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 (UseCompactObjectHeaders) {
1856 if (tkls->offset() == in_bytes(Klass::prototype_header_offset())) {
1857 // The field is Klass::_prototype_header. Return its (constant) value.
1858 assert(this->Opcode() == Op_LoadX, "must load a proper type from _prototype_header");
1859 return TypeX::make(klass->prototype_header());
1860 }
1861 }
1862 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1863 // The field is Klass::_modifier_flags. Return its (constant) value.
1864 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1865 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1866 return TypeInt::make(klass->modifier_flags());
1867 }
1868 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1869 // The field is Klass::_access_flags. Return its (constant) value.
1870 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1871 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1872 return TypeInt::make(klass->access_flags());
1873 }
1874 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1875 // The field is Klass::_layout_helper. Return its constant value if known.
1876 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1877 return TypeInt::make(klass->layout_helper());
1878 }
1879
1880 // No match.
1881 return nullptr;
1882 }
1883
1884 //------------------------------Value-----------------------------------------
1885 const Type* LoadNode::Value(PhaseGVN* phase) const {
1886 // Either input is TOP ==> the result is TOP
1887 Node* mem = in(MemNode::Memory);
1888 const Type *t1 = phase->type(mem);
1889 if (t1 == Type::TOP) return Type::TOP;
1890 Node* adr = in(MemNode::Address);
1891 const TypePtr* tp = phase->type(adr)->isa_ptr();
1892 if (tp == nullptr || tp->empty()) return Type::TOP;
1893 int off = tp->offset();
1894 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1895 Compile* C = phase->C;
1896
1897 // Try to guess loaded type from pointer type
1898 if (tp->isa_aryptr()) {
1899 const TypeAryPtr* ary = tp->is_aryptr();
1900 const Type* t = ary->elem();
1901
1902 // Determine whether the reference is beyond the header or not, by comparing
1903 // the offset against the offset of the start of the array's data.
1904 // Different array types begin at slightly different offsets (12 vs. 16).
1905 // We choose T_BYTE as an example base type that is least restrictive
1906 // as to alignment, which will therefore produce the smallest
1907 // possible base offset.
1908 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1909 const bool off_beyond_header = (off >= min_base_off);
1910
1911 // Try to constant-fold a stable array element.
1912 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1913 // Make sure the reference is not into the header and the offset is constant
1914 ciObject* aobj = ary->const_oop();
1915 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1916 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1917 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1918 stable_dimension,
1919 memory_type(), is_unsigned());
1920 if (con_type != nullptr) {
1921 return con_type;
1922 }
1923 }
1924 }
1925
1926 // Don't do this for integer types. There is only potential profit if
1927 // the element type t is lower than _type; that is, for int types, if _type is
1928 // more restrictive than t. This only happens here if one is short and the other
1929 // char (both 16 bits), and in those cases we've made an intentional decision
1930 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1931 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1932 //
1933 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1934 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1935 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1936 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1937 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1938 // In fact, that could have been the original type of p1, and p1 could have
1939 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1940 // expression (LShiftL quux 3) independently optimized to the constant 8.
1941 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
1942 && (_type->isa_vect() == nullptr)
1943 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1944 // t might actually be lower than _type, if _type is a unique
1945 // concrete subclass of abstract class t.
1946 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
1947 const Type* jt = t->join_speculative(_type);
1948 // In any case, do not allow the join, per se, to empty out the type.
1949 if (jt->empty() && !t->empty()) {
1950 // This can happen if a interface-typed array narrows to a class type.
1951 jt = _type;
1952 }
1953 #ifdef ASSERT
1954 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1955 // The pointers in the autobox arrays are always non-null
1956 Node* base = adr->in(AddPNode::Base);
1957 if ((base != nullptr) && base->is_DecodeN()) {
1958 // Get LoadN node which loads IntegerCache.cache field
1959 base = base->in(1);
1960 }
1961 if ((base != nullptr) && base->is_Con()) {
1962 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1963 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
1964 // It could be narrow oop
1965 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1966 }
1967 }
1968 }
1969 #endif
1970 return jt;
1971 }
1972 }
1973 } else if (tp->base() == Type::InstPtr) {
1974 assert( off != Type::OffsetBot ||
1975 // arrays can be cast to Objects
1976 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1977 // unsafe field access may not have a constant offset
1978 C->has_unsafe_access(),
1979 "Field accesses must be precise" );
1980 // For oop loads, we expect the _type to be precise.
1981
1982 // Optimize loads from constant fields.
1983 const TypeInstPtr* tinst = tp->is_instptr();
1984 ciObject* const_oop = tinst->const_oop();
1985 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
1986 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
1987 if (con_type != nullptr) {
1988 return con_type;
1989 }
1990 }
1991 } else if (tp->base() == Type::KlassPtr) {
1992 assert( off != Type::OffsetBot ||
1993 // arrays can be cast to Objects
1994 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1995 // also allow array-loading from the primary supertype
1996 // array during subtype checks
1997 Opcode() == Op_LoadKlass,
1998 "Field accesses must be precise" );
1999 // For klass/static loads, we expect the _type to be precise
2000 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2001 /* With mirrors being an indirect in the Klass*
2002 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2003 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2004 *
2005 * So check the type and klass of the node before the LoadP.
2006 */
2007 Node* adr2 = adr->in(MemNode::Address);
2008 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2009 if (tkls != nullptr && !StressReflectiveCode) {
2010 ciKlass* klass = tkls->klass();
2011 if (klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2012 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2013 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2014 return TypeInstPtr::make(klass->java_mirror());
2015 }
2016 }
2017 }
2018
2019 const TypeKlassPtr *tkls = tp->isa_klassptr();
2020 if (tkls != nullptr && !StressReflectiveCode) {
2021 ciKlass* klass = tkls->klass();
2022 if (klass->is_loaded() && tkls->klass_is_exact()) {
2023 // We are loading a field from a Klass metaobject whose identity
2024 // is known at compile time (the type is "exact" or "precise").
2025 // Check for fields we know are maintained as constants by the VM.
2026 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2027 // The field is Klass::_super_check_offset. Return its (constant) value.
2028 // (Folds up type checking code.)
2029 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2030 return TypeInt::make(klass->super_check_offset());
2031 }
2032 if (UseCompactObjectHeaders) {
2033 if (tkls->offset() == in_bytes(Klass::prototype_header_offset())) {
2034 // The field is Klass::_prototype_header. Return its (constant) value.
2035 assert(this->Opcode() == Op_LoadX, "must load a proper type from _prototype_header");
2036 return TypeX::make(klass->prototype_header());
2037 }
2038 }
2039 // Compute index into primary_supers array
2040 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2041 // Check for overflowing; use unsigned compare to handle the negative case.
2042 if( depth < ciKlass::primary_super_limit() ) {
2043 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2044 // (Folds up type checking code.)
2045 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2046 ciKlass *ss = klass->super_of_depth(depth);
2047 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
2048 }
2049 const Type* aift = load_array_final_field(tkls, klass);
2050 if (aift != nullptr) return aift;
2051 }
2052
2053 // We can still check if we are loading from the primary_supers array at a
2054 // shallow enough depth. Even though the klass is not exact, entries less
2055 // than or equal to its super depth are correct.
2056 if (klass->is_loaded() ) {
2057 ciType *inner = klass;
2058 while( inner->is_obj_array_klass() )
2059 inner = inner->as_obj_array_klass()->base_element_type();
2060 if( inner->is_instance_klass() &&
2061 !inner->as_instance_klass()->flags().is_interface() ) {
2062 // Compute index into primary_supers array
2063 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2064 // Check for overflowing; use unsigned compare to handle the negative case.
2065 if( depth < ciKlass::primary_super_limit() &&
2066 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
2067 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2068 // (Folds up type checking code.)
2069 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2070 ciKlass *ss = klass->super_of_depth(depth);
2071 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
2072 }
2073 }
2074 }
2075
2076 // If the type is enough to determine that the thing is not an array,
2077 // we can give the layout_helper a positive interval type.
2078 // This will help short-circuit some reflective code.
2079 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
2080 && !klass->is_array_klass() // not directly typed as an array
2081 && !klass->is_interface() // specifically not Serializable & Cloneable
2082 && !klass->is_java_lang_Object() // not the supertype of all T[]
2083 ) {
2084 // Note: When interfaces are reliable, we can narrow the interface
2085 // test to (klass != Serializable && klass != Cloneable).
2086 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2087 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2088 // The key property of this type is that it folds up tests
2089 // for array-ness, since it proves that the layout_helper is positive.
2090 // Thus, a generic value like the basic object layout helper works fine.
2091 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2092 }
2093 }
2094
2095 // If we are loading from a freshly-allocated object, produce a zero,
2096 // if the load is provably beyond the header of the object.
2097 // (Also allow a variable load from a fresh array to produce zero.)
2098 const TypeOopPtr *tinst = tp->isa_oopptr();
2099 bool is_instance = (tinst != nullptr) && tinst->is_known_instance_field();
2100 bool is_boxed_value = (tinst != nullptr) && tinst->is_ptr_to_boxed_value();
2101 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
2102 Node* value = can_see_stored_value(mem,phase);
2103 if (value != nullptr && value->is_Con()) {
2104 assert(value->bottom_type()->higher_equal(_type),"sanity");
2105 return value->bottom_type();
2106 }
2107 }
2108
2109 bool is_vect = (_type->isa_vect() != nullptr);
2110 if (is_instance && !is_vect) {
2111 // If we have an instance type and our memory input is the
2112 // programs's initial memory state, there is no matching store,
2113 // so just return a zero of the appropriate type -
2114 // except if it is vectorized - then we have no zero constant.
2115 Node *mem = in(MemNode::Memory);
2116 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2117 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2118 return Type::get_zero_type(_type->basic_type());
2119 }
2120 }
2121
2122 Node* alloc = is_new_object_mark_load(phase);
2123 if (alloc != nullptr && !(alloc->Opcode() == Op_Allocate && UseBiasedLocking) && !UseCompactObjectHeaders) {
2124 return TypeX::make(markWord::prototype().value());
2125 }
2126
2127 return _type;
2128 }
2129
2130 //------------------------------match_edge-------------------------------------
2131 // Do we Match on this edge index or not? Match only the address.
