1 /*
2 * Copyright (c) 1997, 2026, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2024, Alibaba Group Holding Limited. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "classfile/javaClasses.hpp"
27 #include "code/aotCodeCache.hpp"
28 #include "compiler/compileLog.hpp"
29 #include "gc/shared/barrierSet.hpp"
30 #include "gc/shared/c2/barrierSetC2.hpp"
31 #include "gc/shared/tlab_globals.hpp"
32 #include "memory/allocation.inline.hpp"
33 #include "memory/resourceArea.hpp"
34 #include "oops/objArrayKlass.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/arraycopynode.hpp"
37 #include "opto/cfgnode.hpp"
38 #include "opto/compile.hpp"
39 #include "opto/connode.hpp"
40 #include "opto/convertnode.hpp"
41 #include "opto/loopnode.hpp"
42 #include "opto/machnode.hpp"
43 #include "opto/matcher.hpp"
44 #include "opto/memnode.hpp"
45 #include "opto/mempointer.hpp"
46 #include "opto/mulnode.hpp"
47 #include "opto/narrowptrnode.hpp"
48 #include "opto/opcodes.hpp"
49 #include "opto/phaseX.hpp"
50 #include "opto/regalloc.hpp"
51 #include "opto/regmask.hpp"
52 #include "opto/rootnode.hpp"
53 #include "opto/traceMergeStoresTag.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(_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 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 ->with_offset(t_oop->is_aryptr()->offset())
239 ->is_aryptr();
240 }
241 do_split = mem_t == t_oop;
242 }
243 if (do_split) {
244 // clone the Phi with our address type
245 result = mphi->split_out_instance(t_adr, igvn);
246 } else {
247 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
248 }
249 }
250 return result;
251 }
252
253 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
254 uint alias_idx = phase->C->get_alias_index(tp);
255 Node *mem = mmem;
256 #ifdef ASSERT
257 {
258 // Check that current type is consistent with the alias index used during graph construction
259 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
260 bool consistent = adr_check == nullptr || adr_check->empty() ||
261 phase->C->must_alias(adr_check, alias_idx );
262 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
263 if( !consistent && adr_check != nullptr && !adr_check->empty() &&
264 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
265 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
266 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
267 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
268 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
269 // don't assert if it is dead code.
270 consistent = true;
271 }
272 if( !consistent ) {
273 st->print("alias_idx==%d, adr_check==", alias_idx);
274 if( adr_check == nullptr ) {
275 st->print("null");
276 } else {
277 adr_check->dump();
278 }
279 st->cr();
280 print_alias_types();
281 assert(consistent, "adr_check must match alias idx");
282 }
283 }
284 #endif
285 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
286 // means an array I have not precisely typed yet. Do not do any
287 // alias stuff with it any time soon.
288 const TypeOopPtr *toop = tp->isa_oopptr();
289 if (tp->base() != Type::AnyPtr &&
290 !(toop &&
291 toop->isa_instptr() &&
292 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
293 toop->offset() == Type::OffsetBot)) {
294 // IGVN _delay_transform may be set to true and if that is the case and mmem
295 // is already a registered node then the validation inside transform will
296 // complain.
297 Node* m = mmem;
298 PhaseIterGVN* igvn = phase->is_IterGVN();
299 if (igvn == nullptr || !igvn->delay_transform()) {
300 // compress paths and change unreachable cycles to TOP
301 // If not, we can update the input infinitely along a MergeMem cycle
302 // Equivalent code in PhiNode::Ideal
303 m = phase->transform(mmem);
304 }
305 // If transformed to a MergeMem, get the desired slice
306 // Otherwise the returned node represents memory for every slice
307 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
308 // Update input if it is progress over what we have now
309 }
310 return mem;
311 }
312
313 //--------------------------Ideal_common---------------------------------------
314 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
315 // Unhook non-raw memories from complete (macro-expanded) initializations.
316 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
317 // If our control input is a dead region, kill all below the region
318 Node *ctl = in(MemNode::Control);
319 if (ctl && remove_dead_region(phase, can_reshape))
320 return this;
321 ctl = in(MemNode::Control);
322 // Don't bother trying to transform a dead node
323 if (ctl && ctl->is_top()) return NodeSentinel;
324
325 PhaseIterGVN *igvn = phase->is_IterGVN();
326 // Wait if control on the worklist.
327 if (ctl && can_reshape && igvn != nullptr) {
328 Node* bol = nullptr;
329 Node* cmp = nullptr;
330 if (ctl->in(0)->is_If()) {
331 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
332 bol = ctl->in(0)->in(1);
333 if (bol->is_Bool())
334 cmp = ctl->in(0)->in(1)->in(1);
335 }
336 if (igvn->_worklist.member(ctl) ||
337 (bol != nullptr && igvn->_worklist.member(bol)) ||
338 (cmp != nullptr && igvn->_worklist.member(cmp)) ) {
339 // This control path may be dead.
340 // Delay this memory node transformation until the control is processed.
341 igvn->_worklist.push(this);
342 return NodeSentinel; // caller will return null
343 }
344 }
345 // Ignore if memory is dead, or self-loop
346 Node *mem = in(MemNode::Memory);
347 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return null
348 assert(mem != this, "dead loop in MemNode::Ideal");
349
350 if (can_reshape && igvn != nullptr && igvn->_worklist.member(mem)) {
351 // This memory slice may be dead.
352 // Delay this mem node transformation until the memory is processed.
353 igvn->_worklist.push(this);
354 return NodeSentinel; // caller will return null
355 }
356
357 Node *address = in(MemNode::Address);
358 const Type *t_adr = phase->type(address);
359 if (t_adr == Type::TOP) return NodeSentinel; // caller will return null
360
361 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
362 // Unsafe off-heap access with zero address. Remove access and other control users
363 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
364 assert(ctl != nullptr, "unsafe accesses should be control dependent");
365 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
366 Node* u = ctl->fast_out(i);
367 if (u != ctl) {
368 igvn->rehash_node_delayed(u);
369 int nb = u->replace_edge(ctl, phase->C->top(), igvn);
370 --i, imax -= nb;
371 }
372 }
373 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
374 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
375 phase->C->root()->add_req(halt);
376 return this;
377 }
378
379 if (can_reshape && igvn != nullptr &&
380 (igvn->_worklist.member(address) ||
381 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
382 // The address's base and type may change when the address is processed.
383 // Delay this mem node transformation until the address is processed.
384 igvn->_worklist.push(this);
385 return NodeSentinel; // caller will return null
386 }
387
388 // Do NOT remove or optimize the next lines: ensure a new alias index
389 // is allocated for an oop pointer type before Escape Analysis.
390 // Note: C++ will not remove it since the call has side effect.
391 if (t_adr->isa_oopptr()) {
392 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
393 }
394
395 Node* base = nullptr;
396 if (address->is_AddP()) {
397 base = address->in(AddPNode::Base);
398 }
399 if (base != nullptr && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
400 !t_adr->isa_rawptr()) {
401 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
402 // Skip this node optimization if its address has TOP base.
403 return NodeSentinel; // caller will return null
404 }
405
406 // Avoid independent memory operations
407 Node* old_mem = mem;
408
409 // The code which unhooks non-raw memories from complete (macro-expanded)
410 // initializations was removed. After macro-expansion all stores caught
411 // by Initialize node became raw stores and there is no information
412 // which memory slices they modify. So it is unsafe to move any memory
413 // operation above these stores. Also in most cases hooked non-raw memories
414 // were already unhooked by using information from detect_ptr_independence()
415 // and find_previous_store().
416
417 if (mem->is_MergeMem()) {
418 MergeMemNode* mmem = mem->as_MergeMem();
419 const TypePtr *tp = t_adr->is_ptr();
420
421 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
422 }
423
424 if (mem != old_mem) {
425 set_req_X(MemNode::Memory, mem, phase);
426 if (phase->type(mem) == Type::TOP) return NodeSentinel;
427 return this;
428 }
429
430 // let the subclass continue analyzing...
431 return nullptr;
432 }
433
434 // Helper function for proving some simple control dominations.
435 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
436 // Already assumes that 'dom' is available at 'sub', and that 'sub'
437 // is not a constant (dominated by the method's StartNode).
438 // Used by MemNode::find_previous_store to prove that the
439 // control input of a memory operation predates (dominates)
440 // an allocation it wants to look past.
441 // Returns 'DomResult::Dominate' if all control inputs of 'dom'
442 // dominate 'sub', 'DomResult::NotDominate' if not,
443 // and 'DomResult::EncounteredDeadCode' if we can't decide due to
444 // dead code, but at the end of IGVN, we know the definite result
445 // once the dead code is cleaned up.
446 Node::DomResult MemNode::maybe_all_controls_dominate(Node* dom, Node* sub) {
447 if (dom == nullptr || dom->is_top() || sub == nullptr || sub->is_top()) {
448 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
449 }
450
451 // Check 'dom'. Skip Proj and CatchProj nodes.
452 dom = dom->find_exact_control(dom);
453 if (dom == nullptr || dom->is_top()) {
454 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
455 }
456
457 if (dom == sub) {
458 // For the case when, for example, 'sub' is Initialize and the original
459 // 'dom' is Proj node of the 'sub'.
460 return DomResult::NotDominate;
461 }
462
463 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) {
464 return DomResult::Dominate;
465 }
466
467 // 'dom' dominates 'sub' if its control edge and control edges
468 // of all its inputs dominate or equal to sub's control edge.
469
470 // Currently 'sub' is either Allocate, Initialize or Start nodes.
471 // Or Region for the check in LoadNode::Ideal();
472 // 'sub' should have sub->in(0) != nullptr.
473 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
474 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
475
476 // Get control edge of 'sub'.
477 Node* orig_sub = sub;
478 sub = sub->find_exact_control(sub->in(0));
479 if (sub == nullptr || sub->is_top()) {
480 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
481 }
482
483 assert(sub->is_CFG(), "expecting control");
484
485 if (sub == dom) {
486 return DomResult::Dominate;
487 }
488
489 if (sub->is_Start() || sub->is_Root()) {
490 return DomResult::NotDominate;
491 }
492
493 {
494 // Check all control edges of 'dom'.
495
496 ResourceMark rm;
497 Node_List nlist;
498 Unique_Node_List dom_list;
499
500 dom_list.push(dom);
501 bool only_dominating_controls = false;
502
503 for (uint next = 0; next < dom_list.size(); next++) {
504 Node* n = dom_list.at(next);
505 if (n == orig_sub) {
506 return DomResult::NotDominate; // One of dom's inputs dominated by sub.
507 }
508 if (!n->is_CFG() && n->pinned()) {
509 // Check only own control edge for pinned non-control nodes.
510 n = n->find_exact_control(n->in(0));
511 if (n == nullptr || n->is_top()) {
512 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
513 }
514 assert(n->is_CFG(), "expecting control");
515 dom_list.push(n);
516 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
517 only_dominating_controls = true;
518 } else if (n->is_CFG()) {
519 DomResult dom_result = n->dominates(sub, nlist);
520 if (dom_result == DomResult::Dominate) {
521 only_dominating_controls = true;
522 } else {
523 return dom_result;
524 }
525 } else {
526 // First, own control edge.
527 Node* m = n->find_exact_control(n->in(0));
528 if (m != nullptr) {
529 if (m->is_top()) {
530 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
531 }
532 dom_list.push(m);
533 }
534 // Now, the rest of edges.
535 uint cnt = n->req();
536 for (uint i = 1; i < cnt; i++) {
537 m = n->find_exact_control(n->in(i));
538 if (m == nullptr || m->is_top()) {
539 continue;
540 }
541 dom_list.push(m);
542 }
543 }
544 }
545 return only_dominating_controls ? DomResult::Dominate : DomResult::NotDominate;
546 }
547 }
548
549 //---------------------detect_ptr_independence---------------------------------
550 // Used by MemNode::find_previous_store to prove that two base
551 // pointers are never equal.
552 // The pointers are accompanied by their associated allocations,
553 // if any, which have been previously discovered by the caller.
554 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
555 Node* p2, AllocateNode* a2,
556 PhaseTransform* phase) {
557 // Trivial case: Non-overlapping values. Be careful, we can cast a raw pointer to an oop (e.g. in
558 // the allocation pattern) so joining the types only works if both are oops. join may also give
559 // an incorrect result when both pointers are nullable and the result is supposed to be
560 // TypePtr::NULL_PTR, so we exclude that case.
561 const Type* p1_type = p1->bottom_type();
562 const Type* p2_type = p2->bottom_type();
563 if (p1_type->isa_oopptr() && p2_type->isa_oopptr() &&
564 (!p1_type->maybe_null() || !p2_type->maybe_null()) &&
565 p1_type->join(p2_type)->empty()) {
566 return true;
567 }
568
569 // Attempt to prove that these two pointers cannot be aliased.
570 // They may both manifestly be allocations, and they should differ.
571 // Or, if they are not both allocations, they can be distinct constants.
572 // Otherwise, one is an allocation and the other a pre-existing value.
573 if (a1 == nullptr && a2 == nullptr) { // neither an allocation
574 return (p1 != p2) && p1->is_Con() && p2->is_Con();
575 } else if (a1 != nullptr && a2 != nullptr) { // both allocations
576 return (a1 != a2);
577 } else if (a1 != nullptr) { // one allocation a1
578 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
579 return all_controls_dominate(p2, a1);
580 } else { //(a2 != null) // one allocation a2
581 return all_controls_dominate(p1, a2);
582 }
583 return false;
584 }
585
586
587 // Find an arraycopy ac that produces the memory state represented by parameter mem.
588 // Return ac if
589 // (a) can_see_stored_value=true and ac must have set the value for this load or if
590 // (b) can_see_stored_value=false and ac could have set the value for this load or if
591 // (c) can_see_stored_value=false and ac cannot have set the value for this load.
592 // In case (c) change the parameter mem to the memory input of ac to skip it
593 // when searching stored value.
594 // Otherwise return null.
595 Node* LoadNode::find_previous_arraycopy(PhaseValues* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
596 ArrayCopyNode* ac = find_array_copy_clone(ld_alloc, mem);
597 if (ac != nullptr) {
598 Node* ld_addp = in(MemNode::Address);
599 Node* src = ac->in(ArrayCopyNode::Src);
600 const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
601
602 // This is a load from a cloned array. The corresponding arraycopy ac must
603 // have set the value for the load and we can return ac but only if the load
604 // is known to be within bounds. This is checked below.
605 if (ary_t != nullptr && ld_addp->is_AddP()) {
606 Node* ld_offs = ld_addp->in(AddPNode::Offset);
607 BasicType ary_elem = ary_t->elem()->array_element_basic_type();
608 jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
609 jlong elemsize = type2aelembytes(ary_elem);
610
611 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
612 const TypeInt* sizetype = ary_t->size();
613
614 if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
615 // The load is known to be within bounds. It receives its value from ac.
616 return ac;
617 }
618 // The load is known to be out-of-bounds.
619 }
620 // The load could be out-of-bounds. It must not be hoisted but must remain
621 // dependent on the runtime range check. This is achieved by returning null.
622 } else if (mem->is_Proj() && mem->in(0) != nullptr && mem->in(0)->is_ArrayCopy()) {
623 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
624
625 if (ac->is_arraycopy_validated() ||
626 ac->is_copyof_validated() ||
627 ac->is_copyofrange_validated()) {
628 Node* ld_addp = in(MemNode::Address);
629 if (ld_addp->is_AddP()) {
630 Node* ld_base = ld_addp->in(AddPNode::Address);
631 Node* ld_offs = ld_addp->in(AddPNode::Offset);
632
633 Node* dest = ac->in(ArrayCopyNode::Dest);
634
635 if (dest == ld_base) {
636 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
637 assert(!ld_offs_t->empty(), "dead reference should be checked already");
638 // Take into account vector or unsafe access size
639 jlong ld_size_in_bytes = (jlong)memory_size();
640 jlong offset_hi = ld_offs_t->_hi + ld_size_in_bytes - 1;
641 offset_hi = MIN2(offset_hi, (jlong)(TypeX::MAX->_hi)); // Take care for overflow in 32-bit VM
642 if (ac->modifies(ld_offs_t->_lo, (intptr_t)offset_hi, phase, can_see_stored_value)) {
643 return ac;
644 }
645 if (!can_see_stored_value) {
646 mem = ac->in(TypeFunc::Memory);
647 return ac;
648 }
649 }
650 }
651 }
652 }
653 return nullptr;
654 }
655
656 ArrayCopyNode* MemNode::find_array_copy_clone(Node* ld_alloc, Node* mem) const {
657 if (mem->is_Proj() && mem->in(0) != nullptr && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
658 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
659 if (ld_alloc != nullptr) {
660 // Check if there is an array copy for a clone
661 Node* mb = mem->in(0);
662 ArrayCopyNode* ac = nullptr;
663 if (mb->in(0) != nullptr && mb->in(0)->is_Proj() &&
664 mb->in(0)->in(0) != nullptr && mb->in(0)->in(0)->is_ArrayCopy()) {
665 ac = mb->in(0)->in(0)->as_ArrayCopy();
666 } else {
667 // Step over GC barrier when ReduceInitialCardMarks is disabled
668 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
669 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
670
671 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
672 ac = control_proj_ac->in(0)->as_ArrayCopy();
673 }
674 }
675
676 if (ac != nullptr && ac->is_clonebasic()) {
677 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest));
678 if (alloc != nullptr && alloc == ld_alloc) {
679 return ac;
680 }
681 }
682 }
683 }
684 return nullptr;
685 }
686
687 // The logic for reordering loads and stores uses four steps:
688 // (a) Walk carefully past stores and initializations which we
689 // can prove are independent of this load.
690 // (b) Observe that the next memory state makes an exact match
691 // with self (load or store), and locate the relevant store.
692 // (c) Ensure that, if we were to wire self directly to the store,
693 // the optimizer would fold it up somehow.
694 // (d) Do the rewiring, and return, depending on some other part of
695 // the optimizer to fold up the load.
696 // This routine handles steps (a) and (b). Steps (c) and (d) are
697 // specific to loads and stores, so they are handled by the callers.
698 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
699 //
700 Node* MemNode::find_previous_store(PhaseValues* phase) {
701 Node* ctrl = in(MemNode::Control);
702 Node* adr = in(MemNode::Address);
703 intptr_t offset = 0;
704 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
705 AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
706
707 const TypePtr* adr_type = this->adr_type();
708 if (adr_type == nullptr) {
709 // This means the access is dead
710 return phase->C->top();
711 } else if (adr_type->base() == TypePtr::AnyPtr) {
712 assert(adr_type->ptr() == TypePtr::Null, "MemNode should never access a wide memory");
713 // Give up, this will upset Compile::get_alias_index
714 return nullptr;
715 }
716
717 int alias_idx = phase->C->get_alias_index(adr_type);
718 assert(alias_idx != Compile::AliasIdxTop, "must not be a dead node");
719 assert(alias_idx != Compile::AliasIdxBot || !phase->C->do_aliasing(), "must not be a very wide access");
720
721 if (offset == Type::OffsetBot)
722 return nullptr; // cannot unalias unless there are precise offsets
723
724 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
725 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
726
727 intptr_t size_in_bytes = memory_size();
728
729 Node* mem = in(MemNode::Memory); // start searching here...
730
731 int cnt = 50; // Cycle limiter
732 for (;;) { // While we can dance past unrelated stores...
733 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
734
735 Node* prev = mem;
736 if (mem->is_Store()) {
737 Node* st_adr = mem->in(MemNode::Address);
738 intptr_t st_offset = 0;
739 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
740 if (st_base == nullptr) {
741 // inscrutable pointer
742 break;
743 }
744
745 // If the bases are the same and the offsets are the same, it seems that this is the exact
746 // store we are looking for, the caller will check if the type of the store matches using
747 // MemNode::can_see_stored_value
748 if (st_base == base && st_offset == offset) {
749 return mem; // (b) found the store that this access observes
750 }
751
752 // If it is provable that the memory accessed by mem does not overlap the memory accessed by
753 // this, we may walk past mem.
754 // For raw accesses, 2 accesses are independent if they have the same base and the offsets
755 // say that they do not overlap.
756 // For heap accesses, 2 accesses are independent if either the bases are provably different
757 // at runtime or the offsets say that the accesses do not overlap.
758 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
759 // Raw accesses can only be provably independent if they have the same base
760 break;
761 }
762
763 // If the offsets say that the accesses do not overlap, then it is provable that mem and this
764 // do not overlap. For example, a LoadI from Object+8 is independent from a StoreL into
765 // Object+12, no matter what the bases are.
766 if (st_offset != offset && st_offset != Type::OffsetBot) {
767 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
768 assert(mem->as_Store()->memory_size() <= MAX_STORE, "");
769 if (st_offset >= offset + size_in_bytes ||
770 st_offset <= offset - MAX_STORE ||
771 st_offset <= offset - mem->as_Store()->memory_size()) {
772 // Success: The offsets are provably independent.
773 // (You may ask, why not just test st_offset != offset and be done?
774 // The answer is that stores of different sizes can co-exist
775 // in the same sequence of RawMem effects. We sometimes initialize
776 // a whole 'tile' of array elements with a single jint or jlong.)
777 mem = mem->in(MemNode::Memory);
778 continue; // (a) advance through the independent store
779 }
780 }
781
782 // Same base and overlapping offsets, it seems provable that the accesses overlap, give up
783 if (st_base == base) {
784 break;
785 }
786
787 // Try to prove that 2 different base nodes at compile time are different values at runtime
788 bool known_independent = false;
789 if (detect_ptr_independence(base, alloc, st_base, AllocateNode::Ideal_allocation(st_base), phase)) {
790 known_independent = true;
791 }
792
793 if (known_independent) {
794 mem = mem->in(MemNode::Memory);
795 continue; // (a) advance through the independent store
796 }
797 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
798 InitializeNode* st_init = mem->in(0)->as_Initialize();
799 AllocateNode* st_alloc = st_init->allocation();
800 if (st_alloc == nullptr) {
801 break; // something degenerated
802 }
803 bool known_identical = false;
804 bool known_independent = false;
805 if (alloc == st_alloc) {
806 known_identical = true;
807 } else if (alloc != nullptr) {
808 known_independent = true;
809 } else if (all_controls_dominate(this, st_alloc)) {
810 known_independent = true;
811 }
812
813 if (known_independent) {
814 // The bases are provably independent: Either they are
815 // manifestly distinct allocations, or else the control
816 // of this load dominates the store's allocation.
817 if (alias_idx == Compile::AliasIdxRaw) {
818 mem = st_alloc->in(TypeFunc::Memory);
819 } else {
820 mem = st_init->memory(alias_idx);
821 }
822 continue; // (a) advance through independent store memory
823 }
824
825 // (b) at this point, if we are not looking at a store initializing
826 // the same allocation we are loading from, we lose.
827 if (known_identical) {
828 // From caller, can_see_stored_value will consult find_captured_store.
829 return mem; // let caller handle steps (c), (d)
830 }
831
832 } else if (find_previous_arraycopy(phase, alloc, mem, false) != nullptr) {
833 if (prev != mem) {
834 // Found an arraycopy but it doesn't affect that load
835 continue;
836 }
837 // Found an arraycopy that may affect that load
838 return mem;
839 } else if (mem->is_MergeMem()) {
840 mem = mem->as_MergeMem()->memory_at(alias_idx);
841 continue;
842 } else if (addr_t != nullptr && addr_t->is_known_instance_field()) {
843 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
844 if (mem->is_Proj() && mem->in(0)->is_Call()) {
845 // ArrayCopyNodes processed here as well.
846 CallNode *call = mem->in(0)->as_Call();
847 if (!call->may_modify(addr_t, phase)) {
848 mem = call->in(TypeFunc::Memory);
849 continue; // (a) advance through independent call memory
850 }
851 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
852 ArrayCopyNode* ac = nullptr;
853 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
854 break;
855 }
856 mem = mem->in(0)->in(TypeFunc::Memory);
857 continue; // (a) advance through independent MemBar memory
858 } else if (mem->is_ClearArray()) {
859 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
860 // (the call updated 'mem' value)
861 continue; // (a) advance through independent allocation memory
862 } else {
863 // Can not bypass initialization of the instance
864 // we are looking for.
865 return mem;
866 }
867 }
868 }
869
870 // Unless there is an explicit 'continue', we must bail out here,
871 // because 'mem' is an inscrutable memory state (e.g., a call).
872 break;
873 }
874
875 return nullptr; // bail out
876 }
877
878 //----------------------calculate_adr_type-------------------------------------
879 // Helper function. Notices when the given type of address hits top or bottom.
880 // Also, asserts a cross-check of the type against the expected address type.
881 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
882 if (t == Type::TOP) return nullptr; // does not touch memory any more?
883 #ifdef ASSERT
884 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = nullptr;
885 #endif
886 const TypePtr* tp = t->isa_ptr();
887 if (tp == nullptr) {
888 assert(cross_check == nullptr || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
889 return TypePtr::BOTTOM; // touches lots of memory
890 } else {
891 #ifdef ASSERT
892 // %%%% [phh] We don't check the alias index if cross_check is
893 // TypeRawPtr::BOTTOM. Needs to be investigated.
894 if (cross_check != nullptr &&
895 cross_check != TypePtr::BOTTOM &&
896 cross_check != TypeRawPtr::BOTTOM) {
897 // Recheck the alias index, to see if it has changed (due to a bug).
898 Compile* C = Compile::current();
899 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
900 "must stay in the original alias category");
901 // The type of the address must be contained in the adr_type,
902 // disregarding "null"-ness.
903 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
904 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
905 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
906 "real address must not escape from expected memory type");
907 }
908 #endif
909 return tp;
910 }
911 }
912
913 uint8_t MemNode::barrier_data(const Node* n) {
914 if (n->is_LoadStore()) {
915 return n->as_LoadStore()->barrier_data();
916 } else if (n->is_Mem()) {
917 return n->as_Mem()->barrier_data();
918 }
919 return 0;
920 }
921
922 //=============================================================================
923 // Should LoadNode::Ideal() attempt to remove control edges?
