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 }
771 return 0;
772 }
773
774 AccessAnalyzer::AccessAnalyzer(PhaseValues* phase, MemNode* n)
775 : _phase(phase), _n(n), _memory_size(n->memory_size()), _alias_idx(-1) {
776 Node* adr = _n->in(MemNode::Address);
777 _offset = 0;
778 _base = AddPNode::Ideal_base_and_offset(adr, _phase, _offset);
779 _maybe_raw = MemNode::check_if_adr_maybe_raw(adr);
780 _alloc = AllocateNode::Ideal_allocation(_base);
781 _adr_type = _n->adr_type();
782
783 if (_adr_type != nullptr && _adr_type->base() != TypePtr::AnyPtr) {
784 // Avoid the cases that will upset Compile::get_alias_index
785 _alias_idx = _phase->C->get_alias_index(_adr_type);
786 assert(_alias_idx != Compile::AliasIdxTop, "must not be a dead node");
787 assert(_alias_idx != Compile::AliasIdxBot || !phase->C->do_aliasing(), "must not be a very wide access");
788 }
789 }
790
791 // Decide whether the memory accessed by '_n' and 'other' may overlap. This function may be used
792 // when we want to walk the memory graph to fold a load, or when we want to hoist a load above a
793 // loop when there are no stores that may overlap with the load inside the loop.
794 AccessAnalyzer::AccessIndependence AccessAnalyzer::detect_access_independence(Node* other) const {
795 assert(_phase->type(other) == Type::MEMORY, "must be a memory node %s", other->Name());
796 assert(!other->is_Phi(), "caller must handle Phi");
797
798 if (_adr_type == nullptr) {
799 // This means the access is dead
800 return {false, _phase->C->top()};
801 } else if (_adr_type->base() == TypePtr::AnyPtr) {
802 // An example for this case is an access into the memory address 0 performed using Unsafe
803 assert(_adr_type->ptr() == TypePtr::Null, "MemNode should never access a wide memory");
804 return {false, nullptr};
805 }
806
807 if (_offset == Type::OffsetBot) {
808 // cannot unalias unless there are precise offsets
809 return {false, nullptr};
810 }
811
812 const TypeOopPtr* adr_oop_type = _adr_type->isa_oopptr();
813 Node* prev = other;
814 if (other->is_Store()) {
815 Node* st_adr = other->in(MemNode::Address);
816 intptr_t st_offset = 0;
817 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, _phase, st_offset);
818 if (st_base == nullptr) {
819 // inscrutable pointer
820 return {false, nullptr};
821 }
822
823 // If the bases are the same and the offsets are the same, it seems that this is the exact
824 // store we are looking for, the caller will check if the type of the store matches using
825 // MemNode::can_see_stored_value
826 if (st_base == _base && st_offset == _offset) {
827 return {false, other};
828 }
829
830 // If it is provable that the memory accessed by 'other' does not overlap the memory accessed
831 // by '_n', we may walk past 'other'.
832 // For raw accesses, 2 accesses are independent if they have the same base and the offsets
833 // say that they do not overlap.
834 // For heap accesses, 2 accesses are independent if either the bases are provably different
835 // at runtime or the offsets say that the accesses do not overlap.
836 if ((_maybe_raw || MemNode::check_if_adr_maybe_raw(st_adr)) && st_base != _base) {
837 // Raw accesses can only be provably independent if they have the same base
838 return {false, nullptr};
839 }
840
841 // If the offsets say that the accesses do not overlap, then it is provable that 'other' and
842 // '_n' do not overlap. For example, a LoadI from Object+8 is independent from a StoreL into
843 // Object+12, no matter what the bases are.
844 if (st_offset != _offset && st_offset != Type::OffsetBot) {
845 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
846 assert(other->as_Store()->memory_size() <= MAX_STORE, "");
847 if (st_offset >= _offset + _memory_size ||
848 st_offset <= _offset - MAX_STORE ||
849 st_offset <= _offset - other->as_Store()->memory_size()) {
850 // Success: The offsets are provably independent.
851 // (You may ask, why not just test st_offset != offset and be done?
852 // The answer is that stores of different sizes can co-exist
853 // in the same sequence of RawMem effects. We sometimes initialize
854 // a whole 'tile' of array elements with a single jint or jlong.)
855 return {true, other->in(MemNode::Memory)};
856 }
857 }
858
859 // Same base and overlapping offsets, it seems provable that the accesses overlap, give up
860 if (st_base == _base) {
861 return {false, nullptr};
862 }
863
864 // Try to prove that 2 different base nodes at compile time are different values at runtime
865 bool known_independent = false;
866 if (MemNode::detect_ptr_independence(_base, _alloc, st_base, AllocateNode::Ideal_allocation(st_base), _phase)) {
867 known_independent = true;
868 }
869
870 if (known_independent) {
871 return {true, other->in(MemNode::Memory)};
872 }
873 } else if (other->is_Proj() && other->in(0)->is_Initialize()) {
874 InitializeNode* st_init = other->in(0)->as_Initialize();
875 AllocateNode* st_alloc = st_init->allocation();
876 if (st_alloc == nullptr) {
877 // Something degenerated
878 return {false, nullptr};
879 }
880 bool known_identical = false;
881 bool known_independent = false;
882 if (_alloc == st_alloc) {
883 known_identical = true;
884 } else if (_alloc != nullptr) {
885 known_independent = true;
886 } else if (MemNode::all_controls_dominate(_n, st_alloc)) {
887 known_independent = true;
888 }
889
890 if (known_independent) {
891 // The bases are provably independent: Either they are
892 // manifestly distinct allocations, or else the control
893 // of _n dominates the store's allocation.
894 if (_alias_idx == Compile::AliasIdxRaw) {
895 other = st_alloc->in(TypeFunc::Memory);
896 } else {
897 other = st_init->memory(_alias_idx);
898 }
899 return {true, other};
900 }
901
902 // If we are not looking at a store initializing the same
903 // allocation we are loading from, we lose.
904 if (known_identical) {
905 // From caller, can_see_stored_value will consult find_captured_store.
906 return {false, other};
907 }
908
909 } else if (_n->find_previous_arraycopy(_phase, _alloc, other, false) != nullptr) {
910 // Find an arraycopy that may or may not affect the MemNode
911 return {prev != other, other};
912 } else if (other->is_MergeMem()) {
913 return {true, other->as_MergeMem()->memory_at(_alias_idx)};
914 } else if (adr_oop_type != nullptr && adr_oop_type->is_known_instance_field()) {
915 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
916 if (other->is_Proj() && other->in(0)->is_Call()) {
917 // ArrayCopyNodes processed here as well.
918 CallNode* call = other->in(0)->as_Call();
919 if (!call->may_modify(adr_oop_type, _phase)) {
920 return {true, call->in(TypeFunc::Memory)};
921 }
922 } else if (other->is_Proj() && other->in(0)->is_MemBar()) {
923 ArrayCopyNode* ac = nullptr;
924 if (!ArrayCopyNode::may_modify(adr_oop_type, other->in(0)->as_MemBar(), _phase, ac)) {
925 return {true, other->in(0)->in(TypeFunc::Memory)};
926 }
927 } else if (other->is_ClearArray()) {
928 if (ClearArrayNode::step_through(&other, (uint)adr_oop_type->instance_id(), _phase)) {
929 // (the call updated 'other' value)
930 return {true, other};
931 } else {
932 // Can not bypass initialization of the instance
933 // we are looking for.
934 return {false, other};
935 }
936 }
937 }
938
939 return {false, nullptr};
940 }
941
942 //=============================================================================
943 // Should LoadNode::Ideal() attempt to remove control edges?
944 bool LoadNode::can_remove_control() const {
945 return !has_pinned_control_dependency();
946 }
947 uint LoadNode::size_of() const { return sizeof(*this); }
948 bool LoadNode::cmp(const Node &n) const {
949 LoadNode& load = (LoadNode &)n;
950 return Type::equals(_type, load._type) &&
951 _control_dependency == load._control_dependency &&
952 _mo == load._mo;
953 }
954 const Type *LoadNode::bottom_type() const { return _type; }
955 uint LoadNode::ideal_reg() const {
956 return _type->ideal_reg();
957 }
958
959 #ifndef PRODUCT
960 void LoadNode::dump_spec(outputStream *st) const {
961 MemNode::dump_spec(st);
962 if( !Verbose && !WizardMode ) {
963 // standard dump does this in Verbose and WizardMode
964 st->print(" #"); _type->dump_on(st);
965 }
966 if (in(0) != nullptr && !depends_only_on_test()) {
967 st->print(" (does not depend only on test, ");
968 if (control_dependency() == UnknownControl) {
969 st->print("unknown control");
970 } else if (control_dependency() == Pinned) {
971 st->print("pinned");
972 } else if (adr_type() == TypeRawPtr::BOTTOM) {
973 st->print("raw access");
974 } else {
975 st->print("unknown reason");
976 }
977 st->print(")");
978 }
979 }
980 #endif
981
982 #ifdef ASSERT
983 //----------------------------is_immutable_value-------------------------------
984 // Helper function to allow a raw load without control edge for some cases
985 bool LoadNode::is_immutable_value(Node* adr) {
986 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
987 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) {
988
989 jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1);
990 int offsets[] = {
991 in_bytes(JavaThread::osthread_offset()),
992 in_bytes(JavaThread::threadObj_offset()),
993 in_bytes(JavaThread::vthread_offset()),
994 in_bytes(JavaThread::scopedValueCache_offset()),
995 };
996
997 for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) {
998 if (offset == offsets[i]) {
999 return true;
1000 }
1001 }
1002 }
1003
1004 return false;
1005 }
1006 #endif
1007
1008 //----------------------------LoadNode::make-----------------------------------
1009 // Polymorphic factory method:
1010 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo,
1011 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
1012 Compile* C = gvn.C;
1013 assert(adr->is_top() || C->get_alias_index(gvn.type(adr)->is_ptr()) == C->get_alias_index(adr_type), "adr and adr_type must agree");
1014
1015 // sanity check the alias category against the created node type
1016 assert(!(adr_type->isa_oopptr() &&
1017 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
1018 "use LoadKlassNode instead");
1019 assert(!(adr_type->isa_aryptr() &&
1020 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
1021 "use LoadRangeNode instead");
1022 // Check control edge of raw loads
1023 assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
1024 // oop will be recorded in oop map if load crosses safepoint
1025 rt->isa_oopptr() || is_immutable_value(adr),
1026 "raw memory operations should have control edge");
1027 LoadNode* load = nullptr;
1028 switch (bt) {
1029 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1030 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1031 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1032 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1033 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1034 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break;
1035 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
1036 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break;
1037 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
1038 case T_OBJECT:
1039 case T_NARROWOOP:
1040 #ifdef _LP64
1041 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1042 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
1043 } else
1044 #endif
1045 {
1046 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
1047 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
1048 }
1049 break;
1050 default:
1051 ShouldNotReachHere();
1052 break;
1053 }
1054 assert(load != nullptr, "LoadNode should have been created");
1055 if (unaligned) {
1056 load->set_unaligned_access();
1057 }
1058 if (mismatched) {
1059 load->set_mismatched_access();
1060 }
1061 if (unsafe) {
1062 load->set_unsafe_access();
1063 }
1064 load->set_barrier_data(barrier_data);
1065 if (load->Opcode() == Op_LoadN) {
1066 Node* ld = gvn.transform(load);
1067 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
1068 }
1069
1070 return load;
1071 }
1072
1073 //------------------------------hash-------------------------------------------
1074 uint LoadNode::hash() const {
1075 // unroll addition of interesting fields
1076 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
1077 }
1078
1079 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
1080 if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) {
1081 bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile();
1082 bool is_stable_ary = FoldStableValues &&
1083 (tp != nullptr) && (tp->isa_aryptr() != nullptr) &&
1084 tp->isa_aryptr()->is_stable();
1085
1086 return (eliminate_boxing && non_volatile) || is_stable_ary;
1087 }
1088
1089 return false;
1090 }
1091
1092 // Is the value loaded previously stored by an arraycopy? If so return
1093 // a load node that reads from the source array so we may be able to
1094 // optimize out the ArrayCopy node later.
1095 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
1096 Node* ld_adr = in(MemNode::Address);
1097 intptr_t ld_off = 0;
1098 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
1099 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
1100 if (ac != nullptr) {
1101 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
1102
1103 Node* mem = ac->in(TypeFunc::Memory);
1104 Node* ctl = ac->in(0);
1105 Node* src = ac->in(ArrayCopyNode::Src);
1106
1107 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
1108 return nullptr;
1109 }
1110
1111 // load depends on the tests that validate the arraycopy
1112 LoadNode* ld = clone_pinned();
1113 Node* addp = in(MemNode::Address)->clone();
1114 if (ac->as_ArrayCopy()->is_clonebasic()) {
1115 assert(ld_alloc != nullptr, "need an alloc");
1116 assert(addp->is_AddP(), "address must be addp");
1117 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1118 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1119 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1120 addp->set_req(AddPNode::Base, src);
1121 addp->set_req(AddPNode::Address, src);
1122 } else {
1123 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1124 ac->as_ArrayCopy()->is_copyof_validated() ||
1125 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1126 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1127 addp->set_req(AddPNode::Base, src);
1128 addp->set_req(AddPNode::Address, src);
1129
1130 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1131 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
1132 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
1133
1134 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
1135 uint shift = exact_log2(type2aelembytes(ary_elem));
1136
1137 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1138 #ifdef _LP64
1139 diff = phase->transform(new ConvI2LNode(diff));
1140 #endif
1141 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1142
1143 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1144 addp->set_req(AddPNode::Offset, offset);
1145 }
1146 addp = phase->transform(addp);
1147 #ifdef ASSERT
1148 const TypePtr* adr_type = phase->type(addp)->is_ptr();
1149 ld->_adr_type = adr_type;
1150 #endif
1151 ld->set_req(MemNode::Address, addp);
1152 ld->set_req(0, ctl);
1153 ld->set_req(MemNode::Memory, mem);
1154 return ld;
1155 }
1156 return nullptr;
1157 }
1158
1159
1160 //---------------------------can_see_stored_value------------------------------
1161 // This routine exists to make sure this set of tests is done the same
1162 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
1163 // will change the graph shape in a way which makes memory alive twice at the
1164 // same time (uses the Oracle model of aliasing), then some
1165 // LoadXNode::Identity will fold things back to the equivalence-class model
1166 // of aliasing.
1167 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const {
1168 Node* ld_adr = in(MemNode::Address);
1169 intptr_t ld_off = 0;
1170 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1171 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base);
1172 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1173 Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr;
1174 // This is more general than load from boxing objects.
1175 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1176 uint alias_idx = atp->index();
1177 Node* result = nullptr;
1178 Node* current = st;
1179 // Skip through chains of MemBarNodes checking the MergeMems for
1180 // new states for the slice of this load. Stop once any other
1181 // kind of node is encountered. Loads from final memory can skip
1182 // through any kind of MemBar but normal loads shouldn't skip
1183 // through MemBarAcquire since the could allow them to move out of
1184 // a synchronized region. It is not safe to step over MemBarCPUOrder,
1185 // because alias info above them may be inaccurate (e.g., due to
1186 // mixed/mismatched unsafe accesses).
1187 bool is_final_mem = !atp->is_rewritable();
1188 while (current->is_Proj()) {
1189 int opc = current->in(0)->Opcode();
1190 if ((is_final_mem && (opc == Op_MemBarAcquire ||
1191 opc == Op_MemBarAcquireLock ||
1192 opc == Op_LoadFence)) ||
1193 opc == Op_MemBarRelease ||
1194 opc == Op_StoreFence ||
1195 opc == Op_MemBarReleaseLock ||
1196 opc == Op_MemBarStoreStore ||
1197 opc == Op_StoreStoreFence) {
1198 Node* mem = current->in(0)->in(TypeFunc::Memory);
1199 if (mem->is_MergeMem()) {
1200 MergeMemNode* merge = mem->as_MergeMem();
1201 Node* new_st = merge->memory_at(alias_idx);
1202 if (new_st == merge->base_memory()) {
1203 // Keep searching
1204 current = new_st;
1205 continue;
1206 }
1207 // Save the new memory state for the slice and fall through
1208 // to exit.
1209 result = new_st;
1210 }
1211 }
1212 break;
1213 }
1214 if (result != nullptr) {
1215 st = result;
1216 }
1217 }
1218
1219 // Loop around twice in the case Load -> Initialize -> Store.
1220 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1221 for (int trip = 0; trip <= 1; trip++) {
1222
1223 if (st->is_Store()) {
1224 Node* st_adr = st->in(MemNode::Address);
1225 if (st_adr != ld_adr) {
1226 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1227 intptr_t st_off = 0;
1228 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1229 if (ld_base == nullptr) return nullptr;
1230 if (st_base == nullptr) return nullptr;
1231 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr;
1232 if (ld_off != st_off) return nullptr;
1233 if (ld_off == Type::OffsetBot) return nullptr;
1234 // Same base, same offset.
1235 // Possible improvement for arrays: check index value instead of absolute offset.
1236
1237 // At this point we have proven something like this setup:
1238 // B = << base >>
1239 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1240 // S = StoreQ(AddP( B , #Off), V)
1241 // (Actually, we haven't yet proven the Q's are the same.)
1242 // In other words, we are loading from a casted version of
1243 // the same pointer-and-offset that we stored to.
1244 // Casted version may carry a dependency and it is respected.
1245 // Thus, we are able to replace L by V.
1246 }
1247 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1248 if (store_Opcode() != st->Opcode()) {
1249 return nullptr;
1250 }
1251 // LoadVector/StoreVector needs additional check to ensure the types match.
1252 if (st->is_StoreVector()) {
1253 if ((Opcode() != Op_LoadVector && Opcode() != Op_StoreVector) || st->Opcode() != Op_StoreVector) {
1254 // Some kind of masked access or gather/scatter
1255 return nullptr;
1256 }
1257
1258 const TypeVect* in_vt = st->as_StoreVector()->vect_type();
1259 const TypeVect* out_vt = is_Load() ? as_LoadVector()->vect_type() : as_StoreVector()->vect_type();
1260 if (in_vt != out_vt) {
1261 return nullptr;
1262 }
1263 }
1264 return st->in(MemNode::ValueIn);
1265 }
1266
1267 // A load from a freshly-created object always returns zero.
1268 // (This can happen after LoadNode::Ideal resets the load's memory input
1269 // to find_captured_store, which returned InitializeNode::zero_memory.)
1270 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1271 (st->in(0) == ld_alloc) &&
1272 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1273 // return a zero value for the load's basic type
1274 // (This is one of the few places where a generic PhaseTransform
1275 // can create new nodes. Think of it as lazily manifesting
1276 // virtually pre-existing constants.)
1277 if (value_basic_type() != T_VOID) {
1278 if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) {
1279 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1280 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1281 // by the ArrayCopyNode.
1282 return phase->zerocon(value_basic_type());
1283 }
1284 } else {
1285 // TODO: materialize all-zero vector constant
1286 assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1287 }
1288 }
1289
1290 // A load from an initialization barrier can match a captured store.
1291 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1292 InitializeNode* init = st->in(0)->as_Initialize();
1293 AllocateNode* alloc = init->allocation();
1294 if ((alloc != nullptr) && (alloc == ld_alloc)) {
1295 // examine a captured store value
1296 st = init->find_captured_store(ld_off, memory_size(), phase);
1297 if (st != nullptr) {
1298 continue; // take one more trip around
1299 }
1300 }
1301 }
1302
1303 // Load boxed value from result of valueOf() call is input parameter.
1304 if (this->is_Load() && ld_adr->is_AddP() &&
1305 (tp != nullptr) && tp->is_ptr_to_boxed_value()) {
1306 intptr_t ignore = 0;
1307 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1308 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1309 base = bs->step_over_gc_barrier(base);
1310 if (base != nullptr && base->is_Proj() &&
1311 base->as_Proj()->_con == TypeFunc::Parms &&
1312 base->in(0)->is_CallStaticJava() &&
1313 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1314 return base->in(0)->in(TypeFunc::Parms);
1315 }
1316 }
1317
1318 break;
1319 }
1320
1321 return nullptr;
1322 }
1323
1324 //----------------------is_instance_field_load_with_local_phi------------------
1325 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1326 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1327 in(Address)->is_AddP() ) {
1328 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1329 // Only instances and boxed values.
1330 if( t_oop != nullptr &&
1331 (t_oop->is_ptr_to_boxed_value() ||
1332 t_oop->is_known_instance_field()) &&
1333 t_oop->offset() != Type::OffsetBot &&
1334 t_oop->offset() != Type::OffsetTop) {
1335 return true;
1336 }
1337 }
1338 return false;
1339 }
1340
1341 //------------------------------Identity---------------------------------------
1342 // Loads are identity if previous store is to same address
1343 Node* LoadNode::Identity(PhaseGVN* phase) {
1344 // If the previous store-maker is the right kind of Store, and the store is
1345 // to the same address, then we are equal to the value stored.
1346 Node* mem = in(Memory);
1347 Node* value = can_see_stored_value(mem, phase);
1348 if( value ) {
1349 // byte, short & char stores truncate naturally.
1350 // A load has to load the truncated value which requires
1351 // some sort of masking operation and that requires an
1352 // Ideal call instead of an Identity call.
1353 if (memory_size() < BytesPerInt) {
1354 // If the input to the store does not fit with the load's result type,
1355 // it must be truncated via an Ideal call.
1356 if (!phase->type(value)->higher_equal(phase->type(this)))
1357 return this;
1358 }
1359 // (This works even when value is a Con, but LoadNode::Value
1360 // usually runs first, producing the singleton type of the Con.)
1361 if (!has_pinned_control_dependency() || value->is_Con()) {
1362 return value;
1363 } else {
1364 return this;
1365 }
1366 }
1367
1368 if (has_pinned_control_dependency()) {
1369 return this;
1370 }
1371 // Search for an existing data phi which was generated before for the same
1372 // instance's field to avoid infinite generation of phis in a loop.
1373 Node *region = mem->in(0);
1374 if (is_instance_field_load_with_local_phi(region)) {
1375 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1376 int this_index = phase->C->get_alias_index(addr_t);
1377 int this_offset = addr_t->offset();
1378 int this_iid = addr_t->instance_id();
1379 if (!addr_t->is_known_instance() &&
1380 addr_t->is_ptr_to_boxed_value()) {
1381 // Use _idx of address base (could be Phi node) for boxed values.
1382 intptr_t ignore = 0;
1383 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1384 if (base == nullptr) {
1385 return this;
1386 }
1387 this_iid = base->_idx;
1388 }
1389 const Type* this_type = bottom_type();
1390 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1391 Node* phi = region->fast_out(i);
1392 if (phi->is_Phi() && phi != mem &&
1393 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1394 return phi;
1395 }
1396 }
1397 }
1398
1399 return this;
1400 }
1401
1402 // Construct an equivalent unsigned load.
1403 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1404 BasicType bt = T_ILLEGAL;
1405 const Type* rt = nullptr;
1406 switch (Opcode()) {
1407 case Op_LoadUB: return this;
1408 case Op_LoadUS: return this;
1409 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1410 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1411 default:
1412 assert(false, "no unsigned variant: %s", Name());
1413 return nullptr;
1414 }
1415 const Type* mem_t = gvn.type(in(MemNode::Address));
1416 if (mem_t == Type::TOP) {
1417 return gvn.C->top();
1418 }
1419 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1420 mem_t->is_ptr(), rt, bt, _mo, _control_dependency,
1421 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1422 }
1423
1424 // Construct an equivalent signed load.
