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