1 /*
2 * Copyright (c) 1997, 2026, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2024, Alibaba Group Holding Limited. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "classfile/javaClasses.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "gc/shared/tlab_globals.hpp"
31 #include "memory/allocation.inline.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "oops/objArrayKlass.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/arraycopynode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/compile.hpp"
38 #include "opto/connode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/loopnode.hpp"
41 #include "opto/machnode.hpp"
42 #include "opto/matcher.hpp"
43 #include "opto/memnode.hpp"
44 #include "opto/mempointer.hpp"
45 #include "opto/mulnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opcodes.hpp"
48 #include "opto/phaseX.hpp"
49 #include "opto/regalloc.hpp"
50 #include "opto/regmask.hpp"
51 #include "opto/rootnode.hpp"
52 #include "opto/traceMergeStoresTag.hpp"
53 #include "opto/vectornode.hpp"
54 #include "utilities/align.hpp"
55 #include "utilities/copy.hpp"
56 #include "utilities/macros.hpp"
57 #include "utilities/powerOfTwo.hpp"
58 #include "utilities/vmError.hpp"
59
60 // Portions of code courtesy of Clifford Click
61
62 // Optimization - Graph Style
63
64 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
65
66 //=============================================================================
67 uint MemNode::size_of() const { return sizeof(*this); }
68
69 const TypePtr *MemNode::adr_type() const {
70 Node* adr = in(Address);
71 if (adr == nullptr) return nullptr; // node is dead
72 const TypePtr* cross_check = nullptr;
73 DEBUG_ONLY(cross_check = _adr_type);
74 return calculate_adr_type(adr->bottom_type(), cross_check);
75 }
76
77 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
78 if (adr != nullptr) {
79 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
80 return true;
81 }
82 }
83 return false;
84 }
85
86 #ifndef PRODUCT
87 void MemNode::dump_spec(outputStream *st) const {
88 if (in(Address) == nullptr) return; // node is dead
89 #ifndef ASSERT
90 // fake the missing field
91 const TypePtr* _adr_type = nullptr;
92 if (in(Address) != nullptr)
93 _adr_type = in(Address)->bottom_type()->isa_ptr();
94 #endif
95 dump_adr_type(_adr_type, st);
96
97 Compile* C = Compile::current();
98 if (C->alias_type(_adr_type)->is_volatile()) {
99 st->print(" Volatile!");
100 }
101 if (_unaligned_access) {
102 st->print(" unaligned");
103 }
104 if (_mismatched_access) {
105 st->print(" mismatched");
106 }
107 if (_unsafe_access) {
108 st->print(" unsafe");
109 }
110 }
111
112 void MemNode::dump_adr_type(const TypePtr* adr_type, outputStream* st) {
113 st->print(" @");
114 if (adr_type == nullptr) {
115 st->print("null");
116 } else {
117 adr_type->dump_on(st);
118 Compile* C = Compile::current();
119 Compile::AliasType* atp = nullptr;
120 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
121 if (atp == nullptr)
122 st->print(", idx=?\?;");
123 else if (atp->index() == Compile::AliasIdxBot)
124 st->print(", idx=Bot;");
125 else if (atp->index() == Compile::AliasIdxTop)
126 st->print(", idx=Top;");
127 else if (atp->index() == Compile::AliasIdxRaw)
128 st->print(", idx=Raw;");
129 else {
130 ciField* field = atp->field();
131 if (field) {
132 st->print(", name=");
133 field->print_name_on(st);
134 }
135 st->print(", idx=%d;", atp->index());
136 }
137 }
138 }
139
140 extern void print_alias_types();
141
142 #endif
143
144 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
145 assert((t_oop != nullptr), "sanity");
146 bool is_instance = t_oop->is_known_instance_field();
147 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
148 (load != nullptr) && load->is_Load() &&
149 (phase->is_IterGVN() != nullptr);
150 if (!(is_instance || is_boxed_value_load))
151 return mchain; // don't try to optimize non-instance types
152 uint instance_id = t_oop->instance_id();
153 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
154 Node *prev = nullptr;
155 Node *result = mchain;
156 while (prev != result) {
157 prev = result;
158 if (result == start_mem)
159 break; // hit one of our sentinels
160 // skip over a call which does not affect this memory slice
161 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
162 Node *proj_in = result->in(0);
163 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
164 break; // hit one of our sentinels
165 } else if (proj_in->is_Call()) {
166 // ArrayCopyNodes processed here as well
167 CallNode *call = proj_in->as_Call();
168 if (!call->may_modify(t_oop, phase)) { // returns false for instances
169 result = call->in(TypeFunc::Memory);
170 }
171 } else if (proj_in->is_Initialize()) {
172 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
173 // Stop if this is the initialization for the object instance which
174 // contains this memory slice, otherwise skip over it.
175 if ((alloc == nullptr) || (alloc->_idx == instance_id)) {
176 break;
177 }
178 if (is_instance) {
179 result = proj_in->in(TypeFunc::Memory);
180 } else if (is_boxed_value_load) {
181 Node* klass = alloc->in(AllocateNode::KlassNode);
182 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
183 if (tklass->klass_is_exact() && !tklass->exact_klass()->equals(t_oop->is_instptr()->exact_klass())) {
184 result = proj_in->in(TypeFunc::Memory); // not related allocation
185 }
186 }
187 } else if (proj_in->is_MemBar()) {
188 ArrayCopyNode* ac = nullptr;
189 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
190 break;
191 }
192 result = proj_in->in(TypeFunc::Memory);
193 } else if (proj_in->is_top()) {
194 break; // dead code
195 } else {
196 assert(false, "unexpected projection");
197 }
198 } else if (result->is_ClearArray()) {
199 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
200 // Can not bypass initialization of the instance
201 // we are looking for.
202 break;
203 }
204 // Otherwise skip it (the call updated 'result' value).
205 } else if (result->is_MergeMem()) {
206 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, nullptr, tty);
207 }
208 }
209 return result;
210 }
211
212 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
213 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
214 if (t_oop == nullptr)
215 return mchain; // don't try to optimize non-oop types
216 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
217 bool is_instance = t_oop->is_known_instance_field();
218 PhaseIterGVN *igvn = phase->is_IterGVN();
219 if (is_instance && igvn != nullptr && result->is_Phi()) {
220 PhiNode *mphi = result->as_Phi();
221 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
222 const TypePtr *t = mphi->adr_type();
223 bool do_split = false;
224 // In the following cases, Load memory input can be further optimized based on
225 // its precise address type
226 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ) {
227 do_split = true;
228 } else if (t->isa_oopptr() && !t->is_oopptr()->is_known_instance()) {
229 const TypeOopPtr* mem_t =
230 t->is_oopptr()->cast_to_exactness(true)
231 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
232 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id());
233 if (t_oop->isa_aryptr()) {
234 mem_t = mem_t->is_aryptr()
235 ->cast_to_stable(t_oop->is_aryptr()->is_stable())
236 ->cast_to_size(t_oop->is_aryptr()->size())
237 ->with_offset(t_oop->is_aryptr()->offset())
238 ->is_aryptr();
239 }
240 do_split = mem_t == t_oop;
241 }
242 if (do_split) {
243 // clone the Phi with our address type
244 result = mphi->split_out_instance(t_adr, igvn);
245 } else {
246 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
247 }
248 }
249 return result;
250 }
251
252 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
253 uint alias_idx = phase->C->get_alias_index(tp);
254 Node *mem = mmem;
255 #ifdef ASSERT
256 {
257 // Check that current type is consistent with the alias index used during graph construction
258 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
259 bool consistent = adr_check == nullptr || adr_check->empty() ||
260 phase->C->must_alias(adr_check, alias_idx );
261 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
262 if( !consistent && adr_check != nullptr && !adr_check->empty() &&
263 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
264 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
265 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
266 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
267 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
268 // don't assert if it is dead code.
269 consistent = true;
270 }
271 if( !consistent ) {
272 st->print("alias_idx==%d, adr_check==", alias_idx);
273 if( adr_check == nullptr ) {
274 st->print("null");
275 } else {
276 adr_check->dump();
277 }
278 st->cr();
279 print_alias_types();
280 assert(consistent, "adr_check must match alias idx");
281 }
282 }
283 #endif
284 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
285 // means an array I have not precisely typed yet. Do not do any
286 // alias stuff with it any time soon.
287 const TypeOopPtr *toop = tp->isa_oopptr();
288 if (tp->base() != Type::AnyPtr &&
289 !(toop &&
290 toop->isa_instptr() &&
291 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
292 toop->offset() == Type::OffsetBot)) {
293 // IGVN _delay_transform may be set to true and if that is the case and mmem
294 // is already a registered node then the validation inside transform will
295 // complain.
296 Node* m = mmem;
297 PhaseIterGVN* igvn = phase->is_IterGVN();
298 if (igvn == nullptr || !igvn->delay_transform()) {
299 // compress paths and change unreachable cycles to TOP
300 // If not, we can update the input infinitely along a MergeMem cycle
301 // Equivalent code in PhiNode::Ideal
302 m = phase->transform(mmem);
303 }
304 // If transformed to a MergeMem, get the desired slice
305 // Otherwise the returned node represents memory for every slice
306 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
307 // Update input if it is progress over what we have now
308 }
309 return mem;
310 }
311
312 //--------------------------Ideal_common---------------------------------------
313 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
314 // Unhook non-raw memories from complete (macro-expanded) initializations.
315 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
316 // If our control input is a dead region, kill all below the region
317 Node *ctl = in(MemNode::Control);
318 if (ctl && remove_dead_region(phase, can_reshape))
319 return this;
320 ctl = in(MemNode::Control);
321 // Don't bother trying to transform a dead node
322 if (ctl && ctl->is_top()) return NodeSentinel;
323
324 PhaseIterGVN *igvn = phase->is_IterGVN();
325 // Wait if control on the worklist.
326 if (ctl && can_reshape && igvn != nullptr) {
327 Node* bol = nullptr;
328 Node* cmp = nullptr;
329 if (ctl->in(0)->is_If()) {
330 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
331 bol = ctl->in(0)->in(1);
332 if (bol->is_Bool())
333 cmp = ctl->in(0)->in(1)->in(1);
334 }
335 if (igvn->_worklist.member(ctl) ||
336 (bol != nullptr && igvn->_worklist.member(bol)) ||
337 (cmp != nullptr && igvn->_worklist.member(cmp)) ) {
338 // This control path may be dead.
339 // Delay this memory node transformation until the control is processed.
340 igvn->_worklist.push(this);
341 return NodeSentinel; // caller will return null
342 }
343 }
344 // Ignore if memory is dead, or self-loop
345 Node *mem = in(MemNode::Memory);
346 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return null
347 assert(mem != this, "dead loop in MemNode::Ideal");
348
349 if (can_reshape && igvn != nullptr && igvn->_worklist.member(mem)) {
350 // This memory slice may be dead.
351 // Delay this mem node transformation until the memory is processed.
352 igvn->_worklist.push(this);
353 return NodeSentinel; // caller will return null
354 }
355
356 Node *address = in(MemNode::Address);
357 const Type *t_adr = phase->type(address);
358 if (t_adr == Type::TOP) return NodeSentinel; // caller will return null
359
360 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
361 // Unsafe off-heap access with zero address. Remove access and other control users
362 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
363 assert(ctl != nullptr, "unsafe accesses should be control dependent");
364 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
365 Node* u = ctl->fast_out(i);
366 if (u != ctl) {
367 igvn->rehash_node_delayed(u);
368 int nb = u->replace_edge(ctl, phase->C->top(), igvn);
369 --i, imax -= nb;
370 }
371 }
372 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
373 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
374 phase->C->root()->add_req(halt);
375 return this;
376 }
377
378 if (can_reshape && igvn != nullptr &&
379 (igvn->_worklist.member(address) ||
380 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
381 // The address's base and type may change when the address is processed.
382 // Delay this mem node transformation until the address is processed.
383 igvn->_worklist.push(this);
384 return NodeSentinel; // caller will return null
385 }
386
387 // Do NOT remove or optimize the next lines: ensure a new alias index
388 // is allocated for an oop pointer type before Escape Analysis.
389 // Note: C++ will not remove it since the call has side effect.
390 if (t_adr->isa_oopptr()) {
391 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
392 }
393
394 Node* base = nullptr;
395 if (address->is_AddP()) {
396 base = address->in(AddPNode::Base);
397 }
398 if (base != nullptr && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
399 !t_adr->isa_rawptr()) {
400 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
401 // Skip this node optimization if its address has TOP base.
402 return NodeSentinel; // caller will return null
403 }
404
405 // Avoid independent memory operations
406 Node* old_mem = mem;
407
408 // The code which unhooks non-raw memories from complete (macro-expanded)
409 // initializations was removed. After macro-expansion all stores caught
410 // by Initialize node became raw stores and there is no information
411 // which memory slices they modify. So it is unsafe to move any memory
412 // operation above these stores. Also in most cases hooked non-raw memories
413 // were already unhooked by using information from detect_ptr_independence()
414 // and find_previous_store().
415
416 if (mem->is_MergeMem()) {
417 MergeMemNode* mmem = mem->as_MergeMem();
418 const TypePtr *tp = t_adr->is_ptr();
419
420 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
421 }
422
423 if (mem != old_mem) {
424 set_req_X(MemNode::Memory, mem, phase);
425 if (phase->type(mem) == Type::TOP) return NodeSentinel;
426 return this;
427 }
428
429 // let the subclass continue analyzing...
430 return nullptr;
431 }
432
433 // Helper function for proving some simple control dominations.
434 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
435 // Already assumes that 'dom' is available at 'sub', and that 'sub'
436 // is not a constant (dominated by the method's StartNode).
437 // Used by MemNode::find_previous_store to prove that the
438 // control input of a memory operation predates (dominates)
439 // an allocation it wants to look past.
440 // Returns 'DomResult::Dominate' if all control inputs of 'dom'
441 // dominate 'sub', 'DomResult::NotDominate' if not,
442 // and 'DomResult::EncounteredDeadCode' if we can't decide due to
443 // dead code, but at the end of IGVN, we know the definite result
444 // once the dead code is cleaned up.
445 Node::DomResult MemNode::maybe_all_controls_dominate(Node* dom, Node* sub) {
446 if (dom == nullptr || dom->is_top() || sub == nullptr || sub->is_top()) {
447 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
448 }
449
450 // Check 'dom'. Skip Proj and CatchProj nodes.
451 dom = dom->find_exact_control(dom);
452 if (dom == nullptr || dom->is_top()) {
453 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
454 }
455
456 if (dom == sub) {
457 // For the case when, for example, 'sub' is Initialize and the original
458 // 'dom' is Proj node of the 'sub'.
459 return DomResult::NotDominate;
460 }
461
462 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) {
463 return DomResult::Dominate;
464 }
465
466 // 'dom' dominates 'sub' if its control edge and control edges
467 // of all its inputs dominate or equal to sub's control edge.
468
469 // Currently 'sub' is either Allocate, Initialize or Start nodes.
470 // Or Region for the check in LoadNode::Ideal();
471 // 'sub' should have sub->in(0) != nullptr.
472 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
473 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
474
475 // Get control edge of 'sub'.
476 Node* orig_sub = sub;
477 sub = sub->find_exact_control(sub->in(0));
478 if (sub == nullptr || sub->is_top()) {
479 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
480 }
481
482 assert(sub->is_CFG(), "expecting control");
483
484 if (sub == dom) {
485 return DomResult::Dominate;
486 }
487
488 if (sub->is_Start() || sub->is_Root()) {
489 return DomResult::NotDominate;
490 }
491
492 {
493 // Check all control edges of 'dom'.
494
495 ResourceMark rm;
496 Node_List nlist;
497 Unique_Node_List dom_list;
498
499 dom_list.push(dom);
500 bool only_dominating_controls = false;
501
502 for (uint next = 0; next < dom_list.size(); next++) {
503 Node* n = dom_list.at(next);
504 if (n == orig_sub) {
505 return DomResult::NotDominate; // One of dom's inputs dominated by sub.
506 }
507 if (!n->is_CFG() && n->pinned()) {
508 // Check only own control edge for pinned non-control nodes.
509 n = n->find_exact_control(n->in(0));
510 if (n == nullptr || n->is_top()) {
511 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
512 }
513 assert(n->is_CFG(), "expecting control");
514 dom_list.push(n);
515 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
516 only_dominating_controls = true;
517 } else if (n->is_CFG()) {
518 DomResult dom_result = n->dominates(sub, nlist);
519 if (dom_result == DomResult::Dominate) {
520 only_dominating_controls = true;
521 } else {
522 return dom_result;
523 }
524 } else {
525 // First, own control edge.
526 Node* m = n->find_exact_control(n->in(0));
527 if (m != nullptr) {
528 if (m->is_top()) {
529 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
530 }
531 dom_list.push(m);
532 }
533 // Now, the rest of edges.
534 uint cnt = n->req();
535 for (uint i = 1; i < cnt; i++) {
536 m = n->find_exact_control(n->in(i));
537 if (m == nullptr || m->is_top()) {
538 continue;
539 }
540 dom_list.push(m);
541 }
542 }
543 }
544 return only_dominating_controls ? DomResult::Dominate : DomResult::NotDominate;
545 }
546 }
547
548 //---------------------detect_ptr_independence---------------------------------
549 // Used by MemNode::find_previous_store to prove that two base
550 // pointers are never equal.
551 // The pointers are accompanied by their associated allocations,
552 // if any, which have been previously discovered by the caller.
553 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
554 Node* p2, AllocateNode* a2,
555 PhaseTransform* phase) {
556 // Trivial case: Non-overlapping values. Be careful, we can cast a raw pointer to an oop (e.g. in
557 // the allocation pattern) so joining the types only works if both are oops. join may also give
558 // an incorrect result when both pointers are nullable and the result is supposed to be
559 // TypePtr::NULL_PTR, so we exclude that case.
560 const Type* p1_type = p1->bottom_type();
561 const Type* p2_type = p2->bottom_type();
562 if (p1_type->isa_oopptr() && p2_type->isa_oopptr() &&
563 (!p1_type->maybe_null() || !p2_type->maybe_null()) &&
564 p1_type->join(p2_type)->empty()) {
565 return true;
566 }
567
568 // Attempt to prove that these two pointers cannot be aliased.
569 // They may both manifestly be allocations, and they should differ.
570 // Or, if they are not both allocations, they can be distinct constants.
571 // Otherwise, one is an allocation and the other a pre-existing value.
572 if (a1 == nullptr && a2 == nullptr) { // neither an allocation
573 return (p1 != p2) && p1->is_Con() && p2->is_Con();
574 } else if (a1 != nullptr && a2 != nullptr) { // both allocations
575 return (a1 != a2);
576 } else if (a1 != nullptr) { // one allocation a1
577 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
578 return all_controls_dominate(p2, a1);
579 } else { //(a2 != null) // one allocation a2
580 return all_controls_dominate(p1, a2);
581 }
582 return false;
583 }
584
585 // Find an arraycopy ac that produces the memory state represented by parameter mem.
586 // Return ac if
587 // (a) can_see_stored_value=true and ac must have set the value for this load or if
588 // (b) can_see_stored_value=false and ac could have set the value for this load or if
589 // (c) can_see_stored_value=false and ac cannot have set the value for this load.
590 // In case (c) change the parameter mem to the memory input of ac to skip it
591 // when searching stored value.
592 // Otherwise return null.
593 Node* LoadNode::find_previous_arraycopy(PhaseValues* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
594 ArrayCopyNode* ac = find_array_copy_clone(ld_alloc, mem);
595 if (ac != nullptr) {
596 Node* ld_addp = in(MemNode::Address);
597 Node* src = ac->in(ArrayCopyNode::Src);
598 const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
599
600 // This is a load from a cloned array. The corresponding arraycopy ac must
601 // have set the value for the load and we can return ac but only if the load
602 // is known to be within bounds. This is checked below.
603 if (ary_t != nullptr && ld_addp->is_AddP()) {
604 Node* ld_offs = ld_addp->in(AddPNode::Offset);
605 BasicType ary_elem = ary_t->elem()->array_element_basic_type();
606 jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
607 jlong elemsize = type2aelembytes(ary_elem);
608
609 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
610 const TypeInt* sizetype = ary_t->size();
611
612 if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
613 // The load is known to be within bounds. It receives its value from ac.
614 return ac;
615 }
616 // The load is known to be out-of-bounds.
617 }
618 // The load could be out-of-bounds. It must not be hoisted but must remain
619 // dependent on the runtime range check. This is achieved by returning null.
620 } else if (mem->is_Proj() && mem->in(0) != nullptr && mem->in(0)->is_ArrayCopy()) {
621 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
622
623 if (ac->is_arraycopy_validated() ||
624 ac->is_copyof_validated() ||
625 ac->is_copyofrange_validated()) {
626 Node* ld_addp = in(MemNode::Address);
627 if (ld_addp->is_AddP()) {
628 Node* ld_base = ld_addp->in(AddPNode::Address);
629 Node* ld_offs = ld_addp->in(AddPNode::Offset);
630
631 Node* dest = ac->in(ArrayCopyNode::Dest);
632
633 if (dest == ld_base) {
634 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
635 assert(!ld_offs_t->empty(), "dead reference should be checked already");
636 // Take into account vector or unsafe access size
637 jlong ld_size_in_bytes = (jlong)memory_size();
638 jlong offset_hi = ld_offs_t->_hi + ld_size_in_bytes - 1;
639 offset_hi = MIN2(offset_hi, (jlong)(TypeX::MAX->_hi)); // Take care for overflow in 32-bit VM
640 if (ac->modifies(ld_offs_t->_lo, (intptr_t)offset_hi, phase, can_see_stored_value)) {
641 return ac;
642 }
643 if (!can_see_stored_value) {
644 mem = ac->in(TypeFunc::Memory);
645 return ac;
646 }
647 }
648 }
649 }
650 }
651 return nullptr;
652 }
653
654 ArrayCopyNode* MemNode::find_array_copy_clone(Node* ld_alloc, Node* mem) const {
655 if (mem->is_Proj() && mem->in(0) != nullptr && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
656 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
657 if (ld_alloc != nullptr) {
658 // Check if there is an array copy for a clone
659 Node* mb = mem->in(0);
660 ArrayCopyNode* ac = nullptr;
661 if (mb->in(0) != nullptr && mb->in(0)->is_Proj() &&
662 mb->in(0)->in(0) != nullptr && mb->in(0)->in(0)->is_ArrayCopy()) {
663 ac = mb->in(0)->in(0)->as_ArrayCopy();
664 } else {
665 // Step over GC barrier when ReduceInitialCardMarks is disabled
666 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
667 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
668
669 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
670 ac = control_proj_ac->in(0)->as_ArrayCopy();
671 }
672 }
673
674 if (ac != nullptr && ac->is_clonebasic()) {
675 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest));
676 if (alloc != nullptr && alloc == ld_alloc) {
677 return ac;
678 }
679 }
680 }
681 }
682 return nullptr;
683 }
684
685 // The logic for reordering loads and stores uses four steps:
686 // (a) Walk carefully past stores and initializations which we
687 // can prove are independent of this load.
688 // (b) Observe that the next memory state makes an exact match
689 // with self (load or store), and locate the relevant store.
690 // (c) Ensure that, if we were to wire self directly to the store,
691 // the optimizer would fold it up somehow.
692 // (d) Do the rewiring, and return, depending on some other part of
693 // the optimizer to fold up the load.
694 // This routine handles steps (a) and (b). Steps (c) and (d) are
695 // specific to loads and stores, so they are handled by the callers.
696 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
697 //
698 Node* MemNode::find_previous_store(PhaseValues* phase) {
699 AccessAnalyzer analyzer(phase, this);
700
701 Node* mem = in(MemNode::Memory); // start searching here...
702 int cnt = 50; // Cycle limiter
703 for (;; cnt--) {
704 // While we can dance past unrelated stores...
705 if (phase->type(mem) == Type::TOP) {
706 // Encounter a dead node
707 return phase->C->top();
708 } else if (cnt <= 0) {
709 // Caught in cycle or a complicated dance?
710 return nullptr;
711 } else if (mem->is_Phi()) {
712 return nullptr;
713 }
714
715 AccessAnalyzer::AccessIndependence independence = analyzer.detect_access_independence(mem);
716 if (independence.independent) {
717 // (a) advance through the independent store
718 mem = independence.mem;
719 assert(mem != nullptr, "must not be nullptr");
720 } else {
721 // (b) found the store that this access observes if this is not null
722 // Otherwise, give up if it is null
723 return independence.mem;
724 }
725 }
726
727 return nullptr; // bail out
728 }
729
730 //----------------------calculate_adr_type-------------------------------------
731 // Helper function. Notices when the given type of address hits top or bottom.
732 // Also, asserts a cross-check of the type against the expected address type.
733 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
734 if (t == Type::TOP) return nullptr; // does not touch memory any more?
735 #ifdef ASSERT
736 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = nullptr;
737 #endif
738 const TypePtr* tp = t->isa_ptr();
739 if (tp == nullptr) {
740 assert(cross_check == nullptr || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
741 return TypePtr::BOTTOM; // touches lots of memory
742 } else {
743 #ifdef ASSERT
744 // %%%% [phh] We don't check the alias index if cross_check is
745 // TypeRawPtr::BOTTOM. Needs to be investigated.
746 if (cross_check != nullptr &&
747 cross_check != TypePtr::BOTTOM &&
748 cross_check != TypeRawPtr::BOTTOM) {
749 // Recheck the alias index, to see if it has changed (due to a bug).
750 Compile* C = Compile::current();
751 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
752 "must stay in the original alias category");
753 // The type of the address must be contained in the adr_type,
754 // disregarding "null"-ness.
755 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
756 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
757 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
758 "real address must not escape from expected memory type");
759 }
760 #endif
761 return tp;
762 }
763 }
764
765 uint8_t MemNode::barrier_data(const Node* n) {
766 if (n->is_LoadStore()) {
767 return n->as_LoadStore()->barrier_data();
768 } else if (n->is_Mem()) {
769 return n->as_Mem()->barrier_data();
770 } else if (n->is_DecodeN() && n->in(1)->is_Load()) {
771 return n->in(1)->as_Load()->barrier_data();
772 }
773 return 0;
774 }
775
776 MemNode::MemOrd MemNode::memory_order(const Node* n) {
777 if (n->is_Load()) {
778 return n->as_Load()->memory_order();
779 } else if (n->is_Store()) {
780 return n->as_Store()->memory_order();;
781 }
782 return MemNode::MemOrd::seqcst;
783 }
784
785 AccessAnalyzer::AccessAnalyzer(PhaseValues* phase, MemNode* n)
786 : _phase(phase), _n(n), _memory_size(n->memory_size()), _alias_idx(-1) {
787 Node* adr = _n->in(MemNode::Address);
788 _offset = 0;
789 _base = AddPNode::Ideal_base_and_offset(adr, _phase, _offset);
790 _maybe_raw = MemNode::check_if_adr_maybe_raw(adr);
791 _alloc = AllocateNode::Ideal_allocation(_base);
792 _adr_type = _n->adr_type();
793
794 if (_adr_type != nullptr && _adr_type->base() != TypePtr::AnyPtr) {
795 // Avoid the cases that will upset Compile::get_alias_index
796 _alias_idx = _phase->C->get_alias_index(_adr_type);
797 assert(_alias_idx != Compile::AliasIdxTop, "must not be a dead node");
798 assert(_alias_idx != Compile::AliasIdxBot || !phase->C->do_aliasing(), "must not be a very wide access");
799 }
800 }
801
802 // Decide whether the memory accessed by '_n' and 'other' may overlap. This function may be used
803 // when we want to walk the memory graph to fold a load, or when we want to hoist a load above a
804 // loop when there are no stores that may overlap with the load inside the loop.