1425 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1426 BasicType bt = T_ILLEGAL;
1427 const Type* rt = nullptr;
1428 switch (Opcode()) {
1429 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1430 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1431 case Op_LoadB: // fall through
1432 case Op_LoadS: // fall through
1433 case Op_LoadI: // fall through
1434 case Op_LoadL: return this;
1435 default:
1436 assert(false, "no signed variant: %s", Name());
1437 return nullptr;
1438 }
1439 const Type* mem_t = gvn.type(in(MemNode::Address));
1440 if (mem_t == Type::TOP) {
1441 return gvn.C->top();
1442 }
1443 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1444 mem_t->is_ptr(), rt, bt, _mo, _control_dependency,
1445 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1446 }
1447
1448 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1449 BasicType bt = rt->basic_type();
1450 switch (Opcode()) {
1451 case Op_LoadI: return (bt == T_FLOAT);
1452 case Op_LoadL: return (bt == T_DOUBLE);
1453 case Op_LoadF: return (bt == T_INT);
1454 case Op_LoadD: return (bt == T_LONG);
1455
1456 default: return false;
1457 }
1458 }
1459
1460 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1461 BasicType bt = rt->basic_type();
1462 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1463 bool is_mismatched = is_mismatched_access();
1464 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1465 if (raw_type == nullptr) {
1466 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1467 }
1468 const int op = Opcode();
1469 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) ||
1470 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access());
1471 const Type* mem_t = gvn.type(in(MemNode::Address));
1472 if (mem_t == Type::TOP) {
1473 return gvn.C->top();
1474 }
1475 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1476 mem_t->is_ptr(), rt, bt, _mo, _control_dependency,
1477 require_atomic_access, is_unaligned_access(), is_mismatched);
1478 }
1479
1480 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1481 BasicType bt = vt->basic_type();
1482 switch (Opcode()) {
1483 case Op_StoreI: return (bt == T_FLOAT);
1484 case Op_StoreL: return (bt == T_DOUBLE);
1485 case Op_StoreF: return (bt == T_INT);
1486 case Op_StoreD: return (bt == T_LONG);
1487
1488 default: return false;
1489 }
1490 }
1491
1492 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1493 BasicType bt = vt->basic_type();
1494 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1495 const int op = Opcode();
1496 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) ||
1497 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access());
1498 const Type* mem_t = gvn.type(in(MemNode::Address));
1499 if (mem_t == Type::TOP) {
1500 return gvn.C->top();
1501 }
1502 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1503 mem_t->is_ptr(), val, bt, _mo, require_atomic_access);
1504
1505 bool is_mismatched = is_mismatched_access();
1506 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1507 if (raw_type == nullptr) {
1508 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1509 }
1510 if (is_mismatched) {
1511 st->set_mismatched_access();
1512 }
1513 return st;
1514 }
1515
1516 // We're loading from an object which has autobox behaviour.
1517 // If this object is result of a valueOf call we'll have a phi
1518 // merging a newly allocated object and a load from the cache.
1519 // We want to replace this load with the original incoming
1520 // argument to the valueOf call.
1521 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1522 assert(igvn->C->eliminate_boxing(), "sanity");
1523 intptr_t ignore = 0;
1524 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1525 if ((base == nullptr) || base->is_Phi()) {
1526 // Push the loads from the phi that comes from valueOf up
1527 // through it to allow elimination of the loads and the recovery
1528 // of the original value. It is done in split_through_phi().
1529 return nullptr;
1530 } else if (base->is_Load() ||
1531 (base->is_DecodeN() && base->in(1)->is_Load())) {
1532 // Eliminate the load of boxed value for integer types from the cache
1533 // array by deriving the value from the index into the array.
1534 // Capture the offset of the load and then reverse the computation.
1535
1536 // Get LoadN node which loads a boxing object from 'cache' array.
1537 if (base->is_DecodeN()) {
1538 base = base->in(1);
1539 }
1540 if (!base->in(Address)->is_AddP()) {
1541 return nullptr; // Complex address
1542 }
1543 AddPNode* address = base->in(Address)->as_AddP();
1544 Node* cache_base = address->in(AddPNode::Base);
1545 if ((cache_base != nullptr) && cache_base->is_DecodeN()) {
1546 // Get ConP node which is static 'cache' field.
1547 cache_base = cache_base->in(1);
1548 }
1549 if ((cache_base != nullptr) && cache_base->is_Con()) {
1550 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1551 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
1552 Node* elements[4];
1553 int shift = exact_log2(type2aelembytes(T_OBJECT));
1554 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1555 if (count > 0 && elements[0]->is_Con() &&
1556 (count == 1 ||
1557 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1558 elements[1]->in(2) == igvn->intcon(shift)))) {
1559 ciObjArray* array = base_type->const_oop()->as_obj_array();
1560 // Fetch the box object cache[0] at the base of the array and get its value
1561 ciInstance* box = array->obj_at(0)->as_instance();
1562 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1563 assert(ik->is_box_klass(), "sanity");
1564 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1565 if (ik->nof_nonstatic_fields() == 1) {
1566 // This should be true nonstatic_field_at requires calling
1567 // nof_nonstatic_fields so check it anyway
1568 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1569 BasicType bt = c.basic_type();
1570 // Only integer types have boxing cache.
1571 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1572 bt == T_BYTE || bt == T_SHORT ||
1573 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1574 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1575 if (cache_low != (int)cache_low) {
1576 return nullptr; // should not happen since cache is array indexed by value
1577 }
1578 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1579 if (offset != (int)offset) {
1580 return nullptr; // should not happen since cache is array indexed by value
1581 }
1582 // Add up all the offsets making of the address of the load
1583 Node* result = elements[0];
1584 for (int i = 1; i < count; i++) {
1585 result = igvn->transform(new AddXNode(result, elements[i]));
1586 }
1587 // Remove the constant offset from the address and then
1588 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1589 // remove the scaling of the offset to recover the original index.
1590 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1591 // Peel the shift off directly but wrap it in a dummy node
1592 // since Ideal can't return existing nodes
1593 igvn->_worklist.push(result); // remove dead node later
1594 result = new RShiftXNode(result->in(1), igvn->intcon(0));
1595 } else if (result->is_Add() && result->in(2)->is_Con() &&
1596 result->in(1)->Opcode() == Op_LShiftX &&
1597 result->in(1)->in(2) == igvn->intcon(shift)) {
1598 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1599 // but for boxing cache access we know that X<<Z will not overflow
1600 // (there is range check) so we do this optimizatrion by hand here.
1601 igvn->_worklist.push(result); // remove dead node later
1602 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1603 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1604 } else {
1605 result = new RShiftXNode(result, igvn->intcon(shift));
1606 }
1607 #ifdef _LP64
1608 if (bt != T_LONG) {
1609 result = new ConvL2INode(igvn->transform(result));
1610 }
1611 #else
1612 if (bt == T_LONG) {
1613 result = new ConvI2LNode(igvn->transform(result));
1614 }
1615 #endif
1616 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1617 // Need to preserve unboxing load type if it is unsigned.
1618 switch(this->Opcode()) {
1619 case Op_LoadUB:
1620 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1621 break;
1622 case Op_LoadUS:
1623 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1624 break;
1625 }
1626 return result;
1627 }
1628 }
1629 }
1630 }
1631 }
1632 return nullptr;
1633 }
1634
1635 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1636 Node* region = phi->in(0);
1637 if (region == nullptr) {
1638 return false; // Wait stable graph
1639 }
1640 uint cnt = phi->req();
1641 for (uint i = 1; i < cnt; i++) {
1642 Node* rc = region->in(i);
1643 if (rc == nullptr || phase->type(rc) == Type::TOP)
1644 return false; // Wait stable graph
1645 Node* in = phi->in(i);
1646 if (in == nullptr || phase->type(in) == Type::TOP)
1647 return false; // Wait stable graph
1648 }
1649 return true;
1650 }
1651
1652 //------------------------------split_through_phi------------------------------
1653 // Check whether a call to 'split_through_phi' would split this load through the
1654 // Phi *base*. This method is essentially a copy of the validations performed
1655 // by 'split_through_phi'. The first use of this method was in EA code as part
1656 // of simplification of allocation merges.
1657 // Some differences from original method (split_through_phi):
1658 // - If base->is_CastPP(): base = base->in(1)
1659 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) {
1660 Node* mem = in(Memory);
1661 Node* address = in(Address);
1662 intptr_t ignore = 0;
1663 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1664
1665 if (base->is_CastPP()) {
1666 base = base->in(1);
1667 }
1668
1669 if (req() > 3 || base == nullptr || !base->is_Phi()) {
1670 return false;
1671 }
1672
1673 if (!mem->is_Phi()) {
1674 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1675 return false;
1676 }
1677 } else if (base->in(0) != mem->in(0)) {
1678 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1679 return false;
1680 }
1681 }
1682
1683 return true;
1684 }
1685
1686 //------------------------------split_through_phi------------------------------
1687 // Split instance or boxed field load through Phi.
1688 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) {
1689 if (req() > 3) {
1690 assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs");
1691 // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account
1692 return nullptr;
1693 }
1694 Node* mem = in(Memory);
1695 Node* address = in(Address);
1696 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1697
1698 assert((t_oop != nullptr) &&
1699 (ignore_missing_instance_id ||
1700 t_oop->is_known_instance_field() ||
1701 t_oop->is_ptr_to_boxed_value()), "invalid conditions");
1702
1703 Compile* C = phase->C;
1704 intptr_t ignore = 0;
1705 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1706 bool base_is_phi = (base != nullptr) && base->is_Phi();
1707 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1708 (base != nullptr) && (base == address->in(AddPNode::Base)) &&
1709 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1710
1711 if (!((mem->is_Phi() || base_is_phi) &&
1712 (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) {
1713 return nullptr; // Neither memory or base are Phi
1714 }
1715
1716 if (mem->is_Phi()) {
1717 if (!stable_phi(mem->as_Phi(), phase)) {
1718 return nullptr; // Wait stable graph
1719 }
1720 uint cnt = mem->req();
1721 // Check for loop invariant memory.
1722 if (cnt == 3) {
1723 for (uint i = 1; i < cnt; i++) {
1724 Node* in = mem->in(i);
1725 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1726 if (m == mem) {
1727 if (i == 1) {
1728 // if the first edge was a loop, check second edge too.
1729 // If both are replaceable - we are in an infinite loop
1730 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1731 if (n == mem) {
1732 break;
1733 }
1734 }
1735 set_req(Memory, mem->in(cnt - i));
1736 return this; // made change
1737 }
1738 }
1739 }
1740 }
1741 if (base_is_phi) {
1742 if (!stable_phi(base->as_Phi(), phase)) {
1743 return nullptr; // Wait stable graph
1744 }
1745 uint cnt = base->req();
1746 // Check for loop invariant memory.
1747 if (cnt == 3) {
1748 for (uint i = 1; i < cnt; i++) {
1749 if (base->in(i) == base) {
1750 return nullptr; // Wait stable graph
1751 }
1752 }
1753 }
1754 }
1755
1756 // Split through Phi (see original code in loopopts.cpp).
1757 assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type");
1758
1759 // Do nothing here if Identity will find a value
1760 // (to avoid infinite chain of value phis generation).
1761 if (this != Identity(phase)) {
1762 return nullptr;
1763 }
1764
1765 // Select Region to split through.
1766 Node* region;
1767 DomResult dom_result = DomResult::Dominate;
1768 if (!base_is_phi) {
1769 assert(mem->is_Phi(), "sanity");
1770 region = mem->in(0);
1771 // Skip if the region dominates some control edge of the address.
1772 // We will check `dom_result` later.
1773 dom_result = MemNode::maybe_all_controls_dominate(address, region);
1774 } else if (!mem->is_Phi()) {
1775 assert(base_is_phi, "sanity");
1776 region = base->in(0);
1777 // Skip if the region dominates some control edge of the memory.
1778 // We will check `dom_result` later.
1779 dom_result = MemNode::maybe_all_controls_dominate(mem, region);
1780 } else if (base->in(0) != mem->in(0)) {
1781 assert(base_is_phi && mem->is_Phi(), "sanity");
1782 dom_result = MemNode::maybe_all_controls_dominate(mem, base->in(0));
1783 if (dom_result == DomResult::Dominate) {
1784 region = base->in(0);
1785 } else {
1786 dom_result = MemNode::maybe_all_controls_dominate(address, mem->in(0));
1787 if (dom_result == DomResult::Dominate) {
1788 region = mem->in(0);
1789 }
1790 // Otherwise we encountered a complex graph.
1791 }
1792 } else {
1793 assert(base->in(0) == mem->in(0), "sanity");
1794 region = mem->in(0);
1795 }
1796
1797 PhaseIterGVN* igvn = phase->is_IterGVN();
1798 if (dom_result != DomResult::Dominate) {
1799 if (dom_result == DomResult::EncounteredDeadCode) {
1800 // There is some dead code which eventually will be removed in IGVN.
1801 // Once this is the case, we get an unambiguous dominance result.
1802 // Push the node to the worklist again until the dead code is removed.
1803 igvn->_worklist.push(this);
1804 }
1805 return nullptr;
1806 }
1807
1808 Node* phi = nullptr;
1809 const Type* this_type = this->bottom_type();
1810 if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) {
1811 int this_index = C->get_alias_index(t_oop);
1812 int this_offset = t_oop->offset();
1813 int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx;
1814 phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset);
1815 } else if (ignore_missing_instance_id) {
1816 phi = new PhiNode(region, this_type, nullptr, mem->_idx);
1817 } else {
1818 return nullptr;
1819 }
1820
1821 for (uint i = 1; i < region->req(); i++) {
1822 Node* x;
1823 Node* the_clone = nullptr;
1824 Node* in = region->in(i);
1825 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1826 in != nullptr && in->is_OuterStripMinedLoop()) {
1827 // No node should go in the outer strip mined loop
1828 in = in->in(LoopNode::EntryControl);
1829 }
1830 if (in == nullptr || in == C->top()) {
1831 x = C->top(); // Dead path? Use a dead data op
1832 } else {
1833 x = this->clone(); // Else clone up the data op
1834 the_clone = x; // Remember for possible deletion.
1835 // Alter data node to use pre-phi inputs
1836 if (this->in(0) == region) {
1837 x->set_req(0, in);
1838 } else {
1839 x->set_req(0, nullptr);
1840 }
1841 if (mem->is_Phi() && (mem->in(0) == region)) {
1842 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1843 }
1844 if (address->is_Phi() && address->in(0) == region) {
1845 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1846 }
1847 if (base_is_phi && (base->in(0) == region)) {
1848 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1849 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1850 x->set_req(Address, adr_x);
1851 }
1852 }
1853 // Check for a 'win' on some paths
1854 const Type *t = x->Value(igvn);
1855
1856 bool singleton = t->singleton();
1857
1858 // See comments in PhaseIdealLoop::split_thru_phi().
1859 if (singleton && t == Type::TOP) {
1860 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1861 }
1862
1863 if (singleton) {
1864 x = igvn->makecon(t);
1865 } else {
1866 // We now call Identity to try to simplify the cloned node.
1867 // Note that some Identity methods call phase->type(this).
1868 // Make sure that the type array is big enough for
1869 // our new node, even though we may throw the node away.
1870 // (This tweaking with igvn only works because x is a new node.)
1871 igvn->set_type(x, t);
1872 // If x is a TypeNode, capture any more-precise type permanently into Node
1873 // otherwise it will be not updated during igvn->transform since
1874 // igvn->type(x) is set to x->Value() already.
1875 x->raise_bottom_type(t);
1876 Node* y = x->Identity(igvn);
1877 if (y != x) {
1878 x = y;
1879 } else {
1880 y = igvn->hash_find_insert(x);
1881 if (y) {
1882 x = y;
1883 } else {
1884 // Else x is a new node we are keeping
1885 // We do not need register_new_node_with_optimizer
1886 // because set_type has already been called.
1887 igvn->_worklist.push(x);
1888 }
1889 }
1890 }
1891 if (x != the_clone && the_clone != nullptr) {
1892 igvn->remove_dead_node(the_clone);
1893 }
1894 phi->set_req(i, x);
1895 }
1896 // Record Phi
1897 igvn->register_new_node_with_optimizer(phi);
1898 return phi;
1899 }
1900
1901 AllocateNode* LoadNode::is_new_object_mark_load() const {
1902 if (Opcode() == Op_LoadX) {
1903 Node* address = in(MemNode::Address);
1904 AllocateNode* alloc = AllocateNode::Ideal_allocation(address);
1905 Node* mem = in(MemNode::Memory);
1906 if (alloc != nullptr && mem->is_Proj() &&
1907 mem->in(0) != nullptr &&
1908 mem->in(0) == alloc->initialization() &&
1909 alloc->initialization()->proj_out_or_null(0) != nullptr) {
1910 return alloc;
1911 }
1912 }
1913 return nullptr;
1914 }
1915
1916
1917 //------------------------------Ideal------------------------------------------
1918 // If the load is from Field memory and the pointer is non-null, it might be possible to
1919 // zero out the control input.
1920 // If the offset is constant and the base is an object allocation,
1921 // try to hook me up to the exact initializing store.
1922 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1923 if (has_pinned_control_dependency()) {
1924 return nullptr;
1925 }
1926 Node* p = MemNode::Ideal_common(phase, can_reshape);
1927 if (p) return (p == NodeSentinel) ? nullptr : p;
1928
1929 Node* ctrl = in(MemNode::Control);
1930 Node* address = in(MemNode::Address);
1931
1932 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1933 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1934
1935 // Skip up past a SafePoint control. Cannot do this for Stores because
1936 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1937 if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint &&
1938 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1939 !addr_mark &&
1940 (depends_only_on_test() || has_unknown_control_dependency())) {
1941 ctrl = ctrl->in(0);
1942 set_req(MemNode::Control,ctrl);
1943 return this;
1944 }
1945
1946 intptr_t ignore = 0;
1947 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1948 if (base != nullptr
1949 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1950 // Check for useless control edge in some common special cases
1951 if (in(MemNode::Control) != nullptr
1952 && can_remove_control()
1953 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1954 && all_controls_dominate(base, phase->C->start())) {
1955 // A method-invariant, non-null address (constant or 'this' argument).
1956 set_req(MemNode::Control, nullptr);
1957 return this;
1958 }
1959 }
1960
1961 Node* mem = in(MemNode::Memory);
1962 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1963
1964 if (can_reshape && (addr_t != nullptr)) {
1965 // try to optimize our memory input
1966 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1967 if (opt_mem != mem) {
1968 set_req_X(MemNode::Memory, opt_mem, phase);
1969 if (phase->type( opt_mem ) == Type::TOP) return nullptr;
1970 return this;
1971 }
1972 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1973 if ((t_oop != nullptr) &&
1974 (t_oop->is_known_instance_field() ||
1975 t_oop->is_ptr_to_boxed_value())) {
1976 PhaseIterGVN *igvn = phase->is_IterGVN();
1977 assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true");
1978 if (igvn->_worklist.member(opt_mem)) {
1979 // Delay this transformation until memory Phi is processed.
1980 igvn->_worklist.push(this);
1981 return nullptr;
1982 }
1983 // Split instance field load through Phi.
1984 Node* result = split_through_phi(phase);
1985 if (result != nullptr) return result;
1986
1987 if (t_oop->is_ptr_to_boxed_value()) {
1988 Node* result = eliminate_autobox(igvn);
1989 if (result != nullptr) return result;
1990 }
1991 }
1992 }
1993
1994 // Is there a dominating load that loads the same value? Leave
1995 // anything that is not a load of a field/array element (like
1996 // barriers etc.) alone
1997 if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) {
1998 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1999 Node *use = mem->fast_out(i);
2000 if (use != this &&
2001 use->Opcode() == Opcode() &&
2002 use->in(0) != nullptr &&
2003 use->in(0) != in(0) &&
2004 use->in(Address) == in(Address)) {
2005 Node* ctl = in(0);
2006 for (int i = 0; i < 10 && ctl != nullptr; i++) {
2007 ctl = IfNode::up_one_dom(ctl);
2008 if (ctl == use->in(0)) {
2009 set_req(0, use->in(0));
2010 return this;
2011 }
2012 }
2013 }
2014 }
2015 }
2016
2017 // Check for prior store with a different base or offset; make Load
2018 // independent. Skip through any number of them. Bail out if the stores
2019 // are in an endless dead cycle and report no progress. This is a key
2020 // transform for Reflection. However, if after skipping through the Stores
2021 // we can't then fold up against a prior store do NOT do the transform as
2022 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
2023 // array memory alive twice: once for the hoisted Load and again after the
2024 // bypassed Store. This situation only works if EVERYBODY who does
2025 // anti-dependence work knows how to bypass. I.e. we need all
2026 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
2027 // the alias index stuff. So instead, peek through Stores and IFF we can
2028 // fold up, do so.
2029 Node* prev_mem = find_previous_store(phase);
2030 if (prev_mem != nullptr && prev_mem->is_top()) {
2031 // find_previous_store returns top when the access is dead
2032 return prev_mem;
2033 }
2034 if (prev_mem != nullptr) {
2035 Node* value = can_see_arraycopy_value(prev_mem, phase);
2036 if (value != nullptr) {
2037 return value;
2038 }
2039 }
2040 // Steps (a), (b): Walk past independent stores to find an exact match.
2041 if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) {
2042 // (c) See if we can fold up on the spot, but don't fold up here.
2043 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
2044 // just return a prior value, which is done by Identity calls.
2045 if (can_see_stored_value(prev_mem, phase)) {
2046 // Make ready for step (d):
2047 set_req_X(MemNode::Memory, prev_mem, phase);
2048 return this;
2049 }
2050 }
2051
2052 if (!can_reshape) {
2053 phase->record_for_igvn(this);
2054 }
2055
2056 return nullptr;
2057 }
2058
2059 // Helper to recognize certain Klass fields which are invariant across
2060 // some group of array types (e.g., int[] or all T[] where T < Object).
2061 const Type*
2062 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
2063 ciKlass* klass) const {
2064 assert(!UseCompactObjectHeaders || tkls->offset() != in_bytes(Klass::prototype_header_offset()),
2065 "must not happen");
2066
2067 if (tkls->isa_instklassptr() && tkls->offset() == in_bytes(InstanceKlass::access_flags_offset())) {
2068 // The field is InstanceKlass::_access_flags. Return its (constant) value.
2069 assert(Opcode() == Op_LoadUS, "must load an unsigned short from _access_flags");
2070 ciInstanceKlass* iklass = tkls->is_instklassptr()->instance_klass();
2071 return TypeInt::make(iklass->access_flags());
2072 }
2073 if (tkls->offset() == in_bytes(Klass::misc_flags_offset())) {
2074 // The field is Klass::_misc_flags. Return its (constant) value.
2075 assert(Opcode() == Op_LoadUB, "must load an unsigned byte from _misc_flags");
2076 return TypeInt::make(klass->misc_flags());
2077 }
2078 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
2079 // The field is Klass::_layout_helper. Return its constant value if known.
2080 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2081 return TypeInt::make(klass->layout_helper());
2082 }
2083
2084 // No match.
2085 return nullptr;
2086 }
2087
2088 //------------------------------Value-----------------------------------------
2089 const Type* LoadNode::Value(PhaseGVN* phase) const {
2090 // Either input is TOP ==> the result is TOP
2091 Node* mem = in(MemNode::Memory);
2092 const Type *t1 = phase->type(mem);
2093 if (t1 == Type::TOP) return Type::TOP;
2094 Node* adr = in(MemNode::Address);
2095 const TypePtr* tp = phase->type(adr)->isa_ptr();
2096 if (tp == nullptr || tp->empty()) return Type::TOP;
2097 int off = tp->offset();
2098 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
2099 Compile* C = phase->C;
2100
2101 // If load can see a previous constant store, use that.
2102 Node* value = can_see_stored_value(mem, phase);
2103 if (value != nullptr && value->is_Con()) {
2104 assert(value->bottom_type()->higher_equal(_type), "sanity");
2105 return value->bottom_type();
2106 }
2107
2108 // Try to guess loaded type from pointer type
2109 if (tp->isa_aryptr()) {
2110 const TypeAryPtr* ary = tp->is_aryptr();
2111 const Type* t = ary->elem();
2112
2113 // Determine whether the reference is beyond the header or not, by comparing
2114 // the offset against the offset of the start of the array's data.
2115 // Different array types begin at slightly different offsets (12 vs. 16).
2116 // We choose T_BYTE as an example base type that is least restrictive
2117 // as to alignment, which will therefore produce the smallest
2118 // possible base offset.
2119 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2120 const bool off_beyond_header = (off >= min_base_off);
2121
2122 // Try to constant-fold a stable array element.
2123 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
2124 // Make sure the reference is not into the header and the offset is constant
2125 ciObject* aobj = ary->const_oop();
2126 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
2127 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
2128 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
2129 stable_dimension,
2130 value_basic_type(), is_unsigned());
2131 if (con_type != nullptr) {
2132 return con_type;
2133 }
2134 }
2135 }
2136
2137 // Don't do this for integer types. There is only potential profit if
2138 // the element type t is lower than _type; that is, for int types, if _type is
2139 // more restrictive than t. This only happens here if one is short and the other
2140 // char (both 16 bits), and in those cases we've made an intentional decision
2141 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2142 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2143 //
2144 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2145 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
2146 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2147 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
2148 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2149 // In fact, that could have been the original type of p1, and p1 could have
2150 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2151 // expression (LShiftL quux 3) independently optimized to the constant 8.
2152 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2153 && (_type->isa_vect() == nullptr)
2154 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2155 // t might actually be lower than _type, if _type is a unique
2156 // concrete subclass of abstract class t.
2157 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
2158 const Type* jt = t->join_speculative(_type);
2159 // In any case, do not allow the join, per se, to empty out the type.
2160 if (jt->empty() && !t->empty()) {
2161 // This can happen if a interface-typed array narrows to a class type.
