset, to make memory
5258 // activity as coherent as possible.
5259 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
5260 intptr_t header_size,
5261 Node* size_in_bytes,
5262 PhaseIterGVN* phase) {
5263 assert(!is_complete(), "not already complete");
5264 assert(stores_are_sane(phase), "");
5265 assert(allocation() != nullptr, "must be present");
5266
5267 remove_extra_zeroes();
5268
5269 if (ReduceFieldZeroing || ReduceBulkZeroing)
5270 // reduce instruction count for common initialization patterns
5271 coalesce_subword_stores(header_size, size_in_bytes, phase);
5272
5273 Node* zmem = zero_memory(); // initially zero memory state
5274 Node* inits = zmem; // accumulating a linearized chain of inits
5275 #ifdef ASSERT
5276 intptr_t first_offset = allocation()->minimum_header_size();
5277 intptr_t last_init_off = first_offset; // previous init offset
5278 intptr_t last_init_end = first_offset; // previous init offset+size
5279 intptr_t last_tile_end = first_offset; // previous tile offset+size
5280 #endif
5281 intptr_t zeroes_done = header_size;
5282
5283 bool do_zeroing = true; // we might give up if inits are very sparse
5284 int big_init_gaps = 0; // how many large gaps have we seen?
5285
5286 if (UseTLAB && ZeroTLAB) do_zeroing = false;
5287 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
5288
5289 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5290 Node* st = in(i);
5291 intptr_t st_off = get_store_offset(st, phase);
5292 if (st_off < 0)
5293 break; // unknown junk in the inits
5294 if (st->in(MemNode::Memory) != zmem)
5295 break; // complicated store chains somehow in list
5296
5297 int st_size = st->as_Store()->memory_size();
5298 intptr_t next_init_off = st_off + st_size;
5299
5300 if (do_zeroing && zeroes_done < next_init_off) {
5301 // See if this store needs a zero before it or under it.
5302 intptr_t zeroes_needed = st_off;
5303
5304 if (st_size < BytesPerInt) {
5305 // Look for subword stores which only partially initialize words.
5306 // If we find some, we must lay down some word-level zeroes first,
5307 // underneath the subword stores.
5308 //
5309 // Examples:
5310 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
5311 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
5312 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
5313 //
5314 // Note: coalesce_subword_stores may have already done this,
5315 // if it was prompted by constant non-zero subword initializers.
5316 // But this case can still arise with non-constant stores.
5317
5318 intptr_t next_full_store = find_next_fullword_store(i, phase);
5319
5320 // In the examples above:
5321 // in(i) p q r s x y z
5322 // st_off 12 13 14 15 12 13 14
5323 // st_size 1 1 1 1 1 1 1
5324 // next_full_s. 12 16 16 16 16 16 16
5325 // z's_done 12 16 16 16 12 16 12
5326 // z's_needed 12 16 16 16 16 16 16
5327 // zsize 0 0 0 0 4 0 4
5328 if (next_full_store < 0) {
5329 // Conservative tack: Zero to end of current word.
5330 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
5331 } else {
5332 // Zero to beginning of next fully initialized word.
5333 // Or, don't zero at all, if we are already in that word.
5334 assert(next_full_store >= zeroes_needed, "must go forward");
5335 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
5336 zeroes_needed = next_full_store;
5337 }
5338 }
5339
5340 if (zeroes_needed > zeroes_done) {
5341 intptr_t zsize = zeroes_needed - zeroes_done;
5342 // Do some incremental zeroing on rawmem, in parallel with inits.
5343 zeroes_done = align_down(zeroes_done, BytesPerInt);
5344 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5345 allocation()->in(AllocateNode::DefaultValue),
5346 allocation()->in(AllocateNode::RawDefaultValue),
5347 zeroes_done, zeroes_needed,
5348 phase);
5349 zeroes_done = zeroes_needed;
5350 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
5351 do_zeroing = false; // leave the hole, next time
5352 }
5353 }
5354
5355 // Collect the store and move on:
5356 phase->replace_input_of(st, MemNode::Memory, inits);
5357 inits = st; // put it on the linearized chain
5358 set_req(i, zmem); // unhook from previous position
5359
5360 if (zeroes_done == st_off)
5361 zeroes_done = next_init_off;
5362
5363 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
5364
5365 #ifdef ASSERT
5366 // Various order invariants. Weaker than stores_are_sane because
5367 // a large constant tile can be filled in by smaller non-constant stores.
