1 /*
2 * Copyright (c) 1997, 2025, 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 "ci/ciFlatArrayKlass.hpp"
27 #include "classfile/javaClasses.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/c2/barrierSetC2.hpp"
32 #include "gc/shared/tlab_globals.hpp"
33 #include "memory/allocation.inline.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "oops/objArrayKlass.hpp"
36 #include "opto/addnode.hpp"
37 #include "opto/arraycopynode.hpp"
38 #include "opto/cfgnode.hpp"
39 #include "opto/regalloc.hpp"
40 #include "opto/compile.hpp"
41 #include "opto/connode.hpp"
42 #include "opto/convertnode.hpp"
43 #include "opto/inlinetypenode.hpp"
44 #include "opto/loopnode.hpp"
45 #include "opto/machnode.hpp"
46 #include "opto/matcher.hpp"
47 #include "opto/memnode.hpp"
48 #include "opto/mempointer.hpp"
49 #include "opto/mulnode.hpp"
50 #include "opto/narrowptrnode.hpp"
51 #include "opto/phaseX.hpp"
52 #include "opto/regmask.hpp"
53 #include "opto/rootnode.hpp"
54 #include "opto/traceMergeStoresTag.hpp"
55 #include "opto/vectornode.hpp"
56 #include "utilities/align.hpp"
57 #include "utilities/copy.hpp"
58 #include "utilities/macros.hpp"
59 #include "utilities/powerOfTwo.hpp"
60 #include "utilities/vmError.hpp"
61
62 // Portions of code courtesy of Clifford Click
63
64 // Optimization - Graph Style
65
66 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
67
68 //=============================================================================
69 uint MemNode::size_of() const { return sizeof(*this); }
70
71 const TypePtr *MemNode::adr_type() const {
72 Node* adr = in(Address);
73 if (adr == nullptr) return nullptr; // node is dead
74 const TypePtr* cross_check = nullptr;
75 DEBUG_ONLY(cross_check = _adr_type);
76 return calculate_adr_type(adr->bottom_type(), cross_check);
77 }
78
79 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
80 if (adr != nullptr) {
81 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
82 return true;
83 }
84 }
85 return false;
86 }
87
88 #ifndef PRODUCT
89 void MemNode::dump_spec(outputStream *st) const {
90 if (in(Address) == nullptr) return; // node is dead
91 #ifndef ASSERT
92 // fake the missing field
93 const TypePtr* _adr_type = nullptr;
94 if (in(Address) != nullptr)
95 _adr_type = in(Address)->bottom_type()->isa_ptr();
96 #endif
97 dump_adr_type(this, _adr_type, st);
98
99 Compile* C = Compile::current();
100 if (C->alias_type(_adr_type)->is_volatile()) {
101 st->print(" Volatile!");
102 }
103 if (_unaligned_access) {
104 st->print(" unaligned");
105 }
106 if (_mismatched_access) {
107 st->print(" mismatched");
108 }
109 if (_unsafe_access) {
110 st->print(" unsafe");
111 }
112 }
113
114 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
115 st->print(" @");
116 if (adr_type == nullptr) {
117 st->print("null");
118 } else {
119 adr_type->dump_on(st);
120 Compile* C = Compile::current();
121 Compile::AliasType* atp = nullptr;
122 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
123 if (atp == nullptr)
124 st->print(", idx=?\?;");
125 else if (atp->index() == Compile::AliasIdxBot)
126 st->print(", idx=Bot;");
127 else if (atp->index() == Compile::AliasIdxTop)
128 st->print(", idx=Top;");
129 else if (atp->index() == Compile::AliasIdxRaw)
130 st->print(", idx=Raw;");
131 else {
132 ciField* field = atp->field();
133 if (field) {
134 st->print(", name=");
135 field->print_name_on(st);
136 }
137 st->print(", idx=%d;", atp->index());
138 }
139 }
140 }
141
142 extern void print_alias_types();
143
144 #endif
145
146 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
147 assert((t_oop != nullptr), "sanity");
148 bool is_instance = t_oop->is_known_instance_field();
149 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
150 (load != nullptr) && load->is_Load() &&
151 (phase->is_IterGVN() != nullptr);
152 if (!(is_instance || is_boxed_value_load))
153 return mchain; // don't try to optimize non-instance types
154 uint instance_id = t_oop->instance_id();
155 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
156 Node *prev = nullptr;
157 Node *result = mchain;
158 while (prev != result) {
159 prev = result;
160 if (result == start_mem)
161 break; // hit one of our sentinels
162 // skip over a call which does not affect this memory slice
163 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
164 Node *proj_in = result->in(0);
165 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
166 break; // hit one of our sentinels
167 } else if (proj_in->is_Call()) {
168 // ArrayCopyNodes processed here as well
169 CallNode *call = proj_in->as_Call();
170 if (!call->may_modify(t_oop, phase)) { // returns false for instances
171 result = call->in(TypeFunc::Memory);
172 }
173 } else if (proj_in->is_Initialize()) {
174 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
175 // Stop if this is the initialization for the object instance which
176 // contains this memory slice, otherwise skip over it.
177 if ((alloc == nullptr) || (alloc->_idx == instance_id)) {
178 break;
179 }
180 if (is_instance) {
181 result = proj_in->in(TypeFunc::Memory);
182 } else if (is_boxed_value_load) {
183 Node* klass = alloc->in(AllocateNode::KlassNode);
184 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
185 if (tklass->klass_is_exact() && !tklass->exact_klass()->equals(t_oop->is_instptr()->exact_klass())) {
186 result = proj_in->in(TypeFunc::Memory); // not related allocation
187 }
188 }
189 } else if (proj_in->is_MemBar()) {
190 ArrayCopyNode* ac = nullptr;
191 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
192 break;
193 }
194 result = proj_in->in(TypeFunc::Memory);
195 } else if (proj_in->is_top()) {
196 break; // dead code
197 } else {
198 assert(false, "unexpected projection");
199 }
200 } else if (result->is_ClearArray()) {
201 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
202 // Can not bypass initialization of the instance
203 // we are looking for.
204 break;
205 }
206 // Otherwise skip it (the call updated 'result' value).
207 } else if (result->is_MergeMem()) {
208 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, nullptr, tty);
209 }
210 }
211 return result;
212 }
213
214 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
215 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
216 if (t_oop == nullptr)
217 return mchain; // don't try to optimize non-oop types
218 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
219 bool is_instance = t_oop->is_known_instance_field();
220 PhaseIterGVN *igvn = phase->is_IterGVN();
221 if (is_instance && igvn != nullptr && result->is_Phi()) {
222 PhiNode *mphi = result->as_Phi();
223 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
224 const TypePtr *t = mphi->adr_type();
225 bool do_split = false;
226 // In the following cases, Load memory input can be further optimized based on
227 // its precise address type
228 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ) {
229 do_split = true;
230 } else if (t->isa_oopptr() && !t->is_oopptr()->is_known_instance()) {
231 const TypeOopPtr* mem_t =
232 t->is_oopptr()->cast_to_exactness(true)
233 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
234 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id());
235 if (t_oop->isa_aryptr()) {
236 mem_t = mem_t->is_aryptr()
237 ->cast_to_stable(t_oop->is_aryptr()->is_stable())
238 ->cast_to_size(t_oop->is_aryptr()->size())
239 ->cast_to_not_flat(t_oop->is_aryptr()->is_not_flat())
240 ->cast_to_not_null_free(t_oop->is_aryptr()->is_not_null_free())
241 ->with_offset(t_oop->is_aryptr()->offset())
242 ->is_aryptr();
243 }
244 do_split = mem_t == t_oop;
245 }
246 if (do_split) {
247 // clone the Phi with our address type
248 result = mphi->split_out_instance(t_adr, igvn);
249 } else {
250 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
251 }
252 }
253 return result;
254 }
255
256 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
257 uint alias_idx = phase->C->get_alias_index(tp);
258 Node *mem = mmem;
259 #ifdef ASSERT
260 {
261 // Check that current type is consistent with the alias index used during graph construction
262 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
263 bool consistent = adr_check == nullptr || adr_check->empty() ||
264 phase->C->must_alias(adr_check, alias_idx );
265 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
266 if( !consistent && adr_check != nullptr && !adr_check->empty() &&
267 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
268 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
269 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
270 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
271 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
272 // don't assert if it is dead code.
273 consistent = true;
274 }
275 if( !consistent ) {
276 st->print("alias_idx==%d, adr_check==", alias_idx);
277 if( adr_check == nullptr ) {
278 st->print("null");
279 } else {
280 adr_check->dump();
281 }
282 st->cr();
283 print_alias_types();
284 assert(consistent, "adr_check must match alias idx");
285 }
286 }
287 #endif
288 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
289 // means an array I have not precisely typed yet. Do not do any
290 // alias stuff with it any time soon.
291 const TypeOopPtr *toop = tp->isa_oopptr();
292 if (tp->base() != Type::AnyPtr &&
293 !(toop &&
294 toop->isa_instptr() &&
295 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
296 toop->offset() == Type::OffsetBot)) {
297 // compress paths and change unreachable cycles to TOP
298 // If not, we can update the input infinitely along a MergeMem cycle
299 // Equivalent code in PhiNode::Ideal
300 Node* m = phase->transform(mmem);
301 // If transformed to a MergeMem, get the desired slice
302 // Otherwise the returned node represents memory for every slice
303 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
304 // Update input if it is progress over what we have now
305 }
306 return mem;
307 }
308
309 //--------------------------Ideal_common---------------------------------------
310 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
311 // Unhook non-raw memories from complete (macro-expanded) initializations.
312 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
313 // If our control input is a dead region, kill all below the region
314 Node *ctl = in(MemNode::Control);
315 if (ctl && remove_dead_region(phase, can_reshape))
316 return this;
317 ctl = in(MemNode::Control);
318 // Don't bother trying to transform a dead node
319 if (ctl && ctl->is_top()) return NodeSentinel;
320
321 PhaseIterGVN *igvn = phase->is_IterGVN();
322 // Wait if control on the worklist.
323 if (ctl && can_reshape && igvn != nullptr) {
324 Node* bol = nullptr;
325 Node* cmp = nullptr;
326 if (ctl->in(0)->is_If()) {
327 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
328 bol = ctl->in(0)->in(1);
329 if (bol->is_Bool())
330 cmp = ctl->in(0)->in(1)->in(1);
331 }
332 if (igvn->_worklist.member(ctl) ||
333 (bol != nullptr && igvn->_worklist.member(bol)) ||
334 (cmp != nullptr && igvn->_worklist.member(cmp)) ) {
335 // This control path may be dead.
336 // Delay this memory node transformation until the control is processed.
337 igvn->_worklist.push(this);
338 return NodeSentinel; // caller will return null
339 }
340 }
341 // Ignore if memory is dead, or self-loop
342 Node *mem = in(MemNode::Memory);
343 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return null
344 assert(mem != this, "dead loop in MemNode::Ideal");
345
346 if (can_reshape && igvn != nullptr && igvn->_worklist.member(mem)) {
347 // This memory slice may be dead.
348 // Delay this mem node transformation until the memory is processed.
349 igvn->_worklist.push(this);
350 return NodeSentinel; // caller will return null
351 }
352
353 Node *address = in(MemNode::Address);
354 const Type *t_adr = phase->type(address);
355 if (t_adr == Type::TOP) return NodeSentinel; // caller will return null
356
357 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
358 // Unsafe off-heap access with zero address. Remove access and other control users
359 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
360 assert(ctl != nullptr, "unsafe accesses should be control dependent");
361 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
362 Node* u = ctl->fast_out(i);
363 if (u != ctl) {
364 igvn->rehash_node_delayed(u);
365 int nb = u->replace_edge(ctl, phase->C->top(), igvn);
366 --i, imax -= nb;
367 }
368 }
369 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
370 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
371 phase->C->root()->add_req(halt);
372 return this;
373 }
374
375 if (can_reshape && igvn != nullptr &&
376 (igvn->_worklist.member(address) ||
377 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
378 // The address's base and type may change when the address is processed.
379 // Delay this mem node transformation until the address is processed.
380 igvn->_worklist.push(this);
381 return NodeSentinel; // caller will return null
382 }
383
384 // Do NOT remove or optimize the next lines: ensure a new alias index
385 // is allocated for an oop pointer type before Escape Analysis.
386 // Note: C++ will not remove it since the call has side effect.
387 if (t_adr->isa_oopptr()) {
388 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
389 }
390
391 Node* base = nullptr;
392 if (address->is_AddP()) {
393 base = address->in(AddPNode::Base);
394 }
395 if (base != nullptr && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
396 !t_adr->isa_rawptr()) {
397 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
398 // Skip this node optimization if its address has TOP base.
399 return NodeSentinel; // caller will return null
400 }
401
402 // Avoid independent memory operations
403 Node* old_mem = mem;
404
405 // The code which unhooks non-raw memories from complete (macro-expanded)
406 // initializations was removed. After macro-expansion all stores caught
407 // by Initialize node became raw stores and there is no information
408 // which memory slices they modify. So it is unsafe to move any memory
409 // operation above these stores. Also in most cases hooked non-raw memories
410 // were already unhooked by using information from detect_ptr_independence()
411 // and find_previous_store().
412
413 if (mem->is_MergeMem()) {
414 MergeMemNode* mmem = mem->as_MergeMem();
415 const TypePtr *tp = t_adr->is_ptr();
416
417 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
418 }
419
420 if (mem != old_mem) {
421 set_req_X(MemNode::Memory, mem, phase);
422 if (phase->type(mem) == Type::TOP) return NodeSentinel;
423 return this;
424 }
425
426 // let the subclass continue analyzing...
427 return nullptr;
428 }
429
430 // Helper function for proving some simple control dominations.
431 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
432 // Already assumes that 'dom' is available at 'sub', and that 'sub'
433 // is not a constant (dominated by the method's StartNode).
434 // Used by MemNode::find_previous_store to prove that the
435 // control input of a memory operation predates (dominates)
436 // an allocation it wants to look past.
437 // Returns 'DomResult::Dominate' if all control inputs of 'dom'
438 // dominate 'sub', 'DomResult::NotDominate' if not,
439 // and 'DomResult::EncounteredDeadCode' if we can't decide due to
440 // dead code, but at the end of IGVN, we know the definite result
441 // once the dead code is cleaned up.
442 Node::DomResult MemNode::maybe_all_controls_dominate(Node* dom, Node* sub) {
443 if (dom == nullptr || dom->is_top() || sub == nullptr || sub->is_top()) {
444 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
445 }
446
447 // Check 'dom'. Skip Proj and CatchProj nodes.
448 dom = dom->find_exact_control(dom);
449 if (dom == nullptr || dom->is_top()) {
450 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
451 }
452
453 if (dom == sub) {
454 // For the case when, for example, 'sub' is Initialize and the original
455 // 'dom' is Proj node of the 'sub'.
456 return DomResult::NotDominate;
457 }
458
459 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) {
460 return DomResult::Dominate;
461 }
462
463 // 'dom' dominates 'sub' if its control edge and control edges
464 // of all its inputs dominate or equal to sub's control edge.
465
466 // Currently 'sub' is either Allocate, Initialize or Start nodes.
467 // Or Region for the check in LoadNode::Ideal();
468 // 'sub' should have sub->in(0) != nullptr.
469 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
470 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
471
472 // Get control edge of 'sub'.
473 Node* orig_sub = sub;
474 sub = sub->find_exact_control(sub->in(0));
475 if (sub == nullptr || sub->is_top()) {
476 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
477 }
478
479 assert(sub->is_CFG(), "expecting control");
480
481 if (sub == dom) {
482 return DomResult::Dominate;
483 }
484
485 if (sub->is_Start() || sub->is_Root()) {
486 return DomResult::NotDominate;
487 }
488
489 {
490 // Check all control edges of 'dom'.
491
492 ResourceMark rm;
493 Node_List nlist;
494 Unique_Node_List dom_list;
495
496 dom_list.push(dom);
497 bool only_dominating_controls = false;
498
499 for (uint next = 0; next < dom_list.size(); next++) {
500 Node* n = dom_list.at(next);
501 if (n == orig_sub) {
502 return DomResult::NotDominate; // One of dom's inputs dominated by sub.
503 }
504 if (!n->is_CFG() && n->pinned()) {
505 // Check only own control edge for pinned non-control nodes.
506 n = n->find_exact_control(n->in(0));
507 if (n == nullptr || n->is_top()) {
508 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
509 }
510 assert(n->is_CFG(), "expecting control");
511 dom_list.push(n);
512 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
513 only_dominating_controls = true;
514 } else if (n->is_CFG()) {
515 DomResult dom_result = n->dominates(sub, nlist);
516 if (dom_result == DomResult::Dominate) {
517 only_dominating_controls = true;
518 } else {
519 return dom_result;
520 }
521 } else {
522 // First, own control edge.
523 Node* m = n->find_exact_control(n->in(0));
524 if (m != nullptr) {
525 if (m->is_top()) {
526 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
527 }
528 dom_list.push(m);
529 }
530 // Now, the rest of edges.
531 uint cnt = n->req();
532 for (uint i = 1; i < cnt; i++) {
533 m = n->find_exact_control(n->in(i));
534 if (m == nullptr || m->is_top()) {
535 continue;
536 }
537 dom_list.push(m);
538 }
539 }
540 }
541 return only_dominating_controls ? DomResult::Dominate : DomResult::NotDominate;
542 }
543 }
544
545 //---------------------detect_ptr_independence---------------------------------
546 // Used by MemNode::find_previous_store to prove that two base
547 // pointers are never equal.
548 // The pointers are accompanied by their associated allocations,
549 // if any, which have been previously discovered by the caller.
550 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
551 Node* p2, AllocateNode* a2,
552 PhaseTransform* phase) {
553 // Attempt to prove that these two pointers cannot be aliased.
554 // They may both manifestly be allocations, and they should differ.
555 // Or, if they are not both allocations, they can be distinct constants.
556 // Otherwise, one is an allocation and the other a pre-existing value.
557 if (a1 == nullptr && a2 == nullptr) { // neither an allocation
558 return (p1 != p2) && p1->is_Con() && p2->is_Con();
559 } else if (a1 != nullptr && a2 != nullptr) { // both allocations
560 return (a1 != a2);
561 } else if (a1 != nullptr) { // one allocation a1
562 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
563 return all_controls_dominate(p2, a1);
564 } else { //(a2 != null) // one allocation a2
565 return all_controls_dominate(p1, a2);
566 }
567 return false;
568 }
569
570
571 // Find an arraycopy ac that produces the memory state represented by parameter mem.
572 // Return ac if
573 // (a) can_see_stored_value=true and ac must have set the value for this load or if
574 // (b) can_see_stored_value=false and ac could have set the value for this load or if
575 // (c) can_see_stored_value=false and ac cannot have set the value for this load.
576 // In case (c) change the parameter mem to the memory input of ac to skip it
577 // when searching stored value.
578 // Otherwise return null.
579 Node* LoadNode::find_previous_arraycopy(PhaseValues* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
580 ArrayCopyNode* ac = find_array_copy_clone(ld_alloc, mem);
581 if (ac != nullptr) {
582 Node* ld_addp = in(MemNode::Address);
583 Node* src = ac->in(ArrayCopyNode::Src);
584 const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
585
586 // This is a load from a cloned array. The corresponding arraycopy ac must
587 // have set the value for the load and we can return ac but only if the load
588 // is known to be within bounds. This is checked below.
589 if (ary_t != nullptr && ld_addp->is_AddP()) {
590 Node* ld_offs = ld_addp->in(AddPNode::Offset);
591 BasicType ary_elem = ary_t->elem()->array_element_basic_type();
592 jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
593 jlong elemsize = type2aelembytes(ary_elem);
594
595 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
596 const TypeInt* sizetype = ary_t->size();
597
598 if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
599 // The load is known to be within bounds. It receives its value from ac.
600 return ac;
601 }
602 // The load is known to be out-of-bounds.
603 }
604 // The load could be out-of-bounds. It must not be hoisted but must remain
605 // dependent on the runtime range check. This is achieved by returning null.
606 } else if (mem->is_Proj() && mem->in(0) != nullptr && mem->in(0)->is_ArrayCopy()) {
607 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
608
609 if (ac->is_arraycopy_validated() ||
610 ac->is_copyof_validated() ||
611 ac->is_copyofrange_validated()) {
612 Node* ld_addp = in(MemNode::Address);
613 if (ld_addp->is_AddP()) {
614 Node* ld_base = ld_addp->in(AddPNode::Address);
615 Node* ld_offs = ld_addp->in(AddPNode::Offset);
616
617 Node* dest = ac->in(ArrayCopyNode::Dest);
618
619 if (dest == ld_base) {
620 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
621 assert(!ld_offs_t->empty(), "dead reference should be checked already");
622 // Take into account vector or unsafe access size
623 jlong ld_size_in_bytes = (jlong)memory_size();
624 jlong offset_hi = ld_offs_t->_hi + ld_size_in_bytes - 1;
625 offset_hi = MIN2(offset_hi, (jlong)(TypeX::MAX->_hi)); // Take care for overflow in 32-bit VM
626 if (ac->modifies(ld_offs_t->_lo, (intptr_t)offset_hi, phase, can_see_stored_value)) {
627 return ac;
628 }
629 if (!can_see_stored_value) {
630 mem = ac->in(TypeFunc::Memory);
631 return ac;
632 }
633 }
634 }
635 }
636 }
637 return nullptr;
638 }
639
640 ArrayCopyNode* MemNode::find_array_copy_clone(Node* ld_alloc, Node* mem) const {
641 if (mem->is_Proj() && mem->in(0) != nullptr && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
642 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
643 if (ld_alloc != nullptr) {
644 // Check if there is an array copy for a clone
645 Node* mb = mem->in(0);
646 ArrayCopyNode* ac = nullptr;
647 if (mb->in(0) != nullptr && mb->in(0)->is_Proj() &&
648 mb->in(0)->in(0) != nullptr && mb->in(0)->in(0)->is_ArrayCopy()) {
649 ac = mb->in(0)->in(0)->as_ArrayCopy();
650 } else {
651 // Step over GC barrier when ReduceInitialCardMarks is disabled
652 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
653 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
654
655 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
656 ac = control_proj_ac->in(0)->as_ArrayCopy();
657 }
658 }
659
660 if (ac != nullptr && ac->is_clonebasic()) {
661 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest));
662 if (alloc != nullptr && alloc == ld_alloc) {
663 return ac;
664 }
665 }
666 }
667 }
668 return nullptr;
669 }
670
671 // The logic for reordering loads and stores uses four steps:
672 // (a) Walk carefully past stores and initializations which we
673 // can prove are independent of this load.
674 // (b) Observe that the next memory state makes an exact match
675 // with self (load or store), and locate the relevant store.
676 // (c) Ensure that, if we were to wire self directly to the store,
677 // the optimizer would fold it up somehow.
678 // (d) Do the rewiring, and return, depending on some other part of
679 // the optimizer to fold up the load.
680 // This routine handles steps (a) and (b). Steps (c) and (d) are
681 // specific to loads and stores, so they are handled by the callers.
682 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
683 //
684 Node* MemNode::find_previous_store(PhaseValues* phase) {
685 Node* ctrl = in(MemNode::Control);
686 Node* adr = in(MemNode::Address);
687 intptr_t offset = 0;
688 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
689 AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
690
691 if (offset == Type::OffsetBot)
692 return nullptr; // cannot unalias unless there are precise offsets
693
694 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
695 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
696
697 intptr_t size_in_bytes = memory_size();
698
699 Node* mem = in(MemNode::Memory); // start searching here...
700
701 int cnt = 50; // Cycle limiter
702 for (;;) { // While we can dance past unrelated stores...
703 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
704
705 Node* prev = mem;
706 if (mem->is_Store()) {
707 Node* st_adr = mem->in(MemNode::Address);
708 intptr_t st_offset = 0;
709 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
710 if (st_base == nullptr)
711 break; // inscrutable pointer
712
713 // For raw accesses it's not enough to prove that constant offsets don't intersect.
714 // We need the bases to be the equal in order for the offset check to make sense.
715 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
716 break;
717 }
718
719 if (st_offset != offset && st_offset != Type::OffsetBot) {
720 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
721 assert(mem->as_Store()->memory_size() <= MAX_STORE, "");
722 if (st_offset >= offset + size_in_bytes ||
723 st_offset <= offset - MAX_STORE ||
724 st_offset <= offset - mem->as_Store()->memory_size()) {
725 // Success: The offsets are provably independent.
726 // (You may ask, why not just test st_offset != offset and be done?
727 // The answer is that stores of different sizes can co-exist
728 // in the same sequence of RawMem effects. We sometimes initialize
729 // a whole 'tile' of array elements with a single jint or jlong.)
730 mem = mem->in(MemNode::Memory);
731 continue; // (a) advance through independent store memory
732 }
733 }
734 if (st_base != base &&
735 detect_ptr_independence(base, alloc,
736 st_base,
737 AllocateNode::Ideal_allocation(st_base),
738 phase)) {
739 // Success: The bases are provably independent.
740 mem = mem->in(MemNode::Memory);
741 continue; // (a) advance through independent store memory
742 }
743
744 // (b) At this point, if the bases or offsets do not agree, we lose,
745 // since we have not managed to prove 'this' and 'mem' independent.
746 if (st_base == base && st_offset == offset) {
747 return mem; // let caller handle steps (c), (d)
748 }
749
750 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
751 InitializeNode* st_init = mem->in(0)->as_Initialize();
752 AllocateNode* st_alloc = st_init->allocation();
753 if (st_alloc == nullptr) {
754 break; // something degenerated
755 }
756 bool known_identical = false;
757 bool known_independent = false;
758 if (alloc == st_alloc) {
759 known_identical = true;
760 } else if (alloc != nullptr) {
761 known_independent = true;
762 } else if (all_controls_dominate(this, st_alloc)) {
763 known_independent = true;
764 }
765
766 if (known_independent) {
767 // The bases are provably independent: Either they are
768 // manifestly distinct allocations, or else the control
769 // of this load dominates the store's allocation.
770 int alias_idx = phase->C->get_alias_index(adr_type());
771 if (alias_idx == Compile::AliasIdxRaw) {
772 mem = st_alloc->in(TypeFunc::Memory);
773 } else {
774 mem = st_init->memory(alias_idx);
775 }
776 continue; // (a) advance through independent store memory
777 }
778
779 // (b) at this point, if we are not looking at a store initializing
780 // the same allocation we are loading from, we lose.
781 if (known_identical) {
782 // From caller, can_see_stored_value will consult find_captured_store.
783 return mem; // let caller handle steps (c), (d)
784 }
785
786 } else if (find_previous_arraycopy(phase, alloc, mem, false) != nullptr) {
787 if (prev != mem) {
788 // Found an arraycopy but it doesn't affect that load
789 continue;
790 }
791 // Found an arraycopy that may affect that load
792 return mem;
793 } else if (addr_t != nullptr && addr_t->is_known_instance_field()) {
794 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
795 if (mem->is_Proj() && mem->in(0)->is_Call()) {
796 // ArrayCopyNodes processed here as well.