2162 jt = _type;
2163 }
2164 #ifdef ASSERT
2165 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2166 // The pointers in the autobox arrays are always non-null
2167 Node* base = adr->in(AddPNode::Base);
2168 if ((base != nullptr) && base->is_DecodeN()) {
2169 // Get LoadN node which loads IntegerCache.cache field
2170 base = base->in(1);
2171 }
2172 if ((base != nullptr) && base->is_Con()) {
2173 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2174 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2175 // It could be narrow oop
2176 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2177 }
2178 }
2179 }
2180 #endif
2181 return jt;
2182 }
2183 }
2184 } else if (tp->base() == Type::InstPtr) {
2185 assert( off != Type::OffsetBot ||
2186 // arrays can be cast to Objects
2187 !tp->isa_instptr() ||
2188 tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
2189 // unsafe field access may not have a constant offset
2190 C->has_unsafe_access(),
2191 "Field accesses must be precise" );
2192 // For oop loads, we expect the _type to be precise.
2193
2194 // Optimize loads from constant fields.
2195 const TypeInstPtr* tinst = tp->is_instptr();
2196 ciObject* const_oop = tinst->const_oop();
2197 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2198 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), value_basic_type());
2199 if (con_type != nullptr) {
2200 return con_type;
2201 }
2202 }
2203 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2204 assert(off != Type::OffsetBot ||
2205 !tp->isa_instklassptr() ||
2206 // arrays can be cast to Objects
2207 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2208 // also allow array-loading from the primary supertype
2209 // array during subtype checks
2210 Opcode() == Op_LoadKlass,
2211 "Field accesses must be precise");
2212 // For klass/static loads, we expect the _type to be precise
2213 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2214 /* With mirrors being an indirect in the Klass*
2215 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2216 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2217 *
2218 * So check the type and klass of the node before the LoadP.
2219 */
2220 Node* adr2 = adr->in(MemNode::Address);
2221 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2222 if (tkls != nullptr && !StressReflectiveCode) {
2223 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2224 ciKlass* klass = tkls->exact_klass();
2225 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2226 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2227 return TypeInstPtr::make(klass->java_mirror());
2228 }
2229 }
2230 }
2231
2232 const TypeKlassPtr *tkls = tp->isa_klassptr();
2233 if (tkls != nullptr) {
2234 if (tkls->is_loaded() && tkls->klass_is_exact()) {
2235 ciKlass* klass = tkls->exact_klass();
2236 // We are loading a field from a Klass metaobject whose identity
2237 // is known at compile time (the type is "exact" or "precise").
2238 // Check for fields we know are maintained as constants by the VM.
2239 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2240 // The field is Klass::_super_check_offset. Return its (constant) value.
2241 // (Folds up type checking code.)
2242 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2243 return TypeInt::make(klass->super_check_offset());
2244 }
2245 if (UseCompactObjectHeaders) {
2246 if (tkls->offset() == in_bytes(Klass::prototype_header_offset())) {
2247 // The field is Klass::_prototype_header. Return its (constant) value.
2248 assert(this->Opcode() == Op_LoadX, "must load a proper type from _prototype_header");
2249 return TypeX::make(klass->prototype_header());
2250 }
2251 }
2252 // Compute index into primary_supers array
2253 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2254 // Check for overflowing; use unsigned compare to handle the negative case.
2255 if( depth < ciKlass::primary_super_limit() ) {
2256 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2257 // (Folds up type checking code.)
2258 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2259 ciKlass *ss = klass->super_of_depth(depth);
2260 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2261 }
2262 const Type* aift = load_array_final_field(tkls, klass);
2263 if (aift != nullptr) return aift;
2264 }
2265
2266 // We can still check if we are loading from the primary_supers array at a
2267 // shallow enough depth. Even though the klass is not exact, entries less
2268 // than or equal to its super depth are correct.
2269 if (tkls->is_loaded()) {
2270 ciKlass* klass = nullptr;
2271 if (tkls->isa_instklassptr()) {
2272 klass = tkls->is_instklassptr()->instance_klass();
2273 } else {
2274 int dims;
2275 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims);
2276 if (inner->isa_instklassptr()) {
2277 klass = inner->is_instklassptr()->instance_klass();
2278 klass = ciObjArrayKlass::make(klass, dims);
2279 }
2280 }
2281 if (klass != nullptr) {
2282 // Compute index into primary_supers array
2283 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2284 // Check for overflowing; use unsigned compare to handle the negative case.
2285 if (depth < ciKlass::primary_super_limit() &&
2286 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case
2287 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2288 // (Folds up type checking code.)
2289 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2290 ciKlass *ss = klass->super_of_depth(depth);
2291 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2292 }
2293 }
2294 }
2295
2296 // If the type is enough to determine that the thing is not an array,
2297 // we can give the layout_helper a positive interval type.
2298 // This will help short-circuit some reflective code.
2299 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) &&
2300 tkls->isa_instklassptr() && // not directly typed as an array
2301 !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
2302 ) {
2303 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2304 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2305 // The key property of this type is that it folds up tests
2306 // for array-ness, since it proves that the layout_helper is positive.
2307 // Thus, a generic value like the basic object layout helper works fine.
2308 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2309 }
2310 }
2311
2312 // If we are loading from a freshly-allocated object/array, produce a zero.
2313 // Things to check:
2314 // 1. Load is beyond the header: headers are not guaranteed to be zero
2315 // 2. Load is not vectorized: vectors have no zero constant
2316 // 3. Load has no matching store, i.e. the input is the initial memory state
2317 const TypeOopPtr* tinst = tp->isa_oopptr();
2318 bool is_not_header = (tinst != nullptr) && tinst->is_known_instance_field();
2319 bool is_not_vect = (_type->isa_vect() == nullptr);
2320 if (is_not_header && is_not_vect) {
2321 Node* mem = in(MemNode::Memory);
2322 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2323 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2324 return Type::get_zero_type(_type->basic_type());
2325 }
2326 }
2327
2328 if (!UseCompactObjectHeaders) {
2329 Node* alloc = is_new_object_mark_load();
2330 if (alloc != nullptr) {
2331 return TypeX::make(markWord::prototype().value());
2332 }
2333 }
2334
2335 return _type;
2336 }
2337
2338 //------------------------------match_edge-------------------------------------
2339 // Do we Match on this edge index or not? Match only the address.
2340 uint LoadNode::match_edge(uint idx) const {
2341 return idx == MemNode::Address;
2342 }
2343
2344 //--------------------------LoadBNode::Ideal--------------------------------------
2345 //
2346 // If the previous store is to the same address as this load,
2347 // and the value stored was larger than a byte, replace this load
2348 // with the value stored truncated to a byte. If no truncation is
2349 // needed, the replacement is done in LoadNode::Identity().
2350 //
2351 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2352 Node* mem = in(MemNode::Memory);
2353 Node* value = can_see_stored_value(mem,phase);
2354 if (value != nullptr) {
2355 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2356 if (narrow != value) {
2357 return narrow;
2358 }
2359 }
2360 // Identity call will handle the case where truncation is not needed.
2361 return LoadNode::Ideal(phase, can_reshape);
2362 }
2363
2364 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2365 Node* mem = in(MemNode::Memory);
2366 Node* value = can_see_stored_value(mem,phase);
2367 if (value != nullptr && value->is_Con() &&
2368 !value->bottom_type()->higher_equal(_type)) {
2369 // If the input to the store does not fit with the load's result type,
2370 // it must be truncated. We can't delay until Ideal call since
2371 // a singleton Value is needed for split_thru_phi optimization.
2372 int con = value->get_int();
2373 return TypeInt::make((con << 24) >> 24);
2374 }
2375 return LoadNode::Value(phase);
2376 }
2377
2378 //--------------------------LoadUBNode::Ideal-------------------------------------
2379 //
2380 // If the previous store is to the same address as this load,
2381 // and the value stored was larger than a byte, replace this load
2382 // with the value stored truncated to a byte. If no truncation is
2383 // needed, the replacement is done in LoadNode::Identity().
2384 //
2385 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2386 Node* mem = in(MemNode::Memory);
2387 Node* value = can_see_stored_value(mem, phase);
2388 if (value != nullptr) {
2389 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2390 if (narrow != value) {
2391 return narrow;
2392 }
2393 }
2394 // Identity call will handle the case where truncation is not needed.
2395 return LoadNode::Ideal(phase, can_reshape);
2396 }
2397
2398 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2399 Node* mem = in(MemNode::Memory);
2400 Node* value = can_see_stored_value(mem,phase);
2401 if (value != nullptr && value->is_Con() &&
2402 !value->bottom_type()->higher_equal(_type)) {
2403 // If the input to the store does not fit with the load's result type,
2404 // it must be truncated. We can't delay until Ideal call since
2405 // a singleton Value is needed for split_thru_phi optimization.
2406 int con = value->get_int();
2407 return TypeInt::make(con & 0xFF);
2408 }
2409 return LoadNode::Value(phase);
2410 }
2411
2412 //--------------------------LoadUSNode::Ideal-------------------------------------
2413 //
2414 // If the previous store is to the same address as this load,
2415 // and the value stored was larger than a char, replace this load
2416 // with the value stored truncated to a char. If no truncation is
2417 // needed, the replacement is done in LoadNode::Identity().
2418 //
2419 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2420 Node* mem = in(MemNode::Memory);
2421 Node* value = can_see_stored_value(mem,phase);
2422 if (value != nullptr) {
2423 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2424 if (narrow != value) {
2425 return narrow;
2426 }
2427 }
2428 // Identity call will handle the case where truncation is not needed.
2429 return LoadNode::Ideal(phase, can_reshape);
2430 }
2431
2432 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2433 Node* mem = in(MemNode::Memory);
2434 Node* value = can_see_stored_value(mem,phase);
2435 if (value != nullptr && value->is_Con() &&
2436 !value->bottom_type()->higher_equal(_type)) {
2437 // If the input to the store does not fit with the load's result type,
2438 // it must be truncated. We can't delay until Ideal call since
2439 // a singleton Value is needed for split_thru_phi optimization.
2440 int con = value->get_int();
2441 return TypeInt::make(con & 0xFFFF);
2442 }
2443 return LoadNode::Value(phase);
2444 }
2445
2446 //--------------------------LoadSNode::Ideal--------------------------------------
2447 //
2448 // If the previous store is to the same address as this load,
2449 // and the value stored was larger than a short, replace this load
2450 // with the value stored truncated to a short. If no truncation is
2451 // needed, the replacement is done in LoadNode::Identity().
2452 //
2453 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2454 Node* mem = in(MemNode::Memory);
2455 Node* value = can_see_stored_value(mem,phase);
2456 if (value != nullptr) {
2457 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2458 if (narrow != value) {
2459 return narrow;
2460 }
2461 }
2462 // Identity call will handle the case where truncation is not needed.
2463 return LoadNode::Ideal(phase, can_reshape);
2464 }
2465
2466 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2467 Node* mem = in(MemNode::Memory);
2468 Node* value = can_see_stored_value(mem,phase);
2469 if (value != nullptr && value->is_Con() &&
2470 !value->bottom_type()->higher_equal(_type)) {
2471 // If the input to the store does not fit with the load's result type,
2472 // it must be truncated. We can't delay until Ideal call since
2473 // a singleton Value is needed for split_thru_phi optimization.
2474 int con = value->get_int();
2475 return TypeInt::make((con << 16) >> 16);
2476 }
2477 return LoadNode::Value(phase);
2478 }
2479
2480 //=============================================================================
2481 //----------------------------LoadKlassNode::make------------------------------
2482 // Polymorphic factory method:
2483 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2484 // sanity check the alias category against the created node type
2485 const TypePtr* adr_type = adr->bottom_type()->isa_ptr();
2486 assert(adr_type != nullptr, "expecting TypeKlassPtr");
2487 #ifdef _LP64
2488 if (adr_type->is_ptr_to_narrowklass()) {
2489 assert(UseCompressedClassPointers, "no compressed klasses");
2490 Node* load_klass = gvn.transform(new LoadNKlassNode(mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2491 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2492 }
2493 #endif
2494 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2495 return new LoadKlassNode(mem, adr, at, tk, MemNode::unordered);
2496 }
2497
2498 //------------------------------Value------------------------------------------
2499 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2500 return klass_value_common(phase);
2501 }
2502
2503 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2504 // Either input is TOP ==> the result is TOP
2505 const Type *t1 = phase->type( in(MemNode::Memory) );
2506 if (t1 == Type::TOP) return Type::TOP;
2507 Node *adr = in(MemNode::Address);
2508 const Type *t2 = phase->type( adr );
2509 if (t2 == Type::TOP) return Type::TOP;
2510 const TypePtr *tp = t2->is_ptr();
2511 if (TypePtr::above_centerline(tp->ptr()) ||
2512 tp->ptr() == TypePtr::Null) return Type::TOP;
2513
2514 // Return a more precise klass, if possible
2515 const TypeInstPtr *tinst = tp->isa_instptr();
2516 if (tinst != nullptr) {
2517 ciInstanceKlass* ik = tinst->instance_klass();
2518 int offset = tinst->offset();
2519 if (ik == phase->C->env()->Class_klass()
2520 && (offset == java_lang_Class::klass_offset() ||
2521 offset == java_lang_Class::array_klass_offset())) {
2522 // We are loading a special hidden field from a Class mirror object,
2523 // the field which points to the VM's Klass metaobject.
2524 ciType* t = tinst->java_mirror_type();
2525 // java_mirror_type returns non-null for compile-time Class constants.
2526 if (t != nullptr) {
2527 // constant oop => constant klass
2528 if (offset == java_lang_Class::array_klass_offset()) {
2529 if (t->is_void()) {
2530 // We cannot create a void array. Since void is a primitive type return null
2531 // klass. Users of this result need to do a null check on the returned klass.
2532 return TypePtr::NULL_PTR;
2533 }
2534 return TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);
2535 }
2536 if (!t->is_klass()) {
2537 // a primitive Class (e.g., int.class) has null for a klass field
2538 return TypePtr::NULL_PTR;
2539 }
2540 // Fold up the load of the hidden field
2541 return TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);
2542 }
2543 // non-constant mirror, so we can't tell what's going on
2544 }
2545 if (!tinst->is_loaded())
2546 return _type; // Bail out if not loaded
2547 if (offset == oopDesc::klass_offset_in_bytes()) {
2548 return tinst->as_klass_type(true);
2549 }
2550 }
2551
2552 // Check for loading klass from an array
2553 const TypeAryPtr *tary = tp->isa_aryptr();
2554 if (tary != nullptr &&
2555 tary->offset() == oopDesc::klass_offset_in_bytes()) {
2556 return tary->as_klass_type(true);
2557 }
2558
2559 // Check for loading klass from an array klass
2560 const TypeKlassPtr *tkls = tp->isa_klassptr();
2561 if (tkls != nullptr && !StressReflectiveCode) {
2562 if (!tkls->is_loaded())
2563 return _type; // Bail out if not loaded
2564 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2565 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2566 // // Always returning precise element type is incorrect,
2567 // // e.g., element type could be object and array may contain strings
2568 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2569
2570 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2571 // according to the element type's subclassing.
2572 return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2573 }
2574 if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() &&
2575 tkls->offset() == in_bytes(Klass::super_offset())) {
2576 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super();
2577 // The field is Klass::_super. Return its (constant) value.
2578 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2579 return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR;
2580 }
2581 }
2582
2583 if (tkls != nullptr && !UseSecondarySupersCache
2584 && tkls->offset() == in_bytes(Klass::secondary_super_cache_offset())) {
2585 // Treat Klass::_secondary_super_cache as a constant when the cache is disabled.
2586 return TypePtr::NULL_PTR;
2587 }
2588
2589 // Bailout case
2590 return LoadNode::Value(phase);
2591 }
2592
2593 //------------------------------Identity---------------------------------------
2594 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2595 // Also feed through the klass in Allocate(...klass...)._klass.
2596 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2597 return klass_identity_common(phase);
2598 }
2599
2600 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2601 Node* x = LoadNode::Identity(phase);
2602 if (x != this) return x;
2603
2604 // Take apart the address into an oop and offset.
2605 // Return 'this' if we cannot.
2606 Node* adr = in(MemNode::Address);
2607 intptr_t offset = 0;
2608 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2609 if (base == nullptr) return this;
2610 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2611 if (toop == nullptr) return this;
2612
2613 // Step over potential GC barrier for OopHandle resolve
2614 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2615 if (bs->is_gc_barrier_node(base)) {
2616 base = bs->step_over_gc_barrier(base);
2617 }
2618
2619 // We can fetch the klass directly through an AllocateNode.
2620 // This works even if the klass is not constant (clone or newArray).
2621 if (offset == oopDesc::klass_offset_in_bytes()) {
2622 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2623 if (allocated_klass != nullptr) {
2624 return allocated_klass;
2625 }
2626 }
2627
2628 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2629 // See inline_native_Class_query for occurrences of these patterns.
2630 // Java Example: x.getClass().isAssignableFrom(y)
2631 //
2632 // This improves reflective code, often making the Class
2633 // mirror go completely dead. (Current exception: Class
2634 // mirrors may appear in debug info, but we could clean them out by
2635 // introducing a new debug info operator for Klass.java_mirror).
2636
2637 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass()
2638 && offset == java_lang_Class::klass_offset()) {
2639 if (base->is_Load()) {
2640 Node* base2 = base->in(MemNode::Address);
2641 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2642 Node* adr2 = base2->in(MemNode::Address);
2643 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2644 if (tkls != nullptr && !tkls->empty()
2645 && (tkls->isa_instklassptr() || tkls->isa_aryklassptr())
2646 && adr2->is_AddP()
2647 ) {
2648 int mirror_field = in_bytes(Klass::java_mirror_offset());
2649 if (tkls->offset() == mirror_field) {
2650 #ifdef ASSERT
2651 const TypeKlassPtr* tkls2 = phase->type(adr2->in(AddPNode::Address))->is_klassptr();
2652 assert(tkls2->offset() == 0, "not a load of java_mirror");
2653 #endif
2654 assert(adr2->in(AddPNode::Base)->is_top(), "not an off heap load");
2655 assert(adr2->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(Klass::java_mirror_offset()), "incorrect offset");
2656 return adr2->in(AddPNode::Address);
2657 }
2658 }
2659 }
2660 }
2661 }
2662
2663 return this;
2664 }
2665
2666 LoadNode* LoadNode::clone_pinned() const {
2667 LoadNode* ld = clone()->as_Load();
2668 ld->_control_dependency = UnknownControl;
2669 return ld;
2670 }
2671
2672 // Pin a LoadNode if it carries a dependency on its control input. There are cases when the node
2673 // does not actually have any dependency on its control input. For example, if we have a LoadNode
2674 // being used only outside a loop but it must be scheduled inside the loop, we can clone the node
2675 // for each of its use so that all the clones can be scheduled outside the loop. Then, to prevent
2676 // the clones from being GVN-ed again, we add a control input for each of them at the loop exit. In
2677 // those case, since there is not a dependency between the node and its control input, we do not
2678 // need to pin it.
2679 LoadNode* LoadNode::pin_node_under_control_impl() const {
2680 const TypePtr* adr_type = this->adr_type();
2681 if (adr_type != nullptr && adr_type->isa_aryptr()) {
2682 // Only array accesses have dependencies on their control input
2683 return clone_pinned();
2684 }
2685 return nullptr;
2686 }
2687
2688 //------------------------------Value------------------------------------------
2689 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2690 const Type *t = klass_value_common(phase);
2691 if (t == Type::TOP)
2692 return t;
2693
2694 return t->make_narrowklass();
2695 }
2696
2697 //------------------------------Identity---------------------------------------
2698 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2699 // Also feed through the klass in Allocate(...klass...)._klass.
2700 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2701 Node *x = klass_identity_common(phase);
2702
2703 const Type *t = phase->type( x );
2704 if( t == Type::TOP ) return x;
2705 if( t->isa_narrowklass()) return x;
2706 assert (!t->isa_narrowoop(), "no narrow oop here");
2707
2708 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2709 }
2710
2711 //------------------------------Value-----------------------------------------
2712 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2713 // Either input is TOP ==> the result is TOP
2714 const Type *t1 = phase->type( in(MemNode::Memory) );
2715 if( t1 == Type::TOP ) return Type::TOP;
2716 Node *adr = in(MemNode::Address);
2717 const Type *t2 = phase->type( adr );
2718 if( t2 == Type::TOP ) return Type::TOP;
2719 const TypePtr *tp = t2->is_ptr();
2720 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2721 const TypeAryPtr *tap = tp->isa_aryptr();
2722 if( !tap ) return _type;
2723 return tap->size();
2724 }
2725
2726 //-------------------------------Ideal---------------------------------------
2727 // Feed through the length in AllocateArray(...length...)._length.
2728 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2729 Node* p = MemNode::Ideal_common(phase, can_reshape);
2730 if (p) return (p == NodeSentinel) ? nullptr : p;
2731
2732 // Take apart the address into an oop and offset.
2733 // Return 'this' if we cannot.
2734 Node* adr = in(MemNode::Address);
2735 intptr_t offset = 0;
2736 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2737 if (base == nullptr) return nullptr;
2738 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2739 if (tary == nullptr) return nullptr;
2740
2741 // We can fetch the length directly through an AllocateArrayNode.
2742 // This works even if the length is not constant (clone or newArray).
2743 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2744 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2745 if (alloc != nullptr) {
2746 Node* allocated_length = alloc->Ideal_length();
2747 Node* len = alloc->make_ideal_length(tary, phase);
2748 if (allocated_length != len) {
2749 // New CastII improves on this.
2750 return len;
2751 }
2752 }
2753 }
2754
2755 return nullptr;
2756 }
2757
2758 //------------------------------Identity---------------------------------------
2759 // Feed through the length in AllocateArray(...length...)._length.
2760 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2761 Node* x = LoadINode::Identity(phase);
2762 if (x != this) return x;
2763
2764 // Take apart the address into an oop and offset.
2765 // Return 'this' if we cannot.
2766 Node* adr = in(MemNode::Address);
2767 intptr_t offset = 0;
2768 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2769 if (base == nullptr) return this;
2770 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2771 if (tary == nullptr) return this;
2772
2773 // We can fetch the length directly through an AllocateArrayNode.
2774 // This works even if the length is not constant (clone or newArray).
2775 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2776 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2777 if (alloc != nullptr) {
2778 Node* allocated_length = alloc->Ideal_length();
2779 // Do not allow make_ideal_length to allocate a CastII node.
2780 Node* len = alloc->make_ideal_length(tary, phase, false);
2781 if (allocated_length == len) {
2782 // Return allocated_length only if it would not be improved by a CastII.
2783 return allocated_length;
2784 }
2785 }
2786 }
2787
2788 return this;
2789
2790 }
2791
2792 //=============================================================================
2793 //---------------------------StoreNode::make-----------------------------------
2794 // Polymorphic factory method:
2795 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) {
2796 assert((mo == unordered || mo == release), "unexpected");
2797 Compile* C = gvn.C;
2798 assert(adr_type == nullptr || adr->is_top() || C->get_alias_index(gvn.type(adr)->is_ptr()) == C->get_alias_index(adr_type), "adr and adr_type must agree");
2799 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2800 ctl != nullptr, "raw memory operations should have control edge");
2801
2802 switch (bt) {
2803 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2804 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2805 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2806 case T_CHAR:
2807 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2808 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2809 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2810 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2811 case T_METADATA:
2812 case T_ADDRESS:
2813 case T_OBJECT:
2814 #ifdef _LP64
2815 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2816 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2817 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2818 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2819 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2820 adr->bottom_type()->isa_rawptr())) {
2821 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2822 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2823 }
2824 #endif
2825 {
2826 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2827 }
2828 default:
2829 ShouldNotReachHere();
2830 return (StoreNode*)nullptr;
2831 }
2832 }
2833
2834 //--------------------------bottom_type----------------------------------------
2835 const Type *StoreNode::bottom_type() const {
2836 return Type::MEMORY;
2837 }
2838
2839 //------------------------------hash-------------------------------------------
2840 uint StoreNode::hash() const {
2841 // unroll addition of interesting fields
2842 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2843
2844 // Since they are not commoned, do not hash them:
2845 return NO_HASH;
2846 }
2847
2848 // Link together multiple stores (B/S/C/I) into a longer one.
2849 //
2850 // Example: _store = StoreB[i+3]
2851 //
2852 // RangeCheck[i+0] RangeCheck[i+0]
2853 // StoreB[i+0]
2854 // RangeCheck[i+3] RangeCheck[i+3]
2855 // StoreB[i+1] --> pass: fail:
2856 // StoreB[i+2] StoreI[i+0] StoreB[i+0]
2857 // StoreB[i+3]
2858 //
2859 // The 4 StoreB are merged into a single StoreI node. We have to be careful with RangeCheck[i+1]: before
2860 // the optimization, if this RangeCheck[i+1] fails, then we execute only StoreB[i+0], and then trap. After
2861 // the optimization, the new StoreI[i+0] is on the passing path of RangeCheck[i+3], and StoreB[i+0] on the
2862 // failing path.