924 bool LoadNode::can_remove_control() const {
925 return !has_pinned_control_dependency();
926 }
927 uint LoadNode::size_of() const { return sizeof(*this); }
928 bool LoadNode::cmp(const Node &n) const {
929 LoadNode& load = (LoadNode &)n;
930 return Type::equals(_type, load._type) &&
931 _control_dependency == load._control_dependency &&
932 _mo == load._mo;
933 }
934 const Type *LoadNode::bottom_type() const { return _type; }
935 uint LoadNode::ideal_reg() const {
936 return _type->ideal_reg();
937 }
938
939 #ifndef PRODUCT
940 void LoadNode::dump_spec(outputStream *st) const {
941 MemNode::dump_spec(st);
942 if( !Verbose && !WizardMode ) {
943 // standard dump does this in Verbose and WizardMode
944 st->print(" #"); _type->dump_on(st);
945 }
946 if (in(0) != nullptr && !depends_only_on_test()) {
947 st->print(" (does not depend only on test, ");
948 if (control_dependency() == UnknownControl) {
949 st->print("unknown control");
950 } else if (control_dependency() == Pinned) {
951 st->print("pinned");
952 } else if (adr_type() == TypeRawPtr::BOTTOM) {
953 st->print("raw access");
954 } else {
955 st->print("unknown reason");
956 }
957 st->print(")");
958 }
959 }
960 #endif
961
962 #ifdef ASSERT
963 //----------------------------is_immutable_value-------------------------------
964 // Helper function to allow a raw load without control edge for some cases
965 bool LoadNode::is_immutable_value(Node* adr) {
966 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
967 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) {
968
969 jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1);
970 int offsets[] = {
971 in_bytes(JavaThread::osthread_offset()),
972 in_bytes(JavaThread::threadObj_offset()),
973 in_bytes(JavaThread::vthread_offset()),
974 in_bytes(JavaThread::scopedValueCache_offset()),
975 };
976
977 for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) {
978 if (offset == offsets[i]) {
979 return true;
980 }
981 }
982 }
983
984 return false;
985 }
986 #endif
987
988 //----------------------------LoadNode::make-----------------------------------
989 // Polymorphic factory method:
990 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo,
991 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
992 Compile* C = gvn.C;
993
994 // sanity check the alias category against the created node type
995 assert(!(adr_type->isa_oopptr() &&
996 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
997 "use LoadKlassNode instead");
998 assert(!(adr_type->isa_aryptr() &&
999 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
1000 "use LoadRangeNode instead");
1001 // Check control edge of raw loads
1002 assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
1003 // oop will be recorded in oop map if load crosses safepoint
1004 rt->isa_oopptr() || is_immutable_value(adr),
1005 "raw memory operations should have control edge");
1006 LoadNode* load = nullptr;
1007 switch (bt) {
1008 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1009 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1010 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1011 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1012 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1013 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break;
1014 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
1015 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break;
1016 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
1017 case T_OBJECT:
1018 case T_NARROWOOP:
1019 #ifdef _LP64
1020 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1021 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
1022 } else
1023 #endif
1024 {
1025 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
1026 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
1027 }
1028 break;
1029 default:
1030 ShouldNotReachHere();
1031 break;
1032 }
1033 assert(load != nullptr, "LoadNode should have been created");
1034 if (unaligned) {
1035 load->set_unaligned_access();
1036 }
1037 if (mismatched) {
1038 load->set_mismatched_access();
1039 }
1040 if (unsafe) {
1041 load->set_unsafe_access();
1042 }
1043 load->set_barrier_data(barrier_data);
1044 if (load->Opcode() == Op_LoadN) {
1045 Node* ld = gvn.transform(load);
1046 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
1047 }
1048
1049 return load;
1050 }
1051
1052 //------------------------------hash-------------------------------------------
1053 uint LoadNode::hash() const {
1054 // unroll addition of interesting fields
1055 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
1056 }
1057
1058 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
1059 if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) {
1060 bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile();
1061 bool is_stable_ary = FoldStableValues &&
1062 (tp != nullptr) && (tp->isa_aryptr() != nullptr) &&
1063 tp->isa_aryptr()->is_stable();
1064
1065 return (eliminate_boxing && non_volatile) || is_stable_ary;
1066 }
1067
1068 return false;
1069 }
1070
1071 // Is the value loaded previously stored by an arraycopy? If so return
1072 // a load node that reads from the source array so we may be able to
1073 // optimize out the ArrayCopy node later.
1074 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
1075 Node* ld_adr = in(MemNode::Address);
1076 intptr_t ld_off = 0;
1077 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
1078 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
1079 if (ac != nullptr) {
1080 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
1081
1082 Node* mem = ac->in(TypeFunc::Memory);
1083 Node* ctl = ac->in(0);
1084 Node* src = ac->in(ArrayCopyNode::Src);
1085
1086 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
1087 return nullptr;
1088 }
1089
1090 // load depends on the tests that validate the arraycopy
1091 LoadNode* ld = clone_pinned();
1092 Node* addp = in(MemNode::Address)->clone();
1093 if (ac->as_ArrayCopy()->is_clonebasic()) {
1094 assert(ld_alloc != nullptr, "need an alloc");
1095 assert(addp->is_AddP(), "address must be addp");
1096 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1097 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1098 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1099 addp->set_req(AddPNode::Base, src);
1100 addp->set_req(AddPNode::Address, src);
1101 } else {
1102 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1103 ac->as_ArrayCopy()->is_copyof_validated() ||
1104 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1105 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1106 addp->set_req(AddPNode::Base, src);
1107 addp->set_req(AddPNode::Address, src);
1108
1109 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1110 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
1111 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
1112
1113 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
1114 uint shift = exact_log2(type2aelembytes(ary_elem));
1115
1116 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1117 #ifdef _LP64
1118 diff = phase->transform(new ConvI2LNode(diff));
1119 #endif
1120 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1121
1122 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1123 addp->set_req(AddPNode::Offset, offset);
1124 }
1125 addp = phase->transform(addp);
1126 #ifdef ASSERT
1127 const TypePtr* adr_type = phase->type(addp)->is_ptr();
1128 ld->_adr_type = adr_type;
1129 #endif
1130 ld->set_req(MemNode::Address, addp);
1131 ld->set_req(0, ctl);
1132 ld->set_req(MemNode::Memory, mem);
1133 return ld;
1134 }
1135 return nullptr;
1136 }
1137
1138
1139 //---------------------------can_see_stored_value------------------------------
1140 // This routine exists to make sure this set of tests is done the same
1141 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
1142 // will change the graph shape in a way which makes memory alive twice at the
1143 // same time (uses the Oracle model of aliasing), then some
1144 // LoadXNode::Identity will fold things back to the equivalence-class model
1145 // of aliasing.
1146 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const {
1147 Node* ld_adr = in(MemNode::Address);
1148 intptr_t ld_off = 0;
1149 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1150 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base);
1151 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1152 Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr;
1153 // This is more general than load from boxing objects.
1154 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1155 uint alias_idx = atp->index();
1156 Node* result = nullptr;
1157 Node* current = st;
1158 // Skip through chains of MemBarNodes checking the MergeMems for
1159 // new states for the slice of this load. Stop once any other
1160 // kind of node is encountered. Loads from final memory can skip
1161 // through any kind of MemBar but normal loads shouldn't skip
1162 // through MemBarAcquire since the could allow them to move out of
1163 // a synchronized region. It is not safe to step over MemBarCPUOrder,
1164 // because alias info above them may be inaccurate (e.g., due to
1165 // mixed/mismatched unsafe accesses).
1166 bool is_final_mem = !atp->is_rewritable();
1167 while (current->is_Proj()) {
1168 int opc = current->in(0)->Opcode();
1169 if ((is_final_mem && (opc == Op_MemBarAcquire ||
1170 opc == Op_MemBarAcquireLock ||
1171 opc == Op_LoadFence)) ||
1172 opc == Op_MemBarRelease ||
1173 opc == Op_StoreFence ||
1174 opc == Op_MemBarReleaseLock ||
1175 opc == Op_MemBarStoreStore ||
1176 opc == Op_StoreStoreFence) {
1177 Node* mem = current->in(0)->in(TypeFunc::Memory);
1178 if (mem->is_MergeMem()) {
1179 MergeMemNode* merge = mem->as_MergeMem();
1180 Node* new_st = merge->memory_at(alias_idx);
1181 if (new_st == merge->base_memory()) {
1182 // Keep searching
1183 current = new_st;
1184 continue;
1185 }
1186 // Save the new memory state for the slice and fall through
1187 // to exit.
1188 result = new_st;
1189 }
1190 }
1191 break;
1192 }
1193 if (result != nullptr) {
1194 st = result;
1195 }
1196 }
1197
1198 // Loop around twice in the case Load -> Initialize -> Store.
1199 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1200 for (int trip = 0; trip <= 1; trip++) {
1201
1202 if (st->is_Store()) {
1203 Node* st_adr = st->in(MemNode::Address);
1204 if (st_adr != ld_adr) {
1205 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1206 intptr_t st_off = 0;
1207 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1208 if (ld_base == nullptr) return nullptr;
1209 if (st_base == nullptr) return nullptr;
1210 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr;
1211 if (ld_off != st_off) return nullptr;
1212 if (ld_off == Type::OffsetBot) return nullptr;
1213 // Same base, same offset.
1214 // Possible improvement for arrays: check index value instead of absolute offset.
1215
1216 // At this point we have proven something like this setup:
1217 // B = << base >>
1218 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1219 // S = StoreQ(AddP( B , #Off), V)
1220 // (Actually, we haven't yet proven the Q's are the same.)
1221 // In other words, we are loading from a casted version of
1222 // the same pointer-and-offset that we stored to.
1223 // Casted version may carry a dependency and it is respected.
1224 // Thus, we are able to replace L by V.
1225 }
1226 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1227 if (store_Opcode() != st->Opcode()) {
1228 return nullptr;
1229 }
1230 // LoadVector/StoreVector needs additional check to ensure the types match.
1231 if (st->is_StoreVector()) {
1232 if ((Opcode() != Op_LoadVector && Opcode() != Op_StoreVector) || st->Opcode() != Op_StoreVector) {
1233 // Some kind of masked access or gather/scatter
1234 return nullptr;
1235 }
1236
1237 const TypeVect* in_vt = st->as_StoreVector()->vect_type();
1238 const TypeVect* out_vt = is_Load() ? as_LoadVector()->vect_type() : as_StoreVector()->vect_type();
1239 if (in_vt != out_vt) {
1240 return nullptr;
1241 }
1242 }
1243 return st->in(MemNode::ValueIn);
1244 }
1245
1246 // A load from a freshly-created object always returns zero.
1247 // (This can happen after LoadNode::Ideal resets the load's memory input
1248 // to find_captured_store, which returned InitializeNode::zero_memory.)
1249 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1250 (st->in(0) == ld_alloc) &&
1251 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1252 // return a zero value for the load's basic type
1253 // (This is one of the few places where a generic PhaseTransform
1254 // can create new nodes. Think of it as lazily manifesting
1255 // virtually pre-existing constants.)
1256 if (value_basic_type() != T_VOID) {
1257 if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) {
1258 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1259 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1260 // by the ArrayCopyNode.
1261 return phase->zerocon(value_basic_type());
1262 }
1263 } else {
1264 // TODO: materialize all-zero vector constant
1265 assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1266 }
1267 }
1268
1269 // A load from an initialization barrier can match a captured store.
1270 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1271 InitializeNode* init = st->in(0)->as_Initialize();
1272 AllocateNode* alloc = init->allocation();
1273 if ((alloc != nullptr) && (alloc == ld_alloc)) {
1274 // examine a captured store value
1275 st = init->find_captured_store(ld_off, memory_size(), phase);
1276 if (st != nullptr) {
1277 continue; // take one more trip around
1278 }
1279 }
1280 }
1281
1282 // Load boxed value from result of valueOf() call is input parameter.
1283 if (this->is_Load() && ld_adr->is_AddP() &&
1284 (tp != nullptr) && tp->is_ptr_to_boxed_value()) {
1285 intptr_t ignore = 0;
1286 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1287 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1288 base = bs->step_over_gc_barrier(base);
1289 if (base != nullptr && base->is_Proj() &&
1290 base->as_Proj()->_con == TypeFunc::Parms &&
1291 base->in(0)->is_CallStaticJava() &&
1292 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1293 return base->in(0)->in(TypeFunc::Parms);
1294 }
1295 }
1296
1297 break;
1298 }
1299
1300 return nullptr;
1301 }
1302
1303 //----------------------is_instance_field_load_with_local_phi------------------
1304 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1305 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1306 in(Address)->is_AddP() ) {
1307 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1308 // Only instances and boxed values.
1309 if( t_oop != nullptr &&
1310 (t_oop->is_ptr_to_boxed_value() ||
1311 t_oop->is_known_instance_field()) &&
1312 t_oop->offset() != Type::OffsetBot &&
1313 t_oop->offset() != Type::OffsetTop) {
1314 return true;
1315 }
1316 }
1317 return false;
1318 }
1319
1320 //------------------------------Identity---------------------------------------
1321 // Loads are identity if previous store is to same address
1322 Node* LoadNode::Identity(PhaseGVN* phase) {
1323 // If the previous store-maker is the right kind of Store, and the store is
1324 // to the same address, then we are equal to the value stored.
1325 Node* mem = in(Memory);
1326 Node* value = can_see_stored_value(mem, phase);
1327 if( value ) {
1328 // byte, short & char stores truncate naturally.
1329 // A load has to load the truncated value which requires
1330 // some sort of masking operation and that requires an
1331 // Ideal call instead of an Identity call.
1332 if (memory_size() < BytesPerInt) {
1333 // If the input to the store does not fit with the load's result type,
1334 // it must be truncated via an Ideal call.
1335 if (!phase->type(value)->higher_equal(phase->type(this)))
1336 return this;
1337 }
1338 // (This works even when value is a Con, but LoadNode::Value
1339 // usually runs first, producing the singleton type of the Con.)
1340 if (!has_pinned_control_dependency() || value->is_Con()) {
1341 return value;
1342 } else {
1343 return this;
1344 }
1345 }
1346
1347 if (has_pinned_control_dependency()) {
1348 return this;
1349 }
1350 // Search for an existing data phi which was generated before for the same
1351 // instance's field to avoid infinite generation of phis in a loop.
1352 Node *region = mem->in(0);
1353 if (is_instance_field_load_with_local_phi(region)) {
1354 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1355 int this_index = phase->C->get_alias_index(addr_t);
1356 int this_offset = addr_t->offset();
1357 int this_iid = addr_t->instance_id();
1358 if (!addr_t->is_known_instance() &&
1359 addr_t->is_ptr_to_boxed_value()) {
1360 // Use _idx of address base (could be Phi node) for boxed values.
1361 intptr_t ignore = 0;
1362 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1363 if (base == nullptr) {
1364 return this;
1365 }
1366 this_iid = base->_idx;
1367 }
1368 const Type* this_type = bottom_type();
1369 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1370 Node* phi = region->fast_out(i);
1371 if (phi->is_Phi() && phi != mem &&
1372 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1373 return phi;
1374 }
1375 }
1376 }
1377
1378 return this;
1379 }
1380
1381 // Construct an equivalent unsigned load.
1382 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1383 BasicType bt = T_ILLEGAL;
1384 const Type* rt = nullptr;
1385 switch (Opcode()) {
1386 case Op_LoadUB: return this;
1387 case Op_LoadUS: return this;
1388 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1389 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1390 default:
1391 assert(false, "no unsigned variant: %s", Name());
1392 return nullptr;
1393 }
1394 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1395 raw_adr_type(), rt, bt, _mo, _control_dependency,
1396 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1397 }
1398
1399 // Construct an equivalent signed load.
1400 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1401 BasicType bt = T_ILLEGAL;
1402 const Type* rt = nullptr;
1403 switch (Opcode()) {
1404 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1405 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1406 case Op_LoadB: // fall through
1407 case Op_LoadS: // fall through
1408 case Op_LoadI: // fall through
1409 case Op_LoadL: return this;
1410 default:
1411 assert(false, "no signed variant: %s", Name());
1412 return nullptr;
1413 }
1414 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1415 raw_adr_type(), rt, bt, _mo, _control_dependency,
1416 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1417 }
1418
1419 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1420 BasicType bt = rt->basic_type();
1421 switch (Opcode()) {
1422 case Op_LoadI: return (bt == T_FLOAT);
1423 case Op_LoadL: return (bt == T_DOUBLE);
1424 case Op_LoadF: return (bt == T_INT);
1425 case Op_LoadD: return (bt == T_LONG);
1426
1427 default: return false;
1428 }
1429 }
1430
1431 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1432 BasicType bt = rt->basic_type();
1433 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1434 bool is_mismatched = is_mismatched_access();
1435 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1436 if (raw_type == nullptr) {
1437 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1438 }
1439 const int op = Opcode();
1440 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) ||
1441 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access());
1442 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1443 raw_adr_type(), rt, bt, _mo, _control_dependency,
1444 require_atomic_access, is_unaligned_access(), is_mismatched);
1445 }
1446
1447 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1448 BasicType bt = vt->basic_type();
1449 switch (Opcode()) {
1450 case Op_StoreI: return (bt == T_FLOAT);
1451 case Op_StoreL: return (bt == T_DOUBLE);
1452 case Op_StoreF: return (bt == T_INT);
1453 case Op_StoreD: return (bt == T_LONG);
1454
1455 default: return false;
1456 }
1457 }
1458
1459 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1460 BasicType bt = vt->basic_type();
1461 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1462 const int op = Opcode();
1463 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) ||
1464 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access());
1465 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1466 raw_adr_type(), val, bt, _mo, require_atomic_access);
1467
1468 bool is_mismatched = is_mismatched_access();
1469 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1470 if (raw_type == nullptr) {
1471 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1472 }
1473 if (is_mismatched) {
1474 st->set_mismatched_access();
1475 }
1476 return st;
1477 }
1478
1479 // We're loading from an object which has autobox behaviour.
1480 // If this object is result of a valueOf call we'll have a phi
1481 // merging a newly allocated object and a load from the cache.
1482 // We want to replace this load with the original incoming
1483 // argument to the valueOf call.
1484 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1485 assert(igvn->C->eliminate_boxing(), "sanity");
1486 intptr_t ignore = 0;
1487 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1488 if ((base == nullptr) || base->is_Phi()) {
1489 // Push the loads from the phi that comes from valueOf up
1490 // through it to allow elimination of the loads and the recovery
1491 // of the original value. It is done in split_through_phi().
1492 return nullptr;
1493 } else if (base->is_Load() ||
1494 (base->is_DecodeN() && base->in(1)->is_Load())) {
1495 // Eliminate the load of boxed value for integer types from the cache
1496 // array by deriving the value from the index into the array.
1497 // Capture the offset of the load and then reverse the computation.
1498
1499 // Get LoadN node which loads a boxing object from 'cache' array.
1500 if (base->is_DecodeN()) {
1501 base = base->in(1);
1502 }
1503 if (!base->in(Address)->is_AddP()) {
1504 return nullptr; // Complex address
1505 }
1506 AddPNode* address = base->in(Address)->as_AddP();
1507 Node* cache_base = address->in(AddPNode::Base);
1508 if ((cache_base != nullptr) && cache_base->is_DecodeN()) {
1509 // Get ConP node which is static 'cache' field.
1510 cache_base = cache_base->in(1);
1511 }
1512 if ((cache_base != nullptr) && cache_base->is_Con()) {
1513 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1514 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
1515 Node* elements[4];
1516 int shift = exact_log2(type2aelembytes(T_OBJECT));
1517 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1518 if (count > 0 && elements[0]->is_Con() &&
1519 (count == 1 ||
1520 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1521 elements[1]->in(2) == igvn->intcon(shift)))) {
1522 ciObjArray* array = base_type->const_oop()->as_obj_array();
1523 // Fetch the box object cache[0] at the base of the array and get its value
1524 ciInstance* box = array->obj_at(0)->as_instance();
1525 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1526 assert(ik->is_box_klass(), "sanity");
1527 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1528 if (ik->nof_nonstatic_fields() == 1) {
1529 // This should be true nonstatic_field_at requires calling
1530 // nof_nonstatic_fields so check it anyway
1531 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1532 BasicType bt = c.basic_type();
1533 // Only integer types have boxing cache.
1534 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1535 bt == T_BYTE || bt == T_SHORT ||
1536 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1537 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1538 if (cache_low != (int)cache_low) {
1539 return nullptr; // should not happen since cache is array indexed by value
1540 }
1541 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1542 if (offset != (int)offset) {
1543 return nullptr; // should not happen since cache is array indexed by value
1544 }
1545 // Add up all the offsets making of the address of the load
1546 Node* result = elements[0];
1547 for (int i = 1; i < count; i++) {
1548 result = igvn->transform(new AddXNode(result, elements[i]));
1549 }
1550 // Remove the constant offset from the address and then
1551 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1552 // remove the scaling of the offset to recover the original index.
1553 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1554 // Peel the shift off directly but wrap it in a dummy node
1555 // since Ideal can't return existing nodes
1556 igvn->_worklist.push(result); // remove dead node later
1557 result = new RShiftXNode(result->in(1), igvn->intcon(0));
1558 } else if (result->is_Add() && result->in(2)->is_Con() &&
1559 result->in(1)->Opcode() == Op_LShiftX &&
1560 result->in(1)->in(2) == igvn->intcon(shift)) {
1561 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1562 // but for boxing cache access we know that X<<Z will not overflow
1563 // (there is range check) so we do this optimizatrion by hand here.
1564 igvn->_worklist.push(result); // remove dead node later
1565 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1566 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1567 } else {
1568 result = new RShiftXNode(result, igvn->intcon(shift));
1569 }
1570 #ifdef _LP64
1571 if (bt != T_LONG) {
1572 result = new ConvL2INode(igvn->transform(result));
1573 }
1574 #else
1575 if (bt == T_LONG) {
1576 result = new ConvI2LNode(igvn->transform(result));
1577 }
1578 #endif
1579 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1580 // Need to preserve unboxing load type if it is unsigned.
1581 switch(this->Opcode()) {
1582 case Op_LoadUB:
1583 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1584 break;
1585 case Op_LoadUS:
1586 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1587 break;
1588 }
1589 return result;
1590 }
1591 }
1592 }
1593 }
1594 }
1595 return nullptr;
1596 }
1597
1598 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1599 Node* region = phi->in(0);
1600 if (region == nullptr) {
1601 return false; // Wait stable graph
1602 }
1603 uint cnt = phi->req();
1604 for (uint i = 1; i < cnt; i++) {
1605 Node* rc = region->in(i);
1606 if (rc == nullptr || phase->type(rc) == Type::TOP)
1607 return false; // Wait stable graph
1608 Node* in = phi->in(i);
1609 if (in == nullptr || phase->type(in) == Type::TOP)
1610 return false; // Wait stable graph
1611 }
1612 return true;
1613 }
1614
1615 //------------------------------split_through_phi------------------------------
1616 // Check whether a call to 'split_through_phi' would split this load through the
1617 // Phi *base*. This method is essentially a copy of the validations performed
1618 // by 'split_through_phi'. The first use of this method was in EA code as part
1619 // of simplification of allocation merges.
1620 // Some differences from original method (split_through_phi):
1621 // - If base->is_CastPP(): base = base->in(1)
1622 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) {
1623 Node* mem = in(Memory);
1624 Node* address = in(Address);
1625 intptr_t ignore = 0;
1626 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1627
1628 if (base->is_CastPP()) {
1629 base = base->in(1);
1630 }
1631
1632 if (req() > 3 || base == nullptr || !base->is_Phi()) {
1633 return false;
1634 }
1635
1636 if (!mem->is_Phi()) {
1637 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1638 return false;
1639 }
1640 } else if (base->in(0) != mem->in(0)) {
1641 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1642 return false;
1643 }
1644 }
1645
1646 return true;
1647 }
1648
1649 //------------------------------split_through_phi------------------------------
1650 // Split instance or boxed field load through Phi.
1651 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) {
1652 if (req() > 3) {
1653 assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs");
1654 // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account
1655 return nullptr;
1656 }
1657 Node* mem = in(Memory);
1658 Node* address = in(Address);
1659 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1660
1661 assert((t_oop != nullptr) &&
1662 (ignore_missing_instance_id ||
1663 t_oop->is_known_instance_field() ||
1664 t_oop->is_ptr_to_boxed_value()), "invalid conditions");
1665
1666 Compile* C = phase->C;
1667 intptr_t ignore = 0;
1668 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1669 bool base_is_phi = (base != nullptr) && base->is_Phi();
1670 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1671 (base != nullptr) && (base == address->in(AddPNode::Base)) &&
1672 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1673
1674 if (!((mem->is_Phi() || base_is_phi) &&
1675 (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) {
1676 return nullptr; // Neither memory or base are Phi
1677 }
1678
1679 if (mem->is_Phi()) {
1680 if (!stable_phi(mem->as_Phi(), phase)) {
1681 return nullptr; // Wait stable graph
1682 }
1683 uint cnt = mem->req();
1684 // Check for loop invariant memory.
1685 if (cnt == 3) {
1686 for (uint i = 1; i < cnt; i++) {
1687 Node* in = mem->in(i);
1688 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1689 if (m == mem) {
1690 if (i == 1) {
1691 // if the first edge was a loop, check second edge too.
1692 // If both are replaceable - we are in an infinite loop
1693 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1694 if (n == mem) {
1695 break;
1696 }
1697 }
1698 set_req(Memory, mem->in(cnt - i));
1699 return this; // made change
1700 }
1701 }
1702 }
1703 }
1704 if (base_is_phi) {
1705 if (!stable_phi(base->as_Phi(), phase)) {
1706 return nullptr; // Wait stable graph
1707 }
1708 uint cnt = base->req();
1709 // Check for loop invariant memory.
1710 if (cnt == 3) {
1711 for (uint i = 1; i < cnt; i++) {
1712 if (base->in(i) == base) {
1713 return nullptr; // Wait stable graph
1714 }
1715 }
1716 }
1717 }
1718
1719 // Split through Phi (see original code in loopopts.cpp).
1720 assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type");
1721
1722 // Do nothing here if Identity will find a value
1723 // (to avoid infinite chain of value phis generation).
1724 if (this != Identity(phase)) {
1725 return nullptr;
1726 }
1727
1728 // Select Region to split through.
1729 Node* region;
1730 DomResult dom_result = DomResult::Dominate;
1731 if (!base_is_phi) {
1732 assert(mem->is_Phi(), "sanity");
1733 region = mem->in(0);
1734 // Skip if the region dominates some control edge of the address.
1735 // We will check `dom_result` later.
1736 dom_result = MemNode::maybe_all_controls_dominate(address, region);
1737 } else if (!mem->is_Phi()) {
1738 assert(base_is_phi, "sanity");
1739 region = base->in(0);
1740 // Skip if the region dominates some control edge of the memory.
1741 // We will check `dom_result` later.
1742 dom_result = MemNode::maybe_all_controls_dominate(mem, region);
1743 } else if (base->in(0) != mem->in(0)) {
1744 assert(base_is_phi && mem->is_Phi(), "sanity");
1745 dom_result = MemNode::maybe_all_controls_dominate(mem, base->in(0));
1746 if (dom_result == DomResult::Dominate) {
1747 region = base->in(0);
1748 } else {
1749 dom_result = MemNode::maybe_all_controls_dominate(address, mem->in(0));
1750 if (dom_result == DomResult::Dominate) {
1751 region = mem->in(0);
1752 }
1753 // Otherwise we encountered a complex graph.
1754 }
1755 } else {
1756 assert(base->in(0) == mem->in(0), "sanity");
1757 region = mem->in(0);
1758 }
1759
1760 PhaseIterGVN* igvn = phase->is_IterGVN();
1761 if (dom_result != DomResult::Dominate) {
1762 if (dom_result == DomResult::EncounteredDeadCode) {
1763 // There is some dead code which eventually will be removed in IGVN.