797 CallNode *call = mem->in(0)->as_Call();
798 if (!call->may_modify(addr_t, phase)) {
799 mem = call->in(TypeFunc::Memory);
800 continue; // (a) advance through independent call memory
801 }
802 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
803 ArrayCopyNode* ac = nullptr;
804 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
805 break;
806 }
807 mem = mem->in(0)->in(TypeFunc::Memory);
808 continue; // (a) advance through independent MemBar memory
809 } else if (mem->is_ClearArray()) {
810 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
811 // (the call updated 'mem' value)
812 continue; // (a) advance through independent allocation memory
813 } else {
814 // Can not bypass initialization of the instance
815 // we are looking for.
816 return mem;
817 }
818 } else if (mem->is_MergeMem()) {
819 int alias_idx = phase->C->get_alias_index(adr_type());
820 mem = mem->as_MergeMem()->memory_at(alias_idx);
821 continue; // (a) advance through independent MergeMem memory
822 }
823 }
824
825 // Unless there is an explicit 'continue', we must bail out here,
826 // because 'mem' is an inscrutable memory state (e.g., a call).
827 break;
828 }
829
830 return nullptr; // bail out
831 }
832
833 //----------------------calculate_adr_type-------------------------------------
834 // Helper function. Notices when the given type of address hits top or bottom.
835 // Also, asserts a cross-check of the type against the expected address type.
836 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
837 if (t == Type::TOP) return nullptr; // does not touch memory any more?
838 #ifdef ASSERT
839 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = nullptr;
840 #endif
841 const TypePtr* tp = t->isa_ptr();
842 if (tp == nullptr) {
843 assert(cross_check == nullptr || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
844 return TypePtr::BOTTOM; // touches lots of memory
845 } else {
846 #ifdef ASSERT
847 // %%%% [phh] We don't check the alias index if cross_check is
848 // TypeRawPtr::BOTTOM. Needs to be investigated.
849 if (cross_check != nullptr &&
850 cross_check != TypePtr::BOTTOM &&
851 cross_check != TypeRawPtr::BOTTOM) {
852 // Recheck the alias index, to see if it has changed (due to a bug).
853 Compile* C = Compile::current();
854 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
855 "must stay in the original alias category");
856 // The type of the address must be contained in the adr_type,
857 // disregarding "null"-ness.
858 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
859 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
860 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
861 "real address must not escape from expected memory type");
862 }
863 #endif
864 return tp;
865 }
866 }
867
868 uint8_t MemNode::barrier_data(const Node* n) {
869 if (n->is_LoadStore()) {
870 return n->as_LoadStore()->barrier_data();
871 } else if (n->is_Mem()) {
872 return n->as_Mem()->barrier_data();
873 }
874 return 0;
875 }
876
877 //=============================================================================
878 // Should LoadNode::Ideal() attempt to remove control edges?
879 bool LoadNode::can_remove_control() const {
880 return !has_pinned_control_dependency();
881 }
882 uint LoadNode::size_of() const { return sizeof(*this); }
883 bool LoadNode::cmp(const Node &n) const {
884 LoadNode& load = (LoadNode &)n;
885 return Type::equals(_type, load._type) &&
886 _control_dependency == load._control_dependency &&
887 _mo == load._mo;
888 }
889 const Type *LoadNode::bottom_type() const { return _type; }
890 uint LoadNode::ideal_reg() const {
891 return _type->ideal_reg();
892 }
893
894 #ifndef PRODUCT
895 void LoadNode::dump_spec(outputStream *st) const {
896 MemNode::dump_spec(st);
897 if( !Verbose && !WizardMode ) {
898 // standard dump does this in Verbose and WizardMode
899 st->print(" #"); _type->dump_on(st);
900 }
901 if (!depends_only_on_test()) {
902 st->print(" (does not depend only on test, ");
903 if (control_dependency() == UnknownControl) {
904 st->print("unknown control");
905 } else if (control_dependency() == Pinned) {
906 st->print("pinned");
907 } else if (adr_type() == TypeRawPtr::BOTTOM) {
908 st->print("raw access");
909 } else {
910 st->print("unknown reason");
911 }
912 st->print(")");
913 }
914 }
915 #endif
916
917 #ifdef ASSERT
918 //----------------------------is_immutable_value-------------------------------
919 // Helper function to allow a raw load without control edge for some cases
920 bool LoadNode::is_immutable_value(Node* adr) {
921 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
922 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) {
923
924 jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1);
925 int offsets[] = {
926 in_bytes(JavaThread::osthread_offset()),
927 in_bytes(JavaThread::threadObj_offset()),
928 in_bytes(JavaThread::vthread_offset()),
929 in_bytes(JavaThread::scopedValueCache_offset()),
930 };
931
932 for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) {
933 if (offset == offsets[i]) {
934 return true;
935 }
936 }
937 }
938
939 return false;
940 }
941 #endif
942
943 //----------------------------LoadNode::make-----------------------------------
944 // Polymorphic factory method:
945 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo,
946 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
947 Compile* C = gvn.C;
948
949 // sanity check the alias category against the created node type
950 assert(!(adr_type->isa_oopptr() &&
951 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
952 "use LoadKlassNode instead");
953 assert(!(adr_type->isa_aryptr() &&
954 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
955 "use LoadRangeNode instead");
956 // Check control edge of raw loads
957 assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
958 // oop will be recorded in oop map if load crosses safepoint
959 rt->isa_oopptr() || is_immutable_value(adr),
960 "raw memory operations should have control edge");
961 LoadNode* load = nullptr;
962 switch (bt) {
963 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
964 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
965 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
966 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
967 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
968 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break;
969 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
970 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break;
971 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
972 case T_OBJECT:
973 case T_NARROWOOP:
974 #ifdef _LP64
975 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
976 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
977 } else
978 #endif
979 {
980 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
981 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
982 }
983 break;
984 default:
985 ShouldNotReachHere();
986 break;
987 }
988 assert(load != nullptr, "LoadNode should have been created");
989 if (unaligned) {
990 load->set_unaligned_access();
991 }
992 if (mismatched) {
993 load->set_mismatched_access();
994 }
995 if (unsafe) {
996 load->set_unsafe_access();
997 }
998 load->set_barrier_data(barrier_data);
999 if (load->Opcode() == Op_LoadN) {
1000 Node* ld = gvn.transform(load);
1001 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
1002 }
1003
1004 return load;
1005 }
1006
1007 //------------------------------hash-------------------------------------------
1008 uint LoadNode::hash() const {
1009 // unroll addition of interesting fields
1010 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
1011 }
1012
1013 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
1014 if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) {
1015 bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile();
1016 bool is_stable_ary = FoldStableValues &&
1017 (tp != nullptr) && (tp->isa_aryptr() != nullptr) &&
1018 tp->isa_aryptr()->is_stable();
1019
1020 return (eliminate_boxing && non_volatile) || is_stable_ary || tp->is_inlinetypeptr();
1021 }
1022
1023 return false;
1024 }
1025
1026 LoadNode* LoadNode::pin_array_access_node() const {
1027 const TypePtr* adr_type = this->adr_type();
1028 if (adr_type != nullptr && adr_type->isa_aryptr()) {
1029 return clone_pinned();
1030 }
1031 return nullptr;
1032 }
1033
1034 // Is the value loaded previously stored by an arraycopy? If so return
1035 // a load node that reads from the source array so we may be able to
1036 // optimize out the ArrayCopy node later.
1037 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
1038 Node* ld_adr = in(MemNode::Address);
1039 intptr_t ld_off = 0;
1040 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
1041 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
1042 if (ac != nullptr) {
1043 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
1044
1045 Node* mem = ac->in(TypeFunc::Memory);
1046 Node* ctl = ac->in(0);
1047 Node* src = ac->in(ArrayCopyNode::Src);
1048
1049 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
1050 return nullptr;
1051 }
1052
1053 // load depends on the tests that validate the arraycopy
1054 LoadNode* ld = clone_pinned();
1055 Node* addp = in(MemNode::Address)->clone();
1056 if (ac->as_ArrayCopy()->is_clonebasic()) {
1057 assert(ld_alloc != nullptr, "need an alloc");
1058 assert(addp->is_AddP(), "address must be addp");
1059 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1060 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1061 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1062 addp->set_req(AddPNode::Base, src);
1063 addp->set_req(AddPNode::Address, src);
1064 } else {
1065 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1066 ac->as_ArrayCopy()->is_copyof_validated() ||
1067 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1068 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1069 addp->set_req(AddPNode::Base, src);
1070 addp->set_req(AddPNode::Address, src);
1071
1072 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1073 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
1074 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
1075
1076 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
1077 uint shift = ary_t->is_flat() ? ary_t->flat_log_elem_size() : exact_log2(type2aelembytes(ary_elem));
1078
1079 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1080 #ifdef _LP64
1081 diff = phase->transform(new ConvI2LNode(diff));
1082 #endif
1083 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1084
1085 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1086 addp->set_req(AddPNode::Offset, offset);
1087 }
1088 addp = phase->transform(addp);
1089 #ifdef ASSERT
1090 const TypePtr* adr_type = phase->type(addp)->is_ptr();
1091 ld->_adr_type = adr_type;
1092 #endif
1093 ld->set_req(MemNode::Address, addp);
1094 ld->set_req(0, ctl);
1095 ld->set_req(MemNode::Memory, mem);
1096 return ld;
1097 }
1098 return nullptr;
1099 }
1100
1101 static Node* see_through_inline_type(PhaseValues* phase, const MemNode* load, Node* base, int offset) {
1102 if (!load->is_mismatched_access() && base != nullptr && base->is_InlineType() && offset > oopDesc::klass_offset_in_bytes()) {
1103 InlineTypeNode* vt = base->as_InlineType();
1104 assert(!vt->is_larval(), "must not load from a larval object");
1105 Node* value = vt->field_value_by_offset(offset, true);
1106 assert(value != nullptr, "must see some value");
1107 return value;
1108 }
1109
1110 return nullptr;
1111 }
1112
1113 //---------------------------can_see_stored_value------------------------------
1114 // This routine exists to make sure this set of tests is done the same
1115 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
1116 // will change the graph shape in a way which makes memory alive twice at the
1117 // same time (uses the Oracle model of aliasing), then some
1118 // LoadXNode::Identity will fold things back to the equivalence-class model
1119 // of aliasing.
1120 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const {
1121 Node* ld_adr = in(MemNode::Address);
1122 intptr_t ld_off = 0;
1123 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1124 // Try to see through an InlineTypeNode
1125 // LoadN is special because the input is not compressed
1126 if (Opcode() != Op_LoadN) {
1127 Node* value = see_through_inline_type(phase, this, ld_base, ld_off);
1128 if (value != nullptr) {
1129 return value;
1130 }
1131 }
1132
1133 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base);
1134 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1135 Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr;
1136 // This is more general than load from boxing objects.
1137 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1138 uint alias_idx = atp->index();
1139 Node* result = nullptr;
1140 Node* current = st;
1141 // Skip through chains of MemBarNodes checking the MergeMems for
1142 // new states for the slice of this load. Stop once any other
1143 // kind of node is encountered. Loads from final memory can skip
1144 // through any kind of MemBar but normal loads shouldn't skip
1145 // through MemBarAcquire since the could allow them to move out of
1146 // a synchronized region. It is not safe to step over MemBarCPUOrder,
1147 // because alias info above them may be inaccurate (e.g., due to
1148 // mixed/mismatched unsafe accesses).
1149 bool is_final_mem = !atp->is_rewritable();
1150 while (current->is_Proj()) {
1151 int opc = current->in(0)->Opcode();
1152 if ((is_final_mem && (opc == Op_MemBarAcquire ||
1153 opc == Op_MemBarAcquireLock ||
1154 opc == Op_LoadFence)) ||
1155 opc == Op_MemBarRelease ||
1156 opc == Op_StoreFence ||
1157 opc == Op_MemBarReleaseLock ||
1158 opc == Op_MemBarStoreStore ||
1159 opc == Op_StoreStoreFence) {
1160 Node* mem = current->in(0)->in(TypeFunc::Memory);
1161 if (mem->is_MergeMem()) {
1162 MergeMemNode* merge = mem->as_MergeMem();
1163 Node* new_st = merge->memory_at(alias_idx);
1164 if (new_st == merge->base_memory()) {
1165 // Keep searching
1166 current = new_st;
1167 continue;
1168 }
1169 // Save the new memory state for the slice and fall through
1170 // to exit.
1171 result = new_st;
1172 }
1173 }
1174 break;
1175 }
1176 if (result != nullptr) {
1177 st = result;
1178 }
1179 }
1180
1181 // Loop around twice in the case Load -> Initialize -> Store.
1182 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1183 for (int trip = 0; trip <= 1; trip++) {
1184
1185 if (st->is_Store()) {
1186 Node* st_adr = st->in(MemNode::Address);
1187 if (st_adr != ld_adr) {
1188 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1189 intptr_t st_off = 0;
1190 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1191 if (ld_base == nullptr) return nullptr;
1192 if (st_base == nullptr) return nullptr;
1193 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr;
1194 if (ld_off != st_off) return nullptr;
1195 if (ld_off == Type::OffsetBot) return nullptr;
1196 // Same base, same offset.
1197 // Possible improvement for arrays: check index value instead of absolute offset.
1198
1199 // At this point we have proven something like this setup:
1200 // B = << base >>
1201 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1202 // S = StoreQ(AddP( B , #Off), V)
1203 // (Actually, we haven't yet proven the Q's are the same.)
1204 // In other words, we are loading from a casted version of
1205 // the same pointer-and-offset that we stored to.
1206 // Casted version may carry a dependency and it is respected.
1207 // Thus, we are able to replace L by V.
1208 }
1209 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1210 if (store_Opcode() != st->Opcode()) {
1211 return nullptr;
1212 }
1213 // LoadVector/StoreVector needs additional check to ensure the types match.
1214 if (st->is_StoreVector()) {
1215 const TypeVect* in_vt = st->as_StoreVector()->vect_type();
1216 const TypeVect* out_vt = is_Load() ? as_LoadVector()->vect_type() : as_StoreVector()->vect_type();
1217 if (in_vt != out_vt) {
1218 return nullptr;
1219 }
1220 }
1221 return st->in(MemNode::ValueIn);
1222 }
1223
1224 // A load from a freshly-created object always returns zero.
1225 // (This can happen after LoadNode::Ideal resets the load's memory input
1226 // to find_captured_store, which returned InitializeNode::zero_memory.)
1227 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1228 (st->in(0) == ld_alloc) &&
1229 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1230 // return a zero value for the load's basic type
1231 // (This is one of the few places where a generic PhaseTransform
1232 // can create new nodes. Think of it as lazily manifesting
1233 // virtually pre-existing constants.)
1234 Node* init_value = ld_alloc->in(AllocateNode::InitValue);
1235 if (init_value != nullptr) {
1236 // TODO 8350865 Is this correct for non-all-zero init values? Don't we need field_value_by_offset?
1237 return init_value;
1238 }
1239 assert(ld_alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
1240 if (memory_type() != T_VOID) {
1241 if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) {
1242 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1243 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1244 // by the ArrayCopyNode.
1245 return phase->zerocon(memory_type());
1246 }
1247 } else {
1248 // TODO: materialize all-zero vector constant
1249 assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1250 }
1251 }
1252
1253 // A load from an initialization barrier can match a captured store.
1254 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1255 InitializeNode* init = st->in(0)->as_Initialize();
1256 AllocateNode* alloc = init->allocation();
1257 if ((alloc != nullptr) && (alloc == ld_alloc)) {
1258 // examine a captured store value
1259 st = init->find_captured_store(ld_off, memory_size(), phase);
1260 if (st != nullptr) {
1261 continue; // take one more trip around
1262 }
1263 }
1264 }
1265
1266 // Load boxed value from result of valueOf() call is input parameter.
1267 if (this->is_Load() && ld_adr->is_AddP() &&
1268 (tp != nullptr) && tp->is_ptr_to_boxed_value()) {
1269 intptr_t ignore = 0;
1270 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1271 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1272 base = bs->step_over_gc_barrier(base);
1273 if (base != nullptr && base->is_Proj() &&
1274 base->as_Proj()->_con == TypeFunc::Parms &&
1275 base->in(0)->is_CallStaticJava() &&
1276 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1277 return base->in(0)->in(TypeFunc::Parms);
1278 }
1279 }
1280
1281 break;
1282 }
1283
1284 return nullptr;
1285 }
1286
1287 //----------------------is_instance_field_load_with_local_phi------------------
1288 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1289 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1290 in(Address)->is_AddP() ) {
1291 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1292 // Only instances and boxed values.
1293 if( t_oop != nullptr &&
1294 (t_oop->is_ptr_to_boxed_value() ||
1295 t_oop->is_known_instance_field()) &&
1296 t_oop->offset() != Type::OffsetBot &&
1297 t_oop->offset() != Type::OffsetTop) {
1298 return true;
1299 }
1300 }
1301 return false;
1302 }
1303
1304 //------------------------------Identity---------------------------------------
1305 // Loads are identity if previous store is to same address
1306 Node* LoadNode::Identity(PhaseGVN* phase) {
1307 // If the previous store-maker is the right kind of Store, and the store is
1308 // to the same address, then we are equal to the value stored.
1309 Node* mem = in(Memory);
1310 Node* value = can_see_stored_value(mem, phase);
1311 if( value ) {
1312 // byte, short & char stores truncate naturally.
1313 // A load has to load the truncated value which requires
1314 // some sort of masking operation and that requires an
1315 // Ideal call instead of an Identity call.
1316 if (memory_size() < BytesPerInt) {
1317 // If the input to the store does not fit with the load's result type,
1318 // it must be truncated via an Ideal call.
1319 if (!phase->type(value)->higher_equal(phase->type(this)))
1320 return this;
1321 }
1322 // (This works even when value is a Con, but LoadNode::Value
1323 // usually runs first, producing the singleton type of the Con.)
1324 if (!has_pinned_control_dependency() || value->is_Con()) {
1325 return value;
1326 } else {
1327 return this;
1328 }
1329 }
1330
1331 if (has_pinned_control_dependency()) {
1332 return this;
1333 }
1334 // Search for an existing data phi which was generated before for the same
1335 // instance's field to avoid infinite generation of phis in a loop.
1336 Node *region = mem->in(0);
1337 if (is_instance_field_load_with_local_phi(region)) {
1338 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1339 int this_index = phase->C->get_alias_index(addr_t);
1340 int this_offset = addr_t->offset();
1341 int this_iid = addr_t->instance_id();
1342 if (!addr_t->is_known_instance() &&
1343 addr_t->is_ptr_to_boxed_value()) {
1344 // Use _idx of address base (could be Phi node) for boxed values.
1345 intptr_t ignore = 0;
1346 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1347 if (base == nullptr) {
1348 return this;
1349 }
1350 this_iid = base->_idx;
1351 }
1352 const Type* this_type = bottom_type();
1353 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1354 Node* phi = region->fast_out(i);
1355 if (phi->is_Phi() && phi != mem &&
1356 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1357 return phi;
1358 }
1359 }
1360 }
1361
1362 return this;
1363 }
1364
1365 // Construct an equivalent unsigned load.
1366 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1367 BasicType bt = T_ILLEGAL;
1368 const Type* rt = nullptr;
1369 switch (Opcode()) {
1370 case Op_LoadUB: return this;
1371 case Op_LoadUS: return this;
1372 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1373 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1374 default:
1375 assert(false, "no unsigned variant: %s", Name());
1376 return nullptr;
1377 }
1378 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1379 raw_adr_type(), rt, bt, _mo, _control_dependency,
1380 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1381 }
1382
1383 // Construct an equivalent signed load.
1384 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1385 BasicType bt = T_ILLEGAL;
1386 const Type* rt = nullptr;
1387 switch (Opcode()) {
1388 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1389 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1390 case Op_LoadB: // fall through
1391 case Op_LoadS: // fall through
1392 case Op_LoadI: // fall through
1393 case Op_LoadL: return this;
1394 default:
1395 assert(false, "no signed variant: %s", Name());
1396 return nullptr;
1397 }
1398 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1399 raw_adr_type(), rt, bt, _mo, _control_dependency,
1400 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1401 }
1402
1403 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1404 BasicType bt = rt->basic_type();
1405 switch (Opcode()) {
1406 case Op_LoadI: return (bt == T_FLOAT);
1407 case Op_LoadL: return (bt == T_DOUBLE);
1408 case Op_LoadF: return (bt == T_INT);
1409 case Op_LoadD: return (bt == T_LONG);
1410
1411 default: return false;
1412 }
1413 }
1414
1415 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1416 BasicType bt = rt->basic_type();
1417 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1418 bool is_mismatched = is_mismatched_access();
1419 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1420 if (raw_type == nullptr) {
1421 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1422 }
1423 const int op = Opcode();
1424 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) ||
1425 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access());
1426 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1427 raw_adr_type(), rt, bt, _mo, _control_dependency,
1428 require_atomic_access, is_unaligned_access(), is_mismatched);
1429 }
1430
1431 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1432 BasicType bt = vt->basic_type();
1433 switch (Opcode()) {
1434 case Op_StoreI: return (bt == T_FLOAT);
1435 case Op_StoreL: return (bt == T_DOUBLE);
1436 case Op_StoreF: return (bt == T_INT);
1437 case Op_StoreD: return (bt == T_LONG);
1438
1439 default: return false;
1440 }
1441 }
1442
1443 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1444 BasicType bt = vt->basic_type();
1445 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1446 const int op = Opcode();
1447 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) ||
1448 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access());
1449 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1450 raw_adr_type(), val, bt, _mo, require_atomic_access);
1451
1452 bool is_mismatched = is_mismatched_access();
1453 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1454 if (raw_type == nullptr) {
1455 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1456 }
1457 if (is_mismatched) {
1458 st->set_mismatched_access();
1459 }
1460 return st;
1461 }
1462
1463 // We're loading from an object which has autobox behaviour.
1464 // If this object is result of a valueOf call we'll have a phi
1465 // merging a newly allocated object and a load from the cache.
1466 // We want to replace this load with the original incoming
1467 // argument to the valueOf call.
1468 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1469 assert(igvn->C->eliminate_boxing(), "sanity");
1470 intptr_t ignore = 0;
1471 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1472 if ((base == nullptr) || base->is_Phi()) {
1473 // Push the loads from the phi that comes from valueOf up
1474 // through it to allow elimination of the loads and the recovery
1475 // of the original value. It is done in split_through_phi().
1476 return nullptr;
1477 } else if (base->is_Load() ||
1478 (base->is_DecodeN() && base->in(1)->is_Load())) {
1479 // Eliminate the load of boxed value for integer types from the cache
1480 // array by deriving the value from the index into the array.
1481 // Capture the offset of the load and then reverse the computation.
1482
1483 // Get LoadN node which loads a boxing object from 'cache' array.
1484 if (base->is_DecodeN()) {
1485 base = base->in(1);
1486 }
1487 if (!base->in(Address)->is_AddP()) {
1488 return nullptr; // Complex address
1489 }
1490 AddPNode* address = base->in(Address)->as_AddP();
1491 Node* cache_base = address->in(AddPNode::Base);
1492 if ((cache_base != nullptr) && cache_base->is_DecodeN()) {
1493 // Get ConP node which is static 'cache' field.
1494 cache_base = cache_base->in(1);
1495 }
1496 if ((cache_base != nullptr) && cache_base->is_Con()) {
1497 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1498 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
1499 Node* elements[4];
1500 int shift = exact_log2(type2aelembytes(T_OBJECT));
1501 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1502 if (count > 0 && elements[0]->is_Con() &&
1503 (count == 1 ||
1504 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1505 elements[1]->in(2) == igvn->intcon(shift)))) {
1506 ciObjArray* array = base_type->const_oop()->as_obj_array();
1507 // Fetch the box object cache[0] at the base of the array and get its value
1508 ciInstance* box = array->obj_at(0)->as_instance();
1509 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1510 assert(ik->is_box_klass(), "sanity");
1511 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1512 if (ik->nof_nonstatic_fields() == 1) {
1513 // This should be true nonstatic_field_at requires calling
1514 // nof_nonstatic_fields so check it anyway
1515 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1516 BasicType bt = c.basic_type();
1517 // Only integer types have boxing cache.
1518 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1519 bt == T_BYTE || bt == T_SHORT ||
1520 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1521 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1522 if (cache_low != (int)cache_low) {
1523 return nullptr; // should not happen since cache is array indexed by value
1524 }
1525 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1526 if (offset != (int)offset) {
1527 return nullptr; // should not happen since cache is array indexed by value
1528 }
1529 // Add up all the offsets making of the address of the load
1530 Node* result = elements[0];
1531 for (int i = 1; i < count; i++) {
1532 result = igvn->transform(new AddXNode(result, elements[i]));
1533 }
1534 // Remove the constant offset from the address and then
1535 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1536 // remove the scaling of the offset to recover the original index.
1537 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1538 // Peel the shift off directly but wrap it in a dummy node
1539 // since Ideal can't return existing nodes
1540 igvn->_worklist.push(result); // remove dead node later
1541 result = new RShiftXNode(result->in(1), igvn->intcon(0));
1542 } else if (result->is_Add() && result->in(2)->is_Con() &&
1543 result->in(1)->Opcode() == Op_LShiftX &&
1544 result->in(1)->in(2) == igvn->intcon(shift)) {
1545 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1546 // but for boxing cache access we know that X<<Z will not overflow
1547 // (there is range check) so we do this optimizatrion by hand here.
1548 igvn->_worklist.push(result); // remove dead node later
1549 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1550 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1551 } else {
1552 result = new RShiftXNode(result, igvn->intcon(shift));
1553 }
1554 #ifdef _LP64
1555 if (bt != T_LONG) {
1556 result = new ConvL2INode(igvn->transform(result));
1557 }
1558 #else
1559 if (bt == T_LONG) {
1560 result = new ConvI2LNode(igvn->transform(result));
1561 }
1562 #endif
1563 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1564 // Need to preserve unboxing load type if it is unsigned.
1565 switch(this->Opcode()) {
1566 case Op_LoadUB:
1567 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1568 break;
1569 case Op_LoadUS:
1570 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1571 break;
1572 }
1573 return result;
1574 }
1575 }
1576 }
1577 }
1578 }
1579 return nullptr;
1580 }
1581
1582 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1583 Node* region = phi->in(0);
1584 if (region == nullptr) {
1585 return false; // Wait stable graph
1586 }
1587 uint cnt = phi->req();
1588 for (uint i = 1; i < cnt; i++) {
1589 Node* rc = region->in(i);
1590 if (rc == nullptr || phase->type(rc) == Type::TOP)
1591 return false; // Wait stable graph
1592 Node* in = phi->in(i);
1593 if (in == nullptr || phase->type(in) == Type::TOP)
1594 return false; // Wait stable graph
1595 }
1596 return true;
1597 }
1598
1599 //------------------------------split_through_phi------------------------------
1600 // Check whether a call to 'split_through_phi' would split this load through the
1601 // Phi *base*. This method is essentially a copy of the validations performed
1602 // by 'split_through_phi'. The first use of this method was in EA code as part
1603 // of simplification of allocation merges.