2863 //
2864 // Note: For normal array stores, every store at first has a RangeCheck. But they can be removed with:
2865 // - RCE (RangeCheck Elimination): the RangeChecks in the loop are hoisted out and before the loop,
2866 // and possibly no RangeChecks remain between the stores.
2867 // - RangeCheck smearing: the earlier RangeChecks are adjusted such that they cover later RangeChecks,
2868 // and those later RangeChecks can be removed. Example:
2869 //
2870 // RangeCheck[i+0] RangeCheck[i+0] <- before first store
2871 // StoreB[i+0] StoreB[i+0] <- first store
2872 // RangeCheck[i+1] --> smeared --> RangeCheck[i+3] <- only RC between first and last store
2873 // StoreB[i+1] StoreB[i+1] <- second store
2874 // RangeCheck[i+2] --> removed
2875 // StoreB[i+2] StoreB[i+2]
2876 // RangeCheck[i+3] --> removed
2877 // StoreB[i+3] StoreB[i+3] <- last store
2878 //
2879 // Thus, it is a common pattern that between the first and last store in a chain
2880 // of adjacent stores there remains exactly one RangeCheck, located between the
2881 // first and the second store (e.g. RangeCheck[i+3]).
2882 //
2883 class MergePrimitiveStores : public StackObj {
2884 private:
2885 PhaseGVN* const _phase;
2886 StoreNode* const _store;
2887 // State machine with initial state Unknown
2888 // Allowed transitions:
2889 // Unknown -> Const
2890 // Unknown -> Platform
2891 // Unknown -> Reverse
2892 // Unknown -> NotAdjacent
2893 // Const -> Const
2894 // Const -> NotAdjacent
2895 // Platform -> Platform
2896 // Platform -> NotAdjacent
2897 // Reverse -> Reverse
2898 // Reverse -> NotAdjacent
2899 // NotAdjacent -> NotAdjacent
2900 enum ValueOrder : uint8_t {
2901 Unknown, // Initial state
2902 Const, // Input values are const
2903 Platform, // Platform order
2904 Reverse, // Reverse platform order
2905 NotAdjacent // Not adjacent
2906 };
2907 ValueOrder _value_order;
2908
2909 NOT_PRODUCT( const CHeapBitMap &_trace_tags; )
2910
2911 public:
2912 MergePrimitiveStores(PhaseGVN* phase, StoreNode* store) :
2913 _phase(phase), _store(store), _value_order(ValueOrder::Unknown)
2914 NOT_PRODUCT( COMMA _trace_tags(Compile::current()->directive()->trace_merge_stores_tags()) )
2915 {}
2916
2917 StoreNode* run();
2918
2919 private:
2920 bool is_compatible_store(const StoreNode* other_store) const;
2921 bool is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const;
2922 bool is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const;
2923 static bool is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase);
2924 enum CFGStatus { SuccessNoRangeCheck, SuccessWithRangeCheck, Failure };
2925 static CFGStatus cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store);
2926
2927 class Status {
2928 private:
2929 StoreNode* _found_store;
2930 bool _found_range_check;
2931
2932 Status(StoreNode* found_store, bool found_range_check)
2933 : _found_store(found_store), _found_range_check(found_range_check) {}
2934
2935 public:
2936 StoreNode* found_store() const { return _found_store; }
2937 bool found_range_check() const { return _found_range_check; }
2938 static Status make_failure() { return Status(nullptr, false); }
2939
2940 static Status make(StoreNode* found_store, const CFGStatus cfg_status) {
2941 if (cfg_status == CFGStatus::Failure) {
2942 return Status::make_failure();
2943 }
2944 return Status(found_store, cfg_status == CFGStatus::SuccessWithRangeCheck);
2945 }
2946
2947 #ifndef PRODUCT
2948 void print_on(outputStream* st) const {
2949 if (_found_store == nullptr) {
2950 st->print_cr("None");
2951 } else {
2952 st->print_cr("Found[%d %s, %s]", _found_store->_idx, _found_store->Name(),
2953 _found_range_check ? "RC" : "no-RC");
2954 }
2955 }
2956 #endif
2957 };
2958
2959 enum ValueOrder find_adjacent_input_value_order(const Node* n1, const Node* n2, const int memory_size) const;
2960 Status find_adjacent_use_store(const StoreNode* def_store) const;
2961 Status find_adjacent_def_store(const StoreNode* use_store) const;
2962 Status find_use_store(const StoreNode* def_store) const;
2963 Status find_def_store(const StoreNode* use_store) const;
2964 Status find_use_store_unidirectional(const StoreNode* def_store) const;
2965 Status find_def_store_unidirectional(const StoreNode* use_store) const;
2966
2967 void collect_merge_list(Node_List& merge_list) const;
2968 Node* make_merged_input_value(const Node_List& merge_list);
2969 StoreNode* make_merged_store(const Node_List& merge_list, Node* merged_input_value);
2970
2971 #ifndef PRODUCT
2972 // Access to TraceMergeStores tags
2973 bool is_trace(TraceMergeStores::Tag tag) const {
2974 return _trace_tags.at(tag);
2975 }
2976
2977 bool is_trace_basic() const {
2978 return is_trace(TraceMergeStores::Tag::BASIC);
2979 }
2980
2981 bool is_trace_pointer_parsing() const {
2982 return is_trace(TraceMergeStores::Tag::POINTER_PARSING);
2983 }
2984
2985 bool is_trace_pointer_aliasing() const {
2986 return is_trace(TraceMergeStores::Tag::POINTER_ALIASING);
2987 }
2988
2989 bool is_trace_pointer_adjacency() const {
2990 return is_trace(TraceMergeStores::Tag::POINTER_ADJACENCY);
2991 }
2992
2993 bool is_trace_success() const {
2994 return is_trace(TraceMergeStores::Tag::SUCCESS);
2995 }
2996 #endif
2997
2998 NOT_PRODUCT( void trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const; )
2999 };
3000
3001 StoreNode* MergePrimitiveStores::run() {
3002 // Check for B/S/C/I
3003 int opc = _store->Opcode();
3004 if (opc != Op_StoreB && opc != Op_StoreC && opc != Op_StoreI) {
3005 return nullptr;
3006 }
3007
3008 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] MergePrimitiveStores::run: "); _store->dump(); })
3009
3010 // The _store must be the "last" store in a chain. If we find a use we could merge with
3011 // then that use or a store further down is the "last" store.
3012 Status status_use = find_adjacent_use_store(_store);
3013 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect no use: "); status_use.print_on(tty); })
3014 if (status_use.found_store() != nullptr) {
3015 return nullptr;
3016 }
3017
3018 // Check if we can merge with at least one def, so that we have at least 2 stores to merge.
3019 Status status_def = find_adjacent_def_store(_store);
3020 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect def: "); status_def.print_on(tty); })
3021 Node* def_store = status_def.found_store();
3022 if (def_store == nullptr) {
3023 return nullptr;
3024 }
3025
3026 // Initialize value order
3027 _value_order = find_adjacent_input_value_order(def_store->in(MemNode::ValueIn),
3028 _store->in(MemNode::ValueIn),
3029 _store->memory_size());
3030 assert(_value_order != ValueOrder::NotAdjacent && _value_order != ValueOrder::Unknown, "Order should be checked");
3031
3032 ResourceMark rm;
3033 Node_List merge_list;
3034 collect_merge_list(merge_list);
3035
3036 Node* merged_input_value = make_merged_input_value(merge_list);
3037 if (merged_input_value == nullptr) { return nullptr; }
3038
3039 StoreNode* merged_store = make_merged_store(merge_list, merged_input_value);
3040
3041 NOT_PRODUCT( if (is_trace_success()) { trace(merge_list, merged_input_value, merged_store); } )
3042
3043 return merged_store;
3044 }
3045
3046 // Check compatibility between _store and other_store.
3047 bool MergePrimitiveStores::is_compatible_store(const StoreNode* other_store) const {
3048 int opc = _store->Opcode();
3049 assert(opc == Op_StoreB || opc == Op_StoreC || opc == Op_StoreI, "precondition");
3050
3051 if (other_store == nullptr ||
3052 _store->Opcode() != other_store->Opcode()) {
3053 return false;
3054 }
3055
3056 return true;
3057 }
3058
3059 bool MergePrimitiveStores::is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const {
3060 if (!is_adjacent_input_pair(def_store->in(MemNode::ValueIn),
3061 use_store->in(MemNode::ValueIn),
3062 def_store->memory_size())) {
3063 return false;
3064 }
3065
3066 ResourceMark rm;
3067 #ifndef PRODUCT
3068 const TraceMemPointer trace(is_trace_pointer_parsing(),
3069 is_trace_pointer_aliasing(),
3070 is_trace_pointer_adjacency(),
3071 true);
3072 #endif
3073 const MemPointer pointer_use(use_store NOT_PRODUCT(COMMA trace));
3074 const MemPointer pointer_def(def_store NOT_PRODUCT(COMMA trace));
3075 return pointer_def.is_adjacent_to_and_before(pointer_use);
3076 }
3077
3078 // Check input values n1 and n2 can be merged and return the value order
3079 MergePrimitiveStores::ValueOrder MergePrimitiveStores::find_adjacent_input_value_order(const Node* n1, const Node* n2,
3080 const int memory_size) const {
3081 // Pattern: [n1 = ConI, n2 = ConI]
3082 if (n1->Opcode() == Op_ConI && n2->Opcode() == Op_ConI) {
3083 return ValueOrder::Const;
3084 }
3085
3086 Node const *base_n2;
3087 jint shift_n2;
3088 if (!is_con_RShift(n2, base_n2, shift_n2, _phase)) {
3089 return ValueOrder::NotAdjacent;
3090 }
3091 Node const *base_n1;
3092 jint shift_n1;
3093 if (!is_con_RShift(n1, base_n1, shift_n1, _phase)) {
3094 return ValueOrder::NotAdjacent;
3095 }
3096
3097 int bits_per_store = memory_size * 8;
3098 if (base_n1 != base_n2 ||
3099 abs(shift_n1 - shift_n2) != bits_per_store ||
3100 shift_n1 % bits_per_store != 0) {
3101 // Values are not adjacent
3102 return ValueOrder::NotAdjacent;
3103 }
3104
3105 // Detect value order
3106 #ifdef VM_LITTLE_ENDIAN
3107 return shift_n1 < shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3108 : ValueOrder::Reverse; // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3109 #else
3110 return shift_n1 > shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3111 : ValueOrder::Reverse; // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3112 #endif
3113 }
3114
3115 bool MergePrimitiveStores::is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const {
3116 ValueOrder input_value_order = find_adjacent_input_value_order(n1, n2, memory_size);
3117
3118 switch (input_value_order) {
3119 case ValueOrder::NotAdjacent:
3120 return false;
3121 case ValueOrder::Reverse:
3122 if (memory_size != 1 ||
3123 !Matcher::match_rule_supported(Op_ReverseBytesS) ||
3124 !Matcher::match_rule_supported(Op_ReverseBytesI) ||
3125 !Matcher::match_rule_supported(Op_ReverseBytesL)) {
3126 // ReverseBytes are not supported by platform
3127 return false;
3128 }
3129 // fall-through.
3130 case ValueOrder::Const:
3131 case ValueOrder::Platform:
3132 if (_value_order == ValueOrder::Unknown) {
3133 // Initial state is Unknown, and we find a valid input value order
3134 return true;
3135 }
3136 // The value order can not be changed
3137 return _value_order == input_value_order;
3138 case ValueOrder::Unknown:
3139 default:
3140 ShouldNotReachHere();
3141 }
3142 return false;
3143 }
3144
3145 // Detect pattern: n = base_out >> shift_out
3146 bool MergePrimitiveStores::is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase) {
3147 assert(n != nullptr, "precondition");
3148
3149 int opc = n->Opcode();
3150 if (opc == Op_ConvL2I) {
3151 n = n->in(1);
3152 opc = n->Opcode();
3153 }
3154
3155 if ((opc == Op_RShiftI ||
3156 opc == Op_RShiftL ||
3157 opc == Op_URShiftI ||
3158 opc == Op_URShiftL) &&
3159 n->in(2)->is_ConI()) {
3160 base_out = n->in(1);
3161 shift_out = n->in(2)->get_int();
3162 // The shift must be positive:
3163 return shift_out >= 0;
3164 }
3165
3166 if (phase->type(n)->isa_int() != nullptr ||
3167 phase->type(n)->isa_long() != nullptr) {
3168 // (base >> 0)
3169 base_out = n;
3170 shift_out = 0;
3171 return true;
3172 }
3173 return false;
3174 }
3175
3176 // Check if there is nothing between the two stores, except optionally a RangeCheck leading to an uncommon trap.
3177 MergePrimitiveStores::CFGStatus MergePrimitiveStores::cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store) {
3178 assert(use_store->in(MemNode::Memory) == def_store, "use-def relationship");
3179
3180 Node* ctrl_use = use_store->in(MemNode::Control);
3181 Node* ctrl_def = def_store->in(MemNode::Control);
3182 if (ctrl_use == nullptr || ctrl_def == nullptr) {
3183 return CFGStatus::Failure;
3184 }
3185
3186 if (ctrl_use == ctrl_def) {
3187 // Same ctrl -> no RangeCheck in between.
3188 // Check: use_store must be the only use of def_store.
3189 if (def_store->outcnt() > 1) {
3190 return CFGStatus::Failure;
3191 }
3192 return CFGStatus::SuccessNoRangeCheck;
3193 }
3194
3195 // Different ctrl -> could have RangeCheck in between.
3196 // Check: 1. def_store only has these uses: use_store and MergeMem for uncommon trap, and
3197 // 2. ctrl separated by RangeCheck.
3198 if (def_store->outcnt() != 2) {
3199 return CFGStatus::Failure; // Cannot have exactly these uses: use_store and MergeMem for uncommon trap.
3200 }
3201 int use_store_out_idx = def_store->raw_out(0) == use_store ? 0 : 1;
3202 Node* merge_mem = def_store->raw_out(1 - use_store_out_idx)->isa_MergeMem();
3203 if (merge_mem == nullptr ||
3204 merge_mem->outcnt() != 1) {
3205 return CFGStatus::Failure; // Does not have MergeMem for uncommon trap.
3206 }
3207 if (!ctrl_use->is_IfProj() ||
3208 !ctrl_use->in(0)->is_RangeCheck() ||
3209 ctrl_use->in(0)->outcnt() != 2) {
3210 return CFGStatus::Failure; // Not RangeCheck.
3211 }
3212 IfProjNode* other_proj = ctrl_use->as_IfProj()->other_if_proj();
3213 Node* trap = other_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check);
3214 if (trap != merge_mem->unique_out() ||
3215 ctrl_use->in(0)->in(0) != ctrl_def) {
3216 return CFGStatus::Failure; // Not RangeCheck with merge_mem leading to uncommon trap.
3217 }
3218
3219 return CFGStatus::SuccessWithRangeCheck;
3220 }
3221
3222 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_use_store(const StoreNode* def_store) const {
3223 Status status_use = find_use_store(def_store);
3224 StoreNode* use_store = status_use.found_store();
3225 if (use_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3226 return Status::make_failure();
3227 }
3228 return status_use;
3229 }
3230
3231 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_def_store(const StoreNode* use_store) const {
3232 Status status_def = find_def_store(use_store);
3233 StoreNode* def_store = status_def.found_store();
3234 if (def_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3235 return Status::make_failure();
3236 }
3237 return status_def;
3238 }
3239
3240 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store(const StoreNode* def_store) const {
3241 Status status_use = find_use_store_unidirectional(def_store);
3242
3243 #ifdef ASSERT
3244 StoreNode* use_store = status_use.found_store();
3245 if (use_store != nullptr) {
3246 Status status_def = find_def_store_unidirectional(use_store);
3247 assert(status_def.found_store() == def_store &&
3248 status_def.found_range_check() == status_use.found_range_check(),
3249 "find_use_store and find_def_store must be symmetric");
3250 }
3251 #endif
3252
3253 return status_use;
3254 }
3255
3256 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store(const StoreNode* use_store) const {
3257 Status status_def = find_def_store_unidirectional(use_store);
3258
3259 #ifdef ASSERT
3260 StoreNode* def_store = status_def.found_store();
3261 if (def_store != nullptr) {
3262 Status status_use = find_use_store_unidirectional(def_store);
3263 assert(status_use.found_store() == use_store &&
3264 status_use.found_range_check() == status_def.found_range_check(),
3265 "find_use_store and find_def_store must be symmetric");
3266 }
3267 #endif
3268
3269 return status_def;
3270 }
3271
3272 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store_unidirectional(const StoreNode* def_store) const {
3273 assert(is_compatible_store(def_store), "precondition: must be compatible with _store");
3274
3275 for (DUIterator_Fast imax, i = def_store->fast_outs(imax); i < imax; i++) {
3276 StoreNode* use_store = def_store->fast_out(i)->isa_Store();
3277 if (is_compatible_store(use_store)) {
3278 return Status::make(use_store, cfg_status_for_pair(use_store, def_store));
3279 }
3280 }
3281
3282 return Status::make_failure();
3283 }
3284
3285 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store_unidirectional(const StoreNode* use_store) const {
3286 assert(is_compatible_store(use_store), "precondition: must be compatible with _store");
3287
3288 StoreNode* def_store = use_store->in(MemNode::Memory)->isa_Store();
3289 if (!is_compatible_store(def_store)) {
3290 return Status::make_failure();
3291 }
3292
3293 return Status::make(def_store, cfg_status_for_pair(use_store, def_store));
3294 }
3295
3296 void MergePrimitiveStores::collect_merge_list(Node_List& merge_list) const {
3297 // The merged store can be at most 8 bytes.
3298 const uint merge_list_max_size = 8 / _store->memory_size();
3299 assert(merge_list_max_size >= 2 &&
3300 merge_list_max_size <= 8 &&
3301 is_power_of_2(merge_list_max_size),
3302 "must be 2, 4 or 8");
3303
3304 // Traverse up the chain of adjacent def stores.
3305 StoreNode* current = _store;
3306 merge_list.push(current);
3307 while (current != nullptr && merge_list.size() < merge_list_max_size) {
3308 Status status = find_adjacent_def_store(current);
3309 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] find def: "); status.print_on(tty); })
3310
3311 current = status.found_store();
3312 if (current != nullptr) {
3313 merge_list.push(current);
3314
3315 // We can have at most one RangeCheck.
3316 if (status.found_range_check()) {
3317 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found RangeCheck, stop traversal."); })
3318 break;
3319 }
3320 }
3321 }
3322
3323 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found:"); merge_list.dump(); })
3324
3325 // Truncate the merge_list to a power of 2.
3326 const uint pow2size = round_down_power_of_2(merge_list.size());
3327 assert(pow2size >= 2, "must be merging at least 2 stores");
3328 while (merge_list.size() > pow2size) { merge_list.pop(); }
3329
3330 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] truncated:"); merge_list.dump(); })
3331 }
3332
3333 // Merge the input values of the smaller stores to a single larger input value.
3334 Node* MergePrimitiveStores::make_merged_input_value(const Node_List& merge_list) {
3335 int new_memory_size = _store->memory_size() * merge_list.size();
3336 Node* first = merge_list.at(merge_list.size()-1);
3337 Node* merged_input_value = nullptr;
3338 if (_store->in(MemNode::ValueIn)->Opcode() == Op_ConI) {
3339 assert(_value_order == ValueOrder::Const, "must match");
3340 // Pattern: [ConI, ConI, ...] -> new constant
3341 jlong con = 0;
3342 jlong bits_per_store = _store->memory_size() * 8;
3343 jlong mask = (((jlong)1) << bits_per_store) - 1;
3344 for (uint i = 0; i < merge_list.size(); i++) {
3345 jlong con_i = merge_list.at(i)->in(MemNode::ValueIn)->get_int();
3346 #ifdef VM_LITTLE_ENDIAN
3347 con = con << bits_per_store;
3348 con = con | (mask & con_i);
3349 #else // VM_LITTLE_ENDIAN
3350 con_i = (mask & con_i) << (i * bits_per_store);
3351 con = con | con_i;
3352 #endif // VM_LITTLE_ENDIAN
3353 }
3354 merged_input_value = _phase->longcon(con);
3355 } else {
3356 assert(_value_order == ValueOrder::Platform || _value_order == ValueOrder::Reverse, "must match");
3357 // Pattern: [base >> 24, base >> 16, base >> 8, base] -> base
3358 // | |
3359 // _store first
3360 //
3361 Node* hi = _store->in(MemNode::ValueIn);
3362 Node* lo = first->in(MemNode::ValueIn);
3363 #ifndef VM_LITTLE_ENDIAN
3364 // `_store` and `first` are swapped in the diagram above
3365 swap(hi, lo);
3366 #endif // !VM_LITTLE_ENDIAN
3367 if (_value_order == ValueOrder::Reverse) {
3368 swap(hi, lo);
3369 }
3370 Node const* hi_base;
3371 jint hi_shift;
3372 merged_input_value = lo;
3373 bool is_true = is_con_RShift(hi, hi_base, hi_shift, _phase);
3374 assert(is_true, "must detect con RShift");
3375 if (merged_input_value != hi_base && merged_input_value->Opcode() == Op_ConvL2I) {
3376 // look through
3377 merged_input_value = merged_input_value->in(1);
3378 }
3379 if (merged_input_value != hi_base) {
3380 // merged_input_value is not the base
3381 return nullptr;
3382 }
3383 }
3384
3385 if (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size <= 4) {
3386 // Example:
3387 //
3388 // long base = ...;
3389 // a[0] = (byte)(base >> 0);
3390 // a[1] = (byte)(base >> 8);
3391 //
3392 merged_input_value = _phase->transform(new ConvL2INode(merged_input_value));
3393 }
3394
3395 assert((_phase->type(merged_input_value)->isa_int() != nullptr && new_memory_size <= 4) ||
3396 (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size == 8),
3397 "merged_input_value is either int or long, and new_memory_size is small enough");
3398
3399 if (_value_order == ValueOrder::Reverse) {
3400 assert(_store->memory_size() == 1, "only implemented for bytes");
3401 if (new_memory_size == 8) {
3402 merged_input_value = _phase->transform(new ReverseBytesLNode(merged_input_value));
3403 } else if (new_memory_size == 4) {
3404 merged_input_value = _phase->transform(new ReverseBytesINode(merged_input_value));
3405 } else {
3406 assert(new_memory_size == 2, "sanity check");
3407 merged_input_value = _phase->transform(new ReverseBytesSNode(merged_input_value));
3408 }
3409 }
3410 return merged_input_value;
3411 }
3412
3413 // //
3414 // first_ctrl first_mem first_adr first_ctrl first_mem first_adr //
3415 // | | | | | | //
3416 // | | | | +---------------+ | //
3417 // | | | | | | | //
3418 // | | +---------+ | | +---------------+ //
3419 // | | | | | | | | //
3420 // +--------------+ | | v1 +------------------------------+ | | v1 //
3421 // | | | | | | | | | | | | //
3422 // RangeCheck first_store RangeCheck | | first_store //
3423 // | | | | | | | //
3424 // last_ctrl | +----> unc_trap last_ctrl | | +----> unc_trap //
3425 // | | ===> | | | //
3426 // +--------------+ | a2 v2 | | | //
3427 // | | | | | | | | //
3428 // | second_store | | | //
3429 // | | | | | [v1 v2 ... vn] //
3430 // ... ... | | | | //
3431 // | | | | | v //
3432 // +--------------+ | an vn +--------------+ | | merged_input_value //
3433 // | | | | | | | | //
3434 // last_store (= _store) merged_store //
3435 // //
3436 StoreNode* MergePrimitiveStores::make_merged_store(const Node_List& merge_list, Node* merged_input_value) {
3437 Node* first_store = merge_list.at(merge_list.size()-1);
3438 Node* last_ctrl = _store->in(MemNode::Control); // after (optional) RangeCheck
3439 Node* first_mem = first_store->in(MemNode::Memory);
3440 Node* first_adr = first_store->in(MemNode::Address);
3441
3442 const TypePtr* new_adr_type = _store->adr_type();
3443
3444 int new_memory_size = _store->memory_size() * merge_list.size();
3445 BasicType bt = T_ILLEGAL;
3446 switch (new_memory_size) {
3447 case 2: bt = T_SHORT; break;
3448 case 4: bt = T_INT; break;
3449 case 8: bt = T_LONG; break;
3450 }
3451
3452 StoreNode* merged_store = StoreNode::make(*_phase, last_ctrl, first_mem, first_adr,
3453 new_adr_type, merged_input_value, bt, MemNode::unordered);
3454
3455 // Marking the store mismatched is sufficient to prevent reordering, since array stores
3456 // are all on the same slice. Hence, we need no barriers.