2132 uint LoadNode::match_edge(uint idx) const {
2133 return idx == MemNode::Address;
2134 }
2135
2136 //--------------------------LoadBNode::Ideal--------------------------------------
2137 //
2138 // If the previous store is to the same address as this load,
2139 // and the value stored was larger than a byte, replace this load
2140 // with the value stored truncated to a byte. If no truncation is
2141 // needed, the replacement is done in LoadNode::Identity().
2142 //
2143 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2144 Node* mem = in(MemNode::Memory);
2145 Node* value = can_see_stored_value(mem,phase);
2146 if (value != nullptr) {
2147 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2148 if (narrow != value) {
2149 return narrow;
2150 }
2151 }
2152 // Identity call will handle the case where truncation is not needed.
2153 return LoadNode::Ideal(phase, can_reshape);
2154 }
2155
2156 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2157 Node* mem = in(MemNode::Memory);
2158 Node* value = can_see_stored_value(mem,phase);
2159 if (value != nullptr && value->is_Con() &&
2160 !value->bottom_type()->higher_equal(_type)) {
2161 // If the input to the store does not fit with the load's result type,
2162 // it must be truncated. We can't delay until Ideal call since
2163 // a singleton Value is needed for split_thru_phi optimization.
2164 int con = value->get_int();
2165 return TypeInt::make((con << 24) >> 24);
2166 }
2167 return LoadNode::Value(phase);
2168 }
2169
2170 //--------------------------LoadUBNode::Ideal-------------------------------------
2171 //
2172 // If the previous store is to the same address as this load,
2173 // and the value stored was larger than a byte, replace this load
2174 // with the value stored truncated to a byte. If no truncation is
2175 // needed, the replacement is done in LoadNode::Identity().
2176 //
2177 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2178 Node* mem = in(MemNode::Memory);
2179 Node* value = can_see_stored_value(mem, phase);
2180 if (value != nullptr) {
2181 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2182 if (narrow != value) {
2183 return narrow;
2184 }
2185 }
2186 // Identity call will handle the case where truncation is not needed.
2187 return LoadNode::Ideal(phase, can_reshape);
2188 }
2189
2190 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2191 Node* mem = in(MemNode::Memory);
2192 Node* value = can_see_stored_value(mem,phase);
2193 if (value != nullptr && value->is_Con() &&
2194 !value->bottom_type()->higher_equal(_type)) {
2195 // If the input to the store does not fit with the load's result type,
2196 // it must be truncated. We can't delay until Ideal call since
2197 // a singleton Value is needed for split_thru_phi optimization.
2198 int con = value->get_int();
2199 return TypeInt::make(con & 0xFF);
2200 }
2201 return LoadNode::Value(phase);
2202 }
2203
2204 //--------------------------LoadUSNode::Ideal-------------------------------------
2205 //
2206 // If the previous store is to the same address as this load,
2207 // and the value stored was larger than a char, replace this load
2208 // with the value stored truncated to a char. If no truncation is
2209 // needed, the replacement is done in LoadNode::Identity().
2210 //
2211 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2212 Node* mem = in(MemNode::Memory);
2213 Node* value = can_see_stored_value(mem,phase);
2214 if (value != nullptr) {
2215 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2216 if (narrow != value) {
2217 return narrow;
2218 }
2219 }
2220 // Identity call will handle the case where truncation is not needed.
2221 return LoadNode::Ideal(phase, can_reshape);
2222 }
2223
2224 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2225 Node* mem = in(MemNode::Memory);
2226 Node* value = can_see_stored_value(mem,phase);
2227 if (value != nullptr && value->is_Con() &&
2228 !value->bottom_type()->higher_equal(_type)) {
2229 // If the input to the store does not fit with the load's result type,
2230 // it must be truncated. We can't delay until Ideal call since
2231 // a singleton Value is needed for split_thru_phi optimization.
2232 int con = value->get_int();
2233 return TypeInt::make(con & 0xFFFF);
2234 }
2235 return LoadNode::Value(phase);
2236 }
2237
2238 //--------------------------LoadSNode::Ideal--------------------------------------
2239 //
2240 // If the previous store is to the same address as this load,
2241 // and the value stored was larger than a short, replace this load
2242 // with the value stored truncated to a short. If no truncation is
2243 // needed, the replacement is done in LoadNode::Identity().
2244 //
2245 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2246 Node* mem = in(MemNode::Memory);
2247 Node* value = can_see_stored_value(mem,phase);
2248 if (value != nullptr) {
2249 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2250 if (narrow != value) {
2251 return narrow;
2252 }
2253 }
2254 // Identity call will handle the case where truncation is not needed.
2255 return LoadNode::Ideal(phase, can_reshape);
2256 }
2257
2258 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2259 Node* mem = in(MemNode::Memory);
2260 Node* value = can_see_stored_value(mem,phase);
2261 if (value != nullptr && value->is_Con() &&
2262 !value->bottom_type()->higher_equal(_type)) {
2263 // If the input to the store does not fit with the load's result type,
2264 // it must be truncated. We can't delay until Ideal call since
2265 // a singleton Value is needed for split_thru_phi optimization.
2266 int con = value->get_int();
2267 return TypeInt::make((con << 16) >> 16);
2268 }
2269 return LoadNode::Value(phase);
2270 }
2271
2272 //=============================================================================
2273 //----------------------------LoadKlassNode::make------------------------------
2274 // Polymorphic factory method:
2275 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2276 // sanity check the alias category against the created node type
2277 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2278 assert(adr_type != nullptr, "expecting TypeKlassPtr");
2279 #ifdef _LP64
2280 if (adr_type->is_ptr_to_narrowklass()) {
2281 assert(UseCompressedClassPointers, "no compressed klasses");
2282 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2283 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2284 }
2285 #endif
2286 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2287 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2288 }
2289
2290 //------------------------------Value------------------------------------------
2291 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2292 return klass_value_common(phase);
2293 }
2294
2295 // In most cases, LoadKlassNode does not have the control input set. If the control
2296 // input is set, it must not be removed (by LoadNode::Ideal()).
2297 bool LoadKlassNode::can_remove_control() const {
2298 return false;
2299 }
2300
2301 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2302 // Either input is TOP ==> the result is TOP
2303 const Type *t1 = phase->type( in(MemNode::Memory) );
2304 if (t1 == Type::TOP) return Type::TOP;
2305 Node *adr = in(MemNode::Address);
2306 const Type *t2 = phase->type( adr );
2307 if (t2 == Type::TOP) return Type::TOP;
2308 const TypePtr *tp = t2->is_ptr();
2309 if (TypePtr::above_centerline(tp->ptr()) ||
2310 tp->ptr() == TypePtr::Null) return Type::TOP;
2311
2312 // Return a more precise klass, if possible
2313 const TypeInstPtr *tinst = tp->isa_instptr();
2314 if (tinst != nullptr) {
2315 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2316 int offset = tinst->offset();
2317 if (ik == phase->C->env()->Class_klass()
2318 && (offset == java_lang_Class::klass_offset() ||
2319 offset == java_lang_Class::array_klass_offset())) {
2320 // We are loading a special hidden field from a Class mirror object,
2321 // the field which points to the VM's Klass metaobject.
2322 ciType* t = tinst->java_mirror_type();
2323 // java_mirror_type returns non-null for compile-time Class constants.
2324 if (t != nullptr) {
2325 // constant oop => constant klass
2326 if (offset == java_lang_Class::array_klass_offset()) {
2327 if (t->is_void()) {
2328 // We cannot create a void array. Since void is a primitive type return null
2329 // klass. Users of this result need to do a null check on the returned klass.
2330 return TypePtr::NULL_PTR;
2331 }
2332 return TypeKlassPtr::make(ciArrayKlass::make(t));
2333 }
2334 if (!t->is_klass()) {
2335 // a primitive Class (e.g., int.class) has null for a klass field
2336 return TypePtr::NULL_PTR;
2337 }
2338 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2339 return TypeKlassPtr::make(t->as_klass());
2340 }
2341 // non-constant mirror, so we can't tell what's going on
2342 }
2343 if( !ik->is_loaded() )
2344 return _type; // Bail out if not loaded
2345 if (offset == oopDesc::klass_offset_in_bytes()) {
2346 if (tinst->klass_is_exact()) {
2347 return TypeKlassPtr::make(ik);
2348 }
2349 // See if we can become precise: no subklasses and no interface
2350 // (Note: We need to support verified interfaces.)
2351 if (!ik->is_interface() && !ik->has_subklass()) {
2352 // Add a dependence; if any subclass added we need to recompile
2353 if (!ik->is_final()) {
2354 // %%% should use stronger assert_unique_concrete_subtype instead
2355 phase->C->dependencies()->assert_leaf_type(ik);
2356 }
2357 // Return precise klass
2358 return TypeKlassPtr::make(ik);
2359 }
2360
2361 // Return root of possible klass
2362 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2363 }
2364 }
2365
2366 // Check for loading klass from an array
2367 const TypeAryPtr *tary = tp->isa_aryptr();
2368 if( tary != nullptr ) {
2369 ciKlass *tary_klass = tary->klass();
2370 if (tary_klass != nullptr // can be null when at BOTTOM or TOP
2371 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2372 if (tary->klass_is_exact()) {
2373 return TypeKlassPtr::make(tary_klass);
2374 }
2375 ciArrayKlass *ak = tary->klass()->as_array_klass();
2376 // If the klass is an object array, we defer the question to the
2377 // array component klass.