1604 // Some differences from original method (split_through_phi):
1605 // - If base->is_CastPP(): base = base->in(1)
1606 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) {
1607 Node* mem = in(Memory);
1608 Node* address = in(Address);
1609 intptr_t ignore = 0;
1610 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1611
1612 if (base->is_CastPP()) {
1613 base = base->in(1);
1614 }
1615
1616 if (req() > 3 || base == nullptr || !base->is_Phi()) {
1617 return false;
1618 }
1619
1620 if (!mem->is_Phi()) {
1621 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1622 return false;
1623 }
1624 } else if (base->in(0) != mem->in(0)) {
1625 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1626 return false;
1627 }
1628 }
1629
1630 return true;
1631 }
1632
1633 //------------------------------split_through_phi------------------------------
1634 // Split instance or boxed field load through Phi.
1635 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) {
1636 if (req() > 3) {
1637 assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs");
1638 // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account
1639 return nullptr;
1640 }
1641 Node* mem = in(Memory);
1642 Node* address = in(Address);
1643 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1644
1645 assert((t_oop != nullptr) &&
1646 (ignore_missing_instance_id ||
1647 t_oop->is_known_instance_field() ||
1648 t_oop->is_ptr_to_boxed_value()), "invalid conditions");
1649
1650 Compile* C = phase->C;
1651 intptr_t ignore = 0;
1652 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1653 bool base_is_phi = (base != nullptr) && base->is_Phi();
1654 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1655 (base != nullptr) && (base == address->in(AddPNode::Base)) &&
1656 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1657
1658 if (!((mem->is_Phi() || base_is_phi) &&
1659 (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) {
1660 return nullptr; // Neither memory or base are Phi
1661 }
1662
1663 if (mem->is_Phi()) {
1664 if (!stable_phi(mem->as_Phi(), phase)) {
1665 return nullptr; // Wait stable graph
1666 }
1667 uint cnt = mem->req();
1668 // Check for loop invariant memory.
1669 if (cnt == 3) {
1670 for (uint i = 1; i < cnt; i++) {
1671 Node* in = mem->in(i);
1672 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1673 if (m == mem) {
1674 if (i == 1) {
1675 // if the first edge was a loop, check second edge too.
1676 // If both are replaceable - we are in an infinite loop
1677 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1678 if (n == mem) {
1679 break;
1680 }
1681 }
1682 set_req(Memory, mem->in(cnt - i));
1683 return this; // made change
1684 }
1685 }
1686 }
1687 }
1688 if (base_is_phi) {
1689 if (!stable_phi(base->as_Phi(), phase)) {
1690 return nullptr; // Wait stable graph
1691 }
1692 uint cnt = base->req();
1693 // Check for loop invariant memory.
1694 if (cnt == 3) {
1695 for (uint i = 1; i < cnt; i++) {
1696 if (base->in(i) == base) {
1697 return nullptr; // Wait stable graph
1698 }
1699 }
1700 }
1701 }
1702
1703 // Split through Phi (see original code in loopopts.cpp).
1704 assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type");
1705
1706 // Do nothing here if Identity will find a value
1707 // (to avoid infinite chain of value phis generation).
1708 if (this != Identity(phase)) {
1709 return nullptr;
1710 }
1711
1712 // Select Region to split through.
1713 Node* region;
1714 DomResult dom_result = DomResult::Dominate;
1715 if (!base_is_phi) {
1716 assert(mem->is_Phi(), "sanity");
1717 region = mem->in(0);
1718 // Skip if the region dominates some control edge of the address.
1719 // We will check `dom_result` later.
1720 dom_result = MemNode::maybe_all_controls_dominate(address, region);
1721 } else if (!mem->is_Phi()) {
1722 assert(base_is_phi, "sanity");
1723 region = base->in(0);
1724 // Skip if the region dominates some control edge of the memory.
1725 // We will check `dom_result` later.
1726 dom_result = MemNode::maybe_all_controls_dominate(mem, region);
1727 } else if (base->in(0) != mem->in(0)) {
1728 assert(base_is_phi && mem->is_Phi(), "sanity");
1729 dom_result = MemNode::maybe_all_controls_dominate(mem, base->in(0));
1730 if (dom_result == DomResult::Dominate) {
1731 region = base->in(0);
1732 } else {
1733 dom_result = MemNode::maybe_all_controls_dominate(address, mem->in(0));
1734 if (dom_result == DomResult::Dominate) {
1735 region = mem->in(0);
1736 }
1737 // Otherwise we encountered a complex graph.
1738 }
1739 } else {
1740 assert(base->in(0) == mem->in(0), "sanity");
1741 region = mem->in(0);
1742 }
1743
1744 PhaseIterGVN* igvn = phase->is_IterGVN();
1745 if (dom_result != DomResult::Dominate) {
1746 if (dom_result == DomResult::EncounteredDeadCode) {
1747 // There is some dead code which eventually will be removed in IGVN.
1748 // Once this is the case, we get an unambiguous dominance result.
1749 // Push the node to the worklist again until the dead code is removed.
1750 igvn->_worklist.push(this);
1751 }
1752 return nullptr;
1753 }
1754
1755 Node* phi = nullptr;
1756 const Type* this_type = this->bottom_type();
1757 if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) {
1758 int this_index = C->get_alias_index(t_oop);
1759 int this_offset = t_oop->offset();
1760 int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx;
1761 phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset);
1762 } else if (ignore_missing_instance_id) {
1763 phi = new PhiNode(region, this_type, nullptr, mem->_idx);
1764 } else {
1765 return nullptr;
1766 }
1767
1768 for (uint i = 1; i < region->req(); i++) {
1769 Node* x;
1770 Node* the_clone = nullptr;
1771 Node* in = region->in(i);
1772 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1773 in != nullptr && in->is_OuterStripMinedLoop()) {
1774 // No node should go in the outer strip mined loop
1775 in = in->in(LoopNode::EntryControl);
1776 }
1777 if (in == nullptr || in == C->top()) {
1778 x = C->top(); // Dead path? Use a dead data op
1779 } else {
1780 x = this->clone(); // Else clone up the data op
1781 the_clone = x; // Remember for possible deletion.
1782 // Alter data node to use pre-phi inputs
1783 if (this->in(0) == region) {
1784 x->set_req(0, in);
1785 } else {
1786 x->set_req(0, nullptr);
1787 }
1788 if (mem->is_Phi() && (mem->in(0) == region)) {
1789 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1790 }
1791 if (address->is_Phi() && address->in(0) == region) {
1792 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1793 }
1794 if (base_is_phi && (base->in(0) == region)) {
1795 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1796 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1797 x->set_req(Address, adr_x);
1798 }
1799 }
1800 // Check for a 'win' on some paths
1801 const Type *t = x->Value(igvn);
1802
1803 bool singleton = t->singleton();
1804
1805 // See comments in PhaseIdealLoop::split_thru_phi().
1806 if (singleton && t == Type::TOP) {
1807 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1808 }
1809
1810 if (singleton) {
1811 x = igvn->makecon(t);
1812 } else {
1813 // We now call Identity to try to simplify the cloned node.
1814 // Note that some Identity methods call phase->type(this).
1815 // Make sure that the type array is big enough for
1816 // our new node, even though we may throw the node away.
1817 // (This tweaking with igvn only works because x is a new node.)
1818 igvn->set_type(x, t);
1819 // If x is a TypeNode, capture any more-precise type permanently into Node
1820 // otherwise it will be not updated during igvn->transform since
1821 // igvn->type(x) is set to x->Value() already.
1822 x->raise_bottom_type(t);
1823 Node* y = x->Identity(igvn);
1824 if (y != x) {
1825 x = y;
1826 } else {
1827 y = igvn->hash_find_insert(x);
1828 if (y) {
1829 x = y;
1830 } else {
1831 // Else x is a new node we are keeping
1832 // We do not need register_new_node_with_optimizer
1833 // because set_type has already been called.
1834 igvn->_worklist.push(x);
1835 }
1836 }
1837 }
1838 if (x != the_clone && the_clone != nullptr) {
1839 igvn->remove_dead_node(the_clone);
1840 }
1841 phi->set_req(i, x);
1842 }
1843 // Record Phi
1844 igvn->register_new_node_with_optimizer(phi);
1845 return phi;
1846 }
1847
1848 AllocateNode* LoadNode::is_new_object_mark_load() const {
1849 if (Opcode() == Op_LoadX) {
1850 Node* address = in(MemNode::Address);
1851 AllocateNode* alloc = AllocateNode::Ideal_allocation(address);
1852 Node* mem = in(MemNode::Memory);
1853 if (alloc != nullptr && mem->is_Proj() &&
1854 mem->in(0) != nullptr &&
1855 mem->in(0) == alloc->initialization() &&
1856 alloc->initialization()->proj_out_or_null(0) != nullptr) {
1857 return alloc;
1858 }
1859 }
1860 return nullptr;
1861 }
1862
1863
1864 //------------------------------Ideal------------------------------------------
1865 // If the load is from Field memory and the pointer is non-null, it might be possible to
1866 // zero out the control input.
1867 // If the offset is constant and the base is an object allocation,
1868 // try to hook me up to the exact initializing store.
1869 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1870 if (has_pinned_control_dependency()) {
1871 return nullptr;
1872 }
1873 Node* p = MemNode::Ideal_common(phase, can_reshape);
1874 if (p) return (p == NodeSentinel) ? nullptr : p;
1875
1876 Node* ctrl = in(MemNode::Control);
1877 Node* address = in(MemNode::Address);
1878 bool progress = false;
1879
1880 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1881 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1882
1883 // Skip up past a SafePoint control. Cannot do this for Stores because
1884 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1885 if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint &&
1886 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1887 !addr_mark &&
1888 (depends_only_on_test() || has_unknown_control_dependency())) {
1889 ctrl = ctrl->in(0);
1890 set_req(MemNode::Control,ctrl);
1891 progress = true;
1892 }
1893
1894 intptr_t ignore = 0;
1895 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1896 if (base != nullptr
1897 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1898 // Check for useless control edge in some common special cases
1899 if (in(MemNode::Control) != nullptr
1900 && !(phase->type(address)->is_inlinetypeptr() && is_mismatched_access())
1901 && can_remove_control()
1902 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1903 && all_controls_dominate(base, phase->C->start())) {
1904 // A method-invariant, non-null address (constant or 'this' argument).
1905 set_req(MemNode::Control, nullptr);
1906 progress = true;
1907 }
1908 }
1909
1910 Node* mem = in(MemNode::Memory);
1911 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1912
1913 if (can_reshape && (addr_t != nullptr)) {
1914 // try to optimize our memory input
1915 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1916 if (opt_mem != mem) {
1917 set_req_X(MemNode::Memory, opt_mem, phase);
1918 if (phase->type( opt_mem ) == Type::TOP) return nullptr;
1919 return this;
1920 }
1921 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1922 if ((t_oop != nullptr) &&
1923 (t_oop->is_known_instance_field() ||
1924 t_oop->is_ptr_to_boxed_value())) {
1925 PhaseIterGVN *igvn = phase->is_IterGVN();
1926 assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true");
1927 if (igvn->_worklist.member(opt_mem)) {
1928 // Delay this transformation until memory Phi is processed.
1929 igvn->_worklist.push(this);
1930 return nullptr;
1931 }
1932 // Split instance field load through Phi.
1933 Node* result = split_through_phi(phase);
1934 if (result != nullptr) return result;
1935
1936 if (t_oop->is_ptr_to_boxed_value()) {
1937 Node* result = eliminate_autobox(igvn);
1938 if (result != nullptr) return result;
1939 }
1940 }
1941 }
1942
1943 // Is there a dominating load that loads the same value? Leave
1944 // anything that is not a load of a field/array element (like
1945 // barriers etc.) alone
1946 if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) {
1947 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1948 Node *use = mem->fast_out(i);
1949 if (use != this &&
1950 use->Opcode() == Opcode() &&
1951 use->in(0) != nullptr &&
1952 use->in(0) != in(0) &&
1953 use->in(Address) == in(Address)) {
1954 Node* ctl = in(0);
1955 for (int i = 0; i < 10 && ctl != nullptr; i++) {
1956 ctl = IfNode::up_one_dom(ctl);
1957 if (ctl == use->in(0)) {
1958 set_req(0, use->in(0));
1959 return this;
1960 }
1961 }
1962 }
1963 }
1964 }
1965
1966 // Check for prior store with a different base or offset; make Load
1967 // independent. Skip through any number of them. Bail out if the stores
1968 // are in an endless dead cycle and report no progress. This is a key
1969 // transform for Reflection. However, if after skipping through the Stores
1970 // we can't then fold up against a prior store do NOT do the transform as
1971 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1972 // array memory alive twice: once for the hoisted Load and again after the
1973 // bypassed Store. This situation only works if EVERYBODY who does
1974 // anti-dependence work knows how to bypass. I.e. we need all
1975 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1976 // the alias index stuff. So instead, peek through Stores and IFF we can
1977 // fold up, do so.
1978 Node* prev_mem = find_previous_store(phase);
1979 if (prev_mem != nullptr) {
1980 Node* value = can_see_arraycopy_value(prev_mem, phase);
1981 if (value != nullptr) {
1982 return value;
1983 }
1984 }
1985 // Steps (a), (b): Walk past independent stores to find an exact match.
1986 if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) {
1987 // (c) See if we can fold up on the spot, but don't fold up here.
1988 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1989 // just return a prior value, which is done by Identity calls.
1990 if (can_see_stored_value(prev_mem, phase)) {
1991 // Make ready for step (d):
1992 set_req_X(MemNode::Memory, prev_mem, phase);
1993 return this;
1994 }
1995 }
1996
1997 return progress ? this : nullptr;
1998 }
1999
2000 // Helper to recognize certain Klass fields which are invariant across
2001 // some group of array types (e.g., int[] or all T[] where T < Object).
2002 const Type*
2003 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
2004 ciKlass* klass) const {
2005 assert(!UseCompactObjectHeaders || tkls->offset() != in_bytes(Klass::prototype_header_offset()),
2006 "must not happen");
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(Opcode() == Op_LoadUS, "must load an unsigned short 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(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(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 // If we are loading from a freshly-allocated object, produce a zero,
2043 // if the load is provably beyond the header of the object.
2044 // (Also allow a variable load from a fresh array to produce zero.)
2045 const TypeOopPtr* tinst = tp->isa_oopptr();
2046 bool is_instance = (tinst != nullptr) && tinst->is_known_instance_field();
2047 Node* value = can_see_stored_value(mem, phase);
2048 if (value != nullptr && value->is_Con()) {
2049 assert(value->bottom_type()->higher_equal(_type), "sanity");
2050 return value->bottom_type();
2051 }
2052
2053 // Try to guess loaded type from pointer type
2054 if (tp->isa_aryptr()) {
2055 const TypeAryPtr* ary = tp->is_aryptr();
2056 const Type* t = ary->elem();
2057
2058 // Determine whether the reference is beyond the header or not, by comparing
2059 // the offset against the offset of the start of the array's data.
2060 // Different array types begin at slightly different offsets (12 vs. 16).
2061 // We choose T_BYTE as an example base type that is least restrictive
2062 // as to alignment, which will therefore produce the smallest
2063 // possible base offset.
2064 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2065 const bool off_beyond_header = (off >= min_base_off);
2066
2067 // Try to constant-fold a stable array element.
2068 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
2069 // Make sure the reference is not into the header and the offset is constant
2070 ciObject* aobj = ary->const_oop();
2071 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
2072 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
2073 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
2074 stable_dimension,
2075 memory_type(), is_unsigned());
2076 if (con_type != nullptr) {
2077 return con_type;
2078 }
2079 }
2080 }
2081
2082 // Don't do this for integer types. There is only potential profit if
2083 // the element type t is lower than _type; that is, for int types, if _type is
2084 // more restrictive than t. This only happens here if one is short and the other
2085 // char (both 16 bits), and in those cases we've made an intentional decision
2086 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2087 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2088 //
2089 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2090 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
2091 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2092 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
2093 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2094 // In fact, that could have been the original type of p1, and p1 could have
2095 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2096 // expression (LShiftL quux 3) independently optimized to the constant 8.
2097 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2098 && (_type->isa_vect() == nullptr)
2099 && !ary->is_flat()
2100 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2101 // t might actually be lower than _type, if _type is a unique
2102 // concrete subclass of abstract class t.
2103 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
2104 const Type* jt = t->join_speculative(_type);
2105 // In any case, do not allow the join, per se, to empty out the type.
2106 if (jt->empty() && !t->empty()) {
2107 // This can happen if a interface-typed array narrows to a class type.
2108 jt = _type;
2109 }
2110 #ifdef ASSERT
2111 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2112 // The pointers in the autobox arrays are always non-null
2113 Node* base = adr->in(AddPNode::Base);
2114 if ((base != nullptr) && base->is_DecodeN()) {
2115 // Get LoadN node which loads IntegerCache.cache field
2116 base = base->in(1);
2117 }
2118 if ((base != nullptr) && base->is_Con()) {
2119 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2120 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2121 // It could be narrow oop
2122 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2123 }
2124 }
2125 }
2126 #endif
2127 return jt;
2128 }
2129 }
2130 } else if (tp->base() == Type::InstPtr) {
2131 assert( off != Type::OffsetBot ||
2132 // arrays can be cast to Objects
2133 !tp->isa_instptr() ||
2134 tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
2135 // Default value load
2136 tp->is_instptr()->instance_klass() == ciEnv::current()->Class_klass() ||
2137 // unsafe field access may not have a constant offset
2138 C->has_unsafe_access(),
2139 "Field accesses must be precise" );
2140 // For oop loads, we expect the _type to be precise.
2141
2142 const TypeInstPtr* tinst = tp->is_instptr();
2143 BasicType bt = memory_type();
2144
2145 // Optimize loads from constant fields.
2146 ciObject* const_oop = tinst->const_oop();
2147 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2148 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt);
2149 if (con_type != nullptr) {
2150 return con_type;
2151 }
2152 }
2153 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2154 assert(off != Type::OffsetBot ||
2155 !tp->isa_instklassptr() ||
2156 // arrays can be cast to Objects
2157 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2158 // also allow array-loading from the primary supertype
2159 // array during subtype checks
2160 Opcode() == Op_LoadKlass,
2161 "Field accesses must be precise");
2162 // For klass/static loads, we expect the _type to be precise
2163 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2164 /* With mirrors being an indirect in the Klass*
2165 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2166 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2167 *
2168 * So check the type and klass of the node before the LoadP.
2169 */
2170 Node* adr2 = adr->in(MemNode::Address);
2171 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2172 if (tkls != nullptr && !StressReflectiveCode) {
2173 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2174 ciKlass* klass = tkls->exact_klass();
2175 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2176 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2177 return TypeInstPtr::make(klass->java_mirror());
2178 }
2179 }
2180 }
2181
2182 const TypeKlassPtr *tkls = tp->isa_klassptr();
2183 if (tkls != nullptr) {
2184 if (tkls->is_loaded() && tkls->klass_is_exact()) {
2185 ciKlass* klass = tkls->exact_klass();
2186 // We are loading a field from a Klass metaobject whose identity
2187 // is known at compile time (the type is "exact" or "precise").
2188 // Check for fields we know are maintained as constants by the VM.
2189 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2190 // The field is Klass::_super_check_offset. Return its (constant) value.
2191 // (Folds up type checking code.)
2192 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2193 return TypeInt::make(klass->super_check_offset());
2194 }
2195 if (UseCompactObjectHeaders) { // TODO: Should EnableValhalla also take this path ?
2196 if (tkls->offset() == in_bytes(Klass::prototype_header_offset())) {
2197 // The field is Klass::_prototype_header. Return its (constant) value.
2198 assert(this->Opcode() == Op_LoadX, "must load a proper type from _prototype_header");
2199 return TypeX::make(klass->prototype_header());
2200 }
2201 }
2202 // Compute index into primary_supers array
2203 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2204 // Check for overflowing; use unsigned compare to handle the negative case.
2205 if( depth < ciKlass::primary_super_limit() ) {
2206 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2207 // (Folds up type checking code.)
2208 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2209 ciKlass *ss = klass->super_of_depth(depth);
2210 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2211 }
2212 const Type* aift = load_array_final_field(tkls, klass);
2213 if (aift != nullptr) return aift;
2214 }
2215
2216 // We can still check if we are loading from the primary_supers array at a
2217 // shallow enough depth. Even though the klass is not exact, entries less
2218 // than or equal to its super depth are correct.
2219 if (tkls->is_loaded()) {
2220 ciKlass* klass = nullptr;
2221 if (tkls->isa_instklassptr()) {
2222 klass = tkls->is_instklassptr()->instance_klass();
2223 } else {
2224 int dims;
2225 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims);
2226 if (inner->isa_instklassptr()) {
2227 klass = inner->is_instklassptr()->instance_klass();
2228 klass = ciObjArrayKlass::make(klass, dims);
2229 }
2230 }
2231 if (klass != nullptr) {
2232 // Compute index into primary_supers array
2233 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2234 // Check for overflowing; use unsigned compare to handle the negative case.
2235 if (depth < ciKlass::primary_super_limit() &&
2236 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case
2237 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2238 // (Folds up type checking code.)
2239 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2240 ciKlass *ss = klass->super_of_depth(depth);
2241 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2242 }
2243 }
2244 }
2245
2246 // If the type is enough to determine that the thing is not an array,
2247 // we can give the layout_helper a positive interval type.
2248 // This will help short-circuit some reflective code.
2249 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) &&
2250 tkls->isa_instklassptr() && // not directly typed as an array
2251 !tkls->is_instklassptr()->instance_klass()->is_java_lang_Object() // not the supertype of all T[] and specifically not Serializable & Cloneable
2252 ) {
2253 // Note: When interfaces are reliable, we can narrow the interface
2254 // test to (klass != Serializable && klass != Cloneable).
2255 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2256 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2257 // The key property of this type is that it folds up tests
2258 // for array-ness, since it proves that the layout_helper is positive.
2259 // Thus, a generic value like the basic object layout helper works fine.
2260 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2261 }
2262 }
2263
2264 bool is_vect = (_type->isa_vect() != nullptr);
2265 if (is_instance && !is_vect) {
2266 // If we have an instance type and our memory input is the
2267 // programs's initial memory state, there is no matching store,
2268 // so just return a zero of the appropriate type -
2269 // except if it is vectorized - then we have no zero constant.
2270 Node *mem = in(MemNode::Memory);
2271 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2272 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2273 // TODO 8350865 This is needed for flat array accesses, somehow the memory of the loads bypasses the intrinsic
2274 // Run TestArrays.test6 in Scenario4, we need more tests for this. TestBasicFunctionality::test20 also needs this.
2275 if (tp->isa_aryptr() && tp->is_aryptr()->is_flat() && !UseFieldFlattening) {
2276 return _type;
2277 }
2278 return Type::get_zero_type(_type->basic_type());
2279 }
2280 }
2281 if (!UseCompactObjectHeaders) {
2282 Node* alloc = is_new_object_mark_load();
2283 if (alloc != nullptr) {
2284 if (EnableValhalla) {
2285 // The mark word may contain property bits (inline, flat, null-free)
2286 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2287 const TypeKlassPtr* tkls = phase->type(klass_node)->isa_klassptr();
2288 if (tkls != nullptr && tkls->is_loaded() && tkls->klass_is_exact()) {
2289 return TypeX::make(tkls->exact_klass()->prototype_header());
2290 }
2291 } else {
2292 return TypeX::make(markWord::prototype().value());
2293 }
2294 }
2295 }
2296
2297 return _type;
2298 }
2299
2300 //------------------------------match_edge-------------------------------------
2301 // Do we Match on this edge index or not? Match only the address.
2302 uint LoadNode::match_edge(uint idx) const {
2303 return idx == MemNode::Address;
2304 }
2305
2306 //--------------------------LoadBNode::Ideal--------------------------------------
2307 //
2308 // If the previous store is to the same address as this load,
2309 // and the value stored was larger than a byte, replace this load
2310 // with the value stored truncated to a byte. If no truncation is
2311 // needed, the replacement is done in LoadNode::Identity().
2312 //
2313 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2314 Node* mem = in(MemNode::Memory);
2315 Node* value = can_see_stored_value(mem,phase);
2316 if (value != nullptr) {
2317 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2318 if (narrow != value) {
2319 return narrow;
2320 }
2321 }
2322 // Identity call will handle the case where truncation is not needed.
2323 return LoadNode::Ideal(phase, can_reshape);
2324 }
2325
2326 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2327 Node* mem = in(MemNode::Memory);
2328 Node* value = can_see_stored_value(mem,phase);
2329 if (value != nullptr && value->is_Con() &&
2330 !value->bottom_type()->higher_equal(_type)) {
2331 // If the input to the store does not fit with the load's result type,
2332 // it must be truncated. We can't delay until Ideal call since
2333 // a singleton Value is needed for split_thru_phi optimization.
2334 int con = value->get_int();
2335 return TypeInt::make((con << 24) >> 24);
2336 }
2337 return LoadNode::Value(phase);
2338 }
2339
2340 //--------------------------LoadUBNode::Ideal-------------------------------------
2341 //
2342 // If the previous store is to the same address as this load,
2343 // and the value stored was larger than a byte, replace this load
2344 // with the value stored truncated to a byte. If no truncation is
2345 // needed, the replacement is done in LoadNode::Identity().
2346 //
2347 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2348 Node* mem = in(MemNode::Memory);
2349 Node* value = can_see_stored_value(mem, phase);
2350 if (value != nullptr) {
2351 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2352 if (narrow != value) {
2353 return narrow;
2354 }
2355 }
2356 // Identity call will handle the case where truncation is not needed.
2357 return LoadNode::Ideal(phase, can_reshape);
2358 }
2359
2360 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2361 Node* mem = in(MemNode::Memory);
2362 Node* value = can_see_stored_value(mem,phase);
2363 if (value != nullptr && value->is_Con() &&
2364 !value->bottom_type()->higher_equal(_type)) {
2365 // If the input to the store does not fit with the load's result type,
2366 // it must be truncated. We can't delay until Ideal call since
2367 // a singleton Value is needed for split_thru_phi optimization.
2368 int con = value->get_int();
2369 return TypeInt::make(con & 0xFF);
2370 }
2371 return LoadNode::Value(phase);
2372 }
2373
2374 //--------------------------LoadUSNode::Ideal-------------------------------------
2375 //
2376 // If the previous store is to the same address as this load,
2377 // and the value stored was larger than a char, replace this load
2378 // with the value stored truncated to a char. If no truncation is
2379 // needed, the replacement is done in LoadNode::Identity().
2380 //
2381 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2382 Node* mem = in(MemNode::Memory);
2383 Node* value = can_see_stored_value(mem,phase);
2384 if (value != nullptr) {
2385 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2386 if (narrow != value) {
2387 return narrow;
2388 }
2389 }
2390 // Identity call will handle the case where truncation is not needed.
2391 return LoadNode::Ideal(phase, can_reshape);
2392 }
2393
2394 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2395 Node* mem = in(MemNode::Memory);
2396 Node* value = can_see_stored_value(mem,phase);
2397 if (value != nullptr && value->is_Con() &&
2398 !value->bottom_type()->higher_equal(_type)) {
2399 // If the input to the store does not fit with the load's result type,
2400 // it must be truncated. We can't delay until Ideal call since
2401 // a singleton Value is needed for split_thru_phi optimization.