5368 assert(st_off >= last_init_off, "inits do not reverse");
5369 last_init_off = st_off;
5370 const Type* val = nullptr;
5371 if (st_size >= BytesPerInt &&
5372 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
5373 (int)val->basic_type() < (int)T_OBJECT) {
5374 assert(st_off >= last_tile_end, "tiles do not overlap");
5375 assert(st_off >= last_init_end, "tiles do not overwrite inits");
5376 last_tile_end = MAX2(last_tile_end, next_init_off);
5377 } else {
5378 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
5379 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
5380 assert(st_off >= last_init_end, "inits do not overlap");
5381 last_init_end = next_init_off; // it's a non-tile
5382 }
5383 #endif //ASSERT
5384 }
5385
5386 remove_extra_zeroes(); // clear out all the zmems left over
5387 add_req(inits);
5388
5389 if (!(UseTLAB && ZeroTLAB)) {
5390 // If anything remains to be zeroed, zero it all now.
5391 zeroes_done = align_down(zeroes_done, BytesPerInt);
5392 // if it is the last unused 4 bytes of an instance, forget about it
5393 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
5394 if (zeroes_done + BytesPerLong >= size_limit) {
5395 AllocateNode* alloc = allocation();
5396 assert(alloc != nullptr, "must be present");
5397 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
5398 Node* klass_node = alloc->in(AllocateNode::KlassNode);
5399 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
5400 if (zeroes_done == k->layout_helper())
5401 zeroes_done = size_limit;
5402 }
5403 }
5404 if (zeroes_done < size_limit) {
5405 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5406 allocation()->in(AllocateNode::DefaultValue),
5407 allocation()->in(AllocateNode::RawDefaultValue),
5408 zeroes_done, size_in_bytes, phase);
5409 }
5410 }
5411
5412 set_complete(phase);
5413 return rawmem;
5414 }
5415
5416
5417 #ifdef ASSERT
5418 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
5419 if (is_complete())
5420 return true; // stores could be anything at this point
5421 assert(allocation() != nullptr, "must be present");
5422 intptr_t last_off = allocation()->minimum_header_size();
5423 for (uint i = InitializeNode::RawStores; i < req(); i++) {
5424 Node* st = in(i);
5425 intptr_t st_off = get_store_offset(st, phase);
5426 if (st_off < 0) continue; // ignore dead garbage
5427 if (last_off > st_off) {
5428 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
5429 this->dump(2);
5430 assert(false, "ascending store offsets");
5431 return false;
5432 }
5433 last_off = st_off + st->as_Store()->memory_size();
5434 }
5435 return true;
5436 }
5437 #endif //ASSERT
5438
5439
5440
5441
5442 //============================MergeMemNode=====================================
5443 //
5444 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
5445 // contributing store or call operations. Each contributor provides the memory
5446 // state for a particular "alias type" (see Compile::alias_type). For example,
5447 // if a MergeMem has an input X for alias category #6, then any memory reference
5448 // to alias category #6 may use X as its memory state input, as an exact equivalent
5449 // to using the MergeMem as a whole.
5450 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
5451 //
5452 // (Here, the <N> notation gives the index of the relevant adr_type.)
5453 //
5454 // In one special case (and more cases in the future), alias categories overlap.
5455 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
5456 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
5457 // it is exactly equivalent to that state W:
5458 // MergeMem(<Bot>: W) <==> W
5459 //
5460 // Usually, the merge has more than one input. In that case, where inputs
5461 // overlap (i.e., one is Bot), the narrower alias type determines the memory
5462 // state for that type, and the wider alias type (Bot) fills in everywhere else:
5463 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
5464 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
5465 //
5466 // A merge can take a "wide" memory state as one of its narrow inputs.