2378 if( ak->is_obj_array_klass() ) {
2379 assert( ak->is_loaded(), "" );
2380 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2381 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2382 ciInstanceKlass* ik = base_k->as_instance_klass();
2383 // See if we can become precise: no subklasses and no interface
2384 if (!ik->is_interface() && !ik->has_subklass()) {
2385 // Add a dependence; if any subclass added we need to recompile
2386 if (!ik->is_final()) {
2387 phase->C->dependencies()->assert_leaf_type(ik);
2388 }
2389 // Return precise array klass
2390 return TypeKlassPtr::make(ak);
2391 }
2392 }
2393 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2394 } else { // Found a type-array?
2395 assert( ak->is_type_array_klass(), "" );
2396 return TypeKlassPtr::make(ak); // These are always precise
2397 }
2398 }
2399 }
2400
2401 // Check for loading klass from an array klass
2402 const TypeKlassPtr *tkls = tp->isa_klassptr();
2403 if (tkls != nullptr && !StressReflectiveCode) {
2404 ciKlass* klass = tkls->klass();
2405 if( !klass->is_loaded() )
2406 return _type; // Bail out if not loaded
2407 if( klass->is_obj_array_klass() &&
2408 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2409 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2410 // // Always returning precise element type is incorrect,
2411 // // e.g., element type could be object and array may contain strings
2412 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2413
2414 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2415 // according to the element type's subclassing.
2416 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2417 }
2418 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2419 tkls->offset() == in_bytes(Klass::super_offset())) {
2420 ciKlass* sup = klass->as_instance_klass()->super();
2421 // The field is Klass::_super. Return its (constant) value.
2422 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2423 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2424 }
2425 }
2426
2427 // Bailout case
2428 return LoadNode::Value(phase);
2429 }
2430
2431 //------------------------------Identity---------------------------------------
2432 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2433 // Also feed through the klass in Allocate(...klass...)._klass.
2434 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2435 return klass_identity_common(phase);
2436 }
2437
2438 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2439 Node* x = LoadNode::Identity(phase);
2440 if (x != this) return x;
2441
2442 // Take apart the address into an oop and and offset.
2443 // Return 'this' if we cannot.
2444 Node* adr = in(MemNode::Address);
2445 intptr_t offset = 0;
2446 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2447 if (base == nullptr) return this;
2448 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2449 if (toop == nullptr) return this;
2450
2451 // Step over potential GC barrier for OopHandle resolve
2452 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2453 if (bs->is_gc_barrier_node(base)) {
2454 base = bs->step_over_gc_barrier(base);
2455 }
2456
2457 // We can fetch the klass directly through an AllocateNode.
2458 // This works even if the klass is not constant (clone or newArray).
2459 if (offset == oopDesc::klass_offset_in_bytes()) {
2460 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2461 if (allocated_klass != nullptr) {
2462 return allocated_klass;
2463 }
2464 }
2465
2466 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2467 // See inline_native_Class_query for occurrences of these patterns.
2468 // Java Example: x.getClass().isAssignableFrom(y)
2469 //
2470 // This improves reflective code, often making the Class
2471 // mirror go completely dead. (Current exception: Class
2472 // mirrors may appear in debug info, but we could clean them out by
2473 // introducing a new debug info operator for Klass.java_mirror).
2474
2475 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2476 && offset == java_lang_Class::klass_offset()) {
2477 if (base->is_Load()) {
2478 Node* base2 = base->in(MemNode::Address);
2479 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2480 Node* adr2 = base2->in(MemNode::Address);
2481 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2482 if (tkls != nullptr && !tkls->empty()
2483 && (tkls->klass()->is_instance_klass() ||
2484 tkls->klass()->is_array_klass())
2485 && adr2->is_AddP()
2486 ) {
2487 int mirror_field = in_bytes(Klass::java_mirror_offset());
2488 if (tkls->offset() == mirror_field) {
2489 return adr2->in(AddPNode::Base);
2490 }
2491 }
2492 }
2493 }
2494 }
2495
2496 return this;
2497 }
2498
2499
2500 //------------------------------Value------------------------------------------
2501 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2502 const Type *t = klass_value_common(phase);
2503 if (t == Type::TOP)
2504 return t;
2505
2506 return t->make_narrowklass();
2507 }
2508
2509 //------------------------------Identity---------------------------------------
2510 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2511 // Also feed through the klass in Allocate(...klass...)._klass.
2512 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2513 Node *x = klass_identity_common(phase);
2514
2515 const Type *t = phase->type( x );
2516 if( t == Type::TOP ) return x;
2517 if( t->isa_narrowklass()) return x;
2518 assert (!t->isa_narrowoop(), "no narrow oop here");
2519
2520 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2521 }
2522
2523 //------------------------------Value-----------------------------------------
2524 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2525 // Either input is TOP ==> the result is TOP
2526 const Type *t1 = phase->type( in(MemNode::Memory) );
2527 if( t1 == Type::TOP ) return Type::TOP;
2528 Node *adr = in(MemNode::Address);
2529 const Type *t2 = phase->type( adr );
2530 if( t2 == Type::TOP ) return Type::TOP;
2531 const TypePtr *tp = t2->is_ptr();
2532 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2533 const TypeAryPtr *tap = tp->isa_aryptr();
2534 if( !tap ) return _type;
2535 return tap->size();
2536 }
2537
2538 //-------------------------------Ideal---------------------------------------
2539 // Feed through the length in AllocateArray(...length...)._length.
2540 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2541 Node* p = MemNode::Ideal_common(phase, can_reshape);
2542 if (p) return (p == NodeSentinel) ? nullptr : p;
2543
2544 // Take apart the address into an oop and and offset.
2545 // Return 'this' if we cannot.
2546 Node* adr = in(MemNode::Address);
2547 intptr_t offset = 0;
2548 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2549 if (base == nullptr) return nullptr;
2550 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2551 if (tary == nullptr) return nullptr;
2552
2553 // We can fetch the length directly through an AllocateArrayNode.
2554 // This works even if the length is not constant (clone or newArray).
2555 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2556 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2557 if (alloc != nullptr) {
2558 Node* allocated_length = alloc->Ideal_length();
2559 Node* len = alloc->make_ideal_length(tary, phase);
2560 if (allocated_length != len) {
2561 // New CastII improves on this.
2562 return len;
2563 }
2564 }
2565 }
2566
2567 return nullptr;
2568 }
2569
2570 //------------------------------Identity---------------------------------------
2571 // Feed through the length in AllocateArray(...length...)._length.
2572 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2573 Node* x = LoadINode::Identity(phase);
2574 if (x != this) return x;
2575
2576 // Take apart the address into an oop and and offset.
2577 // Return 'this' if we cannot.
2578 Node* adr = in(MemNode::Address);
2579 intptr_t offset = 0;
2580 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2581 if (base == nullptr) return this;
2582 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2583 if (tary == nullptr) return this;
2584
2585 // We can fetch the length directly through an AllocateArrayNode.
2586 // This works even if the length is not constant (clone or newArray).
2587 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2588 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2589 if (alloc != nullptr) {
2590 Node* allocated_length = alloc->Ideal_length();
2591 // Do not allow make_ideal_length to allocate a CastII node.
2592 Node* len = alloc->make_ideal_length(tary, phase, false);
2593 if (allocated_length == len) {
2594 // Return allocated_length only if it would not be improved by a CastII.
2595 return allocated_length;
2596 }
2597 }
2598 }
2599
2600 return this;
2601
2602 }
2603
2604 //=============================================================================
2605 //---------------------------StoreNode::make-----------------------------------
2606 // Polymorphic factory method:
2607 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) {
2608 assert((mo == unordered || mo == release), "unexpected");
2609 Compile* C = gvn.C;
2610 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2611 ctl != nullptr, "raw memory operations should have control edge");
2612
2613 switch (bt) {
2614 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2615 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2616 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2617 case T_CHAR:
2618 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2619 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2620 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2621 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2622 case T_METADATA:
2623 case T_ADDRESS:
2624 case T_OBJECT:
2625 #ifdef _LP64
2626 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2627 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2628 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2629 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2630 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2631 adr->bottom_type()->isa_rawptr())) {
2632 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2633 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2634 }
2635 #endif
2636 {
2637 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2638 }
2639 default:
2640 ShouldNotReachHere();
2641 return (StoreNode*)nullptr;
2642 }
2643 }
2644
2645 //--------------------------bottom_type----------------------------------------
2646 const Type *StoreNode::bottom_type() const {
2647 return Type::MEMORY;
2648 }
2649
2650 //------------------------------hash-------------------------------------------
2651 uint StoreNode::hash() const {
2652 // unroll addition of interesting fields
2653 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2654
2655 // Since they are not commoned, do not hash them:
2656 return NO_HASH;
2657 }
2658
2659 //------------------------------Ideal------------------------------------------
2660 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2661 // When a store immediately follows a relevant allocation/initialization,
2662 // try to capture it into the initialization, or hoist it above.
2663 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2664 Node* p = MemNode::Ideal_common(phase, can_reshape);
2665 if (p) return (p == NodeSentinel) ? nullptr : p;
2666
2667 Node* mem = in(MemNode::Memory);
2668 Node* address = in(MemNode::Address);
2669 Node* value = in(MemNode::ValueIn);
2670 // Back-to-back stores to same address? Fold em up. Generally
2671 // unsafe if I have intervening uses... Also disallowed for StoreCM
2672 // since they must follow each StoreP operation. Redundant StoreCMs
2673 // are eliminated just before matching in final_graph_reshape.
2674 {
2675 Node* st = mem;
2676 // If Store 'st' has more than one use, we cannot fold 'st' away.
2677 // For example, 'st' might be the final state at a conditional
2678 // return. Or, 'st' might be used by some node which is live at
2679 // the same time 'st' is live, which might be unschedulable. So,
2680 // require exactly ONE user until such time as we clone 'mem' for
2681 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2682 // true).
2683 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2684 // Looking at a dead closed cycle of memory?