2402 int con = value->get_int();
2403 return TypeInt::make(con & 0xFFFF);
2404 }
2405 return LoadNode::Value(phase);
2406 }
2407
2408 //--------------------------LoadSNode::Ideal--------------------------------------
2409 //
2410 // If the previous store is to the same address as this load,
2411 // and the value stored was larger than a short, replace this load
2412 // with the value stored truncated to a short. If no truncation is
2413 // needed, the replacement is done in LoadNode::Identity().
2414 //
2415 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2416 Node* mem = in(MemNode::Memory);
2417 Node* value = can_see_stored_value(mem,phase);
2418 if (value != nullptr) {
2419 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2420 if (narrow != value) {
2421 return narrow;
2422 }
2423 }
2424 // Identity call will handle the case where truncation is not needed.
2425 return LoadNode::Ideal(phase, can_reshape);
2426 }
2427
2428 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2429 Node* mem = in(MemNode::Memory);
2430 Node* value = can_see_stored_value(mem,phase);
2431 if (value != nullptr && value->is_Con() &&
2432 !value->bottom_type()->higher_equal(_type)) {
2433 // If the input to the store does not fit with the load's result type,
2434 // it must be truncated. We can't delay until Ideal call since
2435 // a singleton Value is needed for split_thru_phi optimization.
2436 int con = value->get_int();
2437 return TypeInt::make((con << 16) >> 16);
2438 }
2439 return LoadNode::Value(phase);
2440 }
2441
2442 Node* LoadNNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2443 // Loading from an InlineType, find the input and make an EncodeP
2444 Node* addr = in(Address);
2445 intptr_t offset;
2446 Node* base = AddPNode::Ideal_base_and_offset(addr, phase, offset);
2447 Node* value = see_through_inline_type(phase, this, base, offset);
2448 if (value != nullptr) {
2449 return new EncodePNode(value, type());
2450 }
2451
2452 return LoadNode::Ideal(phase, can_reshape);
2453 }
2454
2455 //=============================================================================
2456 //----------------------------LoadKlassNode::make------------------------------
2457 // Polymorphic factory method:
2458 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2459 // sanity check the alias category against the created node type
2460 const TypePtr* adr_type = adr->bottom_type()->isa_ptr();
2461 assert(adr_type != nullptr, "expecting TypeKlassPtr");
2462 #ifdef _LP64
2463 if (adr_type->is_ptr_to_narrowklass()) {
2464 assert(UseCompressedClassPointers, "no compressed klasses");
2465 Node* load_klass = gvn.transform(new LoadNKlassNode(mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2466 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2467 }
2468 #endif
2469 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2470 return new LoadKlassNode(mem, adr, at, tk, MemNode::unordered);
2471 }
2472
2473 //------------------------------Value------------------------------------------
2474 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2475 return klass_value_common(phase);
2476 }
2477
2478 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2479 // Either input is TOP ==> the result is TOP
2480 const Type *t1 = phase->type( in(MemNode::Memory) );
2481 if (t1 == Type::TOP) return Type::TOP;
2482 Node *adr = in(MemNode::Address);
2483 const Type *t2 = phase->type( adr );
2484 if (t2 == Type::TOP) return Type::TOP;
2485 const TypePtr *tp = t2->is_ptr();
2486 if (TypePtr::above_centerline(tp->ptr()) ||
2487 tp->ptr() == TypePtr::Null) return Type::TOP;
2488
2489 // Return a more precise klass, if possible
2490 const TypeInstPtr *tinst = tp->isa_instptr();
2491 if (tinst != nullptr) {
2492 ciInstanceKlass* ik = tinst->instance_klass();
2493 int offset = tinst->offset();
2494 if (ik == phase->C->env()->Class_klass()
2495 && (offset == java_lang_Class::klass_offset() ||
2496 offset == java_lang_Class::array_klass_offset())) {
2497 // We are loading a special hidden field from a Class mirror object,
2498 // the field which points to the VM's Klass metaobject.
2499 bool is_null_free_array = false;
2500 ciType* t = tinst->java_mirror_type(&is_null_free_array);
2501 // java_mirror_type returns non-null for compile-time Class constants.
2502 if (t != nullptr) {
2503 // constant oop => constant klass
2504 if (offset == java_lang_Class::array_klass_offset()) {
2505 if (t->is_void()) {
2506 // We cannot create a void array. Since void is a primitive type return null
2507 // klass. Users of this result need to do a null check on the returned klass.
2508 return TypePtr::NULL_PTR;
2509 }
2510 const TypeKlassPtr* tklass = TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);
2511 if (is_null_free_array) {
2512 tklass = tklass->is_aryklassptr()->cast_to_null_free();
2513 }
2514 return tklass;
2515 }
2516 if (!t->is_klass()) {
2517 // a primitive Class (e.g., int.class) has null for a klass field
2518 return TypePtr::NULL_PTR;
2519 }
2520 // Fold up the load of the hidden field
2521 const TypeKlassPtr* tklass = TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);
2522 if (is_null_free_array) {
2523 tklass = tklass->is_aryklassptr()->cast_to_null_free();
2524 }
2525 return tklass;
2526 }
2527 // non-constant mirror, so we can't tell what's going on
2528 }
2529 if (!tinst->is_loaded())
2530 return _type; // Bail out if not loaded
2531 if (offset == oopDesc::klass_offset_in_bytes()) {
2532 return tinst->as_klass_type(true);
2533 }
2534 }
2535
2536 // Check for loading klass from an array
2537 const TypeAryPtr* tary = tp->isa_aryptr();
2538 if (tary != nullptr &&
2539 tary->offset() == oopDesc::klass_offset_in_bytes()) {
2540 return tary->as_klass_type(true);
2541 }
2542
2543 // Check for loading klass from an array klass
2544 const TypeKlassPtr *tkls = tp->isa_klassptr();
2545 if (tkls != nullptr && !StressReflectiveCode) {
2546 if (!tkls->is_loaded())
2547 return _type; // Bail out if not loaded
2548 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2549 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2550 // // Always returning precise element type is incorrect,
2551 // // e.g., element type could be object and array may contain strings
2552 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2553
2554 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2555 // according to the element type's subclassing.
2556 return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2557 }
2558 if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() &&
2559 tkls->offset() == in_bytes(Klass::super_offset())) {
2560 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super();
2561 // The field is Klass::_super. Return its (constant) value.
2562 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2563 return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR;
2564 }
2565 }
2566
2567 if (tkls != nullptr && !UseSecondarySupersCache
2568 && tkls->offset() == in_bytes(Klass::secondary_super_cache_offset())) {
2569 // Treat Klass::_secondary_super_cache as a constant when the cache is disabled.
2570 return TypePtr::NULL_PTR;
2571 }
2572
2573 // Bailout case
2574 return LoadNode::Value(phase);
2575 }
2576
2577 //------------------------------Identity---------------------------------------
2578 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2579 // Also feed through the klass in Allocate(...klass...)._klass.
2580 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2581 return klass_identity_common(phase);
2582 }
2583
2584 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2585 Node* x = LoadNode::Identity(phase);
2586 if (x != this) return x;
2587
2588 // Take apart the address into an oop and offset.
2589 // Return 'this' if we cannot.
2590 Node* adr = in(MemNode::Address);
2591 intptr_t offset = 0;
2592 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2593 if (base == nullptr) return this;
2594 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2595 if (toop == nullptr) return this;
2596
2597 // Step over potential GC barrier for OopHandle resolve
2598 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2599 if (bs->is_gc_barrier_node(base)) {
2600 base = bs->step_over_gc_barrier(base);
2601 }
2602
2603 // We can fetch the klass directly through an AllocateNode.
2604 // This works even if the klass is not constant (clone or newArray).
2605 if (offset == oopDesc::klass_offset_in_bytes()) {
2606 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2607 if (allocated_klass != nullptr) {
2608 return allocated_klass;
2609 }
2610 }
2611
2612 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2613 // See inline_native_Class_query for occurrences of these patterns.
2614 // Java Example: x.getClass().isAssignableFrom(y)
2615 //
2616 // This improves reflective code, often making the Class
2617 // mirror go completely dead. (Current exception: Class
2618 // mirrors may appear in debug info, but we could clean them out by
2619 // introducing a new debug info operator for Klass.java_mirror).
2620
2621 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass()
2622 && offset == java_lang_Class::klass_offset()) {
2623 if (base->is_Load()) {
2624 Node* base2 = base->in(MemNode::Address);
2625 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2626 Node* adr2 = base2->in(MemNode::Address);
2627 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2628 if (tkls != nullptr && !tkls->empty()
2629 && (tkls->isa_instklassptr() || tkls->isa_aryklassptr())
2630 && adr2->is_AddP()
2631 ) {
2632 int mirror_field = in_bytes(Klass::java_mirror_offset());
2633 if (tkls->offset() == mirror_field) {
2634 return adr2->in(AddPNode::Base);
2635 }
2636 }
2637 }
2638 }
2639 }
2640
2641 return this;
2642 }
2643
2644 LoadNode* LoadNode::clone_pinned() const {
2645 LoadNode* ld = clone()->as_Load();
2646 ld->_control_dependency = UnknownControl;
2647 return ld;
2648 }
2649
2650
2651 //------------------------------Value------------------------------------------
2652 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2653 const Type *t = klass_value_common(phase);
2654 if (t == Type::TOP)
2655 return t;
2656
2657 return t->make_narrowklass();
2658 }
2659
2660 //------------------------------Identity---------------------------------------
2661 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2662 // Also feed through the klass in Allocate(...klass...)._klass.
2663 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2664 Node *x = klass_identity_common(phase);
2665
2666 const Type *t = phase->type( x );
2667 if( t == Type::TOP ) return x;
2668 if( t->isa_narrowklass()) return x;
2669 assert (!t->isa_narrowoop(), "no narrow oop here");
2670
2671 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2672 }
2673
2674 //------------------------------Value-----------------------------------------
2675 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2676 // Either input is TOP ==> the result is TOP
2677 const Type *t1 = phase->type( in(MemNode::Memory) );
2678 if( t1 == Type::TOP ) return Type::TOP;
2679 Node *adr = in(MemNode::Address);
2680 const Type *t2 = phase->type( adr );
2681 if( t2 == Type::TOP ) return Type::TOP;
2682 const TypePtr *tp = t2->is_ptr();
2683 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2684 const TypeAryPtr *tap = tp->isa_aryptr();
2685 if( !tap ) return _type;
2686 return tap->size();
2687 }
2688
2689 //-------------------------------Ideal---------------------------------------
2690 // Feed through the length in AllocateArray(...length...)._length.
2691 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2692 Node* p = MemNode::Ideal_common(phase, can_reshape);
2693 if (p) return (p == NodeSentinel) ? nullptr : p;
2694
2695 // Take apart the address into an oop and offset.
2696 // Return 'this' if we cannot.
2697 Node* adr = in(MemNode::Address);
2698 intptr_t offset = 0;
2699 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2700 if (base == nullptr) return nullptr;
2701 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2702 if (tary == nullptr) return nullptr;
2703
2704 // We can fetch the length directly through an AllocateArrayNode.
2705 // This works even if the length is not constant (clone or newArray).
2706 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2707 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2708 if (alloc != nullptr) {
2709 Node* allocated_length = alloc->Ideal_length();
2710 Node* len = alloc->make_ideal_length(tary, phase);
2711 if (allocated_length != len) {
2712 // New CastII improves on this.
2713 return len;
2714 }
2715 }
2716 }
2717
2718 return nullptr;
2719 }
2720
2721 //------------------------------Identity---------------------------------------
2722 // Feed through the length in AllocateArray(...length...)._length.
2723 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2724 Node* x = LoadINode::Identity(phase);
2725 if (x != this) return x;
2726
2727 // Take apart the address into an oop and offset.
2728 // Return 'this' if we cannot.
2729 Node* adr = in(MemNode::Address);
2730 intptr_t offset = 0;
2731 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2732 if (base == nullptr) return this;
2733 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2734 if (tary == nullptr) return this;
2735
2736 // We can fetch the length directly through an AllocateArrayNode.
2737 // This works even if the length is not constant (clone or newArray).
2738 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2739 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2740 if (alloc != nullptr) {
2741 Node* allocated_length = alloc->Ideal_length();
2742 // Do not allow make_ideal_length to allocate a CastII node.
2743 Node* len = alloc->make_ideal_length(tary, phase, false);
2744 if (allocated_length == len) {
2745 // Return allocated_length only if it would not be improved by a CastII.
2746 return allocated_length;
2747 }
2748 }
2749 }
2750
2751 return this;
2752
2753 }
2754
2755 //=============================================================================
2756 //---------------------------StoreNode::make-----------------------------------
2757 // Polymorphic factory method:
2758 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) {
2759 assert((mo == unordered || mo == release), "unexpected");
2760 Compile* C = gvn.C;
2761 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2762 ctl != nullptr, "raw memory operations should have control edge");
2763
2764 switch (bt) {
2765 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2766 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2767 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2768 case T_CHAR:
2769 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2770 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2771 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2772 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2773 case T_METADATA:
2774 case T_ADDRESS:
2775 case T_OBJECT:
2776 #ifdef _LP64
2777 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2778 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2779 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2780 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2781 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2782 adr->bottom_type()->isa_rawptr())) {
2783 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2784 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2785 }
2786 #endif
2787 {
2788 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2789 }
2790 default:
2791 ShouldNotReachHere();
2792 return (StoreNode*)nullptr;
2793 }
2794 }
2795
2796 //--------------------------bottom_type----------------------------------------
2797 const Type *StoreNode::bottom_type() const {
2798 return Type::MEMORY;
2799 }
2800
2801 //------------------------------hash-------------------------------------------
2802 uint StoreNode::hash() const {
2803 // unroll addition of interesting fields
2804 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2805
2806 // Since they are not commoned, do not hash them:
2807 return NO_HASH;
2808 }
2809
2810 // Link together multiple stores (B/S/C/I) into a longer one.
2811 //
2812 // Example: _store = StoreB[i+3]
2813 //
2814 // RangeCheck[i+0] RangeCheck[i+0]
2815 // StoreB[i+0]
2816 // RangeCheck[i+3] RangeCheck[i+3]
2817 // StoreB[i+1] --> pass: fail:
2818 // StoreB[i+2] StoreI[i+0] StoreB[i+0]
2819 // StoreB[i+3]
2820 //
2821 // The 4 StoreB are merged into a single StoreI node. We have to be careful with RangeCheck[i+1]: before
2822 // the optimization, if this RangeCheck[i+1] fails, then we execute only StoreB[i+0], and then trap. After
2823 // the optimization, the new StoreI[i+0] is on the passing path of RangeCheck[i+3], and StoreB[i+0] on the
2824 // failing path.
2825 //
2826 // Note: For normal array stores, every store at first has a RangeCheck. But they can be removed with:
2827 // - RCE (RangeCheck Elimination): the RangeChecks in the loop are hoisted out and before the loop,
2828 // and possibly no RangeChecks remain between the stores.
2829 // - RangeCheck smearing: the earlier RangeChecks are adjusted such that they cover later RangeChecks,
2830 // and those later RangeChecks can be removed. Example:
2831 //
2832 // RangeCheck[i+0] RangeCheck[i+0] <- before first store
2833 // StoreB[i+0] StoreB[i+0] <- first store
2834 // RangeCheck[i+1] --> smeared --> RangeCheck[i+3] <- only RC between first and last store
2835 // StoreB[i+1] StoreB[i+1] <- second store
2836 // RangeCheck[i+2] --> removed
2837 // StoreB[i+2] StoreB[i+2]
2838 // RangeCheck[i+3] --> removed
2839 // StoreB[i+3] StoreB[i+3] <- last store
2840 //
2841 // Thus, it is a common pattern that between the first and last store in a chain
2842 // of adjacent stores there remains exactly one RangeCheck, located between the
2843 // first and the second store (e.g. RangeCheck[i+3]).
2844 //
2845 class MergePrimitiveStores : public StackObj {
2846 private:
2847 PhaseGVN* const _phase;
2848 StoreNode* const _store;
2849 // State machine with initial state Unknown
2850 // Allowed transitions:
2851 // Unknown -> Const
2852 // Unknown -> Platform
2853 // Unknown -> Reverse
2854 // Unknown -> NotAdjacent
2855 // Const -> Const
2856 // Const -> NotAdjacent
2857 // Platform -> Platform
2858 // Platform -> NotAdjacent
2859 // Reverse -> Reverse
2860 // Reverse -> NotAdjacent
2861 // NotAdjacent -> NotAdjacent
2862 enum ValueOrder : uint8_t {
2863 Unknown, // Initial state
2864 Const, // Input values are const
2865 Platform, // Platform order
2866 Reverse, // Reverse platform order
2867 NotAdjacent // Not adjacent
2868 };
2869 ValueOrder _value_order;
2870
2871 NOT_PRODUCT( const CHeapBitMap &_trace_tags; )
2872
2873 public:
2874 MergePrimitiveStores(PhaseGVN* phase, StoreNode* store) :
2875 _phase(phase), _store(store), _value_order(ValueOrder::Unknown)
2876 NOT_PRODUCT( COMMA _trace_tags(Compile::current()->directive()->trace_merge_stores_tags()) )
2877 {}
2878
2879 StoreNode* run();
2880
2881 private:
2882 bool is_compatible_store(const StoreNode* other_store) const;
2883 bool is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const;
2884 bool is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const;
2885 static bool is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase);
2886 enum CFGStatus { SuccessNoRangeCheck, SuccessWithRangeCheck, Failure };
2887 static CFGStatus cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store);
2888
2889 class Status {
2890 private:
2891 StoreNode* _found_store;
2892 bool _found_range_check;
2893
2894 Status(StoreNode* found_store, bool found_range_check)
2895 : _found_store(found_store), _found_range_check(found_range_check) {}
2896
2897 public:
2898 StoreNode* found_store() const { return _found_store; }
2899 bool found_range_check() const { return _found_range_check; }
2900 static Status make_failure() { return Status(nullptr, false); }
2901
2902 static Status make(StoreNode* found_store, const CFGStatus cfg_status) {
2903 if (cfg_status == CFGStatus::Failure) {
2904 return Status::make_failure();
2905 }
2906 return Status(found_store, cfg_status == CFGStatus::SuccessWithRangeCheck);
2907 }
2908
2909 #ifndef PRODUCT
2910 void print_on(outputStream* st) const {
2911 if (_found_store == nullptr) {
2912 st->print_cr("None");
2913 } else {
2914 st->print_cr("Found[%d %s, %s]", _found_store->_idx, _found_store->Name(),
2915 _found_range_check ? "RC" : "no-RC");
2916 }
2917 }
2918 #endif
2919 };
2920
2921 enum ValueOrder find_adjacent_input_value_order(const Node* n1, const Node* n2, const int memory_size) const;
2922 Status find_adjacent_use_store(const StoreNode* def_store) const;
2923 Status find_adjacent_def_store(const StoreNode* use_store) const;
2924 Status find_use_store(const StoreNode* def_store) const;
2925 Status find_def_store(const StoreNode* use_store) const;
2926 Status find_use_store_unidirectional(const StoreNode* def_store) const;
2927 Status find_def_store_unidirectional(const StoreNode* use_store) const;
2928
2929 void collect_merge_list(Node_List& merge_list) const;
2930 Node* make_merged_input_value(const Node_List& merge_list);
2931 StoreNode* make_merged_store(const Node_List& merge_list, Node* merged_input_value);
2932
2933 #ifndef PRODUCT
2934 // Access to TraceMergeStores tags
2935 bool is_trace(TraceMergeStores::Tag tag) const {
2936 return _trace_tags.at(tag);
2937 }
2938
2939 bool is_trace_basic() const {
2940 return is_trace(TraceMergeStores::Tag::BASIC);
2941 }
2942
2943 bool is_trace_pointer_parsing() const {
2944 return is_trace(TraceMergeStores::Tag::POINTER_PARSING);
2945 }
2946
2947 bool is_trace_pointer_aliasing() const {
2948 return is_trace(TraceMergeStores::Tag::POINTER_ALIASING);
2949 }
2950
2951 bool is_trace_pointer_adjacency() const {
2952 return is_trace(TraceMergeStores::Tag::POINTER_ADJACENCY);
2953 }
2954
2955 bool is_trace_success() const {
2956 return is_trace(TraceMergeStores::Tag::SUCCESS);
2957 }
2958 #endif
2959
2960 NOT_PRODUCT( void trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const; )
2961 };
2962
2963 StoreNode* MergePrimitiveStores::run() {
2964 // Check for B/S/C/I
2965 int opc = _store->Opcode();
2966 if (opc != Op_StoreB && opc != Op_StoreC && opc != Op_StoreI) {
2967 return nullptr;
2968 }
2969
2970 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] MergePrimitiveStores::run: "); _store->dump(); })
2971
2972 // The _store must be the "last" store in a chain. If we find a use we could merge with
2973 // then that use or a store further down is the "last" store.
2974 Status status_use = find_adjacent_use_store(_store);
2975 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect no use: "); status_use.print_on(tty); })
2976 if (status_use.found_store() != nullptr) {
2977 return nullptr;
2978 }
2979
2980 // Check if we can merge with at least one def, so that we have at least 2 stores to merge.
2981 Status status_def = find_adjacent_def_store(_store);
2982 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect def: "); status_def.print_on(tty); })
2983 Node* def_store = status_def.found_store();
2984 if (def_store == nullptr) {
2985 return nullptr;
2986 }
2987
2988 // Initialize value order
2989 _value_order = find_adjacent_input_value_order(def_store->in(MemNode::ValueIn),
2990 _store->in(MemNode::ValueIn),
2991 _store->memory_size());
2992 assert(_value_order != ValueOrder::NotAdjacent && _value_order != ValueOrder::Unknown, "Order should be checked");
2993
2994 ResourceMark rm;
2995 Node_List merge_list;
2996 collect_merge_list(merge_list);
2997
2998 Node* merged_input_value = make_merged_input_value(merge_list);
2999 if (merged_input_value == nullptr) { return nullptr; }
3000
3001 StoreNode* merged_store = make_merged_store(merge_list, merged_input_value);
3002
3003 NOT_PRODUCT( if (is_trace_success()) { trace(merge_list, merged_input_value, merged_store); } )
3004
3005 return merged_store;
3006 }
3007
3008 // Check compatibility between _store and other_store.
3009 bool MergePrimitiveStores::is_compatible_store(const StoreNode* other_store) const {
3010 int opc = _store->Opcode();
3011 assert(opc == Op_StoreB || opc == Op_StoreC || opc == Op_StoreI, "precondition");
3012
3013 if (other_store == nullptr ||
3014 _store->Opcode() != other_store->Opcode()) {
3015 return false;
3016 }
3017
3018 return true;
3019 }
3020
3021 bool MergePrimitiveStores::is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const {
3022 if (!is_adjacent_input_pair(def_store->in(MemNode::ValueIn),
3023 use_store->in(MemNode::ValueIn),
3024 def_store->memory_size())) {
3025 return false;
3026 }
3027
3028 ResourceMark rm;
3029 #ifndef PRODUCT
3030 const TraceMemPointer trace(is_trace_pointer_parsing(),
3031 is_trace_pointer_aliasing(),
3032 is_trace_pointer_adjacency(),
3033 true);
3034 #endif
3035 const MemPointer pointer_use(use_store NOT_PRODUCT(COMMA trace));
3036 const MemPointer pointer_def(def_store NOT_PRODUCT(COMMA trace));
3037 return pointer_def.is_adjacent_to_and_before(pointer_use);
3038 }
3039
3040 // Check input values n1 and n2 can be merged and return the value order
3041 MergePrimitiveStores::ValueOrder MergePrimitiveStores::find_adjacent_input_value_order(const Node* n1, const Node* n2,
3042 const int memory_size) const {
3043 // Pattern: [n1 = ConI, n2 = ConI]
3044 if (n1->Opcode() == Op_ConI && n2->Opcode() == Op_ConI) {
3045 return ValueOrder::Const;
3046 }
3047
3048 Node const *base_n2;
3049 jint shift_n2;
3050 if (!is_con_RShift(n2, base_n2, shift_n2, _phase)) {
3051 return ValueOrder::NotAdjacent;
3052 }
3053 Node const *base_n1;
3054 jint shift_n1;
3055 if (!is_con_RShift(n1, base_n1, shift_n1, _phase)) {
3056 return ValueOrder::NotAdjacent;
3057 }
3058
3059 int bits_per_store = memory_size * 8;
3060 if (base_n1 != base_n2 ||
3061 abs(shift_n1 - shift_n2) != bits_per_store ||
3062 shift_n1 % bits_per_store != 0) {
3063 // Values are not adjacent
3064 return ValueOrder::NotAdjacent;
3065 }
3066
3067 // Detect value order
3068 #ifdef VM_LITTLE_ENDIAN
3069 return shift_n1 < shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3070 : ValueOrder::Reverse; // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3071 #else
3072 return shift_n1 > shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3073 : ValueOrder::Reverse; // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3074 #endif
3075 }
3076
3077 bool MergePrimitiveStores::is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const {
3078 ValueOrder input_value_order = find_adjacent_input_value_order(n1, n2, memory_size);
3079
3080 switch (input_value_order) {
3081 case ValueOrder::NotAdjacent:
3082 return false;
3083 case ValueOrder::Reverse:
3084 if (memory_size != 1 ||
3085 !Matcher::match_rule_supported(Op_ReverseBytesS) ||
3086 !Matcher::match_rule_supported(Op_ReverseBytesI) ||
3087 !Matcher::match_rule_supported(Op_ReverseBytesL)) {
3088 // ReverseBytes are not supported by platform
3089 return false;
3090 }
3091 // fall-through.
3092 case ValueOrder::Const:
3093 case ValueOrder::Platform:
3094 if (_value_order == ValueOrder::Unknown) {
3095 // Initial state is Unknown, and we find a valid input value order
3096 return true;
3097 }
3098 // The value order can not be changed
3099 return _value_order == input_value_order;
3100 case ValueOrder::Unknown:
3101 default:
3102 ShouldNotReachHere();
3103 }
3104 return false;
3105 }
3106
3107 // Detect pattern: n = base_out >> shift_out
3108 bool MergePrimitiveStores::is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase) {
3109 assert(n != nullptr, "precondition");
3110
3111 int opc = n->Opcode();
3112 if (opc == Op_ConvL2I) {
3113 n = n->in(1);
3114 opc = n->Opcode();
3115 }
3116
3117 if ((opc == Op_RShiftI ||
3118 opc == Op_RShiftL ||
3119 opc == Op_URShiftI ||
3120 opc == Op_URShiftL) &&
3121 n->in(2)->is_ConI()) {
3122 base_out = n->in(1);
3123 shift_out = n->in(2)->get_int();
3124 // The shift must be positive:
3125 return shift_out >= 0;
3126 }
3127
3128 if (phase->type(n)->isa_int() != nullptr ||
3129 phase->type(n)->isa_long() != nullptr) {
3130 // (base >> 0)
3131 base_out = n;
3132 shift_out = 0;
3133 return true;
3134 }
3135 return false;
3136 }
3137
3138 // Check if there is nothing between the two stores, except optionally a RangeCheck leading to an uncommon trap.