5467 // This simply means that the merge observes out only the relevant parts of
5468 // the wide input. That is, wide memory states arriving at narrow merge inputs
5469 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
5470 //
5471 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
5472 // and that memory slices "leak through":
5473 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
5474 //
5475 // But, in such a cascade, repeated memory slices can "block the leak":
5476 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
5477 //
5478 // In the last example, Y is not part of the combined memory state of the
5479 // outermost MergeMem. The system must, of course, prevent unschedulable
5480 // memory states from arising, so you can be sure that the state Y is somehow
5481 // a precursor to state Y'.
5482 //
5483 //
5484 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
5485 // of each MergeMemNode array are exactly the numerical alias indexes, including
5486 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
5487 // Compile::alias_type (and kin) produce and manage these indexes.
5488 //
5489 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
5490 // (Note that this provides quick access to the top node inside MergeMem methods,
5491 // without the need to reach out via TLS to Compile::current.)
5492 //
5493 // As a consequence of what was just described, a MergeMem that represents a full
5494 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
5495 // containing all alias categories.
5496 //
5497 // MergeMem nodes never (?) have control inputs, so in(0) is null.
5498 //
5499 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
5500 // a memory state for the alias type <N>, or else the top node, meaning that
5501 // there is no particular input for that alias type. Note that the length of
5502 // a MergeMem is variable, and may be extended at any time to accommodate new
5503 // memory states at larger alias indexes. When merges grow, they are of course
5504 // filled with "top" in the unused in() positions.
5505 //
5506 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
5507 // (Top was chosen because it works smoothly with passes like GCM.)
5508 //
5509 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
5510 // the type of random VM bits like TLS references.) Since it is always the
5511 // first non-Bot memory slice, some low-level loops use it to initialize an
5512 // index variable: for (i = AliasIdxRaw; i < req(); i++).
5513 //
5514 //
5515 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
5516 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
5517 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
5518 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
5519 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
5520 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
5521 //
5522 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
5523 // really that different from the other memory inputs. An abbreviation called
5524 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
5525 //
5526 //
5527 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
5528 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
5529 // that "emerges though" the base memory will be marked as excluding the alias types
5530 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
5531 //
5532 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
5533 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
5534 //
5535 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
5536 // (It is currently unimplemented.) As you can see, the resulting merge is
5537 // actually a disjoint union of memory states, rather than an overlay.
5538 //
5539
5540 //------------------------------MergeMemNode-----------------------------------
5541 Node* MergeMemNode::make_empty_memory() {
5542 Node* empty_memory = (Node*) Compile::current()->top();
5543 assert(empty_memory->is_top(), "correct sentinel identity");
5544 return empty_memory;
5545 }
5546
5547 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
5548 init_class_id(Class_MergeMem);
5549 // all inputs are nullified in Node::Node(int)
5550 // set_input(0, nullptr); // no control input
5551
5552 // Initialize the edges uniformly to top, for starters.
5553 Node* empty_mem = make_empty_memory();
5554 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
5555 init_req(i,empty_mem);
5556 }
5557 assert(empty_memory() == empty_mem, "");
5558
5559 if( new_base != nullptr && new_base->is_MergeMem() ) {
5560 MergeMemNode* mdef = new_base->as_MergeMem();
5561 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
5562 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
5563 mms.set_memory(mms.memory2());
5564 }
5565 assert(base_memory() == mdef->base_memory(), "");
5566 } else {
5567 set_base_memory(new_base);
5568 }
5569 }
5570
5571 // Make a new, untransformed MergeMem with the same base as 'mem'.
5572 // If mem is itself a MergeMem, populate the result with the same edges.
5573 MergeMemNode* MergeMemNode::make(Node* mem) {
5574 return new MergeMemNode(mem);
5575 }
5576
5577 //------------------------------cmp--------------------------------------------
5578 uint MergeMemNode::hash() const { return NO_HASH; }
5579 bool MergeMemNode::cmp( const Node &n ) const {
5580 return (&n == this); // Always fail except on self
5581 }
5582
5583 //------------------------------Identity---------------------------------------
5584 Node* MergeMemNode::Identity(PhaseGVN* phase) {
5585 // Identity if this merge point does not record any interesting memory
5586 // disambiguations.
5587 Node* base_mem = base_memory();
5588 Node* empty_mem = empty_memory();
5589 if (base_mem != empty_mem) { // Memory path is not dead?