3457 merged_store->set_mismatched_access();
3458
3459 // Constants above may now also be be packed -> put candidate on worklist
3460 _phase->is_IterGVN()->_worklist.push(first_mem);
3461
3462 return merged_store;
3463 }
3464
3465 #ifndef PRODUCT
3466 void MergePrimitiveStores::trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const {
3467 stringStream ss;
3468 ss.print_cr("[TraceMergeStores]: Replace");
3469 for (int i = (int)merge_list.size() - 1; i >= 0; i--) {
3470 merge_list.at(i)->dump("\n", false, &ss);
3471 }
3472 ss.print_cr("[TraceMergeStores]: with");
3473 merged_input_value->dump("\n", false, &ss);
3474 merged_store->dump("\n", false, &ss);
3475 tty->print("%s", ss.as_string());
3476 }
3477 #endif
3478
3479 //------------------------------Ideal------------------------------------------
3480 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
3481 // When a store immediately follows a relevant allocation/initialization,
3482 // try to capture it into the initialization, or hoist it above.
3483 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3484 Node* p = MemNode::Ideal_common(phase, can_reshape);
3485 if (p) return (p == NodeSentinel) ? nullptr : p;
3486
3487 Node* mem = in(MemNode::Memory);
3488 Node* address = in(MemNode::Address);
3489 Node* value = in(MemNode::ValueIn);
3490 // Back-to-back stores to same address? Fold em up. Generally
3491 // unsafe if I have intervening uses.
3492 {
3493 Node* st = mem;
3494 // If Store 'st' has more than one use, we cannot fold 'st' away.
3495 // For example, 'st' might be the final state at a conditional
3496 // return. Or, 'st' might be used by some node which is live at
3497 // the same time 'st' is live, which might be unschedulable. So,
3498 // require exactly ONE user until such time as we clone 'mem' for
3499 // each of 'mem's uses (thus making the exactly-1-user-rule hold
3500 // true).
3501 while (st->is_Store() && st->outcnt() == 1) {
3502 // Looking at a dead closed cycle of memory?
3503 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
3504 assert(Opcode() == st->Opcode() ||
3505 st->Opcode() == Op_StoreVector ||
3506 Opcode() == Op_StoreVector ||
3507 st->Opcode() == Op_StoreVectorScatter ||
3508 Opcode() == Op_StoreVectorScatter ||
3509 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
3510 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
3511 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
3512 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
3513 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
3514
3515 if (st->in(MemNode::Address)->eqv_uncast(address) &&
3516 st->as_Store()->memory_size() <= this->memory_size()) {
3517 Node* use = st->raw_out(0);
3518 if (phase->is_IterGVN()) {
3519 phase->is_IterGVN()->rehash_node_delayed(use);
3520 }
3521 // It's OK to do this in the parser, since DU info is always accurate,
3522 // and the parser always refers to nodes via SafePointNode maps.
3523 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
3524 return this;
3525 }
3526 st = st->in(MemNode::Memory);
3527 }
3528 }
3529
3530
3531 // Capture an unaliased, unconditional, simple store into an initializer.
3532 // Or, if it is independent of the allocation, hoist it above the allocation.
3533 if (ReduceFieldZeroing && /*can_reshape &&*/
3534 mem->is_Proj() && mem->in(0)->is_Initialize()) {
3535 InitializeNode* init = mem->in(0)->as_Initialize();
3536 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
3537 if (offset > 0) {
3538 Node* moved = init->capture_store(this, offset, phase, can_reshape);
3539 // If the InitializeNode captured me, it made a raw copy of me,
3540 // and I need to disappear.
3541 if (moved != nullptr) {
3542 // %%% hack to ensure that Ideal returns a new node:
3543 mem = MergeMemNode::make(mem);
3544 return mem; // fold me away
3545 }
3546 }
3547 }
3548
3549 // Fold reinterpret cast into memory operation:
3550 // StoreX mem (MoveY2X v) => StoreY mem v
3551 if (value->is_Move()) {
3552 const Type* vt = value->in(1)->bottom_type();
3553 if (has_reinterpret_variant(vt)) {
3554 if (phase->C->post_loop_opts_phase()) {
3555 return convert_to_reinterpret_store(*phase, value->in(1), vt);
3556 } else {
3557 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
3558 }
3559 }
3560 }
3561
3562 if (MergeStores && UseUnalignedAccesses) {
3563 if (phase->C->merge_stores_phase()) {
3564 MergePrimitiveStores merge(phase, this);
3565 Node* progress = merge.run();
3566 if (progress != nullptr) { return progress; }
3567 } else {
3568 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
3569 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
3570 // the stores, and we merge the wrong sequence of stores.
3571 // Example:
3572 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
3573 // Apply MergeStores:
3574 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
3575 // Remove more RangeChecks:
3576 // StoreI [ StoreL ] StoreI
3577 // But now it would have been better to do this instead:
3578 // [ StoreL ] [ StoreL ]
3579 phase->C->record_for_merge_stores_igvn(this);
3580 }
3581 }
3582
3583 return nullptr; // No further progress
3584 }
3585
3586 //------------------------------Value-----------------------------------------
3587 const Type* StoreNode::Value(PhaseGVN* phase) const {
3588 // Either input is TOP ==> the result is TOP
3589 const Type *t1 = phase->type( in(MemNode::Memory) );
3590 if( t1 == Type::TOP ) return Type::TOP;
3591 const Type *t2 = phase->type( in(MemNode::Address) );
3592 if( t2 == Type::TOP ) return Type::TOP;
3593 const Type *t3 = phase->type( in(MemNode::ValueIn) );
3594 if( t3 == Type::TOP ) return Type::TOP;
3595 return Type::MEMORY;
3596 }
3597
3598 //------------------------------Identity---------------------------------------
3599 // Remove redundant stores:
3600 // Store(m, p, Load(m, p)) changes to m.
3601 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
3602 Node* StoreNode::Identity(PhaseGVN* phase) {
3603 Node* mem = in(MemNode::Memory);
3604 Node* adr = in(MemNode::Address);
3605 Node* val = in(MemNode::ValueIn);
3606
3607 Node* result = this;
3608
3609 // Load then Store? Then the Store is useless
3610 if (val->is_Load() &&
3611 val->in(MemNode::Address)->eqv_uncast(adr) &&
3612 val->in(MemNode::Memory )->eqv_uncast(mem) &&
3613 val->as_Load()->store_Opcode() == Opcode()) {
3614 if (!is_StoreVector()) {
3615 result = mem;
3616 } else if (Opcode() == Op_StoreVector && val->Opcode() == Op_LoadVector &&
3617 as_StoreVector()->vect_type() == val->as_LoadVector()->vect_type()) {
3618 // Ensure both are not masked accesses or gathers/scatters and vector types are the same
3619 result = mem;
3620 }
3621 }
3622
3623 // Two stores in a row of the same value?
3624 if (result == this &&
3625 mem->is_Store() &&
3626 mem->in(MemNode::Address)->eqv_uncast(adr) &&
3627 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
3628 mem->Opcode() == Opcode()) {
3629 if (!is_StoreVector()) {
3630 result = mem;
3631 } else {
3632 const StoreVectorNode* store_vector = as_StoreVector();
3633 const StoreVectorNode* mem_vector = mem->as_StoreVector();
3634 const Node* store_indices = store_vector->indices();
3635 const Node* mem_indices = mem_vector->indices();
3636 const Node* store_mask = store_vector->mask();
3637 const Node* mem_mask = mem_vector->mask();
3638 // Ensure types, indices, and masks match
3639 if (store_vector->vect_type() == mem_vector->vect_type() &&
3640 ((store_indices == nullptr) == (mem_indices == nullptr) &&
3641 (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) &&
3642 ((store_mask == nullptr) == (mem_mask == nullptr) &&
3643 (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) {
3644 result = mem;
3645 }
3646 }
3647 }
3648
3649 // Store of zero anywhere into a freshly-allocated object?
3650 // Then the store is useless.
3651 // (It must already have been captured by the InitializeNode.)
3652 if (result == this &&
3653 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
3654 // a newly allocated object is already all-zeroes everywhere
3655 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
3656 result = mem;
3657 }
3658
3659 if (result == this) {
3660 // the store may also apply to zero-bits in an earlier object
3661 Node* prev_mem = find_previous_store(phase);
3662 // Steps (a), (b): Walk past independent stores to find an exact match.
3663 if (prev_mem != nullptr) {
3664 if (prev_mem->is_top()) {
3665 // find_previous_store returns top when the access is dead
3666 return prev_mem;
3667 }
3668 Node* prev_val = can_see_stored_value(prev_mem, phase);
3669 if (prev_val != nullptr && prev_val == val) {
3670 // prev_val and val might differ by a cast; it would be good
3671 // to keep the more informative of the two.
3672 result = mem;
3673 }
3674 }
3675 }
3676 }
3677
3678 PhaseIterGVN* igvn = phase->is_IterGVN();
3679 if (result != this && igvn != nullptr) {
3680 MemBarNode* trailing = trailing_membar();
3681 if (trailing != nullptr) {
3682 #ifdef ASSERT
3683 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
3684 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects");
3685 #endif
3686 trailing->remove(igvn);
3687 }
3688 }
3689
3690 return result;
3691 }
3692
3693 //------------------------------match_edge-------------------------------------
3694 // Do we Match on this edge index or not? Match only memory & value
3695 uint StoreNode::match_edge(uint idx) const {
3696 return idx == MemNode::Address || idx == MemNode::ValueIn;
3697 }
3698
3699 //------------------------------cmp--------------------------------------------
3700 // Do not common stores up together. They generally have to be split
3701 // back up anyways, so do not bother.
3702 bool StoreNode::cmp( const Node &n ) const {
3703 return (&n == this); // Always fail except on self
3704 }
3705
3706 //------------------------------Ideal_masked_input-----------------------------
3707 // Check for a useless mask before a partial-word store
3708 // (StoreB ... (AndI valIn conIa) )
3709 // If (conIa & mask == mask) this simplifies to
3710 // (StoreB ... (valIn) )
3711 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
3712 Node *val = in(MemNode::ValueIn);
3713 if( val->Opcode() == Op_AndI ) {
3714 const TypeInt *t = phase->type( val->in(2) )->isa_int();
3715 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
3716 set_req_X(MemNode::ValueIn, val->in(1), phase);
3717 return this;
3718 }
3719 }
3720 return nullptr;
3721 }
3722
3723
3724 //------------------------------Ideal_sign_extended_input----------------------
3725 // Check for useless sign-extension before a partial-word store
3726 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
3727 // If (conIL == conIR && conIR <= num_rejected_bits) this simplifies to
3728 // (StoreB ... (v))
3729 // If (conIL > conIR) under some conditions, it can be simplified into
3730 // (StoreB ... (LShiftI _ v (conIL - conIR)))
3731 // This case happens when the value of the store was itself a left shift, that
3732 // gets merged into the inner left shift of the sign-extension. For instance,
3733 // if we have
3734 // array_of_shorts[0] = (short)(v << 2)
3735 // We get a structure such as:
3736 // (StoreB ... (RShiftI _ (LShiftI _ (LShiftI _ v 2) 16) 16))
3737 // that is simplified into
3738 // (StoreB ... (RShiftI _ (LShiftI _ v 18) 16)).
3739 // It is thus useful to handle cases where conIL > conIR. But this simplification
3740 // does not always hold. Let's see in which cases it's valid.
3741 //
3742 // Let's assume we have the following 32 bits integer v:
3743 // +----------------------------------+
3744 // | v[0..31] |
3745 // +----------------------------------+
3746 // 31 0
3747 // that will be stuffed in 8 bits byte after a shift left and a shift right of
3748 // potentially different magnitudes.
3749 // We denote num_rejected_bits the number of bits of the discarded part. In this
3750 // case, num_rejected_bits == 24.
3751 //
3752 // Statement (proved further below in case analysis):
3753 // Given:
3754 // - 0 <= conIL < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
3755 // - 0 <= conIR < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
3756 // - conIL >= conIR
3757 // - num_rejected_bits >= conIR
3758 // Then this form:
3759 // (RShiftI _ (LShiftI _ v conIL) conIR)
3760 // can be replaced with this form:
3761 // (LShiftI _ v (conIL-conIR))
3762 //
3763 // Note: We only have to show that the non-rejected lowest bits (8 bits for byte)
3764 // have to be correct, as the higher bits are rejected / truncated by the store.
3765 //
3766 // The hypotheses
3767 // 0 <= conIL < BitsPerJavaInteger
3768 // 0 <= conIR < BitsPerJavaInteger
3769 // are ensured by maskShiftAmount (called from ::Ideal of shift nodes). Indeed,
3770 // (v << 31) << 2 must be simplified into 0, not into v << 33 (which is equivalent
3771 // to v << 1).
3772 //
3773 //
3774 // If you don't like case analysis, jump after the conclusion.
3775 // ### Case 1 : conIL == conIR
3776 // ###### Case 1.1: conIL == conIR == num_rejected_bits
3777 // If we do the shift left then right by 24 bits, we get:
3778 // after: v << 24
3779 // +---------+------------------------+
3780 // | v[0..7] | 0 |
3781 // +---------+------------------------+
3782 // 31 24 23 0
3783 // after: (v << 24) >> 24
3784 // +------------------------+---------+
3785 // | sign bit | v[0..7] |
3786 // +------------------------+---------+
3787 // 31 8 7 0
3788 // The non-rejected bits (bits kept by the store, that is the 8 lower bits of the
3789 // result) are the same before and after, so, indeed, simplifying is correct.
3790
3791 // ###### Case 1.2: conIL == conIR < num_rejected_bits
3792 // If we do the shift left then right by 22 bits, we get:
3793 // after: v << 22
3794 // +---------+------------------------+
3795 // | v[0..9] | 0 |
3796 // +---------+------------------------+
3797 // 31 22 21 0
3798 // after: (v << 22) >> 22
3799 // +------------------------+---------+
3800 // | sign bit | v[0..9] |
3801 // +------------------------+---------+
3802 // 31 10 9 0
3803 // The non-rejected bits are the 8 lower bits of v. The bits 8 and 9 of v are still
3804 // present in (v << 22) >> 22 but will be dropped by the store. The simplification is
3805 // still correct.
3806
3807 // ###### But! Case 1.3: conIL == conIR > num_rejected_bits
3808 // If we do the shift left then right by 26 bits, we get:
3809 // after: v << 26
3810 // +---------+------------------------+
3811 // | v[0..5] | 0 |
3812 // +---------+------------------------+
3813 // 31 26 25 0
3814 // after: (v << 26) >> 26
3815 // +------------------------+---------+
3816 // | sign bit | v[0..5] |
3817 // +------------------------+---------+
3818 // 31 6 5 0
3819 // The non-rejected bits are made of
3820 // - 0-5 => the bits 0 to 5 of v
3821 // - 6-7 => the sign bit of v[0..5] (that is v[5])
3822 // Simplifying this as v is not correct.
3823 // The condition conIR <= num_rejected_bits is indeed necessary in Case 1
3824 //
3825 // ### Case 2: conIL > conIR
3826 // ###### Case 2.1: num_rejected_bits == conIR
3827 // We take conIL == 26 for this example.
3828 // after: v << 26
3829 // +---------+------------------------+
3830 // | v[0..5] | 0 |
3831 // +---------+------------------------+
3832 // 31 26 25 0
3833 // after: (v << 26) >> 24
3834 // +------------------+---------+-----+
3835 // | sign bit | v[0..5] | 0 |
3836 // +------------------+---------+-----+
3837 // 31 8 7 2 1 0
3838 // The non-rejected bits are the 8 lower ones of (v << conIL - conIR).
3839 // The bits 6 and 7 of v have been thrown away after the shift left.
3840 // The simplification is still correct.
3841 //
3842 // ###### Case 2.2: num_rejected_bits > conIR.
3843 // Let's say conIL == 26 and conIR == 22.
3844 // after: v << 26
3845 // +---------+------------------------+
3846 // | v[0..5] | 0 |
3847 // +---------+------------------------+
3848 // 31 26 25 0
3849 // after: (v << 26) >> 22
3850 // +------------------+---------+-----+
3851 // | sign bit | v[0..5] | 0 |
3852 // +------------------+---------+-----+
3853 // 31 10 9 4 3 0
3854 // The bits non-rejected by the store are exactly the 8 lower ones of (v << (conIL - conIR)):
3855 // - 0-3 => 0
3856 // - 4-7 => bits 0 to 3 of v
3857 // The simplification is still correct.
3858 // The bits 4 and 5 of v are still present in (v << (conIL - conIR)) but they don't
3859 // matter as they are not in the 8 lower bits: they will be cut out by the store.
3860 //
3861 // ###### But! Case 2.3: num_rejected_bits < conIR.
3862 // Let's see that this case is not as easy to simplify.
3863 // Let's say conIL == 28 and conIR == 26.
3864 // after: v << 28
3865 // +---------+------------------------+
3866 // | v[0..3] | 0 |
3867 // +---------+------------------------+
3868 // 31 28 27 0
3869 // after: (v << 28) >> 26
3870 // +------------------+---------+-----+
3871 // | sign bit | v[0..3] | 0 |
3872 // +------------------+---------+-----+
3873 // 31 6 5 2 1 0
3874 // The non-rejected bits are made of
3875 // - 0-1 => 0
3876 // - 2-5 => the bits 0 to 3 of v
3877 // - 6-7 => the sign bit of v[0..3] (that is v[3])
3878 // Simplifying this as (v << 2) is not correct.
3879 // The condition conIR <= num_rejected_bits is indeed necessary in Case 2.
3880 //
3881 // ### Conclusion:
3882 // Our hypotheses are indeed sufficient:
3883 // - 0 <= conIL < BitsPerJavaInteger
3884 // - 0 <= conIR < BitsPerJavaInteger
3885 // - conIL >= conIR
3886 // - num_rejected_bits >= conIR
3887 //
3888 // ### A rationale without case analysis:
3889 // After the shift left, conIL upper bits of v are discarded and conIL lower bit
3890 // zeroes are added. After the shift right, conIR lower bits of the previous result
3891 // are discarded. If conIL >= conIR, we discard only the zeroes we made up during
3892 // the shift left, but if conIL < conIR, then we discard also lower bits of v. But
3893 // the point of the simplification is to get an expression of the form
3894 // (v << (conIL - conIR)). This expression discard only higher bits of v, thus the
3895 // simplification is not correct if conIL < conIR.
3896 //
3897 // Moreover, after the shift right, the higher bit of (v << conIL) is repeated on the
3898 // conIR higher bits of ((v << conIL) >> conIR), it's the sign-extension. If
3899 // conIR > num_rejected_bits, then at least one artificial copy of this sign bit will
3900 // be in the window of the store. Thus ((v << conIL) >> conIR) is not equivalent to
3901 // (v << (conIL-conIR)) if conIR > num_rejected_bits.
3902 //
3903 // We do not treat the case conIL < conIR here since the point of this function is
3904 // to skip sign-extensions (that is conIL == conIR == num_rejected_bits). The need
3905 // of treating conIL > conIR comes from the cases where the sign-extended value is
3906 // also left-shift expression. Computing the sign-extension of a right-shift expression
3907 // doesn't yield a situation such as
3908 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
3909 // where conIL < conIR.
3910 Node* StoreNode::Ideal_sign_extended_input(PhaseGVN* phase, int num_rejected_bits) {
3911 Node* shr = in(MemNode::ValueIn);
3912 if (shr->Opcode() == Op_RShiftI) {
3913 const TypeInt* conIR = phase->type(shr->in(2))->isa_int();
3914 if (conIR != nullptr && conIR->is_con() && conIR->get_con() >= 0 && conIR->get_con() < BitsPerJavaInteger && conIR->get_con() <= num_rejected_bits) {
3915 Node* shl = shr->in(1);
3916 if (shl->Opcode() == Op_LShiftI) {
3917 const TypeInt* conIL = phase->type(shl->in(2))->isa_int();
3918 if (conIL != nullptr && conIL->is_con() && conIL->get_con() >= 0 && conIL->get_con() < BitsPerJavaInteger) {
3919 if (conIL->get_con() == conIR->get_con()) {
3920 set_req_X(MemNode::ValueIn, shl->in(1), phase);
3921 return this;
3922 }
3923 if (conIL->get_con() > conIR->get_con()) {
3924 Node* new_shl = phase->transform(new LShiftINode(shl->in(1), phase->intcon(conIL->get_con() - conIR->get_con())));
3925 set_req_X(MemNode::ValueIn, new_shl, phase);
3926 return this;
3927 }
3928 }
3929 }
3930 }
3931 }
3932 return nullptr;
3933 }
3934
3935 //------------------------------value_never_loaded-----------------------------------
3936 // Determine whether there are any possible loads of the value stored.
3937 // For simplicity, we actually check if there are any loads from the
3938 // address stored to, not just for loads of the value stored by this node.
3939 //
3940 bool StoreNode::value_never_loaded(PhaseValues* phase) const {
3941 Node *adr = in(Address);
3942 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
3943 if (adr_oop == nullptr)
3944 return false;
3945 if (!adr_oop->is_known_instance_field())
3946 return false; // if not a distinct instance, there may be aliases of the address
3947 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
3948 Node *use = adr->fast_out(i);
3949 if (use->is_Load() || use->is_LoadStore()) {
3950 return false;
3951 }
3952 }
3953 return true;
3954 }
3955
3956 MemBarNode* StoreNode::trailing_membar() const {
3957 if (is_release()) {
3958 MemBarNode* trailing_mb = nullptr;
3959 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3960 Node* u = fast_out(i);
3961 if (u->is_MemBar()) {
3962 if (u->as_MemBar()->trailing_store()) {
3963 assert(u->Opcode() == Op_MemBarVolatile, "");
3964 assert(trailing_mb == nullptr, "only one");
3965 trailing_mb = u->as_MemBar();
3966 #ifdef ASSERT
3967 Node* leading = u->as_MemBar()->leading_membar();
3968 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3969 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
3970 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
3971 #endif
3972 } else {
3973 assert(u->as_MemBar()->standalone(), "");
3974 }
3975 }
3976 }
3977 return trailing_mb;
3978 }
3979 return nullptr;
3980 }
3981
3982
3983 //=============================================================================
3984 //------------------------------Ideal------------------------------------------
3985 // If the store is from an AND mask that leaves the low bits untouched, then
3986 // we can skip the AND operation. If the store is from a sign-extension
3987 // (a left shift, then right shift) we can skip both.
3988 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
3989 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
3990 if( progress != nullptr ) return progress;
3991
3992 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
3993 if( progress != nullptr ) return progress;
3994
3995 // Finally check the default case
3996 return StoreNode::Ideal(phase, can_reshape);
3997 }
3998
3999 //=============================================================================
4000 //------------------------------Ideal------------------------------------------
4001 // If the store is from an AND mask that leaves the low bits untouched, then
4002 // we can skip the AND operation
4003 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
4004 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
4005 if( progress != nullptr ) return progress;
4006
4007 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
4008 if( progress != nullptr ) return progress;
4009
4010 // Finally check the default case
4011 return StoreNode::Ideal(phase, can_reshape);
4012 }
4013
4014 //=============================================================================
4015 //----------------------------------SCMemProjNode------------------------------
4016 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
4017 {
4018 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
4019 return Type::TOP;
4020 }
4021 return bottom_type();
4022 }
4023
4024 //=============================================================================
4025 //----------------------------------LoadStoreNode------------------------------
4026 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
4027 : Node(required),
4028 _type(rt),
4029 _barrier_data(0)
4030 {
4031 init_req(MemNode::Control, c );
4032 init_req(MemNode::Memory , mem);
4033 init_req(MemNode::Address, adr);
4034 init_req(MemNode::ValueIn, val);
4035 init_class_id(Class_LoadStore);
4036 DEBUG_ONLY(_adr_type = at; adr_type();)
4037 }
4038
4039 //------------------------------Value-----------------------------------------
4040 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
4041 // Either input is TOP ==> the result is TOP
4042 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
4043 return Type::TOP;
4044 }
4045 const Type* t = phase->type(in(MemNode::Memory));
4046 if (t == Type::TOP) {
4047 return Type::TOP;
4048 }
4049 t = phase->type(in(MemNode::Address));
4050 if (t == Type::TOP) {
4051 return Type::TOP;
4052 }
4053 t = phase->type(in(MemNode::ValueIn));
4054 if (t == Type::TOP) {
4055 return Type::TOP;
4056 }
4057 return bottom_type();
4058 }
4059
4060 const TypePtr* LoadStoreNode::adr_type() const {
4061 const TypePtr* cross_check = DEBUG_ONLY(_adr_type) NOT_DEBUG(nullptr);
4062 return MemNode::calculate_adr_type(in(MemNode::Address)->bottom_type(), cross_check);
4063 }
4064
4065 uint LoadStoreNode::ideal_reg() const {
4066 return _type->ideal_reg();
4067 }
4068
4069 // This method conservatively checks if the result of a LoadStoreNode is
4070 // used, that is, if it returns true, then it is definitely the case that
4071 // the result of the node is not needed.