3139 MergePrimitiveStores::CFGStatus MergePrimitiveStores::cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store) {
3140 assert(use_store->in(MemNode::Memory) == def_store, "use-def relationship");
3141
3142 Node* ctrl_use = use_store->in(MemNode::Control);
3143 Node* ctrl_def = def_store->in(MemNode::Control);
3144 if (ctrl_use == nullptr || ctrl_def == nullptr) {
3145 return CFGStatus::Failure;
3146 }
3147
3148 if (ctrl_use == ctrl_def) {
3149 // Same ctrl -> no RangeCheck in between.
3150 // Check: use_store must be the only use of def_store.
3151 if (def_store->outcnt() > 1) {
3152 return CFGStatus::Failure;
3153 }
3154 return CFGStatus::SuccessNoRangeCheck;
3155 }
3156
3157 // Different ctrl -> could have RangeCheck in between.
3158 // Check: 1. def_store only has these uses: use_store and MergeMem for uncommon trap, and
3159 // 2. ctrl separated by RangeCheck.
3160 if (def_store->outcnt() != 2) {
3161 return CFGStatus::Failure; // Cannot have exactly these uses: use_store and MergeMem for uncommon trap.
3162 }
3163 int use_store_out_idx = def_store->raw_out(0) == use_store ? 0 : 1;
3164 Node* merge_mem = def_store->raw_out(1 - use_store_out_idx)->isa_MergeMem();
3165 if (merge_mem == nullptr ||
3166 merge_mem->outcnt() != 1) {
3167 return CFGStatus::Failure; // Does not have MergeMem for uncommon trap.
3168 }
3169 if (!ctrl_use->is_IfProj() ||
3170 !ctrl_use->in(0)->is_RangeCheck() ||
3171 ctrl_use->in(0)->outcnt() != 2) {
3172 return CFGStatus::Failure; // Not RangeCheck.
3173 }
3174 ProjNode* other_proj = ctrl_use->as_IfProj()->other_if_proj();
3175 Node* trap = other_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check);
3176 if (trap != merge_mem->unique_out() ||
3177 ctrl_use->in(0)->in(0) != ctrl_def) {
3178 return CFGStatus::Failure; // Not RangeCheck with merge_mem leading to uncommon trap.
3179 }
3180
3181 return CFGStatus::SuccessWithRangeCheck;
3182 }
3183
3184 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_use_store(const StoreNode* def_store) const {
3185 Status status_use = find_use_store(def_store);
3186 StoreNode* use_store = status_use.found_store();
3187 if (use_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3188 return Status::make_failure();
3189 }
3190 return status_use;
3191 }
3192
3193 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_def_store(const StoreNode* use_store) const {
3194 Status status_def = find_def_store(use_store);
3195 StoreNode* def_store = status_def.found_store();
3196 if (def_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3197 return Status::make_failure();
3198 }
3199 return status_def;
3200 }
3201
3202 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store(const StoreNode* def_store) const {
3203 Status status_use = find_use_store_unidirectional(def_store);
3204
3205 #ifdef ASSERT
3206 StoreNode* use_store = status_use.found_store();
3207 if (use_store != nullptr) {
3208 Status status_def = find_def_store_unidirectional(use_store);
3209 assert(status_def.found_store() == def_store &&
3210 status_def.found_range_check() == status_use.found_range_check(),
3211 "find_use_store and find_def_store must be symmetric");
3212 }
3213 #endif
3214
3215 return status_use;
3216 }
3217
3218 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store(const StoreNode* use_store) const {
3219 Status status_def = find_def_store_unidirectional(use_store);
3220
3221 #ifdef ASSERT
3222 StoreNode* def_store = status_def.found_store();
3223 if (def_store != nullptr) {
3224 Status status_use = find_use_store_unidirectional(def_store);
3225 assert(status_use.found_store() == use_store &&
3226 status_use.found_range_check() == status_def.found_range_check(),
3227 "find_use_store and find_def_store must be symmetric");
3228 }
3229 #endif
3230
3231 return status_def;
3232 }
3233
3234 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store_unidirectional(const StoreNode* def_store) const {
3235 assert(is_compatible_store(def_store), "precondition: must be compatible with _store");
3236
3237 for (DUIterator_Fast imax, i = def_store->fast_outs(imax); i < imax; i++) {
3238 StoreNode* use_store = def_store->fast_out(i)->isa_Store();
3239 if (is_compatible_store(use_store)) {
3240 return Status::make(use_store, cfg_status_for_pair(use_store, def_store));
3241 }
3242 }
3243
3244 return Status::make_failure();
3245 }
3246
3247 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store_unidirectional(const StoreNode* use_store) const {
3248 assert(is_compatible_store(use_store), "precondition: must be compatible with _store");
3249
3250 StoreNode* def_store = use_store->in(MemNode::Memory)->isa_Store();
3251 if (!is_compatible_store(def_store)) {
3252 return Status::make_failure();
3253 }
3254
3255 return Status::make(def_store, cfg_status_for_pair(use_store, def_store));
3256 }
3257
3258 void MergePrimitiveStores::collect_merge_list(Node_List& merge_list) const {
3259 // The merged store can be at most 8 bytes.
3260 const uint merge_list_max_size = 8 / _store->memory_size();
3261 assert(merge_list_max_size >= 2 &&
3262 merge_list_max_size <= 8 &&
3263 is_power_of_2(merge_list_max_size),
3264 "must be 2, 4 or 8");
3265
3266 // Traverse up the chain of adjacent def stores.
3267 StoreNode* current = _store;
3268 merge_list.push(current);
3269 while (current != nullptr && merge_list.size() < merge_list_max_size) {
3270 Status status = find_adjacent_def_store(current);
3271 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] find def: "); status.print_on(tty); })
3272
3273 current = status.found_store();
3274 if (current != nullptr) {
3275 merge_list.push(current);
3276
3277 // We can have at most one RangeCheck.
3278 if (status.found_range_check()) {
3279 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found RangeCheck, stop traversal."); })
3280 break;
3281 }
3282 }
3283 }
3284
3285 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found:"); merge_list.dump(); })
3286
3287 // Truncate the merge_list to a power of 2.
3288 const uint pow2size = round_down_power_of_2(merge_list.size());
3289 assert(pow2size >= 2, "must be merging at least 2 stores");
3290 while (merge_list.size() > pow2size) { merge_list.pop(); }
3291
3292 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] truncated:"); merge_list.dump(); })
3293 }
3294
3295 // Merge the input values of the smaller stores to a single larger input value.
3296 Node* MergePrimitiveStores::make_merged_input_value(const Node_List& merge_list) {
3297 int new_memory_size = _store->memory_size() * merge_list.size();
3298 Node* first = merge_list.at(merge_list.size()-1);
3299 Node* merged_input_value = nullptr;
3300 if (_store->in(MemNode::ValueIn)->Opcode() == Op_ConI) {
3301 assert(_value_order == ValueOrder::Const, "must match");
3302 // Pattern: [ConI, ConI, ...] -> new constant
3303 jlong con = 0;
3304 jlong bits_per_store = _store->memory_size() * 8;
3305 jlong mask = (((jlong)1) << bits_per_store) - 1;
3306 for (uint i = 0; i < merge_list.size(); i++) {
3307 jlong con_i = merge_list.at(i)->in(MemNode::ValueIn)->get_int();
3308 #ifdef VM_LITTLE_ENDIAN
3309 con = con << bits_per_store;
3310 con = con | (mask & con_i);
3311 #else // VM_LITTLE_ENDIAN
3312 con_i = (mask & con_i) << (i * bits_per_store);
3313 con = con | con_i;
3314 #endif // VM_LITTLE_ENDIAN
3315 }
3316 merged_input_value = _phase->longcon(con);
3317 } else {
3318 assert(_value_order == ValueOrder::Platform || _value_order == ValueOrder::Reverse, "must match");
3319 // Pattern: [base >> 24, base >> 16, base >> 8, base] -> base
3320 // | |
3321 // _store first
3322 //
3323 Node* hi = _store->in(MemNode::ValueIn);
3324 Node* lo = first->in(MemNode::ValueIn);
3325 #ifndef VM_LITTLE_ENDIAN
3326 // `_store` and `first` are swapped in the diagram above
3327 swap(hi, lo);
3328 #endif // !VM_LITTLE_ENDIAN
3329 if (_value_order == ValueOrder::Reverse) {
3330 swap(hi, lo);
3331 }
3332 Node const* hi_base;
3333 jint hi_shift;
3334 merged_input_value = lo;
3335 bool is_true = is_con_RShift(hi, hi_base, hi_shift, _phase);
3336 assert(is_true, "must detect con RShift");
3337 if (merged_input_value != hi_base && merged_input_value->Opcode() == Op_ConvL2I) {
3338 // look through
3339 merged_input_value = merged_input_value->in(1);
3340 }
3341 if (merged_input_value != hi_base) {
3342 // merged_input_value is not the base
3343 return nullptr;
3344 }
3345 }
3346
3347 if (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size <= 4) {
3348 // Example:
3349 //
3350 // long base = ...;
3351 // a[0] = (byte)(base >> 0);
3352 // a[1] = (byte)(base >> 8);
3353 //
3354 merged_input_value = _phase->transform(new ConvL2INode(merged_input_value));
3355 }
3356
3357 assert((_phase->type(merged_input_value)->isa_int() != nullptr && new_memory_size <= 4) ||
3358 (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size == 8),
3359 "merged_input_value is either int or long, and new_memory_size is small enough");
3360
3361 if (_value_order == ValueOrder::Reverse) {
3362 assert(_store->memory_size() == 1, "only implemented for bytes");
3363 if (new_memory_size == 8) {
3364 merged_input_value = _phase->transform(new ReverseBytesLNode(merged_input_value));
3365 } else if (new_memory_size == 4) {
3366 merged_input_value = _phase->transform(new ReverseBytesINode(merged_input_value));
3367 } else {
3368 assert(new_memory_size == 2, "sanity check");
3369 merged_input_value = _phase->transform(new ReverseBytesSNode(merged_input_value));
3370 }
3371 }
3372 return merged_input_value;
3373 }
3374
3375 // //
3376 // first_ctrl first_mem first_adr first_ctrl first_mem first_adr //
3377 // | | | | | | //
3378 // | | | | +---------------+ | //
3379 // | | | | | | | //
3380 // | | +---------+ | | +---------------+ //
3381 // | | | | | | | | //
3382 // +--------------+ | | v1 +------------------------------+ | | v1 //
3383 // | | | | | | | | | | | | //
3384 // RangeCheck first_store RangeCheck | | first_store //
3385 // | | | | | | | //
3386 // last_ctrl | +----> unc_trap last_ctrl | | +----> unc_trap //
3387 // | | ===> | | | //
3388 // +--------------+ | a2 v2 | | | //
3389 // | | | | | | | | //
3390 // | second_store | | | //
3391 // | | | | | [v1 v2 ... vn] //
3392 // ... ... | | | | //
3393 // | | | | | v //
3394 // +--------------+ | an vn +--------------+ | | merged_input_value //
3395 // | | | | | | | | //
3396 // last_store (= _store) merged_store //
3397 // //
3398 StoreNode* MergePrimitiveStores::make_merged_store(const Node_List& merge_list, Node* merged_input_value) {
3399 Node* first_store = merge_list.at(merge_list.size()-1);
3400 Node* last_ctrl = _store->in(MemNode::Control); // after (optional) RangeCheck
3401 Node* first_mem = first_store->in(MemNode::Memory);
3402 Node* first_adr = first_store->in(MemNode::Address);
3403
3404 const TypePtr* new_adr_type = _store->adr_type();
3405
3406 int new_memory_size = _store->memory_size() * merge_list.size();
3407 BasicType bt = T_ILLEGAL;
3408 switch (new_memory_size) {
3409 case 2: bt = T_SHORT; break;
3410 case 4: bt = T_INT; break;
3411 case 8: bt = T_LONG; break;
3412 }
3413
3414 StoreNode* merged_store = StoreNode::make(*_phase, last_ctrl, first_mem, first_adr,
3415 new_adr_type, merged_input_value, bt, MemNode::unordered);
3416
3417 // Marking the store mismatched is sufficient to prevent reordering, since array stores
3418 // are all on the same slice. Hence, we need no barriers.
3419 merged_store->set_mismatched_access();
3420
3421 // Constants above may now also be be packed -> put candidate on worklist
3422 _phase->is_IterGVN()->_worklist.push(first_mem);
3423
3424 return merged_store;
3425 }
3426
3427 #ifndef PRODUCT
3428 void MergePrimitiveStores::trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const {
3429 stringStream ss;
3430 ss.print_cr("[TraceMergeStores]: Replace");
3431 for (int i = (int)merge_list.size() - 1; i >= 0; i--) {
3432 merge_list.at(i)->dump("\n", false, &ss);
3433 }
3434 ss.print_cr("[TraceMergeStores]: with");
3435 merged_input_value->dump("\n", false, &ss);
3436 merged_store->dump("\n", false, &ss);
3437 tty->print("%s", ss.as_string());
3438 }
3439 #endif
3440
3441 //------------------------------Ideal------------------------------------------
3442 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
3443 // When a store immediately follows a relevant allocation/initialization,
3444 // try to capture it into the initialization, or hoist it above.
3445 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3446 Node* p = MemNode::Ideal_common(phase, can_reshape);
3447 if (p) return (p == NodeSentinel) ? nullptr : p;
3448
3449 Node* mem = in(MemNode::Memory);
3450 Node* address = in(MemNode::Address);
3451 Node* value = in(MemNode::ValueIn);
3452 // Back-to-back stores to same address? Fold em up. Generally
3453 // unsafe if I have intervening uses...
3454 if (phase->C->get_adr_type(phase->C->get_alias_index(adr_type())) != TypeAryPtr::INLINES) {
3455 Node* st = mem;
3456 // If Store 'st' has more than one use, we cannot fold 'st' away.
3457 // For example, 'st' might be the final state at a conditional
3458 // return. Or, 'st' might be used by some node which is live at
3459 // the same time 'st' is live, which might be unschedulable. So,
3460 // require exactly ONE user until such time as we clone 'mem' for
3461 // each of 'mem's uses (thus making the exactly-1-user-rule hold
3462 // true).
3463 while (st->is_Store() && st->outcnt() == 1) {
3464 // Looking at a dead closed cycle of memory?
3465 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
3466 assert(Opcode() == st->Opcode() ||
3467 st->Opcode() == Op_StoreVector ||
3468 Opcode() == Op_StoreVector ||
3469 st->Opcode() == Op_StoreVectorScatter ||
3470 Opcode() == Op_StoreVectorScatter ||
3471 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
3472 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
3473 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
3474 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) ||
3475 (st->adr_type()->isa_aryptr() && st->adr_type()->is_aryptr()->is_flat()) || // TODO 8343835
3476 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
3477 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
3478
3479 if (st->in(MemNode::Address)->eqv_uncast(address) &&
3480 st->as_Store()->memory_size() <= this->memory_size()) {
3481 Node* use = st->raw_out(0);
3482 if (phase->is_IterGVN()) {
3483 phase->is_IterGVN()->rehash_node_delayed(use);
3484 }
3485 // It's OK to do this in the parser, since DU info is always accurate,
3486 // and the parser always refers to nodes via SafePointNode maps.
3487 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
3488 return this;
3489 }
3490 st = st->in(MemNode::Memory);
3491 }
3492 }
3493
3494
3495 // Capture an unaliased, unconditional, simple store into an initializer.
3496 // Or, if it is independent of the allocation, hoist it above the allocation.
3497 if (ReduceFieldZeroing && /*can_reshape &&*/
3498 mem->is_Proj() && mem->in(0)->is_Initialize()) {
3499 InitializeNode* init = mem->in(0)->as_Initialize();
3500 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
3501 if (offset > 0) {
3502 Node* moved = init->capture_store(this, offset, phase, can_reshape);
3503 // If the InitializeNode captured me, it made a raw copy of me,
3504 // and I need to disappear.
3505 if (moved != nullptr) {
3506 // %%% hack to ensure that Ideal returns a new node:
3507 mem = MergeMemNode::make(mem);
3508 return mem; // fold me away
3509 }
3510 }
3511 }
3512
3513 // Fold reinterpret cast into memory operation:
3514 // StoreX mem (MoveY2X v) => StoreY mem v
3515 if (value->is_Move()) {
3516 const Type* vt = value->in(1)->bottom_type();
3517 if (has_reinterpret_variant(vt)) {
3518 if (phase->C->post_loop_opts_phase()) {
3519 return convert_to_reinterpret_store(*phase, value->in(1), vt);
3520 } else {
3521 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
3522 }
3523 }
3524 }
3525
3526 if (MergeStores && UseUnalignedAccesses) {
3527 if (phase->C->merge_stores_phase()) {
3528 MergePrimitiveStores merge(phase, this);
3529 Node* progress = merge.run();
3530 if (progress != nullptr) { return progress; }
3531 } else {
3532 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
3533 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
3534 // the stores, and we merge the wrong sequence of stores.
3535 // Example:
3536 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
3537 // Apply MergeStores:
3538 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
3539 // Remove more RangeChecks:
3540 // StoreI [ StoreL ] StoreI
3541 // But now it would have been better to do this instead:
3542 // [ StoreL ] [ StoreL ]
3543 phase->C->record_for_merge_stores_igvn(this);
3544 }
3545 }
3546
3547 return nullptr; // No further progress
3548 }
3549
3550 //------------------------------Value-----------------------------------------
3551 const Type* StoreNode::Value(PhaseGVN* phase) const {
3552 // Either input is TOP ==> the result is TOP
3553 const Type *t1 = phase->type( in(MemNode::Memory) );
3554 if( t1 == Type::TOP ) return Type::TOP;
3555 const Type *t2 = phase->type( in(MemNode::Address) );
3556 if( t2 == Type::TOP ) return Type::TOP;
3557 const Type *t3 = phase->type( in(MemNode::ValueIn) );
3558 if( t3 == Type::TOP ) return Type::TOP;
3559 return Type::MEMORY;
3560 }
3561
3562 //------------------------------Identity---------------------------------------
3563 // Remove redundant stores:
3564 // Store(m, p, Load(m, p)) changes to m.
3565 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
3566 Node* StoreNode::Identity(PhaseGVN* phase) {
3567 Node* mem = in(MemNode::Memory);
3568 Node* adr = in(MemNode::Address);
3569 Node* val = in(MemNode::ValueIn);
3570
3571 Node* result = this;
3572
3573 // Load then Store? Then the Store is useless
3574 if (val->is_Load() &&
3575 val->in(MemNode::Address)->eqv_uncast(adr) &&
3576 val->in(MemNode::Memory )->eqv_uncast(mem) &&
3577 val->as_Load()->store_Opcode() == Opcode()) {
3578 // Ensure vector type is the same
3579 if (!is_StoreVector() || (mem->is_LoadVector() && as_StoreVector()->vect_type() == mem->as_LoadVector()->vect_type())) {
3580 result = mem;
3581 }
3582 }
3583
3584 // Two stores in a row of the same value?
3585 if (result == this &&
3586 mem->is_Store() &&
3587 mem->in(MemNode::Address)->eqv_uncast(adr) &&
3588 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
3589 mem->Opcode() == Opcode()) {
3590 if (!is_StoreVector()) {
3591 result = mem;
3592 } else {
3593 const StoreVectorNode* store_vector = as_StoreVector();
3594 const StoreVectorNode* mem_vector = mem->as_StoreVector();
3595 const Node* store_indices = store_vector->indices();
3596 const Node* mem_indices = mem_vector->indices();
3597 const Node* store_mask = store_vector->mask();
3598 const Node* mem_mask = mem_vector->mask();
3599 // Ensure types, indices, and masks match
3600 if (store_vector->vect_type() == mem_vector->vect_type() &&
3601 ((store_indices == nullptr) == (mem_indices == nullptr) &&
3602 (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) &&
3603 ((store_mask == nullptr) == (mem_mask == nullptr) &&
3604 (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) {
3605 result = mem;
3606 }
3607 }
3608 }
3609
3610 // Store of zero anywhere into a freshly-allocated object?
3611 // Then the store is useless.
3612 // (It must already have been captured by the InitializeNode.)
3613 if (result == this && ReduceFieldZeroing) {
3614 // a newly allocated object is already all-zeroes everywhere
3615 if (mem->is_Proj() && mem->in(0)->is_Allocate() &&
3616 (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::InitValue) == val)) {
3617 result = mem;
3618 }
3619
3620 if (result == this && phase->type(val)->is_zero_type()) {
3621 // the store may also apply to zero-bits in an earlier object
3622 Node* prev_mem = find_previous_store(phase);
3623 // Steps (a), (b): Walk past independent stores to find an exact match.
3624 if (prev_mem != nullptr) {
3625 Node* prev_val = can_see_stored_value(prev_mem, phase);
3626 if (prev_val != nullptr && prev_val == val) {
3627 // prev_val and val might differ by a cast; it would be good
3628 // to keep the more informative of the two.
3629 result = mem;
3630 }
3631 }
3632 }
3633 }
3634
3635 PhaseIterGVN* igvn = phase->is_IterGVN();
3636 if (result != this && igvn != nullptr) {
3637 MemBarNode* trailing = trailing_membar();
3638 if (trailing != nullptr) {
3639 #ifdef ASSERT
3640 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
3641 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects");
3642 #endif
3643 trailing->remove(igvn);
3644 }
3645 }
3646
3647 return result;
3648 }
3649
3650 //------------------------------match_edge-------------------------------------
3651 // Do we Match on this edge index or not? Match only memory & value
3652 uint StoreNode::match_edge(uint idx) const {
3653 return idx == MemNode::Address || idx == MemNode::ValueIn;
3654 }
3655
3656 //------------------------------cmp--------------------------------------------
3657 // Do not common stores up together. They generally have to be split
3658 // back up anyways, so do not bother.
3659 bool StoreNode::cmp( const Node &n ) const {
3660 return (&n == this); // Always fail except on self
3661 }
3662
3663 //------------------------------Ideal_masked_input-----------------------------
3664 // Check for a useless mask before a partial-word store
3665 // (StoreB ... (AndI valIn conIa) )
3666 // If (conIa & mask == mask) this simplifies to
3667 // (StoreB ... (valIn) )
3668 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
3669 Node *val = in(MemNode::ValueIn);
3670 if( val->Opcode() == Op_AndI ) {
3671 const TypeInt *t = phase->type( val->in(2) )->isa_int();
3672 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
3673 set_req_X(MemNode::ValueIn, val->in(1), phase);
3674 return this;
3675 }
3676 }
3677 return nullptr;
3678 }
3679
3680
3681 //------------------------------Ideal_sign_extended_input----------------------
3682 // Check for useless sign-extension before a partial-word store
3683 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
3684 // If (conIL == conIR && conIR <= num_rejected_bits) this simplifies to
3685 // (StoreB ... (v))
3686 // If (conIL > conIR) under some conditions, it can be simplified into
3687 // (StoreB ... (LShiftI _ v (conIL - conIR)))
3688 // This case happens when the value of the store was itself a left shift, that
3689 // gets merged into the inner left shift of the sign-extension. For instance,
3690 // if we have
3691 // array_of_shorts[0] = (short)(v << 2)
3692 // We get a structure such as:
3693 // (StoreB ... (RShiftI _ (LShiftI _ (LShiftI _ v 2) 16) 16))
3694 // that is simplified into
3695 // (StoreB ... (RShiftI _ (LShiftI _ v 18) 16)).
3696 // It is thus useful to handle cases where conIL > conIR. But this simplification
3697 // does not always hold. Let's see in which cases it's valid.
3698 //
3699 // Let's assume we have the following 32 bits integer v:
3700 // +----------------------------------+
3701 // | v[0..31] |
3702 // +----------------------------------+
3703 // 31 0
3704 // that will be stuffed in 8 bits byte after a shift left and a shift right of
3705 // potentially different magnitudes.
3706 // We denote num_rejected_bits the number of bits of the discarded part. In this
3707 // case, num_rejected_bits == 24.
3708 //
3709 // Statement (proved further below in case analysis):
3710 // Given:
3711 // - 0 <= conIL < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
3712 // - 0 <= conIR < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
3713 // - conIL >= conIR
3714 // - num_rejected_bits >= conIR
3715 // Then this form:
3716 // (RShiftI _ (LShiftI _ v conIL) conIR)
3717 // can be replaced with this form:
3718 // (LShiftI _ v (conIL-conIR))
3719 //
3720 // Note: We only have to show that the non-rejected lowest bits (8 bits for byte)
3721 // have to be correct, as the higher bits are rejected / truncated by the store.
3722 //
3723 // The hypotheses
3724 // 0 <= conIL < BitsPerJavaInteger
3725 // 0 <= conIR < BitsPerJavaInteger
3726 // are ensured by maskShiftAmount (called from ::Ideal of shift nodes). Indeed,
3727 // (v << 31) << 2 must be simplified into 0, not into v << 33 (which is equivalent
3728 // to v << 1).
3729 //
3730 //
3731 // If you don't like case analysis, jump after the conclusion.
3732 // ### Case 1 : conIL == conIR
3733 // ###### Case 1.1: conIL == conIR == num_rejected_bits
3734 // If we do the shift left then right by 24 bits, we get:
3735 // after: v << 24
3736 // +---------+------------------------+
3737 // | v[0..7] | 0 |
3738 // +---------+------------------------+
3739 // 31 24 23 0
3740 // after: (v << 24) >> 24
3741 // +------------------------+---------+
3742 // | sign bit | v[0..7] |
3743 // +------------------------+---------+
3744 // 31 8 7 0
3745 // The non-rejected bits (bits kept by the store, that is the 8 lower bits of the
3746 // result) are the same before and after, so, indeed, simplifying is correct.
3747
3748 // ###### Case 1.2: conIL == conIR < num_rejected_bits
3749 // If we do the shift left then right by 22 bits, we get:
3750 // after: v << 22
3751 // +---------+------------------------+
3752 // | v[0..9] | 0 |
3753 // +---------+------------------------+
3754 // 31 22 21 0
3755 // after: (v << 22) >> 22
3756 // +------------------------+---------+
3757 // | sign bit | v[0..9] |
3758 // +------------------------+---------+
3759 // 31 10 9 0
3760 // The non-rejected bits are the 8 lower bits of v. The bits 8 and 9 of v are still
3761 // present in (v << 22) >> 22 but will be dropped by the store. The simplification is
3762 // still correct.
3763
3764 // ###### But! Case 1.3: conIL == conIR > num_rejected_bits
3765 // If we do the shift left then right by 26 bits, we get:
3766 // after: v << 26
3767 // +---------+------------------------+
3768 // | v[0..5] | 0 |
3769 // +---------+------------------------+
3770 // 31 26 25 0
3771 // after: (v << 26) >> 26
3772 // +------------------------+---------+
3773 // | sign bit | v[0..5] |
3774 // +------------------------+---------+
3775 // 31 6 5 0
3776 // The non-rejected bits are made of
3777 // - 0-5 => the bits 0 to 5 of v
3778 // - 6-7 => the sign bit of v[0..5] (that is v[5])
3779 // Simplifying this as v is not correct.
3780 // The condition conIR <= num_rejected_bits is indeed necessary in Case 1
3781 //
3782 // ### Case 2: conIL > conIR
3783 // ###### Case 2.1: num_rejected_bits == conIR
3784 // We take conIL == 26 for this example.