5590 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5591 Node* mem = in(i);
5592 if (mem != empty_mem && mem != base_mem) {
5593 return this; // Many memory splits; no change
5594 }
5595 }
5596 }
5597 return base_mem; // No memory splits; ID on the one true input
5598 }
5599
5600 //------------------------------Ideal------------------------------------------
5601 // This method is invoked recursively on chains of MergeMem nodes
5602 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
5603 // Remove chain'd MergeMems
5604 //
5605 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
5606 // relative to the "in(Bot)". Since we are patching both at the same time,
5607 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
5608 // but rewrite each "in(i)" relative to the new "in(Bot)".
5609 Node *progress = nullptr;
5610
5611
5612 Node* old_base = base_memory();
5613 Node* empty_mem = empty_memory();
5614 if (old_base == empty_mem)
5615 return nullptr; // Dead memory path.
5616
5617 MergeMemNode* old_mbase;
5618 if (old_base != nullptr && old_base->is_MergeMem())
5619 old_mbase = old_base->as_MergeMem();
5620 else
5621 old_mbase = nullptr;
5622 Node* new_base = old_base;
5623
5624 // simplify stacked MergeMems in base memory
5625 if (old_mbase) new_base = old_mbase->base_memory();
5626
5627 // the base memory might contribute new slices beyond my req()
5628 if (old_mbase) grow_to_match(old_mbase);
5629
5630 // Note: We do not call verify_sparse on entry, because inputs
5631 // can normalize to the base_memory via subsume_node or similar
5632 // mechanisms. This method repairs that damage.
5633
5634 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
5635
5636 // Look at each slice.
5637 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5638 Node* old_in = in(i);
5639 // calculate the old memory value
5640 Node* old_mem = old_in;
5641 if (old_mem == empty_mem) old_mem = old_base;
5642 assert(old_mem == memory_at(i), "");
5643
5644 // maybe update (reslice) the old memory value
5645
5646 // simplify stacked MergeMems
5647 Node* new_mem = old_mem;
5648 MergeMemNode* old_mmem;
5649 if (old_mem != nullptr && old_mem->is_MergeMem())
5650 old_mmem = old_mem->as_MergeMem();
5651 else
5652 old_mmem = nullptr;
5653 if (old_mmem == this) {
5654 // This can happen if loops break up and safepoints disappear.
5655 // A merge of BotPtr (default) with a RawPtr memory derived from a
5656 // safepoint can be rewritten to a merge of the same BotPtr with
5657 // the BotPtr phi coming into the loop. If that phi disappears
5658 // also, we can end up with a self-loop of the mergemem.
5659 // In general, if loops degenerate and memory effects disappear,
5660 // a mergemem can be left looking at itself. This simply means
5661 // that the mergemem's default should be used, since there is
5662 // no longer any apparent effect on this slice.
5663 // Note: If a memory slice is a MergeMem cycle, it is unreachable
5664 // from start. Update the input to TOP.
5665 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
5666 }
5667 else if (old_mmem != nullptr) {
5668 new_mem = old_mmem->memory_at(i);
5669 }
5670 // else preceding memory was not a MergeMem
5671
5672 // maybe store down a new value
5673 Node* new_in = new_mem;
5674 if (new_in == new_base) new_in = empty_mem;
5675
5676 if (new_in != old_in) {
5677 // Warning: Do not combine this "if" with the previous "if"
5678 // A memory slice might have be be rewritten even if it is semantically
5679 // unchanged, if the base_memory value has changed.
5680 set_req_X(i, new_in, phase);
5681 progress = this; // Report progress
5682 }
5683 }
5684
5685 if (new_base != old_base) {
5686 set_req_X(Compile::AliasIdxBot, new_base, phase);
5687 // Don't use set_base_memory(new_base), because we need to update du.