805 AccessAnalyzer::AccessIndependence AccessAnalyzer::detect_access_independence(Node* other) const {
806 assert(_phase->type(other) == Type::MEMORY, "must be a memory node %s", other->Name());
807 assert(!other->is_Phi(), "caller must handle Phi");
808
809 if (_adr_type == nullptr) {
810 // This means the access is dead
811 return {false, _phase->C->top()};
812 } else if (_adr_type->base() == TypePtr::AnyPtr) {
813 // An example for this case is an access into the memory address 0 performed using Unsafe
814 assert(_adr_type->ptr() == TypePtr::Null, "MemNode should never access a wide memory");
815 return {false, nullptr};
816 }
817
818 if (_offset == Type::OffsetBot) {
819 // cannot unalias unless there are precise offsets
820 return {false, nullptr};
821 }
822
823 const TypeOopPtr* adr_oop_type = _adr_type->isa_oopptr();
824 Node* prev = other;
825 if (other->is_Store()) {
826 Node* st_adr = other->in(MemNode::Address);
827 intptr_t st_offset = 0;
828 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, _phase, st_offset);
829 if (st_base == nullptr) {
830 // inscrutable pointer
831 return {false, nullptr};
832 }
833
834 // If the bases are the same and the offsets are the same, it seems that this is the exact
835 // store we are looking for, the caller will check if the type of the store matches using
836 // MemNode::can_see_stored_value
837 if (st_base == _base && st_offset == _offset) {
838 return {false, other};
839 }
840
841 // If it is provable that the memory accessed by 'other' does not overlap the memory accessed
842 // by '_n', we may walk past 'other'.
843 // For raw accesses, 2 accesses are independent if they have the same base and the offsets
844 // say that they do not overlap.
845 // For heap accesses, 2 accesses are independent if either the bases are provably different
846 // at runtime or the offsets say that the accesses do not overlap.
847 if ((_maybe_raw || MemNode::check_if_adr_maybe_raw(st_adr)) && st_base != _base) {
848 // Raw accesses can only be provably independent if they have the same base
849 return {false, nullptr};
850 }
851
852 // If the offsets say that the accesses do not overlap, then it is provable that 'other' and
853 // '_n' do not overlap. For example, a LoadI from Object+8 is independent from a StoreL into
854 // Object+12, no matter what the bases are.
855 if (st_offset != _offset && st_offset != Type::OffsetBot) {
856 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
857 assert(other->as_Store()->memory_size() <= MAX_STORE, "");
858 if (st_offset >= _offset + _memory_size ||
859 st_offset <= _offset - MAX_STORE ||
860 st_offset <= _offset - other->as_Store()->memory_size()) {
861 // Success: The offsets are provably independent.
862 // (You may ask, why not just test st_offset != offset and be done?
863 // The answer is that stores of different sizes can co-exist
864 // in the same sequence of RawMem effects. We sometimes initialize
865 // a whole 'tile' of array elements with a single jint or jlong.)
866 return {true, other->in(MemNode::Memory)};
867 }
868 }
869
870 // Same base and overlapping offsets, it seems provable that the accesses overlap, give up
871 if (st_base == _base) {
872 return {false, nullptr};
873 }
874
875 // Try to prove that 2 different base nodes at compile time are different values at runtime
876 bool known_independent = false;
877 if (MemNode::detect_ptr_independence(_base, _alloc, st_base, AllocateNode::Ideal_allocation(st_base), _phase)) {
878 known_independent = true;
879 }
880
881 if (known_independent) {
882 return {true, other->in(MemNode::Memory)};
883 }
884 } else if (other->is_Proj() && other->in(0)->is_Initialize()) {
885 InitializeNode* st_init = other->in(0)->as_Initialize();
886 AllocateNode* st_alloc = st_init->allocation();
887 if (st_alloc == nullptr) {
888 // Something degenerated
889 return {false, nullptr};
890 }
891 bool known_identical = false;
892 bool known_independent = false;
893 if (_alloc == st_alloc) {
894 known_identical = true;
895 } else if (_alloc != nullptr) {
896 known_independent = true;
897 } else if (MemNode::all_controls_dominate(_n, st_alloc)) {
898 known_independent = true;
899 }
900
901 if (known_independent) {
902 // The bases are provably independent: Either they are
903 // manifestly distinct allocations, or else the control
904 // of _n dominates the store's allocation.
905 if (_alias_idx == Compile::AliasIdxRaw) {
906 other = st_alloc->in(TypeFunc::Memory);
907 } else {
908 other = st_init->memory(_alias_idx);
909 }
910 return {true, other};
911 }
912
913 // If we are not looking at a store initializing the same
914 // allocation we are loading from, we lose.
915 if (known_identical) {
916 // From caller, can_see_stored_value will consult find_captured_store.
917 return {false, other};
918 }
919
920 } else if (_n->find_previous_arraycopy(_phase, _alloc, other, false) != nullptr) {
921 // Find an arraycopy that may or may not affect the MemNode
922 return {prev != other, other};
923 } else if (other->is_MergeMem()) {
924 return {true, other->as_MergeMem()->memory_at(_alias_idx)};
925 } else if (adr_oop_type != nullptr && adr_oop_type->is_known_instance_field()) {
926 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
927 if (other->is_Proj() && other->in(0)->is_Call()) {
928 // ArrayCopyNodes processed here as well.
929 CallNode* call = other->in(0)->as_Call();
930 if (!call->may_modify(adr_oop_type, _phase)) {
931 return {true, call->in(TypeFunc::Memory)};
932 }
933 } else if (other->is_Proj() && other->in(0)->is_MemBar()) {
934 ArrayCopyNode* ac = nullptr;
935 if (!ArrayCopyNode::may_modify(adr_oop_type, other->in(0)->as_MemBar(), _phase, ac)) {
936 return {true, other->in(0)->in(TypeFunc::Memory)};
937 }
938 } else if (other->is_ClearArray()) {
939 if (ClearArrayNode::step_through(&other, (uint)adr_oop_type->instance_id(), _phase)) {
940 // (the call updated 'other' value)
941 return {true, other};
942 } else {
943 // Can not bypass initialization of the instance
944 // we are looking for.
945 return {false, other};
946 }
947 }
948 }
949
950 return {false, nullptr};
951 }
952
953 //=============================================================================
954 // Should LoadNode::Ideal() attempt to remove control edges?
955 bool LoadNode::can_remove_control() const {
956 return !has_pinned_control_dependency();
957 }
958 uint LoadNode::size_of() const { return sizeof(*this); }
959 bool LoadNode::cmp(const Node &n) const {
960 LoadNode& load = (LoadNode &)n;
961 return Type::equals(_type, load._type) &&
962 _control_dependency == load._control_dependency &&
963 _mo == load._mo;
964 }
965 const Type *LoadNode::bottom_type() const { return _type; }
966 uint LoadNode::ideal_reg() const {
967 return _type->ideal_reg();
968 }
969
970 #ifndef PRODUCT
971 void LoadNode::dump_spec(outputStream *st) const {
972 MemNode::dump_spec(st);
973 if( !Verbose && !WizardMode ) {
974 // standard dump does this in Verbose and WizardMode
975 st->print(" #"); _type->dump_on(st);
976 }
977 if (in(0) != nullptr && !depends_only_on_test()) {
978 st->print(" (does not depend only on test, ");
979 if (control_dependency() == UnknownControl) {
980 st->print("unknown control");
981 } else if (control_dependency() == Pinned) {
982 st->print("pinned");
983 } else if (adr_type() == TypeRawPtr::BOTTOM) {
984 st->print("raw access");
985 } else {
986 st->print("unknown reason");
987 }
988 st->print(")");
989 }
990 }
991 #endif
992
993 #ifdef ASSERT
994 //----------------------------is_immutable_value-------------------------------
995 // Helper function to allow a raw load without control edge for some cases
996 bool LoadNode::is_immutable_value(Node* adr) {
997 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
998 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) {
999
1000 jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1);
1001 int offsets[] = {
1002 in_bytes(JavaThread::osthread_offset()),
1003 in_bytes(JavaThread::threadObj_offset()),
1004 in_bytes(JavaThread::vthread_offset()),
1005 in_bytes(JavaThread::scopedValueCache_offset()),
1006 };
1007
1008 for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) {
1009 if (offset == offsets[i]) {
1010 return true;
1011 }
1012 }
1013 }
1014
1015 return false;
1016 }
1017 #endif
1018
1019 //----------------------------LoadNode::make-----------------------------------
1020 // Polymorphic factory method:
1021 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo,
1022 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
1023 Compile* C = gvn.C;
1024
1025 // sanity check the alias category against the created node type
1026 assert(!(adr_type->isa_oopptr() &&
1027 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
1028 "use LoadKlassNode instead");
1029 assert(!(adr_type->isa_aryptr() &&
1030 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
1031 "use LoadRangeNode instead");
1032 // Check control edge of raw loads
1033 assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
1034 // oop will be recorded in oop map if load crosses safepoint
1035 rt->isa_oopptr() || is_immutable_value(adr),
1036 "raw memory operations should have control edge");
1037 LoadNode* load = nullptr;
1038 switch (bt) {
1039 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1040 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1041 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1042 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1043 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1044 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break;
1045 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
1046 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break;
1047 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
1048 case T_OBJECT:
1049 case T_NARROWOOP:
1050 #ifdef _LP64
1051 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1052 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
1053 } else
1054 #endif
1055 {
1056 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
1057 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
1058 }
1059 break;
1060 default:
1061 ShouldNotReachHere();
1062 break;
1063 }
1064 assert(load != nullptr, "LoadNode should have been created");
1065 if (unaligned) {
1066 load->set_unaligned_access();
1067 }
1068 if (mismatched) {
1069 load->set_mismatched_access();
1070 }
1071 if (unsafe) {
1072 load->set_unsafe_access();
1073 }
1074 load->set_barrier_data(barrier_data);
1075 if (load->Opcode() == Op_LoadN) {
1076 Node* ld = gvn.transform(load);
1077 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
1078 }
1079
1080 return load;
1081 }
1082
1083 //------------------------------hash-------------------------------------------
1084 uint LoadNode::hash() const {
1085 // unroll addition of interesting fields
1086 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
1087 }
1088
1089 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
1090 if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) {
1091 bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile();
1092 bool is_stable_ary = FoldStableValues &&
1093 (tp != nullptr) && (tp->isa_aryptr() != nullptr) &&
1094 tp->isa_aryptr()->is_stable();
1095
1096 return (eliminate_boxing && non_volatile) || is_stable_ary;
1097 }
1098
1099 return false;
1100 }
1101
1102 // Is the value loaded previously stored by an arraycopy? If so return
1103 // a load node that reads from the source array so we may be able to
1104 // optimize out the ArrayCopy node later.
1105 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
1106 Node* ld_adr = in(MemNode::Address);
1107 intptr_t ld_off = 0;
1108 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
1109 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
1110 if (ac != nullptr) {
1111 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
1112
1113 Node* mem = ac->in(TypeFunc::Memory);
1114 Node* ctl = ac->in(0);
1115 Node* src = ac->in(ArrayCopyNode::Src);
1116
1117 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
1118 return nullptr;
1119 }
1120
1121 // load depends on the tests that validate the arraycopy
1122 LoadNode* ld = clone_pinned();
1123 Node* addp = in(MemNode::Address)->clone();
1124 if (ac->as_ArrayCopy()->is_clonebasic()) {
1125 assert(ld_alloc != nullptr, "need an alloc");
1126 assert(addp->is_AddP(), "address must be addp");
1127 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1128 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1129 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1130 addp->set_req(AddPNode::Base, src);
1131 addp->set_req(AddPNode::Address, src);
1132 } else {
1133 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1134 ac->as_ArrayCopy()->is_copyof_validated() ||
1135 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1136 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1137 addp->set_req(AddPNode::Base, src);
1138 addp->set_req(AddPNode::Address, src);
1139
1140 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1141 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
1142 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
1143
1144 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
1145 uint shift = exact_log2(type2aelembytes(ary_elem));
1146
1147 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1148 #ifdef _LP64
1149 diff = phase->transform(new ConvI2LNode(diff));
1150 #endif
1151 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1152
1153 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1154 addp->set_req(AddPNode::Offset, offset);
1155 }
1156 addp = phase->transform(addp);
1157 #ifdef ASSERT
1158 const TypePtr* adr_type = phase->type(addp)->is_ptr();
1159 ld->_adr_type = adr_type;
1160 #endif
1161 ld->set_req(MemNode::Address, addp);
1162 ld->set_req(0, ctl);
1163 ld->set_req(MemNode::Memory, mem);
1164 return ld;
1165 }
1166 return nullptr;
1167 }
1168
1169
1170 //---------------------------can_see_stored_value------------------------------
1171 // This routine exists to make sure this set of tests is done the same
1172 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
1173 // will change the graph shape in a way which makes memory alive twice at the
1174 // same time (uses the Oracle model of aliasing), then some
1175 // LoadXNode::Identity will fold things back to the equivalence-class model
1176 // of aliasing.
1177 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const {
1178 Node* ld_adr = in(MemNode::Address);
1179 intptr_t ld_off = 0;
1180 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1181 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base);
1182 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1183 Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr;
1184 // This is more general than load from boxing objects.
1185 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1186 uint alias_idx = atp->index();
1187 Node* result = nullptr;
1188 Node* current = st;
1189 // Skip through chains of MemBarNodes checking the MergeMems for
1190 // new states for the slice of this load. Stop once any other
1191 // kind of node is encountered. Loads from final memory can skip
1192 // through any kind of MemBar but normal loads shouldn't skip
1193 // through MemBarAcquire since the could allow them to move out of
1194 // a synchronized region. It is not safe to step over MemBarCPUOrder,
1195 // because alias info above them may be inaccurate (e.g., due to
1196 // mixed/mismatched unsafe accesses).
1197 bool is_final_mem = !atp->is_rewritable();
1198 while (current->is_Proj()) {
1199 int opc = current->in(0)->Opcode();
1200 if ((is_final_mem && (opc == Op_MemBarAcquire ||
1201 opc == Op_MemBarAcquireLock ||
1202 opc == Op_LoadFence)) ||
1203 opc == Op_MemBarRelease ||
1204 opc == Op_StoreFence ||
1205 opc == Op_MemBarReleaseLock ||
1206 opc == Op_MemBarStoreStore ||
1207 opc == Op_StoreStoreFence) {
1208 Node* mem = current->in(0)->in(TypeFunc::Memory);
1209 if (mem->is_MergeMem()) {
1210 MergeMemNode* merge = mem->as_MergeMem();
1211 Node* new_st = merge->memory_at(alias_idx);
1212 if (new_st == merge->base_memory()) {
1213 // Keep searching
1214 current = new_st;
1215 continue;
1216 }
1217 // Save the new memory state for the slice and fall through
1218 // to exit.
1219 result = new_st;
1220 }
1221 }
1222 break;
1223 }
1224 if (result != nullptr) {
1225 st = result;
1226 }
1227 }
1228
1229 // Loop around twice in the case Load -> Initialize -> Store.
1230 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1231 for (int trip = 0; trip <= 1; trip++) {
1232
1233 if (st->is_Store()) {
1234 Node* st_adr = st->in(MemNode::Address);
1235 if (st_adr != ld_adr) {
1236 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1237 intptr_t st_off = 0;
1238 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1239 if (ld_base == nullptr) return nullptr;
1240 if (st_base == nullptr) return nullptr;
1241 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr;
1242 if (ld_off != st_off) return nullptr;
1243 if (ld_off == Type::OffsetBot) return nullptr;
1244 // Same base, same offset.
1245 // Possible improvement for arrays: check index value instead of absolute offset.
1246
1247 // At this point we have proven something like this setup:
1248 // B = << base >>
1249 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1250 // S = StoreQ(AddP( B , #Off), V)
1251 // (Actually, we haven't yet proven the Q's are the same.)
1252 // In other words, we are loading from a casted version of
1253 // the same pointer-and-offset that we stored to.
1254 // Casted version may carry a dependency and it is respected.
1255 // Thus, we are able to replace L by V.
1256 }
1257 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1258 if (store_Opcode() != st->Opcode()) {
1259 return nullptr;
1260 }
1261 // LoadVector/StoreVector needs additional check to ensure the types match.
1262 if (st->is_StoreVector()) {
1263 if ((Opcode() != Op_LoadVector && Opcode() != Op_StoreVector) || st->Opcode() != Op_StoreVector) {
1264 // Some kind of masked access or gather/scatter
1265 return nullptr;
1266 }
1267
1268 const TypeVect* in_vt = st->as_StoreVector()->vect_type();
1269 const TypeVect* out_vt = is_Load() ? as_LoadVector()->vect_type() : as_StoreVector()->vect_type();
1270 if (in_vt != out_vt) {
1271 return nullptr;
1272 }
1273 }
1274 return st->in(MemNode::ValueIn);
1275 }
1276
1277 // A load from a freshly-created object always returns zero.
1278 // (This can happen after LoadNode::Ideal resets the load's memory input
1279 // to find_captured_store, which returned InitializeNode::zero_memory.)
1280 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1281 (st->in(0) == ld_alloc) &&
1282 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1283 // return a zero value for the load's basic type
1284 // (This is one of the few places where a generic PhaseTransform
1285 // can create new nodes. Think of it as lazily manifesting
1286 // virtually pre-existing constants.)
1287 if (value_basic_type() != T_VOID) {
1288 if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) {
1289 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1290 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1291 // by the ArrayCopyNode.
1292 return phase->zerocon(value_basic_type());
1293 }
1294 } else {
1295 // TODO: materialize all-zero vector constant
1296 assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1297 }
1298 }
1299
1300 // A load from an initialization barrier can match a captured store.
1301 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1302 InitializeNode* init = st->in(0)->as_Initialize();
1303 AllocateNode* alloc = init->allocation();
1304 if ((alloc != nullptr) && (alloc == ld_alloc)) {
1305 // examine a captured store value
1306 st = init->find_captured_store(ld_off, memory_size(), phase);
1307 if (st != nullptr) {
1308 continue; // take one more trip around
1309 }
1310 }
1311 }
1312
1313 // Load boxed value from result of valueOf() call is input parameter.
1314 if (this->is_Load() && ld_adr->is_AddP() &&
1315 (tp != nullptr) && tp->is_ptr_to_boxed_value()) {
1316 intptr_t ignore = 0;
1317 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1318 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1319 base = bs->step_over_gc_barrier(base);
1320 if (base != nullptr && base->is_Proj() &&
1321 base->as_Proj()->_con == TypeFunc::Parms &&
1322 base->in(0)->is_CallStaticJava() &&
1323 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1324 return base->in(0)->in(TypeFunc::Parms);
1325 }
1326 }
1327
1328 break;
1329 }
1330
1331 return nullptr;
1332 }
1333
1334 //----------------------is_instance_field_load_with_local_phi------------------
1335 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1336 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1337 in(Address)->is_AddP() ) {
1338 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1339 // Only instances and boxed values.
1340 if( t_oop != nullptr &&
1341 (t_oop->is_ptr_to_boxed_value() ||
1342 t_oop->is_known_instance_field()) &&
1343 t_oop->offset() != Type::OffsetBot &&
1344 t_oop->offset() != Type::OffsetTop) {
1345 return true;
1346 }
1347 }
1348 return false;
1349 }
1350
1351 //------------------------------Identity---------------------------------------
1352 // Loads are identity if previous store is to same address
1353 Node* LoadNode::Identity(PhaseGVN* phase) {
1354 // If the previous store-maker is the right kind of Store, and the store is
1355 // to the same address, then we are equal to the value stored.
1356 Node* mem = in(Memory);
1357 Node* value = can_see_stored_value(mem, phase);
1358 if( value ) {
1359 // byte, short & char stores truncate naturally.
1360 // A load has to load the truncated value which requires
1361 // some sort of masking operation and that requires an
1362 // Ideal call instead of an Identity call.
1363 if (memory_size() < BytesPerInt) {
1364 // If the input to the store does not fit with the load's result type,
1365 // it must be truncated via an Ideal call.
1366 if (!phase->type(value)->higher_equal(phase->type(this)))
1367 return this;
1368 }
1369 // (This works even when value is a Con, but LoadNode::Value
1370 // usually runs first, producing the singleton type of the Con.)
1371 if (!has_pinned_control_dependency() || value->is_Con()) {
1372 return value;
1373 } else {
1374 return this;
1375 }
1376 }
1377
1378 if (has_pinned_control_dependency()) {
1379 return this;
1380 }
1381 // Search for an existing data phi which was generated before for the same
1382 // instance's field to avoid infinite generation of phis in a loop.
1383 Node *region = mem->in(0);
1384 if (is_instance_field_load_with_local_phi(region)) {
1385 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1386 int this_index = phase->C->get_alias_index(addr_t);
1387 int this_offset = addr_t->offset();
1388 int this_iid = addr_t->instance_id();
1389 if (!addr_t->is_known_instance() &&
1390 addr_t->is_ptr_to_boxed_value()) {
1391 // Use _idx of address base (could be Phi node) for boxed values.
1392 intptr_t ignore = 0;
1393 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1394 if (base == nullptr) {
1395 return this;
1396 }
1397 this_iid = base->_idx;
1398 }
1399 const Type* this_type = bottom_type();
1400 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1401 Node* phi = region->fast_out(i);
1402 if (phi->is_Phi() && phi != mem &&
1403 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1404 return phi;
1405 }
1406 }
1407 }
1408
1409 return this;
1410 }
1411
1412 // Construct an equivalent unsigned load.
1413 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1414 BasicType bt = T_ILLEGAL;
1415 const Type* rt = nullptr;
1416 switch (Opcode()) {
1417 case Op_LoadUB: return this;
1418 case Op_LoadUS: return this;
1419 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1420 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1421 default:
1422 assert(false, "no unsigned variant: %s", Name());
1423 return nullptr;
1424 }
1425 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1426 raw_adr_type(), rt, bt, _mo, _control_dependency,
1427 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1428 }
1429
1430 // Construct an equivalent signed load.
1431 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1432 BasicType bt = T_ILLEGAL;
1433 const Type* rt = nullptr;
1434 switch (Opcode()) {
1435 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1436 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1437 case Op_LoadB: // fall through
1438 case Op_LoadS: // fall through
1439 case Op_LoadI: // fall through
1440 case Op_LoadL: return this;
1441 default:
1442 assert(false, "no signed variant: %s", Name());
1443 return nullptr;
1444 }
1445 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1446 raw_adr_type(), rt, bt, _mo, _control_dependency,
1447 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1448 }
1449
1450 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1451 BasicType bt = rt->basic_type();
1452 switch (Opcode()) {
1453 case Op_LoadI: return (bt == T_FLOAT);
1454 case Op_LoadL: return (bt == T_DOUBLE);
1455 case Op_LoadF: return (bt == T_INT);
1456 case Op_LoadD: return (bt == T_LONG);
1457
1458 default: return false;
1459 }
1460 }
1461
1462 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1463 BasicType bt = rt->basic_type();
1464 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1465 bool is_mismatched = is_mismatched_access();
1466 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1467 if (raw_type == nullptr) {
1468 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1469 }
1470 const int op = Opcode();
1471 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) ||
1472 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access());
1473 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1474 raw_adr_type(), rt, bt, _mo, _control_dependency,
1475 require_atomic_access, is_unaligned_access(), is_mismatched);
1476 }
1477
1478 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1479 BasicType bt = vt->basic_type();
1480 switch (Opcode()) {
1481 case Op_StoreI: return (bt == T_FLOAT);
1482 case Op_StoreL: return (bt == T_DOUBLE);
1483 case Op_StoreF: return (bt == T_INT);
1484 case Op_StoreD: return (bt == T_LONG);
1485
1486 default: return false;
1487 }
1488 }
1489
1490 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1491 BasicType bt = vt->basic_type();
1492 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1493 const int op = Opcode();
1494 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) ||
1495 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access());
1496 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1497 raw_adr_type(), val, bt, _mo, require_atomic_access);
1498
1499 bool is_mismatched = is_mismatched_access();
1500 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1501 if (raw_type == nullptr) {
1502 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1503 }
1504 if (is_mismatched) {
1505 st->set_mismatched_access();
1506 }
1507 return st;
1508 }
1509
1510 // We're loading from an object which has autobox behaviour.
1511 // If this object is result of a valueOf call we'll have a phi
1512 // merging a newly allocated object and a load from the cache.
1513 // We want to replace this load with the original incoming
1514 // argument to the valueOf call.
1515 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1516 assert(igvn->C->eliminate_boxing(), "sanity");
1517 intptr_t ignore = 0;
1518 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1519 if ((base == nullptr) || base->is_Phi()) {
1520 // Push the loads from the phi that comes from valueOf up
1521 // through it to allow elimination of the loads and the recovery
1522 // of the original value. It is done in split_through_phi().
1523 return nullptr;
1524 } else if (base->is_Load() ||
1525 (base->is_DecodeN() && base->in(1)->is_Load())) {
1526 // Eliminate the load of boxed value for integer types from the cache
1527 // array by deriving the value from the index into the array.
1528 // Capture the offset of the load and then reverse the computation.
1529
1530 // Get LoadN node which loads a boxing object from 'cache' array.
1531 if (base->is_DecodeN()) {
1532 base = base->in(1);
1533 }
1534 if (!base->in(Address)->is_AddP()) {
1535 return nullptr; // Complex address
1536 }
1537 AddPNode* address = base->in(Address)->as_AddP();
1538 Node* cache_base = address->in(AddPNode::Base);
1539 if ((cache_base != nullptr) && cache_base->is_DecodeN()) {
1540 // Get ConP node which is static 'cache' field.
1541 cache_base = cache_base->in(1);
1542 }
1543 if ((cache_base != nullptr) && cache_base->is_Con()) {
1544 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1545 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
1546 Node* elements[4];
1547 int shift = exact_log2(type2aelembytes(T_OBJECT));
1548 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1549 if (count > 0 && elements[0]->is_Con() &&
1550 (count == 1 ||
1551 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1552 elements[1]->in(2) == igvn->intcon(shift)))) {
1553 ciObjArray* array = base_type->const_oop()->as_obj_array();
1554 // Fetch the box object cache[0] at the base of the array and get its value
1555 ciInstance* box = array->obj_at(0)->as_instance();
1556 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1557 assert(ik->is_box_klass(), "sanity");
1558 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1559 if (ik->nof_nonstatic_fields() == 1) {
1560 // This should be true nonstatic_field_at requires calling
1561 // nof_nonstatic_fields so check it anyway
1562 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1563 BasicType bt = c.basic_type();
1564 // Only integer types have boxing cache.
1565 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1566 bt == T_BYTE || bt == T_SHORT ||
1567 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1568 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1569 if (cache_low != (int)cache_low) {
1570 return nullptr; // should not happen since cache is array indexed by value
1571 }
1572 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1573 if (offset != (int)offset) {
1574 return nullptr; // should not happen since cache is array indexed by value
1575 }
1576 // Add up all the offsets making of the address of the load
1577 Node* result = elements[0];
1578 for (int i = 1; i < count; i++) {
1579 result = igvn->transform(new AddXNode(result, elements[i]));
1580 }
1581 // Remove the constant offset from the address and then
1582 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1583 // remove the scaling of the offset to recover the original index.
1584 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1585 // Peel the shift off directly but wrap it in a dummy node
1586 // since Ideal can't return existing nodes
1587 igvn->_worklist.push(result); // remove dead node later
1588 result = new RShiftXNode(result->in(1), igvn->intcon(0));
1589 } else if (result->is_Add() && result->in(2)->is_Con() &&
1590 result->in(1)->Opcode() == Op_LShiftX &&
1591 result->in(1)->in(2) == igvn->intcon(shift)) {
1592 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1593 // but for boxing cache access we know that X<<Z will not overflow
1594 // (there is range check) so we do this optimizatrion by hand here.
1595 igvn->_worklist.push(result); // remove dead node later
1596 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1597 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1598 } else {
1599 result = new RShiftXNode(result, igvn->intcon(shift));
1600 }
1601 #ifdef _LP64
1602 if (bt != T_LONG) {
1603 result = new ConvL2INode(igvn->transform(result));
1604 }
1605 #else
1606 if (bt == T_LONG) {
1607 result = new ConvI2LNode(igvn->transform(result));
1608 }
1609 #endif
1610 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1611 // Need to preserve unboxing load type if it is unsigned.