1764 // Once this is the case, we get an unambiguous dominance result.
1765 // Push the node to the worklist again until the dead code is removed.
1766 igvn->_worklist.push(this);
1767 }
1768 return nullptr;
1769 }
1770
1771 Node* phi = nullptr;
1772 const Type* this_type = this->bottom_type();
1773 if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) {
1774 int this_index = C->get_alias_index(t_oop);
1775 int this_offset = t_oop->offset();
1776 int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx;
1777 phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset);
1778 } else if (ignore_missing_instance_id) {
1779 phi = new PhiNode(region, this_type, nullptr, mem->_idx);
1780 } else {
1781 return nullptr;
1782 }
1783
1784 for (uint i = 1; i < region->req(); i++) {
1785 Node* x;
1786 Node* the_clone = nullptr;
1787 Node* in = region->in(i);
1788 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1789 in != nullptr && in->is_OuterStripMinedLoop()) {
1790 // No node should go in the outer strip mined loop
1791 in = in->in(LoopNode::EntryControl);
1792 }
1793 if (in == nullptr || in == C->top()) {
1794 x = C->top(); // Dead path? Use a dead data op
1795 } else {
1796 x = this->clone(); // Else clone up the data op
1797 the_clone = x; // Remember for possible deletion.
1798 // Alter data node to use pre-phi inputs
1799 if (this->in(0) == region) {
1800 x->set_req(0, in);
1801 } else {
1802 x->set_req(0, nullptr);
1803 }
1804 if (mem->is_Phi() && (mem->in(0) == region)) {
1805 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1806 }
1807 if (address->is_Phi() && address->in(0) == region) {
1808 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1809 }
1810 if (base_is_phi && (base->in(0) == region)) {
1811 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1812 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1813 x->set_req(Address, adr_x);
1814 }
1815 }
1816 // Check for a 'win' on some paths
1817 const Type *t = x->Value(igvn);
1818
1819 bool singleton = t->singleton();
1820
1821 // See comments in PhaseIdealLoop::split_thru_phi().
1822 if (singleton && t == Type::TOP) {
1823 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1824 }
1825
1826 if (singleton) {
1827 x = igvn->makecon(t);
1828 } else {
1829 // We now call Identity to try to simplify the cloned node.
1830 // Note that some Identity methods call phase->type(this).
1831 // Make sure that the type array is big enough for
1832 // our new node, even though we may throw the node away.
1833 // (This tweaking with igvn only works because x is a new node.)
1834 igvn->set_type(x, t);
1835 // If x is a TypeNode, capture any more-precise type permanently into Node
1836 // otherwise it will be not updated during igvn->transform since
1837 // igvn->type(x) is set to x->Value() already.
1838 x->raise_bottom_type(t);
1839 Node* y = x->Identity(igvn);
1840 if (y != x) {
1841 x = y;
1842 } else {
1843 y = igvn->hash_find_insert(x);
1844 if (y) {
1845 x = y;
1846 } else {
1847 // Else x is a new node we are keeping
1848 // We do not need register_new_node_with_optimizer
1849 // because set_type has already been called.
1850 igvn->_worklist.push(x);
1851 }
1852 }
1853 }
1854 if (x != the_clone && the_clone != nullptr) {
1855 igvn->remove_dead_node(the_clone);
1856 }
1857 phi->set_req(i, x);
1858 }
1859 // Record Phi
1860 igvn->register_new_node_with_optimizer(phi);
1861 return phi;
1862 }
1863
1864 AllocateNode* LoadNode::is_new_object_mark_load() const {
1865 if (Opcode() == Op_LoadX) {
1866 Node* address = in(MemNode::Address);
1867 AllocateNode* alloc = AllocateNode::Ideal_allocation(address);
1868 Node* mem = in(MemNode::Memory);
1869 if (alloc != nullptr && mem->is_Proj() &&
1870 mem->in(0) != nullptr &&
1871 mem->in(0) == alloc->initialization() &&
1872 alloc->initialization()->proj_out_or_null(0) != nullptr) {
1873 return alloc;
1874 }
1875 }
1876 return nullptr;
1877 }
1878
1879
1880 //------------------------------Ideal------------------------------------------
1881 // If the load is from Field memory and the pointer is non-null, it might be possible to
1882 // zero out the control input.
1883 // If the offset is constant and the base is an object allocation,
1884 // try to hook me up to the exact initializing store.
1885 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1886 if (has_pinned_control_dependency()) {
1887 return nullptr;
1888 }
1889 Node* p = MemNode::Ideal_common(phase, can_reshape);
1890 if (p) return (p == NodeSentinel) ? nullptr : p;
1891
1892 Node* ctrl = in(MemNode::Control);
1893 Node* address = in(MemNode::Address);
1894
1895 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1896 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1897
1898 // Skip up past a SafePoint control. Cannot do this for Stores because
1899 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1900 if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint &&
1901 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1902 !addr_mark &&
1903 (depends_only_on_test() || has_unknown_control_dependency())) {
1904 ctrl = ctrl->in(0);
1905 set_req(MemNode::Control,ctrl);
1906 return this;
1907 }
1908
1909 intptr_t ignore = 0;
1910 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1911 if (base != nullptr
1912 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1913 // Check for useless control edge in some common special cases
1914 if (in(MemNode::Control) != nullptr
1915 && can_remove_control()
1916 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1917 && all_controls_dominate(base, phase->C->start())) {
1918 // A method-invariant, non-null address (constant or 'this' argument).
1919 set_req(MemNode::Control, nullptr);
1920 return this;
1921 }
1922 }
1923
1924 Node* mem = in(MemNode::Memory);
1925 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1926
1927 if (can_reshape && (addr_t != nullptr)) {
1928 // try to optimize our memory input
1929 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1930 if (opt_mem != mem) {
1931 set_req_X(MemNode::Memory, opt_mem, phase);
1932 if (phase->type( opt_mem ) == Type::TOP) return nullptr;
1933 return this;
1934 }
1935 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1936 if ((t_oop != nullptr) &&
1937 (t_oop->is_known_instance_field() ||
1938 t_oop->is_ptr_to_boxed_value())) {
1939 PhaseIterGVN *igvn = phase->is_IterGVN();
1940 assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true");
1941 if (igvn->_worklist.member(opt_mem)) {
1942 // Delay this transformation until memory Phi is processed.
1943 igvn->_worklist.push(this);
1944 return nullptr;
1945 }
1946 // Split instance field load through Phi.
1947 Node* result = split_through_phi(phase);
1948 if (result != nullptr) return result;
1949
1950 if (t_oop->is_ptr_to_boxed_value()) {
1951 Node* result = eliminate_autobox(igvn);
1952 if (result != nullptr) return result;
1953 }
1954 }
1955 }
1956
1957 // Is there a dominating load that loads the same value? Leave
1958 // anything that is not a load of a field/array element (like
1959 // barriers etc.) alone
1960 if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) {
1961 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1962 Node *use = mem->fast_out(i);
1963 if (use != this &&
1964 use->Opcode() == Opcode() &&
1965 use->in(0) != nullptr &&
1966 use->in(0) != in(0) &&
1967 use->in(Address) == in(Address)) {
1968 Node* ctl = in(0);
1969 for (int i = 0; i < 10 && ctl != nullptr; i++) {
1970 ctl = IfNode::up_one_dom(ctl);
1971 if (ctl == use->in(0)) {
1972 set_req(0, use->in(0));
1973 return this;
1974 }
1975 }
1976 }
1977 }
1978 }
1979
1980 // Check for prior store with a different base or offset; make Load
1981 // independent. Skip through any number of them. Bail out if the stores
1982 // are in an endless dead cycle and report no progress. This is a key
1983 // transform for Reflection. However, if after skipping through the Stores
1984 // we can't then fold up against a prior store do NOT do the transform as
1985 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1986 // array memory alive twice: once for the hoisted Load and again after the
1987 // bypassed Store. This situation only works if EVERYBODY who does
1988 // anti-dependence work knows how to bypass. I.e. we need all
1989 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1990 // the alias index stuff. So instead, peek through Stores and IFF we can
1991 // fold up, do so.
1992 Node* prev_mem = find_previous_store(phase);
1993 if (prev_mem != nullptr && prev_mem->is_top()) {
1994 // find_previous_store returns top when the access is dead
1995 return prev_mem;
1996 }
1997 if (prev_mem != nullptr) {
1998 Node* value = can_see_arraycopy_value(prev_mem, phase);
1999 if (value != nullptr) {
2000 return value;
2001 }
2002 }
2003 // Steps (a), (b): Walk past independent stores to find an exact match.
2004 if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) {
2005 // (c) See if we can fold up on the spot, but don't fold up here.
2006 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
2007 // just return a prior value, which is done by Identity calls.
2008 if (can_see_stored_value(prev_mem, phase)) {
2009 // Make ready for step (d):
2010 set_req_X(MemNode::Memory, prev_mem, phase);
2011 return this;
2012 }
2013 }
2014
2015 if (!can_reshape) {
2016 phase->record_for_igvn(this);
2017 }
2018
2019 return nullptr;
2020 }
2021
2022 // Helper to recognize certain Klass fields which are invariant across
2023 // some group of array types (e.g., int[] or all T[] where T < Object).
2024 const Type*
2025 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
2026 ciKlass* klass) const {
2027 assert(!UseCompactObjectHeaders ||
2028 AOTCodeCache::is_on_for_dump() ||
2029 tkls->offset() != in_bytes(Klass::prototype_header_offset()),
2030 "must not happen");
2031
2032 if (tkls->isa_instklassptr() && tkls->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
2033 // The field is InstanceKlass::_access_flags. Return its (constant) value.
2034 assert(Opcode() == Op_LoadUS, "must load an unsigned short from _access_flags");
2035 ciInstanceKlass* iklass = tkls->is_instklassptr()->instance_klass();
2036 return TypeInt::make(iklass->access_flags());
2037 }
2038 if (tkls->offset() == in_bytes(Klass::misc_flags_offset())) {
2039 // The field is Klass::_misc_flags. Return its (constant) value.
2040 assert(Opcode() == Op_LoadUB, "must load an unsigned byte from _misc_flags");
2041 return TypeInt::make(klass->misc_flags());
2042 }
2043 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
2044 // The field is Klass::_layout_helper. Return its constant value if known.
2045 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2046 return TypeInt::make(klass->layout_helper());
2047 }
2048
2049 // No match.
2050 return nullptr;
2051 }
2052
2053 //------------------------------Value-----------------------------------------
2054 const Type* LoadNode::Value(PhaseGVN* phase) const {
2055 // Either input is TOP ==> the result is TOP
2056 Node* mem = in(MemNode::Memory);
2057 const Type *t1 = phase->type(mem);
2058 if (t1 == Type::TOP) return Type::TOP;
2059 Node* adr = in(MemNode::Address);
2060 const TypePtr* tp = phase->type(adr)->isa_ptr();
2061 if (tp == nullptr || tp->empty()) return Type::TOP;
2062 int off = tp->offset();
2063 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
2064 Compile* C = phase->C;
2065
2066 // If load can see a previous constant store, use that.
2067 Node* value = can_see_stored_value(mem, phase);
2068 if (value != nullptr && value->is_Con()) {
2069 assert(value->bottom_type()->higher_equal(_type), "sanity");
2070 return value->bottom_type();
2071 }
2072
2073 // Try to guess loaded type from pointer type
2074 if (tp->isa_aryptr()) {
2075 const TypeAryPtr* ary = tp->is_aryptr();
2076 const Type* t = ary->elem();
2077
2078 // Determine whether the reference is beyond the header or not, by comparing
2079 // the offset against the offset of the start of the array's data.
2080 // Different array types begin at slightly different offsets (12 vs. 16).
2081 // We choose T_BYTE as an example base type that is least restrictive
2082 // as to alignment, which will therefore produce the smallest
2083 // possible base offset.
2084 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2085 const bool off_beyond_header = (off >= min_base_off);
2086
2087 // Try to constant-fold a stable array element.
2088 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
2089 // Make sure the reference is not into the header and the offset is constant
2090 ciObject* aobj = ary->const_oop();
2091 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
2092 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
2093 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
2094 stable_dimension,
2095 value_basic_type(), is_unsigned());
2096 if (con_type != nullptr) {
2097 return con_type;
2098 }
2099 }
2100 }
2101
2102 // Don't do this for integer types. There is only potential profit if
2103 // the element type t is lower than _type; that is, for int types, if _type is
2104 // more restrictive than t. This only happens here if one is short and the other
2105 // char (both 16 bits), and in those cases we've made an intentional decision
2106 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2107 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2108 //
2109 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2110 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
2111 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2112 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
2113 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2114 // In fact, that could have been the original type of p1, and p1 could have
2115 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2116 // expression (LShiftL quux 3) independently optimized to the constant 8.
2117 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2118 && (_type->isa_vect() == nullptr)
2119 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2120 // t might actually be lower than _type, if _type is a unique
2121 // concrete subclass of abstract class t.
2122 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
2123 const Type* jt = t->join_speculative(_type);
2124 // In any case, do not allow the join, per se, to empty out the type.
2125 if (jt->empty() && !t->empty()) {
2126 // This can happen if a interface-typed array narrows to a class type.
2127 jt = _type;
2128 }
2129 #ifdef ASSERT
2130 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2131 // The pointers in the autobox arrays are always non-null
2132 Node* base = adr->in(AddPNode::Base);
2133 if ((base != nullptr) && base->is_DecodeN()) {
2134 // Get LoadN node which loads IntegerCache.cache field
2135 base = base->in(1);
2136 }
2137 if ((base != nullptr) && base->is_Con()) {
2138 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2139 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2140 // It could be narrow oop
2141 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2142 }
2143 }
2144 }
2145 #endif
2146 return jt;
2147 }
2148 }
2149 } else if (tp->base() == Type::InstPtr) {
2150 assert( off != Type::OffsetBot ||
2151 // arrays can be cast to Objects
2152 !tp->isa_instptr() ||
2153 tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
2154 // unsafe field access may not have a constant offset
2155 C->has_unsafe_access(),
2156 "Field accesses must be precise" );
2157 // For oop loads, we expect the _type to be precise.
2158
2159 // Optimize loads from constant fields.
2160 const TypeInstPtr* tinst = tp->is_instptr();
2161 ciObject* const_oop = tinst->const_oop();
2162 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2163 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), value_basic_type());
2164 if (con_type != nullptr) {
2165 return con_type;
2166 }
2167 }
2168 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2169 assert(off != Type::OffsetBot ||
2170 !tp->isa_instklassptr() ||
2171 // arrays can be cast to Objects
2172 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2173 // also allow array-loading from the primary supertype
2174 // array during subtype checks
2175 Opcode() == Op_LoadKlass,
2176 "Field accesses must be precise");
2177 // For klass/static loads, we expect the _type to be precise
2178 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2179 /* With mirrors being an indirect in the Klass*
2180 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2181 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2182 *
2183 * So check the type and klass of the node before the LoadP.
2184 */
2185 Node* adr2 = adr->in(MemNode::Address);
2186 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2187 if (tkls != nullptr && !StressReflectiveCode) {
2188 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2189 ciKlass* klass = tkls->exact_klass();
2190 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2191 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2192 return TypeInstPtr::make(klass->java_mirror());
2193 }
2194 }
2195 }
2196
2197 const TypeKlassPtr *tkls = tp->isa_klassptr();
2198 if (tkls != nullptr) {
2199 if (tkls->is_loaded() && tkls->klass_is_exact()) {
2200 ciKlass* klass = tkls->exact_klass();
2201 // We are loading a field from a Klass metaobject whose identity
2202 // is known at compile time (the type is "exact" or "precise").
2203 // Check for fields we know are maintained as constants by the VM.
2204 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2205 // The field is Klass::_super_check_offset. Return its (constant) value.
2206 // (Folds up type checking code.)
2207 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2208 return TypeInt::make(klass->super_check_offset());
2209 }
2210 // Class encoding in prototype header may change between runs.
2211 // Force loading prototype header when AOT code is generated.
2212 if (UseCompactObjectHeaders && !AOTCodeCache::is_on_for_dump()) {
2213 if (tkls->offset() == in_bytes(Klass::prototype_header_offset())) {
2214 // The field is Klass::_prototype_header. Return its (constant) value.
2215 assert(this->Opcode() == Op_LoadX, "must load a proper type from _prototype_header");
2216 return TypeX::make(klass->prototype_header());
2217 }
2218 }
2219 // Compute index into primary_supers array
2220 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2221 // Check for overflowing; use unsigned compare to handle the negative case.
2222 if( depth < ciKlass::primary_super_limit() ) {
2223 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2224 // (Folds up type checking code.)
2225 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2226 ciKlass *ss = klass->super_of_depth(depth);
2227 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2228 }
2229 const Type* aift = load_array_final_field(tkls, klass);
2230 if (aift != nullptr) return aift;
2231 }
2232
2233 // We can still check if we are loading from the primary_supers array at a
2234 // shallow enough depth. Even though the klass is not exact, entries less
2235 // than or equal to its super depth are correct.
2236 if (tkls->is_loaded()) {
2237 ciKlass* klass = nullptr;
2238 if (tkls->isa_instklassptr()) {
2239 klass = tkls->is_instklassptr()->instance_klass();
2240 } else {
2241 int dims;
2242 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims);
2243 if (inner->isa_instklassptr()) {
2244 klass = inner->is_instklassptr()->instance_klass();
2245 klass = ciObjArrayKlass::make(klass, dims);
2246 }
2247 }
2248 if (klass != nullptr) {
2249 // Compute index into primary_supers array
2250 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2251 // Check for overflowing; use unsigned compare to handle the negative case.
2252 if (depth < ciKlass::primary_super_limit() &&
2253 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case
2254 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2255 // (Folds up type checking code.)
2256 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2257 ciKlass *ss = klass->super_of_depth(depth);
2258 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2259 }
2260 }
2261 }
2262
2263 // If the type is enough to determine that the thing is not an array,
2264 // we can give the layout_helper a positive interval type.
2265 // This will help short-circuit some reflective code.
2266 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) &&
2267 tkls->isa_instklassptr() && // not directly typed as an array
2268 !tkls->is_instklassptr()->might_be_an_array() // not the supertype of all T[] (java.lang.Object) or has an interface that is not Serializable or Cloneable
2269 ) {
2270 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2271 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2272 // The key property of this type is that it folds up tests
2273 // for array-ness, since it proves that the layout_helper is positive.
2274 // Thus, a generic value like the basic object layout helper works fine.
2275 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2276 }
2277 }
2278
2279 // If we are loading from a freshly-allocated object/array, produce a zero.
2280 // Things to check:
2281 // 1. Load is beyond the header: headers are not guaranteed to be zero
2282 // 2. Load is not vectorized: vectors have no zero constant
2283 // 3. Load has no matching store, i.e. the input is the initial memory state
2284 const TypeOopPtr* tinst = tp->isa_oopptr();
2285 bool is_not_header = (tinst != nullptr) && tinst->is_known_instance_field();
2286 bool is_not_vect = (_type->isa_vect() == nullptr);
2287 if (is_not_header && is_not_vect) {
2288 Node* mem = in(MemNode::Memory);
2289 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2290 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2291 return Type::get_zero_type(_type->basic_type());
2292 }
2293 }
2294
2295 if (!UseCompactObjectHeaders) {
2296 Node* alloc = is_new_object_mark_load();
2297 if (alloc != nullptr) {
2298 return TypeX::make(markWord::prototype().value());
2299 }
2300 }
2301
2302 return _type;
2303 }
2304
2305 //------------------------------match_edge-------------------------------------
2306 // Do we Match on this edge index or not? Match only the address.
2307 uint LoadNode::match_edge(uint idx) const {
2308 return idx == MemNode::Address;
2309 }
2310
2311 //--------------------------LoadBNode::Ideal--------------------------------------
2312 //
2313 // If the previous store is to the same address as this load,
2314 // and the value stored was larger than a byte, replace this load
2315 // with the value stored truncated to a byte. If no truncation is
2316 // needed, the replacement is done in LoadNode::Identity().
2317 //
2318 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2319 Node* mem = in(MemNode::Memory);
2320 Node* value = can_see_stored_value(mem,phase);
2321 if (value != nullptr) {
2322 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2323 if (narrow != value) {
2324 return narrow;
2325 }
2326 }
2327 // Identity call will handle the case where truncation is not needed.
2328 return LoadNode::Ideal(phase, can_reshape);
2329 }
2330
2331 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2332 Node* mem = in(MemNode::Memory);
2333 Node* value = can_see_stored_value(mem,phase);
2334 if (value != nullptr && value->is_Con() &&
2335 !value->bottom_type()->higher_equal(_type)) {
2336 // If the input to the store does not fit with the load's result type,
2337 // it must be truncated. We can't delay until Ideal call since
2338 // a singleton Value is needed for split_thru_phi optimization.
2339 int con = value->get_int();
2340 return TypeInt::make((con << 24) >> 24);
2341 }
2342 return LoadNode::Value(phase);
2343 }
2344
2345 //--------------------------LoadUBNode::Ideal-------------------------------------
2346 //
2347 // If the previous store is to the same address as this load,
2348 // and the value stored was larger than a byte, replace this load
2349 // with the value stored truncated to a byte. If no truncation is
2350 // needed, the replacement is done in LoadNode::Identity().
2351 //
2352 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2353 Node* mem = in(MemNode::Memory);
2354 Node* value = can_see_stored_value(mem, phase);
2355 if (value != nullptr) {
2356 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2357 if (narrow != value) {
2358 return narrow;
2359 }
2360 }
2361 // Identity call will handle the case where truncation is not needed.
2362 return LoadNode::Ideal(phase, can_reshape);
2363 }
2364
2365 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2366 Node* mem = in(MemNode::Memory);
2367 Node* value = can_see_stored_value(mem,phase);
2368 if (value != nullptr && value->is_Con() &&
2369 !value->bottom_type()->higher_equal(_type)) {
2370 // If the input to the store does not fit with the load's result type,
2371 // it must be truncated. We can't delay until Ideal call since
2372 // a singleton Value is needed for split_thru_phi optimization.
2373 int con = value->get_int();
2374 return TypeInt::make(con & 0xFF);
2375 }
2376 return LoadNode::Value(phase);
2377 }
2378
2379 //--------------------------LoadUSNode::Ideal-------------------------------------
2380 //
2381 // If the previous store is to the same address as this load,
2382 // and the value stored was larger than a char, replace this load
2383 // with the value stored truncated to a char. If no truncation is
2384 // needed, the replacement is done in LoadNode::Identity().
2385 //
2386 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2387 Node* mem = in(MemNode::Memory);
2388 Node* value = can_see_stored_value(mem,phase);
2389 if (value != nullptr) {
2390 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2391 if (narrow != value) {
2392 return narrow;
2393 }
2394 }
2395 // Identity call will handle the case where truncation is not needed.
2396 return LoadNode::Ideal(phase, can_reshape);
2397 }
2398
2399 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2400 Node* mem = in(MemNode::Memory);
2401 Node* value = can_see_stored_value(mem,phase);
2402 if (value != nullptr && value->is_Con() &&
2403 !value->bottom_type()->higher_equal(_type)) {
2404 // If the input to the store does not fit with the load's result type,
2405 // it must be truncated. We can't delay until Ideal call since
2406 // a singleton Value is needed for split_thru_phi optimization.
2407 int con = value->get_int();
2408 return TypeInt::make(con & 0xFFFF);
2409 }
2410 return LoadNode::Value(phase);
2411 }
2412
2413 //--------------------------LoadSNode::Ideal--------------------------------------
2414 //
2415 // If the previous store is to the same address as this load,
2416 // and the value stored was larger than a short, replace this load
2417 // with the value stored truncated to a short. If no truncation is
2418 // needed, the replacement is done in LoadNode::Identity().
2419 //
2420 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2421 Node* mem = in(MemNode::Memory);
2422 Node* value = can_see_stored_value(mem,phase);
2423 if (value != nullptr) {
2424 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2425 if (narrow != value) {
2426 return narrow;
2427 }
2428 }
2429 // Identity call will handle the case where truncation is not needed.
2430 return LoadNode::Ideal(phase, can_reshape);
2431 }
2432
2433 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2434 Node* mem = in(MemNode::Memory);
2435 Node* value = can_see_stored_value(mem,phase);
2436 if (value != nullptr && value->is_Con() &&
2437 !value->bottom_type()->higher_equal(_type)) {
2438 // If the input to the store does not fit with the load's result type,
2439 // it must be truncated. We can't delay until Ideal call since
2440 // a singleton Value is needed for split_thru_phi optimization.
2441 int con = value->get_int();
2442 return TypeInt::make((con << 16) >> 16);
2443 }
2444 return LoadNode::Value(phase);
2445 }
2446
2447 //=============================================================================
2448 //----------------------------LoadKlassNode::make------------------------------
2449 // Polymorphic factory method:
2450 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2451 // sanity check the alias category against the created node type
2452 const TypePtr* adr_type = adr->bottom_type()->isa_ptr();
2453 assert(adr_type != nullptr, "expecting TypeKlassPtr");
2454 #ifdef _LP64
2455 if (adr_type->is_ptr_to_narrowklass()) {
2456 assert(UseCompressedClassPointers, "no compressed klasses");
2457 Node* load_klass = gvn.transform(new LoadNKlassNode(mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2458 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2459 }
2460 #endif
2461 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2462 return new LoadKlassNode(mem, adr, at, tk, MemNode::unordered);
2463 }
2464
2465 //------------------------------Value------------------------------------------
2466 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2467 return klass_value_common(phase);
2468 }
2469
2470 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2471 // Either input is TOP ==> the result is TOP
2472 const Type *t1 = phase->type( in(MemNode::Memory) );
2473 if (t1 == Type::TOP) return Type::TOP;
2474 Node *adr = in(MemNode::Address);
2475 const Type *t2 = phase->type( adr );
2476 if (t2 == Type::TOP) return Type::TOP;
2477 const TypePtr *tp = t2->is_ptr();
2478 if (TypePtr::above_centerline(tp->ptr()) ||
2479 tp->ptr() == TypePtr::Null) return Type::TOP;
2480
2481 // Return a more precise klass, if possible
2482 const TypeInstPtr *tinst = tp->isa_instptr();
2483 if (tinst != nullptr) {
2484 ciInstanceKlass* ik = tinst->instance_klass();
2485 int offset = tinst->offset();
2486 if (ik == phase->C->env()->Class_klass()
2487 && (offset == java_lang_Class::klass_offset() ||
2488 offset == java_lang_Class::array_klass_offset())) {
2489 // We are loading a special hidden field from a Class mirror object,
2490 // the field which points to the VM's Klass metaobject.
2491 ciType* t = tinst->java_mirror_type();
2492 // java_mirror_type returns non-null for compile-time Class constants.
2493 if (t != nullptr) {
2494 // constant oop => constant klass
2495 if (offset == java_lang_Class::array_klass_offset()) {
2496 if (t->is_void()) {
2497 // We cannot create a void array. Since void is a primitive type return null
2498 // klass. Users of this result need to do a null check on the returned klass.