5688 assert(base_memory() == new_base, "");
5689 progress = this;
5690 }
5691
5692 if( base_memory() == this ) {
5693 // a self cycle indicates this memory path is dead
5694 set_req(Compile::AliasIdxBot, empty_mem);
5695 }
5696
5697 // Resolve external cycles by calling Ideal on a MergeMem base_memory
5698 // Recursion must occur after the self cycle check above
5699 if( base_memory()->is_MergeMem() ) {
5700 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
5701 Node *m = phase->transform(new_mbase); // Rollup any cycles
5702 if( m != nullptr &&
5703 (m->is_top() ||
5704 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
5705 // propagate rollup of dead cycle to self
5706 set_req(Compile::AliasIdxBot, empty_mem);
5707 }
5708 }
5709
5710 if( base_memory() == empty_mem ) {
5711 progress = this;
5712 // Cut inputs during Parse phase only.
5713 // During Optimize phase a dead MergeMem node will be subsumed by Top.
5714 if( !can_reshape ) {
5715 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5716 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
5717 }
5718 }
5719 }
5720
5721 if( !progress && base_memory()->is_Phi() && can_reshape ) {
5722 // Check if PhiNode::Ideal's "Split phis through memory merges"
5723 // transform should be attempted. Look for this->phi->this cycle.
5724 uint merge_width = req();
5725 if (merge_width > Compile::AliasIdxRaw) {
5726 PhiNode* phi = base_memory()->as_Phi();
5727 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
5728 if (phi->in(i) == this) {
5729 phase->is_IterGVN()->_worklist.push(phi);
5730 break;
5731 }
5732 }
5733 }
5734 }
5735
5736 assert(progress || verify_sparse(), "please, no dups of base");
5737 return progress;
5738 }
5739
5740 //-------------------------set_base_memory-------------------------------------
5741 void MergeMemNode::set_base_memory(Node *new_base) {
5742 Node* empty_mem = empty_memory();
5743 set_req(Compile::AliasIdxBot, new_base);
5744 assert(memory_at(req()) == new_base, "must set default memory");
5745 // Clear out other occurrences of new_base:
5746 if (new_base != empty_mem) {
5747 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5748 if (in(i) == new_base) set_req(i, empty_mem);
5749 }
5750 }
5751 }
5752
5753 //------------------------------out_RegMask------------------------------------
5754 const RegMask &MergeMemNode::out_RegMask() const {
5755 return RegMask::Empty;
5756 }
5757
5758 //------------------------------dump_spec--------------------------------------
5759 #ifndef PRODUCT
5760 void MergeMemNode::dump_spec(outputStream *st) const {
5761 st->print(" {");
5762 Node* base_mem = base_memory();
5763 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
5764 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem;
5765 if (mem == base_mem) { st->print(" -"); continue; }
5766 st->print( " N%d:", mem->_idx );
5767 Compile::current()->get_adr_type(i)->dump_on(st);
5768 }
5769 st->print(" }");
5770 }
5771 #endif // !PRODUCT
5772
5773
5774 #ifdef ASSERT
5775 static bool might_be_same(Node* a, Node* b) {
5776 if (a == b) return true;
5777 if (!(a->is_Phi() || b->is_Phi())) return false;
5778 // phis shift around during optimization
5779 return true; // pretty stupid...
5780 }
5781
5782 // verify a narrow slice (either incoming or outgoing)
5783 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
5784 if (!VerifyAliases) return; // don't bother to verify unless requested
5785 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
5786 if (Node::in_dump()) return; // muzzle asserts when printing
5787 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
5788 assert(n != nullptr, "");
5789 // Elide intervening MergeMem's
5790 while (n->is_MergeMem()) {
5791 n = n->as_MergeMem()->memory_at(alias_idx);
5792 }
5793 Compile* C = Compile::current();
5794 const TypePtr* n_adr_type = n->adr_type();
5795 if (n == m->empty_memory()) {
5796 // Implicit copy of base_memory()
5797 } else if (n_adr_type != TypePtr::BOTTOM) {
5798 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type");
5799 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
5800 } else {
5801 // A few places like make_runtime_call "know" that VM calls are narrow,
5802 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
5803 bool expected_wide_mem = false;
5804 if (n == m->base_memory()) {
5805 expected_wide_mem = true;
5806 } else if (alias_idx == Compile::AliasIdxRaw ||
5807 n == m->memory_at(Compile::AliasIdxRaw)) {
5808 expected_wide_mem = true;
5809 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
5810 // memory can "leak through" calls on channels that
5811 // are write-once. Allow this also.