1612 switch(this->Opcode()) {
1613 case Op_LoadUB:
1614 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1615 break;
1616 case Op_LoadUS:
1617 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1618 break;
1619 }
1620 return result;
1621 }
1622 }
1623 }
1624 }
1625 }
1626 return nullptr;
1627 }
1628
1629 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1630 Node* region = phi->in(0);
1631 if (region == nullptr) {
1632 return false; // Wait stable graph
1633 }
1634 uint cnt = phi->req();
1635 for (uint i = 1; i < cnt; i++) {
1636 Node* rc = region->in(i);
1637 if (rc == nullptr || phase->type(rc) == Type::TOP)
1638 return false; // Wait stable graph
1639 Node* in = phi->in(i);
1640 if (in == nullptr || phase->type(in) == Type::TOP)
1641 return false; // Wait stable graph
1642 }
1643 return true;
1644 }
1645
1646 //------------------------------split_through_phi------------------------------
1647 // Check whether a call to 'split_through_phi' would split this load through the
1648 // Phi *base*. This method is essentially a copy of the validations performed
1649 // by 'split_through_phi'. The first use of this method was in EA code as part
1650 // of simplification of allocation merges.
1651 // Some differences from original method (split_through_phi):
1652 // - If base->is_CastPP(): base = base->in(1)
1653 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) {
1654 Node* mem = in(Memory);
1655 Node* address = in(Address);
1656 intptr_t ignore = 0;
1657 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1658
1659 if (base->is_CastPP()) {
1660 base = base->in(1);
1661 }
1662
1663 if (req() > 3 || base == nullptr || !base->is_Phi()) {
1664 return false;
1665 }
1666
1667 if (!mem->is_Phi()) {
1668 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1669 return false;
1670 }
1671 } else if (base->in(0) != mem->in(0)) {
1672 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1673 return false;
1674 }
1675 }
1676
1677 return true;
1678 }
1679
1680 //------------------------------split_through_phi------------------------------
1681 // Split instance or boxed field load through Phi.
1682 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) {
1683 if (req() > 3) {
1684 assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs");
1685 // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account
1686 return nullptr;
1687 }
1688 Node* mem = in(Memory);
1689 Node* address = in(Address);
1690 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1691
1692 assert((t_oop != nullptr) &&
1693 (ignore_missing_instance_id ||
1694 t_oop->is_known_instance_field() ||
1695 t_oop->is_ptr_to_boxed_value()), "invalid conditions");
1696
1697 Compile* C = phase->C;
1698 intptr_t ignore = 0;
1699 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1700 bool base_is_phi = (base != nullptr) && base->is_Phi();
1701 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1702 (base != nullptr) && (base == address->in(AddPNode::Base)) &&
1703 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1704
1705 if (!((mem->is_Phi() || base_is_phi) &&
1706 (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) {
1707 return nullptr; // Neither memory or base are Phi
1708 }
1709
1710 if (mem->is_Phi()) {
1711 if (!stable_phi(mem->as_Phi(), phase)) {
1712 return nullptr; // Wait stable graph
1713 }
1714 uint cnt = mem->req();
1715 // Check for loop invariant memory.
1716 if (cnt == 3) {
1717 for (uint i = 1; i < cnt; i++) {
1718 Node* in = mem->in(i);
1719 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1720 if (m == mem) {
1721 if (i == 1) {
1722 // if the first edge was a loop, check second edge too.
1723 // If both are replaceable - we are in an infinite loop
1724 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1725 if (n == mem) {
1726 break;
1727 }
1728 }
1729 set_req(Memory, mem->in(cnt - i));
1730 return this; // made change
1731 }
1732 }
1733 }
1734 }
1735 if (base_is_phi) {
1736 if (!stable_phi(base->as_Phi(), phase)) {
1737 return nullptr; // Wait stable graph
1738 }
1739 uint cnt = base->req();
1740 // Check for loop invariant memory.
1741 if (cnt == 3) {
1742 for (uint i = 1; i < cnt; i++) {
1743 if (base->in(i) == base) {
1744 return nullptr; // Wait stable graph
1745 }
1746 }
1747 }
1748 }
1749
1750 // Split through Phi (see original code in loopopts.cpp).
1751 assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type");
1752
1753 // Do nothing here if Identity will find a value
1754 // (to avoid infinite chain of value phis generation).
1755 if (this != Identity(phase)) {
1756 return nullptr;
1757 }
1758
1759 // Select Region to split through.
1760 Node* region;
1761 DomResult dom_result = DomResult::Dominate;
1762 if (!base_is_phi) {
1763 assert(mem->is_Phi(), "sanity");
1764 region = mem->in(0);
1765 // Skip if the region dominates some control edge of the address.
1766 // We will check `dom_result` later.
1767 dom_result = MemNode::maybe_all_controls_dominate(address, region);
1768 } else if (!mem->is_Phi()) {
1769 assert(base_is_phi, "sanity");
1770 region = base->in(0);
1771 // Skip if the region dominates some control edge of the memory.
1772 // We will check `dom_result` later.
1773 dom_result = MemNode::maybe_all_controls_dominate(mem, region);
1774 } else if (base->in(0) != mem->in(0)) {
1775 assert(base_is_phi && mem->is_Phi(), "sanity");
1776 dom_result = MemNode::maybe_all_controls_dominate(mem, base->in(0));
1777 if (dom_result == DomResult::Dominate) {
1778 region = base->in(0);
1779 } else {
1780 dom_result = MemNode::maybe_all_controls_dominate(address, mem->in(0));
1781 if (dom_result == DomResult::Dominate) {
1782 region = mem->in(0);
1783 }
1784 // Otherwise we encountered a complex graph.
1785 }
1786 } else {
1787 assert(base->in(0) == mem->in(0), "sanity");
1788 region = mem->in(0);
1789 }
1790
1791 PhaseIterGVN* igvn = phase->is_IterGVN();
1792 if (dom_result != DomResult::Dominate) {
1793 if (dom_result == DomResult::EncounteredDeadCode) {
1794 // There is some dead code which eventually will be removed in IGVN.
1795 // Once this is the case, we get an unambiguous dominance result.
1796 // Push the node to the worklist again until the dead code is removed.
1797 igvn->_worklist.push(this);
1798 }
1799 return nullptr;
1800 }
1801
1802 Node* phi = nullptr;
1803 const Type* this_type = this->bottom_type();
1804 if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) {
1805 int this_index = C->get_alias_index(t_oop);
1806 int this_offset = t_oop->offset();
1807 int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx;
1808 phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset);
1809 } else if (ignore_missing_instance_id) {
1810 phi = new PhiNode(region, this_type, nullptr, mem->_idx);
1811 } else {
1812 return nullptr;
1813 }
1814
1815 for (uint i = 1; i < region->req(); i++) {
1816 Node* x;
1817 Node* the_clone = nullptr;
1818 Node* in = region->in(i);
1819 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1820 in != nullptr && in->is_OuterStripMinedLoop()) {
1821 // No node should go in the outer strip mined loop
1822 in = in->in(LoopNode::EntryControl);
1823 }
1824 if (in == nullptr || in == C->top()) {
1825 x = C->top(); // Dead path? Use a dead data op
1826 } else {
1827 x = this->clone(); // Else clone up the data op
1828 the_clone = x; // Remember for possible deletion.
1829 // Alter data node to use pre-phi inputs
1830 if (this->in(0) == region) {
1831 x->set_req(0, in);
1832 } else {
1833 x->set_req(0, nullptr);
1834 }
1835 if (mem->is_Phi() && (mem->in(0) == region)) {
1836 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1837 }
1838 if (address->is_Phi() && address->in(0) == region) {
1839 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1840 }
1841 if (base_is_phi && (base->in(0) == region)) {
1842 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1843 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1844 x->set_req(Address, adr_x);
1845 }
1846 }
1847 // Check for a 'win' on some paths
1848 const Type *t = x->Value(igvn);
1849
1850 bool singleton = t->singleton();
1851
1852 // See comments in PhaseIdealLoop::split_thru_phi().
1853 if (singleton && t == Type::TOP) {
1854 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1855 }
1856
1857 if (singleton) {
1858 x = igvn->makecon(t);
1859 } else {
1860 // We now call Identity to try to simplify the cloned node.
1861 // Note that some Identity methods call phase->type(this).
1862 // Make sure that the type array is big enough for
1863 // our new node, even though we may throw the node away.
1864 // (This tweaking with igvn only works because x is a new node.)
1865 igvn->set_type(x, t);
1866 // If x is a TypeNode, capture any more-precise type permanently into Node
1867 // otherwise it will be not updated during igvn->transform since
1868 // igvn->type(x) is set to x->Value() already.
1869 x->raise_bottom_type(t);
1870 Node* y = x->Identity(igvn);
1871 if (y != x) {
1872 x = y;
1873 } else {
1874 y = igvn->hash_find_insert(x);
1875 if (y) {
1876 x = y;
1877 } else {
1878 // Else x is a new node we are keeping
1879 // We do not need register_new_node_with_optimizer
1880 // because set_type has already been called.
1881 igvn->_worklist.push(x);
1882 }
1883 }
1884 }
1885 if (x != the_clone && the_clone != nullptr) {
1886 igvn->remove_dead_node(the_clone);
1887 }
1888 phi->set_req(i, x);
1889 }
1890 // Record Phi
1891 igvn->register_new_node_with_optimizer(phi);
1892 return phi;
1893 }
1894
1895 AllocateNode* LoadNode::is_new_object_mark_load() const {
1896 if (Opcode() == Op_LoadX) {
1897 Node* address = in(MemNode::Address);
1898 AllocateNode* alloc = AllocateNode::Ideal_allocation(address);
1899 Node* mem = in(MemNode::Memory);
1900 if (alloc != nullptr && mem->is_Proj() &&
1901 mem->in(0) != nullptr &&
1902 mem->in(0) == alloc->initialization() &&
1903 alloc->initialization()->proj_out_or_null(0) != nullptr) {
1904 return alloc;
1905 }
1906 }
1907 return nullptr;
1908 }
1909
1910
1911 //------------------------------Ideal------------------------------------------
1912 // If the load is from Field memory and the pointer is non-null, it might be possible to
1913 // zero out the control input.
1914 // If the offset is constant and the base is an object allocation,
1915 // try to hook me up to the exact initializing store.
1916 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1917 if (has_pinned_control_dependency()) {
1918 return nullptr;
1919 }
1920 Node* p = MemNode::Ideal_common(phase, can_reshape);
1921 if (p) return (p == NodeSentinel) ? nullptr : p;
1922
1923 Node* ctrl = in(MemNode::Control);
1924 Node* address = in(MemNode::Address);
1925
1926 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1927 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1928
1929 // Skip up past a SafePoint control. Cannot do this for Stores because
1930 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1931 if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint &&
1932 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1933 !addr_mark &&
1934 (depends_only_on_test() || has_unknown_control_dependency())) {
1935 ctrl = ctrl->in(0);
1936 set_req(MemNode::Control,ctrl);
1937 return this;
1938 }
1939
1940 intptr_t ignore = 0;
1941 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1942 if (base != nullptr
1943 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1944 // Check for useless control edge in some common special cases
1945 if (in(MemNode::Control) != nullptr
1946 && can_remove_control()
1947 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1948 && all_controls_dominate(base, phase->C->start())) {
1949 // A method-invariant, non-null address (constant or 'this' argument).
1950 set_req(MemNode::Control, nullptr);
1951 return this;
1952 }
1953 }
1954
1955 Node* mem = in(MemNode::Memory);
1956 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1957
1958 if (can_reshape && (addr_t != nullptr)) {
1959 // try to optimize our memory input
1960 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1961 if (opt_mem != mem) {
1962 set_req_X(MemNode::Memory, opt_mem, phase);
1963 if (phase->type( opt_mem ) == Type::TOP) return nullptr;
1964 return this;
1965 }
1966 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1967 if ((t_oop != nullptr) &&
1968 (t_oop->is_known_instance_field() ||
1969 t_oop->is_ptr_to_boxed_value())) {
1970 PhaseIterGVN *igvn = phase->is_IterGVN();
1971 assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true");
1972 if (igvn->_worklist.member(opt_mem)) {
1973 // Delay this transformation until memory Phi is processed.
1974 igvn->_worklist.push(this);
1975 return nullptr;
1976 }
1977 // Split instance field load through Phi.
1978 Node* result = split_through_phi(phase);
1979 if (result != nullptr) return result;
1980
1981 if (t_oop->is_ptr_to_boxed_value()) {
1982 Node* result = eliminate_autobox(igvn);
1983 if (result != nullptr) return result;
1984 }
1985 }
1986 }
1987
1988 // Is there a dominating load that loads the same value? Leave
1989 // anything that is not a load of a field/array element (like
1990 // barriers etc.) alone
1991 if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) {
1992 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1993 Node *use = mem->fast_out(i);
1994 if (use != this &&
1995 use->Opcode() == Opcode() &&
1996 use->in(0) != nullptr &&
1997 use->in(0) != in(0) &&
1998 use->in(Address) == in(Address)) {
1999 Node* ctl = in(0);
2000 for (int i = 0; i < 10 && ctl != nullptr; i++) {
2001 ctl = IfNode::up_one_dom(ctl);
2002 if (ctl == use->in(0)) {
2003 set_req(0, use->in(0));
2004 return this;
2005 }
2006 }
2007 }
2008 }
2009 }
2010
2011 // Check for prior store with a different base or offset; make Load
2012 // independent. Skip through any number of them. Bail out if the stores
2013 // are in an endless dead cycle and report no progress. This is a key
2014 // transform for Reflection. However, if after skipping through the Stores
2015 // we can't then fold up against a prior store do NOT do the transform as
2016 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
2017 // array memory alive twice: once for the hoisted Load and again after the
2018 // bypassed Store. This situation only works if EVERYBODY who does
2019 // anti-dependence work knows how to bypass. I.e. we need all
2020 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
2021 // the alias index stuff. So instead, peek through Stores and IFF we can
2022 // fold up, do so.
2023 Node* prev_mem = find_previous_store(phase);
2024 if (prev_mem != nullptr && prev_mem->is_top()) {
2025 // find_previous_store returns top when the access is dead
2026 return prev_mem;
2027 }
2028 if (prev_mem != nullptr) {
2029 Node* value = can_see_arraycopy_value(prev_mem, phase);
2030 if (value != nullptr) {
2031 return value;
2032 }
2033 }
2034 // Steps (a), (b): Walk past independent stores to find an exact match.
2035 if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) {
2036 // (c) See if we can fold up on the spot, but don't fold up here.
2037 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
2038 // just return a prior value, which is done by Identity calls.
2039 if (can_see_stored_value(prev_mem, phase)) {
2040 // Make ready for step (d):
2041 set_req_X(MemNode::Memory, prev_mem, phase);
2042 return this;
2043 }
2044 }
2045
2046 if (!can_reshape) {
2047 phase->record_for_igvn(this);
2048 }
2049
2050 return nullptr;
2051 }
2052
2053 // Helper to recognize certain Klass fields which are invariant across
2054 // some group of array types (e.g., int[] or all T[] where T < Object).
2055 const Type*
2056 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
2057 ciKlass* klass) const {
2058 assert(!UseCompactObjectHeaders || tkls->offset() != in_bytes(Klass::prototype_header_offset()),
2059 "must not happen");
2060
2061 if (tkls->isa_instklassptr() && tkls->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
2062 // The field is InstanceKlass::_access_flags. Return its (constant) value.
2063 assert(Opcode() == Op_LoadUS, "must load an unsigned short from _access_flags");
2064 ciInstanceKlass* iklass = tkls->is_instklassptr()->instance_klass();
2065 return TypeInt::make(iklass->access_flags());
2066 }
2067 if (tkls->offset() == in_bytes(Klass::misc_flags_offset())) {
2068 // The field is Klass::_misc_flags. Return its (constant) value.
2069 assert(Opcode() == Op_LoadUB, "must load an unsigned byte from _misc_flags");
2070 return TypeInt::make(klass->misc_flags());
2071 }
2072 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
2073 // The field is Klass::_layout_helper. Return its constant value if known.
2074 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2075 return TypeInt::make(klass->layout_helper());
2076 }
2077
2078 // No match.
2079 return nullptr;
2080 }
2081
2082 //------------------------------Value-----------------------------------------
2083 const Type* LoadNode::Value(PhaseGVN* phase) const {
2084 // Either input is TOP ==> the result is TOP
2085 Node* mem = in(MemNode::Memory);
2086 const Type *t1 = phase->type(mem);
2087 if (t1 == Type::TOP) return Type::TOP;
2088 Node* adr = in(MemNode::Address);
2089 const TypePtr* tp = phase->type(adr)->isa_ptr();
2090 if (tp == nullptr || tp->empty()) return Type::TOP;
2091 int off = tp->offset();
2092 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
2093 Compile* C = phase->C;
2094
2095 // If load can see a previous constant store, use that.
2096 Node* value = can_see_stored_value(mem, phase);
2097 if (value != nullptr && value->is_Con()) {
2098 assert(value->bottom_type()->higher_equal(_type), "sanity");
2099 return value->bottom_type();
2100 }
2101
2102 // Try to guess loaded type from pointer type
2103 if (tp->isa_aryptr()) {
2104 const TypeAryPtr* ary = tp->is_aryptr();
2105 const Type* t = ary->elem();
2106
2107 // Determine whether the reference is beyond the header or not, by comparing
2108 // the offset against the offset of the start of the array's data.
2109 // Different array types begin at slightly different offsets (12 vs. 16).
2110 // We choose T_BYTE as an example base type that is least restrictive
2111 // as to alignment, which will therefore produce the smallest
2112 // possible base offset.
2113 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2114 const bool off_beyond_header = (off >= min_base_off);
2115
2116 // Try to constant-fold a stable array element.
2117 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
2118 // Make sure the reference is not into the header and the offset is constant
2119 ciObject* aobj = ary->const_oop();
2120 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
2121 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
2122 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
2123 stable_dimension,
2124 value_basic_type(), is_unsigned());
2125 if (con_type != nullptr) {
2126 return con_type;
2127 }
2128 }
2129 }
2130
2131 // Don't do this for integer types. There is only potential profit if
2132 // the element type t is lower than _type; that is, for int types, if _type is
2133 // more restrictive than t. This only happens here if one is short and the other
2134 // char (both 16 bits), and in those cases we've made an intentional decision
2135 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2136 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2137 //
2138 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2139 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
2140 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2141 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
2142 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2143 // In fact, that could have been the original type of p1, and p1 could have
2144 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2145 // expression (LShiftL quux 3) independently optimized to the constant 8.
2146 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2147 && (_type->isa_vect() == nullptr)
2148 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2149 // t might actually be lower than _type, if _type is a unique
2150 // concrete subclass of abstract class t.
2151 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
2152 const Type* jt = t->join_speculative(_type);
2153 // In any case, do not allow the join, per se, to empty out the type.
2154 if (jt->empty() && !t->empty()) {
2155 // This can happen if a interface-typed array narrows to a class type.
2156 jt = _type;
2157 }
2158 #ifdef ASSERT
2159 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2160 // The pointers in the autobox arrays are always non-null
2161 Node* base = adr->in(AddPNode::Base);
2162 if ((base != nullptr) && base->is_DecodeN()) {
2163 // Get LoadN node which loads IntegerCache.cache field
2164 base = base->in(1);
2165 }
2166 if ((base != nullptr) && base->is_Con()) {
2167 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2168 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2169 // It could be narrow oop
2170 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2171 }
2172 }
2173 }
2174 #endif
2175 return jt;
2176 }
2177 }
2178 } else if (tp->base() == Type::InstPtr) {
2179 assert( off != Type::OffsetBot ||
2180 // arrays can be cast to Objects
2181 !tp->isa_instptr() ||
2182 tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
2183 // unsafe field access may not have a constant offset
2184 C->has_unsafe_access(),
2185 "Field accesses must be precise" );
2186 // For oop loads, we expect the _type to be precise.
2187
2188 // Optimize loads from constant fields.
2189 const TypeInstPtr* tinst = tp->is_instptr();
2190 ciObject* const_oop = tinst->const_oop();
2191 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2192 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), value_basic_type());
2193 if (con_type != nullptr) {
2194 return con_type;
2195 }
2196 }
2197 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2198 assert(off != Type::OffsetBot ||
2199 !tp->isa_instklassptr() ||
2200 // arrays can be cast to Objects
2201 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2202 // also allow array-loading from the primary supertype
2203 // array during subtype checks
2204 Opcode() == Op_LoadKlass,
2205 "Field accesses must be precise");
2206 // For klass/static loads, we expect the _type to be precise
2207 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2208 /* With mirrors being an indirect in the Klass*
2209 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2210 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2211 *
2212 * So check the type and klass of the node before the LoadP.
2213 */
2214 Node* adr2 = adr->in(MemNode::Address);
2215 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2216 if (tkls != nullptr && !StressReflectiveCode) {
2217 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2218 ciKlass* klass = tkls->exact_klass();
2219 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2220 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2221 return TypeInstPtr::make(klass->java_mirror());
2222 }
2223 }
2224 }
2225
2226 const TypeKlassPtr *tkls = tp->isa_klassptr();
2227 if (tkls != nullptr) {
2228 if (tkls->is_loaded() && tkls->klass_is_exact()) {
2229 ciKlass* klass = tkls->exact_klass();
2230 // We are loading a field from a Klass metaobject whose identity
2231 // is known at compile time (the type is "exact" or "precise").
2232 // Check for fields we know are maintained as constants by the VM.
2233 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2234 // The field is Klass::_super_check_offset. Return its (constant) value.
2235 // (Folds up type checking code.)
2236 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2237 return TypeInt::make(klass->super_check_offset());
2238 }
2239 if (UseCompactObjectHeaders) {
2240 if (tkls->offset() == in_bytes(Klass::prototype_header_offset())) {
2241 // The field is Klass::_prototype_header. Return its (constant) value.
2242 assert(this->Opcode() == Op_LoadX, "must load a proper type from _prototype_header");
2243 return TypeX::make(klass->prototype_header());
2244 }
2245 }
2246 // Compute index into primary_supers array
2247 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2248 // Check for overflowing; use unsigned compare to handle the negative case.
2249 if( depth < ciKlass::primary_super_limit() ) {
2250 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2251 // (Folds up type checking code.)
2252 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2253 ciKlass *ss = klass->super_of_depth(depth);
2254 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2255 }
2256 const Type* aift = load_array_final_field(tkls, klass);
2257 if (aift != nullptr) return aift;
2258 }
2259
2260 // We can still check if we are loading from the primary_supers array at a
2261 // shallow enough depth. Even though the klass is not exact, entries less
2262 // than or equal to its super depth are correct.
2263 if (tkls->is_loaded()) {
2264 ciKlass* klass = nullptr;
2265 if (tkls->isa_instklassptr()) {
2266 klass = tkls->is_instklassptr()->instance_klass();
2267 } else {
2268 int dims;
2269 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims);
2270 if (inner->isa_instklassptr()) {
2271 klass = inner->is_instklassptr()->instance_klass();
2272 klass = ciObjArrayKlass::make(klass, dims);
2273 }
2274 }
2275 if (klass != nullptr) {
2276 // Compute index into primary_supers array
2277 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2278 // Check for overflowing; use unsigned compare to handle the negative case.
2279 if (depth < ciKlass::primary_super_limit() &&
2280 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case
2281 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2282 // (Folds up type checking code.)
2283 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2284 ciKlass *ss = klass->super_of_depth(depth);
2285 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2286 }
2287 }
2288 }
2289
2290 // If the type is enough to determine that the thing is not an array,
2291 // we can give the layout_helper a positive interval type.
2292 // This will help short-circuit some reflective code.
2293 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) &&
2294 tkls->isa_instklassptr() && // not directly typed as an array
2295 !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
2296 ) {
2297 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2298 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2299 // The key property of this type is that it folds up tests
2300 // for array-ness, since it proves that the layout_helper is positive.
2301 // Thus, a generic value like the basic object layout helper works fine.
2302 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2303 }
2304 }
2305
2306 // If we are loading from a freshly-allocated object/array, produce a zero.
2307 // Things to check:
2308 // 1. Load is beyond the header: headers are not guaranteed to be zero
2309 // 2. Load is not vectorized: vectors have no zero constant
2310 // 3. Load has no matching store, i.e. the input is the initial memory state
2311 const TypeOopPtr* tinst = tp->isa_oopptr();
2312 bool is_not_header = (tinst != nullptr) && tinst->is_known_instance_field();
2313 bool is_not_vect = (_type->isa_vect() == nullptr);
2314 if (is_not_header && is_not_vect) {
2315 Node* mem = in(MemNode::Memory);
2316 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2317 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2318 return Type::get_zero_type(_type->basic_type());
2319 }
2320 }
2321
2322 if (!UseCompactObjectHeaders) {
2323 Node* alloc = is_new_object_mark_load();
2324 if (alloc != nullptr) {
2325 return TypeX::make(markWord::prototype().value());
2326 }
2327 }
2328
2329 return _type;
2330 }
2331
2332 //------------------------------match_edge-------------------------------------
2333 // Do we Match on this edge index or not? Match only the address.
2334 uint LoadNode::match_edge(uint idx) const {
2335 return idx == MemNode::Address;
2336 }
2337
2338 //--------------------------LoadBNode::Ideal--------------------------------------
2339 //
2340 // If the previous store is to the same address as this load,
2341 // and the value stored was larger than a byte, replace this load
2342 // with the value stored truncated to a byte. If no truncation is
2343 // needed, the replacement is done in LoadNode::Identity().
2344 //
2345 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2346 Node* mem = in(MemNode::Memory);
2347 Node* value = can_see_stored_value(mem,phase);
2348 if (value != nullptr) {
2349 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2350 if (narrow != value) {
2351 return narrow;
2352 }
2353 }
2354 // Identity call will handle the case where truncation is not needed.
2355 return LoadNode::Ideal(phase, can_reshape);
2356 }
2357
2358 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2359 Node* mem = in(MemNode::Memory);
2360 Node* value = can_see_stored_value(mem,phase);
2361 if (value != nullptr && value->is_Con() &&
2362 !value->bottom_type()->higher_equal(_type)) {
2363 // If the input to the store does not fit with the load's result type,
2364 // it must be truncated. We can't delay until Ideal call since
2365 // a singleton Value is needed for split_thru_phi optimization.
2366 int con = value->get_int();
2367 return TypeInt::make((con << 24) >> 24);
2368 }
2369 return LoadNode::Value(phase);
2370 }
2371
2372 //--------------------------LoadUBNode::Ideal-------------------------------------
2373 //
2374 // If the previous store is to the same address as this load,
2375 // and the value stored was larger than a byte, replace this load
2376 // with the value stored truncated to a byte. If no truncation is
2377 // needed, the replacement is done in LoadNode::Identity().
2378 //
2379 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2380 Node* mem = in(MemNode::Memory);
2381 Node* value = can_see_stored_value(mem, phase);
2382 if (value != nullptr) {
2383 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2384 if (narrow != value) {
2385 return narrow;
2386 }
2387 }
2388 // Identity call will handle the case where truncation is not needed.
2389 return LoadNode::Ideal(phase, can_reshape);
2390 }
2391
2392 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2393 Node* mem = in(MemNode::Memory);
2394 Node* value = can_see_stored_value(mem,phase);
2395 if (value != nullptr && value->is_Con() &&
2396 !value->bottom_type()->higher_equal(_type)) {
2397 // If the input to the store does not fit with the load's result type,
2398 // it must be truncated. We can't delay until Ideal call since
2399 // a singleton Value is needed for split_thru_phi optimization.