2499 return TypePtr::NULL_PTR;
2500 }
2501 return TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);
2502 }
2503 if (!t->is_klass()) {
2504 // a primitive Class (e.g., int.class) has null for a klass field
2505 return TypePtr::NULL_PTR;
2506 }
2507 // Fold up the load of the hidden field
2508 return TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);
2509 }
2510 // non-constant mirror, so we can't tell what's going on
2511 }
2512 if (!tinst->is_loaded())
2513 return _type; // Bail out if not loaded
2514 if (offset == oopDesc::klass_offset_in_bytes()) {
2515 return tinst->as_klass_type(true);
2516 }
2517 }
2518
2519 // Check for loading klass from an array
2520 const TypeAryPtr *tary = tp->isa_aryptr();
2521 if (tary != nullptr &&
2522 tary->offset() == oopDesc::klass_offset_in_bytes()) {
2523 return tary->as_klass_type(true);
2524 }
2525
2526 // Check for loading klass from an array klass
2527 const TypeKlassPtr *tkls = tp->isa_klassptr();
2528 if (tkls != nullptr && !StressReflectiveCode) {
2529 if (!tkls->is_loaded())
2530 return _type; // Bail out if not loaded
2531 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2532 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2533 // // Always returning precise element type is incorrect,
2534 // // e.g., element type could be object and array may contain strings
2535 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2536
2537 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2538 // according to the element type's subclassing.
2539 return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2540 }
2541 if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() &&
2542 tkls->offset() == in_bytes(Klass::super_offset())) {
2543 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super();
2544 // The field is Klass::_super. Return its (constant) value.
2545 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2546 return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR;
2547 }
2548 }
2549
2550 if (tkls != nullptr && !UseSecondarySupersCache
2551 && tkls->offset() == in_bytes(Klass::secondary_super_cache_offset())) {
2552 // Treat Klass::_secondary_super_cache as a constant when the cache is disabled.
2553 return TypePtr::NULL_PTR;
2554 }
2555
2556 // Bailout case
2557 return LoadNode::Value(phase);
2558 }
2559
2560 //------------------------------Identity---------------------------------------
2561 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2562 // Also feed through the klass in Allocate(...klass...)._klass.
2563 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2564 return klass_identity_common(phase);
2565 }
2566
2567 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2568 Node* x = LoadNode::Identity(phase);
2569 if (x != this) return x;
2570
2571 // Take apart the address into an oop and offset.
2572 // Return 'this' if we cannot.
2573 Node* adr = in(MemNode::Address);
2574 intptr_t offset = 0;
2575 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2576 if (base == nullptr) return this;
2577 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2578 if (toop == nullptr) return this;
2579
2580 // Step over potential GC barrier for OopHandle resolve
2581 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2582 if (bs->is_gc_barrier_node(base)) {
2583 base = bs->step_over_gc_barrier(base);
2584 }
2585
2586 // We can fetch the klass directly through an AllocateNode.
2587 // This works even if the klass is not constant (clone or newArray).
2588 if (offset == oopDesc::klass_offset_in_bytes()) {
2589 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2590 if (allocated_klass != nullptr) {
2591 return allocated_klass;
2592 }
2593 }
2594
2595 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2596 // See inline_native_Class_query for occurrences of these patterns.
2597 // Java Example: x.getClass().isAssignableFrom(y)
2598 //
2599 // This improves reflective code, often making the Class
2600 // mirror go completely dead. (Current exception: Class
2601 // mirrors may appear in debug info, but we could clean them out by
2602 // introducing a new debug info operator for Klass.java_mirror).
2603
2604 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass()
2605 && offset == java_lang_Class::klass_offset()) {
2606 if (base->is_Load()) {
2607 Node* base2 = base->in(MemNode::Address);
2608 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2609 Node* adr2 = base2->in(MemNode::Address);
2610 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2611 if (tkls != nullptr && !tkls->empty()
2612 && (tkls->isa_instklassptr() || tkls->isa_aryklassptr())
2613 && adr2->is_AddP()
2614 ) {
2615 int mirror_field = in_bytes(Klass::java_mirror_offset());
2616 if (tkls->offset() == mirror_field) {
2617 #ifdef ASSERT
2618 const TypeKlassPtr* tkls2 = phase->type(adr2->in(AddPNode::Address))->is_klassptr();
2619 assert(tkls2->offset() == 0, "not a load of java_mirror");
2620 #endif
2621 assert(adr2->in(AddPNode::Base)->is_top(), "not an off heap load");
2622 assert(adr2->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(Klass::java_mirror_offset()), "incorrect offset");
2623 return adr2->in(AddPNode::Address);
2624 }
2625 }
2626 }
2627 }
2628 }
2629
2630 return this;
2631 }
2632
2633 LoadNode* LoadNode::clone_pinned() const {
2634 LoadNode* ld = clone()->as_Load();
2635 ld->_control_dependency = UnknownControl;
2636 return ld;
2637 }
2638
2639 // Pin a LoadNode if it carries a dependency on its control input. There are cases when the node
2640 // does not actually have any dependency on its control input. For example, if we have a LoadNode
2641 // being used only outside a loop but it must be scheduled inside the loop, we can clone the node
2642 // for each of its use so that all the clones can be scheduled outside the loop. Then, to prevent
2643 // the clones from being GVN-ed again, we add a control input for each of them at the loop exit. In
2644 // those case, since there is not a dependency between the node and its control input, we do not
2645 // need to pin it.
2646 LoadNode* LoadNode::pin_node_under_control_impl() const {
2647 const TypePtr* adr_type = this->adr_type();
2648 if (adr_type != nullptr && adr_type->isa_aryptr()) {
2649 // Only array accesses have dependencies on their control input
2650 return clone_pinned();
2651 }
2652 return nullptr;
2653 }
2654
2655 //------------------------------Value------------------------------------------
2656 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2657 const Type *t = klass_value_common(phase);
2658 if (t == Type::TOP)
2659 return t;
2660
2661 return t->make_narrowklass();
2662 }
2663
2664 //------------------------------Identity---------------------------------------
2665 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2666 // Also feed through the klass in Allocate(...klass...)._klass.
2667 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2668 Node *x = klass_identity_common(phase);
2669
2670 const Type *t = phase->type( x );
2671 if( t == Type::TOP ) return x;
2672 if( t->isa_narrowklass()) return x;
2673 assert (!t->isa_narrowoop(), "no narrow oop here");
2674
2675 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2676 }
2677
2678 //------------------------------Value-----------------------------------------
2679 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2680 // Either input is TOP ==> the result is TOP
2681 const Type *t1 = phase->type( in(MemNode::Memory) );
2682 if( t1 == Type::TOP ) return Type::TOP;
2683 Node *adr = in(MemNode::Address);
2684 const Type *t2 = phase->type( adr );
2685 if( t2 == Type::TOP ) return Type::TOP;
2686 const TypePtr *tp = t2->is_ptr();
2687 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2688 const TypeAryPtr *tap = tp->isa_aryptr();
2689 if( !tap ) return _type;
2690 return tap->size();
2691 }
2692
2693 //-------------------------------Ideal---------------------------------------
2694 // Feed through the length in AllocateArray(...length...)._length.
2695 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2696 Node* p = MemNode::Ideal_common(phase, can_reshape);
2697 if (p) return (p == NodeSentinel) ? nullptr : p;
2698
2699 // Take apart the address into an oop and offset.
2700 // Return 'this' if we cannot.
2701 Node* adr = in(MemNode::Address);
2702 intptr_t offset = 0;
2703 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2704 if (base == nullptr) return nullptr;
2705 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2706 if (tary == nullptr) return nullptr;
2707
2708 // We can fetch the length directly through an AllocateArrayNode.
2709 // This works even if the length is not constant (clone or newArray).
2710 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2711 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2712 if (alloc != nullptr) {
2713 Node* allocated_length = alloc->Ideal_length();
2714 Node* len = alloc->make_ideal_length(tary, phase);
2715 if (allocated_length != len) {
2716 // New CastII improves on this.
2717 return len;
2718 }
2719 }
2720 }
2721
2722 return nullptr;
2723 }
2724
2725 //------------------------------Identity---------------------------------------
2726 // Feed through the length in AllocateArray(...length...)._length.
2727 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2728 Node* x = LoadINode::Identity(phase);
2729 if (x != this) return x;
2730
2731 // Take apart the address into an oop and offset.
2732 // Return 'this' if we cannot.
2733 Node* adr = in(MemNode::Address);
2734 intptr_t offset = 0;
2735 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2736 if (base == nullptr) return this;
2737 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2738 if (tary == nullptr) return this;
2739
2740 // We can fetch the length directly through an AllocateArrayNode.
2741 // This works even if the length is not constant (clone or newArray).
2742 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2743 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2744 if (alloc != nullptr) {
2745 Node* allocated_length = alloc->Ideal_length();
2746 // Do not allow make_ideal_length to allocate a CastII node.
2747 Node* len = alloc->make_ideal_length(tary, phase, false);
2748 if (allocated_length == len) {
2749 // Return allocated_length only if it would not be improved by a CastII.
2750 return allocated_length;
2751 }
2752 }
2753 }
2754
2755 return this;
2756
2757 }
2758
2759 //=============================================================================
2760 //---------------------------StoreNode::make-----------------------------------
2761 // Polymorphic factory method:
2762 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) {
2763 assert((mo == unordered || mo == release), "unexpected");
2764 Compile* C = gvn.C;
2765 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2766 ctl != nullptr, "raw memory operations should have control edge");
2767
2768 switch (bt) {
2769 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2770 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2771 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2772 case T_CHAR:
2773 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2774 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2775 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2776 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2777 case T_METADATA:
2778 case T_ADDRESS:
2779 case T_OBJECT:
2780 #ifdef _LP64
2781 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2782 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2783 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2784 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2785 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2786 adr->bottom_type()->isa_rawptr())) {
2787 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2788 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2789 }
2790 #endif
2791 {
2792 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2793 }
2794 default:
2795 ShouldNotReachHere();
2796 return (StoreNode*)nullptr;
2797 }
2798 }
2799
2800 //--------------------------bottom_type----------------------------------------
2801 const Type *StoreNode::bottom_type() const {
2802 return Type::MEMORY;
2803 }
2804
2805 //------------------------------hash-------------------------------------------
2806 uint StoreNode::hash() const {
2807 // unroll addition of interesting fields
2808 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2809
2810 // Since they are not commoned, do not hash them:
2811 return NO_HASH;
2812 }
2813
2814 // Link together multiple stores (B/S/C/I) into a longer one.
2815 //
2816 // Example: _store = StoreB[i+3]
2817 //
2818 // RangeCheck[i+0] RangeCheck[i+0]
2819 // StoreB[i+0]
2820 // RangeCheck[i+3] RangeCheck[i+3]
2821 // StoreB[i+1] --> pass: fail:
2822 // StoreB[i+2] StoreI[i+0] StoreB[i+0]
2823 // StoreB[i+3]
2824 //
2825 // The 4 StoreB are merged into a single StoreI node. We have to be careful with RangeCheck[i+1]: before
2826 // the optimization, if this RangeCheck[i+1] fails, then we execute only StoreB[i+0], and then trap. After
2827 // the optimization, the new StoreI[i+0] is on the passing path of RangeCheck[i+3], and StoreB[i+0] on the
2828 // failing path.
2829 //
2830 // Note: For normal array stores, every store at first has a RangeCheck. But they can be removed with:
2831 // - RCE (RangeCheck Elimination): the RangeChecks in the loop are hoisted out and before the loop,
2832 // and possibly no RangeChecks remain between the stores.
2833 // - RangeCheck smearing: the earlier RangeChecks are adjusted such that they cover later RangeChecks,
2834 // and those later RangeChecks can be removed. Example:
2835 //
2836 // RangeCheck[i+0] RangeCheck[i+0] <- before first store
2837 // StoreB[i+0] StoreB[i+0] <- first store
2838 // RangeCheck[i+1] --> smeared --> RangeCheck[i+3] <- only RC between first and last store
2839 // StoreB[i+1] StoreB[i+1] <- second store
2840 // RangeCheck[i+2] --> removed
2841 // StoreB[i+2] StoreB[i+2]
2842 // RangeCheck[i+3] --> removed
2843 // StoreB[i+3] StoreB[i+3] <- last store
2844 //
2845 // Thus, it is a common pattern that between the first and last store in a chain
2846 // of adjacent stores there remains exactly one RangeCheck, located between the
2847 // first and the second store (e.g. RangeCheck[i+3]).
2848 //
2849 class MergePrimitiveStores : public StackObj {
2850 private:
2851 PhaseGVN* const _phase;
2852 StoreNode* const _store;
2853 // State machine with initial state Unknown
2854 // Allowed transitions:
2855 // Unknown -> Const
2856 // Unknown -> Platform
2857 // Unknown -> Reverse
2858 // Unknown -> NotAdjacent
2859 // Const -> Const
2860 // Const -> NotAdjacent
2861 // Platform -> Platform
2862 // Platform -> NotAdjacent
2863 // Reverse -> Reverse
2864 // Reverse -> NotAdjacent
2865 // NotAdjacent -> NotAdjacent
2866 enum ValueOrder : uint8_t {
2867 Unknown, // Initial state
2868 Const, // Input values are const
2869 Platform, // Platform order
2870 Reverse, // Reverse platform order
2871 NotAdjacent // Not adjacent
2872 };
2873 ValueOrder _value_order;
2874
2875 NOT_PRODUCT( const CHeapBitMap &_trace_tags; )
2876
2877 public:
2878 MergePrimitiveStores(PhaseGVN* phase, StoreNode* store) :
2879 _phase(phase), _store(store), _value_order(ValueOrder::Unknown)
2880 NOT_PRODUCT( COMMA _trace_tags(Compile::current()->directive()->trace_merge_stores_tags()) )
2881 {}
2882
2883 StoreNode* run();
2884
2885 private:
2886 bool is_compatible_store(const StoreNode* other_store) const;
2887 bool is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const;
2888 bool is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const;
2889 static bool is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase);
2890 enum CFGStatus { SuccessNoRangeCheck, SuccessWithRangeCheck, Failure };
2891 static CFGStatus cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store);
2892
2893 class Status {
2894 private:
2895 StoreNode* _found_store;
2896 bool _found_range_check;
2897
2898 Status(StoreNode* found_store, bool found_range_check)
2899 : _found_store(found_store), _found_range_check(found_range_check) {}
2900
2901 public:
2902 StoreNode* found_store() const { return _found_store; }
2903 bool found_range_check() const { return _found_range_check; }
2904 static Status make_failure() { return Status(nullptr, false); }
2905
2906 static Status make(StoreNode* found_store, const CFGStatus cfg_status) {
2907 if (cfg_status == CFGStatus::Failure) {
2908 return Status::make_failure();
2909 }
2910 return Status(found_store, cfg_status == CFGStatus::SuccessWithRangeCheck);
2911 }
2912
2913 #ifndef PRODUCT
2914 void print_on(outputStream* st) const {
2915 if (_found_store == nullptr) {
2916 st->print_cr("None");
2917 } else {
2918 st->print_cr("Found[%d %s, %s]", _found_store->_idx, _found_store->Name(),
2919 _found_range_check ? "RC" : "no-RC");
2920 }
2921 }
2922 #endif
2923 };
2924
2925 enum ValueOrder find_adjacent_input_value_order(const Node* n1, const Node* n2, const int memory_size) const;
2926 Status find_adjacent_use_store(const StoreNode* def_store) const;
2927 Status find_adjacent_def_store(const StoreNode* use_store) const;
2928 Status find_use_store(const StoreNode* def_store) const;
2929 Status find_def_store(const StoreNode* use_store) const;
2930 Status find_use_store_unidirectional(const StoreNode* def_store) const;
2931 Status find_def_store_unidirectional(const StoreNode* use_store) const;
2932
2933 void collect_merge_list(Node_List& merge_list) const;
2934 Node* make_merged_input_value(const Node_List& merge_list);
2935 StoreNode* make_merged_store(const Node_List& merge_list, Node* merged_input_value);
2936
2937 #ifndef PRODUCT
2938 // Access to TraceMergeStores tags
2939 bool is_trace(TraceMergeStores::Tag tag) const {
2940 return _trace_tags.at(tag);
2941 }
2942
2943 bool is_trace_basic() const {
2944 return is_trace(TraceMergeStores::Tag::BASIC);
2945 }
2946
2947 bool is_trace_pointer_parsing() const {
2948 return is_trace(TraceMergeStores::Tag::POINTER_PARSING);
2949 }
2950
2951 bool is_trace_pointer_aliasing() const {
2952 return is_trace(TraceMergeStores::Tag::POINTER_ALIASING);
2953 }
2954
2955 bool is_trace_pointer_adjacency() const {
2956 return is_trace(TraceMergeStores::Tag::POINTER_ADJACENCY);
2957 }
2958
2959 bool is_trace_success() const {
2960 return is_trace(TraceMergeStores::Tag::SUCCESS);
2961 }
2962 #endif
2963
2964 NOT_PRODUCT( void trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const; )
2965 };
2966
2967 StoreNode* MergePrimitiveStores::run() {
2968 // Check for B/S/C/I
2969 int opc = _store->Opcode();
2970 if (opc != Op_StoreB && opc != Op_StoreC && opc != Op_StoreI) {
2971 return nullptr;
2972 }
2973
2974 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] MergePrimitiveStores::run: "); _store->dump(); })
2975
2976 // The _store must be the "last" store in a chain. If we find a use we could merge with
2977 // then that use or a store further down is the "last" store.
2978 Status status_use = find_adjacent_use_store(_store);
2979 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect no use: "); status_use.print_on(tty); })
2980 if (status_use.found_store() != nullptr) {
2981 return nullptr;
2982 }
2983
2984 // Check if we can merge with at least one def, so that we have at least 2 stores to merge.
2985 Status status_def = find_adjacent_def_store(_store);
2986 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect def: "); status_def.print_on(tty); })
2987 Node* def_store = status_def.found_store();
2988 if (def_store == nullptr) {
2989 return nullptr;
2990 }
2991
2992 // Initialize value order
2993 _value_order = find_adjacent_input_value_order(def_store->in(MemNode::ValueIn),
2994 _store->in(MemNode::ValueIn),
2995 _store->memory_size());
2996 assert(_value_order != ValueOrder::NotAdjacent && _value_order != ValueOrder::Unknown, "Order should be checked");
2997
2998 ResourceMark rm;
2999 Node_List merge_list;
3000 collect_merge_list(merge_list);
3001
3002 Node* merged_input_value = make_merged_input_value(merge_list);
3003 if (merged_input_value == nullptr) { return nullptr; }
3004
3005 StoreNode* merged_store = make_merged_store(merge_list, merged_input_value);
3006
3007 NOT_PRODUCT( if (is_trace_success()) { trace(merge_list, merged_input_value, merged_store); } )
3008
3009 return merged_store;
3010 }
3011
3012 // Check compatibility between _store and other_store.
3013 bool MergePrimitiveStores::is_compatible_store(const StoreNode* other_store) const {
3014 int opc = _store->Opcode();
3015 assert(opc == Op_StoreB || opc == Op_StoreC || opc == Op_StoreI, "precondition");
3016
3017 if (other_store == nullptr ||
3018 _store->Opcode() != other_store->Opcode()) {
3019 return false;
3020 }
3021
3022 return true;
3023 }
3024
3025 bool MergePrimitiveStores::is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const {
3026 if (!is_adjacent_input_pair(def_store->in(MemNode::ValueIn),
3027 use_store->in(MemNode::ValueIn),
3028 def_store->memory_size())) {
3029 return false;
3030 }
3031
3032 ResourceMark rm;
3033 #ifndef PRODUCT
3034 const TraceMemPointer trace(is_trace_pointer_parsing(),
3035 is_trace_pointer_aliasing(),
3036 is_trace_pointer_adjacency(),
3037 true);
3038 #endif
3039 const MemPointer pointer_use(use_store NOT_PRODUCT(COMMA trace));
3040 const MemPointer pointer_def(def_store NOT_PRODUCT(COMMA trace));
3041 return pointer_def.is_adjacent_to_and_before(pointer_use);
3042 }
3043
3044 // Check input values n1 and n2 can be merged and return the value order
3045 MergePrimitiveStores::ValueOrder MergePrimitiveStores::find_adjacent_input_value_order(const Node* n1, const Node* n2,
3046 const int memory_size) const {
3047 // Pattern: [n1 = ConI, n2 = ConI]
3048 if (n1->Opcode() == Op_ConI && n2->Opcode() == Op_ConI) {
3049 return ValueOrder::Const;
3050 }
3051
3052 Node const *base_n2;
3053 jint shift_n2;
3054 if (!is_con_RShift(n2, base_n2, shift_n2, _phase)) {
3055 return ValueOrder::NotAdjacent;
3056 }
3057 Node const *base_n1;
3058 jint shift_n1;
3059 if (!is_con_RShift(n1, base_n1, shift_n1, _phase)) {
3060 return ValueOrder::NotAdjacent;
3061 }
3062
3063 int bits_per_store = memory_size * 8;
3064 if (base_n1 != base_n2 ||
3065 abs(shift_n1 - shift_n2) != bits_per_store ||
3066 shift_n1 % bits_per_store != 0) {
3067 // Values are not adjacent
3068 return ValueOrder::NotAdjacent;
3069 }
3070
3071 // Detect value order
3072 #ifdef VM_LITTLE_ENDIAN
3073 return shift_n1 < shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3074 : ValueOrder::Reverse; // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3075 #else
3076 return shift_n1 > shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3077 : ValueOrder::Reverse; // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3078 #endif
3079 }
3080
3081 bool MergePrimitiveStores::is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const {
3082 ValueOrder input_value_order = find_adjacent_input_value_order(n1, n2, memory_size);
3083
3084 switch (input_value_order) {
3085 case ValueOrder::NotAdjacent:
3086 return false;
3087 case ValueOrder::Reverse:
3088 if (memory_size != 1 ||
3089 !Matcher::match_rule_supported(Op_ReverseBytesS) ||
3090 !Matcher::match_rule_supported(Op_ReverseBytesI) ||
3091 !Matcher::match_rule_supported(Op_ReverseBytesL)) {
3092 // ReverseBytes are not supported by platform
3093 return false;
3094 }
3095 // fall-through.
3096 case ValueOrder::Const:
3097 case ValueOrder::Platform:
3098 if (_value_order == ValueOrder::Unknown) {
3099 // Initial state is Unknown, and we find a valid input value order
3100 return true;
3101 }
3102 // The value order can not be changed
3103 return _value_order == input_value_order;
3104 case ValueOrder::Unknown:
3105 default:
3106 ShouldNotReachHere();
3107 }
3108 return false;
3109 }
3110
3111 // Detect pattern: n = base_out >> shift_out
3112 bool MergePrimitiveStores::is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase) {
3113 assert(n != nullptr, "precondition");
3114
3115 int opc = n->Opcode();
3116 if (opc == Op_ConvL2I) {
3117 n = n->in(1);
3118 opc = n->Opcode();
3119 }
3120
3121 if ((opc == Op_RShiftI ||
3122 opc == Op_RShiftL ||
3123 opc == Op_URShiftI ||
3124 opc == Op_URShiftL) &&
3125 n->in(2)->is_ConI()) {
3126 base_out = n->in(1);
3127 shift_out = n->in(2)->get_int();
3128 // The shift must be positive:
3129 return shift_out >= 0;
3130 }
3131
3132 if (phase->type(n)->isa_int() != nullptr ||
3133 phase->type(n)->isa_long() != nullptr) {
3134 // (base >> 0)
3135 base_out = n;
3136 shift_out = 0;
3137 return true;
3138 }
3139 return false;
3140 }
3141
3142 // Check if there is nothing between the two stores, except optionally a RangeCheck leading to an uncommon trap.
3143 MergePrimitiveStores::CFGStatus MergePrimitiveStores::cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store) {
3144 assert(use_store->in(MemNode::Memory) == def_store, "use-def relationship");
3145
3146 Node* ctrl_use = use_store->in(MemNode::Control);
3147 Node* ctrl_def = def_store->in(MemNode::Control);
3148 if (ctrl_use == nullptr || ctrl_def == nullptr) {
3149 return CFGStatus::Failure;
3150 }
3151
3152 if (ctrl_use == ctrl_def) {
3153 // Same ctrl -> no RangeCheck in between.
3154 // Check: use_store must be the only use of def_store.
3155 if (def_store->outcnt() > 1) {
3156 return CFGStatus::Failure;
3157 }
3158 return CFGStatus::SuccessNoRangeCheck;
3159 }
3160
3161 // Different ctrl -> could have RangeCheck in between.
3162 // Check: 1. def_store only has these uses: use_store and MergeMem for uncommon trap, and
3163 // 2. ctrl separated by RangeCheck.
3164 if (def_store->outcnt() != 2) {
3165 return CFGStatus::Failure; // Cannot have exactly these uses: use_store and MergeMem for uncommon trap.
3166 }
3167 int use_store_out_idx = def_store->raw_out(0) == use_store ? 0 : 1;
3168 Node* merge_mem = def_store->raw_out(1 - use_store_out_idx)->isa_MergeMem();
3169 if (merge_mem == nullptr ||
3170 merge_mem->outcnt() != 1) {
3171 return CFGStatus::Failure; // Does not have MergeMem for uncommon trap.
3172 }
3173 if (!ctrl_use->is_IfProj() ||
3174 !ctrl_use->in(0)->is_RangeCheck() ||
3175 ctrl_use->in(0)->outcnt() != 2) {
3176 return CFGStatus::Failure; // Not RangeCheck.
3177 }
3178 IfProjNode* other_proj = ctrl_use->as_IfProj()->other_if_proj();
3179 Node* trap = other_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check);
3180 if (trap != merge_mem->unique_out() ||
3181 ctrl_use->in(0)->in(0) != ctrl_def) {
3182 return CFGStatus::Failure; // Not RangeCheck with merge_mem leading to uncommon trap.
