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