2685 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2686 assert(Opcode() == st->Opcode() ||
2687 st->Opcode() == Op_StoreVector ||
2688 Opcode() == Op_StoreVector ||
2689 st->Opcode() == Op_StoreVectorScatter ||
2690 Opcode() == Op_StoreVectorScatter ||
2691 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2692 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2693 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2694 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2695 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2696
2697 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2698 st->as_Store()->memory_size() <= this->memory_size()) {
2699 Node* use = st->raw_out(0);
2700 if (phase->is_IterGVN()) {
2701 phase->is_IterGVN()->rehash_node_delayed(use);
2702 }
2703 // It's OK to do this in the parser, since DU info is always accurate,
2704 // and the parser always refers to nodes via SafePointNode maps.
2705 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
2706 return this;
2707 }
2708 st = st->in(MemNode::Memory);
2709 }
2710 }
2711
2712
2713 // Capture an unaliased, unconditional, simple store into an initializer.
2714 // Or, if it is independent of the allocation, hoist it above the allocation.
2715 if (ReduceFieldZeroing && /*can_reshape &&*/
2716 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2717 InitializeNode* init = mem->in(0)->as_Initialize();
2718 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2719 if (offset > 0) {
2720 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2721 // If the InitializeNode captured me, it made a raw copy of me,
2722 // and I need to disappear.
2723 if (moved != nullptr) {
2724 // %%% hack to ensure that Ideal returns a new node:
2725 mem = MergeMemNode::make(mem);
2726 return mem; // fold me away
2727 }
2728 }
2729 }
2730
2731 // Fold reinterpret cast into memory operation:
2732 // StoreX mem (MoveY2X v) => StoreY mem v
2733 if (value->is_Move()) {
2734 const Type* vt = value->in(1)->bottom_type();
2735 if (has_reinterpret_variant(vt)) {
2736 if (phase->C->post_loop_opts_phase()) {
2737 return convert_to_reinterpret_store(*phase, value->in(1), vt);
2738 } else {
2739 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
2740 }
2741 }
2742 }
2743
2744 return nullptr; // No further progress
2745 }
2746
2747 //------------------------------Value-----------------------------------------
2748 const Type* StoreNode::Value(PhaseGVN* phase) const {
2749 // Either input is TOP ==> the result is TOP
2750 const Type *t1 = phase->type( in(MemNode::Memory) );
2751 if( t1 == Type::TOP ) return Type::TOP;
2752 const Type *t2 = phase->type( in(MemNode::Address) );
2753 if( t2 == Type::TOP ) return Type::TOP;
2754 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2755 if( t3 == Type::TOP ) return Type::TOP;
2756 return Type::MEMORY;
2757 }
2758
2759 //------------------------------Identity---------------------------------------
2760 // Remove redundant stores:
2761 // Store(m, p, Load(m, p)) changes to m.
2762 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2763 Node* StoreNode::Identity(PhaseGVN* phase) {
2764 Node* mem = in(MemNode::Memory);
2765 Node* adr = in(MemNode::Address);
2766 Node* val = in(MemNode::ValueIn);
2767
2768 Node* result = this;
2769
2770 // Load then Store? Then the Store is useless
2771 if (val->is_Load() &&
2772 val->in(MemNode::Address)->eqv_uncast(adr) &&
2773 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2774 val->as_Load()->store_Opcode() == Opcode()) {
2775 result = mem;
2776 }
2777
2778 // Two stores in a row of the same value?
2779 if (result == this &&
2780 mem->is_Store() &&
2781 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2782 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2783 mem->Opcode() == Opcode()) {
2784 result = mem;
2785 }
2786
2787 // Store of zero anywhere into a freshly-allocated object?
2788 // Then the store is useless.
2789 // (It must already have been captured by the InitializeNode.)
2790 if (result == this &&
2791 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2792 // a newly allocated object is already all-zeroes everywhere
2793 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2794 result = mem;
2795 }
2796
2797 if (result == this) {
2798 // the store may also apply to zero-bits in an earlier object
2799 Node* prev_mem = find_previous_store(phase);
2800 // Steps (a), (b): Walk past independent stores to find an exact match.
2801 if (prev_mem != nullptr) {
2802 Node* prev_val = can_see_stored_value(prev_mem, phase);
2803 if (prev_val != nullptr && prev_val == val) {
2804 // prev_val and val might differ by a cast; it would be good
2805 // to keep the more informative of the two.
2806 result = mem;
2807 }
2808 }
2809 }
2810 }
2811
2812 PhaseIterGVN* igvn = phase->is_IterGVN();
2813 if (result != this && igvn != nullptr) {
2814 MemBarNode* trailing = trailing_membar();
2815 if (trailing != nullptr) {
2816 #ifdef ASSERT
2817 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2818 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects");
2819 #endif
2820 trailing->remove(igvn);
2821 }
2822 }
2823
2824 return result;
2825 }
2826
2827 //------------------------------match_edge-------------------------------------
2828 // Do we Match on this edge index or not? Match only memory & value
2829 uint StoreNode::match_edge(uint idx) const {
2830 return idx == MemNode::Address || idx == MemNode::ValueIn;
2831 }
2832
2833 //------------------------------cmp--------------------------------------------
2834 // Do not common stores up together. They generally have to be split
2835 // back up anyways, so do not bother.
2836 bool StoreNode::cmp( const Node &n ) const {
2837 return (&n == this); // Always fail except on self
2838 }
2839
2840 //------------------------------Ideal_masked_input-----------------------------
2841 // Check for a useless mask before a partial-word store
2842 // (StoreB ... (AndI valIn conIa) )
2843 // If (conIa & mask == mask) this simplifies to
2844 // (StoreB ... (valIn) )
2845 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2846 Node *val = in(MemNode::ValueIn);
2847 if( val->Opcode() == Op_AndI ) {
2848 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2849 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2850 set_req_X(MemNode::ValueIn, val->in(1), phase);
2851 return this;
2852 }
2853 }
2854 return nullptr;
2855 }
2856
2857
2858 //------------------------------Ideal_sign_extended_input----------------------
2859 // Check for useless sign-extension before a partial-word store
2860 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2861 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2862 // (StoreB ... (valIn) )
2863 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2864 Node *val = in(MemNode::ValueIn);
2865 if( val->Opcode() == Op_RShiftI ) {
2866 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2867 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2868 Node *shl = val->in(1);
2869 if( shl->Opcode() == Op_LShiftI ) {
2870 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2871 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2872 set_req_X(MemNode::ValueIn, shl->in(1), phase);
2873 return this;
2874 }
2875 }
2876 }
2877 }
2878 return nullptr;
2879 }
2880
2881 //------------------------------value_never_loaded-----------------------------------
2882 // Determine whether there are any possible loads of the value stored.
2883 // For simplicity, we actually check if there are any loads from the
2884 // address stored to, not just for loads of the value stored by this node.
2885 //
2886 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2887 Node *adr = in(Address);
2888 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2889 if (adr_oop == nullptr)
2890 return false;
2891 if (!adr_oop->is_known_instance_field())
2892 return false; // if not a distinct instance, there may be aliases of the address
2893 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2894 Node *use = adr->fast_out(i);
2895 if (use->is_Load() || use->is_LoadStore()) {
2896 return false;
2897 }
2898 }
2899 return true;
2900 }
2901
2902 MemBarNode* StoreNode::trailing_membar() const {
2903 if (is_release()) {
2904 MemBarNode* trailing_mb = nullptr;
2905 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2906 Node* u = fast_out(i);
2907 if (u->is_MemBar()) {
2908 if (u->as_MemBar()->trailing_store()) {
2909 assert(u->Opcode() == Op_MemBarVolatile, "");
2910 assert(trailing_mb == nullptr, "only one");
2911 trailing_mb = u->as_MemBar();
2912 #ifdef ASSERT
2913 Node* leading = u->as_MemBar()->leading_membar();
2914 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2915 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2916 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2917 #endif
2918 } else {
2919 assert(u->as_MemBar()->standalone(), "");
2920 }
2921 }
2922 }
2923 return trailing_mb;
2924 }
2925 return nullptr;
2926 }
2927
2928
2929 //=============================================================================
2930 //------------------------------Ideal------------------------------------------
2931 // If the store is from an AND mask that leaves the low bits untouched, then
2932 // we can skip the AND operation. If the store is from a sign-extension
2933 // (a left shift, then right shift) we can skip both.
2934 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2935 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2936 if( progress != nullptr ) return progress;
2937
2938 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2939 if( progress != nullptr ) return progress;
2940
2941 // Finally check the default case
2942 return StoreNode::Ideal(phase, can_reshape);
2943 }
2944
2945 //=============================================================================
2946 //------------------------------Ideal------------------------------------------
2947 // If the store is from an AND mask that leaves the low bits untouched, then
2948 // we can skip the AND operation
2949 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2950 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2951 if( progress != nullptr ) return progress;
2952
2953 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2954 if( progress != nullptr ) return progress;
2955
2956 // Finally check the default case
2957 return StoreNode::Ideal(phase, can_reshape);
2958 }
2959
2960 //=============================================================================
2961 //------------------------------Identity---------------------------------------
2962 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2963 // No need to card mark when storing a null ptr
2964 Node* my_store = in(MemNode::OopStore);
2965 if (my_store->is_Store()) {
2966 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2967 if( t1 == TypePtr::NULL_PTR ) {
2968 return in(MemNode::Memory);
2969 }
2970 }
2971 return this;
2972 }
2973
2974 //=============================================================================
2975 //------------------------------Ideal---------------------------------------
2976 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2977 Node* progress = StoreNode::Ideal(phase, can_reshape);
2978 if (progress != nullptr) return progress;
2979
2980 Node* my_store = in(MemNode::OopStore);
2981 if (my_store->is_MergeMem()) {
2982 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2983 set_req_X(MemNode::OopStore, mem, phase);
2984 return this;
2985 }
2986
2987 return nullptr;
2988 }
2989
2990 //------------------------------Value-----------------------------------------
2991 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2992 // Either input is TOP ==> the result is TOP (checked in StoreNode::Value).