3183 }
3184
3185 return CFGStatus::SuccessWithRangeCheck;
3186 }
3187
3188 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_use_store(const StoreNode* def_store) const {
3189 Status status_use = find_use_store(def_store);
3190 StoreNode* use_store = status_use.found_store();
3191 if (use_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3192 return Status::make_failure();
3193 }
3194 return status_use;
3195 }
3196
3197 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_def_store(const StoreNode* use_store) const {
3198 Status status_def = find_def_store(use_store);
3199 StoreNode* def_store = status_def.found_store();
3200 if (def_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3201 return Status::make_failure();
3202 }
3203 return status_def;
3204 }
3205
3206 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store(const StoreNode* def_store) const {
3207 Status status_use = find_use_store_unidirectional(def_store);
3208
3209 #ifdef ASSERT
3210 StoreNode* use_store = status_use.found_store();
3211 if (use_store != nullptr) {
3212 Status status_def = find_def_store_unidirectional(use_store);
3213 assert(status_def.found_store() == def_store &&
3214 status_def.found_range_check() == status_use.found_range_check(),
3215 "find_use_store and find_def_store must be symmetric");
3216 }
3217 #endif
3218
3219 return status_use;
3220 }
3221
3222 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store(const StoreNode* use_store) const {
3223 Status status_def = find_def_store_unidirectional(use_store);
3224
3225 #ifdef ASSERT
3226 StoreNode* def_store = status_def.found_store();
3227 if (def_store != nullptr) {
3228 Status status_use = find_use_store_unidirectional(def_store);
3229 assert(status_use.found_store() == use_store &&
3230 status_use.found_range_check() == status_def.found_range_check(),
3231 "find_use_store and find_def_store must be symmetric");
3232 }
3233 #endif
3234
3235 return status_def;
3236 }
3237
3238 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store_unidirectional(const StoreNode* def_store) const {
3239 assert(is_compatible_store(def_store), "precondition: must be compatible with _store");
3240
3241 for (DUIterator_Fast imax, i = def_store->fast_outs(imax); i < imax; i++) {
3242 StoreNode* use_store = def_store->fast_out(i)->isa_Store();
3243 if (is_compatible_store(use_store)) {
3244 return Status::make(use_store, cfg_status_for_pair(use_store, def_store));
3245 }
3246 }
3247
3248 return Status::make_failure();
3249 }
3250
3251 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store_unidirectional(const StoreNode* use_store) const {
3252 assert(is_compatible_store(use_store), "precondition: must be compatible with _store");
3253
3254 StoreNode* def_store = use_store->in(MemNode::Memory)->isa_Store();
3255 if (!is_compatible_store(def_store)) {
3256 return Status::make_failure();
3257 }
3258
3259 return Status::make(def_store, cfg_status_for_pair(use_store, def_store));
3260 }
3261
3262 void MergePrimitiveStores::collect_merge_list(Node_List& merge_list) const {
3263 // The merged store can be at most 8 bytes.
3264 const uint merge_list_max_size = 8 / _store->memory_size();
3265 assert(merge_list_max_size >= 2 &&
3266 merge_list_max_size <= 8 &&
3267 is_power_of_2(merge_list_max_size),
3268 "must be 2, 4 or 8");
3269
3270 // Traverse up the chain of adjacent def stores.
3271 StoreNode* current = _store;
3272 merge_list.push(current);
3273 while (current != nullptr && merge_list.size() < merge_list_max_size) {
3274 Status status = find_adjacent_def_store(current);
3275 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] find def: "); status.print_on(tty); })
3276
3277 current = status.found_store();
3278 if (current != nullptr) {
3279 merge_list.push(current);
3280
3281 // We can have at most one RangeCheck.
3282 if (status.found_range_check()) {
3283 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found RangeCheck, stop traversal."); })
3284 break;
3285 }
3286 }
3287 }
3288
3289 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found:"); merge_list.dump(); })
3290
3291 // Truncate the merge_list to a power of 2.
3292 const uint pow2size = round_down_power_of_2(merge_list.size());
3293 assert(pow2size >= 2, "must be merging at least 2 stores");
3294 while (merge_list.size() > pow2size) { merge_list.pop(); }
3295
3296 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] truncated:"); merge_list.dump(); })
3297 }
3298
3299 // Merge the input values of the smaller stores to a single larger input value.
3300 Node* MergePrimitiveStores::make_merged_input_value(const Node_List& merge_list) {
3301 int new_memory_size = _store->memory_size() * merge_list.size();
3302 Node* first = merge_list.at(merge_list.size()-1);
3303 Node* merged_input_value = nullptr;
3304 if (_store->in(MemNode::ValueIn)->Opcode() == Op_ConI) {
3305 assert(_value_order == ValueOrder::Const, "must match");
3306 // Pattern: [ConI, ConI, ...] -> new constant
3307 jlong con = 0;
3308 jlong bits_per_store = _store->memory_size() * 8;
3309 jlong mask = (((jlong)1) << bits_per_store) - 1;
3310 for (uint i = 0; i < merge_list.size(); i++) {
3311 jlong con_i = merge_list.at(i)->in(MemNode::ValueIn)->get_int();
3312 #ifdef VM_LITTLE_ENDIAN
3313 con = con << bits_per_store;
3314 con = con | (mask & con_i);
3315 #else // VM_LITTLE_ENDIAN
3316 con_i = (mask & con_i) << (i * bits_per_store);
3317 con = con | con_i;
3318 #endif // VM_LITTLE_ENDIAN
3319 }
3320 merged_input_value = _phase->longcon(con);
3321 } else {
3322 assert(_value_order == ValueOrder::Platform || _value_order == ValueOrder::Reverse, "must match");
3323 // Pattern: [base >> 24, base >> 16, base >> 8, base] -> base
3324 // | |
3325 // _store first
3326 //
3327 Node* hi = _store->in(MemNode::ValueIn);
3328 Node* lo = first->in(MemNode::ValueIn);
3329 #ifndef VM_LITTLE_ENDIAN
3330 // `_store` and `first` are swapped in the diagram above
3331 swap(hi, lo);
3332 #endif // !VM_LITTLE_ENDIAN
3333 if (_value_order == ValueOrder::Reverse) {
3334 swap(hi, lo);
3335 }
3336 Node const* hi_base;
3337 jint hi_shift;
3338 merged_input_value = lo;
3339 bool is_true = is_con_RShift(hi, hi_base, hi_shift, _phase);
3340 assert(is_true, "must detect con RShift");
3341 if (merged_input_value != hi_base && merged_input_value->Opcode() == Op_ConvL2I) {
3342 // look through
3343 merged_input_value = merged_input_value->in(1);
3344 }
3345 if (merged_input_value != hi_base) {
3346 // merged_input_value is not the base
3347 return nullptr;
3348 }
3349 }
3350
3351 if (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size <= 4) {
3352 // Example:
3353 //
3354 // long base = ...;
3355 // a[0] = (byte)(base >> 0);
3356 // a[1] = (byte)(base >> 8);
3357 //
3358 merged_input_value = _phase->transform(new ConvL2INode(merged_input_value));
3359 }
3360
3361 assert((_phase->type(merged_input_value)->isa_int() != nullptr && new_memory_size <= 4) ||
3362 (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size == 8),
3363 "merged_input_value is either int or long, and new_memory_size is small enough");
3364
3365 if (_value_order == ValueOrder::Reverse) {
3366 assert(_store->memory_size() == 1, "only implemented for bytes");
3367 if (new_memory_size == 8) {
3368 merged_input_value = _phase->transform(new ReverseBytesLNode(merged_input_value));
3369 } else if (new_memory_size == 4) {
3370 merged_input_value = _phase->transform(new ReverseBytesINode(merged_input_value));
3371 } else {
3372 assert(new_memory_size == 2, "sanity check");
3373 merged_input_value = _phase->transform(new ReverseBytesSNode(merged_input_value));
3374 }
3375 }
3376 return merged_input_value;
3377 }
3378
3379 // //
3380 // first_ctrl first_mem first_adr first_ctrl first_mem first_adr //
3381 // | | | | | | //
3382 // | | | | +---------------+ | //
3383 // | | | | | | | //
3384 // | | +---------+ | | +---------------+ //
3385 // | | | | | | | | //
3386 // +--------------+ | | v1 +------------------------------+ | | v1 //
3387 // | | | | | | | | | | | | //
3388 // RangeCheck first_store RangeCheck | | first_store //
3389 // | | | | | | | //
3390 // last_ctrl | +----> unc_trap last_ctrl | | +----> unc_trap //
3391 // | | ===> | | | //
3392 // +--------------+ | a2 v2 | | | //
3393 // | | | | | | | | //
3394 // | second_store | | | //
3395 // | | | | | [v1 v2 ... vn] //
3396 // ... ... | | | | //
3397 // | | | | | v //
3398 // +--------------+ | an vn +--------------+ | | merged_input_value //
3399 // | | | | | | | | //
3400 // last_store (= _store) merged_store //
3401 // //
3402 StoreNode* MergePrimitiveStores::make_merged_store(const Node_List& merge_list, Node* merged_input_value) {
3403 Node* first_store = merge_list.at(merge_list.size()-1);
3404 Node* last_ctrl = _store->in(MemNode::Control); // after (optional) RangeCheck
3405 Node* first_mem = first_store->in(MemNode::Memory);
3406 Node* first_adr = first_store->in(MemNode::Address);
3407
3408 const TypePtr* new_adr_type = _store->adr_type();
3409
3410 int new_memory_size = _store->memory_size() * merge_list.size();
3411 BasicType bt = T_ILLEGAL;
3412 switch (new_memory_size) {
3413 case 2: bt = T_SHORT; break;
3414 case 4: bt = T_INT; break;
3415 case 8: bt = T_LONG; break;
3416 }
3417
3418 StoreNode* merged_store = StoreNode::make(*_phase, last_ctrl, first_mem, first_adr,
3419 new_adr_type, merged_input_value, bt, MemNode::unordered);
3420
3421 // Marking the store mismatched is sufficient to prevent reordering, since array stores
3422 // are all on the same slice. Hence, we need no barriers.
3423 merged_store->set_mismatched_access();
3424
3425 // Constants above may now also be be packed -> put candidate on worklist
3426 _phase->is_IterGVN()->_worklist.push(first_mem);
3427
3428 return merged_store;
3429 }
3430
3431 #ifndef PRODUCT
3432 void MergePrimitiveStores::trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const {
3433 stringStream ss;
3434 ss.print_cr("[TraceMergeStores]: Replace");
3435 for (int i = (int)merge_list.size() - 1; i >= 0; i--) {
3436 merge_list.at(i)->dump("\n", false, &ss);
3437 }
3438 ss.print_cr("[TraceMergeStores]: with");
3439 merged_input_value->dump("\n", false, &ss);
3440 merged_store->dump("\n", false, &ss);
3441 tty->print("%s", ss.as_string());
3442 }
3443 #endif
3444
3445 //------------------------------Ideal------------------------------------------
3446 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
3447 // When a store immediately follows a relevant allocation/initialization,
3448 // try to capture it into the initialization, or hoist it above.
3449 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3450 Node* p = MemNode::Ideal_common(phase, can_reshape);
3451 if (p) return (p == NodeSentinel) ? nullptr : p;
3452
3453 Node* mem = in(MemNode::Memory);
3454 Node* address = in(MemNode::Address);
3455 Node* value = in(MemNode::ValueIn);
3456 // Back-to-back stores to same address? Fold em up. Generally
3457 // unsafe if I have intervening uses.
3458 {
3459 Node* st = mem;
3460 // If Store 'st' has more than one use, we cannot fold 'st' away.
3461 // For example, 'st' might be the final state at a conditional
3462 // return. Or, 'st' might be used by some node which is live at
3463 // the same time 'st' is live, which might be unschedulable. So,
3464 // require exactly ONE user until such time as we clone 'mem' for
3465 // each of 'mem's uses (thus making the exactly-1-user-rule hold
3466 // true).
3467 while (st->is_Store() && st->outcnt() == 1) {
3468 // Looking at a dead closed cycle of memory?
3469 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
3470 assert(Opcode() == st->Opcode() ||
3471 st->Opcode() == Op_StoreVector ||
3472 Opcode() == Op_StoreVector ||
3473 st->Opcode() == Op_StoreVectorScatter ||
3474 Opcode() == Op_StoreVectorScatter ||
3475 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
3476 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
3477 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
3478 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
3479 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
3480
3481 if (st->in(MemNode::Address)->eqv_uncast(address) &&
3482 st->as_Store()->memory_size() <= this->memory_size()) {
3483 Node* use = st->raw_out(0);
3484 if (phase->is_IterGVN()) {
3485 phase->is_IterGVN()->rehash_node_delayed(use);
3486 }
3487 // It's OK to do this in the parser, since DU info is always accurate,
3488 // and the parser always refers to nodes via SafePointNode maps.
3489 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
3490 return this;
3491 }
3492 st = st->in(MemNode::Memory);
3493 }
3494 }
3495
3496
3497 // Capture an unaliased, unconditional, simple store into an initializer.
3498 // Or, if it is independent of the allocation, hoist it above the allocation.
3499 if (ReduceFieldZeroing && /*can_reshape &&*/
3500 mem->is_Proj() && mem->in(0)->is_Initialize()) {
3501 InitializeNode* init = mem->in(0)->as_Initialize();
3502 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
3503 if (offset > 0) {
3504 Node* moved = init->capture_store(this, offset, phase, can_reshape);
3505 // If the InitializeNode captured me, it made a raw copy of me,
3506 // and I need to disappear.
3507 if (moved != nullptr) {
3508 // %%% hack to ensure that Ideal returns a new node:
3509 mem = MergeMemNode::make(mem);
3510 return mem; // fold me away
3511 }
3512 }
3513 }
3514
3515 // Fold reinterpret cast into memory operation:
3516 // StoreX mem (MoveY2X v) => StoreY mem v
3517 if (value->is_Move()) {
3518 const Type* vt = value->in(1)->bottom_type();
3519 if (has_reinterpret_variant(vt)) {
3520 if (phase->C->post_loop_opts_phase()) {
3521 return convert_to_reinterpret_store(*phase, value->in(1), vt);
3522 } else {
3523 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
3524 }
3525 }
3526 }
3527
3528 if (MergeStores && UseUnalignedAccesses) {
3529 if (phase->C->merge_stores_phase()) {
3530 MergePrimitiveStores merge(phase, this);
3531 Node* progress = merge.run();
3532 if (progress != nullptr) { return progress; }
3533 } else {
3534 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
3535 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
3536 // the stores, and we merge the wrong sequence of stores.
3537 // Example:
3538 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
3539 // Apply MergeStores:
3540 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
3541 // Remove more RangeChecks:
3542 // StoreI [ StoreL ] StoreI
3543 // But now it would have been better to do this instead:
3544 // [ StoreL ] [ StoreL ]
3545 phase->C->record_for_merge_stores_igvn(this);
3546 }
3547 }
3548
3549 return nullptr; // No further progress
3550 }
3551
3552 //------------------------------Value-----------------------------------------
3553 const Type* StoreNode::Value(PhaseGVN* phase) const {
3554 // Either input is TOP ==> the result is TOP
3555 const Type *t1 = phase->type( in(MemNode::Memory) );
3556 if( t1 == Type::TOP ) return Type::TOP;
3557 const Type *t2 = phase->type( in(MemNode::Address) );
3558 if( t2 == Type::TOP ) return Type::TOP;
3559 const Type *t3 = phase->type( in(MemNode::ValueIn) );
3560 if( t3 == Type::TOP ) return Type::TOP;
3561 return Type::MEMORY;
3562 }
3563
3564 //------------------------------Identity---------------------------------------
3565 // Remove redundant stores:
3566 // Store(m, p, Load(m, p)) changes to m.
3567 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
3568 Node* StoreNode::Identity(PhaseGVN* phase) {
3569 Node* mem = in(MemNode::Memory);
3570 Node* adr = in(MemNode::Address);
3571 Node* val = in(MemNode::ValueIn);
3572
3573 Node* result = this;
3574
3575 // Load then Store? Then the Store is useless
3576 if (val->is_Load() &&
3577 val->in(MemNode::Address)->eqv_uncast(adr) &&
3578 val->in(MemNode::Memory )->eqv_uncast(mem) &&
3579 val->as_Load()->store_Opcode() == Opcode()) {
3580 if (!is_StoreVector()) {
3581 result = mem;
3582 } else if (Opcode() == Op_StoreVector && val->Opcode() == Op_LoadVector &&
3583 as_StoreVector()->vect_type() == val->as_LoadVector()->vect_type()) {
3584 // Ensure both are not masked accesses or gathers/scatters and vector types are the same
3585 result = mem;
3586 }
3587 }
3588
3589 // Two stores in a row of the same value?
3590 if (result == this &&
3591 mem->is_Store() &&
3592 mem->in(MemNode::Address)->eqv_uncast(adr) &&
3593 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
3594 mem->Opcode() == Opcode()) {
3595 if (!is_StoreVector()) {
3596 result = mem;
3597 } else {
3598 const StoreVectorNode* store_vector = as_StoreVector();
3599 const StoreVectorNode* mem_vector = mem->as_StoreVector();
3600 const Node* store_indices = store_vector->indices();
3601 const Node* mem_indices = mem_vector->indices();
3602 const Node* store_mask = store_vector->mask();
3603 const Node* mem_mask = mem_vector->mask();
3604 // Ensure types, indices, and masks match
3605 if (store_vector->vect_type() == mem_vector->vect_type() &&
3606 ((store_indices == nullptr) == (mem_indices == nullptr) &&
3607 (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) &&
3608 ((store_mask == nullptr) == (mem_mask == nullptr) &&
3609 (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) {
3610 result = mem;
3611 }
3612 }
3613 }
3614
3615 // Store of zero anywhere into a freshly-allocated object?
3616 // Then the store is useless.
3617 // (It must already have been captured by the InitializeNode.)
3618 if (result == this &&
3619 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
3620 // a newly allocated object is already all-zeroes everywhere
3621 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
3622 result = mem;
3623 }
3624
3625 if (result == this) {
3626 // the store may also apply to zero-bits in an earlier object
3627 Node* prev_mem = find_previous_store(phase);
3628 // Steps (a), (b): Walk past independent stores to find an exact match.
3629 if (prev_mem != nullptr) {
3630 if (prev_mem->is_top()) {
3631 // find_previous_store returns top when the access is dead
3632 return prev_mem;
3633 }
3634 Node* prev_val = can_see_stored_value(prev_mem, phase);
3635 if (prev_val != nullptr && prev_val == val) {
3636 // prev_val and val might differ by a cast; it would be good
3637 // to keep the more informative of the two.
3638 result = mem;
3639 }
3640 }
3641 }
3642 }
3643
3644 PhaseIterGVN* igvn = phase->is_IterGVN();
3645 if (result != this && igvn != nullptr) {
3646 MemBarNode* trailing = trailing_membar();
3647 if (trailing != nullptr) {
3648 #ifdef ASSERT
3649 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
3650 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects");
3651 #endif
3652 trailing->remove(igvn);
3653 }
3654 }
3655
3656 return result;
3657 }
3658
3659 //------------------------------match_edge-------------------------------------
3660 // Do we Match on this edge index or not? Match only memory & value
3661 uint StoreNode::match_edge(uint idx) const {
3662 return idx == MemNode::Address || idx == MemNode::ValueIn;
3663 }
3664
3665 //------------------------------cmp--------------------------------------------
3666 // Do not common stores up together. They generally have to be split
3667 // back up anyways, so do not bother.
3668 bool StoreNode::cmp( const Node &n ) const {
3669 return (&n == this); // Always fail except on self
3670 }
3671
3672 //------------------------------Ideal_masked_input-----------------------------
3673 // Check for a useless mask before a partial-word store
3674 // (StoreB ... (AndI valIn conIa) )
3675 // If (conIa & mask == mask) this simplifies to
3676 // (StoreB ... (valIn) )
3677 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
3678 Node *val = in(MemNode::ValueIn);
3679 if( val->Opcode() == Op_AndI ) {
3680 const TypeInt *t = phase->type( val->in(2) )->isa_int();
3681 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
3682 set_req_X(MemNode::ValueIn, val->in(1), phase);
3683 return this;
3684 }
3685 }
3686 return nullptr;
3687 }
3688
3689
3690 //------------------------------Ideal_sign_extended_input----------------------
3691 // Check for useless sign-extension before a partial-word store
3692 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
3693 // If (conIL == conIR && conIR <= num_rejected_bits) this simplifies to
3694 // (StoreB ... (v))
3695 // If (conIL > conIR) under some conditions, it can be simplified into
3696 // (StoreB ... (LShiftI _ v (conIL - conIR)))
3697 // This case happens when the value of the store was itself a left shift, that
3698 // gets merged into the inner left shift of the sign-extension. For instance,
3699 // if we have
3700 // array_of_shorts[0] = (short)(v << 2)
3701 // We get a structure such as:
3702 // (StoreB ... (RShiftI _ (LShiftI _ (LShiftI _ v 2) 16) 16))
3703 // that is simplified into
3704 // (StoreB ... (RShiftI _ (LShiftI _ v 18) 16)).
3705 // It is thus useful to handle cases where conIL > conIR. But this simplification
3706 // does not always hold. Let's see in which cases it's valid.
3707 //
3708 // Let's assume we have the following 32 bits integer v:
3709 // +----------------------------------+
3710 // | v[0..31] |
3711 // +----------------------------------+
3712 // 31 0
3713 // that will be stuffed in 8 bits byte after a shift left and a shift right of
3714 // potentially different magnitudes.
3715 // We denote num_rejected_bits the number of bits of the discarded part. In this
3716 // case, num_rejected_bits == 24.
3717 //
3718 // Statement (proved further below in case analysis):
3719 // Given:
3720 // - 0 <= conIL < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
3721 // - 0 <= conIR < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
3722 // - conIL >= conIR
3723 // - num_rejected_bits >= conIR
3724 // Then this form:
3725 // (RShiftI _ (LShiftI _ v conIL) conIR)
3726 // can be replaced with this form:
3727 // (LShiftI _ v (conIL-conIR))
3728 //
3729 // Note: We only have to show that the non-rejected lowest bits (8 bits for byte)
3730 // have to be correct, as the higher bits are rejected / truncated by the store.
3731 //
3732 // The hypotheses
3733 // 0 <= conIL < BitsPerJavaInteger
3734 // 0 <= conIR < BitsPerJavaInteger
3735 // are ensured by maskShiftAmount (called from ::Ideal of shift nodes). Indeed,
3736 // (v << 31) << 2 must be simplified into 0, not into v << 33 (which is equivalent
3737 // to v << 1).
3738 //
3739 //
3740 // If you don't like case analysis, jump after the conclusion.
3741 // ### Case 1 : conIL == conIR
3742 // ###### Case 1.1: conIL == conIR == num_rejected_bits
3743 // If we do the shift left then right by 24 bits, we get:
3744 // after: v << 24
3745 // +---------+------------------------+
3746 // | v[0..7] | 0 |
3747 // +---------+------------------------+
3748 // 31 24 23 0
3749 // after: (v << 24) >> 24
3750 // +------------------------+---------+
3751 // | sign bit | v[0..7] |
3752 // +------------------------+---------+
3753 // 31 8 7 0
3754 // The non-rejected bits (bits kept by the store, that is the 8 lower bits of the
3755 // result) are the same before and after, so, indeed, simplifying is correct.
3756
3757 // ###### Case 1.2: conIL == conIR < num_rejected_bits
3758 // If we do the shift left then right by 22 bits, we get:
3759 // after: v << 22
3760 // +---------+------------------------+
3761 // | v[0..9] | 0 |
3762 // +---------+------------------------+
3763 // 31 22 21 0
3764 // after: (v << 22) >> 22
3765 // +------------------------+---------+
3766 // | sign bit | v[0..9] |
3767 // +------------------------+---------+
3768 // 31 10 9 0
3769 // The non-rejected bits are the 8 lower bits of v. The bits 8 and 9 of v are still
3770 // present in (v << 22) >> 22 but will be dropped by the store. The simplification is
3771 // still correct.
3772
3773 // ###### But! Case 1.3: conIL == conIR > num_rejected_bits
3774 // If we do the shift left then right by 26 bits, we get:
3775 // after: v << 26
3776 // +---------+------------------------+
3777 // | v[0..5] | 0 |
3778 // +---------+------------------------+
3779 // 31 26 25 0
3780 // after: (v << 26) >> 26
3781 // +------------------------+---------+
3782 // | sign bit | v[0..5] |
3783 // +------------------------+---------+
3784 // 31 6 5 0
3785 // The non-rejected bits are made of
3786 // - 0-5 => the bits 0 to 5 of v
3787 // - 6-7 => the sign bit of v[0..5] (that is v[5])
3788 // Simplifying this as v is not correct.
3789 // The condition conIR <= num_rejected_bits is indeed necessary in Case 1
3790 //
3791 // ### Case 2: conIL > conIR
3792 // ###### Case 2.1: num_rejected_bits == conIR
3793 // We take conIL == 26 for this example.
3794 // after: v << 26
3795 // +---------+------------------------+
3796 // | v[0..5] | 0 |
3797 // +---------+------------------------+
3798 // 31 26 25 0
3799 // after: (v << 26) >> 24
3800 // +------------------+---------+-----+
3801 // | sign bit | v[0..5] | 0 |
3802 // +------------------+---------+-----+
3803 // 31 8 7 2 1 0
3804 // The non-rejected bits are the 8 lower ones of (v << conIL - conIR).
3805 // The bits 6 and 7 of v have been thrown away after the shift left.
3806 // The simplification is still correct.
3807 //
3808 // ###### Case 2.2: num_rejected_bits > conIR.
3809 // Let's say conIL == 26 and conIR == 22.
3810 // after: v << 26
3811 // +---------+------------------------+
3812 // | v[0..5] | 0 |
3813 // +---------+------------------------+
3814 // 31 26 25 0
3815 // after: (v << 26) >> 22
3816 // +------------------+---------+-----+
3817 // | sign bit | v[0..5] | 0 |
3818 // +------------------+---------+-----+
3819 // 31 10 9 4 3 0
3820 // The bits non-rejected by the store are exactly the 8 lower ones of (v << (conIL - conIR)):
3821 // - 0-3 => 0
3822 // - 4-7 => bits 0 to 3 of v
3823 // The simplification is still correct.
3824 // The bits 4 and 5 of v are still present in (v << (conIL - conIR)) but they don't
3825 // matter as they are not in the 8 lower bits: they will be cut out by the store.
3826 //
3827 // ###### But! Case 2.3: num_rejected_bits < conIR.
3828 // Let's see that this case is not as easy to simplify.
3829 // Let's say conIL == 28 and conIR == 26.
3830 // after: v << 28
3831 // +---------+------------------------+
3832 // | v[0..3] | 0 |
3833 // +---------+------------------------+
3834 // 31 28 27 0
3835 // after: (v << 28) >> 26
3836 // +------------------+---------+-----+
3837 // | sign bit | v[0..3] | 0 |
3838 // +------------------+---------+-----+
3839 // 31 6 5 2 1 0
3840 // The non-rejected bits are made of
3841 // - 0-1 => 0
3842 // - 2-5 => the bits 0 to 3 of v
3843 // - 6-7 => the sign bit of v[0..3] (that is v[3])
3844 // Simplifying this as (v << 2) is not correct.
3845 // The condition conIR <= num_rejected_bits is indeed necessary in Case 2.
3846 //
3847 // ### Conclusion:
3848 // Our hypotheses are indeed sufficient:
3849 // - 0 <= conIL < BitsPerJavaInteger
3850 // - 0 <= conIR < BitsPerJavaInteger
3851 // - conIL >= conIR
3852 // - num_rejected_bits >= conIR
3853 //
3854 // ### A rationale without case analysis:
3855 // After the shift left, conIL upper bits of v are discarded and conIL lower bit
3856 // zeroes are added. After the shift right, conIR lower bits of the previous result
3857 // are discarded. If conIL >= conIR, we discard only the zeroes we made up during
3858 // the shift left, but if conIL < conIR, then we discard also lower bits of v. But
3859 // the point of the simplification is to get an expression of the form
3860 // (v << (conIL - conIR)). This expression discard only higher bits of v, thus the
3861 // simplification is not correct if conIL < conIR.