4072 // For example, GetAndAdd can be matched into a lock_add instead of a
4073 // lock_xadd if the result of LoadStoreNode::result_not_used() is true
4074 bool LoadStoreNode::result_not_used() const {
4075 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4076 Node *x = fast_out(i);
4077 if (x->Opcode() == Op_SCMemProj) {
4078 continue;
4079 }
4080 if (x->bottom_type() == TypeTuple::MEMBAR &&
4081 !x->is_Call() &&
4082 x->Opcode() != Op_Blackhole) {
4083 continue;
4084 }
4085 return false;
4086 }
4087 return true;
4088 }
4089
4090 MemBarNode* LoadStoreNode::trailing_membar() const {
4091 MemBarNode* trailing = nullptr;
4092 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4093 Node* u = fast_out(i);
4094 if (u->is_MemBar()) {
4095 if (u->as_MemBar()->trailing_load_store()) {
4096 assert(u->Opcode() == Op_MemBarAcquire, "");
4097 assert(trailing == nullptr, "only one");
4098 trailing = u->as_MemBar();
4099 #ifdef ASSERT
4100 Node* leading = trailing->leading_membar();
4101 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
4102 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
4103 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
4104 #endif
4105 } else {
4106 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
4107 }
4108 }
4109 }
4110
4111 return trailing;
4112 }
4113
4114 uint LoadStoreNode::size_of() const { return sizeof(*this); }
4115
4116 //=============================================================================
4117 //----------------------------------LoadStoreConditionalNode--------------------
4118 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) {
4119 init_req(ExpectedIn, ex );
4120 }
4121
4122 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
4123 // Either input is TOP ==> the result is TOP
4124 const Type* t = phase->type(in(ExpectedIn));
4125 if (t == Type::TOP) {
4126 return Type::TOP;
4127 }
4128 return LoadStoreNode::Value(phase);
4129 }
4130
4131 //=============================================================================
4132 //-------------------------------adr_type--------------------------------------
4133 const TypePtr* ClearArrayNode::adr_type() const {
4134 Node *adr = in(3);
4135 if (adr == nullptr) return nullptr; // node is dead
4136 return MemNode::calculate_adr_type(adr->bottom_type());
4137 }
4138
4139 //------------------------------match_edge-------------------------------------
4140 // Do we Match on this edge index or not? Do not match memory
4141 uint ClearArrayNode::match_edge(uint idx) const {
4142 return idx > 1;
4143 }
4144
4145 //------------------------------Identity---------------------------------------
4146 // Clearing a zero length array does nothing
4147 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
4148 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
4149 }
4150
4151 //------------------------------Idealize---------------------------------------
4152 // Clearing a short array is faster with stores
4153 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4154 // Already know this is a large node, do not try to ideal it
4155 if (_is_large) return nullptr;
4156
4157 const int unit = BytesPerLong;
4158 const TypeX* t = phase->type(in(2))->isa_intptr_t();
4159 if (!t) return nullptr;
4160 if (!t->is_con()) return nullptr;
4161 intptr_t raw_count = t->get_con();
4162 intptr_t size = raw_count;
4163 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
4164 // Clearing nothing uses the Identity call.
4165 // Negative clears are possible on dead ClearArrays
4166 // (see jck test stmt114.stmt11402.val).
4167 if (size <= 0 || size % unit != 0) return nullptr;
4168 intptr_t count = size / unit;
4169 // Length too long; communicate this to matchers and assemblers.
4170 // Assemblers are responsible to produce fast hardware clears for it.
4171 if (size > InitArrayShortSize) {
4172 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
4173 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
4174 return nullptr;
4175 }
4176 if (!IdealizeClearArrayNode) return nullptr;
4177 Node *mem = in(1);
4178 if( phase->type(mem)==Type::TOP ) return nullptr;
4179 Node *adr = in(3);
4180 const Type* at = phase->type(adr);
4181 if( at==Type::TOP ) return nullptr;
4182 const TypePtr* atp = at->isa_ptr();
4183 // adjust atp to be the correct array element address type
4184 if (atp == nullptr) atp = TypePtr::BOTTOM;
4185 else atp = atp->add_offset(Type::OffsetBot);
4186 // Get base for derived pointer purposes
4187 if( adr->Opcode() != Op_AddP ) Unimplemented();
4188 Node *base = adr->in(1);
4189
4190 Node *zero = phase->makecon(TypeLong::ZERO);
4191 Node *off = phase->MakeConX(BytesPerLong);
4192 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
4193 count--;
4194 while( count-- ) {
4195 mem = phase->transform(mem);
4196 adr = phase->transform(new AddPNode(base,adr,off));
4197 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
4198 }
4199 return mem;
4200 }
4201
4202 //----------------------------step_through----------------------------------
4203 // Return allocation input memory edge if it is different instance
4204 // or itself if it is the one we are looking for.
4205 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
4206 Node* n = *np;
4207 assert(n->is_ClearArray(), "sanity");
4208 intptr_t offset;
4209 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
4210 // This method is called only before Allocate nodes are expanded
4211 // during macro nodes expansion. Before that ClearArray nodes are
4212 // only generated in PhaseMacroExpand::generate_arraycopy() (before
4213 // Allocate nodes are expanded) which follows allocations.
4214 assert(alloc != nullptr, "should have allocation");
4215 if (alloc->_idx == instance_id) {
4216 // Can not bypass initialization of the instance we are looking for.
4217 return false;
4218 }
4219 // Otherwise skip it.
4220 InitializeNode* init = alloc->initialization();
4221 if (init != nullptr)
4222 *np = init->in(TypeFunc::Memory);
4223 else
4224 *np = alloc->in(TypeFunc::Memory);
4225 return true;
4226 }
4227
4228 Node* ClearArrayNode::make_address(Node* dest, Node* offset, bool raw_base, PhaseGVN* phase) {
4229 Node* base = dest;
4230 if (raw_base) {
4231 // May be called as part of the initialization of a just allocated object
4232 base = phase->C->top();
4233 }
4234 return phase->transform(new AddPNode(base, dest, offset));
4235 }
4236
4237 //----------------------------clear_memory-------------------------------------
4238 // Generate code to initialize object storage to zero.
4239 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4240 intptr_t start_offset,
4241 Node* end_offset,
4242 bool raw_base,
4243 PhaseGVN* phase) {
4244 intptr_t offset = start_offset;
4245
4246 int unit = BytesPerLong;
4247 if ((offset % unit) != 0) {
4248 Node* adr = make_address(dest, phase->MakeConX(offset), raw_base, phase);
4249 const TypePtr* atp = TypeRawPtr::BOTTOM;
4250 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4251 mem = phase->transform(mem);
4252 offset += BytesPerInt;
4253 }
4254 assert((offset % unit) == 0, "");
4255
4256 // Initialize the remaining stuff, if any, with a ClearArray.
4257 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, raw_base, phase);
4258 }
4259
4260 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4261 Node* start_offset,
4262 Node* end_offset,
4263 bool raw_base,
4264 PhaseGVN* phase) {
4265 if (start_offset == end_offset) {
4266 // nothing to do
4267 return mem;
4268 }
4269
4270 int unit = BytesPerLong;
4271 Node* zbase = start_offset;
4272 Node* zend = end_offset;
4273
4274 // Scale to the unit required by the CPU:
4275 if (!Matcher::init_array_count_is_in_bytes) {
4276 Node* shift = phase->intcon(exact_log2(unit));
4277 zbase = phase->transform(new URShiftXNode(zbase, shift) );
4278 zend = phase->transform(new URShiftXNode(zend, shift) );
4279 }
4280
4281 // Bulk clear double-words
4282 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
4283 Node* adr = make_address(dest, start_offset, raw_base, phase);
4284 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
4285 return phase->transform(mem);
4286 }
4287
4288 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4289 intptr_t start_offset,
4290 intptr_t end_offset,
4291 bool raw_base,
4292 PhaseGVN* phase) {
4293 if (start_offset == end_offset) {
4294 // nothing to do
4295 return mem;
4296 }
4297
4298 assert((end_offset % BytesPerInt) == 0, "odd end offset");
4299 intptr_t done_offset = end_offset;
4300 if ((done_offset % BytesPerLong) != 0) {
4301 done_offset -= BytesPerInt;
4302 }
4303 if (done_offset > start_offset) {
4304 mem = clear_memory(ctl, mem, dest,
4305 start_offset, phase->MakeConX(done_offset), raw_base, phase);
4306 }
4307 if (done_offset < end_offset) { // emit the final 32-bit store
4308 Node* adr = make_address(dest, phase->MakeConX(done_offset), raw_base, phase);
4309 const TypePtr* atp = TypeRawPtr::BOTTOM;
4310 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4311 mem = phase->transform(mem);
4312 done_offset += BytesPerInt;
4313 }
4314 assert(done_offset == end_offset, "");
4315 return mem;
4316 }
4317
4318 //=============================================================================
4319 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
4320 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
4321 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
4322 #ifdef ASSERT
4323 , _pair_idx(0)
4324 #endif
4325 {
4326 init_class_id(Class_MemBar);
4327 Node* top = C->top();
4328 init_req(TypeFunc::I_O,top);
4329 init_req(TypeFunc::FramePtr,top);
4330 init_req(TypeFunc::ReturnAdr,top);
4331 if (precedent != nullptr)
4332 init_req(TypeFunc::Parms, precedent);
4333 }
4334
4335 //------------------------------cmp--------------------------------------------
4336 uint MemBarNode::hash() const { return NO_HASH; }
4337 bool MemBarNode::cmp( const Node &n ) const {
4338 return (&n == this); // Always fail except on self
4339 }
4340
4341 //------------------------------make-------------------------------------------
4342 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
4343 switch (opcode) {
4344 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
4345 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
4346 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
4347 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
4348 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
4349 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn);
4350 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
4351 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
4352 case Op_MemBarStoreLoad: return new MemBarStoreLoadNode(C, atp, pn);
4353 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
4354 case Op_MemBarFull: return new MemBarFullNode(C, atp, pn);
4355 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
4356 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
4357 case Op_Initialize: return new InitializeNode(C, atp, pn);
4358 default: ShouldNotReachHere(); return nullptr;
4359 }
4360 }
4361
4362 void MemBarNode::remove(PhaseIterGVN *igvn) {
4363 assert(outcnt() > 0 && (outcnt() <= 2 || Opcode() == Op_Initialize), "Only one or two out edges allowed");
4364 if (trailing_store() || trailing_load_store()) {
4365 MemBarNode* leading = leading_membar();
4366 if (leading != nullptr) {
4367 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
4368 leading->remove(igvn);
4369 }
4370 }
4371 if (proj_out_or_null(TypeFunc::Control) != nullptr) {
4372 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
4373 }
4374 if (is_Initialize()) {
4375 as_Initialize()->replace_mem_projs_by(in(TypeFunc::Memory), igvn);
4376 } else {
4377 if (proj_out_or_null(TypeFunc::Memory) != nullptr) {
4378 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
4379 }
4380 }
4381 }
4382
4383 //------------------------------Ideal------------------------------------------
4384 // Return a node which is more "ideal" than the current node. Strip out
4385 // control copies
4386 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4387 if (remove_dead_region(phase, can_reshape)) return this;
4388 // Don't bother trying to transform a dead node
4389 if (in(0) && in(0)->is_top()) {
4390 return nullptr;
4391 }
4392
4393 bool progress = false;
4394 // Eliminate volatile MemBars for scalar replaced objects.
4395 if (can_reshape && req() == (Precedent+1)) {
4396 bool eliminate = false;
4397 int opc = Opcode();
4398 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
4399 // Volatile field loads and stores.
4400 Node* my_mem = in(MemBarNode::Precedent);
4401 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
4402 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
4403 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
4404 // replace this Precedent (decodeN) with the Load instead.
4405 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
4406 Node* load_node = my_mem->in(1);
4407 set_req(MemBarNode::Precedent, load_node);
4408 phase->is_IterGVN()->_worklist.push(my_mem);
4409 my_mem = load_node;
4410 } else {
4411 assert(my_mem->unique_out() == this, "sanity");
4412 assert(!trailing_load_store(), "load store node can't be eliminated");
4413 del_req(Precedent);
4414 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
4415 my_mem = nullptr;
4416 }
4417 progress = true;
4418 }
4419 if (my_mem != nullptr && my_mem->is_Mem()) {
4420 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
4421 // Check for scalar replaced object reference.
4422 if( t_oop != nullptr && t_oop->is_known_instance_field() &&
4423 t_oop->offset() != Type::OffsetBot &&
4424 t_oop->offset() != Type::OffsetTop) {
4425 eliminate = true;
4426 }
4427 }
4428 } else if (opc == Op_MemBarRelease || (UseStoreStoreForCtor && opc == Op_MemBarStoreStore)) {
4429 // Final field stores.
4430 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent));
4431 if ((alloc != nullptr) && alloc->is_Allocate() &&
4432 alloc->as_Allocate()->does_not_escape_thread()) {
4433 // The allocated object does not escape.
4434 eliminate = true;
4435 }
4436 }
4437 if (eliminate) {
4438 // Replace MemBar projections by its inputs.
4439 PhaseIterGVN* igvn = phase->is_IterGVN();
4440 remove(igvn);
4441 // Must return either the original node (now dead) or a new node
4442 // (Do not return a top here, since that would break the uniqueness of top.)
4443 return new ConINode(TypeInt::ZERO);
4444 }
4445 }
4446 return progress ? this : nullptr;
4447 }
4448
4449 //------------------------------Value------------------------------------------
4450 const Type* MemBarNode::Value(PhaseGVN* phase) const {
4451 if( !in(0) ) return Type::TOP;
4452 if( phase->type(in(0)) == Type::TOP )
4453 return Type::TOP;
4454 return TypeTuple::MEMBAR;
4455 }
4456
4457 //------------------------------match------------------------------------------
4458 // Construct projections for memory.
4459 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
4460 switch (proj->_con) {
4461 case TypeFunc::Control:
4462 case TypeFunc::Memory:
4463 return new MachProjNode(this, proj->_con, RegMask::EMPTY, MachProjNode::unmatched_proj);
4464 }
4465 ShouldNotReachHere();
4466 return nullptr;
4467 }
4468
4469 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4470 trailing->_kind = TrailingStore;
4471 leading->_kind = LeadingStore;
4472 #ifdef ASSERT
4473 trailing->_pair_idx = leading->_idx;
4474 leading->_pair_idx = leading->_idx;
4475 #endif
4476 }
4477
4478 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4479 trailing->_kind = TrailingLoadStore;
4480 leading->_kind = LeadingLoadStore;
4481 #ifdef ASSERT
4482 trailing->_pair_idx = leading->_idx;
4483 leading->_pair_idx = leading->_idx;
4484 #endif
4485 }
4486
4487 MemBarNode* MemBarNode::trailing_membar() const {
4488 ResourceMark rm;
4489 Node* trailing = (Node*)this;
4490 VectorSet seen;
4491 Node_Stack multis(0);
4492 do {
4493 Node* c = trailing;
4494 uint i = 0;
4495 do {
4496 trailing = nullptr;
4497 for (; i < c->outcnt(); i++) {
4498 Node* next = c->raw_out(i);
4499 if (next != c && next->is_CFG()) {
4500 if (c->is_MultiBranch()) {
4501 if (multis.node() == c) {
4502 multis.set_index(i+1);
4503 } else {
4504 multis.push(c, i+1);
4505 }
4506 }
4507 trailing = next;
4508 break;
4509 }
4510 }
4511 if (trailing != nullptr && !seen.test_set(trailing->_idx)) {
4512 break;
4513 }
4514 while (multis.size() > 0) {
4515 c = multis.node();
4516 i = multis.index();
4517 if (i < c->req()) {
4518 break;
4519 }
4520 multis.pop();
4521 }
4522 } while (multis.size() > 0);
4523 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
4524
4525 MemBarNode* mb = trailing->as_MemBar();
4526 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
4527 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
4528 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
4529 return mb;
4530 }
4531
4532 MemBarNode* MemBarNode::leading_membar() const {
4533 ResourceMark rm;
4534 VectorSet seen;
4535 Node_Stack regions(0);
4536 Node* leading = in(0);
4537 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
4538 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) {
4539 leading = nullptr;
4540 while (regions.size() > 0 && leading == nullptr) {
4541 Node* r = regions.node();
4542 uint i = regions.index();
4543 if (i < r->req()) {
4544 leading = r->in(i);
4545 regions.set_index(i+1);
4546 } else {
4547 regions.pop();
4548 }
4549 }
4550 if (leading == nullptr) {
4551 assert(regions.size() == 0, "all paths should have been tried");
4552 return nullptr;
4553 }
4554 }
4555 if (leading->is_Region()) {
4556 regions.push(leading, 2);
4557 leading = leading->in(1);
4558 } else {
4559 leading = leading->in(0);
4560 }
4561 }
4562 #ifdef ASSERT
4563 Unique_Node_List wq;
4564 wq.push((Node*)this);
4565 uint found = 0;
4566 for (uint i = 0; i < wq.size(); i++) {
4567 Node* n = wq.at(i);
4568 if (n->is_Region()) {
4569 for (uint j = 1; j < n->req(); j++) {
4570 Node* in = n->in(j);
4571 if (in != nullptr && !in->is_top()) {
4572 wq.push(in);
4573 }
4574 }
4575 } else {
4576 if (n->is_MemBar() && n->as_MemBar()->leading()) {
4577 assert(n == leading, "consistency check failed");
4578 found++;
4579 } else {
4580 Node* in = n->in(0);
4581 if (in != nullptr && !in->is_top()) {
4582 wq.push(in);
4583 }
4584 }
4585 }
4586 }
4587 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed");
4588 #endif
4589 if (leading == nullptr) {
4590 return nullptr;
4591 }
4592 MemBarNode* mb = leading->as_MemBar();
4593 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
4594 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
4595 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
4596 return mb;
4597 }
4598
4599
4600 //===========================InitializeNode====================================
4601 // SUMMARY:
4602 // This node acts as a memory barrier on raw memory, after some raw stores.
4603 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
4604 // The Initialize can 'capture' suitably constrained stores as raw inits.
4605 // It can coalesce related raw stores into larger units (called 'tiles').
4606 // It can avoid zeroing new storage for memory units which have raw inits.
4607 // At macro-expansion, it is marked 'complete', and does not optimize further.
4608 //
4609 // EXAMPLE:
4610 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
4611 // ctl = incoming control; mem* = incoming memory
4612 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
4613 // First allocate uninitialized memory and fill in the header:
4614 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
4615 // ctl := alloc.Control; mem* := alloc.Memory*
4616 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
4617 // Then initialize to zero the non-header parts of the raw memory block:
4618 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
4619 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
4620 // After the initialize node executes, the object is ready for service:
4621 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
4622 // Suppose its body is immediately initialized as {1,2}:
4623 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4624 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4625 // mem.SLICE(#short[*]) := store2
4626 //
4627 // DETAILS:
4628 // An InitializeNode collects and isolates object initialization after
4629 // an AllocateNode and before the next possible safepoint. As a
4630 // memory barrier (MemBarNode), it keeps critical stores from drifting
4631 // down past any safepoint or any publication of the allocation.
4632 // Before this barrier, a newly-allocated object may have uninitialized bits.
4633 // After this barrier, it may be treated as a real oop, and GC is allowed.
4634 //
4635 // The semantics of the InitializeNode include an implicit zeroing of
4636 // the new object from object header to the end of the object.
4637 // (The object header and end are determined by the AllocateNode.)
4638 //
4639 // Certain stores may be added as direct inputs to the InitializeNode.
4640 // These stores must update raw memory, and they must be to addresses
4641 // derived from the raw address produced by AllocateNode, and with
4642 // a constant offset. They must be ordered by increasing offset.
4643 // The first one is at in(RawStores), the last at in(req()-1).
4644 // Unlike most memory operations, they are not linked in a chain,
4645 // but are displayed in parallel as users of the rawmem output of
4646 // the allocation.
4647 //
4648 // (See comments in InitializeNode::capture_store, which continue
4649 // the example given above.)
4650 //
4651 // When the associated Allocate is macro-expanded, the InitializeNode
4652 // may be rewritten to optimize collected stores. A ClearArrayNode
4653 // may also be created at that point to represent any required zeroing.
4654 // The InitializeNode is then marked 'complete', prohibiting further
4655 // capturing of nearby memory operations.
4656 //
4657 // During macro-expansion, all captured initializations which store
4658 // constant values of 32 bits or smaller are coalesced (if advantageous)
4659 // into larger 'tiles' 32 or 64 bits. This allows an object to be
4660 // initialized in fewer memory operations. Memory words which are
4661 // covered by neither tiles nor non-constant stores are pre-zeroed
4662 // by explicit stores of zero. (The code shape happens to do all
4663 // zeroing first, then all other stores, with both sequences occurring
4664 // in order of ascending offsets.)
4665 //
4666 // Alternatively, code may be inserted between an AllocateNode and its
4667 // InitializeNode, to perform arbitrary initialization of the new object.
4668 // E.g., the object copying intrinsics insert complex data transfers here.
4669 // The initialization must then be marked as 'complete' disable the
4670 // built-in zeroing semantics and the collection of initializing stores.
4671 //
4672 // While an InitializeNode is incomplete, reads from the memory state
4673 // produced by it are optimizable if they match the control edge and
4674 // new oop address associated with the allocation/initialization.
4675 // They return a stored value (if the offset matches) or else zero.
4676 // A write to the memory state, if it matches control and address,
4677 // and if it is to a constant offset, may be 'captured' by the
4678 // InitializeNode. It is cloned as a raw memory operation and rewired
4679 // inside the initialization, to the raw oop produced by the allocation.
4680 // Operations on addresses which are provably distinct (e.g., to
4681 // other AllocateNodes) are allowed to bypass the initialization.
4682 //
4683 // The effect of all this is to consolidate object initialization
4684 // (both arrays and non-arrays, both piecewise and bulk) into a
4685 // single location, where it can be optimized as a unit.
4686 //
4687 // Only stores with an offset less than TrackedInitializationLimit words
4688 // will be considered for capture by an InitializeNode. This puts a
4689 // reasonable limit on the complexity of optimized initializations.
4690
4691 //---------------------------InitializeNode------------------------------------
4692 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
4693 : MemBarNode(C, adr_type, rawoop),
4694 _is_complete(Incomplete), _does_not_escape(false)
4695 {
4696 init_class_id(Class_Initialize);
4697
4698 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
4699 assert(in(RawAddress) == rawoop, "proper init");
4700 // Note: allocation() can be null, for secondary initialization barriers
4701 }
4702
4703 // Since this node is not matched, it will be processed by the
4704 // register allocator. Declare that there are no constraints
4705 // on the allocation of the RawAddress edge.
4706 const RegMask &InitializeNode::in_RegMask(uint idx) const {
4707 // This edge should be set to top, by the set_complete. But be conservative.
4708 if (idx == InitializeNode::RawAddress)
4709 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
4710 return RegMask::EMPTY;
4711 }
4712
4713 Node* InitializeNode::memory(uint alias_idx) {
4714 Node* mem = in(Memory);
4715 if (mem->is_MergeMem()) {
4716 return mem->as_MergeMem()->memory_at(alias_idx);
4717 } else {
4718 // incoming raw memory is not split
4719 return mem;
4720 }
4721 }
4722
4723 bool InitializeNode::is_non_zero() {
4724 if (is_complete()) return false;
4725 remove_extra_zeroes();
4726 return (req() > RawStores);
4727 }
4728
4729 void InitializeNode::set_complete(PhaseGVN* phase) {
4730 assert(!is_complete(), "caller responsibility");
4731 _is_complete = Complete;
4732
4733 // After this node is complete, it contains a bunch of
4734 // raw-memory initializations. There is no need for
4735 // it to have anything to do with non-raw memory effects.
4736 // Therefore, tell all non-raw users to re-optimize themselves,
4737 // after skipping the memory effects of this initialization.
4738 PhaseIterGVN* igvn = phase->is_IterGVN();
4739 if (igvn) igvn->add_users_to_worklist(this);
4740 }
4741
4742 // convenience function
4743 // return false if the init contains any stores already
4744 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
4745 InitializeNode* init = initialization();
4746 if (init == nullptr || init->is_complete()) return false;
4747 init->remove_extra_zeroes();
4748 // for now, if this allocation has already collected any inits, bail:
4749 if (init->is_non_zero()) return false;
4750 init->set_complete(phase);
4751 return true;
4752 }
4753
4754 void InitializeNode::remove_extra_zeroes() {
4755 if (req() == RawStores) return;
4756 Node* zmem = zero_memory();
4757 uint fill = RawStores;
4758 for (uint i = fill; i < req(); i++) {
4759 Node* n = in(i);
4760 if (n->is_top() || n == zmem) continue; // skip
4761 if (fill < i) set_req(fill, n); // compact
4762 ++fill;
4763 }
4764 // delete any empty spaces created:
4765 while (fill < req()) {
4766 del_req(fill);
4767 }
4768 }
4769
4770 // Helper for remembering which stores go with which offsets.