3785 // after: v << 26
3786 // +---------+------------------------+
3787 // | v[0..5] | 0 |
3788 // +---------+------------------------+
3789 // 31 26 25 0
3790 // after: (v << 26) >> 24
3791 // +------------------+---------+-----+
3792 // | sign bit | v[0..5] | 0 |
3793 // +------------------+---------+-----+
3794 // 31 8 7 2 1 0
3795 // The non-rejected bits are the 8 lower ones of (v << conIL - conIR).
3796 // The bits 6 and 7 of v have been thrown away after the shift left.
3797 // The simplification is still correct.
3798 //
3799 // ###### Case 2.2: num_rejected_bits > conIR.
3800 // Let's say conIL == 26 and conIR == 22.
3801 // after: v << 26
3802 // +---------+------------------------+
3803 // | v[0..5] | 0 |
3804 // +---------+------------------------+
3805 // 31 26 25 0
3806 // after: (v << 26) >> 22
3807 // +------------------+---------+-----+
3808 // | sign bit | v[0..5] | 0 |
3809 // +------------------+---------+-----+
3810 // 31 10 9 4 3 0
3811 // The bits non-rejected by the store are exactly the 8 lower ones of (v << (conIL - conIR)):
3812 // - 0-3 => 0
3813 // - 4-7 => bits 0 to 3 of v
3814 // The simplification is still correct.
3815 // The bits 4 and 5 of v are still present in (v << (conIL - conIR)) but they don't
3816 // matter as they are not in the 8 lower bits: they will be cut out by the store.
3817 //
3818 // ###### But! Case 2.3: num_rejected_bits < conIR.
3819 // Let's see that this case is not as easy to simplify.
3820 // Let's say conIL == 28 and conIR == 26.
3821 // after: v << 28
3822 // +---------+------------------------+
3823 // | v[0..3] | 0 |
3824 // +---------+------------------------+
3825 // 31 28 27 0
3826 // after: (v << 28) >> 26
3827 // +------------------+---------+-----+
3828 // | sign bit | v[0..3] | 0 |
3829 // +------------------+---------+-----+
3830 // 31 6 5 2 1 0
3831 // The non-rejected bits are made of
3832 // - 0-1 => 0
3833 // - 2-5 => the bits 0 to 3 of v
3834 // - 6-7 => the sign bit of v[0..3] (that is v[3])
3835 // Simplifying this as (v << 2) is not correct.
3836 // The condition conIR <= num_rejected_bits is indeed necessary in Case 2.
3837 //
3838 // ### Conclusion:
3839 // Our hypotheses are indeed sufficient:
3840 // - 0 <= conIL < BitsPerJavaInteger
3841 // - 0 <= conIR < BitsPerJavaInteger
3842 // - conIL >= conIR
3843 // - num_rejected_bits >= conIR
3844 //
3845 // ### A rationale without case analysis:
3846 // After the shift left, conIL upper bits of v are discarded and conIL lower bit
3847 // zeroes are added. After the shift right, conIR lower bits of the previous result
3848 // are discarded. If conIL >= conIR, we discard only the zeroes we made up during
3849 // the shift left, but if conIL < conIR, then we discard also lower bits of v. But
3850 // the point of the simplification is to get an expression of the form
3851 // (v << (conIL - conIR)). This expression discard only higher bits of v, thus the
3852 // simplification is not correct if conIL < conIR.
3853 //
3854 // Moreover, after the shift right, the higher bit of (v << conIL) is repeated on the
3855 // conIR higher bits of ((v << conIL) >> conIR), it's the sign-extension. If
3856 // conIR > num_rejected_bits, then at least one artificial copy of this sign bit will
3857 // be in the window of the store. Thus ((v << conIL) >> conIR) is not equivalent to
3858 // (v << (conIL-conIR)) if conIR > num_rejected_bits.
3859 //
3860 // We do not treat the case conIL < conIR here since the point of this function is
3861 // to skip sign-extensions (that is conIL == conIR == num_rejected_bits). The need
3862 // of treating conIL > conIR comes from the cases where the sign-extended value is
3863 // also left-shift expression. Computing the sign-extension of a right-shift expression
3864 // doesn't yield a situation such as
3865 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
3866 // where conIL < conIR.
3867 Node* StoreNode::Ideal_sign_extended_input(PhaseGVN* phase, int num_rejected_bits) {
3868 Node* shr = in(MemNode::ValueIn);
3869 if (shr->Opcode() == Op_RShiftI) {
3870 const TypeInt* conIR = phase->type(shr->in(2))->isa_int();
3871 if (conIR != nullptr && conIR->is_con() && conIR->get_con() >= 0 && conIR->get_con() < BitsPerJavaInteger && conIR->get_con() <= num_rejected_bits) {
3872 Node* shl = shr->in(1);
3873 if (shl->Opcode() == Op_LShiftI) {
3874 const TypeInt* conIL = phase->type(shl->in(2))->isa_int();
3875 if (conIL != nullptr && conIL->is_con() && conIL->get_con() >= 0 && conIL->get_con() < BitsPerJavaInteger) {
3876 if (conIL->get_con() == conIR->get_con()) {
3877 set_req_X(MemNode::ValueIn, shl->in(1), phase);
3878 return this;
3879 }
3880 if (conIL->get_con() > conIR->get_con()) {
3881 Node* new_shl = phase->transform(new LShiftINode(shl->in(1), phase->intcon(conIL->get_con() - conIR->get_con())));
3882 set_req_X(MemNode::ValueIn, new_shl, phase);
3883 return this;
3884 }
3885 }
3886 }
3887 }
3888 }
3889 return nullptr;
3890 }
3891
3892 //------------------------------value_never_loaded-----------------------------------
3893 // Determine whether there are any possible loads of the value stored.
3894 // For simplicity, we actually check if there are any loads from the
3895 // address stored to, not just for loads of the value stored by this node.
3896 //
3897 bool StoreNode::value_never_loaded(PhaseValues* phase) const {
3898 Node *adr = in(Address);
3899 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
3900 if (adr_oop == nullptr)
3901 return false;
3902 if (!adr_oop->is_known_instance_field())
3903 return false; // if not a distinct instance, there may be aliases of the address
3904 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
3905 Node *use = adr->fast_out(i);
3906 if (use->is_Load() || use->is_LoadStore()) {
3907 return false;
3908 }
3909 }
3910 return true;
3911 }
3912
3913 MemBarNode* StoreNode::trailing_membar() const {
3914 if (is_release()) {
3915 MemBarNode* trailing_mb = nullptr;
3916 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3917 Node* u = fast_out(i);
3918 if (u->is_MemBar()) {
3919 if (u->as_MemBar()->trailing_store()) {
3920 assert(u->Opcode() == Op_MemBarVolatile, "");
3921 assert(trailing_mb == nullptr, "only one");
3922 trailing_mb = u->as_MemBar();
3923 #ifdef ASSERT
3924 Node* leading = u->as_MemBar()->leading_membar();
3925 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3926 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
3927 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
3928 #endif
3929 } else {
3930 assert(u->as_MemBar()->standalone(), "");
3931 }
3932 }
3933 }
3934 return trailing_mb;
3935 }
3936 return nullptr;
3937 }
3938
3939
3940 //=============================================================================
3941 //------------------------------Ideal------------------------------------------
3942 // If the store is from an AND mask that leaves the low bits untouched, then
3943 // we can skip the AND operation. If the store is from a sign-extension
3944 // (a left shift, then right shift) we can skip both.
3945 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
3946 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
3947 if( progress != nullptr ) return progress;
3948
3949 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
3950 if( progress != nullptr ) return progress;
3951
3952 // Finally check the default case
3953 return StoreNode::Ideal(phase, can_reshape);
3954 }
3955
3956 //=============================================================================
3957 //------------------------------Ideal------------------------------------------
3958 // If the store is from an AND mask that leaves the low bits untouched, then
3959 // we can skip the AND operation
3960 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
3961 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
3962 if( progress != nullptr ) return progress;
3963
3964 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
3965 if( progress != nullptr ) return progress;
3966
3967 // Finally check the default case
3968 return StoreNode::Ideal(phase, can_reshape);
3969 }
3970
3971 //=============================================================================
3972 //----------------------------------SCMemProjNode------------------------------
3973 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
3974 {
3975 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
3976 return Type::TOP;
3977 }
3978 return bottom_type();
3979 }
3980
3981 //=============================================================================
3982 //----------------------------------LoadStoreNode------------------------------
3983 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
3984 : Node(required),
3985 _type(rt),
3986 _adr_type(at),
3987 _barrier_data(0)
3988 {
3989 init_req(MemNode::Control, c );
3990 init_req(MemNode::Memory , mem);
3991 init_req(MemNode::Address, adr);
3992 init_req(MemNode::ValueIn, val);
3993 init_class_id(Class_LoadStore);
3994 }
3995
3996 //------------------------------Value-----------------------------------------
3997 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
3998 // Either input is TOP ==> the result is TOP
3999 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
4000 return Type::TOP;
4001 }
4002 const Type* t = phase->type(in(MemNode::Memory));
4003 if (t == Type::TOP) {
4004 return Type::TOP;
4005 }
4006 t = phase->type(in(MemNode::Address));
4007 if (t == Type::TOP) {
4008 return Type::TOP;
4009 }
4010 t = phase->type(in(MemNode::ValueIn));
4011 if (t == Type::TOP) {
4012 return Type::TOP;
4013 }
4014 return bottom_type();
4015 }
4016
4017 uint LoadStoreNode::ideal_reg() const {
4018 return _type->ideal_reg();
4019 }
4020
4021 // This method conservatively checks if the result of a LoadStoreNode is
4022 // used, that is, if it returns true, then it is definitely the case that
4023 // the result of the node is not needed.
4024 // For example, GetAndAdd can be matched into a lock_add instead of a
4025 // lock_xadd if the result of LoadStoreNode::result_not_used() is true
4026 bool LoadStoreNode::result_not_used() const {
4027 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4028 Node *x = fast_out(i);
4029 if (x->Opcode() == Op_SCMemProj) {
4030 continue;
4031 }
4032 if (x->bottom_type() == TypeTuple::MEMBAR &&
4033 !x->is_Call() &&
4034 x->Opcode() != Op_Blackhole) {
4035 continue;
4036 }
4037 return false;
4038 }
4039 return true;
4040 }
4041
4042 MemBarNode* LoadStoreNode::trailing_membar() const {
4043 MemBarNode* trailing = nullptr;
4044 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4045 Node* u = fast_out(i);
4046 if (u->is_MemBar()) {
4047 if (u->as_MemBar()->trailing_load_store()) {
4048 assert(u->Opcode() == Op_MemBarAcquire, "");
4049 assert(trailing == nullptr, "only one");
4050 trailing = u->as_MemBar();
4051 #ifdef ASSERT
4052 Node* leading = trailing->leading_membar();
4053 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
4054 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
4055 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
4056 #endif
4057 } else {
4058 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
4059 }
4060 }
4061 }
4062
4063 return trailing;
4064 }
4065
4066 uint LoadStoreNode::size_of() const { return sizeof(*this); }
4067
4068 //=============================================================================
4069 //----------------------------------LoadStoreConditionalNode--------------------
4070 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) {
4071 init_req(ExpectedIn, ex );
4072 }
4073
4074 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
4075 // Either input is TOP ==> the result is TOP
4076 const Type* t = phase->type(in(ExpectedIn));
4077 if (t == Type::TOP) {
4078 return Type::TOP;
4079 }
4080 return LoadStoreNode::Value(phase);
4081 }
4082
4083 //=============================================================================
4084 //-------------------------------adr_type--------------------------------------
4085 const TypePtr* ClearArrayNode::adr_type() const {
4086 Node *adr = in(3);
4087 if (adr == nullptr) return nullptr; // node is dead
4088 return MemNode::calculate_adr_type(adr->bottom_type());
4089 }
4090
4091 //------------------------------match_edge-------------------------------------
4092 // Do we Match on this edge index or not? Do not match memory
4093 uint ClearArrayNode::match_edge(uint idx) const {
4094 return idx > 1;
4095 }
4096
4097 //------------------------------Identity---------------------------------------
4098 // Clearing a zero length array does nothing
4099 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
4100 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
4101 }
4102
4103 //------------------------------Idealize---------------------------------------
4104 // Clearing a short array is faster with stores
4105 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4106 // Already know this is a large node, do not try to ideal it
4107 if (_is_large) return nullptr;
4108
4109 const int unit = BytesPerLong;
4110 const TypeX* t = phase->type(in(2))->isa_intptr_t();
4111 if (!t) return nullptr;
4112 if (!t->is_con()) return nullptr;
4113 intptr_t raw_count = t->get_con();
4114 intptr_t size = raw_count;
4115 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
4116 // Clearing nothing uses the Identity call.
4117 // Negative clears are possible on dead ClearArrays
4118 // (see jck test stmt114.stmt11402.val).
4119 if (size <= 0 || size % unit != 0) return nullptr;
4120 intptr_t count = size / unit;
4121 // Length too long; communicate this to matchers and assemblers.
4122 // Assemblers are responsible to produce fast hardware clears for it.
4123 if (size > InitArrayShortSize) {
4124 return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true);
4125 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
4126 return nullptr;
4127 }
4128 if (!IdealizeClearArrayNode) return nullptr;
4129 Node *mem = in(1);
4130 if( phase->type(mem)==Type::TOP ) return nullptr;
4131 Node *adr = in(3);
4132 const Type* at = phase->type(adr);
4133 if( at==Type::TOP ) return nullptr;
4134 const TypePtr* atp = at->isa_ptr();
4135 // adjust atp to be the correct array element address type
4136 if (atp == nullptr) atp = TypePtr::BOTTOM;
4137 else atp = atp->add_offset(Type::OffsetBot);
4138 // Get base for derived pointer purposes
4139 if( adr->Opcode() != Op_AddP ) Unimplemented();
4140 Node *base = adr->in(1);
4141
4142 Node *val = in(4);
4143 Node *off = phase->MakeConX(BytesPerLong);
4144 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
4145 count--;
4146 while( count-- ) {
4147 mem = phase->transform(mem);
4148 adr = phase->transform(new AddPNode(base,adr,off));
4149 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
4150 }
4151 return mem;
4152 }
4153
4154 //----------------------------step_through----------------------------------
4155 // Return allocation input memory edge if it is different instance
4156 // or itself if it is the one we are looking for.
4157 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
4158 Node* n = *np;
4159 assert(n->is_ClearArray(), "sanity");
4160 intptr_t offset;
4161 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
4162 // This method is called only before Allocate nodes are expanded
4163 // during macro nodes expansion. Before that ClearArray nodes are
4164 // only generated in PhaseMacroExpand::generate_arraycopy() (before
4165 // Allocate nodes are expanded) which follows allocations.
4166 assert(alloc != nullptr, "should have allocation");
4167 if (alloc->_idx == instance_id) {
4168 // Can not bypass initialization of the instance we are looking for.
4169 return false;
4170 }
4171 // Otherwise skip it.
4172 InitializeNode* init = alloc->initialization();
4173 if (init != nullptr)
4174 *np = init->in(TypeFunc::Memory);
4175 else
4176 *np = alloc->in(TypeFunc::Memory);
4177 return true;
4178 }
4179
4180 //----------------------------clear_memory-------------------------------------
4181 // Generate code to initialize object storage to zero.
4182 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4183 Node* val,
4184 Node* raw_val,
4185 intptr_t start_offset,
4186 Node* end_offset,
4187 PhaseGVN* phase) {
4188 intptr_t offset = start_offset;
4189
4190 int unit = BytesPerLong;
4191 if ((offset % unit) != 0) {
4192 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
4193 adr = phase->transform(adr);
4194 const TypePtr* atp = TypeRawPtr::BOTTOM;
4195 if (val != nullptr) {
4196 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
4197 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
4198 } else {
4199 assert(raw_val == nullptr, "val may not be null");
4200 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4201 }
4202 mem = phase->transform(mem);
4203 offset += BytesPerInt;
4204 }
4205 assert((offset % unit) == 0, "");
4206
4207 // Initialize the remaining stuff, if any, with a ClearArray.
4208 return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, phase);
4209 }
4210
4211 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4212 Node* raw_val,
4213 Node* start_offset,
4214 Node* end_offset,
4215 PhaseGVN* phase) {
4216 if (start_offset == end_offset) {
4217 // nothing to do
4218 return mem;
4219 }
4220
4221 int unit = BytesPerLong;
4222 Node* zbase = start_offset;
4223 Node* zend = end_offset;
4224
4225 // Scale to the unit required by the CPU:
4226 if (!Matcher::init_array_count_is_in_bytes) {
4227 Node* shift = phase->intcon(exact_log2(unit));
4228 zbase = phase->transform(new URShiftXNode(zbase, shift) );
4229 zend = phase->transform(new URShiftXNode(zend, shift) );
4230 }
4231
4232 // Bulk clear double-words
4233 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
4234 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
4235 if (raw_val == nullptr) {
4236 raw_val = phase->MakeConX(0);
4237 }
4238 mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false);
4239 return phase->transform(mem);
4240 }
4241
4242 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4243 Node* val,
4244 Node* raw_val,
4245 intptr_t start_offset,
4246 intptr_t end_offset,
4247 PhaseGVN* phase) {
4248 if (start_offset == end_offset) {
4249 // nothing to do
4250 return mem;
4251 }
4252
4253 assert((end_offset % BytesPerInt) == 0, "odd end offset");
4254 intptr_t done_offset = end_offset;
4255 if ((done_offset % BytesPerLong) != 0) {
4256 done_offset -= BytesPerInt;
4257 }
4258 if (done_offset > start_offset) {
4259 mem = clear_memory(ctl, mem, dest, val, raw_val,
4260 start_offset, phase->MakeConX(done_offset), phase);
4261 }
4262 if (done_offset < end_offset) { // emit the final 32-bit store
4263 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
4264 adr = phase->transform(adr);
4265 const TypePtr* atp = TypeRawPtr::BOTTOM;
4266 if (val != nullptr) {
4267 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
4268 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
4269 } else {
4270 assert(raw_val == nullptr, "val may not be null");
4271 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4272 }
4273 mem = phase->transform(mem);
4274 done_offset += BytesPerInt;
4275 }
4276 assert(done_offset == end_offset, "");
4277 return mem;
4278 }
4279
4280 //=============================================================================
4281 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
4282 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
4283 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
4284 #ifdef ASSERT
4285 , _pair_idx(0)
4286 #endif
4287 {
4288 init_class_id(Class_MemBar);
4289 Node* top = C->top();
4290 init_req(TypeFunc::I_O,top);
4291 init_req(TypeFunc::FramePtr,top);
4292 init_req(TypeFunc::ReturnAdr,top);
4293 if (precedent != nullptr)
4294 init_req(TypeFunc::Parms, precedent);
4295 }
4296
4297 //------------------------------cmp--------------------------------------------
4298 uint MemBarNode::hash() const { return NO_HASH; }
4299 bool MemBarNode::cmp( const Node &n ) const {
4300 return (&n == this); // Always fail except on self
4301 }
4302
4303 //------------------------------make-------------------------------------------
4304 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
4305 switch (opcode) {
4306 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
4307 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
4308 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
4309 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
4310 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
4311 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn);
4312 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
4313 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
4314 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
4315 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
4316 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
4317 case Op_Initialize: return new InitializeNode(C, atp, pn);
4318 default: ShouldNotReachHere(); return nullptr;
4319 }
4320 }
4321
4322 void MemBarNode::remove(PhaseIterGVN *igvn) {
4323 if (outcnt() != 2) {
4324 assert(Opcode() == Op_Initialize, "Only seen when there are no use of init memory");
4325 assert(outcnt() == 1, "Only control then");
4326 }
4327 if (trailing_store() || trailing_load_store()) {
4328 MemBarNode* leading = leading_membar();
4329 if (leading != nullptr) {
4330 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
4331 leading->remove(igvn);
4332 }
4333 }
4334 if (proj_out_or_null(TypeFunc::Memory) != nullptr) {
4335 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
4336 }
4337 if (proj_out_or_null(TypeFunc::Control) != nullptr) {
4338 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
4339 }
4340 }
4341
4342 //------------------------------Ideal------------------------------------------
4343 // Return a node which is more "ideal" than the current node. Strip out
4344 // control copies
4345 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4346 if (remove_dead_region(phase, can_reshape)) return this;
4347 // Don't bother trying to transform a dead node
4348 if (in(0) && in(0)->is_top()) {
4349 return nullptr;
4350 }
4351
4352 bool progress = false;
4353 // Eliminate volatile MemBars for scalar replaced objects.
4354 if (can_reshape && req() == (Precedent+1)) {
4355 bool eliminate = false;
4356 int opc = Opcode();
4357 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
4358 // Volatile field loads and stores.
4359 Node* my_mem = in(MemBarNode::Precedent);
4360 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
4361 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
4362 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
4363 // replace this Precedent (decodeN) with the Load instead.
4364 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
4365 Node* load_node = my_mem->in(1);
4366 set_req(MemBarNode::Precedent, load_node);
4367 phase->is_IterGVN()->_worklist.push(my_mem);
4368 my_mem = load_node;
4369 } else {
4370 assert(my_mem->unique_out() == this, "sanity");
4371 assert(!trailing_load_store(), "load store node can't be eliminated");
4372 del_req(Precedent);
4373 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
4374 my_mem = nullptr;
4375 }
4376 progress = true;
4377 }
4378 if (my_mem != nullptr && my_mem->is_Mem()) {
4379 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
4380 // Check for scalar replaced object reference.
4381 if( t_oop != nullptr && t_oop->is_known_instance_field() &&
4382 t_oop->offset() != Type::OffsetBot &&
4383 t_oop->offset() != Type::OffsetTop) {
4384 eliminate = true;
4385 }
4386 }
4387 } else if (opc == Op_MemBarRelease || (UseStoreStoreForCtor && opc == Op_MemBarStoreStore)) {
4388 // Final field stores.
4389 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent));
4390 if ((alloc != nullptr) && alloc->is_Allocate() &&
4391 alloc->as_Allocate()->does_not_escape_thread()) {
4392 // The allocated object does not escape.
4393 eliminate = true;
4394 }
4395 }
4396 if (eliminate) {
4397 // Replace MemBar projections by its inputs.
4398 PhaseIterGVN* igvn = phase->is_IterGVN();
4399 remove(igvn);
4400 // Must return either the original node (now dead) or a new node
4401 // (Do not return a top here, since that would break the uniqueness of top.)
4402 return new ConINode(TypeInt::ZERO);
4403 }
4404 }
4405 return progress ? this : nullptr;
4406 }
4407
4408 //------------------------------Value------------------------------------------
4409 const Type* MemBarNode::Value(PhaseGVN* phase) const {
4410 if( !in(0) ) return Type::TOP;
4411 if( phase->type(in(0)) == Type::TOP )
4412 return Type::TOP;
4413 return TypeTuple::MEMBAR;
4414 }
4415
4416 //------------------------------match------------------------------------------
4417 // Construct projections for memory.
4418 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) {
4419 switch (proj->_con) {
4420 case TypeFunc::Control:
4421 case TypeFunc::Memory:
4422 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
4423 }
4424 ShouldNotReachHere();
4425 return nullptr;
4426 }
4427
4428 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4429 trailing->_kind = TrailingStore;
4430 leading->_kind = LeadingStore;
4431 #ifdef ASSERT
4432 trailing->_pair_idx = leading->_idx;
4433 leading->_pair_idx = leading->_idx;
4434 #endif
4435 }
4436
4437 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4438 trailing->_kind = TrailingLoadStore;
4439 leading->_kind = LeadingLoadStore;
4440 #ifdef ASSERT
4441 trailing->_pair_idx = leading->_idx;
4442 leading->_pair_idx = leading->_idx;
4443 #endif
4444 }
4445
4446 MemBarNode* MemBarNode::trailing_membar() const {
4447 ResourceMark rm;
4448 Node* trailing = (Node*)this;
4449 VectorSet seen;
4450 Node_Stack multis(0);
4451 do {
4452 Node* c = trailing;
4453 uint i = 0;
4454 do {
4455 trailing = nullptr;
4456 for (; i < c->outcnt(); i++) {
4457 Node* next = c->raw_out(i);
4458 if (next != c && next->is_CFG()) {
4459 if (c->is_MultiBranch()) {
4460 if (multis.node() == c) {
4461 multis.set_index(i+1);
4462 } else {
4463 multis.push(c, i+1);
4464 }
4465 }
4466 trailing = next;
4467 break;
4468 }
4469 }
4470 if (trailing != nullptr && !seen.test_set(trailing->_idx)) {
4471 break;
4472 }
4473 while (multis.size() > 0) {
4474 c = multis.node();
4475 i = multis.index();
4476 if (i < c->req()) {
4477 break;
4478 }
4479 multis.pop();
4480 }
4481 } while (multis.size() > 0);
4482 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
4483
4484 MemBarNode* mb = trailing->as_MemBar();
4485 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
4486 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
4487 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
4488 return mb;
4489 }
4490
4491 MemBarNode* MemBarNode::leading_membar() const {
4492 ResourceMark rm;
4493 VectorSet seen;
4494 Node_Stack regions(0);
4495 Node* leading = in(0);
4496 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
4497 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) {
4498 leading = nullptr;
4499 while (regions.size() > 0 && leading == nullptr) {
4500 Node* r = regions.node();
4501 uint i = regions.index();
4502 if (i < r->req()) {
4503 leading = r->in(i);
4504 regions.set_index(i+1);
4505 } else {
4506 regions.pop();
4507 }
4508 }
4509 if (leading == nullptr) {
4510 assert(regions.size() == 0, "all paths should have been tried");
4511 return nullptr;
4512 }
4513 }
4514 if (leading->is_Region()) {
4515 regions.push(leading, 2);
4516 leading = leading->in(1);
4517 } else {
4518 leading = leading->in(0);
4519 }
4520 }
4521 #ifdef ASSERT
4522 Unique_Node_List wq;
4523 wq.push((Node*)this);
4524 uint found = 0;
4525 for (uint i = 0; i < wq.size(); i++) {
4526 Node* n = wq.at(i);
4527 if (n->is_Region()) {
4528 for (uint j = 1; j < n->req(); j++) {
4529 Node* in = n->in(j);
4530 if (in != nullptr && !in->is_top()) {
4531 wq.push(in);
4532 }
4533 }
4534 } else {
4535 if (n->is_MemBar() && n->as_MemBar()->leading()) {
4536 assert(n == leading, "consistency check failed");
4537 found++;
4538 } else {
4539 Node* in = n->in(0);
4540 if (in != nullptr && !in->is_top()) {
4541 wq.push(in);
4542 }
4543 }
4544 }
4545 }
4546 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed");
4547 #endif
4548 if (leading == nullptr) {
4549 return nullptr;
4550 }
4551 MemBarNode* mb = leading->as_MemBar();
4552 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
4553 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
4554 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
4555 return mb;
4556 }
4557
4558
4559 //===========================InitializeNode====================================
4560 // SUMMARY:
4561 // This node acts as a memory barrier on raw memory, after some raw stores.
4562 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
4563 // The Initialize can 'capture' suitably constrained stores as raw inits.
4564 // It can coalesce related raw stores into larger units (called 'tiles').
4565 // It can avoid zeroing new storage for memory units which have raw inits.
4566 // At macro-expansion, it is marked 'complete', and does not optimize further.