3862 //
3863 // Moreover, after the shift right, the higher bit of (v << conIL) is repeated on the
3864 // conIR higher bits of ((v << conIL) >> conIR), it's the sign-extension. If
3865 // conIR > num_rejected_bits, then at least one artificial copy of this sign bit will
3866 // be in the window of the store. Thus ((v << conIL) >> conIR) is not equivalent to
3867 // (v << (conIL-conIR)) if conIR > num_rejected_bits.
3868 //
3869 // We do not treat the case conIL < conIR here since the point of this function is
3870 // to skip sign-extensions (that is conIL == conIR == num_rejected_bits). The need
3871 // of treating conIL > conIR comes from the cases where the sign-extended value is
3872 // also left-shift expression. Computing the sign-extension of a right-shift expression
3873 // doesn't yield a situation such as
3874 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
3875 // where conIL < conIR.
3876 Node* StoreNode::Ideal_sign_extended_input(PhaseGVN* phase, int num_rejected_bits) {
3877 Node* shr = in(MemNode::ValueIn);
3878 if (shr->Opcode() == Op_RShiftI) {
3879 const TypeInt* conIR = phase->type(shr->in(2))->isa_int();
3880 if (conIR != nullptr && conIR->is_con() && conIR->get_con() >= 0 && conIR->get_con() < BitsPerJavaInteger && conIR->get_con() <= num_rejected_bits) {
3881 Node* shl = shr->in(1);
3882 if (shl->Opcode() == Op_LShiftI) {
3883 const TypeInt* conIL = phase->type(shl->in(2))->isa_int();
3884 if (conIL != nullptr && conIL->is_con() && conIL->get_con() >= 0 && conIL->get_con() < BitsPerJavaInteger) {
3885 if (conIL->get_con() == conIR->get_con()) {
3886 set_req_X(MemNode::ValueIn, shl->in(1), phase);
3887 return this;
3888 }
3889 if (conIL->get_con() > conIR->get_con()) {
3890 Node* new_shl = phase->transform(new LShiftINode(shl->in(1), phase->intcon(conIL->get_con() - conIR->get_con())));
3891 set_req_X(MemNode::ValueIn, new_shl, phase);
3892 return this;
3893 }
3894 }
3895 }
3896 }
3897 }
3898 return nullptr;
3899 }
3900
3901 //------------------------------value_never_loaded-----------------------------------
3902 // Determine whether there are any possible loads of the value stored.
3903 // For simplicity, we actually check if there are any loads from the
3904 // address stored to, not just for loads of the value stored by this node.
3905 //
3906 bool StoreNode::value_never_loaded(PhaseValues* phase) const {
3907 Node *adr = in(Address);
3908 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
3909 if (adr_oop == nullptr)
3910 return false;
3911 if (!adr_oop->is_known_instance_field())
3912 return false; // if not a distinct instance, there may be aliases of the address
3913 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
3914 Node *use = adr->fast_out(i);
3915 if (use->is_Load() || use->is_LoadStore()) {
3916 return false;
3917 }
3918 }
3919 return true;
3920 }
3921
3922 MemBarNode* StoreNode::trailing_membar() const {
3923 if (is_release()) {
3924 MemBarNode* trailing_mb = nullptr;
3925 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3926 Node* u = fast_out(i);
3927 if (u->is_MemBar()) {
3928 if (u->as_MemBar()->trailing_store()) {
3929 assert(u->Opcode() == Op_MemBarVolatile, "");
3930 assert(trailing_mb == nullptr, "only one");
3931 trailing_mb = u->as_MemBar();
3932 #ifdef ASSERT
3933 Node* leading = u->as_MemBar()->leading_membar();
3934 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3935 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
3936 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
3937 #endif
3938 } else {
3939 assert(u->as_MemBar()->standalone(), "");
3940 }
3941 }
3942 }
3943 return trailing_mb;
3944 }
3945 return nullptr;
3946 }
3947
3948
3949 //=============================================================================
3950 //------------------------------Ideal------------------------------------------
3951 // If the store is from an AND mask that leaves the low bits untouched, then
3952 // we can skip the AND operation. If the store is from a sign-extension
3953 // (a left shift, then right shift) we can skip both.
3954 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
3955 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
3956 if( progress != nullptr ) return progress;
3957
3958 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
3959 if( progress != nullptr ) return progress;
3960
3961 // Finally check the default case
3962 return StoreNode::Ideal(phase, can_reshape);
3963 }
3964
3965 //=============================================================================
3966 //------------------------------Ideal------------------------------------------
3967 // If the store is from an AND mask that leaves the low bits untouched, then
3968 // we can skip the AND operation
3969 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
3970 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
3971 if( progress != nullptr ) return progress;
3972
3973 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
3974 if( progress != nullptr ) return progress;
3975
3976 // Finally check the default case
3977 return StoreNode::Ideal(phase, can_reshape);
3978 }
3979
3980 //=============================================================================
3981 //----------------------------------SCMemProjNode------------------------------
3982 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
3983 {
3984 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
3985 return Type::TOP;
3986 }
3987 return bottom_type();
3988 }
3989
3990 //=============================================================================
3991 //----------------------------------LoadStoreNode------------------------------
3992 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
3993 : Node(required),
3994 _type(rt),
3995 _barrier_data(0)
3996 {
3997 init_req(MemNode::Control, c );
3998 init_req(MemNode::Memory , mem);
3999 init_req(MemNode::Address, adr);
4000 init_req(MemNode::ValueIn, val);
4001 init_class_id(Class_LoadStore);
4002 DEBUG_ONLY(_adr_type = at; adr_type();)
4003 }
4004
4005 //------------------------------Value-----------------------------------------
4006 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
4007 // Either input is TOP ==> the result is TOP
4008 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
4009 return Type::TOP;
4010 }
4011 const Type* t = phase->type(in(MemNode::Memory));
4012 if (t == Type::TOP) {
4013 return Type::TOP;
4014 }
4015 t = phase->type(in(MemNode::Address));
4016 if (t == Type::TOP) {
4017 return Type::TOP;
4018 }
4019 t = phase->type(in(MemNode::ValueIn));
4020 if (t == Type::TOP) {
4021 return Type::TOP;
4022 }
4023 return bottom_type();
4024 }
4025
4026 const TypePtr* LoadStoreNode::adr_type() const {
4027 const TypePtr* cross_check = DEBUG_ONLY(_adr_type) NOT_DEBUG(nullptr);
4028 return MemNode::calculate_adr_type(in(MemNode::Address)->bottom_type(), cross_check);
4029 }
4030
4031 uint LoadStoreNode::ideal_reg() const {
4032 return _type->ideal_reg();
4033 }
4034
4035 // This method conservatively checks if the result of a LoadStoreNode is
4036 // used, that is, if it returns true, then it is definitely the case that
4037 // the result of the node is not needed.
4038 // For example, GetAndAdd can be matched into a lock_add instead of a
4039 // lock_xadd if the result of LoadStoreNode::result_not_used() is true
4040 bool LoadStoreNode::result_not_used() const {
4041 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4042 Node *x = fast_out(i);
4043 if (x->Opcode() == Op_SCMemProj) {
4044 continue;
4045 }
4046 if (x->bottom_type() == TypeTuple::MEMBAR &&
4047 !x->is_Call() &&
4048 x->Opcode() != Op_Blackhole) {
4049 continue;
4050 }
4051 return false;
4052 }
4053 return true;
4054 }
4055
4056 MemBarNode* LoadStoreNode::trailing_membar() const {
4057 MemBarNode* trailing = nullptr;
4058 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4059 Node* u = fast_out(i);
4060 if (u->is_MemBar()) {
4061 if (u->as_MemBar()->trailing_load_store()) {
4062 assert(u->Opcode() == Op_MemBarAcquire, "");
4063 assert(trailing == nullptr, "only one");
4064 trailing = u->as_MemBar();
4065 #ifdef ASSERT
4066 Node* leading = trailing->leading_membar();
4067 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
4068 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
4069 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
4070 #endif
4071 } else {
4072 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
4073 }
4074 }
4075 }
4076
4077 return trailing;
4078 }
4079
4080 uint LoadStoreNode::size_of() const { return sizeof(*this); }
4081
4082 //=============================================================================
4083 //----------------------------------LoadStoreConditionalNode--------------------
4084 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) {
4085 init_req(ExpectedIn, ex );
4086 }
4087
4088 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
4089 // Either input is TOP ==> the result is TOP
4090 const Type* t = phase->type(in(ExpectedIn));
4091 if (t == Type::TOP) {
4092 return Type::TOP;
4093 }
4094 return LoadStoreNode::Value(phase);
4095 }
4096
4097 //=============================================================================
4098 //-------------------------------adr_type--------------------------------------
4099 const TypePtr* ClearArrayNode::adr_type() const {
4100 Node *adr = in(3);
4101 if (adr == nullptr) return nullptr; // node is dead
4102 return MemNode::calculate_adr_type(adr->bottom_type());
4103 }
4104
4105 //------------------------------match_edge-------------------------------------
4106 // Do we Match on this edge index or not? Do not match memory
4107 uint ClearArrayNode::match_edge(uint idx) const {
4108 return idx > 1;
4109 }
4110
4111 //------------------------------Identity---------------------------------------
4112 // Clearing a zero length array does nothing
4113 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
4114 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
4115 }
4116
4117 //------------------------------Idealize---------------------------------------
4118 // Clearing a short array is faster with stores
4119 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4120 // Already know this is a large node, do not try to ideal it
4121 if (_is_large) return nullptr;
4122
4123 const int unit = BytesPerLong;
4124 const TypeX* t = phase->type(in(2))->isa_intptr_t();
4125 if (!t) return nullptr;
4126 if (!t->is_con()) return nullptr;
4127 intptr_t raw_count = t->get_con();
4128 intptr_t size = raw_count;
4129 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
4130 // Clearing nothing uses the Identity call.
4131 // Negative clears are possible on dead ClearArrays
4132 // (see jck test stmt114.stmt11402.val).
4133 if (size <= 0 || size % unit != 0) return nullptr;
4134 intptr_t count = size / unit;
4135 // Length too long; communicate this to matchers and assemblers.
4136 // Assemblers are responsible to produce fast hardware clears for it.
4137 if (size > InitArrayShortSize) {
4138 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
4139 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
4140 return nullptr;
4141 }
4142 if (!IdealizeClearArrayNode) return nullptr;
4143 Node *mem = in(1);
4144 if( phase->type(mem)==Type::TOP ) return nullptr;
4145 Node *adr = in(3);
4146 const Type* at = phase->type(adr);
4147 if( at==Type::TOP ) return nullptr;
4148 const TypePtr* atp = at->isa_ptr();
4149 // adjust atp to be the correct array element address type
4150 if (atp == nullptr) atp = TypePtr::BOTTOM;
4151 else atp = atp->add_offset(Type::OffsetBot);
4152 // Get base for derived pointer purposes
4153 if( adr->Opcode() != Op_AddP ) Unimplemented();
4154 Node *base = adr->in(1);
4155
4156 Node *zero = phase->makecon(TypeLong::ZERO);
4157 Node *off = phase->MakeConX(BytesPerLong);
4158 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
4159 count--;
4160 while( count-- ) {
4161 mem = phase->transform(mem);
4162 adr = phase->transform(new AddPNode(base,adr,off));
4163 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
4164 }
4165 return mem;
4166 }
4167
4168 //----------------------------step_through----------------------------------
4169 // Return allocation input memory edge if it is different instance
4170 // or itself if it is the one we are looking for.
4171 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
4172 Node* n = *np;
4173 assert(n->is_ClearArray(), "sanity");
4174 intptr_t offset;
4175 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
4176 // This method is called only before Allocate nodes are expanded
4177 // during macro nodes expansion. Before that ClearArray nodes are
4178 // only generated in PhaseMacroExpand::generate_arraycopy() (before
4179 // Allocate nodes are expanded) which follows allocations.
4180 assert(alloc != nullptr, "should have allocation");
4181 if (alloc->_idx == instance_id) {
4182 // Can not bypass initialization of the instance we are looking for.
4183 return false;
4184 }
4185 // Otherwise skip it.
4186 InitializeNode* init = alloc->initialization();
4187 if (init != nullptr)
4188 *np = init->in(TypeFunc::Memory);
4189 else
4190 *np = alloc->in(TypeFunc::Memory);
4191 return true;
4192 }
4193
4194 Node* ClearArrayNode::make_address(Node* dest, Node* offset, bool raw_base, PhaseGVN* phase) {
4195 Node* base = dest;
4196 if (raw_base) {
4197 // May be called as part of the initialization of a just allocated object
4198 base = phase->C->top();
4199 }
4200 return phase->transform(new AddPNode(base, dest, offset));
4201 }
4202
4203 //----------------------------clear_memory-------------------------------------
4204 // Generate code to initialize object storage to zero.
4205 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4206 intptr_t start_offset,
4207 Node* end_offset,
4208 bool raw_base,
4209 PhaseGVN* phase) {
4210 intptr_t offset = start_offset;
4211
4212 int unit = BytesPerLong;
4213 if ((offset % unit) != 0) {
4214 Node* adr = make_address(dest, phase->MakeConX(offset), raw_base, phase);
4215 const TypePtr* atp = TypeRawPtr::BOTTOM;
4216 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4217 mem = phase->transform(mem);
4218 offset += BytesPerInt;
4219 }
4220 assert((offset % unit) == 0, "");
4221
4222 // Initialize the remaining stuff, if any, with a ClearArray.
4223 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, raw_base, phase);
4224 }
4225
4226 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4227 Node* start_offset,
4228 Node* end_offset,
4229 bool raw_base,
4230 PhaseGVN* phase) {
4231 if (start_offset == end_offset) {
4232 // nothing to do
4233 return mem;
4234 }
4235
4236 int unit = BytesPerLong;
4237 Node* zbase = start_offset;
4238 Node* zend = end_offset;
4239
4240 // Scale to the unit required by the CPU:
4241 if (!Matcher::init_array_count_is_in_bytes) {
4242 Node* shift = phase->intcon(exact_log2(unit));
4243 zbase = phase->transform(new URShiftXNode(zbase, shift) );
4244 zend = phase->transform(new URShiftXNode(zend, shift) );
4245 }
4246
4247 // Bulk clear double-words
4248 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
4249 Node* adr = make_address(dest, start_offset, raw_base, phase);
4250 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
4251 return phase->transform(mem);
4252 }
4253
4254 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4255 intptr_t start_offset,
4256 intptr_t end_offset,
4257 bool raw_base,
4258 PhaseGVN* phase) {
4259 if (start_offset == end_offset) {
4260 // nothing to do
4261 return mem;
4262 }
4263
4264 assert((end_offset % BytesPerInt) == 0, "odd end offset");
4265 intptr_t done_offset = end_offset;
4266 if ((done_offset % BytesPerLong) != 0) {
4267 done_offset -= BytesPerInt;
4268 }
4269 if (done_offset > start_offset) {
4270 mem = clear_memory(ctl, mem, dest,
4271 start_offset, phase->MakeConX(done_offset), raw_base, phase);
4272 }
4273 if (done_offset < end_offset) { // emit the final 32-bit store
4274 Node* adr = make_address(dest, phase->MakeConX(done_offset), raw_base, phase);
4275 const TypePtr* atp = TypeRawPtr::BOTTOM;
4276 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4277 mem = phase->transform(mem);
4278 done_offset += BytesPerInt;
4279 }
4280 assert(done_offset == end_offset, "");
4281 return mem;
4282 }
4283
4284 //=============================================================================
4285 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
4286 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
4287 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
4288 #ifdef ASSERT
4289 , _pair_idx(0)
4290 #endif
4291 {
4292 init_class_id(Class_MemBar);
4293 Node* top = C->top();
4294 init_req(TypeFunc::I_O,top);
4295 init_req(TypeFunc::FramePtr,top);
4296 init_req(TypeFunc::ReturnAdr,top);
4297 if (precedent != nullptr)
4298 init_req(TypeFunc::Parms, precedent);
4299 }
4300
4301 //------------------------------cmp--------------------------------------------
4302 uint MemBarNode::hash() const { return NO_HASH; }
4303 bool MemBarNode::cmp( const Node &n ) const {
4304 return (&n == this); // Always fail except on self
4305 }
4306
4307 //------------------------------make-------------------------------------------
4308 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
4309 switch (opcode) {
4310 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
4311 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
4312 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
4313 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
4314 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
4315 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn);
4316 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
4317 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
4318 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
4319 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
4320 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
4321 case Op_Initialize: return new InitializeNode(C, atp, pn);
4322 default: ShouldNotReachHere(); return nullptr;
4323 }
4324 }
4325
4326 void MemBarNode::remove(PhaseIterGVN *igvn) {
4327 assert(outcnt() > 0 && (outcnt() <= 2 || Opcode() == Op_Initialize), "Only one or two out edges allowed");
4328 if (trailing_store() || trailing_load_store()) {
4329 MemBarNode* leading = leading_membar();
4330 if (leading != nullptr) {
4331 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
4332 leading->remove(igvn);
4333 }
4334 }
4335 if (proj_out_or_null(TypeFunc::Control) != nullptr) {
4336 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
4337 }
4338 if (is_Initialize()) {
4339 as_Initialize()->replace_mem_projs_by(in(TypeFunc::Memory), igvn);
4340 } else {
4341 if (proj_out_or_null(TypeFunc::Memory) != nullptr) {
4342 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
4343 }
4344 }
4345 }
4346
4347 //------------------------------Ideal------------------------------------------
4348 // Return a node which is more "ideal" than the current node. Strip out
4349 // control copies
4350 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4351 if (remove_dead_region(phase, can_reshape)) return this;
4352 // Don't bother trying to transform a dead node
4353 if (in(0) && in(0)->is_top()) {
4354 return nullptr;
4355 }
4356
4357 bool progress = false;
4358 // Eliminate volatile MemBars for scalar replaced objects.
4359 if (can_reshape && req() == (Precedent+1)) {
4360 bool eliminate = false;
4361 int opc = Opcode();
4362 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
4363 // Volatile field loads and stores.
4364 Node* my_mem = in(MemBarNode::Precedent);
4365 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
4366 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
4367 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
4368 // replace this Precedent (decodeN) with the Load instead.
4369 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
4370 Node* load_node = my_mem->in(1);
4371 set_req(MemBarNode::Precedent, load_node);
4372 phase->is_IterGVN()->_worklist.push(my_mem);
4373 my_mem = load_node;
4374 } else {
4375 assert(my_mem->unique_out() == this, "sanity");
4376 assert(!trailing_load_store(), "load store node can't be eliminated");
4377 del_req(Precedent);
4378 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
4379 my_mem = nullptr;
4380 }
4381 progress = true;
4382 }
4383 if (my_mem != nullptr && my_mem->is_Mem()) {
4384 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
4385 // Check for scalar replaced object reference.
4386 if( t_oop != nullptr && t_oop->is_known_instance_field() &&
4387 t_oop->offset() != Type::OffsetBot &&
4388 t_oop->offset() != Type::OffsetTop) {
4389 eliminate = true;
4390 }
4391 }
4392 } else if (opc == Op_MemBarRelease || (UseStoreStoreForCtor && opc == Op_MemBarStoreStore)) {
4393 // Final field stores.
4394 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent));
4395 if ((alloc != nullptr) && alloc->is_Allocate() &&
4396 alloc->as_Allocate()->does_not_escape_thread()) {
4397 // The allocated object does not escape.
4398 eliminate = true;
4399 }
4400 }
4401 if (eliminate) {
4402 // Replace MemBar projections by its inputs.
4403 PhaseIterGVN* igvn = phase->is_IterGVN();
4404 remove(igvn);
4405 // Must return either the original node (now dead) or a new node
4406 // (Do not return a top here, since that would break the uniqueness of top.)
4407 return new ConINode(TypeInt::ZERO);
4408 }
4409 }
4410 return progress ? this : nullptr;
4411 }
4412
4413 //------------------------------Value------------------------------------------
4414 const Type* MemBarNode::Value(PhaseGVN* phase) const {
4415 if( !in(0) ) return Type::TOP;
4416 if( phase->type(in(0)) == Type::TOP )
4417 return Type::TOP;
4418 return TypeTuple::MEMBAR;
4419 }
4420
4421 //------------------------------match------------------------------------------
4422 // Construct projections for memory.
4423 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
4424 switch (proj->_con) {
4425 case TypeFunc::Control:
4426 case TypeFunc::Memory:
4427 return new MachProjNode(this, proj->_con, RegMask::EMPTY, MachProjNode::unmatched_proj);
4428 }
4429 ShouldNotReachHere();
4430 return nullptr;
4431 }
4432
4433 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4434 trailing->_kind = TrailingStore;
4435 leading->_kind = LeadingStore;
4436 #ifdef ASSERT
4437 trailing->_pair_idx = leading->_idx;
4438 leading->_pair_idx = leading->_idx;
4439 #endif
4440 }
4441
4442 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4443 trailing->_kind = TrailingLoadStore;
4444 leading->_kind = LeadingLoadStore;
4445 #ifdef ASSERT
4446 trailing->_pair_idx = leading->_idx;
4447 leading->_pair_idx = leading->_idx;
4448 #endif
4449 }
4450
4451 MemBarNode* MemBarNode::trailing_membar() const {
4452 ResourceMark rm;
4453 Node* trailing = (Node*)this;
4454 VectorSet seen;
4455 Node_Stack multis(0);
4456 do {
4457 Node* c = trailing;
4458 uint i = 0;
4459 do {
4460 trailing = nullptr;
4461 for (; i < c->outcnt(); i++) {
4462 Node* next = c->raw_out(i);
4463 if (next != c && next->is_CFG()) {
4464 if (c->is_MultiBranch()) {
4465 if (multis.node() == c) {
4466 multis.set_index(i+1);
4467 } else {
4468 multis.push(c, i+1);
4469 }
4470 }
4471 trailing = next;
4472 break;
4473 }
4474 }
4475 if (trailing != nullptr && !seen.test_set(trailing->_idx)) {
4476 break;
4477 }
4478 while (multis.size() > 0) {
4479 c = multis.node();
4480 i = multis.index();
4481 if (i < c->req()) {
4482 break;
4483 }
4484 multis.pop();
4485 }
4486 } while (multis.size() > 0);
4487 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
4488
4489 MemBarNode* mb = trailing->as_MemBar();
4490 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
4491 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
4492 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
4493 return mb;
4494 }
4495
4496 MemBarNode* MemBarNode::leading_membar() const {
4497 ResourceMark rm;
4498 VectorSet seen;
4499 Node_Stack regions(0);
4500 Node* leading = in(0);
4501 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
4502 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) {
4503 leading = nullptr;
4504 while (regions.size() > 0 && leading == nullptr) {
4505 Node* r = regions.node();
4506 uint i = regions.index();
4507 if (i < r->req()) {
4508 leading = r->in(i);
4509 regions.set_index(i+1);
4510 } else {
4511 regions.pop();
4512 }
4513 }
4514 if (leading == nullptr) {
4515 assert(regions.size() == 0, "all paths should have been tried");
4516 return nullptr;
4517 }
4518 }
4519 if (leading->is_Region()) {
4520 regions.push(leading, 2);
4521 leading = leading->in(1);
4522 } else {
4523 leading = leading->in(0);
4524 }
4525 }
4526 #ifdef ASSERT
4527 Unique_Node_List wq;
4528 wq.push((Node*)this);
4529 uint found = 0;
4530 for (uint i = 0; i < wq.size(); i++) {
4531 Node* n = wq.at(i);
4532 if (n->is_Region()) {
4533 for (uint j = 1; j < n->req(); j++) {
4534 Node* in = n->in(j);
4535 if (in != nullptr && !in->is_top()) {
4536 wq.push(in);
4537 }
4538 }
4539 } else {
4540 if (n->is_MemBar() && n->as_MemBar()->leading()) {
4541 assert(n == leading, "consistency check failed");
4542 found++;
4543 } else {
4544 Node* in = n->in(0);
4545 if (in != nullptr && !in->is_top()) {
4546 wq.push(in);
4547 }
4548 }
4549 }
4550 }
4551 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed");
4552 #endif
4553 if (leading == nullptr) {
4554 return nullptr;
4555 }
4556 MemBarNode* mb = leading->as_MemBar();
4557 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
4558 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
4559 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
4560 return mb;
4561 }
4562
4563
4564 //===========================InitializeNode====================================
4565 // SUMMARY:
4566 // This node acts as a memory barrier on raw memory, after some raw stores.
4567 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
4568 // The Initialize can 'capture' suitably constrained stores as raw inits.
4569 // It can coalesce related raw stores into larger units (called 'tiles').
4570 // It can avoid zeroing new storage for memory units which have raw inits.
4571 // At macro-expansion, it is marked 'complete', and does not optimize further.
4572 //
4573 // EXAMPLE:
4574 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
4575 // ctl = incoming control; mem* = incoming memory
4576 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
4577 // First allocate uninitialized memory and fill in the header:
4578 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
4579 // ctl := alloc.Control; mem* := alloc.Memory*
4580 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
4581 // Then initialize to zero the non-header parts of the raw memory block:
4582 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
4583 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
4584 // After the initialize node executes, the object is ready for service:
4585 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
4586 // Suppose its body is immediately initialized as {1,2}:
4587 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4588 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4589 // mem.SLICE(#short[*]) := store2
4590 //
4591 // DETAILS:
4592 // An InitializeNode collects and isolates object initialization after
4593 // an AllocateNode and before the next possible safepoint. As a
4594 // memory barrier (MemBarNode), it keeps critical stores from drifting
4595 // down past any safepoint or any publication of the allocation.
4596 // Before this barrier, a newly-allocated object may have uninitialized bits.
4597 // After this barrier, it may be treated as a real oop, and GC is allowed.
4598 //
4599 // The semantics of the InitializeNode include an implicit zeroing of
4600 // the new object from object header to the end of the object.
4601 // (The object header and end are determined by the AllocateNode.)
4602 //
4603 // Certain stores may be added as direct inputs to the InitializeNode.
4604 // These stores must update raw memory, and they must be to addresses
4605 // derived from the raw address produced by AllocateNode, and with
4606 // a constant offset. They must be ordered by increasing offset.
4607 // The first one is at in(RawStores), the last at in(req()-1).
4608 // Unlike most memory operations, they are not linked in a chain,
4609 // but are displayed in parallel as users of the rawmem output of
4610 // the allocation.
4611 //
4612 // (See comments in InitializeNode::capture_store, which continue
4613 // the example given above.)
4614 //
4615 // When the associated Allocate is macro-expanded, the InitializeNode
4616 // may be rewritten to optimize collected stores. A ClearArrayNode
4617 // may also be created at that point to represent any required zeroing.