4771 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) {
4772 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
4773 intptr_t offset = -1;
4774 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
4775 phase, offset);
4776 if (base == nullptr) return -1; // something is dead,
4777 if (offset < 0) return -1; // dead, dead
4778 return offset;
4779 }
4780
4781 // Helper for proving that an initialization expression is
4782 // "simple enough" to be folded into an object initialization.
4783 // Attempts to prove that a store's initial value 'n' can be captured
4784 // within the initialization without creating a vicious cycle, such as:
4785 // { Foo p = new Foo(); p.next = p; }
4786 // True for constants and parameters and small combinations thereof.
4787 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
4788 ResourceMark rm;
4789 Unique_Node_List worklist;
4790 worklist.push(value);
4791
4792 uint complexity_limit = 20;
4793 for (uint j = 0; j < worklist.size(); j++) {
4794 if (j >= complexity_limit) {
4795 return false; // Bail out if processed too many nodes
4796 }
4797
4798 Node* n = worklist.at(j);
4799 if (n == nullptr) continue; // (can this really happen?)
4800 if (n->is_Proj()) n = n->in(0);
4801 if (n == this) return false; // found a cycle
4802 if (n->is_Con()) continue;
4803 if (n->is_Start()) continue; // params, etc., are OK
4804 if (n->is_Root()) continue; // even better
4805
4806 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
4807 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
4808 continue;
4809 }
4810
4811 Node* ctl = n->in(0);
4812 if (ctl != nullptr && !ctl->is_top()) {
4813 if (ctl->is_Proj()) ctl = ctl->in(0);
4814 if (ctl == this) return false;
4815
4816 // If we already know that the enclosing memory op is pinned right after
4817 // the init, then any control flow that the store has picked up
4818 // must have preceded the init, or else be equal to the init.
4819 // Even after loop optimizations (which might change control edges)
4820 // a store is never pinned *before* the availability of its inputs.
4821 if (!MemNode::all_controls_dominate(n, this)) {
4822 return false; // failed to prove a good control
4823 }
4824 }
4825
4826 // Check data edges for possible dependencies on 'this'.
4827 for (uint i = 1; i < n->req(); i++) {
4828 Node* m = n->in(i);
4829 if (m == nullptr || m == n || m->is_top()) continue;
4830
4831 // Only process data inputs once
4832 worklist.push(m);
4833 }
4834 }
4835
4836 return true;
4837 }
4838
4839 // Here are all the checks a Store must pass before it can be moved into
4840 // an initialization. Returns zero if a check fails.
4841 // On success, returns the (constant) offset to which the store applies,
4842 // within the initialized memory.
4843 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
4844 const int FAIL = 0;
4845 if (st->req() != MemNode::ValueIn + 1)
4846 return FAIL; // an inscrutable StoreNode (card mark?)
4847 Node* ctl = st->in(MemNode::Control);
4848 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this))
4849 return FAIL; // must be unconditional after the initialization
4850 Node* mem = st->in(MemNode::Memory);
4851 if (!(mem->is_Proj() && mem->in(0) == this))
4852 return FAIL; // must not be preceded by other stores
4853 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4854 if ((st->Opcode() == Op_StoreP || st->Opcode() == Op_StoreN) &&
4855 !bs->can_initialize_object(st)) {
4856 return FAIL;
4857 }
4858 Node* adr = st->in(MemNode::Address);
4859 intptr_t offset;
4860 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
4861 if (alloc == nullptr)
4862 return FAIL; // inscrutable address
4863 if (alloc != allocation())
4864 return FAIL; // wrong allocation! (store needs to float up)
4865 int size_in_bytes = st->memory_size();
4866 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
4867 return FAIL; // mismatched access
4868 }
4869 Node* val = st->in(MemNode::ValueIn);
4870
4871 if (!detect_init_independence(val, phase))
4872 return FAIL; // stored value must be 'simple enough'
4873
4874 // The Store can be captured only if nothing after the allocation
4875 // and before the Store is using the memory location that the store
4876 // overwrites.
4877 bool failed = false;
4878 // If is_complete_with_arraycopy() is true the shape of the graph is
4879 // well defined and is safe so no need for extra checks.
4880 if (!is_complete_with_arraycopy()) {
4881 // We are going to look at each use of the memory state following
4882 // the allocation to make sure nothing reads the memory that the
4883 // Store writes.
4884 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
4885 int alias_idx = phase->C->get_alias_index(t_adr);
4886 ResourceMark rm;
4887 Unique_Node_List mems;
4888 mems.push(mem);
4889 Node* unique_merge = nullptr;
4890 for (uint next = 0; next < mems.size(); ++next) {
4891 Node *m = mems.at(next);
4892 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
4893 Node *n = m->fast_out(j);
4894 if (n->outcnt() == 0) {
4895 continue;
4896 }
4897 if (n == st) {
4898 continue;
4899 } else if (n->in(0) != nullptr && n->in(0) != ctl) {
4900 // If the control of this use is different from the control
4901 // of the Store which is right after the InitializeNode then
4902 // this node cannot be between the InitializeNode and the
4903 // Store.
4904 continue;
4905 } else if (n->is_MergeMem()) {
4906 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
4907 // We can hit a MergeMemNode (that will likely go away
4908 // later) that is a direct use of the memory state
4909 // following the InitializeNode on the same slice as the
4910 // store node that we'd like to capture. We need to check
4911 // the uses of the MergeMemNode.
4912 mems.push(n);
4913 }
4914 } else if (n->is_Mem()) {
4915 Node* other_adr = n->in(MemNode::Address);
4916 if (other_adr == adr) {
4917 failed = true;
4918 break;
4919 } else {
4920 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
4921 if (other_t_adr != nullptr) {
4922 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
4923 if (other_alias_idx == alias_idx) {
4924 // A load from the same memory slice as the store right
4925 // after the InitializeNode. We check the control of the
4926 // object/array that is loaded from. If it's the same as
4927 // the store control then we cannot capture the store.
4928 assert(!n->is_Store(), "2 stores to same slice on same control?");
4929 Node* base = other_adr;
4930 assert(base->is_AddP(), "should be addp but is %s", base->Name());
4931 base = base->in(AddPNode::Base);
4932 if (base != nullptr) {
4933 base = base->uncast();
4934 if (base->is_Proj() && base->in(0) == alloc) {
4935 failed = true;
4936 break;
4937 }
4938 }
4939 }
4940 }
4941 }
4942 } else {
4943 failed = true;
4944 break;
4945 }
4946 }
4947 }
4948 }
4949 if (failed) {
4950 if (!can_reshape) {
4951 // We decided we couldn't capture the store during parsing. We
4952 // should try again during the next IGVN once the graph is
4953 // cleaner.
4954 phase->C->record_for_igvn(st);
4955 }
4956 return FAIL;
4957 }
4958
4959 return offset; // success
4960 }
4961
4962 // Find the captured store in(i) which corresponds to the range
4963 // [start..start+size) in the initialized object.
4964 // If there is one, return its index i. If there isn't, return the
4965 // negative of the index where it should be inserted.
4966 // Return 0 if the queried range overlaps an initialization boundary
4967 // or if dead code is encountered.
4968 // If size_in_bytes is zero, do not bother with overlap checks.
4969 int InitializeNode::captured_store_insertion_point(intptr_t start,
4970 int size_in_bytes,
4971 PhaseValues* phase) {
4972 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
4973
4974 if (is_complete())
4975 return FAIL; // arraycopy got here first; punt
4976
4977 assert(allocation() != nullptr, "must be present");
4978
4979 // no negatives, no header fields:
4980 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
4981
4982 // after a certain size, we bail out on tracking all the stores:
4983 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4984 if (start >= ti_limit) return FAIL;
4985
4986 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
4987 if (i >= limit) return -(int)i; // not found; here is where to put it
4988
4989 Node* st = in(i);
4990 intptr_t st_off = get_store_offset(st, phase);
4991 if (st_off < 0) {
4992 if (st != zero_memory()) {
4993 return FAIL; // bail out if there is dead garbage
4994 }
4995 } else if (st_off > start) {
4996 // ...we are done, since stores are ordered
4997 if (st_off < start + size_in_bytes) {
4998 return FAIL; // the next store overlaps
4999 }
5000 return -(int)i; // not found; here is where to put it
5001 } else if (st_off < start) {
5002 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
5003 if (size_in_bytes != 0 &&
5004 start < st_off + MAX_STORE &&
5005 start < st_off + st->as_Store()->memory_size()) {
5006 return FAIL; // the previous store overlaps
5007 }
5008 } else {
5009 if (size_in_bytes != 0 &&
5010 st->as_Store()->memory_size() != size_in_bytes) {
5011 return FAIL; // mismatched store size
5012 }
5013 return i;
5014 }
5015
5016 ++i;
5017 }
5018 }
5019
5020 // Look for a captured store which initializes at the offset 'start'
5021 // with the given size. If there is no such store, and no other
5022 // initialization interferes, then return zero_memory (the memory
5023 // projection of the AllocateNode).
5024 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
5025 PhaseValues* phase) {
5026 assert(stores_are_sane(phase), "");
5027 int i = captured_store_insertion_point(start, size_in_bytes, phase);
5028 if (i == 0) {
5029 return nullptr; // something is dead
5030 } else if (i < 0) {
5031 return zero_memory(); // just primordial zero bits here
5032 } else {
5033 Node* st = in(i); // here is the store at this position
5034 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
5035 return st;
5036 }
5037 }
5038
5039 // Create, as a raw pointer, an address within my new object at 'offset'.
5040 Node* InitializeNode::make_raw_address(intptr_t offset,
5041 PhaseGVN* phase) {
5042 Node* addr = in(RawAddress);
5043 if (offset != 0) {
5044 Compile* C = phase->C;
5045 addr = phase->transform( new AddPNode(C->top(), addr,
5046 phase->MakeConX(offset)) );
5047 }
5048 return addr;
5049 }
5050
5051 // Clone the given store, converting it into a raw store
5052 // initializing a field or element of my new object.
5053 // Caller is responsible for retiring the original store,
5054 // with subsume_node or the like.
5055 //
5056 // From the example above InitializeNode::InitializeNode,
5057 // here are the old stores to be captured:
5058 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
5059 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
5060 //
5061 // Here is the changed code; note the extra edges on init:
5062 // alloc = (Allocate ...)
5063 // rawoop = alloc.RawAddress
5064 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
5065 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
5066 // init = (Initialize alloc.Control alloc.Memory rawoop
5067 // rawstore1 rawstore2)
5068 //
5069 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
5070 PhaseGVN* phase, bool can_reshape) {
5071 assert(stores_are_sane(phase), "");
5072
5073 if (start < 0) return nullptr;
5074 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
5075
5076 Compile* C = phase->C;
5077 int size_in_bytes = st->memory_size();
5078 int i = captured_store_insertion_point(start, size_in_bytes, phase);
5079 if (i == 0) return nullptr; // bail out
5080 Node* prev_mem = nullptr; // raw memory for the captured store
5081 if (i > 0) {
5082 prev_mem = in(i); // there is a pre-existing store under this one
5083 set_req(i, C->top()); // temporarily disconnect it
5084 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
5085 } else {
5086 i = -i; // no pre-existing store
5087 prev_mem = zero_memory(); // a slice of the newly allocated object
5088 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
5089 set_req(--i, C->top()); // reuse this edge; it has been folded away
5090 else
5091 ins_req(i, C->top()); // build a new edge
5092 }
5093 Node* new_st = st->clone_with_adr_type(TypeRawPtr::BOTTOM);
5094 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
5095 new_st->set_req(MemNode::Control, in(Control));
5096 new_st->set_req(MemNode::Memory, prev_mem);
5097 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
5098 bs->eliminate_gc_barrier_data(new_st);
5099 new_st = phase->transform(new_st);
5100
5101 // At this point, new_st might have swallowed a pre-existing store
5102 // at the same offset, or perhaps new_st might have disappeared,
5103 // if it redundantly stored the same value (or zero to fresh memory).
5104
5105 // In any case, wire it in:
5106 PhaseIterGVN* igvn = phase->is_IterGVN();
5107 if (igvn) {
5108 igvn->rehash_node_delayed(this);
5109 }
5110 set_req(i, new_st);
5111
5112 // The caller may now kill the old guy.
5113 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
5114 assert(check_st == new_st || check_st == nullptr, "must be findable");
5115 assert(!is_complete(), "");
5116 return new_st;
5117 }
5118
5119 static bool store_constant(jlong* tiles, int num_tiles,
5120 intptr_t st_off, int st_size,
5121 jlong con) {
5122 if ((st_off & (st_size-1)) != 0)
5123 return false; // strange store offset (assume size==2**N)
5124 address addr = (address)tiles + st_off;
5125 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
5126 switch (st_size) {
5127 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
5128 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
5129 case sizeof(jint): *(jint*) addr = (jint) con; break;
5130 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
5131 default: return false; // strange store size (detect size!=2**N here)
5132 }
5133 return true; // return success to caller
5134 }
5135
5136 // Coalesce subword constants into int constants and possibly
5137 // into long constants. The goal, if the CPU permits,
5138 // is to initialize the object with a small number of 64-bit tiles.
5139 // Also, convert floating-point constants to bit patterns.
5140 // Non-constants are not relevant to this pass.
5141 //
5142 // In terms of the running example on InitializeNode::InitializeNode
5143 // and InitializeNode::capture_store, here is the transformation
5144 // of rawstore1 and rawstore2 into rawstore12:
5145 // alloc = (Allocate ...)
5146 // rawoop = alloc.RawAddress
5147 // tile12 = 0x00010002
5148 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
5149 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
5150 //
5151 void
5152 InitializeNode::coalesce_subword_stores(intptr_t header_size,
5153 Node* size_in_bytes,
5154 PhaseGVN* phase) {
5155 Compile* C = phase->C;
5156
5157 assert(stores_are_sane(phase), "");
5158 // Note: After this pass, they are not completely sane,
5159 // since there may be some overlaps.
5160
5161 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
5162
5163 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
5164 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
5165 size_limit = MIN2(size_limit, ti_limit);
5166 size_limit = align_up(size_limit, BytesPerLong);
5167 int num_tiles = size_limit / BytesPerLong;
5168
5169 // allocate space for the tile map:
5170 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
5171 jlong tiles_buf[small_len];
5172 Node* nodes_buf[small_len];
5173 jlong inits_buf[small_len];
5174 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
5175 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5176 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
5177 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
5178 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
5179 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5180 // tiles: exact bitwise model of all primitive constants
5181 // nodes: last constant-storing node subsumed into the tiles model
5182 // inits: which bytes (in each tile) are touched by any initializations
5183
5184 //// Pass A: Fill in the tile model with any relevant stores.
5185
5186 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
5187 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
5188 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
5189 Node* zmem = zero_memory(); // initially zero memory state
5190 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5191 Node* st = in(i);
5192 intptr_t st_off = get_store_offset(st, phase);
5193
5194 // Figure out the store's offset and constant value:
5195 if (st_off < header_size) continue; //skip (ignore header)
5196 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
5197 int st_size = st->as_Store()->memory_size();
5198 if (st_off + st_size > size_limit) break;
5199
5200 // Record which bytes are touched, whether by constant or not.
5201 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
5202 continue; // skip (strange store size)
5203
5204 const Type* val = phase->type(st->in(MemNode::ValueIn));
5205 if (!val->singleton()) continue; //skip (non-con store)
5206 BasicType type = val->basic_type();
5207
5208 jlong con = 0;
5209 switch (type) {
5210 case T_INT: con = val->is_int()->get_con(); break;
5211 case T_LONG: con = val->is_long()->get_con(); break;
5212 case T_FLOAT: con = jint_cast(val->getf()); break;
5213 case T_DOUBLE: con = jlong_cast(val->getd()); break;
5214 default: continue; //skip (odd store type)
5215 }
5216
5217 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
5218 st->Opcode() == Op_StoreL) {
5219 continue; // This StoreL is already optimal.
5220 }
5221
5222 // Store down the constant.
5223 store_constant(tiles, num_tiles, st_off, st_size, con);
5224
5225 intptr_t j = st_off >> LogBytesPerLong;
5226
5227 if (type == T_INT && st_size == BytesPerInt
5228 && (st_off & BytesPerInt) == BytesPerInt) {
5229 jlong lcon = tiles[j];
5230 if (!Matcher::isSimpleConstant64(lcon) &&
5231 st->Opcode() == Op_StoreI) {
5232 // This StoreI is already optimal by itself.
5233 jint* intcon = (jint*) &tiles[j];
5234 intcon[1] = 0; // undo the store_constant()
5235
5236 // If the previous store is also optimal by itself, back up and
5237 // undo the action of the previous loop iteration... if we can.
5238 // But if we can't, just let the previous half take care of itself.
5239 st = nodes[j];
5240 st_off -= BytesPerInt;
5241 con = intcon[0];
5242 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) {
5243 assert(st_off >= header_size, "still ignoring header");
5244 assert(get_store_offset(st, phase) == st_off, "must be");
5245 assert(in(i-1) == zmem, "must be");
5246 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
5247 assert(con == tcon->is_int()->get_con(), "must be");
5248 // Undo the effects of the previous loop trip, which swallowed st:
5249 intcon[0] = 0; // undo store_constant()
5250 set_req(i-1, st); // undo set_req(i, zmem)
5251 nodes[j] = nullptr; // undo nodes[j] = st
5252 --old_subword; // undo ++old_subword
5253 }
5254 continue; // This StoreI is already optimal.
5255 }
5256 }
5257
5258 // This store is not needed.
5259 set_req(i, zmem);
5260 nodes[j] = st; // record for the moment
5261 if (st_size < BytesPerLong) // something has changed
5262 ++old_subword; // includes int/float, but who's counting...
5263 else ++old_long;
5264 }
5265
5266 if ((old_subword + old_long) == 0)
5267 return; // nothing more to do
5268
5269 //// Pass B: Convert any non-zero tiles into optimal constant stores.
5270 // Be sure to insert them before overlapping non-constant stores.
5271 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
5272 for (int j = 0; j < num_tiles; j++) {
5273 jlong con = tiles[j];
5274 jlong init = inits[j];
5275 if (con == 0) continue;
5276 jint con0, con1; // split the constant, address-wise
5277 jint init0, init1; // split the init map, address-wise
5278 { union { jlong con; jint intcon[2]; } u;
5279 u.con = con;
5280 con0 = u.intcon[0];
5281 con1 = u.intcon[1];
5282 u.con = init;
5283 init0 = u.intcon[0];
5284 init1 = u.intcon[1];
5285 }
5286
5287 Node* old = nodes[j];
5288 assert(old != nullptr, "need the prior store");
5289 intptr_t offset = (j * BytesPerLong);
5290
5291 bool split = !Matcher::isSimpleConstant64(con);
5292
5293 if (offset < header_size) {
5294 assert(offset + BytesPerInt >= header_size, "second int counts");
5295 assert(*(jint*)&tiles[j] == 0, "junk in header");
5296 split = true; // only the second word counts
5297 // Example: int a[] = { 42 ... }
5298 } else if (con0 == 0 && init0 == -1) {
5299 split = true; // first word is covered by full inits
5300 // Example: int a[] = { ... foo(), 42 ... }
5301 } else if (con1 == 0 && init1 == -1) {
5302 split = true; // second word is covered by full inits
5303 // Example: int a[] = { ... 42, foo() ... }
5304 }
5305
5306 // Here's a case where init0 is neither 0 nor -1:
5307 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
5308 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
5309 // In this case the tile is not split; it is (jlong)42.
5310 // The big tile is stored down, and then the foo() value is inserted.
5311 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
5312
5313 Node* ctl = old->in(MemNode::Control);
5314 Node* adr = make_raw_address(offset, phase);
5315 const TypePtr* atp = TypeRawPtr::BOTTOM;
5316
5317 // One or two coalesced stores to plop down.
5318 Node* st[2];
5319 intptr_t off[2];
5320 int nst = 0;
5321 if (!split) {
5322 ++new_long;
5323 off[nst] = offset;
5324 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5325 phase->longcon(con), T_LONG, MemNode::unordered);
5326 } else {
5327 // Omit either if it is a zero.
5328 if (con0 != 0) {
5329 ++new_int;
5330 off[nst] = offset;
5331 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5332 phase->intcon(con0), T_INT, MemNode::unordered);
5333 }
5334 if (con1 != 0) {
5335 ++new_int;
5336 offset += BytesPerInt;
5337 adr = make_raw_address(offset, phase);
5338 off[nst] = offset;
5339 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5340 phase->intcon(con1), T_INT, MemNode::unordered);
5341 }
5342 }
5343
5344 // Insert second store first, then the first before the second.
5345 // Insert each one just before any overlapping non-constant stores.
5346 while (nst > 0) {
5347 Node* st1 = st[--nst];
5348 C->copy_node_notes_to(st1, old);
5349 st1 = phase->transform(st1);
5350 offset = off[nst];
5351 assert(offset >= header_size, "do not smash header");
5352 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
5353 guarantee(ins_idx != 0, "must re-insert constant store");
5354 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
5355 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
5356 set_req(--ins_idx, st1);
5357 else
5358 ins_req(ins_idx, st1);
5359 }
5360 }
5361
5362 if (PrintCompilation && WizardMode)
5363 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
5364 old_subword, old_long, new_int, new_long);
5365 if (C->log() != nullptr)
5366 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
5367 old_subword, old_long, new_int, new_long);
5368
5369 // Clean up any remaining occurrences of zmem:
5370 remove_extra_zeroes();
5371 }
5372
5373 // Explore forward from in(start) to find the first fully initialized
5374 // word, and return its offset. Skip groups of subword stores which
5375 // together initialize full words. If in(start) is itself part of a
5376 // fully initialized word, return the offset of in(start). If there
5377 // are no following full-word stores, or if something is fishy, return
5378 // a negative value.
5379 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
5380 int int_map = 0;
5381 intptr_t int_map_off = 0;
5382 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
5383
5384 for (uint i = start, limit = req(); i < limit; i++) {
5385 Node* st = in(i);
5386
5387 intptr_t st_off = get_store_offset(st, phase);
5388 if (st_off < 0) break; // return conservative answer
5389
5390 int st_size = st->as_Store()->memory_size();
5391 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
5392 return st_off; // we found a complete word init
5393 }
5394
5395 // update the map:
5396
5397 intptr_t this_int_off = align_down(st_off, BytesPerInt);
5398 if (this_int_off != int_map_off) {
5399 // reset the map:
5400 int_map = 0;
5401 int_map_off = this_int_off;
5402 }
5403
5404 int subword_off = st_off - this_int_off;
5405 int_map |= right_n_bits(st_size) << subword_off;
5406 if ((int_map & FULL_MAP) == FULL_MAP) {
5407 return this_int_off; // we found a complete word init
5408 }
5409
5410 // Did this store hit or cross the word boundary?
5411 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
5412 if (next_int_off == this_int_off + BytesPerInt) {
5413 // We passed the current int, without fully initializing it.
5414 int_map_off = next_int_off;
5415 int_map >>= BytesPerInt;
5416 } else if (next_int_off > this_int_off + BytesPerInt) {
5417 // We passed the current and next int.
5418 return this_int_off + BytesPerInt;
5419 }
5420 }
5421
5422 return -1;
5423 }
5424
5425
5426 // Called when the associated AllocateNode is expanded into CFG.
5427 // At this point, we may perform additional optimizations.
5428 // Linearize the stores by ascending offset, to make memory
5429 // activity as coherent as possible.
5430 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
5431 intptr_t header_size,
5432 Node* size_in_bytes,
5433 PhaseIterGVN* phase) {
5434 assert(!is_complete(), "not already complete");
5435 assert(stores_are_sane(phase), "");
5436 assert(allocation() != nullptr, "must be present");
5437
5438 remove_extra_zeroes();
5439
5440 if (ReduceFieldZeroing || ReduceBulkZeroing)
5441 // reduce instruction count for common initialization patterns
5442 coalesce_subword_stores(header_size, size_in_bytes, phase);
5443
5444 Node* zmem = zero_memory(); // initially zero memory state
5445 Node* inits = zmem; // accumulating a linearized chain of inits
5446 #ifdef ASSERT
5447 intptr_t first_offset = allocation()->minimum_header_size();
5448 intptr_t last_init_off = first_offset; // previous init offset
5449 intptr_t last_init_end = first_offset; // previous init offset+size
5450 intptr_t last_tile_end = first_offset; // previous tile offset+size
5451 #endif
5452 intptr_t zeroes_done = header_size;
5453
5454 bool do_zeroing = true; // we might give up if inits are very sparse
5455 int big_init_gaps = 0; // how many large gaps have we seen?