2400 int con = value->get_int();
2401 return TypeInt::make(con & 0xFF);
2402 }
2403 return LoadNode::Value(phase);
2404 }
2405
2406 //--------------------------LoadUSNode::Ideal-------------------------------------
2407 //
2408 // If the previous store is to the same address as this load,
2409 // and the value stored was larger than a char, replace this load
2410 // with the value stored truncated to a char. If no truncation is
2411 // needed, the replacement is done in LoadNode::Identity().
2412 //
2413 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2414 Node* mem = in(MemNode::Memory);
2415 Node* value = can_see_stored_value(mem,phase);
2416 if (value != nullptr) {
2417 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2418 if (narrow != value) {
2419 return narrow;
2420 }
2421 }
2422 // Identity call will handle the case where truncation is not needed.
2423 return LoadNode::Ideal(phase, can_reshape);
2424 }
2425
2426 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2427 Node* mem = in(MemNode::Memory);
2428 Node* value = can_see_stored_value(mem,phase);
2429 if (value != nullptr && value->is_Con() &&
2430 !value->bottom_type()->higher_equal(_type)) {
2431 // If the input to the store does not fit with the load's result type,
2432 // it must be truncated. We can't delay until Ideal call since
2433 // a singleton Value is needed for split_thru_phi optimization.
2434 int con = value->get_int();
2435 return TypeInt::make(con & 0xFFFF);
2436 }
2437 return LoadNode::Value(phase);
2438 }
2439
2440 //--------------------------LoadSNode::Ideal--------------------------------------
2441 //
2442 // If the previous store is to the same address as this load,
2443 // and the value stored was larger than a short, replace this load
2444 // with the value stored truncated to a short. If no truncation is
2445 // needed, the replacement is done in LoadNode::Identity().
2446 //
2447 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2448 Node* mem = in(MemNode::Memory);
2449 Node* value = can_see_stored_value(mem,phase);
2450 if (value != nullptr) {
2451 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2452 if (narrow != value) {
2453 return narrow;
2454 }
2455 }
2456 // Identity call will handle the case where truncation is not needed.
2457 return LoadNode::Ideal(phase, can_reshape);
2458 }
2459
2460 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2461 Node* mem = in(MemNode::Memory);
2462 Node* value = can_see_stored_value(mem,phase);
2463 if (value != nullptr && value->is_Con() &&
2464 !value->bottom_type()->higher_equal(_type)) {
2465 // If the input to the store does not fit with the load's result type,
2466 // it must be truncated. We can't delay until Ideal call since
2467 // a singleton Value is needed for split_thru_phi optimization.
2468 int con = value->get_int();
2469 return TypeInt::make((con << 16) >> 16);
2470 }
2471 return LoadNode::Value(phase);
2472 }
2473
2474 //=============================================================================
2475 //----------------------------LoadKlassNode::make------------------------------
2476 // Polymorphic factory method:
2477 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2478 // sanity check the alias category against the created node type
2479 const TypePtr* adr_type = adr->bottom_type()->isa_ptr();
2480 assert(adr_type != nullptr, "expecting TypeKlassPtr");
2481 #ifdef _LP64
2482 if (adr_type->is_ptr_to_narrowklass()) {
2483 assert(UseCompressedClassPointers, "no compressed klasses");
2484 Node* load_klass = gvn.transform(new LoadNKlassNode(mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2485 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2486 }
2487 #endif
2488 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2489 return new LoadKlassNode(mem, adr, at, tk, MemNode::unordered);
2490 }
2491
2492 //------------------------------Value------------------------------------------
2493 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2494 return klass_value_common(phase);
2495 }
2496
2497 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2498 // Either input is TOP ==> the result is TOP
2499 const Type *t1 = phase->type( in(MemNode::Memory) );
2500 if (t1 == Type::TOP) return Type::TOP;
2501 Node *adr = in(MemNode::Address);
2502 const Type *t2 = phase->type( adr );
2503 if (t2 == Type::TOP) return Type::TOP;
2504 const TypePtr *tp = t2->is_ptr();
2505 if (TypePtr::above_centerline(tp->ptr()) ||
2506 tp->ptr() == TypePtr::Null) return Type::TOP;
2507
2508 // Return a more precise klass, if possible
2509 const TypeInstPtr *tinst = tp->isa_instptr();
2510 if (tinst != nullptr) {
2511 ciInstanceKlass* ik = tinst->instance_klass();
2512 int offset = tinst->offset();
2513 if (ik == phase->C->env()->Class_klass()
2514 && (offset == java_lang_Class::klass_offset() ||
2515 offset == java_lang_Class::array_klass_offset())) {
2516 // We are loading a special hidden field from a Class mirror object,
2517 // the field which points to the VM's Klass metaobject.
2518 ciType* t = tinst->java_mirror_type();
2519 // java_mirror_type returns non-null for compile-time Class constants.
2520 if (t != nullptr) {
2521 // constant oop => constant klass
2522 if (offset == java_lang_Class::array_klass_offset()) {
2523 if (t->is_void()) {
2524 // We cannot create a void array. Since void is a primitive type return null
2525 // klass. Users of this result need to do a null check on the returned klass.
2526 return TypePtr::NULL_PTR;
2527 }
2528 return TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);
2529 }
2530 if (!t->is_klass()) {
2531 // a primitive Class (e.g., int.class) has null for a klass field
2532 return TypePtr::NULL_PTR;
2533 }
2534 // Fold up the load of the hidden field
2535 return TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);
2536 }
2537 // non-constant mirror, so we can't tell what's going on
2538 }
2539 if (!tinst->is_loaded())
2540 return _type; // Bail out if not loaded
2541 if (offset == oopDesc::klass_offset_in_bytes()) {
2542 return tinst->as_klass_type(true);
2543 }
2544 }
2545
2546 // Check for loading klass from an array
2547 const TypeAryPtr *tary = tp->isa_aryptr();
2548 if (tary != nullptr &&
2549 tary->offset() == oopDesc::klass_offset_in_bytes()) {
2550 return tary->as_klass_type(true);
2551 }
2552
2553 // Check for loading klass from an array klass
2554 const TypeKlassPtr *tkls = tp->isa_klassptr();
2555 if (tkls != nullptr && !StressReflectiveCode) {
2556 if (!tkls->is_loaded())
2557 return _type; // Bail out if not loaded
2558 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2559 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2560 // // Always returning precise element type is incorrect,
2561 // // e.g., element type could be object and array may contain strings
2562 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2563
2564 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2565 // according to the element type's subclassing.
2566 return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2567 }
2568 if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() &&
2569 tkls->offset() == in_bytes(Klass::super_offset())) {
2570 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super();
2571 // The field is Klass::_super. Return its (constant) value.
2572 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2573 return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR;
2574 }
2575 }
2576
2577 if (tkls != nullptr && !UseSecondarySupersCache
2578 && tkls->offset() == in_bytes(Klass::secondary_super_cache_offset())) {
2579 // Treat Klass::_secondary_super_cache as a constant when the cache is disabled.
2580 return TypePtr::NULL_PTR;
2581 }
2582
2583 // Bailout case
2584 return LoadNode::Value(phase);
2585 }
2586
2587 //------------------------------Identity---------------------------------------
2588 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2589 // Also feed through the klass in Allocate(...klass...)._klass.
2590 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2591 return klass_identity_common(phase);
2592 }
2593
2594 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2595 Node* x = LoadNode::Identity(phase);
2596 if (x != this) return x;
2597
2598 // Take apart the address into an oop and offset.
2599 // Return 'this' if we cannot.
2600 Node* adr = in(MemNode::Address);
2601 intptr_t offset = 0;
2602 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2603 if (base == nullptr) return this;
2604 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2605 if (toop == nullptr) return this;
2606
2607 // Step over potential GC barrier for OopHandle resolve
2608 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2609 if (bs->is_gc_barrier_node(base)) {
2610 base = bs->step_over_gc_barrier(base);
2611 }
2612
2613 // We can fetch the klass directly through an AllocateNode.
2614 // This works even if the klass is not constant (clone or newArray).
2615 if (offset == oopDesc::klass_offset_in_bytes()) {
2616 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2617 if (allocated_klass != nullptr) {
2618 return allocated_klass;
2619 }
2620 }
2621
2622 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2623 // See inline_native_Class_query for occurrences of these patterns.
2624 // Java Example: x.getClass().isAssignableFrom(y)
2625 //
2626 // This improves reflective code, often making the Class
2627 // mirror go completely dead. (Current exception: Class
2628 // mirrors may appear in debug info, but we could clean them out by
2629 // introducing a new debug info operator for Klass.java_mirror).
2630
2631 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass()
2632 && offset == java_lang_Class::klass_offset()) {
2633 if (base->is_Load()) {
2634 Node* base2 = base->in(MemNode::Address);
2635 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2636 Node* adr2 = base2->in(MemNode::Address);
2637 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2638 if (tkls != nullptr && !tkls->empty()
2639 && (tkls->isa_instklassptr() || tkls->isa_aryklassptr())
2640 && adr2->is_AddP()
2641 ) {
2642 int mirror_field = in_bytes(Klass::java_mirror_offset());
2643 if (tkls->offset() == mirror_field) {
2644 #ifdef ASSERT
2645 const TypeKlassPtr* tkls2 = phase->type(adr2->in(AddPNode::Address))->is_klassptr();
2646 assert(tkls2->offset() == 0, "not a load of java_mirror");
2647 #endif
2648 assert(adr2->in(AddPNode::Base)->is_top(), "not an off heap load");
2649 assert(adr2->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(Klass::java_mirror_offset()), "incorrect offset");
2650 return adr2->in(AddPNode::Address);
2651 }
2652 }
2653 }
2654 }
2655 }
2656
2657 return this;
2658 }
2659
2660 LoadNode* LoadNode::clone_pinned() const {
2661 LoadNode* ld = clone()->as_Load();
2662 ld->_control_dependency = UnknownControl;
2663 return ld;
2664 }
2665
2666 // Pin a LoadNode if it carries a dependency on its control input. There are cases when the node
2667 // does not actually have any dependency on its control input. For example, if we have a LoadNode
2668 // being used only outside a loop but it must be scheduled inside the loop, we can clone the node
2669 // for each of its use so that all the clones can be scheduled outside the loop. Then, to prevent
2670 // the clones from being GVN-ed again, we add a control input for each of them at the loop exit. In
2671 // those case, since there is not a dependency between the node and its control input, we do not
2672 // need to pin it.
2673 LoadNode* LoadNode::pin_node_under_control_impl() const {
2674 const TypePtr* adr_type = this->adr_type();
2675 if (adr_type != nullptr && adr_type->isa_aryptr()) {
2676 // Only array accesses have dependencies on their control input
2677 return clone_pinned();
2678 }
2679 return nullptr;
2680 }
2681
2682 //------------------------------Value------------------------------------------
2683 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2684 const Type *t = klass_value_common(phase);
2685 if (t == Type::TOP)
2686 return t;
2687
2688 return t->make_narrowklass();
2689 }
2690
2691 //------------------------------Identity---------------------------------------
2692 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2693 // Also feed through the klass in Allocate(...klass...)._klass.
2694 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2695 Node *x = klass_identity_common(phase);
2696
2697 const Type *t = phase->type( x );
2698 if( t == Type::TOP ) return x;
2699 if( t->isa_narrowklass()) return x;
2700 assert (!t->isa_narrowoop(), "no narrow oop here");
2701
2702 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2703 }
2704
2705 //------------------------------Value-----------------------------------------
2706 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2707 // Either input is TOP ==> the result is TOP
2708 const Type *t1 = phase->type( in(MemNode::Memory) );
2709 if( t1 == Type::TOP ) return Type::TOP;
2710 Node *adr = in(MemNode::Address);
2711 const Type *t2 = phase->type( adr );
2712 if( t2 == Type::TOP ) return Type::TOP;
2713 const TypePtr *tp = t2->is_ptr();
2714 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2715 const TypeAryPtr *tap = tp->isa_aryptr();
2716 if( !tap ) return _type;
2717 return tap->size();
2718 }
2719
2720 //-------------------------------Ideal---------------------------------------
2721 // Feed through the length in AllocateArray(...length...)._length.
2722 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2723 Node* p = MemNode::Ideal_common(phase, can_reshape);
2724 if (p) return (p == NodeSentinel) ? nullptr : p;
2725
2726 // Take apart the address into an oop and offset.
2727 // Return 'this' if we cannot.
2728 Node* adr = in(MemNode::Address);
2729 intptr_t offset = 0;
2730 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2731 if (base == nullptr) return nullptr;
2732 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2733 if (tary == nullptr) return nullptr;
2734
2735 // We can fetch the length directly through an AllocateArrayNode.
2736 // This works even if the length is not constant (clone or newArray).
2737 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2738 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2739 if (alloc != nullptr) {
2740 Node* allocated_length = alloc->Ideal_length();
2741 Node* len = alloc->make_ideal_length(tary, phase);
2742 if (allocated_length != len) {
2743 // New CastII improves on this.
2744 return len;
2745 }
2746 }
2747 }
2748
2749 return nullptr;
2750 }
2751
2752 //------------------------------Identity---------------------------------------
2753 // Feed through the length in AllocateArray(...length...)._length.
2754 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2755 Node* x = LoadINode::Identity(phase);
2756 if (x != this) return x;
2757
2758 // Take apart the address into an oop and offset.
2759 // Return 'this' if we cannot.
2760 Node* adr = in(MemNode::Address);
2761 intptr_t offset = 0;
2762 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2763 if (base == nullptr) return this;
2764 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2765 if (tary == nullptr) return this;
2766
2767 // We can fetch the length directly through an AllocateArrayNode.
2768 // This works even if the length is not constant (clone or newArray).
2769 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2770 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2771 if (alloc != nullptr) {
2772 Node* allocated_length = alloc->Ideal_length();
2773 // Do not allow make_ideal_length to allocate a CastII node.
2774 Node* len = alloc->make_ideal_length(tary, phase, false);
2775 if (allocated_length == len) {
2776 // Return allocated_length only if it would not be improved by a CastII.
2777 return allocated_length;
2778 }
2779 }
2780 }
2781
2782 return this;
2783
2784 }
2785
2786 //=============================================================================
2787 //---------------------------StoreNode::make-----------------------------------
2788 // Polymorphic factory method:
2789 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) {
2790 assert((mo == unordered || mo == release), "unexpected");
2791 Compile* C = gvn.C;
2792 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2793 ctl != nullptr, "raw memory operations should have control edge");
2794
2795 switch (bt) {
2796 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2797 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2798 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2799 case T_CHAR:
2800 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2801 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2802 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2803 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2804 case T_METADATA:
2805 case T_ADDRESS:
2806 case T_OBJECT:
2807 #ifdef _LP64
2808 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2809 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2810 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2811 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2812 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2813 adr->bottom_type()->isa_rawptr())) {
2814 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2815 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2816 }
2817 #endif
2818 {
2819 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2820 }
2821 default:
2822 ShouldNotReachHere();
2823 return (StoreNode*)nullptr;
2824 }
2825 }
2826
2827 //--------------------------bottom_type----------------------------------------
2828 const Type *StoreNode::bottom_type() const {
2829 return Type::MEMORY;
2830 }
2831
2832 //------------------------------hash-------------------------------------------
2833 uint StoreNode::hash() const {
2834 // unroll addition of interesting fields
2835 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2836
2837 // Since they are not commoned, do not hash them:
2838 return NO_HASH;
2839 }
2840
2841 // Link together multiple stores (B/S/C/I) into a longer one.
2842 //
2843 // Example: _store = StoreB[i+3]
2844 //
2845 // RangeCheck[i+0] RangeCheck[i+0]
2846 // StoreB[i+0]
2847 // RangeCheck[i+3] RangeCheck[i+3]
2848 // StoreB[i+1] --> pass: fail:
2849 // StoreB[i+2] StoreI[i+0] StoreB[i+0]
2850 // StoreB[i+3]
2851 //
2852 // The 4 StoreB are merged into a single StoreI node. We have to be careful with RangeCheck[i+1]: before
2853 // the optimization, if this RangeCheck[i+1] fails, then we execute only StoreB[i+0], and then trap. After
2854 // the optimization, the new StoreI[i+0] is on the passing path of RangeCheck[i+3], and StoreB[i+0] on the
2855 // failing path.
2856 //
2857 // Note: For normal array stores, every store at first has a RangeCheck. But they can be removed with:
2858 // - RCE (RangeCheck Elimination): the RangeChecks in the loop are hoisted out and before the loop,
2859 // and possibly no RangeChecks remain between the stores.
2860 // - RangeCheck smearing: the earlier RangeChecks are adjusted such that they cover later RangeChecks,
2861 // and those later RangeChecks can be removed. Example:
2862 //
2863 // RangeCheck[i+0] RangeCheck[i+0] <- before first store
2864 // StoreB[i+0] StoreB[i+0] <- first store
2865 // RangeCheck[i+1] --> smeared --> RangeCheck[i+3] <- only RC between first and last store
2866 // StoreB[i+1] StoreB[i+1] <- second store
2867 // RangeCheck[i+2] --> removed
2868 // StoreB[i+2] StoreB[i+2]
2869 // RangeCheck[i+3] --> removed
2870 // StoreB[i+3] StoreB[i+3] <- last store
2871 //
2872 // Thus, it is a common pattern that between the first and last store in a chain
2873 // of adjacent stores there remains exactly one RangeCheck, located between the
2874 // first and the second store (e.g. RangeCheck[i+3]).
2875 //
2876 class MergePrimitiveStores : public StackObj {
2877 private:
2878 PhaseGVN* const _phase;
2879 StoreNode* const _store;
2880 // State machine with initial state Unknown
2881 // Allowed transitions:
2882 // Unknown -> Const
2883 // Unknown -> Platform
2884 // Unknown -> Reverse
2885 // Unknown -> NotAdjacent
2886 // Const -> Const
2887 // Const -> NotAdjacent
2888 // Platform -> Platform
2889 // Platform -> NotAdjacent
2890 // Reverse -> Reverse
2891 // Reverse -> NotAdjacent
2892 // NotAdjacent -> NotAdjacent
2893 enum ValueOrder : uint8_t {
2894 Unknown, // Initial state
2895 Const, // Input values are const
2896 Platform, // Platform order
2897 Reverse, // Reverse platform order
2898 NotAdjacent // Not adjacent
2899 };
2900 ValueOrder _value_order;
2901
2902 NOT_PRODUCT( const CHeapBitMap &_trace_tags; )
2903
2904 public:
2905 MergePrimitiveStores(PhaseGVN* phase, StoreNode* store) :
2906 _phase(phase), _store(store), _value_order(ValueOrder::Unknown)
2907 NOT_PRODUCT( COMMA _trace_tags(Compile::current()->directive()->trace_merge_stores_tags()) )
2908 {}
2909
2910 StoreNode* run();
2911
2912 private:
2913 bool is_compatible_store(const StoreNode* other_store) const;
2914 bool is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const;
2915 bool is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const;
2916 static bool is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase);
2917 enum CFGStatus { SuccessNoRangeCheck, SuccessWithRangeCheck, Failure };
2918 static CFGStatus cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store);
2919
2920 class Status {
2921 private:
2922 StoreNode* _found_store;
2923 bool _found_range_check;
2924
2925 Status(StoreNode* found_store, bool found_range_check)
2926 : _found_store(found_store), _found_range_check(found_range_check) {}
2927
2928 public:
2929 StoreNode* found_store() const { return _found_store; }
2930 bool found_range_check() const { return _found_range_check; }
2931 static Status make_failure() { return Status(nullptr, false); }
2932
2933 static Status make(StoreNode* found_store, const CFGStatus cfg_status) {
2934 if (cfg_status == CFGStatus::Failure) {
2935 return Status::make_failure();
2936 }
2937 return Status(found_store, cfg_status == CFGStatus::SuccessWithRangeCheck);
2938 }
2939
2940 #ifndef PRODUCT
2941 void print_on(outputStream* st) const {
2942 if (_found_store == nullptr) {
2943 st->print_cr("None");
2944 } else {
2945 st->print_cr("Found[%d %s, %s]", _found_store->_idx, _found_store->Name(),
2946 _found_range_check ? "RC" : "no-RC");
2947 }
2948 }
2949 #endif
2950 };
2951
2952 enum ValueOrder find_adjacent_input_value_order(const Node* n1, const Node* n2, const int memory_size) const;
2953 Status find_adjacent_use_store(const StoreNode* def_store) const;
2954 Status find_adjacent_def_store(const StoreNode* use_store) const;
2955 Status find_use_store(const StoreNode* def_store) const;
2956 Status find_def_store(const StoreNode* use_store) const;
2957 Status find_use_store_unidirectional(const StoreNode* def_store) const;
2958 Status find_def_store_unidirectional(const StoreNode* use_store) const;
2959
2960 void collect_merge_list(Node_List& merge_list) const;
2961 Node* make_merged_input_value(const Node_List& merge_list);
2962 StoreNode* make_merged_store(const Node_List& merge_list, Node* merged_input_value);
2963
2964 #ifndef PRODUCT
2965 // Access to TraceMergeStores tags
2966 bool is_trace(TraceMergeStores::Tag tag) const {
2967 return _trace_tags.at(tag);
2968 }
2969
2970 bool is_trace_basic() const {
2971 return is_trace(TraceMergeStores::Tag::BASIC);
2972 }
2973
2974 bool is_trace_pointer_parsing() const {
2975 return is_trace(TraceMergeStores::Tag::POINTER_PARSING);
2976 }
2977
2978 bool is_trace_pointer_aliasing() const {
2979 return is_trace(TraceMergeStores::Tag::POINTER_ALIASING);
2980 }
2981
2982 bool is_trace_pointer_adjacency() const {
2983 return is_trace(TraceMergeStores::Tag::POINTER_ADJACENCY);
2984 }
2985
2986 bool is_trace_success() const {
2987 return is_trace(TraceMergeStores::Tag::SUCCESS);
2988 }
2989 #endif
2990
2991 NOT_PRODUCT( void trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const; )
2992 };
2993
2994 StoreNode* MergePrimitiveStores::run() {
2995 // Check for B/S/C/I
2996 int opc = _store->Opcode();
2997 if (opc != Op_StoreB && opc != Op_StoreC && opc != Op_StoreI) {
2998 return nullptr;
2999 }
3000
3001 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] MergePrimitiveStores::run: "); _store->dump(); })
3002
3003 // The _store must be the "last" store in a chain. If we find a use we could merge with
3004 // then that use or a store further down is the "last" store.
3005 Status status_use = find_adjacent_use_store(_store);
3006 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect no use: "); status_use.print_on(tty); })
3007 if (status_use.found_store() != nullptr) {
3008 return nullptr;
3009 }
3010
3011 // Check if we can merge with at least one def, so that we have at least 2 stores to merge.
3012 Status status_def = find_adjacent_def_store(_store);
3013 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect def: "); status_def.print_on(tty); })
3014 Node* def_store = status_def.found_store();
3015 if (def_store == nullptr) {
3016 return nullptr;
3017 }
3018
3019 // Initialize value order
3020 _value_order = find_adjacent_input_value_order(def_store->in(MemNode::ValueIn),
3021 _store->in(MemNode::ValueIn),
3022 _store->memory_size());
3023 assert(_value_order != ValueOrder::NotAdjacent && _value_order != ValueOrder::Unknown, "Order should be checked");
3024
3025 ResourceMark rm;
3026 Node_List merge_list;
3027 collect_merge_list(merge_list);
3028
3029 Node* merged_input_value = make_merged_input_value(merge_list);
3030 if (merged_input_value == nullptr) { return nullptr; }
3031
3032 StoreNode* merged_store = make_merged_store(merge_list, merged_input_value);
3033
3034 NOT_PRODUCT( if (is_trace_success()) { trace(merge_list, merged_input_value, merged_store); } )
3035
3036 return merged_store;
3037 }
3038
3039 // Check compatibility between _store and other_store.
3040 bool MergePrimitiveStores::is_compatible_store(const StoreNode* other_store) const {
3041 int opc = _store->Opcode();
3042 assert(opc == Op_StoreB || opc == Op_StoreC || opc == Op_StoreI, "precondition");
3043
3044 if (other_store == nullptr ||
3045 _store->Opcode() != other_store->Opcode()) {
3046 return false;
3047 }
3048
3049 return true;
3050 }
3051
3052 bool MergePrimitiveStores::is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const {
3053 if (!is_adjacent_input_pair(def_store->in(MemNode::ValueIn),
3054 use_store->in(MemNode::ValueIn),
3055 def_store->memory_size())) {
3056 return false;
3057 }
3058
3059 ResourceMark rm;
3060 #ifndef PRODUCT
3061 const TraceMemPointer trace(is_trace_pointer_parsing(),
3062 is_trace_pointer_aliasing(),
3063 is_trace_pointer_adjacency(),
3064 true);
3065 #endif
3066 const MemPointer pointer_use(use_store NOT_PRODUCT(COMMA trace));
3067 const MemPointer pointer_def(def_store NOT_PRODUCT(COMMA trace));
3068 return pointer_def.is_adjacent_to_and_before(pointer_use);
3069 }
3070
3071 // Check input values n1 and n2 can be merged and return the value order
3072 MergePrimitiveStores::ValueOrder MergePrimitiveStores::find_adjacent_input_value_order(const Node* n1, const Node* n2,
3073 const int memory_size) const {
3074 // Pattern: [n1 = ConI, n2 = ConI]
3075 if (n1->Opcode() == Op_ConI && n2->Opcode() == Op_ConI) {
3076 return ValueOrder::Const;
3077 }
3078
3079 Node const *base_n2;
3080 jint shift_n2;
3081 if (!is_con_RShift(n2, base_n2, shift_n2, _phase)) {
3082 return ValueOrder::NotAdjacent;
3083 }
3084 Node const *base_n1;
3085 jint shift_n1;
3086 if (!is_con_RShift(n1, base_n1, shift_n1, _phase)) {
3087 return ValueOrder::NotAdjacent;
3088 }
3089
3090 int bits_per_store = memory_size * 8;
3091 if (base_n1 != base_n2 ||
3092 abs(shift_n1 - shift_n2) != bits_per_store ||
3093 shift_n1 % bits_per_store != 0) {
3094 // Values are not adjacent
3095 return ValueOrder::NotAdjacent;
3096 }
3097
3098 // Detect value order
3099 #ifdef VM_LITTLE_ENDIAN
3100 return shift_n1 < shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3101 : ValueOrder::Reverse; // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3102 #else
3103 return shift_n1 > shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3104 : ValueOrder::Reverse; // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3105 #endif
3106 }
3107
3108 bool MergePrimitiveStores::is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const {
3109 ValueOrder input_value_order = find_adjacent_input_value_order(n1, n2, memory_size);
3110
3111 switch (input_value_order) {
3112 case ValueOrder::NotAdjacent:
3113 return false;
3114 case ValueOrder::Reverse:
3115 if (memory_size != 1 ||
3116 !Matcher::match_rule_supported(Op_ReverseBytesS) ||
3117 !Matcher::match_rule_supported(Op_ReverseBytesI) ||
3118 !Matcher::match_rule_supported(Op_ReverseBytesL)) {
3119 // ReverseBytes are not supported by platform
3120 return false;
3121 }
3122 // fall-through.
3123 case ValueOrder::Const:
3124 case ValueOrder::Platform:
3125 if (_value_order == ValueOrder::Unknown) {
3126 // Initial state is Unknown, and we find a valid input value order
3127 return true;
3128 }
3129 // The value order can not be changed
3130 return _value_order == input_value_order;
3131 case ValueOrder::Unknown:
3132 default:
3133 ShouldNotReachHere();
3134 }
3135 return false;
3136 }
3137
3138 // Detect pattern: n = base_out >> shift_out
3139 bool MergePrimitiveStores::is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase) {
3140 assert(n != nullptr, "precondition");
3141
3142 int opc = n->Opcode();
3143 if (opc == Op_ConvL2I) {
3144 n = n->in(1);
3145 opc = n->Opcode();
3146 }
3147
3148 if ((opc == Op_RShiftI ||
3149 opc == Op_RShiftL ||
3150 opc == Op_URShiftI ||
3151 opc == Op_URShiftL) &&
3152 n->in(2)->is_ConI()) {
3153 base_out = n->in(1);
3154 shift_out = n->in(2)->get_int();
3155 // The shift must be positive:
3156 return shift_out >= 0;
3157 }
3158
3159 if (phase->type(n)->isa_int() != nullptr ||
3160 phase->type(n)->isa_long() != nullptr) {
3161 // (base >> 0)
3162 base_out = n;
3163 shift_out = 0;
3164 return true;
3165 }
3166 return false;
3167 }
3168
3169 // Check if there is nothing between the two stores, except optionally a RangeCheck leading to an uncommon trap.