4567 //
4568 // EXAMPLE:
4569 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
4570 // ctl = incoming control; mem* = incoming memory
4571 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
4572 // First allocate uninitialized memory and fill in the header:
4573 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
4574 // ctl := alloc.Control; mem* := alloc.Memory*
4575 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
4576 // Then initialize to zero the non-header parts of the raw memory block:
4577 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
4578 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
4579 // After the initialize node executes, the object is ready for service:
4580 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
4581 // Suppose its body is immediately initialized as {1,2}:
4582 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4583 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4584 // mem.SLICE(#short[*]) := store2
4585 //
4586 // DETAILS:
4587 // An InitializeNode collects and isolates object initialization after
4588 // an AllocateNode and before the next possible safepoint. As a
4589 // memory barrier (MemBarNode), it keeps critical stores from drifting
4590 // down past any safepoint or any publication of the allocation.
4591 // Before this barrier, a newly-allocated object may have uninitialized bits.
4592 // After this barrier, it may be treated as a real oop, and GC is allowed.
4593 //
4594 // The semantics of the InitializeNode include an implicit zeroing of
4595 // the new object from object header to the end of the object.
4596 // (The object header and end are determined by the AllocateNode.)
4597 //
4598 // Certain stores may be added as direct inputs to the InitializeNode.
4599 // These stores must update raw memory, and they must be to addresses
4600 // derived from the raw address produced by AllocateNode, and with
4601 // a constant offset. They must be ordered by increasing offset.
4602 // The first one is at in(RawStores), the last at in(req()-1).
4603 // Unlike most memory operations, they are not linked in a chain,
4604 // but are displayed in parallel as users of the rawmem output of
4605 // the allocation.
4606 //
4607 // (See comments in InitializeNode::capture_store, which continue
4608 // the example given above.)
4609 //
4610 // When the associated Allocate is macro-expanded, the InitializeNode
4611 // may be rewritten to optimize collected stores. A ClearArrayNode
4612 // may also be created at that point to represent any required zeroing.
4613 // The InitializeNode is then marked 'complete', prohibiting further
4614 // capturing of nearby memory operations.
4615 //
4616 // During macro-expansion, all captured initializations which store
4617 // constant values of 32 bits or smaller are coalesced (if advantageous)
4618 // into larger 'tiles' 32 or 64 bits. This allows an object to be
4619 // initialized in fewer memory operations. Memory words which are
4620 // covered by neither tiles nor non-constant stores are pre-zeroed
4621 // by explicit stores of zero. (The code shape happens to do all
4622 // zeroing first, then all other stores, with both sequences occurring
4623 // in order of ascending offsets.)
4624 //
4625 // Alternatively, code may be inserted between an AllocateNode and its
4626 // InitializeNode, to perform arbitrary initialization of the new object.
4627 // E.g., the object copying intrinsics insert complex data transfers here.
4628 // The initialization must then be marked as 'complete' disable the
4629 // built-in zeroing semantics and the collection of initializing stores.
4630 //
4631 // While an InitializeNode is incomplete, reads from the memory state
4632 // produced by it are optimizable if they match the control edge and
4633 // new oop address associated with the allocation/initialization.
4634 // They return a stored value (if the offset matches) or else zero.
4635 // A write to the memory state, if it matches control and address,
4636 // and if it is to a constant offset, may be 'captured' by the
4637 // InitializeNode. It is cloned as a raw memory operation and rewired
4638 // inside the initialization, to the raw oop produced by the allocation.
4639 // Operations on addresses which are provably distinct (e.g., to
4640 // other AllocateNodes) are allowed to bypass the initialization.
4641 //
4642 // The effect of all this is to consolidate object initialization
4643 // (both arrays and non-arrays, both piecewise and bulk) into a
4644 // single location, where it can be optimized as a unit.
4645 //
4646 // Only stores with an offset less than TrackedInitializationLimit words
4647 // will be considered for capture by an InitializeNode. This puts a
4648 // reasonable limit on the complexity of optimized initializations.
4649
4650 //---------------------------InitializeNode------------------------------------
4651 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
4652 : MemBarNode(C, adr_type, rawoop),
4653 _is_complete(Incomplete), _does_not_escape(false)
4654 {
4655 init_class_id(Class_Initialize);
4656
4657 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
4658 assert(in(RawAddress) == rawoop, "proper init");
4659 // Note: allocation() can be null, for secondary initialization barriers
4660 }
4661
4662 // Since this node is not matched, it will be processed by the
4663 // register allocator. Declare that there are no constraints
4664 // on the allocation of the RawAddress edge.
4665 const RegMask &InitializeNode::in_RegMask(uint idx) const {
4666 // This edge should be set to top, by the set_complete. But be conservative.
4667 if (idx == InitializeNode::RawAddress)
4668 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
4669 return RegMask::Empty;
4670 }
4671
4672 Node* InitializeNode::memory(uint alias_idx) {
4673 Node* mem = in(Memory);
4674 if (mem->is_MergeMem()) {
4675 return mem->as_MergeMem()->memory_at(alias_idx);
4676 } else {
4677 // incoming raw memory is not split
4678 return mem;
4679 }
4680 }
4681
4682 bool InitializeNode::is_non_zero() {
4683 if (is_complete()) return false;
4684 remove_extra_zeroes();
4685 return (req() > RawStores);
4686 }
4687
4688 void InitializeNode::set_complete(PhaseGVN* phase) {
4689 assert(!is_complete(), "caller responsibility");
4690 _is_complete = Complete;
4691
4692 // After this node is complete, it contains a bunch of
4693 // raw-memory initializations. There is no need for
4694 // it to have anything to do with non-raw memory effects.
4695 // Therefore, tell all non-raw users to re-optimize themselves,
4696 // after skipping the memory effects of this initialization.
4697 PhaseIterGVN* igvn = phase->is_IterGVN();
4698 if (igvn) igvn->add_users_to_worklist(this);
4699 }
4700
4701 // convenience function
4702 // return false if the init contains any stores already
4703 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
4704 InitializeNode* init = initialization();
4705 if (init == nullptr || init->is_complete()) {
4706 return false;
4707 }
4708 init->remove_extra_zeroes();
4709 // for now, if this allocation has already collected any inits, bail:
4710 if (init->is_non_zero()) return false;
4711 init->set_complete(phase);
4712 return true;
4713 }
4714
4715 void InitializeNode::remove_extra_zeroes() {
4716 if (req() == RawStores) return;
4717 Node* zmem = zero_memory();
4718 uint fill = RawStores;
4719 for (uint i = fill; i < req(); i++) {
4720 Node* n = in(i);
4721 if (n->is_top() || n == zmem) continue; // skip
4722 if (fill < i) set_req(fill, n); // compact
4723 ++fill;
4724 }
4725 // delete any empty spaces created:
4726 while (fill < req()) {
4727 del_req(fill);
4728 }
4729 }
4730
4731 // Helper for remembering which stores go with which offsets.
4732 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) {
4733 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
4734 intptr_t offset = -1;
4735 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
4736 phase, offset);
4737 if (base == nullptr) return -1; // something is dead,
4738 if (offset < 0) return -1; // dead, dead
4739 return offset;
4740 }
4741
4742 // Helper for proving that an initialization expression is
4743 // "simple enough" to be folded into an object initialization.
4744 // Attempts to prove that a store's initial value 'n' can be captured
4745 // within the initialization without creating a vicious cycle, such as:
4746 // { Foo p = new Foo(); p.next = p; }
4747 // True for constants and parameters and small combinations thereof.
4748 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
4749 ResourceMark rm;
4750 Unique_Node_List worklist;
4751 worklist.push(value);
4752
4753 uint complexity_limit = 20;
4754 for (uint j = 0; j < worklist.size(); j++) {
4755 if (j >= complexity_limit) {
4756 return false; // Bail out if processed too many nodes
4757 }
4758
4759 Node* n = worklist.at(j);
4760 if (n == nullptr) continue; // (can this really happen?)
4761 if (n->is_Proj()) n = n->in(0);
4762 if (n == this) return false; // found a cycle
4763 if (n->is_Con()) continue;
4764 if (n->is_Start()) continue; // params, etc., are OK
4765 if (n->is_Root()) continue; // even better
4766
4767 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
4768 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
4769 continue;
4770 }
4771
4772 Node* ctl = n->in(0);
4773 if (ctl != nullptr && !ctl->is_top()) {
4774 if (ctl->is_Proj()) ctl = ctl->in(0);
4775 if (ctl == this) return false;
4776
4777 // If we already know that the enclosing memory op is pinned right after
4778 // the init, then any control flow that the store has picked up
4779 // must have preceded the init, or else be equal to the init.
4780 // Even after loop optimizations (which might change control edges)
4781 // a store is never pinned *before* the availability of its inputs.
4782 if (!MemNode::all_controls_dominate(n, this)) {
4783 return false; // failed to prove a good control
4784 }
4785 }
4786
4787 // Check data edges for possible dependencies on 'this'.
4788 for (uint i = 1; i < n->req(); i++) {
4789 Node* m = n->in(i);
4790 if (m == nullptr || m == n || m->is_top()) continue;
4791
4792 // Only process data inputs once
4793 worklist.push(m);
4794 }
4795 }
4796
4797 return true;
4798 }
4799
4800 // Here are all the checks a Store must pass before it can be moved into
4801 // an initialization. Returns zero if a check fails.
4802 // On success, returns the (constant) offset to which the store applies,
4803 // within the initialized memory.
4804 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
4805 const int FAIL = 0;
4806 if (st->req() != MemNode::ValueIn + 1)
4807 return FAIL; // an inscrutable StoreNode (card mark?)
4808 Node* ctl = st->in(MemNode::Control);
4809 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this))
4810 return FAIL; // must be unconditional after the initialization
4811 Node* mem = st->in(MemNode::Memory);
4812 if (!(mem->is_Proj() && mem->in(0) == this))
4813 return FAIL; // must not be preceded by other stores
4814 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4815 if ((st->Opcode() == Op_StoreP || st->Opcode() == Op_StoreN) &&
4816 !bs->can_initialize_object(st)) {
4817 return FAIL;
4818 }
4819 Node* adr = st->in(MemNode::Address);
4820 intptr_t offset;
4821 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
4822 if (alloc == nullptr)
4823 return FAIL; // inscrutable address
4824 if (alloc != allocation())
4825 return FAIL; // wrong allocation! (store needs to float up)
4826 int size_in_bytes = st->memory_size();
4827 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
4828 return FAIL; // mismatched access
4829 }
4830 Node* val = st->in(MemNode::ValueIn);
4831
4832 if (!detect_init_independence(val, phase))
4833 return FAIL; // stored value must be 'simple enough'
4834
4835 // The Store can be captured only if nothing after the allocation
4836 // and before the Store is using the memory location that the store
4837 // overwrites.
4838 bool failed = false;
4839 // If is_complete_with_arraycopy() is true the shape of the graph is
4840 // well defined and is safe so no need for extra checks.
4841 if (!is_complete_with_arraycopy()) {
4842 // We are going to look at each use of the memory state following
4843 // the allocation to make sure nothing reads the memory that the
4844 // Store writes.
4845 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
4846 int alias_idx = phase->C->get_alias_index(t_adr);
4847 ResourceMark rm;
4848 Unique_Node_List mems;
4849 mems.push(mem);
4850 Node* unique_merge = nullptr;
4851 for (uint next = 0; next < mems.size(); ++next) {
4852 Node *m = mems.at(next);
4853 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
4854 Node *n = m->fast_out(j);
4855 if (n->outcnt() == 0) {
4856 continue;
4857 }
4858 if (n == st) {
4859 continue;
4860 } else if (n->in(0) != nullptr && n->in(0) != ctl) {
4861 // If the control of this use is different from the control
4862 // of the Store which is right after the InitializeNode then
4863 // this node cannot be between the InitializeNode and the
4864 // Store.
4865 continue;
4866 } else if (n->is_MergeMem()) {
4867 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
4868 // We can hit a MergeMemNode (that will likely go away
4869 // later) that is a direct use of the memory state
4870 // following the InitializeNode on the same slice as the
4871 // store node that we'd like to capture. We need to check
4872 // the uses of the MergeMemNode.
4873 mems.push(n);
4874 }
4875 } else if (n->is_Mem()) {
4876 Node* other_adr = n->in(MemNode::Address);
4877 if (other_adr == adr) {
4878 failed = true;
4879 break;
4880 } else {
4881 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
4882 if (other_t_adr != nullptr) {
4883 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
4884 if (other_alias_idx == alias_idx) {
4885 // A load from the same memory slice as the store right
4886 // after the InitializeNode. We check the control of the
4887 // object/array that is loaded from. If it's the same as
4888 // the store control then we cannot capture the store.
4889 assert(!n->is_Store(), "2 stores to same slice on same control?");
4890 Node* base = other_adr;
4891 if (base->is_Phi()) {
4892 // In rare case, base may be a PhiNode and it may read
4893 // the same memory slice between InitializeNode and store.
4894 failed = true;
4895 break;
4896 }
4897 assert(base->is_AddP(), "should be addp but is %s", base->Name());
4898 base = base->in(AddPNode::Base);
4899 if (base != nullptr) {
4900 base = base->uncast();
4901 if (base->is_Proj() && base->in(0) == alloc) {
4902 failed = true;
4903 break;
4904 }
4905 }
4906 }
4907 }
4908 }
4909 } else {
4910 failed = true;
4911 break;
4912 }
4913 }
4914 }
4915 }
4916 if (failed) {
4917 if (!can_reshape) {
4918 // We decided we couldn't capture the store during parsing. We
4919 // should try again during the next IGVN once the graph is
4920 // cleaner.
4921 phase->C->record_for_igvn(st);
4922 }
4923 return FAIL;
4924 }
4925
4926 return offset; // success
4927 }
4928
4929 // Find the captured store in(i) which corresponds to the range
4930 // [start..start+size) in the initialized object.
4931 // If there is one, return its index i. If there isn't, return the
4932 // negative of the index where it should be inserted.
4933 // Return 0 if the queried range overlaps an initialization boundary
4934 // or if dead code is encountered.
4935 // If size_in_bytes is zero, do not bother with overlap checks.
4936 int InitializeNode::captured_store_insertion_point(intptr_t start,
4937 int size_in_bytes,
4938 PhaseValues* phase) {
4939 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
4940
4941 if (is_complete())
4942 return FAIL; // arraycopy got here first; punt
4943
4944 assert(allocation() != nullptr, "must be present");
4945
4946 // no negatives, no header fields:
4947 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
4948
4949 // after a certain size, we bail out on tracking all the stores:
4950 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4951 if (start >= ti_limit) return FAIL;
4952
4953 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
4954 if (i >= limit) return -(int)i; // not found; here is where to put it
4955
4956 Node* st = in(i);
4957 intptr_t st_off = get_store_offset(st, phase);
4958 if (st_off < 0) {
4959 if (st != zero_memory()) {
4960 return FAIL; // bail out if there is dead garbage
4961 }
4962 } else if (st_off > start) {
4963 // ...we are done, since stores are ordered
4964 if (st_off < start + size_in_bytes) {
4965 return FAIL; // the next store overlaps
4966 }
4967 return -(int)i; // not found; here is where to put it
4968 } else if (st_off < start) {
4969 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
4970 if (size_in_bytes != 0 &&
4971 start < st_off + MAX_STORE &&
4972 start < st_off + st->as_Store()->memory_size()) {
4973 return FAIL; // the previous store overlaps
4974 }
4975 } else {
4976 if (size_in_bytes != 0 &&
4977 st->as_Store()->memory_size() != size_in_bytes) {
4978 return FAIL; // mismatched store size
4979 }
4980 return i;
4981 }
4982
4983 ++i;
4984 }
4985 }
4986
4987 // Look for a captured store which initializes at the offset 'start'
4988 // with the given size. If there is no such store, and no other
4989 // initialization interferes, then return zero_memory (the memory
4990 // projection of the AllocateNode).
4991 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
4992 PhaseValues* phase) {
4993 assert(stores_are_sane(phase), "");
4994 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4995 if (i == 0) {
4996 return nullptr; // something is dead
4997 } else if (i < 0) {
4998 return zero_memory(); // just primordial zero bits here
4999 } else {
5000 Node* st = in(i); // here is the store at this position
5001 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
5002 return st;
5003 }
5004 }
5005
5006 // Create, as a raw pointer, an address within my new object at 'offset'.
5007 Node* InitializeNode::make_raw_address(intptr_t offset,
5008 PhaseGVN* phase) {
5009 Node* addr = in(RawAddress);
5010 if (offset != 0) {
5011 Compile* C = phase->C;
5012 addr = phase->transform( new AddPNode(C->top(), addr,
5013 phase->MakeConX(offset)) );
5014 }
5015 return addr;
5016 }
5017
5018 // Clone the given store, converting it into a raw store
5019 // initializing a field or element of my new object.
5020 // Caller is responsible for retiring the original store,
5021 // with subsume_node or the like.
5022 //
5023 // From the example above InitializeNode::InitializeNode,
5024 // here are the old stores to be captured:
5025 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
5026 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
5027 //
5028 // Here is the changed code; note the extra edges on init:
5029 // alloc = (Allocate ...)
5030 // rawoop = alloc.RawAddress
5031 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
5032 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
5033 // init = (Initialize alloc.Control alloc.Memory rawoop
5034 // rawstore1 rawstore2)
5035 //
5036 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
5037 PhaseGVN* phase, bool can_reshape) {
5038 assert(stores_are_sane(phase), "");
5039
5040 if (start < 0) return nullptr;
5041 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
5042
5043 Compile* C = phase->C;
5044 int size_in_bytes = st->memory_size();
5045 int i = captured_store_insertion_point(start, size_in_bytes, phase);
5046 if (i == 0) return nullptr; // bail out
5047 Node* prev_mem = nullptr; // raw memory for the captured store
5048 if (i > 0) {
5049 prev_mem = in(i); // there is a pre-existing store under this one
5050 set_req(i, C->top()); // temporarily disconnect it
5051 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
5052 } else {
5053 i = -i; // no pre-existing store
5054 prev_mem = zero_memory(); // a slice of the newly allocated object
5055 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
5056 set_req(--i, C->top()); // reuse this edge; it has been folded away
5057 else
5058 ins_req(i, C->top()); // build a new edge
5059 }
5060 Node* new_st = st->clone();
5061 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
5062 new_st->set_req(MemNode::Control, in(Control));
5063 new_st->set_req(MemNode::Memory, prev_mem);
5064 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
5065 bs->eliminate_gc_barrier_data(new_st);
5066 new_st = phase->transform(new_st);
5067
5068 // At this point, new_st might have swallowed a pre-existing store
5069 // at the same offset, or perhaps new_st might have disappeared,
5070 // if it redundantly stored the same value (or zero to fresh memory).
5071
5072 // In any case, wire it in:
5073 PhaseIterGVN* igvn = phase->is_IterGVN();
5074 if (igvn) {
5075 igvn->rehash_node_delayed(this);
5076 }
5077 set_req(i, new_st);
5078
5079 // The caller may now kill the old guy.
5080 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
5081 assert(check_st == new_st || check_st == nullptr, "must be findable");
5082 assert(!is_complete(), "");
5083 return new_st;
5084 }
5085
5086 static bool store_constant(jlong* tiles, int num_tiles,
5087 intptr_t st_off, int st_size,
5088 jlong con) {
5089 if ((st_off & (st_size-1)) != 0)
5090 return false; // strange store offset (assume size==2**N)
5091 address addr = (address)tiles + st_off;
5092 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
5093 switch (st_size) {
5094 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
5095 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
5096 case sizeof(jint): *(jint*) addr = (jint) con; break;
5097 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
5098 default: return false; // strange store size (detect size!=2**N here)
5099 }
5100 return true; // return success to caller
5101 }
5102
5103 // Coalesce subword constants into int constants and possibly
5104 // into long constants. The goal, if the CPU permits,
5105 // is to initialize the object with a small number of 64-bit tiles.
5106 // Also, convert floating-point constants to bit patterns.
5107 // Non-constants are not relevant to this pass.
5108 //
5109 // In terms of the running example on InitializeNode::InitializeNode
5110 // and InitializeNode::capture_store, here is the transformation
5111 // of rawstore1 and rawstore2 into rawstore12:
5112 // alloc = (Allocate ...)
5113 // rawoop = alloc.RawAddress
5114 // tile12 = 0x00010002
5115 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
5116 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
5117 //
5118 void
5119 InitializeNode::coalesce_subword_stores(intptr_t header_size,
5120 Node* size_in_bytes,
5121 PhaseGVN* phase) {
5122 Compile* C = phase->C;
5123
5124 assert(stores_are_sane(phase), "");
5125 // Note: After this pass, they are not completely sane,
5126 // since there may be some overlaps.
5127
5128 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
5129
5130 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
5131 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
5132 size_limit = MIN2(size_limit, ti_limit);
5133 size_limit = align_up(size_limit, BytesPerLong);
5134 int num_tiles = size_limit / BytesPerLong;
5135
5136 // allocate space for the tile map:
5137 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
5138 jlong tiles_buf[small_len];
5139 Node* nodes_buf[small_len];
5140 jlong inits_buf[small_len];
5141 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
5142 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5143 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
5144 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
5145 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
5146 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5147 // tiles: exact bitwise model of all primitive constants
5148 // nodes: last constant-storing node subsumed into the tiles model
5149 // inits: which bytes (in each tile) are touched by any initializations
5150
5151 //// Pass A: Fill in the tile model with any relevant stores.
5152
5153 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
5154 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
5155 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
5156 Node* zmem = zero_memory(); // initially zero memory state
5157 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5158 Node* st = in(i);
5159 intptr_t st_off = get_store_offset(st, phase);
5160
5161 // Figure out the store's offset and constant value:
5162 if (st_off < header_size) continue; //skip (ignore header)
5163 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
5164 int st_size = st->as_Store()->memory_size();
5165 if (st_off + st_size > size_limit) break;
5166
5167 // Record which bytes are touched, whether by constant or not.
5168 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
5169 continue; // skip (strange store size)
5170
5171 const Type* val = phase->type(st->in(MemNode::ValueIn));
5172 if (!val->singleton()) continue; //skip (non-con store)
5173 BasicType type = val->basic_type();
5174
5175 jlong con = 0;
5176 switch (type) {
5177 case T_INT: con = val->is_int()->get_con(); break;
5178 case T_LONG: con = val->is_long()->get_con(); break;
5179 case T_FLOAT: con = jint_cast(val->getf()); break;
5180 case T_DOUBLE: con = jlong_cast(val->getd()); break;
5181 default: continue; //skip (odd store type)
5182 }
5183
5184 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
5185 st->Opcode() == Op_StoreL) {
5186 continue; // This StoreL is already optimal.
5187 }
5188
5189 // Store down the constant.
5190 store_constant(tiles, num_tiles, st_off, st_size, con);
5191
5192 intptr_t j = st_off >> LogBytesPerLong;
5193
5194 if (type == T_INT && st_size == BytesPerInt
5195 && (st_off & BytesPerInt) == BytesPerInt) {
5196 jlong lcon = tiles[j];
5197 if (!Matcher::isSimpleConstant64(lcon) &&
5198 st->Opcode() == Op_StoreI) {
5199 // This StoreI is already optimal by itself.
5200 jint* intcon = (jint*) &tiles[j];
5201 intcon[1] = 0; // undo the store_constant()
5202
5203 // If the previous store is also optimal by itself, back up and
5204 // undo the action of the previous loop iteration... if we can.
5205 // But if we can't, just let the previous half take care of itself.
5206 st = nodes[j];
5207 st_off -= BytesPerInt;
5208 con = intcon[0];
5209 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) {
5210 assert(st_off >= header_size, "still ignoring header");
5211 assert(get_store_offset(st, phase) == st_off, "must be");
5212 assert(in(i-1) == zmem, "must be");
5213 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
5214 assert(con == tcon->is_int()->get_con(), "must be");
5215 // Undo the effects of the previous loop trip, which swallowed st:
5216 intcon[0] = 0; // undo store_constant()
5217 set_req(i-1, st); // undo set_req(i, zmem)
5218 nodes[j] = nullptr; // undo nodes[j] = st
5219 --old_subword; // undo ++old_subword
5220 }
5221 continue; // This StoreI is already optimal.
5222 }
5223 }
5224
5225 // This store is not needed.
5226 set_req(i, zmem);
5227 nodes[j] = st; // record for the moment
5228 if (st_size < BytesPerLong) // something has changed
5229 ++old_subword; // includes int/float, but who's counting...
5230 else ++old_long;
5231 }
5232
5233 if ((old_subword + old_long) == 0)
5234 return; // nothing more to do
5235
5236 //// Pass B: Convert any non-zero tiles into optimal constant stores.
5237 // Be sure to insert them before overlapping non-constant stores.
5238 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
5239 for (int j = 0; j < num_tiles; j++) {
5240 jlong con = tiles[j];
5241 jlong init = inits[j];
5242 if (con == 0) continue;
5243 jint con0, con1; // split the constant, address-wise
5244 jint init0, init1; // split the init map, address-wise
5245 { union { jlong con; jint intcon[2]; } u;
5246 u.con = con;
5247 con0 = u.intcon[0];
5248 con1 = u.intcon[1];
5249 u.con = init;
5250 init0 = u.intcon[0];
5251 init1 = u.intcon[1];
5252 }
5253
5254 Node* old = nodes[j];
5255 assert(old != nullptr, "need the prior store");
5256 intptr_t offset = (j * BytesPerLong);
5257
5258 bool split = !Matcher::isSimpleConstant64(con);
5259
5260 if (offset < header_size) {
5261 assert(offset + BytesPerInt >= header_size, "second int counts");
5262 assert(*(jint*)&tiles[j] == 0, "junk in header");
5263 split = true; // only the second word counts
5264 // Example: int a[] = { 42 ... }
5265 } else if (con0 == 0 && init0 == -1) {
5266 split = true; // first word is covered by full inits
5267 // Example: int a[] = { ... foo(), 42 ... }
5268 } else if (con1 == 0 && init1 == -1) {
5269 split = true; // second word is covered by full inits
5270 // Example: int a[] = { ... 42, foo() ... }
5271 }
5272
5273 // Here's a case where init0 is neither 0 nor -1:
5274 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
5275 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
5276 // In this case the tile is not split; it is (jlong)42.
5277 // The big tile is stored down, and then the foo() value is inserted.
5278 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
5279
5280 Node* ctl = old->in(MemNode::Control);
5281 Node* adr = make_raw_address(offset, phase);
5282 const TypePtr* atp = TypeRawPtr::BOTTOM;
5283
5284 // One or two coalesced stores to plop down.
5285 Node* st[2];
5286 intptr_t off[2];
5287 int nst = 0;
5288 if (!split) {
5289 ++new_long;
5290 off[nst] = offset;
5291 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5292 phase->longcon(con), T_LONG, MemNode::unordered);
5293 } else {
5294 // Omit either if it is a zero.
5295 if (con0 != 0) {
5296 ++new_int;
5297 off[nst] = offset;
5298 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5299 phase->intcon(con0), T_INT, MemNode::unordered);
5300 }
5301 if (con1 != 0) {
5302 ++new_int;
5303 offset += BytesPerInt;
5304 adr = make_raw_address(offset, phase);
5305 off[nst] = offset;
5306 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5307 phase->intcon(con1), T_INT, MemNode::unordered);
5308 }
5309 }
5310
5311 // Insert second store first, then the first before the second.