5456
5457 if (UseTLAB && ZeroTLAB) do_zeroing = false;
5458 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
5459
5460 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5461 Node* st = in(i);
5462 intptr_t st_off = get_store_offset(st, phase);
5463 if (st_off < 0)
5464 break; // unknown junk in the inits
5465 if (st->in(MemNode::Memory) != zmem)
5466 break; // complicated store chains somehow in list
5467
5468 int st_size = st->as_Store()->memory_size();
5469 intptr_t next_init_off = st_off + st_size;
5470
5471 if (do_zeroing && zeroes_done < next_init_off) {
5472 // See if this store needs a zero before it or under it.
5473 intptr_t zeroes_needed = st_off;
5474
5475 if (st_size < BytesPerInt) {
5476 // Look for subword stores which only partially initialize words.
5477 // If we find some, we must lay down some word-level zeroes first,
5478 // underneath the subword stores.
5479 //
5480 // Examples:
5481 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
5482 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
5483 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
5484 //
5485 // Note: coalesce_subword_stores may have already done this,
5486 // if it was prompted by constant non-zero subword initializers.
5487 // But this case can still arise with non-constant stores.
5488
5489 intptr_t next_full_store = find_next_fullword_store(i, phase);
5490
5491 // In the examples above:
5492 // in(i) p q r s x y z
5493 // st_off 12 13 14 15 12 13 14
5494 // st_size 1 1 1 1 1 1 1
5495 // next_full_s. 12 16 16 16 16 16 16
5496 // z's_done 12 16 16 16 12 16 12
5497 // z's_needed 12 16 16 16 16 16 16
5498 // zsize 0 0 0 0 4 0 4
5499 if (next_full_store < 0) {
5500 // Conservative tack: Zero to end of current word.
5501 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
5502 } else {
5503 // Zero to beginning of next fully initialized word.
5504 // Or, don't zero at all, if we are already in that word.
5505 assert(next_full_store >= zeroes_needed, "must go forward");
5506 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
5507 zeroes_needed = next_full_store;
5508 }
5509 }
5510
5511 if (zeroes_needed > zeroes_done) {
5512 intptr_t zsize = zeroes_needed - zeroes_done;
5513 // Do some incremental zeroing on rawmem, in parallel with inits.
5514 zeroes_done = align_down(zeroes_done, BytesPerInt);
5515 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5516 zeroes_done, zeroes_needed,
5517 true,
5518 phase);
5519 zeroes_done = zeroes_needed;
5520 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
5521 do_zeroing = false; // leave the hole, next time
5522 }
5523 }
5524
5525 // Collect the store and move on:
5526 phase->replace_input_of(st, MemNode::Memory, inits);
5527 inits = st; // put it on the linearized chain
5528 set_req(i, zmem); // unhook from previous position
5529
5530 if (zeroes_done == st_off)
5531 zeroes_done = next_init_off;
5532
5533 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
5534
5535 #ifdef ASSERT
5536 // Various order invariants. Weaker than stores_are_sane because
5537 // a large constant tile can be filled in by smaller non-constant stores.
5538 assert(st_off >= last_init_off, "inits do not reverse");
5539 last_init_off = st_off;
5540 const Type* val = nullptr;
5541 if (st_size >= BytesPerInt &&
5542 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
5543 (int)val->basic_type() < (int)T_OBJECT) {
5544 assert(st_off >= last_tile_end, "tiles do not overlap");
5545 assert(st_off >= last_init_end, "tiles do not overwrite inits");
5546 last_tile_end = MAX2(last_tile_end, next_init_off);
5547 } else {
5548 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
5549 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
5550 assert(st_off >= last_init_end, "inits do not overlap");
5551 last_init_end = next_init_off; // it's a non-tile
5552 }
5553 #endif //ASSERT
5554 }
5555
5556 remove_extra_zeroes(); // clear out all the zmems left over
5557 add_req(inits);
5558
5559 if (!(UseTLAB && ZeroTLAB)) {
5560 // If anything remains to be zeroed, zero it all now.
5561 zeroes_done = align_down(zeroes_done, BytesPerInt);
5562 // if it is the last unused 4 bytes of an instance, forget about it
5563 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
5564 if (zeroes_done + BytesPerLong >= size_limit) {
5565 AllocateNode* alloc = allocation();
5566 assert(alloc != nullptr, "must be present");
5567 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
5568 Node* klass_node = alloc->in(AllocateNode::KlassNode);
5569 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
5570 if (zeroes_done == k->layout_helper())
5571 zeroes_done = size_limit;
5572 }
5573 }
5574 if (zeroes_done < size_limit) {
5575 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5576 zeroes_done, size_in_bytes, true, phase);
5577 }
5578 }
5579
5580 set_complete(phase);
5581 return rawmem;
5582 }
5583
5584 void InitializeNode::replace_mem_projs_by(Node* mem, Compile* C) {
5585 auto replace_proj = [&](ProjNode* proj) {
5586 C->gvn_replace_by(proj, mem);
5587 return CONTINUE;
5588 };
5589 apply_to_projs(replace_proj, TypeFunc::Memory);
5590 }
5591
5592 void InitializeNode::replace_mem_projs_by(Node* mem, PhaseIterGVN* igvn) {
5593 DUIterator_Fast imax, i = fast_outs(imax);
5594 auto replace_proj = [&](ProjNode* proj) {
5595 igvn->replace_node(proj, mem);
5596 --i; --imax;
5597 return CONTINUE;
5598 };
5599 apply_to_projs(imax, i, replace_proj, TypeFunc::Memory);
5600 }
5601
5602 bool InitializeNode::already_has_narrow_mem_proj_with_adr_type(const TypePtr* adr_type) const {
5603 auto find_proj = [&](ProjNode* proj) {
5604 if (proj->adr_type() == adr_type) {
5605 return BREAK_AND_RETURN_CURRENT_PROJ;
5606 }
5607 return CONTINUE;
5608 };
5609 DUIterator_Fast imax, i = fast_outs(imax);
5610 return apply_to_narrow_mem_projs_any_iterator(UsesIteratorFast(imax, i, this), find_proj) != nullptr;
5611 }
5612
5613 MachProjNode* InitializeNode::mem_mach_proj() const {
5614 auto find_proj = [](ProjNode* proj) {
5615 if (proj->is_MachProj()) {
5616 return BREAK_AND_RETURN_CURRENT_PROJ;
5617 }
5618 return CONTINUE;
5619 };
5620 ProjNode* proj = apply_to_projs(find_proj, TypeFunc::Memory);
5621 if (proj == nullptr) {
5622 return nullptr;
5623 }
5624 return proj->as_MachProj();
5625 }
5626
5627 #ifdef ASSERT
5628 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
5629 if (is_complete())
5630 return true; // stores could be anything at this point
5631 assert(allocation() != nullptr, "must be present");
5632 intptr_t last_off = allocation()->minimum_header_size();
5633 for (uint i = InitializeNode::RawStores; i < req(); i++) {
5634 Node* st = in(i);
5635 intptr_t st_off = get_store_offset(st, phase);
5636 if (st_off < 0) continue; // ignore dead garbage
5637 if (last_off > st_off) {
5638 tty->print_cr("*** bad store offset at %d: %zd > %zd", i, last_off, st_off);
5639 this->dump(2);
5640 assert(false, "ascending store offsets");
5641 return false;
5642 }
5643 last_off = st_off + st->as_Store()->memory_size();
5644 }
5645 return true;
5646 }
5647 #endif //ASSERT
5648
5649
5650
5651
5652 //============================MergeMemNode=====================================
5653 //
5654 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
5655 // contributing store or call operations. Each contributor provides the memory
5656 // state for a particular "alias type" (see Compile::alias_type). For example,
5657 // if a MergeMem has an input X for alias category #6, then any memory reference
5658 // to alias category #6 may use X as its memory state input, as an exact equivalent
5659 // to using the MergeMem as a whole.
5660 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
5661 //
5662 // (Here, the <N> notation gives the index of the relevant adr_type.)
5663 //
5664 // In one special case (and more cases in the future), alias categories overlap.
5665 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
5666 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
5667 // it is exactly equivalent to that state W:
5668 // MergeMem(<Bot>: W) <==> W
5669 //
5670 // Usually, the merge has more than one input. In that case, where inputs
5671 // overlap (i.e., one is Bot), the narrower alias type determines the memory
5672 // state for that type, and the wider alias type (Bot) fills in everywhere else:
5673 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
5674 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
5675 //
5676 // A merge can take a "wide" memory state as one of its narrow inputs.
5677 // This simply means that the merge observes out only the relevant parts of
5678 // the wide input. That is, wide memory states arriving at narrow merge inputs
5679 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
5680 //
5681 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
5682 // and that memory slices "leak through":
5683 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
5684 //
5685 // But, in such a cascade, repeated memory slices can "block the leak":
5686 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
5687 //
5688 // In the last example, Y is not part of the combined memory state of the
5689 // outermost MergeMem. The system must, of course, prevent unschedulable
5690 // memory states from arising, so you can be sure that the state Y is somehow
5691 // a precursor to state Y'.
5692 //
5693 //
5694 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
5695 // of each MergeMemNode array are exactly the numerical alias indexes, including
5696 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
5697 // Compile::alias_type (and kin) produce and manage these indexes.
5698 //
5699 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
5700 // (Note that this provides quick access to the top node inside MergeMem methods,
5701 // without the need to reach out via TLS to Compile::current.)
5702 //
5703 // As a consequence of what was just described, a MergeMem that represents a full
5704 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
5705 // containing all alias categories.
5706 //
5707 // MergeMem nodes never (?) have control inputs, so in(0) is null.
5708 //
5709 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
5710 // a memory state for the alias type <N>, or else the top node, meaning that
5711 // there is no particular input for that alias type. Note that the length of
5712 // a MergeMem is variable, and may be extended at any time to accommodate new
5713 // memory states at larger alias indexes. When merges grow, they are of course
5714 // filled with "top" in the unused in() positions.
5715 //
5716 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
5717 // (Top was chosen because it works smoothly with passes like GCM.)
5718 //
5719 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
5720 // the type of random VM bits like TLS references.) Since it is always the
5721 // first non-Bot memory slice, some low-level loops use it to initialize an
5722 // index variable: for (i = AliasIdxRaw; i < req(); i++).
5723 //
5724 //
5725 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
5726 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
5727 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
5728 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
5729 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
5730 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
5731 //
5732 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
5733 // really that different from the other memory inputs. An abbreviation called
5734 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
5735 //
5736 //
5737 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
5738 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
5739 // that "emerges though" the base memory will be marked as excluding the alias types
5740 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
5741 //
5742 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
5743 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
5744 //
5745 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
5746 // (It is currently unimplemented.) As you can see, the resulting merge is
5747 // actually a disjoint union of memory states, rather than an overlay.
5748 //
5749
5750 //------------------------------MergeMemNode-----------------------------------
5751 Node* MergeMemNode::make_empty_memory() {
5752 Node* empty_memory = (Node*) Compile::current()->top();
5753 assert(empty_memory->is_top(), "correct sentinel identity");
5754 return empty_memory;
5755 }
5756
5757 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
5758 init_class_id(Class_MergeMem);
5759 // all inputs are nullified in Node::Node(int)
5760 // set_input(0, nullptr); // no control input
5761
5762 // Initialize the edges uniformly to top, for starters.
5763 Node* empty_mem = make_empty_memory();
5764 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
5765 init_req(i,empty_mem);
5766 }
5767 assert(empty_memory() == empty_mem, "");
5768
5769 if( new_base != nullptr && new_base->is_MergeMem() ) {
5770 MergeMemNode* mdef = new_base->as_MergeMem();
5771 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
5772 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
5773 mms.set_memory(mms.memory2());
5774 }
5775 assert(base_memory() == mdef->base_memory(), "");
5776 } else {
5777 set_base_memory(new_base);
5778 }
5779 }
5780
5781 // Make a new, untransformed MergeMem with the same base as 'mem'.
5782 // If mem is itself a MergeMem, populate the result with the same edges.
5783 MergeMemNode* MergeMemNode::make(Node* mem) {
5784 return new MergeMemNode(mem);
5785 }
5786
5787 //------------------------------cmp--------------------------------------------
5788 uint MergeMemNode::hash() const { return NO_HASH; }
5789 bool MergeMemNode::cmp( const Node &n ) const {
5790 return (&n == this); // Always fail except on self
5791 }
5792
5793 //------------------------------Identity---------------------------------------
5794 Node* MergeMemNode::Identity(PhaseGVN* phase) {
5795 // Identity if this merge point does not record any interesting memory
5796 // disambiguations.
5797 Node* base_mem = base_memory();
5798 Node* empty_mem = empty_memory();
5799 if (base_mem != empty_mem) { // Memory path is not dead?
5800 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5801 Node* mem = in(i);
5802 if (mem != empty_mem && mem != base_mem) {
5803 return this; // Many memory splits; no change
5804 }
5805 }
5806 }
5807 return base_mem; // No memory splits; ID on the one true input
5808 }
5809
5810 //------------------------------Ideal------------------------------------------
5811 // This method is invoked recursively on chains of MergeMem nodes
5812 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
5813 // Remove chain'd MergeMems
5814 //
5815 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
5816 // relative to the "in(Bot)". Since we are patching both at the same time,
5817 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
5818 // but rewrite each "in(i)" relative to the new "in(Bot)".
5819 Node *progress = nullptr;
5820
5821
5822 Node* old_base = base_memory();
5823 Node* empty_mem = empty_memory();
5824 if (old_base == empty_mem)
5825 return nullptr; // Dead memory path.
5826
5827 MergeMemNode* old_mbase;
5828 if (old_base != nullptr && old_base->is_MergeMem())
5829 old_mbase = old_base->as_MergeMem();
5830 else
5831 old_mbase = nullptr;
5832 Node* new_base = old_base;
5833
5834 // simplify stacked MergeMems in base memory
5835 if (old_mbase) new_base = old_mbase->base_memory();
5836
5837 // the base memory might contribute new slices beyond my req()
5838 if (old_mbase) grow_to_match(old_mbase);
5839
5840 // Note: We do not call verify_sparse on entry, because inputs
5841 // can normalize to the base_memory via subsume_node or similar
5842 // mechanisms. This method repairs that damage.
5843
5844 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
5845
5846 // Look at each slice.
5847 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5848 Node* old_in = in(i);
5849 // calculate the old memory value
5850 Node* old_mem = old_in;
5851 if (old_mem == empty_mem) old_mem = old_base;
5852 assert(old_mem == memory_at(i), "");
5853
5854 // maybe update (reslice) the old memory value
5855
5856 // simplify stacked MergeMems
5857 Node* new_mem = old_mem;
5858 MergeMemNode* old_mmem;
5859 if (old_mem != nullptr && old_mem->is_MergeMem())
5860 old_mmem = old_mem->as_MergeMem();
5861 else
5862 old_mmem = nullptr;
5863 if (old_mmem == this) {
5864 // This can happen if loops break up and safepoints disappear.
5865 // A merge of BotPtr (default) with a RawPtr memory derived from a
5866 // safepoint can be rewritten to a merge of the same BotPtr with
5867 // the BotPtr phi coming into the loop. If that phi disappears
5868 // also, we can end up with a self-loop of the mergemem.
5869 // In general, if loops degenerate and memory effects disappear,
5870 // a mergemem can be left looking at itself. This simply means
5871 // that the mergemem's default should be used, since there is
5872 // no longer any apparent effect on this slice.
5873 // Note: If a memory slice is a MergeMem cycle, it is unreachable
5874 // from start. Update the input to TOP.
5875 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
5876 }
5877 else if (old_mmem != nullptr) {
5878 new_mem = old_mmem->memory_at(i);
5879 }
5880 // else preceding memory was not a MergeMem
5881
5882 // maybe store down a new value
5883 Node* new_in = new_mem;
5884 if (new_in == new_base) new_in = empty_mem;
5885
5886 if (new_in != old_in) {
5887 // Warning: Do not combine this "if" with the previous "if"
5888 // A memory slice might have be be rewritten even if it is semantically
5889 // unchanged, if the base_memory value has changed.
5890 set_req_X(i, new_in, phase);
5891 progress = this; // Report progress
5892 }
5893 }
5894
5895 if (new_base != old_base) {
5896 set_req_X(Compile::AliasIdxBot, new_base, phase);
5897 // Don't use set_base_memory(new_base), because we need to update du.
5898 assert(base_memory() == new_base, "");
5899 progress = this;
5900 }
5901
5902 if( base_memory() == this ) {
5903 // a self cycle indicates this memory path is dead
5904 set_req(Compile::AliasIdxBot, empty_mem);
5905 }
5906
5907 // Resolve external cycles by calling Ideal on a MergeMem base_memory
5908 // Recursion must occur after the self cycle check above
5909 if( base_memory()->is_MergeMem() ) {
5910 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
5911 Node *m = phase->transform(new_mbase); // Rollup any cycles
5912 if( m != nullptr &&
5913 (m->is_top() ||
5914 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
5915 // propagate rollup of dead cycle to self
5916 set_req(Compile::AliasIdxBot, empty_mem);
5917 }
5918 }
5919
5920 if( base_memory() == empty_mem ) {
5921 progress = this;
5922 // Cut inputs during Parse phase only.
5923 // During Optimize phase a dead MergeMem node will be subsumed by Top.
5924 if( !can_reshape ) {
5925 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5926 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
5927 }
5928 }
5929 }
5930
5931 if( !progress && base_memory()->is_Phi() && can_reshape ) {
5932 // Check if PhiNode::Ideal's "Split phis through memory merges"
5933 // transform should be attempted. Look for this->phi->this cycle.
5934 uint merge_width = req();
5935 if (merge_width > Compile::AliasIdxRaw) {
5936 PhiNode* phi = base_memory()->as_Phi();
5937 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
5938 if (phi->in(i) == this) {
5939 phase->is_IterGVN()->_worklist.push(phi);
5940 break;
5941 }
5942 }
5943 }
5944 }
5945
5946 assert(progress || verify_sparse(), "please, no dups of base");
5947 return progress;
5948 }
5949
5950 //-------------------------set_base_memory-------------------------------------
5951 void MergeMemNode::set_base_memory(Node *new_base) {
5952 Node* empty_mem = empty_memory();
5953 set_req(Compile::AliasIdxBot, new_base);
5954 assert(memory_at(req()) == new_base, "must set default memory");
5955 // Clear out other occurrences of new_base:
5956 if (new_base != empty_mem) {
5957 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5958 if (in(i) == new_base) set_req(i, empty_mem);
5959 }
5960 }
5961 }
5962
5963 //------------------------------out_RegMask------------------------------------
5964 const RegMask &MergeMemNode::out_RegMask() const {
5965 return RegMask::EMPTY;
5966 }
5967
5968 //------------------------------dump_spec--------------------------------------
5969 #ifndef PRODUCT
5970 void MergeMemNode::dump_spec(outputStream *st) const {
5971 st->print(" {");
5972 Node* base_mem = base_memory();
5973 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
5974 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem;
5975 if (mem == base_mem) { st->print(" -"); continue; }
5976 st->print( " N%d:", mem->_idx );
5977 Compile::current()->get_adr_type(i)->dump_on(st);
5978 }
5979 st->print(" }");
5980 }
5981 #endif // !PRODUCT
5982
5983
5984 #ifdef ASSERT
5985 static bool might_be_same(Node* a, Node* b) {
5986 if (a == b) return true;
5987 if (!(a->is_Phi() || b->is_Phi())) return false;
5988 // phis shift around during optimization
5989 return true; // pretty stupid...
5990 }
5991
5992 // verify a narrow slice (either incoming or outgoing)
5993 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
5994 if (!VerifyAliases) return; // don't bother to verify unless requested
5995 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
5996 if (Node::in_dump()) return; // muzzle asserts when printing
5997 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
5998 assert(n != nullptr, "");
5999 // Elide intervening MergeMem's
6000 while (n->is_MergeMem()) {
6001 n = n->as_MergeMem()->memory_at(alias_idx);
6002 }
6003 Compile* C = Compile::current();
6004 const TypePtr* n_adr_type = n->adr_type();
6005 if (n == m->empty_memory()) {
6006 // Implicit copy of base_memory()
6007 } else if (n_adr_type != TypePtr::BOTTOM) {
6008 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type");
6009 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
6010 } else {
6011 // A few places like make_runtime_call "know" that VM calls are narrow,
6012 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
6013 bool expected_wide_mem = false;
6014 if (n == m->base_memory()) {
6015 expected_wide_mem = true;
6016 } else if (alias_idx == Compile::AliasIdxRaw ||
6017 n == m->memory_at(Compile::AliasIdxRaw)) {
6018 expected_wide_mem = true;
6019 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
6020 // memory can "leak through" calls on channels that
6021 // are write-once. Allow this also.
6022 expected_wide_mem = true;
6023 }
6024 assert(expected_wide_mem, "expected narrow slice replacement");
6025 }
6026 }
6027 #else // !ASSERT
6028 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
6029 #endif
6030
6031
6032 //-----------------------------memory_at---------------------------------------
6033 Node* MergeMemNode::memory_at(uint alias_idx) const {
6034 assert(alias_idx >= Compile::AliasIdxRaw ||
6035 (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()),
6036 "must avoid base_memory and AliasIdxTop");
6037
6038 // Otherwise, it is a narrow slice.
6039 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
6040 if (is_empty_memory(n)) {
6041 // the array is sparse; empty slots are the "top" node
6042 n = base_memory();
6043 assert(Node::in_dump()
6044 || n == nullptr || n->bottom_type() == Type::TOP
6045 || n->adr_type() == nullptr // address is TOP
6046 || n->adr_type() == TypePtr::BOTTOM
6047 || n->adr_type() == TypeRawPtr::BOTTOM
6048 || n->is_NarrowMemProj()
6049 || !Compile::current()->do_aliasing(),
6050 "must be a wide memory");
6051 // do_aliasing == false if we are organizing the memory states manually.
6052 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
6053 } else {
6054 // make sure the stored slice is sane
6055 #ifdef ASSERT
6056 if (VMError::is_error_reported() || Node::in_dump()) {
6057 } else if (might_be_same(n, base_memory())) {
6058 // Give it a pass: It is a mostly harmless repetition of the base.
6059 // This can arise normally from node subsumption during optimization.
6060 } else {
6061 verify_memory_slice(this, alias_idx, n);
6062 }
6063 #endif
6064 }
6065 return n;
6066 }
6067
6068 //---------------------------set_memory_at-------------------------------------
6069 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
6070 verify_memory_slice(this, alias_idx, n);
6071 Node* empty_mem = empty_memory();
6072 if (n == base_memory()) n = empty_mem; // collapse default
6073 uint need_req = alias_idx+1;
6074 if (req() < need_req) {
6075 if (n == empty_mem) return; // already the default, so do not grow me
6076 // grow the sparse array
6077 do {
6078 add_req(empty_mem);
6079 } while (req() < need_req);
6080 }
6081 set_req( alias_idx, n );
6082 }
6083
6084
6085
6086 //--------------------------iteration_setup------------------------------------
6087 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
6088 if (other != nullptr) {
6089 grow_to_match(other);
6090 // invariant: the finite support of mm2 is within mm->req()
6091 #ifdef ASSERT
6092 for (uint i = req(); i < other->req(); i++) {
6093 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
6094 }
6095 #endif
6096 }
6097 // Replace spurious copies of base_memory by top.
6098 Node* base_mem = base_memory();
6099 if (base_mem != nullptr && !base_mem->is_top()) {
6100 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
6101 if (in(i) == base_mem)
6102 set_req(i, empty_memory());
6103 }
6104 }
6105 }
6106
6107 //---------------------------grow_to_match-------------------------------------
6108 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
6109 Node* empty_mem = empty_memory();
6110 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
6111 // look for the finite support of the other memory
6112 for (uint i = other->req(); --i >= req(); ) {
6113 if (other->in(i) != empty_mem) {
6114 uint new_len = i+1;
6115 while (req() < new_len) add_req(empty_mem);
6116 break;
6117 }
6118 }
6119 }
6120
6121 //---------------------------verify_sparse-------------------------------------
6122 #ifndef PRODUCT
6123 bool MergeMemNode::verify_sparse() const {
6124 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
6125 Node* base_mem = base_memory();
6126 // The following can happen in degenerate cases, since empty==top.
6127 if (is_empty_memory(base_mem)) return true;
6128 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6129 assert(in(i) != nullptr, "sane slice");
6130 if (in(i) == base_mem) return false; // should have been the sentinel value!
6131 }
6132 return true;
6133 }
6134
6135 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
6136 Node* n;
6137 n = mm->in(idx);
6138 if (mem == n) return true; // might be empty_memory()
6139 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
6140 if (mem == n) return true;
6141 return false;
6142 }
6143 #endif // !PRODUCT