3170 MergePrimitiveStores::CFGStatus MergePrimitiveStores::cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store) {
3171 assert(use_store->in(MemNode::Memory) == def_store, "use-def relationship");
3172
3173 Node* ctrl_use = use_store->in(MemNode::Control);
3174 Node* ctrl_def = def_store->in(MemNode::Control);
3175 if (ctrl_use == nullptr || ctrl_def == nullptr) {
3176 return CFGStatus::Failure;
3177 }
3178
3179 if (ctrl_use == ctrl_def) {
3180 // Same ctrl -> no RangeCheck in between.
3181 // Check: use_store must be the only use of def_store.
3182 if (def_store->outcnt() > 1) {
3183 return CFGStatus::Failure;
3184 }
3185 return CFGStatus::SuccessNoRangeCheck;
3186 }
3187
3188 // Different ctrl -> could have RangeCheck in between.
3189 // Check: 1. def_store only has these uses: use_store and MergeMem for uncommon trap, and
3190 // 2. ctrl separated by RangeCheck.
3191 if (def_store->outcnt() != 2) {
3192 return CFGStatus::Failure; // Cannot have exactly these uses: use_store and MergeMem for uncommon trap.
3193 }
3194 int use_store_out_idx = def_store->raw_out(0) == use_store ? 0 : 1;
3195 Node* merge_mem = def_store->raw_out(1 - use_store_out_idx)->isa_MergeMem();
3196 if (merge_mem == nullptr ||
3197 merge_mem->outcnt() != 1) {
3198 return CFGStatus::Failure; // Does not have MergeMem for uncommon trap.
3199 }
3200 if (!ctrl_use->is_IfProj() ||
3201 !ctrl_use->in(0)->is_RangeCheck() ||
3202 ctrl_use->in(0)->outcnt() != 2) {
3203 return CFGStatus::Failure; // Not RangeCheck.
3204 }
3205 IfProjNode* other_proj = ctrl_use->as_IfProj()->other_if_proj();
3206 Node* trap = other_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check);
3207 if (trap != merge_mem->unique_out() ||
3208 ctrl_use->in(0)->in(0) != ctrl_def) {
3209 return CFGStatus::Failure; // Not RangeCheck with merge_mem leading to uncommon trap.
3210 }
3211
3212 return CFGStatus::SuccessWithRangeCheck;
3213 }
3214
3215 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_use_store(const StoreNode* def_store) const {
3216 Status status_use = find_use_store(def_store);
3217 StoreNode* use_store = status_use.found_store();
3218 if (use_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3219 return Status::make_failure();
3220 }
3221 return status_use;
3222 }
3223
3224 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_def_store(const StoreNode* use_store) const {
3225 Status status_def = find_def_store(use_store);
3226 StoreNode* def_store = status_def.found_store();
3227 if (def_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3228 return Status::make_failure();
3229 }
3230 return status_def;
3231 }
3232
3233 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store(const StoreNode* def_store) const {
3234 Status status_use = find_use_store_unidirectional(def_store);
3235
3236 #ifdef ASSERT
3237 StoreNode* use_store = status_use.found_store();
3238 if (use_store != nullptr) {
3239 Status status_def = find_def_store_unidirectional(use_store);
3240 assert(status_def.found_store() == def_store &&
3241 status_def.found_range_check() == status_use.found_range_check(),
3242 "find_use_store and find_def_store must be symmetric");
3243 }
3244 #endif
3245
3246 return status_use;
3247 }
3248
3249 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store(const StoreNode* use_store) const {
3250 Status status_def = find_def_store_unidirectional(use_store);
3251
3252 #ifdef ASSERT
3253 StoreNode* def_store = status_def.found_store();
3254 if (def_store != nullptr) {
3255 Status status_use = find_use_store_unidirectional(def_store);
3256 assert(status_use.found_store() == use_store &&
3257 status_use.found_range_check() == status_def.found_range_check(),
3258 "find_use_store and find_def_store must be symmetric");
3259 }
3260 #endif
3261
3262 return status_def;
3263 }
3264
3265 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store_unidirectional(const StoreNode* def_store) const {
3266 assert(is_compatible_store(def_store), "precondition: must be compatible with _store");
3267
3268 for (DUIterator_Fast imax, i = def_store->fast_outs(imax); i < imax; i++) {
3269 StoreNode* use_store = def_store->fast_out(i)->isa_Store();
3270 if (is_compatible_store(use_store)) {
3271 return Status::make(use_store, cfg_status_for_pair(use_store, def_store));
3272 }
3273 }
3274
3275 return Status::make_failure();
3276 }
3277
3278 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store_unidirectional(const StoreNode* use_store) const {
3279 assert(is_compatible_store(use_store), "precondition: must be compatible with _store");
3280
3281 StoreNode* def_store = use_store->in(MemNode::Memory)->isa_Store();
3282 if (!is_compatible_store(def_store)) {
3283 return Status::make_failure();
3284 }
3285
3286 return Status::make(def_store, cfg_status_for_pair(use_store, def_store));
3287 }
3288
3289 void MergePrimitiveStores::collect_merge_list(Node_List& merge_list) const {
3290 // The merged store can be at most 8 bytes.
3291 const uint merge_list_max_size = 8 / _store->memory_size();
3292 assert(merge_list_max_size >= 2 &&
3293 merge_list_max_size <= 8 &&
3294 is_power_of_2(merge_list_max_size),
3295 "must be 2, 4 or 8");
3296
3297 // Traverse up the chain of adjacent def stores.
3298 StoreNode* current = _store;
3299 merge_list.push(current);
3300 while (current != nullptr && merge_list.size() < merge_list_max_size) {
3301 Status status = find_adjacent_def_store(current);
3302 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] find def: "); status.print_on(tty); })
3303
3304 current = status.found_store();
3305 if (current != nullptr) {
3306 merge_list.push(current);
3307
3308 // We can have at most one RangeCheck.
3309 if (status.found_range_check()) {
3310 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found RangeCheck, stop traversal."); })
3311 break;
3312 }
3313 }
3314 }
3315
3316 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found:"); merge_list.dump(); })
3317
3318 // Truncate the merge_list to a power of 2.
3319 const uint pow2size = round_down_power_of_2(merge_list.size());
3320 assert(pow2size >= 2, "must be merging at least 2 stores");
3321 while (merge_list.size() > pow2size) { merge_list.pop(); }
3322
3323 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] truncated:"); merge_list.dump(); })
3324 }
3325
3326 // Merge the input values of the smaller stores to a single larger input value.
3327 Node* MergePrimitiveStores::make_merged_input_value(const Node_List& merge_list) {
3328 int new_memory_size = _store->memory_size() * merge_list.size();
3329 Node* first = merge_list.at(merge_list.size()-1);
3330 Node* merged_input_value = nullptr;
3331 if (_store->in(MemNode::ValueIn)->Opcode() == Op_ConI) {
3332 assert(_value_order == ValueOrder::Const, "must match");
3333 // Pattern: [ConI, ConI, ...] -> new constant
3334 jlong con = 0;
3335 jlong bits_per_store = _store->memory_size() * 8;
3336 jlong mask = (((jlong)1) << bits_per_store) - 1;
3337 for (uint i = 0; i < merge_list.size(); i++) {
3338 jlong con_i = merge_list.at(i)->in(MemNode::ValueIn)->get_int();
3339 #ifdef VM_LITTLE_ENDIAN
3340 con = con << bits_per_store;
3341 con = con | (mask & con_i);
3342 #else // VM_LITTLE_ENDIAN
3343 con_i = (mask & con_i) << (i * bits_per_store);
3344 con = con | con_i;
3345 #endif // VM_LITTLE_ENDIAN
3346 }
3347 merged_input_value = _phase->longcon(con);
3348 } else {
3349 assert(_value_order == ValueOrder::Platform || _value_order == ValueOrder::Reverse, "must match");
3350 // Pattern: [base >> 24, base >> 16, base >> 8, base] -> base
3351 // | |
3352 // _store first
3353 //
3354 Node* hi = _store->in(MemNode::ValueIn);
3355 Node* lo = first->in(MemNode::ValueIn);
3356 #ifndef VM_LITTLE_ENDIAN
3357 // `_store` and `first` are swapped in the diagram above
3358 swap(hi, lo);
3359 #endif // !VM_LITTLE_ENDIAN
3360 if (_value_order == ValueOrder::Reverse) {
3361 swap(hi, lo);
3362 }
3363 Node const* hi_base;
3364 jint hi_shift;
3365 merged_input_value = lo;
3366 bool is_true = is_con_RShift(hi, hi_base, hi_shift, _phase);
3367 assert(is_true, "must detect con RShift");
3368 if (merged_input_value != hi_base && merged_input_value->Opcode() == Op_ConvL2I) {
3369 // look through
3370 merged_input_value = merged_input_value->in(1);
3371 }
3372 if (merged_input_value != hi_base) {
3373 // merged_input_value is not the base
3374 return nullptr;
3375 }
3376 }
3377
3378 if (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size <= 4) {
3379 // Example:
3380 //
3381 // long base = ...;
3382 // a[0] = (byte)(base >> 0);
3383 // a[1] = (byte)(base >> 8);
3384 //
3385 merged_input_value = _phase->transform(new ConvL2INode(merged_input_value));
3386 }
3387
3388 assert((_phase->type(merged_input_value)->isa_int() != nullptr && new_memory_size <= 4) ||
3389 (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size == 8),
3390 "merged_input_value is either int or long, and new_memory_size is small enough");
3391
3392 if (_value_order == ValueOrder::Reverse) {
3393 assert(_store->memory_size() == 1, "only implemented for bytes");
3394 if (new_memory_size == 8) {
3395 merged_input_value = _phase->transform(new ReverseBytesLNode(merged_input_value));
3396 } else if (new_memory_size == 4) {
3397 merged_input_value = _phase->transform(new ReverseBytesINode(merged_input_value));
3398 } else {
3399 assert(new_memory_size == 2, "sanity check");
3400 merged_input_value = _phase->transform(new ReverseBytesSNode(merged_input_value));
3401 }
3402 }
3403 return merged_input_value;
3404 }
3405
3406 // //
3407 // first_ctrl first_mem first_adr first_ctrl first_mem first_adr //
3408 // | | | | | | //
3409 // | | | | +---------------+ | //
3410 // | | | | | | | //
3411 // | | +---------+ | | +---------------+ //
3412 // | | | | | | | | //
3413 // +--------------+ | | v1 +------------------------------+ | | v1 //
3414 // | | | | | | | | | | | | //
3415 // RangeCheck first_store RangeCheck | | first_store //
3416 // | | | | | | | //
3417 // last_ctrl | +----> unc_trap last_ctrl | | +----> unc_trap //
3418 // | | ===> | | | //
3419 // +--------------+ | a2 v2 | | | //
3420 // | | | | | | | | //
3421 // | second_store | | | //
3422 // | | | | | [v1 v2 ... vn] //
3423 // ... ... | | | | //
3424 // | | | | | v //
3425 // +--------------+ | an vn +--------------+ | | merged_input_value //
3426 // | | | | | | | | //
3427 // last_store (= _store) merged_store //
3428 // //
3429 StoreNode* MergePrimitiveStores::make_merged_store(const Node_List& merge_list, Node* merged_input_value) {
3430 Node* first_store = merge_list.at(merge_list.size()-1);
3431 Node* last_ctrl = _store->in(MemNode::Control); // after (optional) RangeCheck
3432 Node* first_mem = first_store->in(MemNode::Memory);
3433 Node* first_adr = first_store->in(MemNode::Address);
3434
3435 const TypePtr* new_adr_type = _store->adr_type();
3436
3437 int new_memory_size = _store->memory_size() * merge_list.size();
3438 BasicType bt = T_ILLEGAL;
3439 switch (new_memory_size) {
3440 case 2: bt = T_SHORT; break;
3441 case 4: bt = T_INT; break;
3442 case 8: bt = T_LONG; break;
3443 }
3444
3445 StoreNode* merged_store = StoreNode::make(*_phase, last_ctrl, first_mem, first_adr,
3446 new_adr_type, merged_input_value, bt, MemNode::unordered);
3447
3448 // Marking the store mismatched is sufficient to prevent reordering, since array stores
3449 // are all on the same slice. Hence, we need no barriers.
3450 merged_store->set_mismatched_access();
3451
3452 // Constants above may now also be be packed -> put candidate on worklist
3453 _phase->is_IterGVN()->_worklist.push(first_mem);
3454
3455 return merged_store;
3456 }
3457
3458 #ifndef PRODUCT
3459 void MergePrimitiveStores::trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const {
3460 stringStream ss;
3461 ss.print_cr("[TraceMergeStores]: Replace");
3462 for (int i = (int)merge_list.size() - 1; i >= 0; i--) {
3463 merge_list.at(i)->dump("\n", false, &ss);
3464 }
3465 ss.print_cr("[TraceMergeStores]: with");
3466 merged_input_value->dump("\n", false, &ss);
3467 merged_store->dump("\n", false, &ss);
3468 tty->print("%s", ss.as_string());
3469 }
3470 #endif
3471
3472 //------------------------------Ideal------------------------------------------
3473 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
3474 // When a store immediately follows a relevant allocation/initialization,
3475 // try to capture it into the initialization, or hoist it above.
3476 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3477 Node* p = MemNode::Ideal_common(phase, can_reshape);
3478 if (p) return (p == NodeSentinel) ? nullptr : p;
3479
3480 Node* mem = in(MemNode::Memory);
3481 Node* address = in(MemNode::Address);
3482 Node* value = in(MemNode::ValueIn);
3483 // Back-to-back stores to same address? Fold em up. Generally
3484 // unsafe if I have intervening uses.
3485 {
3486 Node* st = mem;
3487 // If Store 'st' has more than one use, we cannot fold 'st' away.
3488 // For example, 'st' might be the final state at a conditional
3489 // return. Or, 'st' might be used by some node which is live at
3490 // the same time 'st' is live, which might be unschedulable. So,
3491 // require exactly ONE user until such time as we clone 'mem' for
3492 // each of 'mem's uses (thus making the exactly-1-user-rule hold
3493 // true).
3494 while (st->is_Store() && st->outcnt() == 1) {
3495 // Looking at a dead closed cycle of memory?
3496 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
3497 assert(Opcode() == st->Opcode() ||
3498 st->Opcode() == Op_StoreVector ||
3499 Opcode() == Op_StoreVector ||
3500 st->Opcode() == Op_StoreVectorScatter ||
3501 Opcode() == Op_StoreVectorScatter ||
3502 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
3503 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
3504 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
3505 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
3506 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
3507
3508 if (st->in(MemNode::Address)->eqv_uncast(address) &&
3509 st->as_Store()->memory_size() <= this->memory_size()) {
3510 Node* use = st->raw_out(0);
3511 if (phase->is_IterGVN()) {
3512 phase->is_IterGVN()->rehash_node_delayed(use);
3513 }
3514 // It's OK to do this in the parser, since DU info is always accurate,
3515 // and the parser always refers to nodes via SafePointNode maps.
3516 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
3517 return this;
3518 }
3519 st = st->in(MemNode::Memory);
3520 }
3521 }
3522
3523
3524 // Capture an unaliased, unconditional, simple store into an initializer.
3525 // Or, if it is independent of the allocation, hoist it above the allocation.
3526 if (ReduceFieldZeroing && /*can_reshape &&*/
3527 mem->is_Proj() && mem->in(0)->is_Initialize()) {
3528 InitializeNode* init = mem->in(0)->as_Initialize();
3529 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
3530 if (offset > 0) {
3531 Node* moved = init->capture_store(this, offset, phase, can_reshape);
3532 // If the InitializeNode captured me, it made a raw copy of me,
3533 // and I need to disappear.
3534 if (moved != nullptr) {
3535 // %%% hack to ensure that Ideal returns a new node:
3536 mem = MergeMemNode::make(mem);
3537 return mem; // fold me away
3538 }
3539 }
3540 }
3541
3542 // Fold reinterpret cast into memory operation:
3543 // StoreX mem (MoveY2X v) => StoreY mem v
3544 if (value->is_Move()) {
3545 const Type* vt = value->in(1)->bottom_type();
3546 if (has_reinterpret_variant(vt)) {
3547 if (phase->C->post_loop_opts_phase()) {
3548 return convert_to_reinterpret_store(*phase, value->in(1), vt);
3549 } else {
3550 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
3551 }
3552 }
3553 }
3554
3555 if (MergeStores && UseUnalignedAccesses) {
3556 if (phase->C->merge_stores_phase()) {
3557 MergePrimitiveStores merge(phase, this);
3558 Node* progress = merge.run();
3559 if (progress != nullptr) { return progress; }
3560 } else {
3561 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
3562 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
3563 // the stores, and we merge the wrong sequence of stores.
3564 // Example:
3565 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
3566 // Apply MergeStores:
3567 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
3568 // Remove more RangeChecks:
3569 // StoreI [ StoreL ] StoreI
3570 // But now it would have been better to do this instead:
3571 // [ StoreL ] [ StoreL ]
3572 phase->C->record_for_merge_stores_igvn(this);
3573 }
3574 }
3575
3576 return nullptr; // No further progress
3577 }
3578
3579 //------------------------------Value-----------------------------------------
3580 const Type* StoreNode::Value(PhaseGVN* phase) const {
3581 // Either input is TOP ==> the result is TOP
3582 const Type *t1 = phase->type( in(MemNode::Memory) );
3583 if( t1 == Type::TOP ) return Type::TOP;
3584 const Type *t2 = phase->type( in(MemNode::Address) );
3585 if( t2 == Type::TOP ) return Type::TOP;
3586 const Type *t3 = phase->type( in(MemNode::ValueIn) );
3587 if( t3 == Type::TOP ) return Type::TOP;
3588 return Type::MEMORY;
3589 }
3590
3591 //------------------------------Identity---------------------------------------
3592 // Remove redundant stores:
3593 // Store(m, p, Load(m, p)) changes to m.
3594 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
3595 Node* StoreNode::Identity(PhaseGVN* phase) {
3596 Node* mem = in(MemNode::Memory);
3597 Node* adr = in(MemNode::Address);
3598 Node* val = in(MemNode::ValueIn);
3599
3600 Node* result = this;
3601
3602 // Load then Store? Then the Store is useless
3603 if (val->is_Load() &&
3604 val->in(MemNode::Address)->eqv_uncast(adr) &&
3605 val->in(MemNode::Memory )->eqv_uncast(mem) &&
3606 val->as_Load()->store_Opcode() == Opcode()) {
3607 if (!is_StoreVector()) {
3608 result = mem;
3609 } else if (Opcode() == Op_StoreVector && val->Opcode() == Op_LoadVector &&
3610 as_StoreVector()->vect_type() == val->as_LoadVector()->vect_type()) {
3611 // Ensure both are not masked accesses or gathers/scatters and vector types are the same
3612 result = mem;
3613 }
3614 }
3615
3616 // Two stores in a row of the same value?
3617 if (result == this &&
3618 mem->is_Store() &&
3619 mem->in(MemNode::Address)->eqv_uncast(adr) &&
3620 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
3621 mem->Opcode() == Opcode()) {
3622 if (!is_StoreVector()) {
3623 result = mem;
3624 } else {
3625 const StoreVectorNode* store_vector = as_StoreVector();
3626 const StoreVectorNode* mem_vector = mem->as_StoreVector();
3627 const Node* store_indices = store_vector->indices();
3628 const Node* mem_indices = mem_vector->indices();
3629 const Node* store_mask = store_vector->mask();
3630 const Node* mem_mask = mem_vector->mask();
3631 // Ensure types, indices, and masks match
3632 if (store_vector->vect_type() == mem_vector->vect_type() &&
3633 ((store_indices == nullptr) == (mem_indices == nullptr) &&
3634 (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) &&
3635 ((store_mask == nullptr) == (mem_mask == nullptr) &&
3636 (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) {
3637 result = mem;
3638 }
3639 }
3640 }
3641
3642 // Store of zero anywhere into a freshly-allocated object?
3643 // Then the store is useless.
3644 // (It must already have been captured by the InitializeNode.)
3645 if (result == this &&
3646 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
3647 // a newly allocated object is already all-zeroes everywhere
3648 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
3649 result = mem;
3650 }
3651
3652 if (result == this) {
3653 // the store may also apply to zero-bits in an earlier object
3654 Node* prev_mem = find_previous_store(phase);
3655 // Steps (a), (b): Walk past independent stores to find an exact match.
3656 if (prev_mem != nullptr) {
3657 if (prev_mem->is_top()) {
3658 // find_previous_store returns top when the access is dead
3659 return prev_mem;
3660 }
3661 Node* prev_val = can_see_stored_value(prev_mem, phase);
3662 if (prev_val != nullptr && prev_val == val) {
3663 // prev_val and val might differ by a cast; it would be good
3664 // to keep the more informative of the two.
3665 result = mem;
3666 }
3667 }
3668 }
3669 }
3670
3671 PhaseIterGVN* igvn = phase->is_IterGVN();
3672 if (result != this && igvn != nullptr) {
3673 MemBarNode* trailing = trailing_membar();
3674 if (trailing != nullptr) {
3675 #ifdef ASSERT
3676 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
3677 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects");
3678 #endif
3679 trailing->remove(igvn);
3680 }
3681 }
3682
3683 return result;
3684 }
3685
3686 //------------------------------match_edge-------------------------------------
3687 // Do we Match on this edge index or not? Match only memory & value
3688 uint StoreNode::match_edge(uint idx) const {
3689 return idx == MemNode::Address || idx == MemNode::ValueIn;
3690 }
3691
3692 //------------------------------cmp--------------------------------------------
3693 // Do not common stores up together. They generally have to be split
3694 // back up anyways, so do not bother.
3695 bool StoreNode::cmp( const Node &n ) const {
3696 return (&n == this); // Always fail except on self
3697 }
3698
3699 //------------------------------Ideal_masked_input-----------------------------
3700 // Check for a useless mask before a partial-word store
3701 // (StoreB ... (AndI valIn conIa) )
3702 // If (conIa & mask == mask) this simplifies to
3703 // (StoreB ... (valIn) )
3704 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
3705 Node *val = in(MemNode::ValueIn);
3706 if( val->Opcode() == Op_AndI ) {
3707 const TypeInt *t = phase->type( val->in(2) )->isa_int();
3708 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
3709 set_req_X(MemNode::ValueIn, val->in(1), phase);
3710 return this;
3711 }
3712 }
3713 return nullptr;
3714 }
3715
3716
3717 //------------------------------Ideal_sign_extended_input----------------------
3718 // Check for useless sign-extension before a partial-word store
3719 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
3720 // If (conIL == conIR && conIR <= num_rejected_bits) this simplifies to
3721 // (StoreB ... (v))
3722 // If (conIL > conIR) under some conditions, it can be simplified into
3723 // (StoreB ... (LShiftI _ v (conIL - conIR)))
3724 // This case happens when the value of the store was itself a left shift, that
3725 // gets merged into the inner left shift of the sign-extension. For instance,
3726 // if we have
3727 // array_of_shorts[0] = (short)(v << 2)
3728 // We get a structure such as:
3729 // (StoreB ... (RShiftI _ (LShiftI _ (LShiftI _ v 2) 16) 16))
3730 // that is simplified into
3731 // (StoreB ... (RShiftI _ (LShiftI _ v 18) 16)).
3732 // It is thus useful to handle cases where conIL > conIR. But this simplification
3733 // does not always hold. Let's see in which cases it's valid.
3734 //
3735 // Let's assume we have the following 32 bits integer v:
3736 // +----------------------------------+
3737 // | v[0..31] |
3738 // +----------------------------------+
3739 // 31 0
3740 // that will be stuffed in 8 bits byte after a shift left and a shift right of
3741 // potentially different magnitudes.
3742 // We denote num_rejected_bits the number of bits of the discarded part. In this
3743 // case, num_rejected_bits == 24.
3744 //
3745 // Statement (proved further below in case analysis):
3746 // Given:
3747 // - 0 <= conIL < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
3748 // - 0 <= conIR < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
3749 // - conIL >= conIR
3750 // - num_rejected_bits >= conIR
3751 // Then this form:
3752 // (RShiftI _ (LShiftI _ v conIL) conIR)
3753 // can be replaced with this form:
3754 // (LShiftI _ v (conIL-conIR))
3755 //
3756 // Note: We only have to show that the non-rejected lowest bits (8 bits for byte)
3757 // have to be correct, as the higher bits are rejected / truncated by the store.
3758 //
3759 // The hypotheses
3760 // 0 <= conIL < BitsPerJavaInteger
3761 // 0 <= conIR < BitsPerJavaInteger
3762 // are ensured by maskShiftAmount (called from ::Ideal of shift nodes). Indeed,
3763 // (v << 31) << 2 must be simplified into 0, not into v << 33 (which is equivalent
3764 // to v << 1).
3765 //
3766 //
3767 // If you don't like case analysis, jump after the conclusion.
3768 // ### Case 1 : conIL == conIR
3769 // ###### Case 1.1: conIL == conIR == num_rejected_bits
3770 // If we do the shift left then right by 24 bits, we get:
3771 // after: v << 24
3772 // +---------+------------------------+
3773 // | v[0..7] | 0 |
3774 // +---------+------------------------+
3775 // 31 24 23 0
3776 // after: (v << 24) >> 24
3777 // +------------------------+---------+
3778 // | sign bit | v[0..7] |
3779 // +------------------------+---------+
3780 // 31 8 7 0
3781 // The non-rejected bits (bits kept by the store, that is the 8 lower bits of the
3782 // result) are the same before and after, so, indeed, simplifying is correct.
3783
3784 // ###### Case 1.2: conIL == conIR < num_rejected_bits
3785 // If we do the shift left then right by 22 bits, we get:
3786 // after: v << 22
3787 // +---------+------------------------+
3788 // | v[0..9] | 0 |
3789 // +---------+------------------------+
3790 // 31 22 21 0
3791 // after: (v << 22) >> 22
3792 // +------------------------+---------+
3793 // | sign bit | v[0..9] |
3794 // +------------------------+---------+
3795 // 31 10 9 0
3796 // The non-rejected bits are the 8 lower bits of v. The bits 8 and 9 of v are still
3797 // present in (v << 22) >> 22 but will be dropped by the store. The simplification is
3798 // still correct.
3799
3800 // ###### But! Case 1.3: conIL == conIR > num_rejected_bits
3801 // If we do the shift left then right by 26 bits, we get:
3802 // after: v << 26
3803 // +---------+------------------------+
3804 // | v[0..5] | 0 |
3805 // +---------+------------------------+
3806 // 31 26 25 0
3807 // after: (v << 26) >> 26
3808 // +------------------------+---------+
3809 // | sign bit | v[0..5] |
3810 // +------------------------+---------+
3811 // 31 6 5 0
3812 // The non-rejected bits are made of
3813 // - 0-5 => the bits 0 to 5 of v
3814 // - 6-7 => the sign bit of v[0..5] (that is v[5])
3815 // Simplifying this as v is not correct.
3816 // The condition conIR <= num_rejected_bits is indeed necessary in Case 1
3817 //
3818 // ### Case 2: conIL > conIR
3819 // ###### Case 2.1: num_rejected_bits == conIR
3820 // We take conIL == 26 for this example.