2993 // If extra input is TOP ==> the result is TOP
2994 const Type* t = phase->type(in(MemNode::OopStore));
2995 if (t == Type::TOP) {
2996 return Type::TOP;
2997 }
2998 return StoreNode::Value(phase);
2999 }
3000
3001
3002 //=============================================================================
3003 //----------------------------------SCMemProjNode------------------------------
3004 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
3005 {
3006 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
3007 return Type::TOP;
3008 }
3009 return bottom_type();
3010 }
3011
3012 //=============================================================================
3013 //----------------------------------LoadStoreNode------------------------------
3014 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
3015 : Node(required),
3016 _type(rt),
3017 _adr_type(at),
3018 _barrier_data(0)
3019 {
3020 init_req(MemNode::Control, c );
3021 init_req(MemNode::Memory , mem);
3022 init_req(MemNode::Address, adr);
3023 init_req(MemNode::ValueIn, val);
3024 init_class_id(Class_LoadStore);
3025 }
3026
3027 //------------------------------Value-----------------------------------------
3028 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
3029 // Either input is TOP ==> the result is TOP
3030 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
3031 return Type::TOP;
3032 }
3033 const Type* t = phase->type(in(MemNode::Memory));
3034 if (t == Type::TOP) {
3035 return Type::TOP;
3036 }
3037 t = phase->type(in(MemNode::Address));
3038 if (t == Type::TOP) {
3039 return Type::TOP;
3040 }
3041 t = phase->type(in(MemNode::ValueIn));
3042 if (t == Type::TOP) {
3043 return Type::TOP;
3044 }
3045 return bottom_type();
3046 }
3047
3048 uint LoadStoreNode::ideal_reg() const {
3049 return _type->ideal_reg();
3050 }
3051
3052 bool LoadStoreNode::result_not_used() const {
3053 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
3054 Node *x = fast_out(i);
3055 if (x->Opcode() == Op_SCMemProj) continue;
3056 return false;
3057 }
3058 return true;
3059 }
3060
3061 MemBarNode* LoadStoreNode::trailing_membar() const {
3062 MemBarNode* trailing = nullptr;
3063 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3064 Node* u = fast_out(i);
3065 if (u->is_MemBar()) {
3066 if (u->as_MemBar()->trailing_load_store()) {
3067 assert(u->Opcode() == Op_MemBarAcquire, "");
3068 assert(trailing == nullptr, "only one");
3069 trailing = u->as_MemBar();
3070 #ifdef ASSERT
3071 Node* leading = trailing->leading_membar();
3072 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3073 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
3074 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
3075 #endif
3076 } else {
3077 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
3078 }
3079 }
3080 }
3081
3082 return trailing;
3083 }
3084
3085 uint LoadStoreNode::size_of() const { return sizeof(*this); }
3086
3087 //=============================================================================
3088 //----------------------------------LoadStoreConditionalNode--------------------
3089 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) {
3090 init_req(ExpectedIn, ex );
3091 }
3092
3093 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
3094 // Either input is TOP ==> the result is TOP
3095 const Type* t = phase->type(in(ExpectedIn));
3096 if (t == Type::TOP) {
3097 return Type::TOP;
3098 }
3099 return LoadStoreNode::Value(phase);
3100 }
3101
3102 //=============================================================================
3103 //-------------------------------adr_type--------------------------------------
3104 const TypePtr* ClearArrayNode::adr_type() const {
3105 Node *adr = in(3);
3106 if (adr == nullptr) return nullptr; // node is dead
3107 return MemNode::calculate_adr_type(adr->bottom_type());
3108 }
3109
3110 //------------------------------match_edge-------------------------------------
3111 // Do we Match on this edge index or not? Do not match memory
3112 uint ClearArrayNode::match_edge(uint idx) const {
3113 return idx > 1;
3114 }
3115
3116 //------------------------------Identity---------------------------------------
3117 // Clearing a zero length array does nothing
3118 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
3119 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
3120 }
3121
3122 //------------------------------Idealize---------------------------------------
3123 // Clearing a short array is faster with stores
3124 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3125 // Already know this is a large node, do not try to ideal it
3126 if (_is_large) return nullptr;
3127
3128 const int unit = BytesPerLong;
3129 const TypeX* t = phase->type(in(2))->isa_intptr_t();
3130 if (!t) return nullptr;
3131 if (!t->is_con()) return nullptr;
3132 intptr_t raw_count = t->get_con();
3133 intptr_t size = raw_count;
3134 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
3135 // Clearing nothing uses the Identity call.
3136 // Negative clears are possible on dead ClearArrays
3137 // (see jck test stmt114.stmt11402.val).
3138 if (size <= 0 || size % unit != 0) return nullptr;
3139 intptr_t count = size / unit;
3140 // Length too long; communicate this to matchers and assemblers.
3141 // Assemblers are responsible to produce fast hardware clears for it.
3142 if (size > InitArrayShortSize) {
3143 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
3144 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
3145 return nullptr;
3146 }
3147 if (!IdealizeClearArrayNode) return nullptr;
3148 Node *mem = in(1);
3149 if( phase->type(mem)==Type::TOP ) return nullptr;
3150 Node *adr = in(3);
3151 const Type* at = phase->type(adr);
3152 if( at==Type::TOP ) return nullptr;
3153 const TypePtr* atp = at->isa_ptr();
3154 // adjust atp to be the correct array element address type
3155 if (atp == nullptr) atp = TypePtr::BOTTOM;
3156 else atp = atp->add_offset(Type::OffsetBot);
3157 // Get base for derived pointer purposes
3158 if( adr->Opcode() != Op_AddP ) Unimplemented();
3159 Node *base = adr->in(1);
3160
3161 Node *zero = phase->makecon(TypeLong::ZERO);
3162 Node *off = phase->MakeConX(BytesPerLong);
3163 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3164 count--;
3165 while( count-- ) {
3166 mem = phase->transform(mem);
3167 adr = phase->transform(new AddPNode(base,adr,off));
3168 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3169 }
3170 return mem;
3171 }
3172
3173 //----------------------------step_through----------------------------------
3174 // Return allocation input memory edge if it is different instance
3175 // or itself if it is the one we are looking for.
3176 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
3177 Node* n = *np;
3178 assert(n->is_ClearArray(), "sanity");
3179 intptr_t offset;
3180 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
3181 // This method is called only before Allocate nodes are expanded
3182 // during macro nodes expansion. Before that ClearArray nodes are
3183 // only generated in PhaseMacroExpand::generate_arraycopy() (before
3184 // Allocate nodes are expanded) which follows allocations.
3185 assert(alloc != nullptr, "should have allocation");
3186 if (alloc->_idx == instance_id) {
3187 // Can not bypass initialization of the instance we are looking for.
3188 return false;
3189 }
3190 // Otherwise skip it.
3191 InitializeNode* init = alloc->initialization();
3192 if (init != nullptr)
3193 *np = init->in(TypeFunc::Memory);
3194 else
3195 *np = alloc->in(TypeFunc::Memory);
3196 return true;
3197 }
3198
3199 //----------------------------clear_memory-------------------------------------
3200 // Generate code to initialize object storage to zero.
3201 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3202 intptr_t start_offset,
3203 Node* end_offset,
3204 PhaseGVN* phase) {
3205 intptr_t offset = start_offset;
3206
3207 int unit = BytesPerLong;
3208 if ((offset % unit) != 0) {
3209 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
3210 adr = phase->transform(adr);
3211 const TypePtr* atp = TypeRawPtr::BOTTOM;
3212 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3213 mem = phase->transform(mem);
3214 offset += BytesPerInt;
3215 }
3216 assert((offset % unit) == 0, "");
3217
3218 // Initialize the remaining stuff, if any, with a ClearArray.
3219 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
3220 }
3221
3222 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3223 Node* start_offset,
3224 Node* end_offset,
3225 PhaseGVN* phase) {
3226 if (start_offset == end_offset) {
3227 // nothing to do
3228 return mem;
3229 }
3230
3231 int unit = BytesPerLong;
3232 Node* zbase = start_offset;
3233 Node* zend = end_offset;
3234
3235 // Scale to the unit required by the CPU:
3236 if (!Matcher::init_array_count_is_in_bytes) {
3237 Node* shift = phase->intcon(exact_log2(unit));
3238 zbase = phase->transform(new URShiftXNode(zbase, shift) );
3239 zend = phase->transform(new URShiftXNode(zend, shift) );
3240 }
3241
3242 // Bulk clear double-words
3243 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
3244 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
3245 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
3246 return phase->transform(mem);
3247 }
3248
3249 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3250 intptr_t start_offset,
3251 intptr_t end_offset,
3252 PhaseGVN* phase) {
3253 if (start_offset == end_offset) {
3254 // nothing to do
3255 return mem;
3256 }
3257
3258 assert((end_offset % BytesPerInt) == 0, "odd end offset");
3259 intptr_t done_offset = end_offset;
3260 if ((done_offset % BytesPerLong) != 0) {
3261 done_offset -= BytesPerInt;
3262 }
3263 if (done_offset > start_offset) {
3264 mem = clear_memory(ctl, mem, dest,
3265 start_offset, phase->MakeConX(done_offset), phase);
3266 }
3267 if (done_offset < end_offset) { // emit the final 32-bit store
3268 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3269 adr = phase->transform(adr);
3270 const TypePtr* atp = TypeRawPtr::BOTTOM;
3271 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3272 mem = phase->transform(mem);
3273 done_offset += BytesPerInt;
3274 }
3275 assert(done_offset == end_offset, "");
3276 return mem;
3277 }
3278
3279 //=============================================================================
3280 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3281 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
3282 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3283 #ifdef ASSERT
3284 , _pair_idx(0)
3285 #endif
3286 {
3287 init_class_id(Class_MemBar);
3288 Node* top = C->top();
3289 init_req(TypeFunc::I_O,top);
3290 init_req(TypeFunc::FramePtr,top);
3291 init_req(TypeFunc::ReturnAdr,top);
3292 if (precedent != nullptr)
3293 init_req(TypeFunc::Parms, precedent);
3294 }
3295
3296 //------------------------------cmp--------------------------------------------
3297 uint MemBarNode::hash() const { return NO_HASH; }
3298 bool MemBarNode::cmp( const Node &n ) const {
3299 return (&n == this); // Always fail except on self
3300 }
3301
3302 //------------------------------make-------------------------------------------
3303 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3304 switch (opcode) {
3305 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
3306 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
3307 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
3308 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
3309 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
3310 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn);
3311 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3312 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3313 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
3314 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
3315 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
3316 case Op_Initialize: return new InitializeNode(C, atp, pn);
3317 default: ShouldNotReachHere(); return nullptr;
3318 }
3319 }
3320
3321 void MemBarNode::remove(PhaseIterGVN *igvn) {
3322 if (outcnt() != 2) {
3323 assert(Opcode() == Op_Initialize, "Only seen when there are no use of init memory");
3324 assert(outcnt() == 1, "Only control then");
3325 }
3326 if (trailing_store() || trailing_load_store()) {
3327 MemBarNode* leading = leading_membar();
3328 if (leading != nullptr) {
3329 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3330 leading->remove(igvn);
3331 }
3332 }
3333 if (proj_out_or_null(TypeFunc::Memory) != nullptr) {
3334 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3335 }
3336 if (proj_out_or_null(TypeFunc::Control) != nullptr) {
3337 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3338 }
3339 }
3340
3341 //------------------------------Ideal------------------------------------------
3342 // Return a node which is more "ideal" than the current node. Strip out
3343 // control copies
3344 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3345 if (remove_dead_region(phase, can_reshape)) return this;
3346 // Don't bother trying to transform a dead node
3347 if (in(0) && in(0)->is_top()) {
3348 return nullptr;
3349 }
3350
3351 bool progress = false;
3352 // Eliminate volatile MemBars for scalar replaced objects.