5812 expected_wide_mem = true;
5813 }
5814 assert(expected_wide_mem, "expected narrow slice replacement");
5815 }
5816 }
5817 #else // !ASSERT
5818 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
5819 #endif
5820
5821
5822 //-----------------------------memory_at---------------------------------------
5823 Node* MergeMemNode::memory_at(uint alias_idx) const {
5824 assert(alias_idx >= Compile::AliasIdxRaw ||
5825 (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()),
5826 "must avoid base_memory and AliasIdxTop");
5827
5828 // Otherwise, it is a narrow slice.
5829 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
5830 if (is_empty_memory(n)) {
5831 // the array is sparse; empty slots are the "top" node
5832 n = base_memory();
5833 assert(Node::in_dump()
5834 || n == nullptr || n->bottom_type() == Type::TOP
5835 || n->adr_type() == nullptr // address is TOP
5836 || n->adr_type() == TypePtr::BOTTOM
5837 || n->adr_type() == TypeRawPtr::BOTTOM
5838 || !Compile::current()->do_aliasing(),
5839 "must be a wide memory");
5840 // do_aliasing == false if we are organizing the memory states manually.
5841 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
5842 } else {
5843 // make sure the stored slice is sane
5844 #ifdef ASSERT
5845 if (VMError::is_error_reported() || Node::in_dump()) {
5846 } else if (might_be_same(n, base_memory())) {
5847 // Give it a pass: It is a mostly harmless repetition of the base.
5848 // This can arise normally from node subsumption during optimization.
5849 } else {
5850 verify_memory_slice(this, alias_idx, n);
5851 }
5852 #endif
5853 }
5854 return n;
5855 }
5856
5857 //---------------------------set_memory_at-------------------------------------
5858 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
5859 verify_memory_slice(this, alias_idx, n);
5860 Node* empty_mem = empty_memory();
5861 if (n == base_memory()) n = empty_mem; // collapse default
5862 uint need_req = alias_idx+1;
5863 if (req() < need_req) {
5864 if (n == empty_mem) return; // already the default, so do not grow me
5865 // grow the sparse array
5866 do {
5867 add_req(empty_mem);
5868 } while (req() < need_req);
5869 }
5870 set_req( alias_idx, n );
5871 }
5872
5873
5874
5875 //--------------------------iteration_setup------------------------------------
5876 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
5877 if (other != nullptr) {
5878 grow_to_match(other);
5879 // invariant: the finite support of mm2 is within mm->req()
5880 #ifdef ASSERT
5881 for (uint i = req(); i < other->req(); i++) {
5882 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
5883 }
5884 #endif
5885 }
5886 // Replace spurious copies of base_memory by top.
5887 Node* base_mem = base_memory();
5888 if (base_mem != nullptr && !base_mem->is_top()) {
5889 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
5890 if (in(i) == base_mem)
5891 set_req(i, empty_memory());
5892 }
5893 }
5894 }
5895
5896 //---------------------------grow_to_match-------------------------------------
5897 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
5898 Node* empty_mem = empty_memory();
5899 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
5900 // look for the finite support of the other memory
5901 for (uint i = other->req(); --i >= req(); ) {
5902 if (other->in(i) != empty_mem) {
5903 uint new_len = i+1;
5904 while (req() < new_len) add_req(empty_mem);
5905 break;
5906 }
5907 }
5908 }
5909
5910 //---------------------------verify_sparse-------------------------------------
5911 #ifndef PRODUCT
5912 bool MergeMemNode::verify_sparse() const {
5913 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
5914 Node* base_mem = base_memory();
5915 // The following can happen in degenerate cases, since empty==top.
5916 if (is_empty_memory(base_mem)) return true;
5917 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5918 assert(in(i) != nullptr, "sane slice");
5919 if (in(i) == base_mem) return false; // should have been the sentinel value!
5920 }
5921 return true;
5922 }
5923
5924 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
5925 Node* n;
5926 n = mm->in(idx);
5927 if (mem == n) return true; // might be empty_memory()
5928 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
5929 if (mem == n) return true;
5930 return false;
5931 }
5932 #endif // !PRODUCT