5312 // Insert each one just before any overlapping non-constant stores.
5313 while (nst > 0) {
5314 Node* st1 = st[--nst];
5315 C->copy_node_notes_to(st1, old);
5316 st1 = phase->transform(st1);
5317 offset = off[nst];
5318 assert(offset >= header_size, "do not smash header");
5319 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
5320 guarantee(ins_idx != 0, "must re-insert constant store");
5321 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
5322 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
5323 set_req(--ins_idx, st1);
5324 else
5325 ins_req(ins_idx, st1);
5326 }
5327 }
5328
5329 if (PrintCompilation && WizardMode)
5330 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
5331 old_subword, old_long, new_int, new_long);
5332 if (C->log() != nullptr)
5333 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
5334 old_subword, old_long, new_int, new_long);
5335
5336 // Clean up any remaining occurrences of zmem:
5337 remove_extra_zeroes();
5338 }
5339
5340 // Explore forward from in(start) to find the first fully initialized
5341 // word, and return its offset. Skip groups of subword stores which
5342 // together initialize full words. If in(start) is itself part of a
5343 // fully initialized word, return the offset of in(start). If there
5344 // are no following full-word stores, or if something is fishy, return
5345 // a negative value.
5346 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
5347 int int_map = 0;
5348 intptr_t int_map_off = 0;
5349 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
5350
5351 for (uint i = start, limit = req(); i < limit; i++) {
5352 Node* st = in(i);
5353
5354 intptr_t st_off = get_store_offset(st, phase);
5355 if (st_off < 0) break; // return conservative answer
5356
5357 int st_size = st->as_Store()->memory_size();
5358 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
5359 return st_off; // we found a complete word init
5360 }
5361
5362 // update the map:
5363
5364 intptr_t this_int_off = align_down(st_off, BytesPerInt);
5365 if (this_int_off != int_map_off) {
5366 // reset the map:
5367 int_map = 0;
5368 int_map_off = this_int_off;
5369 }
5370
5371 int subword_off = st_off - this_int_off;
5372 int_map |= right_n_bits(st_size) << subword_off;
5373 if ((int_map & FULL_MAP) == FULL_MAP) {
5374 return this_int_off; // we found a complete word init
5375 }
5376
5377 // Did this store hit or cross the word boundary?
5378 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
5379 if (next_int_off == this_int_off + BytesPerInt) {
5380 // We passed the current int, without fully initializing it.
5381 int_map_off = next_int_off;
5382 int_map >>= BytesPerInt;
5383 } else if (next_int_off > this_int_off + BytesPerInt) {
5384 // We passed the current and next int.
5385 return this_int_off + BytesPerInt;
5386 }
5387 }
5388
5389 return -1;
5390 }
5391
5392
5393 // Called when the associated AllocateNode is expanded into CFG.
5394 // At this point, we may perform additional optimizations.
5395 // Linearize the stores by ascending offset, to make memory
5396 // activity as coherent as possible.
5397 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
5398 intptr_t header_size,
5399 Node* size_in_bytes,
5400 PhaseIterGVN* phase) {
5401 assert(!is_complete(), "not already complete");
5402 assert(stores_are_sane(phase), "");
5403 assert(allocation() != nullptr, "must be present");
5404
5405 remove_extra_zeroes();
5406
5407 if (ReduceFieldZeroing || ReduceBulkZeroing)
5408 // reduce instruction count for common initialization patterns
5409 coalesce_subword_stores(header_size, size_in_bytes, phase);
5410
5411 Node* zmem = zero_memory(); // initially zero memory state
5412 Node* inits = zmem; // accumulating a linearized chain of inits
5413 #ifdef ASSERT
5414 intptr_t first_offset = allocation()->minimum_header_size();
5415 intptr_t last_init_off = first_offset; // previous init offset
5416 intptr_t last_init_end = first_offset; // previous init offset+size
5417 intptr_t last_tile_end = first_offset; // previous tile offset+size
5418 #endif
5419 intptr_t zeroes_done = header_size;
5420
5421 bool do_zeroing = true; // we might give up if inits are very sparse
5422 int big_init_gaps = 0; // how many large gaps have we seen?
5423
5424 if (UseTLAB && ZeroTLAB) do_zeroing = false;
5425 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
5426
5427 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5428 Node* st = in(i);
5429 intptr_t st_off = get_store_offset(st, phase);
5430 if (st_off < 0)
5431 break; // unknown junk in the inits
5432 if (st->in(MemNode::Memory) != zmem)
5433 break; // complicated store chains somehow in list
5434
5435 int st_size = st->as_Store()->memory_size();
5436 intptr_t next_init_off = st_off + st_size;
5437
5438 if (do_zeroing && zeroes_done < next_init_off) {
5439 // See if this store needs a zero before it or under it.
5440 intptr_t zeroes_needed = st_off;
5441
5442 if (st_size < BytesPerInt) {
5443 // Look for subword stores which only partially initialize words.
5444 // If we find some, we must lay down some word-level zeroes first,
5445 // underneath the subword stores.
5446 //
5447 // Examples:
5448 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
5449 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
5450 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
5451 //
5452 // Note: coalesce_subword_stores may have already done this,
5453 // if it was prompted by constant non-zero subword initializers.
5454 // But this case can still arise with non-constant stores.
5455
5456 intptr_t next_full_store = find_next_fullword_store(i, phase);
5457
5458 // In the examples above:
5459 // in(i) p q r s x y z
5460 // st_off 12 13 14 15 12 13 14
5461 // st_size 1 1 1 1 1 1 1
5462 // next_full_s. 12 16 16 16 16 16 16
5463 // z's_done 12 16 16 16 12 16 12
5464 // z's_needed 12 16 16 16 16 16 16
5465 // zsize 0 0 0 0 4 0 4
5466 if (next_full_store < 0) {
5467 // Conservative tack: Zero to end of current word.
5468 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
5469 } else {
5470 // Zero to beginning of next fully initialized word.
5471 // Or, don't zero at all, if we are already in that word.
5472 assert(next_full_store >= zeroes_needed, "must go forward");
5473 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
5474 zeroes_needed = next_full_store;
5475 }
5476 }
5477
5478 if (zeroes_needed > zeroes_done) {
5479 intptr_t zsize = zeroes_needed - zeroes_done;
5480 // Do some incremental zeroing on rawmem, in parallel with inits.
5481 zeroes_done = align_down(zeroes_done, BytesPerInt);
5482 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5483 allocation()->in(AllocateNode::InitValue),
5484 allocation()->in(AllocateNode::RawInitValue),
5485 zeroes_done, zeroes_needed,
5486 phase);
5487 zeroes_done = zeroes_needed;
5488 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
5489 do_zeroing = false; // leave the hole, next time
5490 }
5491 }
5492
5493 // Collect the store and move on:
5494 phase->replace_input_of(st, MemNode::Memory, inits);
5495 inits = st; // put it on the linearized chain
5496 set_req(i, zmem); // unhook from previous position
5497
5498 if (zeroes_done == st_off)
5499 zeroes_done = next_init_off;
5500
5501 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
5502
5503 #ifdef ASSERT
5504 // Various order invariants. Weaker than stores_are_sane because
5505 // a large constant tile can be filled in by smaller non-constant stores.
5506 assert(st_off >= last_init_off, "inits do not reverse");
5507 last_init_off = st_off;
5508 const Type* val = nullptr;
5509 if (st_size >= BytesPerInt &&
5510 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
5511 (int)val->basic_type() < (int)T_OBJECT) {
5512 assert(st_off >= last_tile_end, "tiles do not overlap");
5513 assert(st_off >= last_init_end, "tiles do not overwrite inits");
5514 last_tile_end = MAX2(last_tile_end, next_init_off);
5515 } else {
5516 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
5517 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
5518 assert(st_off >= last_init_end, "inits do not overlap");
5519 last_init_end = next_init_off; // it's a non-tile
5520 }
5521 #endif //ASSERT
5522 }
5523
5524 remove_extra_zeroes(); // clear out all the zmems left over
5525 add_req(inits);
5526
5527 if (!(UseTLAB && ZeroTLAB)) {
5528 // If anything remains to be zeroed, zero it all now.
5529 zeroes_done = align_down(zeroes_done, BytesPerInt);
5530 // if it is the last unused 4 bytes of an instance, forget about it
5531 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
5532 if (zeroes_done + BytesPerLong >= size_limit) {
5533 AllocateNode* alloc = allocation();
5534 assert(alloc != nullptr, "must be present");
5535 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
5536 Node* klass_node = alloc->in(AllocateNode::KlassNode);
5537 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
5538 if (zeroes_done == k->layout_helper())
5539 zeroes_done = size_limit;
5540 }
5541 }
5542 if (zeroes_done < size_limit) {
5543 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5544 allocation()->in(AllocateNode::InitValue),
5545 allocation()->in(AllocateNode::RawInitValue),
5546 zeroes_done, size_in_bytes, phase);
5547 }
5548 }
5549
5550 set_complete(phase);
5551 return rawmem;
5552 }
5553
5554
5555 #ifdef ASSERT
5556 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
5557 if (is_complete())
5558 return true; // stores could be anything at this point
5559 assert(allocation() != nullptr, "must be present");
5560 intptr_t last_off = allocation()->minimum_header_size();
5561 for (uint i = InitializeNode::RawStores; i < req(); i++) {
5562 Node* st = in(i);
5563 intptr_t st_off = get_store_offset(st, phase);
5564 if (st_off < 0) continue; // ignore dead garbage
5565 if (last_off > st_off) {
5566 tty->print_cr("*** bad store offset at %d: %zd > %zd", i, last_off, st_off);
5567 this->dump(2);
5568 assert(false, "ascending store offsets");
5569 return false;
5570 }
5571 last_off = st_off + st->as_Store()->memory_size();
5572 }
5573 return true;
5574 }
5575 #endif //ASSERT
5576
5577
5578
5579
5580 //============================MergeMemNode=====================================
5581 //
5582 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
5583 // contributing store or call operations. Each contributor provides the memory
5584 // state for a particular "alias type" (see Compile::alias_type). For example,
5585 // if a MergeMem has an input X for alias category #6, then any memory reference
5586 // to alias category #6 may use X as its memory state input, as an exact equivalent
5587 // to using the MergeMem as a whole.
5588 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
5589 //
5590 // (Here, the <N> notation gives the index of the relevant adr_type.)
5591 //
5592 // In one special case (and more cases in the future), alias categories overlap.
5593 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
5594 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
5595 // it is exactly equivalent to that state W:
5596 // MergeMem(<Bot>: W) <==> W
5597 //
5598 // Usually, the merge has more than one input. In that case, where inputs
5599 // overlap (i.e., one is Bot), the narrower alias type determines the memory
5600 // state for that type, and the wider alias type (Bot) fills in everywhere else:
5601 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
5602 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
5603 //
5604 // A merge can take a "wide" memory state as one of its narrow inputs.
5605 // This simply means that the merge observes out only the relevant parts of
5606 // the wide input. That is, wide memory states arriving at narrow merge inputs
5607 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
5608 //
5609 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
5610 // and that memory slices "leak through":
5611 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
5612 //
5613 // But, in such a cascade, repeated memory slices can "block the leak":
5614 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
5615 //
5616 // In the last example, Y is not part of the combined memory state of the
5617 // outermost MergeMem. The system must, of course, prevent unschedulable
5618 // memory states from arising, so you can be sure that the state Y is somehow
5619 // a precursor to state Y'.
5620 //
5621 //
5622 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
5623 // of each MergeMemNode array are exactly the numerical alias indexes, including
5624 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
5625 // Compile::alias_type (and kin) produce and manage these indexes.
5626 //
5627 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
5628 // (Note that this provides quick access to the top node inside MergeMem methods,
5629 // without the need to reach out via TLS to Compile::current.)
5630 //
5631 // As a consequence of what was just described, a MergeMem that represents a full
5632 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
5633 // containing all alias categories.
5634 //
5635 // MergeMem nodes never (?) have control inputs, so in(0) is null.
5636 //
5637 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
5638 // a memory state for the alias type <N>, or else the top node, meaning that
5639 // there is no particular input for that alias type. Note that the length of
5640 // a MergeMem is variable, and may be extended at any time to accommodate new
5641 // memory states at larger alias indexes. When merges grow, they are of course
5642 // filled with "top" in the unused in() positions.
5643 //
5644 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
5645 // (Top was chosen because it works smoothly with passes like GCM.)
5646 //
5647 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
5648 // the type of random VM bits like TLS references.) Since it is always the
5649 // first non-Bot memory slice, some low-level loops use it to initialize an
5650 // index variable: for (i = AliasIdxRaw; i < req(); i++).
5651 //
5652 //
5653 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
5654 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
5655 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
5656 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
5657 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
5658 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
5659 //
5660 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
5661 // really that different from the other memory inputs. An abbreviation called
5662 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
5663 //
5664 //
5665 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
5666 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
5667 // that "emerges though" the base memory will be marked as excluding the alias types
5668 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
5669 //
5670 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
5671 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
5672 //
5673 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
5674 // (It is currently unimplemented.) As you can see, the resulting merge is
5675 // actually a disjoint union of memory states, rather than an overlay.
5676 //
5677
5678 //------------------------------MergeMemNode-----------------------------------
5679 Node* MergeMemNode::make_empty_memory() {
5680 Node* empty_memory = (Node*) Compile::current()->top();
5681 assert(empty_memory->is_top(), "correct sentinel identity");
5682 return empty_memory;
5683 }
5684
5685 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
5686 init_class_id(Class_MergeMem);
5687 // all inputs are nullified in Node::Node(int)
5688 // set_input(0, nullptr); // no control input
5689
5690 // Initialize the edges uniformly to top, for starters.
5691 Node* empty_mem = make_empty_memory();
5692 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
5693 init_req(i,empty_mem);
5694 }
5695 assert(empty_memory() == empty_mem, "");
5696
5697 if( new_base != nullptr && new_base->is_MergeMem() ) {
5698 MergeMemNode* mdef = new_base->as_MergeMem();
5699 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
5700 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
5701 mms.set_memory(mms.memory2());
5702 }
5703 assert(base_memory() == mdef->base_memory(), "");
5704 } else {
5705 set_base_memory(new_base);
5706 }
5707 }
5708
5709 // Make a new, untransformed MergeMem with the same base as 'mem'.
5710 // If mem is itself a MergeMem, populate the result with the same edges.
5711 MergeMemNode* MergeMemNode::make(Node* mem) {
5712 return new MergeMemNode(mem);
5713 }
5714
5715 //------------------------------cmp--------------------------------------------
5716 uint MergeMemNode::hash() const { return NO_HASH; }
5717 bool MergeMemNode::cmp( const Node &n ) const {
5718 return (&n == this); // Always fail except on self
5719 }
5720
5721 //------------------------------Identity---------------------------------------
5722 Node* MergeMemNode::Identity(PhaseGVN* phase) {
5723 // Identity if this merge point does not record any interesting memory
5724 // disambiguations.
5725 Node* base_mem = base_memory();
5726 Node* empty_mem = empty_memory();
5727 if (base_mem != empty_mem) { // Memory path is not dead?
5728 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5729 Node* mem = in(i);
5730 if (mem != empty_mem && mem != base_mem) {
5731 return this; // Many memory splits; no change
5732 }
5733 }
5734 }
5735 return base_mem; // No memory splits; ID on the one true input
5736 }
5737
5738 //------------------------------Ideal------------------------------------------
5739 // This method is invoked recursively on chains of MergeMem nodes
5740 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
5741 // Remove chain'd MergeMems
5742 //
5743 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
5744 // relative to the "in(Bot)". Since we are patching both at the same time,
5745 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
5746 // but rewrite each "in(i)" relative to the new "in(Bot)".
5747 Node *progress = nullptr;
5748
5749
5750 Node* old_base = base_memory();
5751 Node* empty_mem = empty_memory();
5752 if (old_base == empty_mem)
5753 return nullptr; // Dead memory path.
5754
5755 MergeMemNode* old_mbase;
5756 if (old_base != nullptr && old_base->is_MergeMem())
5757 old_mbase = old_base->as_MergeMem();
5758 else
5759 old_mbase = nullptr;
5760 Node* new_base = old_base;
5761
5762 // simplify stacked MergeMems in base memory
5763 if (old_mbase) new_base = old_mbase->base_memory();
5764
5765 // the base memory might contribute new slices beyond my req()
5766 if (old_mbase) grow_to_match(old_mbase);
5767
5768 // Note: We do not call verify_sparse on entry, because inputs
5769 // can normalize to the base_memory via subsume_node or similar
5770 // mechanisms. This method repairs that damage.
5771
5772 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
5773
5774 // Look at each slice.
5775 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5776 Node* old_in = in(i);
5777 // calculate the old memory value
5778 Node* old_mem = old_in;
5779 if (old_mem == empty_mem) old_mem = old_base;
5780 assert(old_mem == memory_at(i), "");
5781
5782 // maybe update (reslice) the old memory value
5783
5784 // simplify stacked MergeMems
5785 Node* new_mem = old_mem;
5786 MergeMemNode* old_mmem;
5787 if (old_mem != nullptr && old_mem->is_MergeMem())
5788 old_mmem = old_mem->as_MergeMem();
5789 else
5790 old_mmem = nullptr;
5791 if (old_mmem == this) {
5792 // This can happen if loops break up and safepoints disappear.
5793 // A merge of BotPtr (default) with a RawPtr memory derived from a
5794 // safepoint can be rewritten to a merge of the same BotPtr with
5795 // the BotPtr phi coming into the loop. If that phi disappears
5796 // also, we can end up with a self-loop of the mergemem.
5797 // In general, if loops degenerate and memory effects disappear,
5798 // a mergemem can be left looking at itself. This simply means
5799 // that the mergemem's default should be used, since there is
5800 // no longer any apparent effect on this slice.
5801 // Note: If a memory slice is a MergeMem cycle, it is unreachable
5802 // from start. Update the input to TOP.
5803 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
5804 }
5805 else if (old_mmem != nullptr) {
5806 new_mem = old_mmem->memory_at(i);
5807 }
5808 // else preceding memory was not a MergeMem
5809
5810 // maybe store down a new value
5811 Node* new_in = new_mem;
5812 if (new_in == new_base) new_in = empty_mem;
5813
5814 if (new_in != old_in) {
5815 // Warning: Do not combine this "if" with the previous "if"
5816 // A memory slice might have be be rewritten even if it is semantically
5817 // unchanged, if the base_memory value has changed.
5818 set_req_X(i, new_in, phase);
5819 progress = this; // Report progress
5820 }
5821 }
5822
5823 if (new_base != old_base) {
5824 set_req_X(Compile::AliasIdxBot, new_base, phase);
5825 // Don't use set_base_memory(new_base), because we need to update du.
5826 assert(base_memory() == new_base, "");
5827 progress = this;
5828 }
5829
5830 if( base_memory() == this ) {
5831 // a self cycle indicates this memory path is dead
5832 set_req(Compile::AliasIdxBot, empty_mem);
5833 }
5834
5835 // Resolve external cycles by calling Ideal on a MergeMem base_memory
5836 // Recursion must occur after the self cycle check above
5837 if( base_memory()->is_MergeMem() ) {
5838 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
5839 Node *m = phase->transform(new_mbase); // Rollup any cycles
5840 if( m != nullptr &&
5841 (m->is_top() ||
5842 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
5843 // propagate rollup of dead cycle to self
5844 set_req(Compile::AliasIdxBot, empty_mem);
5845 }
5846 }
5847
5848 if( base_memory() == empty_mem ) {
5849 progress = this;
5850 // Cut inputs during Parse phase only.
5851 // During Optimize phase a dead MergeMem node will be subsumed by Top.
5852 if( !can_reshape ) {
5853 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5854 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
5855 }
5856 }
5857 }
5858
5859 if( !progress && base_memory()->is_Phi() && can_reshape ) {
5860 // Check if PhiNode::Ideal's "Split phis through memory merges"
5861 // transform should be attempted. Look for this->phi->this cycle.
5862 uint merge_width = req();
5863 if (merge_width > Compile::AliasIdxRaw) {
5864 PhiNode* phi = base_memory()->as_Phi();
5865 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
5866 if (phi->in(i) == this) {
5867 phase->is_IterGVN()->_worklist.push(phi);
5868 break;
5869 }
5870 }
5871 }
5872 }
5873
5874 assert(progress || verify_sparse(), "please, no dups of base");
5875 return progress;
5876 }
5877
5878 //-------------------------set_base_memory-------------------------------------
5879 void MergeMemNode::set_base_memory(Node *new_base) {
5880 Node* empty_mem = empty_memory();
5881 set_req(Compile::AliasIdxBot, new_base);
5882 assert(memory_at(req()) == new_base, "must set default memory");
5883 // Clear out other occurrences of new_base:
5884 if (new_base != empty_mem) {
5885 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5886 if (in(i) == new_base) set_req(i, empty_mem);
5887 }
5888 }
5889 }
5890
5891 //------------------------------out_RegMask------------------------------------
5892 const RegMask &MergeMemNode::out_RegMask() const {
5893 return RegMask::Empty;
5894 }
5895
5896 //------------------------------dump_spec--------------------------------------
5897 #ifndef PRODUCT
5898 void MergeMemNode::dump_spec(outputStream *st) const {
5899 st->print(" {");
5900 Node* base_mem = base_memory();
5901 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
5902 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem;
5903 if (mem == base_mem) { st->print(" -"); continue; }
5904 st->print( " N%d:", mem->_idx );
5905 Compile::current()->get_adr_type(i)->dump_on(st);
5906 }
5907 st->print(" }");
5908 }
5909 #endif // !PRODUCT
5910
5911
5912 #ifdef ASSERT
5913 static bool might_be_same(Node* a, Node* b) {
5914 if (a == b) return true;
5915 if (!(a->is_Phi() || b->is_Phi())) return false;
5916 // phis shift around during optimization
5917 return true; // pretty stupid...
5918 }
5919
5920 // verify a narrow slice (either incoming or outgoing)
5921 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
5922 if (!VerifyAliases) return; // don't bother to verify unless requested
5923 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
5924 if (Node::in_dump()) return; // muzzle asserts when printing
5925 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
5926 assert(n != nullptr, "");
5927 // Elide intervening MergeMem's
5928 while (n->is_MergeMem()) {
5929 n = n->as_MergeMem()->memory_at(alias_idx);
5930 }
5931 Compile* C = Compile::current();
5932 const TypePtr* n_adr_type = n->adr_type();
5933 if (n == m->empty_memory()) {
5934 // Implicit copy of base_memory()
5935 } else if (n_adr_type != TypePtr::BOTTOM) {
5936 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type");
5937 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
5938 } else {
5939 // A few places like make_runtime_call "know" that VM calls are narrow,
5940 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
5941 bool expected_wide_mem = false;
5942 if (n == m->base_memory()) {
5943 expected_wide_mem = true;
5944 } else if (alias_idx == Compile::AliasIdxRaw ||
5945 n == m->memory_at(Compile::AliasIdxRaw)) {
5946 expected_wide_mem = true;
5947 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
5948 // memory can "leak through" calls on channels that
5949 // are write-once. Allow this also.
5950 expected_wide_mem = true;
5951 }
5952 assert(expected_wide_mem, "expected narrow slice replacement");
5953 }
5954 }
5955 #else // !ASSERT
5956 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
5957 #endif
5958
5959
5960 //-----------------------------memory_at---------------------------------------
5961 Node* MergeMemNode::memory_at(uint alias_idx) const {
5962 assert(alias_idx >= Compile::AliasIdxRaw ||
5963 (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()),
5964 "must avoid base_memory and AliasIdxTop");
5965
5966 // Otherwise, it is a narrow slice.
5967 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
5968 if (is_empty_memory(n)) {
5969 // the array is sparse; empty slots are the "top" node
5970 n = base_memory();
5971 assert(Node::in_dump()
5972 || n == nullptr || n->bottom_type() == Type::TOP
5973 || n->adr_type() == nullptr // address is TOP
5974 || n->adr_type() == TypePtr::BOTTOM
5975 || n->adr_type() == TypeRawPtr::BOTTOM
5976 || !Compile::current()->do_aliasing(),
5977 "must be a wide memory");
5978 // do_aliasing == false if we are organizing the memory states manually.
5979 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
5980 } else {
5981 // make sure the stored slice is sane
5982 #ifdef ASSERT
5983 if (VMError::is_error_reported() || Node::in_dump()) {
5984 } else if (might_be_same(n, base_memory())) {
5985 // Give it a pass: It is a mostly harmless repetition of the base.
5986 // This can arise normally from node subsumption during optimization.
5987 } else {
5988 verify_memory_slice(this, alias_idx, n);
5989 }
5990 #endif
5991 }
5992 return n;
5993 }
5994
5995 //---------------------------set_memory_at-------------------------------------
5996 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
5997 verify_memory_slice(this, alias_idx, n);
5998 Node* empty_mem = empty_memory();
5999 if (n == base_memory()) n = empty_mem; // collapse default
6000 uint need_req = alias_idx+1;
6001 if (req() < need_req) {
6002 if (n == empty_mem) return; // already the default, so do not grow me
6003 // grow the sparse array
6004 do {
6005 add_req(empty_mem);
6006 } while (req() < need_req);
6007 }
6008 set_req( alias_idx, n );
6009 }
6010
6011
6012
6013 //--------------------------iteration_setup------------------------------------
6014 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
6015 if (other != nullptr) {
6016 grow_to_match(other);
6017 // invariant: the finite support of mm2 is within mm->req()
6018 #ifdef ASSERT
6019 for (uint i = req(); i < other->req(); i++) {
6020 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
6021 }
6022 #endif
6023 }
6024 // Replace spurious copies of base_memory by top.
6025 Node* base_mem = base_memory();
6026 if (base_mem != nullptr && !base_mem->is_top()) {
6027 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
6028 if (in(i) == base_mem)
6029 set_req(i, empty_memory());
6030 }
6031 }
6032 }
6033
6034 //---------------------------grow_to_match-------------------------------------
6035 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
6036 Node* empty_mem = empty_memory();
6037 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
6038 // look for the finite support of the other memory
6039 for (uint i = other->req(); --i >= req(); ) {
6040 if (other->in(i) != empty_mem) {
6041 uint new_len = i+1;
6042 while (req() < new_len) add_req(empty_mem);
6043 break;
6044 }
6045 }
6046 }
6047
6048 //---------------------------verify_sparse-------------------------------------
6049 #ifndef PRODUCT
6050 bool MergeMemNode::verify_sparse() const {
6051 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
6052 Node* base_mem = base_memory();
6053 // The following can happen in degenerate cases, since empty==top.
6054 if (is_empty_memory(base_mem)) return true;
6055 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6056 assert(in(i) != nullptr, "sane slice");
6057 if (in(i) == base_mem) return false; // should have been the sentinel value!
6058 }
6059 return true;
6060 }
6061
6062 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
6063 Node* n;
6064 n = mm->in(idx);
6065 if (mem == n) return true; // might be empty_memory()
6066 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
6067 if (mem == n) return true;
6068 return false;
6069 }
6070 #endif // !PRODUCT