3353 if (can_reshape && req() == (Precedent+1)) {
3354 bool eliminate = false;
3355 int opc = Opcode();
3356 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3357 // Volatile field loads and stores.
3358 Node* my_mem = in(MemBarNode::Precedent);
3359 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3360 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3361 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3362 // replace this Precedent (decodeN) with the Load instead.
3363 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3364 Node* load_node = my_mem->in(1);
3365 set_req(MemBarNode::Precedent, load_node);
3366 phase->is_IterGVN()->_worklist.push(my_mem);
3367 my_mem = load_node;
3368 } else {
3369 assert(my_mem->unique_out() == this, "sanity");
3370 del_req(Precedent);
3371 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3372 my_mem = nullptr;
3373 }
3374 progress = true;
3375 }
3376 if (my_mem != nullptr && my_mem->is_Mem()) {
3377 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3378 // Check for scalar replaced object reference.
3379 if( t_oop != nullptr && t_oop->is_known_instance_field() &&
3380 t_oop->offset() != Type::OffsetBot &&
3381 t_oop->offset() != Type::OffsetTop) {
3382 eliminate = true;
3383 }
3384 }
3385 } else if (opc == Op_MemBarRelease) {
3386 // Final field stores.
3387 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3388 if ((alloc != nullptr) && alloc->is_Allocate() &&
3389 alloc->as_Allocate()->does_not_escape_thread()) {
3390 // The allocated object does not escape.
3391 eliminate = true;
3392 }
3393 }
3394 if (eliminate) {
3395 // Replace MemBar projections by its inputs.
3396 PhaseIterGVN* igvn = phase->is_IterGVN();
3397 remove(igvn);
3398 // Must return either the original node (now dead) or a new node
3399 // (Do not return a top here, since that would break the uniqueness of top.)
3400 return new ConINode(TypeInt::ZERO);
3401 }
3402 }
3403 return progress ? this : nullptr;
3404 }
3405
3406 //------------------------------Value------------------------------------------
3407 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3408 if( !in(0) ) return Type::TOP;
3409 if( phase->type(in(0)) == Type::TOP )
3410 return Type::TOP;
3411 return TypeTuple::MEMBAR;
3412 }
3413
3414 //------------------------------match------------------------------------------
3415 // Construct projections for memory.
3416 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3417 switch (proj->_con) {
3418 case TypeFunc::Control:
3419 case TypeFunc::Memory:
3420 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3421 }
3422 ShouldNotReachHere();
3423 return nullptr;
3424 }
3425
3426 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3427 trailing->_kind = TrailingStore;
3428 leading->_kind = LeadingStore;
3429 #ifdef ASSERT
3430 trailing->_pair_idx = leading->_idx;
3431 leading->_pair_idx = leading->_idx;
3432 #endif
3433 }
3434
3435 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3436 trailing->_kind = TrailingLoadStore;
3437 leading->_kind = LeadingLoadStore;
3438 #ifdef ASSERT
3439 trailing->_pair_idx = leading->_idx;
3440 leading->_pair_idx = leading->_idx;
3441 #endif
3442 }
3443
3444 MemBarNode* MemBarNode::trailing_membar() const {
3445 ResourceMark rm;
3446 Node* trailing = (Node*)this;
3447 VectorSet seen;
3448 Node_Stack multis(0);
3449 do {
3450 Node* c = trailing;
3451 uint i = 0;
3452 do {
3453 trailing = nullptr;
3454 for (; i < c->outcnt(); i++) {
3455 Node* next = c->raw_out(i);
3456 if (next != c && next->is_CFG()) {
3457 if (c->is_MultiBranch()) {
3458 if (multis.node() == c) {
3459 multis.set_index(i+1);
3460 } else {
3461 multis.push(c, i+1);
3462 }
3463 }
3464 trailing = next;
3465 break;
3466 }
3467 }
3468 if (trailing != nullptr && !seen.test_set(trailing->_idx)) {
3469 break;
3470 }
3471 while (multis.size() > 0) {
3472 c = multis.node();
3473 i = multis.index();
3474 if (i < c->req()) {
3475 break;
3476 }
3477 multis.pop();
3478 }
3479 } while (multis.size() > 0);
3480 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
3481
3482 MemBarNode* mb = trailing->as_MemBar();
3483 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3484 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3485 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3486 return mb;
3487 }
3488
3489 MemBarNode* MemBarNode::leading_membar() const {
3490 ResourceMark rm;
3491 VectorSet seen;
3492 Node_Stack regions(0);
3493 Node* leading = in(0);
3494 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3495 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) {
3496 leading = nullptr;
3497 while (regions.size() > 0 && leading == nullptr) {
3498 Node* r = regions.node();
3499 uint i = regions.index();
3500 if (i < r->req()) {
3501 leading = r->in(i);
3502 regions.set_index(i+1);
3503 } else {
3504 regions.pop();
3505 }
3506 }
3507 if (leading == nullptr) {
3508 assert(regions.size() == 0, "all paths should have been tried");
3509 return nullptr;
3510 }
3511 }
3512 if (leading->is_Region()) {
3513 regions.push(leading, 2);
3514 leading = leading->in(1);
3515 } else {
3516 leading = leading->in(0);
3517 }
3518 }
3519 #ifdef ASSERT
3520 Unique_Node_List wq;
3521 wq.push((Node*)this);
3522 uint found = 0;
3523 for (uint i = 0; i < wq.size(); i++) {
3524 Node* n = wq.at(i);
3525 if (n->is_Region()) {
3526 for (uint j = 1; j < n->req(); j++) {
3527 Node* in = n->in(j);
3528 if (in != nullptr && !in->is_top()) {
3529 wq.push(in);
3530 }
3531 }
3532 } else {
3533 if (n->is_MemBar() && n->as_MemBar()->leading()) {
3534 assert(n == leading, "consistency check failed");
3535 found++;
3536 } else {
3537 Node* in = n->in(0);
3538 if (in != nullptr && !in->is_top()) {
3539 wq.push(in);
3540 }
3541 }
3542 }
3543 }
3544 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed");
3545 #endif
3546 if (leading == nullptr) {
3547 return nullptr;
3548 }
3549 MemBarNode* mb = leading->as_MemBar();
3550 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3551 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3552 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3553 return mb;
3554 }
3555
3556
3557 //===========================InitializeNode====================================
3558 // SUMMARY:
3559 // This node acts as a memory barrier on raw memory, after some raw stores.
3560 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3561 // The Initialize can 'capture' suitably constrained stores as raw inits.
3562 // It can coalesce related raw stores into larger units (called 'tiles').
3563 // It can avoid zeroing new storage for memory units which have raw inits.
3564 // At macro-expansion, it is marked 'complete', and does not optimize further.
3565 //
3566 // EXAMPLE:
3567 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3568 // ctl = incoming control; mem* = incoming memory
3569 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3570 // First allocate uninitialized memory and fill in the header:
3571 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3572 // ctl := alloc.Control; mem* := alloc.Memory*
3573 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3574 // Then initialize to zero the non-header parts of the raw memory block:
3575 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3576 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3577 // After the initialize node executes, the object is ready for service:
3578 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3579 // Suppose its body is immediately initialized as {1,2}:
3580 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3581 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3582 // mem.SLICE(#short[*]) := store2
3583 //
3584 // DETAILS:
3585 // An InitializeNode collects and isolates object initialization after
3586 // an AllocateNode and before the next possible safepoint. As a
3587 // memory barrier (MemBarNode), it keeps critical stores from drifting
3588 // down past any safepoint or any publication of the allocation.
3589 // Before this barrier, a newly-allocated object may have uninitialized bits.
3590 // After this barrier, it may be treated as a real oop, and GC is allowed.
3591 //
3592 // The semantics of the InitializeNode include an implicit zeroing of
3593 // the new object from object header to the end of the object.
3594 // (The object header and end are determined by the AllocateNode.)
3595 //
3596 // Certain stores may be added as direct inputs to the InitializeNode.
3597 // These stores must update raw memory, and they must be to addresses
3598 // derived from the raw address produced by AllocateNode, and with
3599 // a constant offset. They must be ordered by increasing offset.
3600 // The first one is at in(RawStores), the last at in(req()-1).
3601 // Unlike most memory operations, they are not linked in a chain,
3602 // but are displayed in parallel as users of the rawmem output of
3603 // the allocation.
3604 //
3605 // (See comments in InitializeNode::capture_store, which continue
3606 // the example given above.)