3821 // after: v << 26
3822 // +---------+------------------------+
3823 // | v[0..5] | 0 |
3824 // +---------+------------------------+
3825 // 31 26 25 0
3826 // after: (v << 26) >> 24
3827 // +------------------+---------+-----+
3828 // | sign bit | v[0..5] | 0 |
3829 // +------------------+---------+-----+
3830 // 31 8 7 2 1 0
3831 // The non-rejected bits are the 8 lower ones of (v << conIL - conIR).
3832 // The bits 6 and 7 of v have been thrown away after the shift left.
3833 // The simplification is still correct.
3834 //
3835 // ###### Case 2.2: num_rejected_bits > conIR.
3836 // Let's say conIL == 26 and conIR == 22.
3837 // after: v << 26
3838 // +---------+------------------------+
3839 // | v[0..5] | 0 |
3840 // +---------+------------------------+
3841 // 31 26 25 0
3842 // after: (v << 26) >> 22
3843 // +------------------+---------+-----+
3844 // | sign bit | v[0..5] | 0 |
3845 // +------------------+---------+-----+
3846 // 31 10 9 4 3 0
3847 // The bits non-rejected by the store are exactly the 8 lower ones of (v << (conIL - conIR)):
3848 // - 0-3 => 0
3849 // - 4-7 => bits 0 to 3 of v
3850 // The simplification is still correct.
3851 // The bits 4 and 5 of v are still present in (v << (conIL - conIR)) but they don't
3852 // matter as they are not in the 8 lower bits: they will be cut out by the store.
3853 //
3854 // ###### But! Case 2.3: num_rejected_bits < conIR.
3855 // Let's see that this case is not as easy to simplify.
3856 // Let's say conIL == 28 and conIR == 26.
3857 // after: v << 28
3858 // +---------+------------------------+
3859 // | v[0..3] | 0 |
3860 // +---------+------------------------+
3861 // 31 28 27 0
3862 // after: (v << 28) >> 26
3863 // +------------------+---------+-----+
3864 // | sign bit | v[0..3] | 0 |
3865 // +------------------+---------+-----+
3866 // 31 6 5 2 1 0
3867 // The non-rejected bits are made of
3868 // - 0-1 => 0
3869 // - 2-5 => the bits 0 to 3 of v
3870 // - 6-7 => the sign bit of v[0..3] (that is v[3])
3871 // Simplifying this as (v << 2) is not correct.
3872 // The condition conIR <= num_rejected_bits is indeed necessary in Case 2.
3873 //
3874 // ### Conclusion:
3875 // Our hypotheses are indeed sufficient:
3876 // - 0 <= conIL < BitsPerJavaInteger
3877 // - 0 <= conIR < BitsPerJavaInteger
3878 // - conIL >= conIR
3879 // - num_rejected_bits >= conIR
3880 //
3881 // ### A rationale without case analysis:
3882 // After the shift left, conIL upper bits of v are discarded and conIL lower bit
3883 // zeroes are added. After the shift right, conIR lower bits of the previous result
3884 // are discarded. If conIL >= conIR, we discard only the zeroes we made up during
3885 // the shift left, but if conIL < conIR, then we discard also lower bits of v. But
3886 // the point of the simplification is to get an expression of the form
3887 // (v << (conIL - conIR)). This expression discard only higher bits of v, thus the
3888 // simplification is not correct if conIL < conIR.
3889 //
3890 // Moreover, after the shift right, the higher bit of (v << conIL) is repeated on the
3891 // conIR higher bits of ((v << conIL) >> conIR), it's the sign-extension. If
3892 // conIR > num_rejected_bits, then at least one artificial copy of this sign bit will
3893 // be in the window of the store. Thus ((v << conIL) >> conIR) is not equivalent to
3894 // (v << (conIL-conIR)) if conIR > num_rejected_bits.
3895 //
3896 // We do not treat the case conIL < conIR here since the point of this function is
3897 // to skip sign-extensions (that is conIL == conIR == num_rejected_bits). The need
3898 // of treating conIL > conIR comes from the cases where the sign-extended value is
3899 // also left-shift expression. Computing the sign-extension of a right-shift expression
3900 // doesn't yield a situation such as
3901 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
3902 // where conIL < conIR.
3903 Node* StoreNode::Ideal_sign_extended_input(PhaseGVN* phase, int num_rejected_bits) {
3904 Node* shr = in(MemNode::ValueIn);
3905 if (shr->Opcode() == Op_RShiftI) {
3906 const TypeInt* conIR = phase->type(shr->in(2))->isa_int();
3907 if (conIR != nullptr && conIR->is_con() && conIR->get_con() >= 0 && conIR->get_con() < BitsPerJavaInteger && conIR->get_con() <= num_rejected_bits) {
3908 Node* shl = shr->in(1);
3909 if (shl->Opcode() == Op_LShiftI) {
3910 const TypeInt* conIL = phase->type(shl->in(2))->isa_int();
3911 if (conIL != nullptr && conIL->is_con() && conIL->get_con() >= 0 && conIL->get_con() < BitsPerJavaInteger) {
3912 if (conIL->get_con() == conIR->get_con()) {
3913 set_req_X(MemNode::ValueIn, shl->in(1), phase);
3914 return this;
3915 }
3916 if (conIL->get_con() > conIR->get_con()) {
3917 Node* new_shl = phase->transform(new LShiftINode(shl->in(1), phase->intcon(conIL->get_con() - conIR->get_con())));
3918 set_req_X(MemNode::ValueIn, new_shl, phase);
3919 return this;
3920 }
3921 }
3922 }
3923 }
3924 }
3925 return nullptr;
3926 }
3927
3928 //------------------------------value_never_loaded-----------------------------------
3929 // Determine whether there are any possible loads of the value stored.
3930 // For simplicity, we actually check if there are any loads from the
3931 // address stored to, not just for loads of the value stored by this node.
3932 //
3933 bool StoreNode::value_never_loaded(PhaseValues* phase) const {
3934 Node *adr = in(Address);
3935 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
3936 if (adr_oop == nullptr)
3937 return false;
3938 if (!adr_oop->is_known_instance_field())
3939 return false; // if not a distinct instance, there may be aliases of the address
3940 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
3941 Node *use = adr->fast_out(i);
3942 if (use->is_Load() || use->is_LoadStore()) {
3943 return false;
3944 }
3945 }
3946 return true;
3947 }
3948
3949 MemBarNode* StoreNode::trailing_membar() const {
3950 if (is_release()) {
3951 MemBarNode* trailing_mb = nullptr;
3952 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3953 Node* u = fast_out(i);
3954 if (u->is_MemBar()) {
3955 if (u->as_MemBar()->trailing_store()) {
3956 assert(u->Opcode() == Op_MemBarVolatile, "");
3957 assert(trailing_mb == nullptr, "only one");
3958 trailing_mb = u->as_MemBar();
3959 #ifdef ASSERT
3960 Node* leading = u->as_MemBar()->leading_membar();
3961 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3962 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
3963 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
3964 #endif
3965 } else {
3966 assert(u->as_MemBar()->standalone(), "");
3967 }
3968 }
3969 }
3970 return trailing_mb;
3971 }
3972 return nullptr;
3973 }
3974
3975
3976 //=============================================================================
3977 //------------------------------Ideal------------------------------------------
3978 // If the store is from an AND mask that leaves the low bits untouched, then
3979 // we can skip the AND operation. If the store is from a sign-extension
3980 // (a left shift, then right shift) we can skip both.
3981 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
3982 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
3983 if( progress != nullptr ) return progress;
3984
3985 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
3986 if( progress != nullptr ) return progress;
3987
3988 // Finally check the default case
3989 return StoreNode::Ideal(phase, can_reshape);
3990 }
3991
3992 //=============================================================================
3993 //------------------------------Ideal------------------------------------------
3994 // If the store is from an AND mask that leaves the low bits untouched, then
3995 // we can skip the AND operation
3996 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
3997 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
3998 if( progress != nullptr ) return progress;
3999
4000 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
4001 if( progress != nullptr ) return progress;
4002
4003 // Finally check the default case
4004 return StoreNode::Ideal(phase, can_reshape);
4005 }
4006
4007 //=============================================================================
4008 //----------------------------------SCMemProjNode------------------------------
4009 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
4010 {
4011 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
4012 return Type::TOP;
4013 }
4014 return bottom_type();
4015 }
4016
4017 //=============================================================================
4018 //----------------------------------LoadStoreNode------------------------------
4019 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
4020 : Node(required),
4021 _type(rt),
4022 _barrier_data(0)
4023 {
4024 init_req(MemNode::Control, c );
4025 init_req(MemNode::Memory , mem);
4026 init_req(MemNode::Address, adr);
4027 init_req(MemNode::ValueIn, val);
4028 init_class_id(Class_LoadStore);
4029 DEBUG_ONLY(_adr_type = at; adr_type();)
4030 }
4031
4032 //------------------------------Value-----------------------------------------
4033 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
4034 // Either input is TOP ==> the result is TOP
4035 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
4036 return Type::TOP;
4037 }
4038 const Type* t = phase->type(in(MemNode::Memory));
4039 if (t == Type::TOP) {
4040 return Type::TOP;
4041 }
4042 t = phase->type(in(MemNode::Address));
4043 if (t == Type::TOP) {
4044 return Type::TOP;
4045 }
4046 t = phase->type(in(MemNode::ValueIn));
4047 if (t == Type::TOP) {
4048 return Type::TOP;
4049 }
4050 return bottom_type();
4051 }
4052
4053 const TypePtr* LoadStoreNode::adr_type() const {
4054 const TypePtr* cross_check = DEBUG_ONLY(_adr_type) NOT_DEBUG(nullptr);
4055 return MemNode::calculate_adr_type(in(MemNode::Address)->bottom_type(), cross_check);
4056 }
4057
4058 uint LoadStoreNode::ideal_reg() const {
4059 return _type->ideal_reg();
4060 }
4061
4062 // This method conservatively checks if the result of a LoadStoreNode is
4063 // used, that is, if it returns true, then it is definitely the case that
4064 // the result of the node is not needed.
4065 // For example, GetAndAdd can be matched into a lock_add instead of a
4066 // lock_xadd if the result of LoadStoreNode::result_not_used() is true
4067 bool LoadStoreNode::result_not_used() const {
4068 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4069 Node *x = fast_out(i);
4070 if (x->Opcode() == Op_SCMemProj) {
4071 continue;
4072 }
4073 if (x->bottom_type() == TypeTuple::MEMBAR &&
4074 !x->is_Call() &&
4075 x->Opcode() != Op_Blackhole) {
4076 continue;
4077 }
4078 return false;
4079 }
4080 return true;
4081 }
4082
4083 MemBarNode* LoadStoreNode::trailing_membar() const {
4084 MemBarNode* trailing = nullptr;
4085 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4086 Node* u = fast_out(i);
4087 if (u->is_MemBar()) {
4088 if (u->as_MemBar()->trailing_load_store()) {
4089 assert(u->Opcode() == Op_MemBarAcquire, "");
4090 assert(trailing == nullptr, "only one");
4091 trailing = u->as_MemBar();
4092 #ifdef ASSERT
4093 Node* leading = trailing->leading_membar();
4094 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
4095 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
4096 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
4097 #endif
4098 } else {
4099 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
4100 }
4101 }
4102 }
4103
4104 return trailing;
4105 }
4106
4107 uint LoadStoreNode::size_of() const { return sizeof(*this); }
4108
4109 //=============================================================================
4110 //----------------------------------LoadStoreConditionalNode--------------------
4111 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) {
4112 init_req(ExpectedIn, ex );
4113 }
4114
4115 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
4116 // Either input is TOP ==> the result is TOP
4117 const Type* t = phase->type(in(ExpectedIn));
4118 if (t == Type::TOP) {
4119 return Type::TOP;
4120 }
4121 return LoadStoreNode::Value(phase);
4122 }
4123
4124 //=============================================================================
4125 //-------------------------------adr_type--------------------------------------
4126 const TypePtr* ClearArrayNode::adr_type() const {
4127 Node *adr = in(3);
4128 if (adr == nullptr) return nullptr; // node is dead
4129 return MemNode::calculate_adr_type(adr->bottom_type());
4130 }
4131
4132 //------------------------------match_edge-------------------------------------
4133 // Do we Match on this edge index or not? Do not match memory
4134 uint ClearArrayNode::match_edge(uint idx) const {
4135 return idx > 1;
4136 }
4137
4138 //------------------------------Identity---------------------------------------
4139 // Clearing a zero length array does nothing
4140 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
4141 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
4142 }
4143
4144 //------------------------------Idealize---------------------------------------
4145 // Clearing a short array is faster with stores
4146 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4147 // Already know this is a large node, do not try to ideal it
4148 if (_is_large) return nullptr;
4149
4150 const int unit = BytesPerLong;
4151 const TypeX* t = phase->type(in(2))->isa_intptr_t();
4152 if (!t) return nullptr;
4153 if (!t->is_con()) return nullptr;
4154 intptr_t raw_count = t->get_con();
4155 intptr_t size = raw_count;
4156 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
4157 // Clearing nothing uses the Identity call.
4158 // Negative clears are possible on dead ClearArrays
4159 // (see jck test stmt114.stmt11402.val).
4160 if (size <= 0 || size % unit != 0) return nullptr;
4161 intptr_t count = size / unit;
4162 // Length too long; communicate this to matchers and assemblers.
4163 // Assemblers are responsible to produce fast hardware clears for it.
4164 if (size > InitArrayShortSize) {
4165 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
4166 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
4167 return nullptr;
4168 }
4169 if (!IdealizeClearArrayNode) return nullptr;
4170 Node *mem = in(1);
4171 if( phase->type(mem)==Type::TOP ) return nullptr;
4172 Node *adr = in(3);
4173 const Type* at = phase->type(adr);
4174 if( at==Type::TOP ) return nullptr;
4175 const TypePtr* atp = at->isa_ptr();
4176 // adjust atp to be the correct array element address type
4177 if (atp == nullptr) atp = TypePtr::BOTTOM;
4178 else atp = atp->add_offset(Type::OffsetBot);
4179 // Get base for derived pointer purposes
4180 if( adr->Opcode() != Op_AddP ) Unimplemented();
4181 Node *base = adr->in(1);
4182
4183 Node *zero = phase->makecon(TypeLong::ZERO);
4184 Node *off = phase->MakeConX(BytesPerLong);
4185 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
4186 count--;
4187 while( count-- ) {
4188 mem = phase->transform(mem);
4189 adr = phase->transform(new AddPNode(base,adr,off));
4190 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
4191 }
4192 return mem;
4193 }
4194
4195 //----------------------------step_through----------------------------------
4196 // Return allocation input memory edge if it is different instance
4197 // or itself if it is the one we are looking for.
4198 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
4199 Node* n = *np;
4200 assert(n->is_ClearArray(), "sanity");
4201 intptr_t offset;
4202 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
4203 // This method is called only before Allocate nodes are expanded
4204 // during macro nodes expansion. Before that ClearArray nodes are
4205 // only generated in PhaseMacroExpand::generate_arraycopy() (before
4206 // Allocate nodes are expanded) which follows allocations.
4207 assert(alloc != nullptr, "should have allocation");
4208 if (alloc->_idx == instance_id) {
4209 // Can not bypass initialization of the instance we are looking for.
4210 return false;
4211 }
4212 // Otherwise skip it.
4213 InitializeNode* init = alloc->initialization();
4214 if (init != nullptr)
4215 *np = init->in(TypeFunc::Memory);
4216 else
4217 *np = alloc->in(TypeFunc::Memory);
4218 return true;
4219 }
4220
4221 Node* ClearArrayNode::make_address(Node* dest, Node* offset, bool raw_base, PhaseGVN* phase) {
4222 Node* base = dest;
4223 if (raw_base) {
4224 // May be called as part of the initialization of a just allocated object
4225 base = phase->C->top();
4226 }
4227 return phase->transform(new AddPNode(base, dest, offset));
4228 }
4229
4230 //----------------------------clear_memory-------------------------------------
4231 // Generate code to initialize object storage to zero.
4232 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4233 intptr_t start_offset,
4234 Node* end_offset,
4235 bool raw_base,
4236 PhaseGVN* phase) {
4237 intptr_t offset = start_offset;
4238
4239 int unit = BytesPerLong;
4240 if ((offset % unit) != 0) {
4241 Node* adr = make_address(dest, phase->MakeConX(offset), raw_base, phase);
4242 const TypePtr* atp = TypeRawPtr::BOTTOM;
4243 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4244 mem = phase->transform(mem);
4245 offset += BytesPerInt;
4246 }
4247 assert((offset % unit) == 0, "");
4248
4249 // Initialize the remaining stuff, if any, with a ClearArray.
4250 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, raw_base, phase);
4251 }
4252
4253 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4254 Node* start_offset,
4255 Node* end_offset,
4256 bool raw_base,
4257 PhaseGVN* phase) {
4258 if (start_offset == end_offset) {
4259 // nothing to do
4260 return mem;
4261 }
4262
4263 int unit = BytesPerLong;
4264 Node* zbase = start_offset;
4265 Node* zend = end_offset;
4266
4267 // Scale to the unit required by the CPU:
4268 if (!Matcher::init_array_count_is_in_bytes) {
4269 Node* shift = phase->intcon(exact_log2(unit));
4270 zbase = phase->transform(new URShiftXNode(zbase, shift) );
4271 zend = phase->transform(new URShiftXNode(zend, shift) );
4272 }
4273
4274 // Bulk clear double-words
4275 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
4276 Node* adr = make_address(dest, start_offset, raw_base, phase);
4277 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
4278 return phase->transform(mem);
4279 }
4280
4281 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4282 intptr_t start_offset,
4283 intptr_t end_offset,
4284 bool raw_base,
4285 PhaseGVN* phase) {
4286 if (start_offset == end_offset) {
4287 // nothing to do
4288 return mem;
4289 }
4290
4291 assert((end_offset % BytesPerInt) == 0, "odd end offset");
4292 intptr_t done_offset = end_offset;
4293 if ((done_offset % BytesPerLong) != 0) {
4294 done_offset -= BytesPerInt;
4295 }
4296 if (done_offset > start_offset) {
4297 mem = clear_memory(ctl, mem, dest,
4298 start_offset, phase->MakeConX(done_offset), raw_base, phase);
4299 }
4300 if (done_offset < end_offset) { // emit the final 32-bit store
4301 Node* adr = make_address(dest, phase->MakeConX(done_offset), raw_base, phase);
4302 const TypePtr* atp = TypeRawPtr::BOTTOM;
4303 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4304 mem = phase->transform(mem);
4305 done_offset += BytesPerInt;
4306 }
4307 assert(done_offset == end_offset, "");
4308 return mem;
4309 }
4310
4311 //=============================================================================
4312 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
4313 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
4314 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
4315 #ifdef ASSERT
4316 , _pair_idx(0)
4317 #endif
4318 {
4319 init_class_id(Class_MemBar);
4320 Node* top = C->top();
4321 init_req(TypeFunc::I_O,top);
4322 init_req(TypeFunc::FramePtr,top);
4323 init_req(TypeFunc::ReturnAdr,top);
4324 if (precedent != nullptr)
4325 init_req(TypeFunc::Parms, precedent);
4326 }
4327
4328 //------------------------------cmp--------------------------------------------
4329 uint MemBarNode::hash() const { return NO_HASH; }
4330 bool MemBarNode::cmp( const Node &n ) const {
4331 return (&n == this); // Always fail except on self
4332 }
4333
4334 //------------------------------make-------------------------------------------
4335 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
4336 switch (opcode) {
4337 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
4338 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
4339 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
4340 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
4341 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
4342 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn);
4343 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
4344 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
4345 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
4346 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
4347 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
4348 case Op_Initialize: return new InitializeNode(C, atp, pn);
4349 default: ShouldNotReachHere(); return nullptr;
4350 }
4351 }
4352
4353 void MemBarNode::remove(PhaseIterGVN *igvn) {
4354 assert(outcnt() > 0 && (outcnt() <= 2 || Opcode() == Op_Initialize), "Only one or two out edges allowed");
4355 if (trailing_store() || trailing_load_store()) {
4356 MemBarNode* leading = leading_membar();
4357 if (leading != nullptr) {
4358 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
4359 leading->remove(igvn);
4360 }
4361 }
4362 if (proj_out_or_null(TypeFunc::Control) != nullptr) {
4363 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
4364 }
4365 if (is_Initialize()) {
4366 as_Initialize()->replace_mem_projs_by(in(TypeFunc::Memory), igvn);
4367 } else {
4368 if (proj_out_or_null(TypeFunc::Memory) != nullptr) {
4369 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
4370 }
4371 }
4372 }
4373
4374 //------------------------------Ideal------------------------------------------
4375 // Return a node which is more "ideal" than the current node. Strip out
4376 // control copies
4377 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4378 if (remove_dead_region(phase, can_reshape)) return this;
4379 // Don't bother trying to transform a dead node
4380 if (in(0) && in(0)->is_top()) {
4381 return nullptr;
4382 }
4383
4384 bool progress = false;
4385 // Eliminate volatile MemBars for scalar replaced objects.
4386 if (can_reshape && req() == (Precedent+1)) {
4387 bool eliminate = false;
4388 int opc = Opcode();
4389 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
4390 // Volatile field loads and stores.
4391 Node* my_mem = in(MemBarNode::Precedent);
4392 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
4393 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
4394 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
4395 // replace this Precedent (decodeN) with the Load instead.
4396 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
4397 Node* load_node = my_mem->in(1);
4398 set_req(MemBarNode::Precedent, load_node);
4399 phase->is_IterGVN()->_worklist.push(my_mem);
4400 my_mem = load_node;
4401 } else {
4402 assert(my_mem->unique_out() == this, "sanity");
4403 assert(!trailing_load_store(), "load store node can't be eliminated");
4404 del_req(Precedent);
4405 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
4406 my_mem = nullptr;
4407 }
4408 progress = true;
4409 }
4410 if (my_mem != nullptr && my_mem->is_Mem()) {
4411 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
4412 // Check for scalar replaced object reference.
4413 if( t_oop != nullptr && t_oop->is_known_instance_field() &&
4414 t_oop->offset() != Type::OffsetBot &&
4415 t_oop->offset() != Type::OffsetTop) {
4416 eliminate = true;
4417 }
4418 }
4419 } else if (opc == Op_MemBarRelease || (UseStoreStoreForCtor && opc == Op_MemBarStoreStore)) {
4420 // Final field stores.
4421 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent));
4422 if ((alloc != nullptr) && alloc->is_Allocate() &&
4423 alloc->as_Allocate()->does_not_escape_thread()) {
4424 // The allocated object does not escape.
4425 eliminate = true;
4426 }
4427 }
4428 if (eliminate) {
4429 // Replace MemBar projections by its inputs.
4430 PhaseIterGVN* igvn = phase->is_IterGVN();
4431 remove(igvn);
4432 // Must return either the original node (now dead) or a new node
4433 // (Do not return a top here, since that would break the uniqueness of top.)
4434 return new ConINode(TypeInt::ZERO);
4435 }
4436 }
4437 return progress ? this : nullptr;
4438 }
4439
4440 //------------------------------Value------------------------------------------
4441 const Type* MemBarNode::Value(PhaseGVN* phase) const {
4442 if( !in(0) ) return Type::TOP;
4443 if( phase->type(in(0)) == Type::TOP )
4444 return Type::TOP;
4445 return TypeTuple::MEMBAR;
4446 }
4447
4448 //------------------------------match------------------------------------------
4449 // Construct projections for memory.
4450 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
4451 switch (proj->_con) {
4452 case TypeFunc::Control:
4453 case TypeFunc::Memory:
4454 return new MachProjNode(this, proj->_con, RegMask::EMPTY, MachProjNode::unmatched_proj);
4455 }
4456 ShouldNotReachHere();
4457 return nullptr;
4458 }
4459
4460 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4461 trailing->_kind = TrailingStore;
4462 leading->_kind = LeadingStore;
4463 #ifdef ASSERT
4464 trailing->_pair_idx = leading->_idx;
4465 leading->_pair_idx = leading->_idx;
4466 #endif
4467 }
4468
4469 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4470 trailing->_kind = TrailingLoadStore;
4471 leading->_kind = LeadingLoadStore;
4472 #ifdef ASSERT
4473 trailing->_pair_idx = leading->_idx;
4474 leading->_pair_idx = leading->_idx;
4475 #endif
4476 }
4477
4478 MemBarNode* MemBarNode::trailing_membar() const {
4479 ResourceMark rm;
4480 Node* trailing = (Node*)this;
4481 VectorSet seen;
4482 Node_Stack multis(0);
4483 do {
4484 Node* c = trailing;
4485 uint i = 0;
4486 do {
4487 trailing = nullptr;
4488 for (; i < c->outcnt(); i++) {
4489 Node* next = c->raw_out(i);
4490 if (next != c && next->is_CFG()) {
4491 if (c->is_MultiBranch()) {
4492 if (multis.node() == c) {
4493 multis.set_index(i+1);
4494 } else {
4495 multis.push(c, i+1);
4496 }
4497 }
4498 trailing = next;
4499 break;
4500 }
4501 }
4502 if (trailing != nullptr && !seen.test_set(trailing->_idx)) {
4503 break;
4504 }
4505 while (multis.size() > 0) {
4506 c = multis.node();
4507 i = multis.index();
4508 if (i < c->req()) {
4509 break;
4510 }
4511 multis.pop();
4512 }
4513 } while (multis.size() > 0);
4514 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
4515
4516 MemBarNode* mb = trailing->as_MemBar();
4517 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
4518 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
4519 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
4520 return mb;
4521 }
4522
4523 MemBarNode* MemBarNode::leading_membar() const {
4524 ResourceMark rm;
4525 VectorSet seen;
4526 Node_Stack regions(0);
4527 Node* leading = in(0);
4528 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
4529 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) {
4530 leading = nullptr;
4531 while (regions.size() > 0 && leading == nullptr) {
4532 Node* r = regions.node();
4533 uint i = regions.index();
4534 if (i < r->req()) {
4535 leading = r->in(i);
4536 regions.set_index(i+1);
4537 } else {
4538 regions.pop();
4539 }
4540 }
4541 if (leading == nullptr) {
4542 assert(regions.size() == 0, "all paths should have been tried");
4543 return nullptr;
4544 }
4545 }
4546 if (leading->is_Region()) {
4547 regions.push(leading, 2);
4548 leading = leading->in(1);
4549 } else {
4550 leading = leading->in(0);
4551 }
4552 }
4553 #ifdef ASSERT
4554 Unique_Node_List wq;
4555 wq.push((Node*)this);
4556 uint found = 0;
4557 for (uint i = 0; i < wq.size(); i++) {
4558 Node* n = wq.at(i);
4559 if (n->is_Region()) {
4560 for (uint j = 1; j < n->req(); j++) {
4561 Node* in = n->in(j);
4562 if (in != nullptr && !in->is_top()) {
4563 wq.push(in);
4564 }
4565 }
4566 } else {
4567 if (n->is_MemBar() && n->as_MemBar()->leading()) {
4568 assert(n == leading, "consistency check failed");
4569 found++;
4570 } else {
4571 Node* in = n->in(0);
4572 if (in != nullptr && !in->is_top()) {
4573 wq.push(in);
4574 }
4575 }
4576 }
4577 }
4578 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed");
4579 #endif
4580 if (leading == nullptr) {
4581 return nullptr;
4582 }
4583 MemBarNode* mb = leading->as_MemBar();
4584 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
4585 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
4586 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
4587 return mb;
4588 }
4589
4590
4591 //===========================InitializeNode====================================
4592 // SUMMARY:
4593 // This node acts as a memory barrier on raw memory, after some raw stores.
4594 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
4595 // The Initialize can 'capture' suitably constrained stores as raw inits.
4596 // It can coalesce related raw stores into larger units (called 'tiles').
4597 // It can avoid zeroing new storage for memory units which have raw inits.
4598 // At macro-expansion, it is marked 'complete', and does not optimize further.