4618 // The InitializeNode is then marked 'complete', prohibiting further
4619 // capturing of nearby memory operations.
4620 //
4621 // During macro-expansion, all captured initializations which store
4622 // constant values of 32 bits or smaller are coalesced (if advantageous)
4623 // into larger 'tiles' 32 or 64 bits. This allows an object to be
4624 // initialized in fewer memory operations. Memory words which are
4625 // covered by neither tiles nor non-constant stores are pre-zeroed
4626 // by explicit stores of zero. (The code shape happens to do all
4627 // zeroing first, then all other stores, with both sequences occurring
4628 // in order of ascending offsets.)
4629 //
4630 // Alternatively, code may be inserted between an AllocateNode and its
4631 // InitializeNode, to perform arbitrary initialization of the new object.
4632 // E.g., the object copying intrinsics insert complex data transfers here.
4633 // The initialization must then be marked as 'complete' disable the
4634 // built-in zeroing semantics and the collection of initializing stores.
4635 //
4636 // While an InitializeNode is incomplete, reads from the memory state
4637 // produced by it are optimizable if they match the control edge and
4638 // new oop address associated with the allocation/initialization.
4639 // They return a stored value (if the offset matches) or else zero.
4640 // A write to the memory state, if it matches control and address,
4641 // and if it is to a constant offset, may be 'captured' by the
4642 // InitializeNode. It is cloned as a raw memory operation and rewired
4643 // inside the initialization, to the raw oop produced by the allocation.
4644 // Operations on addresses which are provably distinct (e.g., to
4645 // other AllocateNodes) are allowed to bypass the initialization.
4646 //
4647 // The effect of all this is to consolidate object initialization
4648 // (both arrays and non-arrays, both piecewise and bulk) into a
4649 // single location, where it can be optimized as a unit.
4650 //
4651 // Only stores with an offset less than TrackedInitializationLimit words
4652 // will be considered for capture by an InitializeNode. This puts a
4653 // reasonable limit on the complexity of optimized initializations.
4654
4655 //---------------------------InitializeNode------------------------------------
4656 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
4657 : MemBarNode(C, adr_type, rawoop),
4658 _is_complete(Incomplete), _does_not_escape(false)
4659 {
4660 init_class_id(Class_Initialize);
4661
4662 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
4663 assert(in(RawAddress) == rawoop, "proper init");
4664 // Note: allocation() can be null, for secondary initialization barriers
4665 }
4666
4667 // Since this node is not matched, it will be processed by the
4668 // register allocator. Declare that there are no constraints
4669 // on the allocation of the RawAddress edge.
4670 const RegMask &InitializeNode::in_RegMask(uint idx) const {
4671 // This edge should be set to top, by the set_complete. But be conservative.
4672 if (idx == InitializeNode::RawAddress)
4673 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
4674 return RegMask::EMPTY;
4675 }
4676
4677 Node* InitializeNode::memory(uint alias_idx) {
4678 Node* mem = in(Memory);
4679 if (mem->is_MergeMem()) {
4680 return mem->as_MergeMem()->memory_at(alias_idx);
4681 } else {
4682 // incoming raw memory is not split
4683 return mem;
4684 }
4685 }
4686
4687 bool InitializeNode::is_non_zero() {
4688 if (is_complete()) return false;
4689 remove_extra_zeroes();
4690 return (req() > RawStores);
4691 }
4692
4693 void InitializeNode::set_complete(PhaseGVN* phase) {
4694 assert(!is_complete(), "caller responsibility");
4695 _is_complete = Complete;
4696
4697 // After this node is complete, it contains a bunch of
4698 // raw-memory initializations. There is no need for
4699 // it to have anything to do with non-raw memory effects.
4700 // Therefore, tell all non-raw users to re-optimize themselves,
4701 // after skipping the memory effects of this initialization.
4702 PhaseIterGVN* igvn = phase->is_IterGVN();
4703 if (igvn) igvn->add_users_to_worklist(this);
4704 }
4705
4706 // convenience function
4707 // return false if the init contains any stores already
4708 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
4709 InitializeNode* init = initialization();
4710 if (init == nullptr || init->is_complete()) return false;
4711 init->remove_extra_zeroes();
4712 // for now, if this allocation has already collected any inits, bail:
4713 if (init->is_non_zero()) return false;
4714 init->set_complete(phase);
4715 return true;
4716 }
4717
4718 void InitializeNode::remove_extra_zeroes() {
4719 if (req() == RawStores) return;
4720 Node* zmem = zero_memory();
4721 uint fill = RawStores;
4722 for (uint i = fill; i < req(); i++) {
4723 Node* n = in(i);
4724 if (n->is_top() || n == zmem) continue; // skip
4725 if (fill < i) set_req(fill, n); // compact
4726 ++fill;
4727 }
4728 // delete any empty spaces created:
4729 while (fill < req()) {
4730 del_req(fill);
4731 }
4732 }
4733
4734 // Helper for remembering which stores go with which offsets.
4735 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) {
4736 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
4737 intptr_t offset = -1;
4738 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
4739 phase, offset);
4740 if (base == nullptr) return -1; // something is dead,
4741 if (offset < 0) return -1; // dead, dead
4742 return offset;
4743 }
4744
4745 // Helper for proving that an initialization expression is
4746 // "simple enough" to be folded into an object initialization.
4747 // Attempts to prove that a store's initial value 'n' can be captured
4748 // within the initialization without creating a vicious cycle, such as:
4749 // { Foo p = new Foo(); p.next = p; }
4750 // True for constants and parameters and small combinations thereof.
4751 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
4752 ResourceMark rm;
4753 Unique_Node_List worklist;
4754 worklist.push(value);
4755
4756 uint complexity_limit = 20;
4757 for (uint j = 0; j < worklist.size(); j++) {
4758 if (j >= complexity_limit) {
4759 return false; // Bail out if processed too many nodes
4760 }
4761
4762 Node* n = worklist.at(j);
4763 if (n == nullptr) continue; // (can this really happen?)
4764 if (n->is_Proj()) n = n->in(0);
4765 if (n == this) return false; // found a cycle
4766 if (n->is_Con()) continue;
4767 if (n->is_Start()) continue; // params, etc., are OK
4768 if (n->is_Root()) continue; // even better
4769
4770 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
4771 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
4772 continue;
4773 }
4774
4775 Node* ctl = n->in(0);
4776 if (ctl != nullptr && !ctl->is_top()) {
4777 if (ctl->is_Proj()) ctl = ctl->in(0);
4778 if (ctl == this) return false;
4779
4780 // If we already know that the enclosing memory op is pinned right after
4781 // the init, then any control flow that the store has picked up
4782 // must have preceded the init, or else be equal to the init.
4783 // Even after loop optimizations (which might change control edges)
4784 // a store is never pinned *before* the availability of its inputs.
4785 if (!MemNode::all_controls_dominate(n, this)) {
4786 return false; // failed to prove a good control
4787 }
4788 }
4789
4790 // Check data edges for possible dependencies on 'this'.
4791 for (uint i = 1; i < n->req(); i++) {
4792 Node* m = n->in(i);
4793 if (m == nullptr || m == n || m->is_top()) continue;
4794
4795 // Only process data inputs once
4796 worklist.push(m);
4797 }
4798 }
4799
4800 return true;
4801 }
4802
4803 // Here are all the checks a Store must pass before it can be moved into
4804 // an initialization. Returns zero if a check fails.
4805 // On success, returns the (constant) offset to which the store applies,
4806 // within the initialized memory.
4807 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
4808 const int FAIL = 0;
4809 if (st->req() != MemNode::ValueIn + 1)
4810 return FAIL; // an inscrutable StoreNode (card mark?)
4811 Node* ctl = st->in(MemNode::Control);
4812 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this))
4813 return FAIL; // must be unconditional after the initialization
4814 Node* mem = st->in(MemNode::Memory);
4815 if (!(mem->is_Proj() && mem->in(0) == this))
4816 return FAIL; // must not be preceded by other stores
4817 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4818 if ((st->Opcode() == Op_StoreP || st->Opcode() == Op_StoreN) &&
4819 !bs->can_initialize_object(st)) {
4820 return FAIL;
4821 }
4822 Node* adr = st->in(MemNode::Address);
4823 intptr_t offset;
4824 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
4825 if (alloc == nullptr)
4826 return FAIL; // inscrutable address
4827 if (alloc != allocation())
4828 return FAIL; // wrong allocation! (store needs to float up)
4829 int size_in_bytes = st->memory_size();
4830 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
4831 return FAIL; // mismatched access
4832 }
4833 Node* val = st->in(MemNode::ValueIn);
4834
4835 if (!detect_init_independence(val, phase))
4836 return FAIL; // stored value must be 'simple enough'
4837
4838 // The Store can be captured only if nothing after the allocation
4839 // and before the Store is using the memory location that the store
4840 // overwrites.
4841 bool failed = false;
4842 // If is_complete_with_arraycopy() is true the shape of the graph is
4843 // well defined and is safe so no need for extra checks.
4844 if (!is_complete_with_arraycopy()) {
4845 // We are going to look at each use of the memory state following
4846 // the allocation to make sure nothing reads the memory that the
4847 // Store writes.
4848 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
4849 int alias_idx = phase->C->get_alias_index(t_adr);
4850 ResourceMark rm;
4851 Unique_Node_List mems;
4852 mems.push(mem);
4853 Node* unique_merge = nullptr;
4854 for (uint next = 0; next < mems.size(); ++next) {
4855 Node *m = mems.at(next);
4856 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
4857 Node *n = m->fast_out(j);
4858 if (n->outcnt() == 0) {
4859 continue;
4860 }
4861 if (n == st) {
4862 continue;
4863 } else if (n->in(0) != nullptr && n->in(0) != ctl) {
4864 // If the control of this use is different from the control
4865 // of the Store which is right after the InitializeNode then
4866 // this node cannot be between the InitializeNode and the
4867 // Store.
4868 continue;
4869 } else if (n->is_MergeMem()) {
4870 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
4871 // We can hit a MergeMemNode (that will likely go away
4872 // later) that is a direct use of the memory state
4873 // following the InitializeNode on the same slice as the
4874 // store node that we'd like to capture. We need to check
4875 // the uses of the MergeMemNode.
4876 mems.push(n);
4877 }
4878 } else if (n->is_Mem()) {
4879 Node* other_adr = n->in(MemNode::Address);
4880 if (other_adr == adr) {
4881 failed = true;
4882 break;
4883 } else {
4884 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
4885 if (other_t_adr != nullptr) {
4886 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
4887 if (other_alias_idx == alias_idx) {
4888 // A load from the same memory slice as the store right
4889 // after the InitializeNode. We check the control of the
4890 // object/array that is loaded from. If it's the same as
4891 // the store control then we cannot capture the store.
4892 assert(!n->is_Store(), "2 stores to same slice on same control?");
4893 Node* base = other_adr;
4894 assert(base->is_AddP(), "should be addp but is %s", base->Name());
4895 base = base->in(AddPNode::Base);
4896 if (base != nullptr) {
4897 base = base->uncast();
4898 if (base->is_Proj() && base->in(0) == alloc) {
4899 failed = true;
4900 break;
4901 }
4902 }
4903 }
4904 }
4905 }
4906 } else {
4907 failed = true;
4908 break;
4909 }
4910 }
4911 }
4912 }
4913 if (failed) {
4914 if (!can_reshape) {
4915 // We decided we couldn't capture the store during parsing. We
4916 // should try again during the next IGVN once the graph is
4917 // cleaner.
4918 phase->C->record_for_igvn(st);
4919 }
4920 return FAIL;
4921 }
4922
4923 return offset; // success
4924 }
4925
4926 // Find the captured store in(i) which corresponds to the range
4927 // [start..start+size) in the initialized object.
4928 // If there is one, return its index i. If there isn't, return the
4929 // negative of the index where it should be inserted.
4930 // Return 0 if the queried range overlaps an initialization boundary
4931 // or if dead code is encountered.
4932 // If size_in_bytes is zero, do not bother with overlap checks.
4933 int InitializeNode::captured_store_insertion_point(intptr_t start,
4934 int size_in_bytes,
4935 PhaseValues* phase) {
4936 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
4937
4938 if (is_complete())
4939 return FAIL; // arraycopy got here first; punt
4940
4941 assert(allocation() != nullptr, "must be present");
4942
4943 // no negatives, no header fields:
4944 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
4945
4946 // after a certain size, we bail out on tracking all the stores:
4947 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4948 if (start >= ti_limit) return FAIL;
4949
4950 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
4951 if (i >= limit) return -(int)i; // not found; here is where to put it
4952
4953 Node* st = in(i);
4954 intptr_t st_off = get_store_offset(st, phase);
4955 if (st_off < 0) {
4956 if (st != zero_memory()) {
4957 return FAIL; // bail out if there is dead garbage
4958 }
4959 } else if (st_off > start) {
4960 // ...we are done, since stores are ordered
4961 if (st_off < start + size_in_bytes) {
4962 return FAIL; // the next store overlaps
4963 }
4964 return -(int)i; // not found; here is where to put it
4965 } else if (st_off < start) {
4966 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
4967 if (size_in_bytes != 0 &&
4968 start < st_off + MAX_STORE &&
4969 start < st_off + st->as_Store()->memory_size()) {
4970 return FAIL; // the previous store overlaps
4971 }
4972 } else {
4973 if (size_in_bytes != 0 &&
4974 st->as_Store()->memory_size() != size_in_bytes) {
4975 return FAIL; // mismatched store size
4976 }
4977 return i;
4978 }
4979
4980 ++i;
4981 }
4982 }
4983
4984 // Look for a captured store which initializes at the offset 'start'
4985 // with the given size. If there is no such store, and no other
4986 // initialization interferes, then return zero_memory (the memory
4987 // projection of the AllocateNode).
4988 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
4989 PhaseValues* phase) {
4990 assert(stores_are_sane(phase), "");
4991 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4992 if (i == 0) {
4993 return nullptr; // something is dead
4994 } else if (i < 0) {
4995 return zero_memory(); // just primordial zero bits here
4996 } else {
4997 Node* st = in(i); // here is the store at this position
4998 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
4999 return st;
5000 }
5001 }
5002
5003 // Create, as a raw pointer, an address within my new object at 'offset'.
5004 Node* InitializeNode::make_raw_address(intptr_t offset,
5005 PhaseGVN* phase) {
5006 Node* addr = in(RawAddress);
5007 if (offset != 0) {
5008 Compile* C = phase->C;
5009 addr = phase->transform( new AddPNode(C->top(), addr,
5010 phase->MakeConX(offset)) );
5011 }
5012 return addr;
5013 }
5014
5015 // Clone the given store, converting it into a raw store
5016 // initializing a field or element of my new object.
5017 // Caller is responsible for retiring the original store,
5018 // with subsume_node or the like.
5019 //
5020 // From the example above InitializeNode::InitializeNode,
5021 // here are the old stores to be captured:
5022 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
5023 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
5024 //
5025 // Here is the changed code; note the extra edges on init:
5026 // alloc = (Allocate ...)
5027 // rawoop = alloc.RawAddress
5028 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
5029 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
5030 // init = (Initialize alloc.Control alloc.Memory rawoop
5031 // rawstore1 rawstore2)
5032 //
5033 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
5034 PhaseGVN* phase, bool can_reshape) {
5035 assert(stores_are_sane(phase), "");
5036
5037 if (start < 0) return nullptr;
5038 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
5039
5040 Compile* C = phase->C;
5041 int size_in_bytes = st->memory_size();
5042 int i = captured_store_insertion_point(start, size_in_bytes, phase);
5043 if (i == 0) return nullptr; // bail out
5044 Node* prev_mem = nullptr; // raw memory for the captured store
5045 if (i > 0) {
5046 prev_mem = in(i); // there is a pre-existing store under this one
5047 set_req(i, C->top()); // temporarily disconnect it
5048 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
5049 } else {
5050 i = -i; // no pre-existing store
5051 prev_mem = zero_memory(); // a slice of the newly allocated object
5052 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
5053 set_req(--i, C->top()); // reuse this edge; it has been folded away
5054 else
5055 ins_req(i, C->top()); // build a new edge
5056 }
5057 Node* new_st = st->clone();
5058 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
5059 new_st->set_req(MemNode::Control, in(Control));
5060 new_st->set_req(MemNode::Memory, prev_mem);
5061 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
5062 bs->eliminate_gc_barrier_data(new_st);
5063 new_st = phase->transform(new_st);
5064
5065 // At this point, new_st might have swallowed a pre-existing store
5066 // at the same offset, or perhaps new_st might have disappeared,
5067 // if it redundantly stored the same value (or zero to fresh memory).
5068
5069 // In any case, wire it in:
5070 PhaseIterGVN* igvn = phase->is_IterGVN();
5071 if (igvn) {
5072 igvn->rehash_node_delayed(this);
5073 }
5074 set_req(i, new_st);
5075
5076 // The caller may now kill the old guy.
5077 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
5078 assert(check_st == new_st || check_st == nullptr, "must be findable");
5079 assert(!is_complete(), "");
5080 return new_st;
5081 }
5082
5083 static bool store_constant(jlong* tiles, int num_tiles,
5084 intptr_t st_off, int st_size,
5085 jlong con) {
5086 if ((st_off & (st_size-1)) != 0)
5087 return false; // strange store offset (assume size==2**N)
5088 address addr = (address)tiles + st_off;
5089 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
5090 switch (st_size) {
5091 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
5092 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
5093 case sizeof(jint): *(jint*) addr = (jint) con; break;
5094 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
5095 default: return false; // strange store size (detect size!=2**N here)
5096 }
5097 return true; // return success to caller
5098 }
5099
5100 // Coalesce subword constants into int constants and possibly
5101 // into long constants. The goal, if the CPU permits,
5102 // is to initialize the object with a small number of 64-bit tiles.
5103 // Also, convert floating-point constants to bit patterns.
5104 // Non-constants are not relevant to this pass.
5105 //
5106 // In terms of the running example on InitializeNode::InitializeNode
5107 // and InitializeNode::capture_store, here is the transformation
5108 // of rawstore1 and rawstore2 into rawstore12:
5109 // alloc = (Allocate ...)
5110 // rawoop = alloc.RawAddress
5111 // tile12 = 0x00010002
5112 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
5113 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
5114 //
5115 void
5116 InitializeNode::coalesce_subword_stores(intptr_t header_size,
5117 Node* size_in_bytes,
5118 PhaseGVN* phase) {
5119 Compile* C = phase->C;
5120
5121 assert(stores_are_sane(phase), "");
5122 // Note: After this pass, they are not completely sane,
5123 // since there may be some overlaps.
5124
5125 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
5126
5127 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
5128 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
5129 size_limit = MIN2(size_limit, ti_limit);
5130 size_limit = align_up(size_limit, BytesPerLong);
5131 int num_tiles = size_limit / BytesPerLong;
5132
5133 // allocate space for the tile map:
5134 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
5135 jlong tiles_buf[small_len];
5136 Node* nodes_buf[small_len];
5137 jlong inits_buf[small_len];
5138 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
5139 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5140 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
5141 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
5142 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
5143 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5144 // tiles: exact bitwise model of all primitive constants
5145 // nodes: last constant-storing node subsumed into the tiles model
5146 // inits: which bytes (in each tile) are touched by any initializations
5147
5148 //// Pass A: Fill in the tile model with any relevant stores.
5149
5150 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
5151 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
5152 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
5153 Node* zmem = zero_memory(); // initially zero memory state
5154 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5155 Node* st = in(i);
5156 intptr_t st_off = get_store_offset(st, phase);
5157
5158 // Figure out the store's offset and constant value:
5159 if (st_off < header_size) continue; //skip (ignore header)
5160 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
5161 int st_size = st->as_Store()->memory_size();
5162 if (st_off + st_size > size_limit) break;
5163
5164 // Record which bytes are touched, whether by constant or not.
5165 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
5166 continue; // skip (strange store size)
5167
5168 const Type* val = phase->type(st->in(MemNode::ValueIn));
5169 if (!val->singleton()) continue; //skip (non-con store)
5170 BasicType type = val->basic_type();
5171
5172 jlong con = 0;
5173 switch (type) {
5174 case T_INT: con = val->is_int()->get_con(); break;
5175 case T_LONG: con = val->is_long()->get_con(); break;
5176 case T_FLOAT: con = jint_cast(val->getf()); break;
5177 case T_DOUBLE: con = jlong_cast(val->getd()); break;
5178 default: continue; //skip (odd store type)
5179 }
5180
5181 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
5182 st->Opcode() == Op_StoreL) {
5183 continue; // This StoreL is already optimal.
5184 }
5185
5186 // Store down the constant.
5187 store_constant(tiles, num_tiles, st_off, st_size, con);
5188
5189 intptr_t j = st_off >> LogBytesPerLong;
5190
5191 if (type == T_INT && st_size == BytesPerInt
5192 && (st_off & BytesPerInt) == BytesPerInt) {
5193 jlong lcon = tiles[j];
5194 if (!Matcher::isSimpleConstant64(lcon) &&
5195 st->Opcode() == Op_StoreI) {
5196 // This StoreI is already optimal by itself.
5197 jint* intcon = (jint*) &tiles[j];
5198 intcon[1] = 0; // undo the store_constant()
5199
5200 // If the previous store is also optimal by itself, back up and
5201 // undo the action of the previous loop iteration... if we can.
5202 // But if we can't, just let the previous half take care of itself.
5203 st = nodes[j];
5204 st_off -= BytesPerInt;
5205 con = intcon[0];
5206 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) {
5207 assert(st_off >= header_size, "still ignoring header");
5208 assert(get_store_offset(st, phase) == st_off, "must be");
5209 assert(in(i-1) == zmem, "must be");
5210 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
5211 assert(con == tcon->is_int()->get_con(), "must be");
5212 // Undo the effects of the previous loop trip, which swallowed st:
5213 intcon[0] = 0; // undo store_constant()
5214 set_req(i-1, st); // undo set_req(i, zmem)
5215 nodes[j] = nullptr; // undo nodes[j] = st
5216 --old_subword; // undo ++old_subword
5217 }
5218 continue; // This StoreI is already optimal.
5219 }
5220 }
5221
5222 // This store is not needed.
5223 set_req(i, zmem);
5224 nodes[j] = st; // record for the moment
5225 if (st_size < BytesPerLong) // something has changed
5226 ++old_subword; // includes int/float, but who's counting...
5227 else ++old_long;
5228 }
5229
5230 if ((old_subword + old_long) == 0)
5231 return; // nothing more to do
5232
5233 //// Pass B: Convert any non-zero tiles into optimal constant stores.
5234 // Be sure to insert them before overlapping non-constant stores.
5235 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
5236 for (int j = 0; j < num_tiles; j++) {
5237 jlong con = tiles[j];
5238 jlong init = inits[j];
5239 if (con == 0) continue;
5240 jint con0, con1; // split the constant, address-wise
5241 jint init0, init1; // split the init map, address-wise
5242 { union { jlong con; jint intcon[2]; } u;
5243 u.con = con;
5244 con0 = u.intcon[0];
5245 con1 = u.intcon[1];
5246 u.con = init;
5247 init0 = u.intcon[0];
5248 init1 = u.intcon[1];
5249 }
5250
5251 Node* old = nodes[j];
5252 assert(old != nullptr, "need the prior store");
5253 intptr_t offset = (j * BytesPerLong);
5254
5255 bool split = !Matcher::isSimpleConstant64(con);
5256
5257 if (offset < header_size) {
5258 assert(offset + BytesPerInt >= header_size, "second int counts");
5259 assert(*(jint*)&tiles[j] == 0, "junk in header");
5260 split = true; // only the second word counts
5261 // Example: int a[] = { 42 ... }
5262 } else if (con0 == 0 && init0 == -1) {
5263 split = true; // first word is covered by full inits
5264 // Example: int a[] = { ... foo(), 42 ... }
5265 } else if (con1 == 0 && init1 == -1) {
5266 split = true; // second word is covered by full inits
5267 // Example: int a[] = { ... 42, foo() ... }
5268 }
5269
5270 // Here's a case where init0 is neither 0 nor -1:
5271 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
5272 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
5273 // In this case the tile is not split; it is (jlong)42.
5274 // The big tile is stored down, and then the foo() value is inserted.
5275 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
5276
5277 Node* ctl = old->in(MemNode::Control);
5278 Node* adr = make_raw_address(offset, phase);
5279 const TypePtr* atp = TypeRawPtr::BOTTOM;
5280
5281 // One or two coalesced stores to plop down.
5282 Node* st[2];
5283 intptr_t off[2];
5284 int nst = 0;
5285 if (!split) {
5286 ++new_long;
5287 off[nst] = offset;
5288 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5289 phase->longcon(con), T_LONG, MemNode::unordered);
5290 } else {
5291 // Omit either if it is a zero.
5292 if (con0 != 0) {
5293 ++new_int;
5294 off[nst] = offset;
5295 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5296 phase->intcon(con0), T_INT, MemNode::unordered);
5297 }
5298 if (con1 != 0) {
5299 ++new_int;
5300 offset += BytesPerInt;
5301 adr = make_raw_address(offset, phase);
5302 off[nst] = offset;
5303 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5304 phase->intcon(con1), T_INT, MemNode::unordered);
5305 }
5306 }
5307
5308 // Insert second store first, then the first before the second.
5309 // Insert each one just before any overlapping non-constant stores.
5310 while (nst > 0) {
5311 Node* st1 = st[--nst];
5312 C->copy_node_notes_to(st1, old);
5313 st1 = phase->transform(st1);
5314 offset = off[nst];
5315 assert(offset >= header_size, "do not smash header");
5316 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
5317 guarantee(ins_idx != 0, "must re-insert constant store");
5318 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
5319 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
5320 set_req(--ins_idx, st1);
5321 else
5322 ins_req(ins_idx, st1);
5323 }
5324 }
5325
5326 if (PrintCompilation && WizardMode)
5327 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
5328 old_subword, old_long, new_int, new_long);
5329 if (C->log() != nullptr)
5330 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
5331 old_subword, old_long, new_int, new_long);
5332
5333 // Clean up any remaining occurrences of zmem:
5334 remove_extra_zeroes();
5335 }
5336
5337 // Explore forward from in(start) to find the first fully initialized
5338 // word, and return its offset. Skip groups of subword stores which
5339 // together initialize full words. If in(start) is itself part of a
5340 // fully initialized word, return the offset of in(start). If there
5341 // are no following full-word stores, or if something is fishy, return
5342 // a negative value.
5343 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
5344 int int_map = 0;
5345 intptr_t int_map_off = 0;
5346 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
5347
5348 for (uint i = start, limit = req(); i < limit; i++) {
5349 Node* st = in(i);
5350
5351 intptr_t st_off = get_store_offset(st, phase);
5352 if (st_off < 0) break; // return conservative answer
5353
5354 int st_size = st->as_Store()->memory_size();
5355 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
5356 return st_off; // we found a complete word init
5357 }
5358
5359 // update the map:
5360
5361 intptr_t this_int_off = align_down(st_off, BytesPerInt);
5362 if (this_int_off != int_map_off) {
5363 // reset the map:
5364 int_map = 0;
5365 int_map_off = this_int_off;
5366 }
5367
5368 int subword_off = st_off - this_int_off;
5369 int_map |= right_n_bits(st_size) << subword_off;
5370 if ((int_map & FULL_MAP) == FULL_MAP) {
5371 return this_int_off; // we found a complete word init
5372 }
5373
5374 // Did this store hit or cross the word boundary?