3607 //
3608 // When the associated Allocate is macro-expanded, the InitializeNode
3609 // may be rewritten to optimize collected stores. A ClearArrayNode
3610 // may also be created at that point to represent any required zeroing.
3611 // The InitializeNode is then marked 'complete', prohibiting further
3612 // capturing of nearby memory operations.
3613 //
3614 // During macro-expansion, all captured initializations which store
3615 // constant values of 32 bits or smaller are coalesced (if advantageous)
3616 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3617 // initialized in fewer memory operations. Memory words which are
3618 // covered by neither tiles nor non-constant stores are pre-zeroed
3619 // by explicit stores of zero. (The code shape happens to do all
3620 // zeroing first, then all other stores, with both sequences occurring
3621 // in order of ascending offsets.)
3622 //
3623 // Alternatively, code may be inserted between an AllocateNode and its
3624 // InitializeNode, to perform arbitrary initialization of the new object.
3625 // E.g., the object copying intrinsics insert complex data transfers here.
3626 // The initialization must then be marked as 'complete' disable the
3627 // built-in zeroing semantics and the collection of initializing stores.
3628 //
3629 // While an InitializeNode is incomplete, reads from the memory state
3630 // produced by it are optimizable if they match the control edge and
3631 // new oop address associated with the allocation/initialization.
3632 // They return a stored value (if the offset matches) or else zero.
3633 // A write to the memory state, if it matches control and address,
3634 // and if it is to a constant offset, may be 'captured' by the
3635 // InitializeNode. It is cloned as a raw memory operation and rewired
3636 // inside the initialization, to the raw oop produced by the allocation.
3637 // Operations on addresses which are provably distinct (e.g., to
3638 // other AllocateNodes) are allowed to bypass the initialization.
3639 //
3640 // The effect of all this is to consolidate object initialization
3641 // (both arrays and non-arrays, both piecewise and bulk) into a
3642 // single location, where it can be optimized as a unit.
3643 //
3644 // Only stores with an offset less than TrackedInitializationLimit words
3645 // will be considered for capture by an InitializeNode. This puts a
3646 // reasonable limit on the complexity of optimized initializations.
3647
3648 //---------------------------InitializeNode------------------------------------
3649 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3650 : MemBarNode(C, adr_type, rawoop),
3651 _is_complete(Incomplete), _does_not_escape(false)
3652 {
3653 init_class_id(Class_Initialize);
3654
3655 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3656 assert(in(RawAddress) == rawoop, "proper init");
3657 // Note: allocation() can be null, for secondary initialization barriers
3658 }
3659
3660 // Since this node is not matched, it will be processed by the
3661 // register allocator. Declare that there are no constraints
3662 // on the allocation of the RawAddress edge.
3663 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3664 // This edge should be set to top, by the set_complete. But be conservative.
3665 if (idx == InitializeNode::RawAddress)
3666 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3667 return RegMask::Empty;
3668 }
3669
3670 Node* InitializeNode::memory(uint alias_idx) {
3671 Node* mem = in(Memory);
3672 if (mem->is_MergeMem()) {
3673 return mem->as_MergeMem()->memory_at(alias_idx);
3674 } else {
3675 // incoming raw memory is not split
3676 return mem;
3677 }
3678 }
3679
3680 bool InitializeNode::is_non_zero() {
3681 if (is_complete()) return false;
3682 remove_extra_zeroes();
3683 return (req() > RawStores);
3684 }
3685
3686 void InitializeNode::set_complete(PhaseGVN* phase) {
3687 assert(!is_complete(), "caller responsibility");
3688 _is_complete = Complete;
3689
3690 // After this node is complete, it contains a bunch of
3691 // raw-memory initializations. There is no need for
3692 // it to have anything to do with non-raw memory effects.
3693 // Therefore, tell all non-raw users to re-optimize themselves,
3694 // after skipping the memory effects of this initialization.
3695 PhaseIterGVN* igvn = phase->is_IterGVN();
3696 if (igvn) igvn->add_users_to_worklist(this);
3697 }
3698
3699 // convenience function
3700 // return false if the init contains any stores already
3701 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3702 InitializeNode* init = initialization();
3703 if (init == nullptr || init->is_complete()) return false;
3704 init->remove_extra_zeroes();
3705 // for now, if this allocation has already collected any inits, bail:
3706 if (init->is_non_zero()) return false;
3707 init->set_complete(phase);
3708 return true;
3709 }
3710
3711 void InitializeNode::remove_extra_zeroes() {
3712 if (req() == RawStores) return;
3713 Node* zmem = zero_memory();
3714 uint fill = RawStores;
3715 for (uint i = fill; i < req(); i++) {
3716 Node* n = in(i);
3717 if (n->is_top() || n == zmem) continue; // skip
3718 if (fill < i) set_req(fill, n); // compact
3719 ++fill;
3720 }
3721 // delete any empty spaces created:
3722 while (fill < req()) {
3723 del_req(fill);
3724 }
3725 }
3726
3727 // Helper for remembering which stores go with which offsets.
3728 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3729 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3730 intptr_t offset = -1;
3731 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3732 phase, offset);
3733 if (base == nullptr) return -1; // something is dead,
3734 if (offset < 0) return -1; // dead, dead
3735 return offset;
3736 }
3737
3738 // Helper for proving that an initialization expression is
3739 // "simple enough" to be folded into an object initialization.
3740 // Attempts to prove that a store's initial value 'n' can be captured
3741 // within the initialization without creating a vicious cycle, such as:
3742 // { Foo p = new Foo(); p.next = p; }
3743 // True for constants and parameters and small combinations thereof.
3744 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
3745 ResourceMark rm;
3746 Unique_Node_List worklist;
3747 worklist.push(value);
3748
3749 uint complexity_limit = 20;
3750 for (uint j = 0; j < worklist.size(); j++) {
3751 if (j >= complexity_limit) {
3752 return false; // Bail out if processed too many nodes
3753 }
3754
3755 Node* n = worklist.at(j);
3756 if (n == nullptr) continue; // (can this really happen?)
3757 if (n->is_Proj()) n = n->in(0);
3758 if (n == this) return false; // found a cycle
3759 if (n->is_Con()) continue;
3760 if (n->is_Start()) continue; // params, etc., are OK
3761 if (n->is_Root()) continue; // even better
3762
3763 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
3764 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
3765 continue;
3766 }
3767
3768 Node* ctl = n->in(0);
3769 if (ctl != nullptr && !ctl->is_top()) {
3770 if (ctl->is_Proj()) ctl = ctl->in(0);
3771 if (ctl == this) return false;
3772
3773 // If we already know that the enclosing memory op is pinned right after
3774 // the init, then any control flow that the store has picked up
3775 // must have preceded the init, or else be equal to the init.
3776 // Even after loop optimizations (which might change control edges)
3777 // a store is never pinned *before* the availability of its inputs.
3778 if (!MemNode::all_controls_dominate(n, this))
3779 return false; // failed to prove a good control
3780 }
3781
3782 // Check data edges for possible dependencies on 'this'.
3783 for (uint i = 1; i < n->req(); i++) {
3784 Node* m = n->in(i);
3785 if (m == nullptr || m == n || m->is_top()) continue;
3786
3787 // Only process data inputs once
3788 worklist.push(m);
3789 }
3790 }
3791
3792 return true;
3793 }
3794
3795 // Here are all the checks a Store must pass before it can be moved into
3796 // an initialization. Returns zero if a check fails.
3797 // On success, returns the (constant) offset to which the store applies,
3798 // within the initialized memory.
3799 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
3800 const int FAIL = 0;
3801 if (st->req() != MemNode::ValueIn + 1)
3802 return FAIL; // an inscrutable StoreNode (card mark?)
3803 Node* ctl = st->in(MemNode::Control);
3804 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this))
3805 return FAIL; // must be unconditional after the initialization
3806 Node* mem = st->in(MemNode::Memory);
3807 if (!(mem->is_Proj() && mem->in(0) == this))
3808 return FAIL; // must not be preceded by other stores
3809 Node* adr = st->in(MemNode::Address);
3810 intptr_t offset;
3811 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3812 if (alloc == nullptr)
3813 return FAIL; // inscrutable address
3814 if (alloc != allocation())
3815 return FAIL; // wrong allocation! (store needs to float up)
3816 int size_in_bytes = st->memory_size();
3817 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
3818 return FAIL; // mismatched access
3819 }
3820 Node* val = st->in(MemNode::ValueIn);
3821
3822 if (!detect_init_independence(val, phase))
3823 return FAIL; // stored value must be 'simple enough'
3824
3825 // The Store can be captured only if nothing after the allocation
3826 // and before the Store is using the memory location that the store
3827 // overwrites.
3828 bool failed = false;
3829 // If is_complete_with_arraycopy() is true the shape of the graph is
3830 // well defined and is safe so no need for extra checks.
3831 if (!is_complete_with_arraycopy()) {
3832 // We are going to look at each use of the memory state following
3833 // the allocation to make sure nothing reads the memory that the
3834 // Store writes.
3835 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3836 int alias_idx = phase->C->get_alias_index(t_adr);
3837 ResourceMark rm;
3838 Unique_Node_List mems;
3839 mems.push(mem);
3840 Node* unique_merge = nullptr;
3841 for (uint next = 0; next < mems.size(); ++next) {
3842 Node *m = mems.at(next);
3843 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3844 Node *n = m->fast_out(j);
3845 if (n->outcnt() == 0) {
3846 continue;
3847 }
3848 if (n == st) {
3849 continue;
3850 } else if (n->in(0) != nullptr && n->in(0) != ctl) {
3851 // If the control of this use is different from the control
3852 // of the Store which is right after the InitializeNode then
3853 // this node cannot be between the InitializeNode and the
3854 // Store.
3855 continue;
3856 } else if (n->is_MergeMem()) {
3857 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3858 // We can hit a MergeMemNode (that will likely go away
3859 // later) that is a direct use of the memory state
3860 // following the InitializeNode on the same slice as the
3861 // store node that we'd like to capture. We need to check
3862 // the uses of the MergeMemNode.