4599 //
4600 // EXAMPLE:
4601 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
4602 // ctl = incoming control; mem* = incoming memory
4603 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
4604 // First allocate uninitialized memory and fill in the header:
4605 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
4606 // ctl := alloc.Control; mem* := alloc.Memory*
4607 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
4608 // Then initialize to zero the non-header parts of the raw memory block:
4609 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
4610 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
4611 // After the initialize node executes, the object is ready for service:
4612 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
4613 // Suppose its body is immediately initialized as {1,2}:
4614 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4615 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4616 // mem.SLICE(#short[*]) := store2
4617 //
4618 // DETAILS:
4619 // An InitializeNode collects and isolates object initialization after
4620 // an AllocateNode and before the next possible safepoint. As a
4621 // memory barrier (MemBarNode), it keeps critical stores from drifting
4622 // down past any safepoint or any publication of the allocation.
4623 // Before this barrier, a newly-allocated object may have uninitialized bits.
4624 // After this barrier, it may be treated as a real oop, and GC is allowed.
4625 //
4626 // The semantics of the InitializeNode include an implicit zeroing of
4627 // the new object from object header to the end of the object.
4628 // (The object header and end are determined by the AllocateNode.)
4629 //
4630 // Certain stores may be added as direct inputs to the InitializeNode.
4631 // These stores must update raw memory, and they must be to addresses
4632 // derived from the raw address produced by AllocateNode, and with
4633 // a constant offset. They must be ordered by increasing offset.
4634 // The first one is at in(RawStores), the last at in(req()-1).
4635 // Unlike most memory operations, they are not linked in a chain,
4636 // but are displayed in parallel as users of the rawmem output of
4637 // the allocation.
4638 //
4639 // (See comments in InitializeNode::capture_store, which continue
4640 // the example given above.)
4641 //
4642 // When the associated Allocate is macro-expanded, the InitializeNode
4643 // may be rewritten to optimize collected stores. A ClearArrayNode
4644 // may also be created at that point to represent any required zeroing.
4645 // The InitializeNode is then marked 'complete', prohibiting further
4646 // capturing of nearby memory operations.
4647 //
4648 // During macro-expansion, all captured initializations which store
4649 // constant values of 32 bits or smaller are coalesced (if advantageous)
4650 // into larger 'tiles' 32 or 64 bits. This allows an object to be
4651 // initialized in fewer memory operations. Memory words which are
4652 // covered by neither tiles nor non-constant stores are pre-zeroed
4653 // by explicit stores of zero. (The code shape happens to do all
4654 // zeroing first, then all other stores, with both sequences occurring
4655 // in order of ascending offsets.)
4656 //
4657 // Alternatively, code may be inserted between an AllocateNode and its
4658 // InitializeNode, to perform arbitrary initialization of the new object.
4659 // E.g., the object copying intrinsics insert complex data transfers here.
4660 // The initialization must then be marked as 'complete' disable the
4661 // built-in zeroing semantics and the collection of initializing stores.
4662 //
4663 // While an InitializeNode is incomplete, reads from the memory state
4664 // produced by it are optimizable if they match the control edge and
4665 // new oop address associated with the allocation/initialization.
4666 // They return a stored value (if the offset matches) or else zero.
4667 // A write to the memory state, if it matches control and address,
4668 // and if it is to a constant offset, may be 'captured' by the
4669 // InitializeNode. It is cloned as a raw memory operation and rewired
4670 // inside the initialization, to the raw oop produced by the allocation.
4671 // Operations on addresses which are provably distinct (e.g., to
4672 // other AllocateNodes) are allowed to bypass the initialization.
4673 //
4674 // The effect of all this is to consolidate object initialization
4675 // (both arrays and non-arrays, both piecewise and bulk) into a
4676 // single location, where it can be optimized as a unit.
4677 //
4678 // Only stores with an offset less than TrackedInitializationLimit words
4679 // will be considered for capture by an InitializeNode. This puts a
4680 // reasonable limit on the complexity of optimized initializations.
4681
4682 //---------------------------InitializeNode------------------------------------
4683 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
4684 : MemBarNode(C, adr_type, rawoop),
4685 _is_complete(Incomplete), _does_not_escape(false)
4686 {
4687 init_class_id(Class_Initialize);
4688
4689 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
4690 assert(in(RawAddress) == rawoop, "proper init");
4691 // Note: allocation() can be null, for secondary initialization barriers
4692 }
4693
4694 // Since this node is not matched, it will be processed by the
4695 // register allocator. Declare that there are no constraints
4696 // on the allocation of the RawAddress edge.
4697 const RegMask &InitializeNode::in_RegMask(uint idx) const {
4698 // This edge should be set to top, by the set_complete. But be conservative.
4699 if (idx == InitializeNode::RawAddress)
4700 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
4701 return RegMask::EMPTY;
4702 }
4703
4704 Node* InitializeNode::memory(uint alias_idx) {
4705 Node* mem = in(Memory);
4706 if (mem->is_MergeMem()) {
4707 return mem->as_MergeMem()->memory_at(alias_idx);
4708 } else {
4709 // incoming raw memory is not split
4710 return mem;
4711 }
4712 }
4713
4714 bool InitializeNode::is_non_zero() {
4715 if (is_complete()) return false;
4716 remove_extra_zeroes();
4717 return (req() > RawStores);
4718 }
4719
4720 void InitializeNode::set_complete(PhaseGVN* phase) {
4721 assert(!is_complete(), "caller responsibility");
4722 _is_complete = Complete;
4723
4724 // After this node is complete, it contains a bunch of
4725 // raw-memory initializations. There is no need for
4726 // it to have anything to do with non-raw memory effects.
4727 // Therefore, tell all non-raw users to re-optimize themselves,
4728 // after skipping the memory effects of this initialization.
4729 PhaseIterGVN* igvn = phase->is_IterGVN();
4730 if (igvn) igvn->add_users_to_worklist(this);
4731 }
4732
4733 // convenience function
4734 // return false if the init contains any stores already
4735 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
4736 InitializeNode* init = initialization();
4737 if (init == nullptr || init->is_complete()) return false;
4738 init->remove_extra_zeroes();
4739 // for now, if this allocation has already collected any inits, bail:
4740 if (init->is_non_zero()) return false;
4741 init->set_complete(phase);
4742 return true;
4743 }
4744
4745 void InitializeNode::remove_extra_zeroes() {
4746 if (req() == RawStores) return;
4747 Node* zmem = zero_memory();
4748 uint fill = RawStores;
4749 for (uint i = fill; i < req(); i++) {
4750 Node* n = in(i);
4751 if (n->is_top() || n == zmem) continue; // skip
4752 if (fill < i) set_req(fill, n); // compact
4753 ++fill;
4754 }
4755 // delete any empty spaces created:
4756 while (fill < req()) {
4757 del_req(fill);
4758 }
4759 }
4760
4761 // Helper for remembering which stores go with which offsets.
4762 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) {
4763 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
4764 intptr_t offset = -1;
4765 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
4766 phase, offset);
4767 if (base == nullptr) return -1; // something is dead,
4768 if (offset < 0) return -1; // dead, dead
4769 return offset;
4770 }
4771
4772 // Helper for proving that an initialization expression is
4773 // "simple enough" to be folded into an object initialization.
4774 // Attempts to prove that a store's initial value 'n' can be captured
4775 // within the initialization without creating a vicious cycle, such as:
4776 // { Foo p = new Foo(); p.next = p; }
4777 // True for constants and parameters and small combinations thereof.
4778 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
4779 ResourceMark rm;
4780 Unique_Node_List worklist;
4781 worklist.push(value);
4782
4783 uint complexity_limit = 20;
4784 for (uint j = 0; j < worklist.size(); j++) {
4785 if (j >= complexity_limit) {
4786 return false; // Bail out if processed too many nodes
4787 }
4788
4789 Node* n = worklist.at(j);
4790 if (n == nullptr) continue; // (can this really happen?)
4791 if (n->is_Proj()) n = n->in(0);
4792 if (n == this) return false; // found a cycle
4793 if (n->is_Con()) continue;
4794 if (n->is_Start()) continue; // params, etc., are OK
4795 if (n->is_Root()) continue; // even better
4796
4797 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
4798 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
4799 continue;
4800 }
4801
4802 Node* ctl = n->in(0);
4803 if (ctl != nullptr && !ctl->is_top()) {
4804 if (ctl->is_Proj()) ctl = ctl->in(0);
4805 if (ctl == this) return false;
4806
4807 // If we already know that the enclosing memory op is pinned right after
4808 // the init, then any control flow that the store has picked up
4809 // must have preceded the init, or else be equal to the init.
4810 // Even after loop optimizations (which might change control edges)
4811 // a store is never pinned *before* the availability of its inputs.
4812 if (!MemNode::all_controls_dominate(n, this)) {
4813 return false; // failed to prove a good control
4814 }
4815 }
4816
4817 // Check data edges for possible dependencies on 'this'.
4818 for (uint i = 1; i < n->req(); i++) {
4819 Node* m = n->in(i);
4820 if (m == nullptr || m == n || m->is_top()) continue;
4821
4822 // Only process data inputs once
4823 worklist.push(m);
4824 }
4825 }
4826
4827 return true;
4828 }
4829
4830 // Here are all the checks a Store must pass before it can be moved into
4831 // an initialization. Returns zero if a check fails.
4832 // On success, returns the (constant) offset to which the store applies,
4833 // within the initialized memory.
4834 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
4835 const int FAIL = 0;
4836 if (st->req() != MemNode::ValueIn + 1)
4837 return FAIL; // an inscrutable StoreNode (card mark?)
4838 Node* ctl = st->in(MemNode::Control);
4839 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this))
4840 return FAIL; // must be unconditional after the initialization
4841 Node* mem = st->in(MemNode::Memory);
4842 if (!(mem->is_Proj() && mem->in(0) == this))
4843 return FAIL; // must not be preceded by other stores
4844 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4845 if ((st->Opcode() == Op_StoreP || st->Opcode() == Op_StoreN) &&
4846 !bs->can_initialize_object(st)) {
4847 return FAIL;
4848 }
4849 Node* adr = st->in(MemNode::Address);
4850 intptr_t offset;
4851 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
4852 if (alloc == nullptr)
4853 return FAIL; // inscrutable address
4854 if (alloc != allocation())
4855 return FAIL; // wrong allocation! (store needs to float up)
4856 int size_in_bytes = st->memory_size();
4857 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
4858 return FAIL; // mismatched access
4859 }
4860 Node* val = st->in(MemNode::ValueIn);
4861
4862 if (!detect_init_independence(val, phase))
4863 return FAIL; // stored value must be 'simple enough'
4864
4865 // The Store can be captured only if nothing after the allocation
4866 // and before the Store is using the memory location that the store
4867 // overwrites.
4868 bool failed = false;
4869 // If is_complete_with_arraycopy() is true the shape of the graph is
4870 // well defined and is safe so no need for extra checks.
4871 if (!is_complete_with_arraycopy()) {
4872 // We are going to look at each use of the memory state following
4873 // the allocation to make sure nothing reads the memory that the
4874 // Store writes.
4875 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
4876 int alias_idx = phase->C->get_alias_index(t_adr);
4877 ResourceMark rm;
4878 Unique_Node_List mems;
4879 mems.push(mem);
4880 Node* unique_merge = nullptr;
4881 for (uint next = 0; next < mems.size(); ++next) {
4882 Node *m = mems.at(next);
4883 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
4884 Node *n = m->fast_out(j);
4885 if (n->outcnt() == 0) {
4886 continue;
4887 }
4888 if (n == st) {
4889 continue;
4890 } else if (n->in(0) != nullptr && n->in(0) != ctl) {
4891 // If the control of this use is different from the control
4892 // of the Store which is right after the InitializeNode then
4893 // this node cannot be between the InitializeNode and the
4894 // Store.
4895 continue;
4896 } else if (n->is_MergeMem()) {
4897 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
4898 // We can hit a MergeMemNode (that will likely go away
4899 // later) that is a direct use of the memory state
4900 // following the InitializeNode on the same slice as the
4901 // store node that we'd like to capture. We need to check
4902 // the uses of the MergeMemNode.
4903 mems.push(n);
4904 }
4905 } else if (n->is_Mem()) {
4906 Node* other_adr = n->in(MemNode::Address);
4907 if (other_adr == adr) {
4908 failed = true;
4909 break;
4910 } else {
4911 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
4912 if (other_t_adr != nullptr) {
4913 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
4914 if (other_alias_idx == alias_idx) {
4915 // A load from the same memory slice as the store right
4916 // after the InitializeNode. We check the control of the
4917 // object/array that is loaded from. If it's the same as
4918 // the store control then we cannot capture the store.
4919 assert(!n->is_Store(), "2 stores to same slice on same control?");
4920 Node* base = other_adr;
4921 assert(base->is_AddP(), "should be addp but is %s", base->Name());
4922 base = base->in(AddPNode::Base);
4923 if (base != nullptr) {
4924 base = base->uncast();
4925 if (base->is_Proj() && base->in(0) == alloc) {
4926 failed = true;
4927 break;
4928 }
4929 }
4930 }
4931 }
4932 }
4933 } else {
4934 failed = true;
4935 break;
4936 }
4937 }
4938 }
4939 }
4940 if (failed) {
4941 if (!can_reshape) {
4942 // We decided we couldn't capture the store during parsing. We
4943 // should try again during the next IGVN once the graph is
4944 // cleaner.
4945 phase->C->record_for_igvn(st);
4946 }
4947 return FAIL;
4948 }
4949
4950 return offset; // success
4951 }
4952
4953 // Find the captured store in(i) which corresponds to the range
4954 // [start..start+size) in the initialized object.
4955 // If there is one, return its index i. If there isn't, return the
4956 // negative of the index where it should be inserted.
4957 // Return 0 if the queried range overlaps an initialization boundary
4958 // or if dead code is encountered.
4959 // If size_in_bytes is zero, do not bother with overlap checks.
4960 int InitializeNode::captured_store_insertion_point(intptr_t start,
4961 int size_in_bytes,
4962 PhaseValues* phase) {
4963 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
4964
4965 if (is_complete())
4966 return FAIL; // arraycopy got here first; punt
4967
4968 assert(allocation() != nullptr, "must be present");
4969
4970 // no negatives, no header fields:
4971 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
4972
4973 // after a certain size, we bail out on tracking all the stores:
4974 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4975 if (start >= ti_limit) return FAIL;
4976
4977 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
4978 if (i >= limit) return -(int)i; // not found; here is where to put it
4979
4980 Node* st = in(i);
4981 intptr_t st_off = get_store_offset(st, phase);
4982 if (st_off < 0) {
4983 if (st != zero_memory()) {
4984 return FAIL; // bail out if there is dead garbage
4985 }
4986 } else if (st_off > start) {
4987 // ...we are done, since stores are ordered
4988 if (st_off < start + size_in_bytes) {
4989 return FAIL; // the next store overlaps
4990 }
4991 return -(int)i; // not found; here is where to put it
4992 } else if (st_off < start) {
4993 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
4994 if (size_in_bytes != 0 &&
4995 start < st_off + MAX_STORE &&
4996 start < st_off + st->as_Store()->memory_size()) {
4997 return FAIL; // the previous store overlaps
4998 }
4999 } else {
5000 if (size_in_bytes != 0 &&
5001 st->as_Store()->memory_size() != size_in_bytes) {
5002 return FAIL; // mismatched store size
5003 }
5004 return i;
5005 }
5006
5007 ++i;
5008 }
5009 }
5010
5011 // Look for a captured store which initializes at the offset 'start'
5012 // with the given size. If there is no such store, and no other
5013 // initialization interferes, then return zero_memory (the memory
5014 // projection of the AllocateNode).
5015 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
5016 PhaseValues* phase) {
5017 assert(stores_are_sane(phase), "");
5018 int i = captured_store_insertion_point(start, size_in_bytes, phase);
5019 if (i == 0) {
5020 return nullptr; // something is dead
5021 } else if (i < 0) {
5022 return zero_memory(); // just primordial zero bits here
5023 } else {
5024 Node* st = in(i); // here is the store at this position
5025 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
5026 return st;
5027 }
5028 }
5029
5030 // Create, as a raw pointer, an address within my new object at 'offset'.
5031 Node* InitializeNode::make_raw_address(intptr_t offset,
5032 PhaseGVN* phase) {
5033 Node* addr = in(RawAddress);
5034 if (offset != 0) {
5035 Compile* C = phase->C;
5036 addr = phase->transform( new AddPNode(C->top(), addr,
5037 phase->MakeConX(offset)) );
5038 }
5039 return addr;
5040 }
5041
5042 // Clone the given store, converting it into a raw store
5043 // initializing a field or element of my new object.
5044 // Caller is responsible for retiring the original store,
5045 // with subsume_node or the like.
5046 //
5047 // From the example above InitializeNode::InitializeNode,
5048 // here are the old stores to be captured:
5049 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
5050 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
5051 //
5052 // Here is the changed code; note the extra edges on init:
5053 // alloc = (Allocate ...)
5054 // rawoop = alloc.RawAddress
5055 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
5056 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
5057 // init = (Initialize alloc.Control alloc.Memory rawoop
5058 // rawstore1 rawstore2)
5059 //
5060 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
5061 PhaseGVN* phase, bool can_reshape) {
5062 assert(stores_are_sane(phase), "");
5063
5064 if (start < 0) return nullptr;
5065 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
5066
5067 Compile* C = phase->C;
5068 int size_in_bytes = st->memory_size();
5069 int i = captured_store_insertion_point(start, size_in_bytes, phase);
5070 if (i == 0) return nullptr; // bail out
5071 Node* prev_mem = nullptr; // raw memory for the captured store
5072 if (i > 0) {
5073 prev_mem = in(i); // there is a pre-existing store under this one
5074 set_req(i, C->top()); // temporarily disconnect it
5075 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
5076 } else {
5077 i = -i; // no pre-existing store
5078 prev_mem = zero_memory(); // a slice of the newly allocated object
5079 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
5080 set_req(--i, C->top()); // reuse this edge; it has been folded away
5081 else
5082 ins_req(i, C->top()); // build a new edge
5083 }
5084 Node* new_st = st->clone();
5085 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
5086 new_st->set_req(MemNode::Control, in(Control));
5087 new_st->set_req(MemNode::Memory, prev_mem);
5088 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
5089 bs->eliminate_gc_barrier_data(new_st);
5090 new_st = phase->transform(new_st);
5091
5092 // At this point, new_st might have swallowed a pre-existing store
5093 // at the same offset, or perhaps new_st might have disappeared,
5094 // if it redundantly stored the same value (or zero to fresh memory).
5095
5096 // In any case, wire it in:
5097 PhaseIterGVN* igvn = phase->is_IterGVN();
5098 if (igvn) {
5099 igvn->rehash_node_delayed(this);
5100 }
5101 set_req(i, new_st);
5102
5103 // The caller may now kill the old guy.
5104 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
5105 assert(check_st == new_st || check_st == nullptr, "must be findable");
5106 assert(!is_complete(), "");
5107 return new_st;
5108 }
5109
5110 static bool store_constant(jlong* tiles, int num_tiles,
5111 intptr_t st_off, int st_size,
5112 jlong con) {
5113 if ((st_off & (st_size-1)) != 0)
5114 return false; // strange store offset (assume size==2**N)
5115 address addr = (address)tiles + st_off;
5116 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
5117 switch (st_size) {
5118 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
5119 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
5120 case sizeof(jint): *(jint*) addr = (jint) con; break;
5121 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
5122 default: return false; // strange store size (detect size!=2**N here)
5123 }
5124 return true; // return success to caller
5125 }
5126
5127 // Coalesce subword constants into int constants and possibly
5128 // into long constants. The goal, if the CPU permits,
5129 // is to initialize the object with a small number of 64-bit tiles.
5130 // Also, convert floating-point constants to bit patterns.
5131 // Non-constants are not relevant to this pass.
5132 //
5133 // In terms of the running example on InitializeNode::InitializeNode
5134 // and InitializeNode::capture_store, here is the transformation
5135 // of rawstore1 and rawstore2 into rawstore12:
5136 // alloc = (Allocate ...)
5137 // rawoop = alloc.RawAddress
5138 // tile12 = 0x00010002
5139 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
5140 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
5141 //
5142 void
5143 InitializeNode::coalesce_subword_stores(intptr_t header_size,
5144 Node* size_in_bytes,
5145 PhaseGVN* phase) {
5146 Compile* C = phase->C;
5147
5148 assert(stores_are_sane(phase), "");
5149 // Note: After this pass, they are not completely sane,
5150 // since there may be some overlaps.
5151
5152 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
5153
5154 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
5155 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
5156 size_limit = MIN2(size_limit, ti_limit);
5157 size_limit = align_up(size_limit, BytesPerLong);
5158 int num_tiles = size_limit / BytesPerLong;
5159
5160 // allocate space for the tile map:
5161 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
5162 jlong tiles_buf[small_len];
5163 Node* nodes_buf[small_len];
5164 jlong inits_buf[small_len];
5165 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
5166 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5167 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
5168 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
5169 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
5170 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5171 // tiles: exact bitwise model of all primitive constants
5172 // nodes: last constant-storing node subsumed into the tiles model
5173 // inits: which bytes (in each tile) are touched by any initializations
5174
5175 //// Pass A: Fill in the tile model with any relevant stores.
5176
5177 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
5178 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
5179 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
5180 Node* zmem = zero_memory(); // initially zero memory state
5181 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5182 Node* st = in(i);
5183 intptr_t st_off = get_store_offset(st, phase);
5184
5185 // Figure out the store's offset and constant value:
5186 if (st_off < header_size) continue; //skip (ignore header)
5187 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
5188 int st_size = st->as_Store()->memory_size();
5189 if (st_off + st_size > size_limit) break;
5190
5191 // Record which bytes are touched, whether by constant or not.
5192 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
5193 continue; // skip (strange store size)
5194
5195 const Type* val = phase->type(st->in(MemNode::ValueIn));
5196 if (!val->singleton()) continue; //skip (non-con store)
5197 BasicType type = val->basic_type();
5198
5199 jlong con = 0;
5200 switch (type) {
5201 case T_INT: con = val->is_int()->get_con(); break;
5202 case T_LONG: con = val->is_long()->get_con(); break;
5203 case T_FLOAT: con = jint_cast(val->getf()); break;
5204 case T_DOUBLE: con = jlong_cast(val->getd()); break;
5205 default: continue; //skip (odd store type)
5206 }
5207
5208 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
5209 st->Opcode() == Op_StoreL) {
5210 continue; // This StoreL is already optimal.
5211 }
5212
5213 // Store down the constant.
5214 store_constant(tiles, num_tiles, st_off, st_size, con);
5215
5216 intptr_t j = st_off >> LogBytesPerLong;
5217
5218 if (type == T_INT && st_size == BytesPerInt
5219 && (st_off & BytesPerInt) == BytesPerInt) {
5220 jlong lcon = tiles[j];
5221 if (!Matcher::isSimpleConstant64(lcon) &&
5222 st->Opcode() == Op_StoreI) {
5223 // This StoreI is already optimal by itself.
5224 jint* intcon = (jint*) &tiles[j];
5225 intcon[1] = 0; // undo the store_constant()
5226
5227 // If the previous store is also optimal by itself, back up and
5228 // undo the action of the previous loop iteration... if we can.
5229 // But if we can't, just let the previous half take care of itself.
5230 st = nodes[j];
5231 st_off -= BytesPerInt;
5232 con = intcon[0];
5233 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) {
5234 assert(st_off >= header_size, "still ignoring header");
5235 assert(get_store_offset(st, phase) == st_off, "must be");
5236 assert(in(i-1) == zmem, "must be");
5237 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
5238 assert(con == tcon->is_int()->get_con(), "must be");
5239 // Undo the effects of the previous loop trip, which swallowed st:
5240 intcon[0] = 0; // undo store_constant()
5241 set_req(i-1, st); // undo set_req(i, zmem)
5242 nodes[j] = nullptr; // undo nodes[j] = st
5243 --old_subword; // undo ++old_subword
5244 }
5245 continue; // This StoreI is already optimal.
5246 }
5247 }
5248
5249 // This store is not needed.
5250 set_req(i, zmem);
5251 nodes[j] = st; // record for the moment
5252 if (st_size < BytesPerLong) // something has changed
5253 ++old_subword; // includes int/float, but who's counting...
5254 else ++old_long;
5255 }
5256
5257 if ((old_subword + old_long) == 0)
5258 return; // nothing more to do
5259
5260 //// Pass B: Convert any non-zero tiles into optimal constant stores.
5261 // Be sure to insert them before overlapping non-constant stores.
5262 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
5263 for (int j = 0; j < num_tiles; j++) {
5264 jlong con = tiles[j];
5265 jlong init = inits[j];
5266 if (con == 0) continue;
5267 jint con0, con1; // split the constant, address-wise
5268 jint init0, init1; // split the init map, address-wise
5269 { union { jlong con; jint intcon[2]; } u;
5270 u.con = con;
5271 con0 = u.intcon[0];
5272 con1 = u.intcon[1];
5273 u.con = init;
5274 init0 = u.intcon[0];
5275 init1 = u.intcon[1];
5276 }
5277
5278 Node* old = nodes[j];
5279 assert(old != nullptr, "need the prior store");
5280 intptr_t offset = (j * BytesPerLong);
5281
5282 bool split = !Matcher::isSimpleConstant64(con);
5283
5284 if (offset < header_size) {
5285 assert(offset + BytesPerInt >= header_size, "second int counts");
5286 assert(*(jint*)&tiles[j] == 0, "junk in header");
5287 split = true; // only the second word counts
5288 // Example: int a[] = { 42 ... }
5289 } else if (con0 == 0 && init0 == -1) {
5290 split = true; // first word is covered by full inits
5291 // Example: int a[] = { ... foo(), 42 ... }
5292 } else if (con1 == 0 && init1 == -1) {
5293 split = true; // second word is covered by full inits
5294 // Example: int a[] = { ... 42, foo() ... }
5295 }
5296
5297 // Here's a case where init0 is neither 0 nor -1:
5298 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
5299 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
5300 // In this case the tile is not split; it is (jlong)42.
5301 // The big tile is stored down, and then the foo() value is inserted.
5302 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
5303
5304 Node* ctl = old->in(MemNode::Control);
5305 Node* adr = make_raw_address(offset, phase);
5306 const TypePtr* atp = TypeRawPtr::BOTTOM;
5307
5308 // One or two coalesced stores to plop down.
5309 Node* st[2];
5310 intptr_t off[2];
5311 int nst = 0;
5312 if (!split) {
5313 ++new_long;
5314 off[nst] = offset;
5315 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5316 phase->longcon(con), T_LONG, MemNode::unordered);
5317 } else {
5318 // Omit either if it is a zero.
5319 if (con0 != 0) {
5320 ++new_int;
5321 off[nst] = offset;
5322 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5323 phase->intcon(con0), T_INT, MemNode::unordered);
5324 }
5325 if (con1 != 0) {
5326 ++new_int;
5327 offset += BytesPerInt;
5328 adr = make_raw_address(offset, phase);
5329 off[nst] = offset;
5330 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5331 phase->intcon(con1), T_INT, MemNode::unordered);
5332 }
5333 }
5334
5335 // Insert second store first, then the first before the second.
5336 // Insert each one just before any overlapping non-constant stores.
5337 while (nst > 0) {
5338 Node* st1 = st[--nst];
5339 C->copy_node_notes_to(st1, old);
5340 st1 = phase->transform(st1);
5341 offset = off[nst];
5342 assert(offset >= header_size, "do not smash header");
5343 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
5344 guarantee(ins_idx != 0, "must re-insert constant store");
5345 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
5346 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
5347 set_req(--ins_idx, st1);
5348 else
5349 ins_req(ins_idx, st1);
5350 }
5351 }
5352
5353 if (PrintCompilation && WizardMode)
5354 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
5355 old_subword, old_long, new_int, new_long);
5356 if (C->log() != nullptr)
5357 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
5358 old_subword, old_long, new_int, new_long);
5359
5360 // Clean up any remaining occurrences of zmem:
5361 remove_extra_zeroes();
5362 }
5363
5364 // Explore forward from in(start) to find the first fully initialized
5365 // word, and return its offset. Skip groups of subword stores which
5366 // together initialize full words. If in(start) is itself part of a
5367 // fully initialized word, return the offset of in(start). If there
5368 // are no following full-word stores, or if something is fishy, return
5369 // a negative value.