5375 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
5376 if (next_int_off == this_int_off + BytesPerInt) {
5377 // We passed the current int, without fully initializing it.
5378 int_map_off = next_int_off;
5379 int_map >>= BytesPerInt;
5380 } else if (next_int_off > this_int_off + BytesPerInt) {
5381 // We passed the current and next int.
5382 return this_int_off + BytesPerInt;
5383 }
5384 }
5385
5386 return -1;
5387 }
5388
5389
5390 // Called when the associated AllocateNode is expanded into CFG.
5391 // At this point, we may perform additional optimizations.
5392 // Linearize the stores by ascending offset, to make memory
5393 // activity as coherent as possible.
5394 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
5395 intptr_t header_size,
5396 Node* size_in_bytes,
5397 PhaseIterGVN* phase) {
5398 assert(!is_complete(), "not already complete");
5399 assert(stores_are_sane(phase), "");
5400 assert(allocation() != nullptr, "must be present");
5401
5402 remove_extra_zeroes();
5403
5404 if (ReduceFieldZeroing || ReduceBulkZeroing)
5405 // reduce instruction count for common initialization patterns
5406 coalesce_subword_stores(header_size, size_in_bytes, phase);
5407
5408 Node* zmem = zero_memory(); // initially zero memory state
5409 Node* inits = zmem; // accumulating a linearized chain of inits
5410 #ifdef ASSERT
5411 intptr_t first_offset = allocation()->minimum_header_size();
5412 intptr_t last_init_off = first_offset; // previous init offset
5413 intptr_t last_init_end = first_offset; // previous init offset+size
5414 intptr_t last_tile_end = first_offset; // previous tile offset+size
5415 #endif
5416 intptr_t zeroes_done = header_size;
5417
5418 bool do_zeroing = true; // we might give up if inits are very sparse
5419 int big_init_gaps = 0; // how many large gaps have we seen?
5420
5421 if (UseTLAB && ZeroTLAB) do_zeroing = false;
5422 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
5423
5424 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5425 Node* st = in(i);
5426 intptr_t st_off = get_store_offset(st, phase);
5427 if (st_off < 0)
5428 break; // unknown junk in the inits
5429 if (st->in(MemNode::Memory) != zmem)
5430 break; // complicated store chains somehow in list
5431
5432 int st_size = st->as_Store()->memory_size();
5433 intptr_t next_init_off = st_off + st_size;
5434
5435 if (do_zeroing && zeroes_done < next_init_off) {
5436 // See if this store needs a zero before it or under it.
5437 intptr_t zeroes_needed = st_off;
5438
5439 if (st_size < BytesPerInt) {
5440 // Look for subword stores which only partially initialize words.
5441 // If we find some, we must lay down some word-level zeroes first,
5442 // underneath the subword stores.
5443 //
5444 // Examples:
5445 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
5446 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
5447 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
5448 //
5449 // Note: coalesce_subword_stores may have already done this,
5450 // if it was prompted by constant non-zero subword initializers.
5451 // But this case can still arise with non-constant stores.
5452
5453 intptr_t next_full_store = find_next_fullword_store(i, phase);
5454
5455 // In the examples above:
5456 // in(i) p q r s x y z
5457 // st_off 12 13 14 15 12 13 14
5458 // st_size 1 1 1 1 1 1 1
5459 // next_full_s. 12 16 16 16 16 16 16
5460 // z's_done 12 16 16 16 12 16 12
5461 // z's_needed 12 16 16 16 16 16 16
5462 // zsize 0 0 0 0 4 0 4
5463 if (next_full_store < 0) {
5464 // Conservative tack: Zero to end of current word.
5465 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
5466 } else {
5467 // Zero to beginning of next fully initialized word.
5468 // Or, don't zero at all, if we are already in that word.
5469 assert(next_full_store >= zeroes_needed, "must go forward");
5470 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
5471 zeroes_needed = next_full_store;
5472 }
5473 }
5474
5475 if (zeroes_needed > zeroes_done) {
5476 intptr_t zsize = zeroes_needed - zeroes_done;
5477 // Do some incremental zeroing on rawmem, in parallel with inits.
5478 zeroes_done = align_down(zeroes_done, BytesPerInt);
5479 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5480 zeroes_done, zeroes_needed,
5481 true,
5482 phase);
5483 zeroes_done = zeroes_needed;
5484 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
5485 do_zeroing = false; // leave the hole, next time
5486 }
5487 }
5488
5489 // Collect the store and move on:
5490 phase->replace_input_of(st, MemNode::Memory, inits);
5491 inits = st; // put it on the linearized chain
5492 set_req(i, zmem); // unhook from previous position
5493
5494 if (zeroes_done == st_off)
5495 zeroes_done = next_init_off;
5496
5497 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
5498
5499 #ifdef ASSERT
5500 // Various order invariants. Weaker than stores_are_sane because
5501 // a large constant tile can be filled in by smaller non-constant stores.
5502 assert(st_off >= last_init_off, "inits do not reverse");
5503 last_init_off = st_off;
5504 const Type* val = nullptr;
5505 if (st_size >= BytesPerInt &&
5506 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
5507 (int)val->basic_type() < (int)T_OBJECT) {
5508 assert(st_off >= last_tile_end, "tiles do not overlap");
5509 assert(st_off >= last_init_end, "tiles do not overwrite inits");
5510 last_tile_end = MAX2(last_tile_end, next_init_off);
5511 } else {
5512 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
5513 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
5514 assert(st_off >= last_init_end, "inits do not overlap");
5515 last_init_end = next_init_off; // it's a non-tile
5516 }
5517 #endif //ASSERT
5518 }
5519
5520 remove_extra_zeroes(); // clear out all the zmems left over
5521 add_req(inits);
5522
5523 if (!(UseTLAB && ZeroTLAB)) {
5524 // If anything remains to be zeroed, zero it all now.
5525 zeroes_done = align_down(zeroes_done, BytesPerInt);
5526 // if it is the last unused 4 bytes of an instance, forget about it
5527 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
5528 if (zeroes_done + BytesPerLong >= size_limit) {
5529 AllocateNode* alloc = allocation();
5530 assert(alloc != nullptr, "must be present");
5531 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
5532 Node* klass_node = alloc->in(AllocateNode::KlassNode);
5533 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
5534 if (zeroes_done == k->layout_helper())
5535 zeroes_done = size_limit;
5536 }
5537 }
5538 if (zeroes_done < size_limit) {
5539 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5540 zeroes_done, size_in_bytes, true, phase);
5541 }
5542 }
5543
5544 set_complete(phase);
5545 return rawmem;
5546 }
5547
5548 void InitializeNode::replace_mem_projs_by(Node* mem, Compile* C) {
5549 auto replace_proj = [&](ProjNode* proj) {
5550 C->gvn_replace_by(proj, mem);
5551 return CONTINUE;
5552 };
5553 apply_to_projs(replace_proj, TypeFunc::Memory);
5554 }
5555
5556 void InitializeNode::replace_mem_projs_by(Node* mem, PhaseIterGVN* igvn) {
5557 DUIterator_Fast imax, i = fast_outs(imax);
5558 auto replace_proj = [&](ProjNode* proj) {
5559 igvn->replace_node(proj, mem);
5560 --i; --imax;
5561 return CONTINUE;
5562 };
5563 apply_to_projs(imax, i, replace_proj, TypeFunc::Memory);
5564 }
5565
5566 bool InitializeNode::already_has_narrow_mem_proj_with_adr_type(const TypePtr* adr_type) const {
5567 auto find_proj = [&](ProjNode* proj) {
5568 if (proj->adr_type() == adr_type) {
5569 return BREAK_AND_RETURN_CURRENT_PROJ;
5570 }
5571 return CONTINUE;
5572 };
5573 DUIterator_Fast imax, i = fast_outs(imax);
5574 return apply_to_narrow_mem_projs_any_iterator(UsesIteratorFast(imax, i, this), find_proj) != nullptr;
5575 }
5576
5577 MachProjNode* InitializeNode::mem_mach_proj() const {
5578 auto find_proj = [](ProjNode* proj) {
5579 if (proj->is_MachProj()) {
5580 return BREAK_AND_RETURN_CURRENT_PROJ;
5581 }
5582 return CONTINUE;
5583 };
5584 ProjNode* proj = apply_to_projs(find_proj, TypeFunc::Memory);
5585 if (proj == nullptr) {
5586 return nullptr;
5587 }
5588 return proj->as_MachProj();
5589 }
5590
5591 #ifdef ASSERT
5592 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
5593 if (is_complete())
5594 return true; // stores could be anything at this point
5595 assert(allocation() != nullptr, "must be present");
5596 intptr_t last_off = allocation()->minimum_header_size();
5597 for (uint i = InitializeNode::RawStores; i < req(); i++) {
5598 Node* st = in(i);
5599 intptr_t st_off = get_store_offset(st, phase);
5600 if (st_off < 0) continue; // ignore dead garbage
5601 if (last_off > st_off) {
5602 tty->print_cr("*** bad store offset at %d: %zd > %zd", i, last_off, st_off);
5603 this->dump(2);
5604 assert(false, "ascending store offsets");
5605 return false;
5606 }
5607 last_off = st_off + st->as_Store()->memory_size();
5608 }
5609 return true;
5610 }
5611 #endif //ASSERT
5612
5613
5614
5615
5616 //============================MergeMemNode=====================================
5617 //
5618 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
5619 // contributing store or call operations. Each contributor provides the memory
5620 // state for a particular "alias type" (see Compile::alias_type). For example,
5621 // if a MergeMem has an input X for alias category #6, then any memory reference
5622 // to alias category #6 may use X as its memory state input, as an exact equivalent
5623 // to using the MergeMem as a whole.
5624 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
5625 //
5626 // (Here, the <N> notation gives the index of the relevant adr_type.)
5627 //
5628 // In one special case (and more cases in the future), alias categories overlap.
5629 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
5630 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
5631 // it is exactly equivalent to that state W:
5632 // MergeMem(<Bot>: W) <==> W
5633 //
5634 // Usually, the merge has more than one input. In that case, where inputs
5635 // overlap (i.e., one is Bot), the narrower alias type determines the memory
5636 // state for that type, and the wider alias type (Bot) fills in everywhere else:
5637 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
5638 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
5639 //
5640 // A merge can take a "wide" memory state as one of its narrow inputs.
5641 // This simply means that the merge observes out only the relevant parts of
5642 // the wide input. That is, wide memory states arriving at narrow merge inputs
5643 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
5644 //
5645 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
5646 // and that memory slices "leak through":
5647 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
5648 //
5649 // But, in such a cascade, repeated memory slices can "block the leak":
5650 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
5651 //
5652 // In the last example, Y is not part of the combined memory state of the
5653 // outermost MergeMem. The system must, of course, prevent unschedulable
5654 // memory states from arising, so you can be sure that the state Y is somehow
5655 // a precursor to state Y'.
5656 //
5657 //
5658 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
5659 // of each MergeMemNode array are exactly the numerical alias indexes, including
5660 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
5661 // Compile::alias_type (and kin) produce and manage these indexes.
5662 //
5663 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
5664 // (Note that this provides quick access to the top node inside MergeMem methods,
5665 // without the need to reach out via TLS to Compile::current.)
5666 //
5667 // As a consequence of what was just described, a MergeMem that represents a full
5668 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
5669 // containing all alias categories.
5670 //
5671 // MergeMem nodes never (?) have control inputs, so in(0) is null.
5672 //
5673 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
5674 // a memory state for the alias type <N>, or else the top node, meaning that
5675 // there is no particular input for that alias type. Note that the length of
5676 // a MergeMem is variable, and may be extended at any time to accommodate new
5677 // memory states at larger alias indexes. When merges grow, they are of course
5678 // filled with "top" in the unused in() positions.
5679 //
5680 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
5681 // (Top was chosen because it works smoothly with passes like GCM.)
5682 //
5683 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
5684 // the type of random VM bits like TLS references.) Since it is always the
5685 // first non-Bot memory slice, some low-level loops use it to initialize an
5686 // index variable: for (i = AliasIdxRaw; i < req(); i++).
5687 //
5688 //
5689 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
5690 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
5691 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
5692 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
5693 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
5694 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
5695 //
5696 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
5697 // really that different from the other memory inputs. An abbreviation called
5698 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
5699 //
5700 //
5701 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
5702 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
5703 // that "emerges though" the base memory will be marked as excluding the alias types
5704 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
5705 //
5706 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
5707 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
5708 //
5709 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
5710 // (It is currently unimplemented.) As you can see, the resulting merge is
5711 // actually a disjoint union of memory states, rather than an overlay.
5712 //
5713
5714 //------------------------------MergeMemNode-----------------------------------
5715 Node* MergeMemNode::make_empty_memory() {
5716 Node* empty_memory = (Node*) Compile::current()->top();
5717 assert(empty_memory->is_top(), "correct sentinel identity");
5718 return empty_memory;
5719 }
5720
5721 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
5722 init_class_id(Class_MergeMem);
5723 // all inputs are nullified in Node::Node(int)
5724 // set_input(0, nullptr); // no control input
5725
5726 // Initialize the edges uniformly to top, for starters.
5727 Node* empty_mem = make_empty_memory();
5728 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
5729 init_req(i,empty_mem);
5730 }
5731 assert(empty_memory() == empty_mem, "");
5732
5733 if( new_base != nullptr && new_base->is_MergeMem() ) {
5734 MergeMemNode* mdef = new_base->as_MergeMem();
5735 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
5736 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
5737 mms.set_memory(mms.memory2());
5738 }
5739 assert(base_memory() == mdef->base_memory(), "");
5740 } else {
5741 set_base_memory(new_base);
5742 }
5743 }
5744
5745 // Make a new, untransformed MergeMem with the same base as 'mem'.
5746 // If mem is itself a MergeMem, populate the result with the same edges.
5747 MergeMemNode* MergeMemNode::make(Node* mem) {
5748 return new MergeMemNode(mem);
5749 }
5750
5751 //------------------------------cmp--------------------------------------------
5752 uint MergeMemNode::hash() const { return NO_HASH; }
5753 bool MergeMemNode::cmp( const Node &n ) const {
5754 return (&n == this); // Always fail except on self
5755 }
5756
5757 //------------------------------Identity---------------------------------------
5758 Node* MergeMemNode::Identity(PhaseGVN* phase) {
5759 // Identity if this merge point does not record any interesting memory
5760 // disambiguations.
5761 Node* base_mem = base_memory();
5762 Node* empty_mem = empty_memory();
5763 if (base_mem != empty_mem) { // Memory path is not dead?
5764 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5765 Node* mem = in(i);
5766 if (mem != empty_mem && mem != base_mem) {
5767 return this; // Many memory splits; no change
5768 }
5769 }
5770 }
5771 return base_mem; // No memory splits; ID on the one true input
5772 }
5773
5774 //------------------------------Ideal------------------------------------------
5775 // This method is invoked recursively on chains of MergeMem nodes
5776 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
5777 // Remove chain'd MergeMems
5778 //
5779 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
5780 // relative to the "in(Bot)". Since we are patching both at the same time,
5781 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
5782 // but rewrite each "in(i)" relative to the new "in(Bot)".
5783 Node *progress = nullptr;
5784
5785
5786 Node* old_base = base_memory();
5787 Node* empty_mem = empty_memory();
5788 if (old_base == empty_mem)
5789 return nullptr; // Dead memory path.
5790
5791 MergeMemNode* old_mbase;
5792 if (old_base != nullptr && old_base->is_MergeMem())
5793 old_mbase = old_base->as_MergeMem();
5794 else
5795 old_mbase = nullptr;
5796 Node* new_base = old_base;
5797
5798 // simplify stacked MergeMems in base memory
5799 if (old_mbase) new_base = old_mbase->base_memory();
5800
5801 // the base memory might contribute new slices beyond my req()
5802 if (old_mbase) grow_to_match(old_mbase);
5803
5804 // Note: We do not call verify_sparse on entry, because inputs
5805 // can normalize to the base_memory via subsume_node or similar
5806 // mechanisms. This method repairs that damage.
5807
5808 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
5809
5810 // Look at each slice.
5811 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5812 Node* old_in = in(i);
5813 // calculate the old memory value
5814 Node* old_mem = old_in;
5815 if (old_mem == empty_mem) old_mem = old_base;
5816 assert(old_mem == memory_at(i), "");
5817
5818 // maybe update (reslice) the old memory value
5819
5820 // simplify stacked MergeMems
5821 Node* new_mem = old_mem;
5822 MergeMemNode* old_mmem;
5823 if (old_mem != nullptr && old_mem->is_MergeMem())
5824 old_mmem = old_mem->as_MergeMem();
5825 else
5826 old_mmem = nullptr;
5827 if (old_mmem == this) {
5828 // This can happen if loops break up and safepoints disappear.
5829 // A merge of BotPtr (default) with a RawPtr memory derived from a
5830 // safepoint can be rewritten to a merge of the same BotPtr with
5831 // the BotPtr phi coming into the loop. If that phi disappears
5832 // also, we can end up with a self-loop of the mergemem.
5833 // In general, if loops degenerate and memory effects disappear,
5834 // a mergemem can be left looking at itself. This simply means
5835 // that the mergemem's default should be used, since there is
5836 // no longer any apparent effect on this slice.
5837 // Note: If a memory slice is a MergeMem cycle, it is unreachable
5838 // from start. Update the input to TOP.
5839 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
5840 }
5841 else if (old_mmem != nullptr) {
5842 new_mem = old_mmem->memory_at(i);
5843 }
5844 // else preceding memory was not a MergeMem
5845
5846 // maybe store down a new value
5847 Node* new_in = new_mem;
5848 if (new_in == new_base) new_in = empty_mem;
5849
5850 if (new_in != old_in) {
5851 // Warning: Do not combine this "if" with the previous "if"
5852 // A memory slice might have be be rewritten even if it is semantically
5853 // unchanged, if the base_memory value has changed.
5854 set_req_X(i, new_in, phase);
5855 progress = this; // Report progress
5856 }
5857 }
5858
5859 if (new_base != old_base) {
5860 set_req_X(Compile::AliasIdxBot, new_base, phase);
5861 // Don't use set_base_memory(new_base), because we need to update du.
5862 assert(base_memory() == new_base, "");
5863 progress = this;
5864 }
5865
5866 if( base_memory() == this ) {
5867 // a self cycle indicates this memory path is dead
5868 set_req(Compile::AliasIdxBot, empty_mem);
5869 }
5870
5871 // Resolve external cycles by calling Ideal on a MergeMem base_memory
5872 // Recursion must occur after the self cycle check above
5873 if( base_memory()->is_MergeMem() ) {
5874 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
5875 Node *m = phase->transform(new_mbase); // Rollup any cycles
5876 if( m != nullptr &&
5877 (m->is_top() ||
5878 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
5879 // propagate rollup of dead cycle to self
5880 set_req(Compile::AliasIdxBot, empty_mem);
5881 }
5882 }
5883
5884 if( base_memory() == empty_mem ) {
5885 progress = this;
5886 // Cut inputs during Parse phase only.
5887 // During Optimize phase a dead MergeMem node will be subsumed by Top.
5888 if( !can_reshape ) {
5889 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5890 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
5891 }
5892 }
5893 }
5894
5895 if( !progress && base_memory()->is_Phi() && can_reshape ) {
5896 // Check if PhiNode::Ideal's "Split phis through memory merges"
5897 // transform should be attempted. Look for this->phi->this cycle.
5898 uint merge_width = req();
5899 if (merge_width > Compile::AliasIdxRaw) {
5900 PhiNode* phi = base_memory()->as_Phi();
5901 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
5902 if (phi->in(i) == this) {
5903 phase->is_IterGVN()->_worklist.push(phi);
5904 break;
5905 }
5906 }
5907 }
5908 }
5909
5910 assert(progress || verify_sparse(), "please, no dups of base");
5911 return progress;
5912 }
5913
5914 //-------------------------set_base_memory-------------------------------------
5915 void MergeMemNode::set_base_memory(Node *new_base) {
5916 Node* empty_mem = empty_memory();
5917 set_req(Compile::AliasIdxBot, new_base);
5918 assert(memory_at(req()) == new_base, "must set default memory");
5919 // Clear out other occurrences of new_base:
5920 if (new_base != empty_mem) {
5921 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5922 if (in(i) == new_base) set_req(i, empty_mem);
5923 }
5924 }
5925 }
5926
5927 //------------------------------out_RegMask------------------------------------
5928 const RegMask &MergeMemNode::out_RegMask() const {
5929 return RegMask::EMPTY;
5930 }
5931
5932 //------------------------------dump_spec--------------------------------------
5933 #ifndef PRODUCT
5934 void MergeMemNode::dump_spec(outputStream *st) const {
5935 st->print(" {");
5936 Node* base_mem = base_memory();
5937 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
5938 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem;
5939 if (mem == base_mem) { st->print(" -"); continue; }
5940 st->print( " N%d:", mem->_idx );
5941 Compile::current()->get_adr_type(i)->dump_on(st);
5942 }
5943 st->print(" }");
5944 }
5945 #endif // !PRODUCT
5946
5947
5948 #ifdef ASSERT
5949 static bool might_be_same(Node* a, Node* b) {
5950 if (a == b) return true;
5951 if (!(a->is_Phi() || b->is_Phi())) return false;
5952 // phis shift around during optimization
5953 return true; // pretty stupid...
5954 }
5955
5956 // verify a narrow slice (either incoming or outgoing)
5957 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
5958 if (!VerifyAliases) return; // don't bother to verify unless requested
5959 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
5960 if (Node::in_dump()) return; // muzzle asserts when printing
5961 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
5962 assert(n != nullptr, "");
5963 // Elide intervening MergeMem's
5964 while (n->is_MergeMem()) {
5965 n = n->as_MergeMem()->memory_at(alias_idx);
5966 }
5967 Compile* C = Compile::current();
5968 const TypePtr* n_adr_type = n->adr_type();
5969 if (n == m->empty_memory()) {
5970 // Implicit copy of base_memory()
5971 } else if (n_adr_type != TypePtr::BOTTOM) {
5972 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type");
5973 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
5974 } else {
5975 // A few places like make_runtime_call "know" that VM calls are narrow,
5976 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
5977 bool expected_wide_mem = false;
5978 if (n == m->base_memory()) {
5979 expected_wide_mem = true;
5980 } else if (alias_idx == Compile::AliasIdxRaw ||
5981 n == m->memory_at(Compile::AliasIdxRaw)) {
5982 expected_wide_mem = true;
5983 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
5984 // memory can "leak through" calls on channels that
5985 // are write-once. Allow this also.
5986 expected_wide_mem = true;
5987 }
5988 assert(expected_wide_mem, "expected narrow slice replacement");
5989 }
5990 }
5991 #else // !ASSERT
5992 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
5993 #endif
5994
5995
5996 //-----------------------------memory_at---------------------------------------
5997 Node* MergeMemNode::memory_at(uint alias_idx) const {
5998 assert(alias_idx >= Compile::AliasIdxRaw ||
5999 (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()),
6000 "must avoid base_memory and AliasIdxTop");
6001
6002 // Otherwise, it is a narrow slice.
6003 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
6004 if (is_empty_memory(n)) {
6005 // the array is sparse; empty slots are the "top" node
6006 n = base_memory();
6007 assert(Node::in_dump()
6008 || n == nullptr || n->bottom_type() == Type::TOP
6009 || n->adr_type() == nullptr // address is TOP
6010 || n->adr_type() == TypePtr::BOTTOM
6011 || n->adr_type() == TypeRawPtr::BOTTOM
6012 || n->is_NarrowMemProj()
6013 || !Compile::current()->do_aliasing(),
6014 "must be a wide memory");
6015 // do_aliasing == false if we are organizing the memory states manually.
6016 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
6017 } else {
6018 // make sure the stored slice is sane
6019 #ifdef ASSERT
6020 if (VMError::is_error_reported() || Node::in_dump()) {
6021 } else if (might_be_same(n, base_memory())) {
6022 // Give it a pass: It is a mostly harmless repetition of the base.
6023 // This can arise normally from node subsumption during optimization.
6024 } else {
6025 verify_memory_slice(this, alias_idx, n);
6026 }
6027 #endif
6028 }
6029 return n;
6030 }
6031
6032 //---------------------------set_memory_at-------------------------------------
6033 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
6034 verify_memory_slice(this, alias_idx, n);
6035 Node* empty_mem = empty_memory();
6036 if (n == base_memory()) n = empty_mem; // collapse default
6037 uint need_req = alias_idx+1;
6038 if (req() < need_req) {
6039 if (n == empty_mem) return; // already the default, so do not grow me
6040 // grow the sparse array
6041 do {
6042 add_req(empty_mem);
6043 } while (req() < need_req);
6044 }
6045 set_req( alias_idx, n );
6046 }
6047
6048
6049
6050 //--------------------------iteration_setup------------------------------------
6051 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
6052 if (other != nullptr) {
6053 grow_to_match(other);
6054 // invariant: the finite support of mm2 is within mm->req()
6055 #ifdef ASSERT
6056 for (uint i = req(); i < other->req(); i++) {
6057 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
6058 }
6059 #endif
6060 }
6061 // Replace spurious copies of base_memory by top.
6062 Node* base_mem = base_memory();
6063 if (base_mem != nullptr && !base_mem->is_top()) {
6064 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
6065 if (in(i) == base_mem)
6066 set_req(i, empty_memory());
6067 }
6068 }
6069 }
6070
6071 //---------------------------grow_to_match-------------------------------------
6072 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
6073 Node* empty_mem = empty_memory();
6074 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
6075 // look for the finite support of the other memory
6076 for (uint i = other->req(); --i >= req(); ) {
6077 if (other->in(i) != empty_mem) {
6078 uint new_len = i+1;
6079 while (req() < new_len) add_req(empty_mem);
6080 break;
6081 }
6082 }
6083 }
6084
6085 //---------------------------verify_sparse-------------------------------------
6086 #ifndef PRODUCT
6087 bool MergeMemNode::verify_sparse() const {
6088 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
6089 Node* base_mem = base_memory();
6090 // The following can happen in degenerate cases, since empty==top.
6091 if (is_empty_memory(base_mem)) return true;
6092 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6093 assert(in(i) != nullptr, "sane slice");
6094 if (in(i) == base_mem) return false; // should have been the sentinel value!
6095 }
6096 return true;
6097 }
6098
6099 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
6100 Node* n;
6101 n = mm->in(idx);
6102 if (mem == n) return true; // might be empty_memory()
6103 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
6104 if (mem == n) return true;
6105 return false;
6106 }
6107 #endif // !PRODUCT