5370 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
5371 int int_map = 0;
5372 intptr_t int_map_off = 0;
5373 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
5374
5375 for (uint i = start, limit = req(); i < limit; i++) {
5376 Node* st = in(i);
5377
5378 intptr_t st_off = get_store_offset(st, phase);
5379 if (st_off < 0) break; // return conservative answer
5380
5381 int st_size = st->as_Store()->memory_size();
5382 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
5383 return st_off; // we found a complete word init
5384 }
5385
5386 // update the map:
5387
5388 intptr_t this_int_off = align_down(st_off, BytesPerInt);
5389 if (this_int_off != int_map_off) {
5390 // reset the map:
5391 int_map = 0;
5392 int_map_off = this_int_off;
5393 }
5394
5395 int subword_off = st_off - this_int_off;
5396 int_map |= right_n_bits(st_size) << subword_off;
5397 if ((int_map & FULL_MAP) == FULL_MAP) {
5398 return this_int_off; // we found a complete word init
5399 }
5400
5401 // Did this store hit or cross the word boundary?
5402 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
5403 if (next_int_off == this_int_off + BytesPerInt) {
5404 // We passed the current int, without fully initializing it.
5405 int_map_off = next_int_off;
5406 int_map >>= BytesPerInt;
5407 } else if (next_int_off > this_int_off + BytesPerInt) {
5408 // We passed the current and next int.
5409 return this_int_off + BytesPerInt;
5410 }
5411 }
5412
5413 return -1;
5414 }
5415
5416
5417 // Called when the associated AllocateNode is expanded into CFG.
5418 // At this point, we may perform additional optimizations.
5419 // Linearize the stores by ascending offset, to make memory
5420 // activity as coherent as possible.
5421 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
5422 intptr_t header_size,
5423 Node* size_in_bytes,
5424 PhaseIterGVN* phase) {
5425 assert(!is_complete(), "not already complete");
5426 assert(stores_are_sane(phase), "");
5427 assert(allocation() != nullptr, "must be present");
5428
5429 remove_extra_zeroes();
5430
5431 if (ReduceFieldZeroing || ReduceBulkZeroing)
5432 // reduce instruction count for common initialization patterns
5433 coalesce_subword_stores(header_size, size_in_bytes, phase);
5434
5435 Node* zmem = zero_memory(); // initially zero memory state
5436 Node* inits = zmem; // accumulating a linearized chain of inits
5437 #ifdef ASSERT
5438 intptr_t first_offset = allocation()->minimum_header_size();
5439 intptr_t last_init_off = first_offset; // previous init offset
5440 intptr_t last_init_end = first_offset; // previous init offset+size
5441 intptr_t last_tile_end = first_offset; // previous tile offset+size
5442 #endif
5443 intptr_t zeroes_done = header_size;
5444
5445 bool do_zeroing = true; // we might give up if inits are very sparse
5446 int big_init_gaps = 0; // how many large gaps have we seen?
5447
5448 if (UseTLAB && ZeroTLAB) do_zeroing = false;
5449 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
5450
5451 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5452 Node* st = in(i);
5453 intptr_t st_off = get_store_offset(st, phase);
5454 if (st_off < 0)
5455 break; // unknown junk in the inits
5456 if (st->in(MemNode::Memory) != zmem)
5457 break; // complicated store chains somehow in list
5458
5459 int st_size = st->as_Store()->memory_size();
5460 intptr_t next_init_off = st_off + st_size;
5461
5462 if (do_zeroing && zeroes_done < next_init_off) {
5463 // See if this store needs a zero before it or under it.
5464 intptr_t zeroes_needed = st_off;
5465
5466 if (st_size < BytesPerInt) {
5467 // Look for subword stores which only partially initialize words.
5468 // If we find some, we must lay down some word-level zeroes first,
5469 // underneath the subword stores.
5470 //
5471 // Examples:
5472 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
5473 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
5474 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
5475 //
5476 // Note: coalesce_subword_stores may have already done this,
5477 // if it was prompted by constant non-zero subword initializers.
5478 // But this case can still arise with non-constant stores.
5479
5480 intptr_t next_full_store = find_next_fullword_store(i, phase);
5481
5482 // In the examples above:
5483 // in(i) p q r s x y z
5484 // st_off 12 13 14 15 12 13 14
5485 // st_size 1 1 1 1 1 1 1
5486 // next_full_s. 12 16 16 16 16 16 16
5487 // z's_done 12 16 16 16 12 16 12
5488 // z's_needed 12 16 16 16 16 16 16
5489 // zsize 0 0 0 0 4 0 4
5490 if (next_full_store < 0) {
5491 // Conservative tack: Zero to end of current word.
5492 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
5493 } else {
5494 // Zero to beginning of next fully initialized word.
5495 // Or, don't zero at all, if we are already in that word.
5496 assert(next_full_store >= zeroes_needed, "must go forward");
5497 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
5498 zeroes_needed = next_full_store;
5499 }
5500 }
5501
5502 if (zeroes_needed > zeroes_done) {
5503 intptr_t zsize = zeroes_needed - zeroes_done;
5504 // Do some incremental zeroing on rawmem, in parallel with inits.
5505 zeroes_done = align_down(zeroes_done, BytesPerInt);
5506 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5507 zeroes_done, zeroes_needed,
5508 true,
5509 phase);
5510 zeroes_done = zeroes_needed;
5511 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
5512 do_zeroing = false; // leave the hole, next time
5513 }
5514 }
5515
5516 // Collect the store and move on:
5517 phase->replace_input_of(st, MemNode::Memory, inits);
5518 inits = st; // put it on the linearized chain
5519 set_req(i, zmem); // unhook from previous position
5520
5521 if (zeroes_done == st_off)
5522 zeroes_done = next_init_off;
5523
5524 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
5525
5526 #ifdef ASSERT
5527 // Various order invariants. Weaker than stores_are_sane because
5528 // a large constant tile can be filled in by smaller non-constant stores.
5529 assert(st_off >= last_init_off, "inits do not reverse");
5530 last_init_off = st_off;
5531 const Type* val = nullptr;
5532 if (st_size >= BytesPerInt &&
5533 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
5534 (int)val->basic_type() < (int)T_OBJECT) {
5535 assert(st_off >= last_tile_end, "tiles do not overlap");
5536 assert(st_off >= last_init_end, "tiles do not overwrite inits");
5537 last_tile_end = MAX2(last_tile_end, next_init_off);
5538 } else {
5539 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
5540 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
5541 assert(st_off >= last_init_end, "inits do not overlap");
5542 last_init_end = next_init_off; // it's a non-tile
5543 }
5544 #endif //ASSERT
5545 }
5546
5547 remove_extra_zeroes(); // clear out all the zmems left over
5548 add_req(inits);
5549
5550 if (!(UseTLAB && ZeroTLAB)) {
5551 // If anything remains to be zeroed, zero it all now.
5552 zeroes_done = align_down(zeroes_done, BytesPerInt);
5553 // if it is the last unused 4 bytes of an instance, forget about it
5554 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
5555 if (zeroes_done + BytesPerLong >= size_limit) {
5556 AllocateNode* alloc = allocation();
5557 assert(alloc != nullptr, "must be present");
5558 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
5559 Node* klass_node = alloc->in(AllocateNode::KlassNode);
5560 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
5561 if (zeroes_done == k->layout_helper())
5562 zeroes_done = size_limit;
5563 }
5564 }
5565 if (zeroes_done < size_limit) {
5566 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5567 zeroes_done, size_in_bytes, true, phase);
5568 }
5569 }
5570
5571 set_complete(phase);
5572 return rawmem;
5573 }
5574
5575 void InitializeNode::replace_mem_projs_by(Node* mem, Compile* C) {
5576 auto replace_proj = [&](ProjNode* proj) {
5577 C->gvn_replace_by(proj, mem);
5578 return CONTINUE;
5579 };
5580 apply_to_projs(replace_proj, TypeFunc::Memory);
5581 }
5582
5583 void InitializeNode::replace_mem_projs_by(Node* mem, PhaseIterGVN* igvn) {
5584 DUIterator_Fast imax, i = fast_outs(imax);
5585 auto replace_proj = [&](ProjNode* proj) {
5586 igvn->replace_node(proj, mem);
5587 --i; --imax;
5588 return CONTINUE;
5589 };
5590 apply_to_projs(imax, i, replace_proj, TypeFunc::Memory);
5591 }
5592
5593 bool InitializeNode::already_has_narrow_mem_proj_with_adr_type(const TypePtr* adr_type) const {
5594 auto find_proj = [&](ProjNode* proj) {
5595 if (proj->adr_type() == adr_type) {
5596 return BREAK_AND_RETURN_CURRENT_PROJ;
5597 }
5598 return CONTINUE;
5599 };
5600 DUIterator_Fast imax, i = fast_outs(imax);
5601 return apply_to_narrow_mem_projs_any_iterator(UsesIteratorFast(imax, i, this), find_proj) != nullptr;
5602 }
5603
5604 MachProjNode* InitializeNode::mem_mach_proj() const {
5605 auto find_proj = [](ProjNode* proj) {
5606 if (proj->is_MachProj()) {
5607 return BREAK_AND_RETURN_CURRENT_PROJ;
5608 }
5609 return CONTINUE;
5610 };
5611 ProjNode* proj = apply_to_projs(find_proj, TypeFunc::Memory);
5612 if (proj == nullptr) {
5613 return nullptr;
5614 }
5615 return proj->as_MachProj();
5616 }
5617
5618 #ifdef ASSERT
5619 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
5620 if (is_complete())
5621 return true; // stores could be anything at this point
5622 assert(allocation() != nullptr, "must be present");
5623 intptr_t last_off = allocation()->minimum_header_size();
5624 for (uint i = InitializeNode::RawStores; i < req(); i++) {
5625 Node* st = in(i);
5626 intptr_t st_off = get_store_offset(st, phase);
5627 if (st_off < 0) continue; // ignore dead garbage
5628 if (last_off > st_off) {
5629 tty->print_cr("*** bad store offset at %d: %zd > %zd", i, last_off, st_off);
5630 this->dump(2);
5631 assert(false, "ascending store offsets");
5632 return false;
5633 }
5634 last_off = st_off + st->as_Store()->memory_size();
5635 }
5636 return true;
5637 }
5638 #endif //ASSERT
5639
5640
5641
5642
5643 //============================MergeMemNode=====================================
5644 //
5645 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
5646 // contributing store or call operations. Each contributor provides the memory
5647 // state for a particular "alias type" (see Compile::alias_type). For example,
5648 // if a MergeMem has an input X for alias category #6, then any memory reference
5649 // to alias category #6 may use X as its memory state input, as an exact equivalent
5650 // to using the MergeMem as a whole.
5651 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
5652 //
5653 // (Here, the <N> notation gives the index of the relevant adr_type.)
5654 //
5655 // In one special case (and more cases in the future), alias categories overlap.
5656 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
5657 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
5658 // it is exactly equivalent to that state W:
5659 // MergeMem(<Bot>: W) <==> W
5660 //
5661 // Usually, the merge has more than one input. In that case, where inputs
5662 // overlap (i.e., one is Bot), the narrower alias type determines the memory
5663 // state for that type, and the wider alias type (Bot) fills in everywhere else:
5664 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
5665 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
5666 //
5667 // A merge can take a "wide" memory state as one of its narrow inputs.
5668 // This simply means that the merge observes out only the relevant parts of
5669 // the wide input. That is, wide memory states arriving at narrow merge inputs
5670 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
5671 //
5672 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
5673 // and that memory slices "leak through":
5674 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
5675 //
5676 // But, in such a cascade, repeated memory slices can "block the leak":
5677 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
5678 //
5679 // In the last example, Y is not part of the combined memory state of the
5680 // outermost MergeMem. The system must, of course, prevent unschedulable
5681 // memory states from arising, so you can be sure that the state Y is somehow
5682 // a precursor to state Y'.
5683 //
5684 //
5685 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
5686 // of each MergeMemNode array are exactly the numerical alias indexes, including
5687 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
5688 // Compile::alias_type (and kin) produce and manage these indexes.
5689 //
5690 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
5691 // (Note that this provides quick access to the top node inside MergeMem methods,
5692 // without the need to reach out via TLS to Compile::current.)
5693 //
5694 // As a consequence of what was just described, a MergeMem that represents a full
5695 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
5696 // containing all alias categories.
5697 //
5698 // MergeMem nodes never (?) have control inputs, so in(0) is null.
5699 //
5700 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
5701 // a memory state for the alias type <N>, or else the top node, meaning that
5702 // there is no particular input for that alias type. Note that the length of
5703 // a MergeMem is variable, and may be extended at any time to accommodate new
5704 // memory states at larger alias indexes. When merges grow, they are of course
5705 // filled with "top" in the unused in() positions.
5706 //
5707 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
5708 // (Top was chosen because it works smoothly with passes like GCM.)
5709 //
5710 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
5711 // the type of random VM bits like TLS references.) Since it is always the
5712 // first non-Bot memory slice, some low-level loops use it to initialize an
5713 // index variable: for (i = AliasIdxRaw; i < req(); i++).
5714 //
5715 //
5716 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
5717 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
5718 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
5719 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
5720 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
5721 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
5722 //
5723 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
5724 // really that different from the other memory inputs. An abbreviation called
5725 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
5726 //
5727 //
5728 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
5729 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
5730 // that "emerges though" the base memory will be marked as excluding the alias types
5731 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
5732 //
5733 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
5734 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
5735 //
5736 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
5737 // (It is currently unimplemented.) As you can see, the resulting merge is
5738 // actually a disjoint union of memory states, rather than an overlay.
5739 //
5740
5741 //------------------------------MergeMemNode-----------------------------------
5742 Node* MergeMemNode::make_empty_memory() {
5743 Node* empty_memory = (Node*) Compile::current()->top();
5744 assert(empty_memory->is_top(), "correct sentinel identity");
5745 return empty_memory;
5746 }
5747
5748 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
5749 init_class_id(Class_MergeMem);
5750 // all inputs are nullified in Node::Node(int)
5751 // set_input(0, nullptr); // no control input
5752
5753 // Initialize the edges uniformly to top, for starters.
5754 Node* empty_mem = make_empty_memory();
5755 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
5756 init_req(i,empty_mem);
5757 }
5758 assert(empty_memory() == empty_mem, "");
5759
5760 if( new_base != nullptr && new_base->is_MergeMem() ) {
5761 MergeMemNode* mdef = new_base->as_MergeMem();
5762 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
5763 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
5764 mms.set_memory(mms.memory2());
5765 }
5766 assert(base_memory() == mdef->base_memory(), "");
5767 } else {
5768 set_base_memory(new_base);
5769 }
5770 }
5771
5772 // Make a new, untransformed MergeMem with the same base as 'mem'.
5773 // If mem is itself a MergeMem, populate the result with the same edges.
5774 MergeMemNode* MergeMemNode::make(Node* mem) {
5775 return new MergeMemNode(mem);
5776 }
5777
5778 //------------------------------cmp--------------------------------------------
5779 uint MergeMemNode::hash() const { return NO_HASH; }
5780 bool MergeMemNode::cmp( const Node &n ) const {
5781 return (&n == this); // Always fail except on self
5782 }
5783
5784 //------------------------------Identity---------------------------------------
5785 Node* MergeMemNode::Identity(PhaseGVN* phase) {
5786 // Identity if this merge point does not record any interesting memory
5787 // disambiguations.
5788 Node* base_mem = base_memory();
5789 Node* empty_mem = empty_memory();
5790 if (base_mem != empty_mem) { // Memory path is not dead?
5791 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5792 Node* mem = in(i);
5793 if (mem != empty_mem && mem != base_mem) {
5794 return this; // Many memory splits; no change
5795 }
5796 }
5797 }
5798 return base_mem; // No memory splits; ID on the one true input
5799 }
5800
5801 //------------------------------Ideal------------------------------------------
5802 // This method is invoked recursively on chains of MergeMem nodes
5803 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
5804 // Remove chain'd MergeMems
5805 //
5806 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
5807 // relative to the "in(Bot)". Since we are patching both at the same time,
5808 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
5809 // but rewrite each "in(i)" relative to the new "in(Bot)".
5810 Node *progress = nullptr;
5811
5812
5813 Node* old_base = base_memory();
5814 Node* empty_mem = empty_memory();
5815 if (old_base == empty_mem)
5816 return nullptr; // Dead memory path.
5817
5818 MergeMemNode* old_mbase;
5819 if (old_base != nullptr && old_base->is_MergeMem())
5820 old_mbase = old_base->as_MergeMem();
5821 else
5822 old_mbase = nullptr;
5823 Node* new_base = old_base;
5824
5825 // simplify stacked MergeMems in base memory
5826 if (old_mbase) new_base = old_mbase->base_memory();
5827
5828 // the base memory might contribute new slices beyond my req()
5829 if (old_mbase) grow_to_match(old_mbase);
5830
5831 // Note: We do not call verify_sparse on entry, because inputs
5832 // can normalize to the base_memory via subsume_node or similar
5833 // mechanisms. This method repairs that damage.
5834
5835 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
5836
5837 // Look at each slice.
5838 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5839 Node* old_in = in(i);
5840 // calculate the old memory value
5841 Node* old_mem = old_in;
5842 if (old_mem == empty_mem) old_mem = old_base;
5843 assert(old_mem == memory_at(i), "");
5844
5845 // maybe update (reslice) the old memory value
5846
5847 // simplify stacked MergeMems
5848 Node* new_mem = old_mem;
5849 MergeMemNode* old_mmem;
5850 if (old_mem != nullptr && old_mem->is_MergeMem())
5851 old_mmem = old_mem->as_MergeMem();
5852 else
5853 old_mmem = nullptr;
5854 if (old_mmem == this) {
5855 // This can happen if loops break up and safepoints disappear.
5856 // A merge of BotPtr (default) with a RawPtr memory derived from a
5857 // safepoint can be rewritten to a merge of the same BotPtr with
5858 // the BotPtr phi coming into the loop. If that phi disappears
5859 // also, we can end up with a self-loop of the mergemem.
5860 // In general, if loops degenerate and memory effects disappear,
5861 // a mergemem can be left looking at itself. This simply means
5862 // that the mergemem's default should be used, since there is
5863 // no longer any apparent effect on this slice.
5864 // Note: If a memory slice is a MergeMem cycle, it is unreachable
5865 // from start. Update the input to TOP.
5866 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
5867 }
5868 else if (old_mmem != nullptr) {
5869 new_mem = old_mmem->memory_at(i);
5870 }
5871 // else preceding memory was not a MergeMem
5872
5873 // maybe store down a new value
5874 Node* new_in = new_mem;
5875 if (new_in == new_base) new_in = empty_mem;
5876
5877 if (new_in != old_in) {
5878 // Warning: Do not combine this "if" with the previous "if"
5879 // A memory slice might have be be rewritten even if it is semantically
5880 // unchanged, if the base_memory value has changed.
5881 set_req_X(i, new_in, phase);
5882 progress = this; // Report progress
5883 }
5884 }
5885
5886 if (new_base != old_base) {
5887 set_req_X(Compile::AliasIdxBot, new_base, phase);
5888 // Don't use set_base_memory(new_base), because we need to update du.
5889 assert(base_memory() == new_base, "");
5890 progress = this;
5891 }
5892
5893 if( base_memory() == this ) {
5894 // a self cycle indicates this memory path is dead
5895 set_req(Compile::AliasIdxBot, empty_mem);
5896 }
5897
5898 // Resolve external cycles by calling Ideal on a MergeMem base_memory
5899 // Recursion must occur after the self cycle check above
5900 if( base_memory()->is_MergeMem() ) {
5901 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
5902 Node *m = phase->transform(new_mbase); // Rollup any cycles
5903 if( m != nullptr &&
5904 (m->is_top() ||
5905 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
5906 // propagate rollup of dead cycle to self
5907 set_req(Compile::AliasIdxBot, empty_mem);
5908 }
5909 }
5910
5911 if( base_memory() == empty_mem ) {
5912 progress = this;
5913 // Cut inputs during Parse phase only.
5914 // During Optimize phase a dead MergeMem node will be subsumed by Top.
5915 if( !can_reshape ) {
5916 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5917 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
5918 }
5919 }
5920 }
5921
5922 if( !progress && base_memory()->is_Phi() && can_reshape ) {
5923 // Check if PhiNode::Ideal's "Split phis through memory merges"
5924 // transform should be attempted. Look for this->phi->this cycle.
5925 uint merge_width = req();
5926 if (merge_width > Compile::AliasIdxRaw) {
5927 PhiNode* phi = base_memory()->as_Phi();
5928 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
5929 if (phi->in(i) == this) {
5930 phase->is_IterGVN()->_worklist.push(phi);
5931 break;
5932 }
5933 }
5934 }
5935 }
5936
5937 assert(progress || verify_sparse(), "please, no dups of base");
5938 return progress;
5939 }
5940
5941 //-------------------------set_base_memory-------------------------------------
5942 void MergeMemNode::set_base_memory(Node *new_base) {
5943 Node* empty_mem = empty_memory();
5944 set_req(Compile::AliasIdxBot, new_base);
5945 assert(memory_at(req()) == new_base, "must set default memory");
5946 // Clear out other occurrences of new_base:
5947 if (new_base != empty_mem) {
5948 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5949 if (in(i) == new_base) set_req(i, empty_mem);
5950 }
5951 }
5952 }
5953
5954 //------------------------------out_RegMask------------------------------------
5955 const RegMask &MergeMemNode::out_RegMask() const {
5956 return RegMask::EMPTY;
5957 }
5958
5959 //------------------------------dump_spec--------------------------------------
5960 #ifndef PRODUCT
5961 void MergeMemNode::dump_spec(outputStream *st) const {
5962 st->print(" {");
5963 Node* base_mem = base_memory();
5964 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
5965 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem;
5966 if (mem == base_mem) { st->print(" -"); continue; }
5967 st->print( " N%d:", mem->_idx );
5968 Compile::current()->get_adr_type(i)->dump_on(st);
5969 }
5970 st->print(" }");
5971 }
5972 #endif // !PRODUCT
5973
5974
5975 #ifdef ASSERT
5976 static bool might_be_same(Node* a, Node* b) {
5977 if (a == b) return true;
5978 if (!(a->is_Phi() || b->is_Phi())) return false;
5979 // phis shift around during optimization
5980 return true; // pretty stupid...
5981 }
5982
5983 // verify a narrow slice (either incoming or outgoing)
5984 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
5985 if (!VerifyAliases) return; // don't bother to verify unless requested
5986 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
5987 if (Node::in_dump()) return; // muzzle asserts when printing
5988 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
5989 assert(n != nullptr, "");
5990 // Elide intervening MergeMem's
5991 while (n->is_MergeMem()) {
5992 n = n->as_MergeMem()->memory_at(alias_idx);
5993 }
5994 Compile* C = Compile::current();
5995 const TypePtr* n_adr_type = n->adr_type();
5996 if (n == m->empty_memory()) {
5997 // Implicit copy of base_memory()
5998 } else if (n_adr_type != TypePtr::BOTTOM) {
5999 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type");
6000 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
6001 } else {
6002 // A few places like make_runtime_call "know" that VM calls are narrow,
6003 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
6004 bool expected_wide_mem = false;
6005 if (n == m->base_memory()) {
6006 expected_wide_mem = true;
6007 } else if (alias_idx == Compile::AliasIdxRaw ||
6008 n == m->memory_at(Compile::AliasIdxRaw)) {
6009 expected_wide_mem = true;
6010 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
6011 // memory can "leak through" calls on channels that
6012 // are write-once. Allow this also.
6013 expected_wide_mem = true;
6014 }
6015 assert(expected_wide_mem, "expected narrow slice replacement");
6016 }
6017 }
6018 #else // !ASSERT
6019 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
6020 #endif
6021
6022
6023 //-----------------------------memory_at---------------------------------------
6024 Node* MergeMemNode::memory_at(uint alias_idx) const {
6025 assert(alias_idx >= Compile::AliasIdxRaw ||
6026 (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()),
6027 "must avoid base_memory and AliasIdxTop");
6028
6029 // Otherwise, it is a narrow slice.
6030 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
6031 if (is_empty_memory(n)) {
6032 // the array is sparse; empty slots are the "top" node
6033 n = base_memory();
6034 assert(Node::in_dump()
6035 || n == nullptr || n->bottom_type() == Type::TOP
6036 || n->adr_type() == nullptr // address is TOP
6037 || n->adr_type() == TypePtr::BOTTOM
6038 || n->adr_type() == TypeRawPtr::BOTTOM
6039 || n->is_NarrowMemProj()
6040 || !Compile::current()->do_aliasing(),
6041 "must be a wide memory");
6042 // do_aliasing == false if we are organizing the memory states manually.
6043 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
6044 } else {
6045 // make sure the stored slice is sane
6046 #ifdef ASSERT
6047 if (VMError::is_error_reported() || Node::in_dump()) {
6048 } else if (might_be_same(n, base_memory())) {
6049 // Give it a pass: It is a mostly harmless repetition of the base.
6050 // This can arise normally from node subsumption during optimization.
6051 } else {
6052 verify_memory_slice(this, alias_idx, n);
6053 }
6054 #endif
6055 }
6056 return n;
6057 }
6058
6059 //---------------------------set_memory_at-------------------------------------
6060 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
6061 verify_memory_slice(this, alias_idx, n);
6062 Node* empty_mem = empty_memory();
6063 if (n == base_memory()) n = empty_mem; // collapse default
6064 uint need_req = alias_idx+1;
6065 if (req() < need_req) {
6066 if (n == empty_mem) return; // already the default, so do not grow me
6067 // grow the sparse array
6068 do {
6069 add_req(empty_mem);
6070 } while (req() < need_req);
6071 }
6072 set_req( alias_idx, n );
6073 }
6074
6075
6076
6077 //--------------------------iteration_setup------------------------------------
6078 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
6079 if (other != nullptr) {
6080 grow_to_match(other);
6081 // invariant: the finite support of mm2 is within mm->req()
6082 #ifdef ASSERT
6083 for (uint i = req(); i < other->req(); i++) {
6084 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
6085 }
6086 #endif
6087 }
6088 // Replace spurious copies of base_memory by top.
6089 Node* base_mem = base_memory();
6090 if (base_mem != nullptr && !base_mem->is_top()) {
6091 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
6092 if (in(i) == base_mem)
6093 set_req(i, empty_memory());
6094 }
6095 }
6096 }
6097
6098 //---------------------------grow_to_match-------------------------------------
6099 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
6100 Node* empty_mem = empty_memory();
6101 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
6102 // look for the finite support of the other memory
6103 for (uint i = other->req(); --i >= req(); ) {
6104 if (other->in(i) != empty_mem) {
6105 uint new_len = i+1;
6106 while (req() < new_len) add_req(empty_mem);
6107 break;
6108 }
6109 }
6110 }
6111
6112 //---------------------------verify_sparse-------------------------------------
6113 #ifndef PRODUCT
6114 bool MergeMemNode::verify_sparse() const {
6115 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
6116 Node* base_mem = base_memory();
6117 // The following can happen in degenerate cases, since empty==top.
6118 if (is_empty_memory(base_mem)) return true;
6119 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6120 assert(in(i) != nullptr, "sane slice");
6121 if (in(i) == base_mem) return false; // should have been the sentinel value!
6122 }
6123 return true;
6124 }
6125
6126 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
6127 Node* n;
6128 n = mm->in(idx);
6129 if (mem == n) return true; // might be empty_memory()
6130 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
6131 if (mem == n) return true;
6132 return false;
6133 }
6134 #endif // !PRODUCT