3863 mems.push(n);
3864 }
3865 } else if (n->is_Mem()) {
3866 Node* other_adr = n->in(MemNode::Address);
3867 if (other_adr == adr) {
3868 failed = true;
3869 break;
3870 } else {
3871 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3872 if (other_t_adr != nullptr) {
3873 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3874 if (other_alias_idx == alias_idx) {
3875 // A load from the same memory slice as the store right
3876 // after the InitializeNode. We check the control of the
3877 // object/array that is loaded from. If it's the same as
3878 // the store control then we cannot capture the store.
3879 assert(!n->is_Store(), "2 stores to same slice on same control?");
3880 Node* base = other_adr;
3881 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3882 base = base->in(AddPNode::Base);
3883 if (base != nullptr) {
3884 base = base->uncast();
3885 if (base->is_Proj() && base->in(0) == alloc) {
3886 failed = true;
3887 break;
3888 }
3889 }
3890 }
3891 }
3892 }
3893 } else {
3894 failed = true;
3895 break;
3896 }
3897 }
3898 }
3899 }
3900 if (failed) {
3901 if (!can_reshape) {
3902 // We decided we couldn't capture the store during parsing. We
3903 // should try again during the next IGVN once the graph is
3904 // cleaner.
3905 phase->C->record_for_igvn(st);
3906 }
3907 return FAIL;
3908 }
3909
3910 return offset; // success
3911 }
3912
3913 // Find the captured store in(i) which corresponds to the range
3914 // [start..start+size) in the initialized object.
3915 // If there is one, return its index i. If there isn't, return the
3916 // negative of the index where it should be inserted.
3917 // Return 0 if the queried range overlaps an initialization boundary
3918 // or if dead code is encountered.
3919 // If size_in_bytes is zero, do not bother with overlap checks.
3920 int InitializeNode::captured_store_insertion_point(intptr_t start,
3921 int size_in_bytes,
3922 PhaseTransform* phase) {
3923 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
3924
3925 if (is_complete())
3926 return FAIL; // arraycopy got here first; punt
3927
3928 assert(allocation() != nullptr, "must be present");
3929
3930 // no negatives, no header fields:
3931 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3932
3933 // after a certain size, we bail out on tracking all the stores:
3934 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3935 if (start >= ti_limit) return FAIL;
3936
3937 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3938 if (i >= limit) return -(int)i; // not found; here is where to put it
3939
3940 Node* st = in(i);
3941 intptr_t st_off = get_store_offset(st, phase);
3942 if (st_off < 0) {
3943 if (st != zero_memory()) {
3944 return FAIL; // bail out if there is dead garbage
3945 }
3946 } else if (st_off > start) {
3947 // ...we are done, since stores are ordered
3948 if (st_off < start + size_in_bytes) {
3949 return FAIL; // the next store overlaps
3950 }
3951 return -(int)i; // not found; here is where to put it
3952 } else if (st_off < start) {
3953 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
3954 if (size_in_bytes != 0 &&
3955 start < st_off + MAX_STORE &&
3956 start < st_off + st->as_Store()->memory_size()) {
3957 return FAIL; // the previous store overlaps
3958 }
3959 } else {
3960 if (size_in_bytes != 0 &&
3961 st->as_Store()->memory_size() != size_in_bytes) {
3962 return FAIL; // mismatched store size
3963 }
3964 return i;
3965 }
3966
3967 ++i;
3968 }
3969 }
3970
3971 // Look for a captured store which initializes at the offset 'start'
3972 // with the given size. If there is no such store, and no other
3973 // initialization interferes, then return zero_memory (the memory
3974 // projection of the AllocateNode).
3975 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3976 PhaseTransform* phase) {
3977 assert(stores_are_sane(phase), "");
3978 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3979 if (i == 0) {
3980 return nullptr; // something is dead
3981 } else if (i < 0) {
3982 return zero_memory(); // just primordial zero bits here
3983 } else {
3984 Node* st = in(i); // here is the store at this position
3985 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3986 return st;
3987 }
3988 }
3989
3990 // Create, as a raw pointer, an address within my new object at 'offset'.
3991 Node* InitializeNode::make_raw_address(intptr_t offset,
3992 PhaseTransform* phase) {
3993 Node* addr = in(RawAddress);
3994 if (offset != 0) {
3995 Compile* C = phase->C;
3996 addr = phase->transform( new AddPNode(C->top(), addr,
3997 phase->MakeConX(offset)) );
3998 }
3999 return addr;
4000 }
4001
4002 // Clone the given store, converting it into a raw store
4003 // initializing a field or element of my new object.
4004 // Caller is responsible for retiring the original store,
4005 // with subsume_node or the like.
4006 //
4007 // From the example above InitializeNode::InitializeNode,
4008 // here are the old stores to be captured:
4009 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4010 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4011 //
4012 // Here is the changed code; note the extra edges on init:
4013 // alloc = (Allocate ...)
4014 // rawoop = alloc.RawAddress
4015 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
4016 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
4017 // init = (Initialize alloc.Control alloc.Memory rawoop
4018 // rawstore1 rawstore2)
4019 //
4020 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
4021 PhaseGVN* phase, bool can_reshape) {
4022 assert(stores_are_sane(phase), "");
4023
4024 if (start < 0) return nullptr;
4025 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
4026
4027 Compile* C = phase->C;
4028 int size_in_bytes = st->memory_size();
4029 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4030 if (i == 0) return nullptr; // bail out
4031 Node* prev_mem = nullptr; // raw memory for the captured store
4032 if (i > 0) {
4033 prev_mem = in(i); // there is a pre-existing store under this one
4034 set_req(i, C->top()); // temporarily disconnect it
4035 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
4036 } else {
4037 i = -i; // no pre-existing store
4038 prev_mem = zero_memory(); // a slice of the newly allocated object
4039 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
4040 set_req(--i, C->top()); // reuse this edge; it has been folded away
4041 else
4042 ins_req(i, C->top()); // build a new edge
4043 }
4044 Node* new_st = st->clone();
4045 new_st->set_req(MemNode::Control, in(Control));
4046 new_st->set_req(MemNode::Memory, prev_mem);
4047 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
4048 new_st = phase->transform(new_st);
4049
4050 // At this point, new_st might have swallowed a pre-existing store
4051 // at the same offset, or perhaps new_st might have disappeared,
4052 // if it redundantly stored the same value (or zero to fresh memory).
4053
4054 // In any case, wire it in:
4055 PhaseIterGVN* igvn = phase->is_IterGVN();
4056 if (igvn) {
4057 igvn->rehash_node_delayed(this);
4058 }
4059 set_req(i, new_st);
4060
4061 // The caller may now kill the old guy.
4062 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
4063 assert(check_st == new_st || check_st == nullptr, "must be findable");
4064 assert(!is_complete(), "");
4065 return new_st;
4066 }
4067
4068 static bool store_constant(jlong* tiles, int num_tiles,
4069 intptr_t st_off, int st_size,
4070 jlong con) {
4071 if ((st_off & (st_size-1)) != 0)
4072 return false; // strange store offset (assume size==2**N)
4073 address addr = (address)tiles + st_off;
4074 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
4075 switch (st_size) {
4076 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
4077 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
4078 case sizeof(jint): *(jint*) addr = (jint) con; break;
4079 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
4080 default: return false; // strange store size (detect size!=2**N here)
4081 }
4082 return true; // return success to caller
4083 }
4084
4085 // Coalesce subword constants into int constants and possibly
4086 // into long constants. The goal, if the CPU permits,
4087 // is to initialize the object with a small number of 64-bit tiles.
4088 // Also, convert floating-point constants to bit patterns.
4089 // Non-constants are not relevant to this pass.
4090 //
4091 // In terms of the running example on InitializeNode::InitializeNode
4092 // and InitializeNode::capture_store, here is the transformation
4093 // of rawstore1 and rawstore2 into rawstore12:
4094 // alloc = (Allocate ...)
4095 // rawoop = alloc.RawAddress
4096 // tile12 = 0x00010002
4097 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
4098 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
4099 //
4100 void
4101 InitializeNode::coalesce_subword_stores(intptr_t header_size,
4102 Node* size_in_bytes,
4103 PhaseGVN* phase) {
4104 Compile* C = phase->C;
4105
4106 assert(stores_are_sane(phase), "");
4107 // Note: After this pass, they are not completely sane,
4108 // since there may be some overlaps.
4109
4110 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
4111
4112 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4113 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
4114 size_limit = MIN2(size_limit, ti_limit);
4115 size_limit = align_up(size_limit, BytesPerLong);
4116 int num_tiles = size_limit / BytesPerLong;
4117
4118 // allocate space for the tile map:
4119 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
4120 jlong tiles_buf[small_len];
4121 Node* nodes_buf[small_len];
4122 jlong inits_buf[small_len];
4123 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
4124 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4125 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
4126 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
4127 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
4128 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4129 // tiles: exact bitwise model of all primitive constants
4130 // nodes: last constant-storing node subsumed into the tiles model
4131 // inits: which bytes (in each tile) are touched by any initializations
4132
4133 //// Pass A: Fill in the tile model with any relevant stores.
4134
4135 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
4136 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
4137 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
4138 Node* zmem = zero_memory(); // initially zero memory state
4139 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4140 Node* st = in(i);
4141 intptr_t st_off = get_store_offset(st, phase);
4142
4143 // Figure out the store's offset and constant value:
4144 if (st_off < header_size) continue; //skip (ignore header)
4145 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
4146 int st_size = st->as_Store()->memory_size();
4147 if (st_off + st_size > size_limit) break;
4148
4149 // Record which bytes are touched, whether by constant or not.
4150 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
4151 continue; // skip (strange store size)
4152
4153 const Type* val = phase->type(st->in(MemNode::ValueIn));
4154 if (!val->singleton()) continue; //skip (non-con store)
4155 BasicType type = val->basic_type();
4156
4157 jlong con = 0;
4158 switch (type) {
4159 case T_INT: con = val->is_int()->get_con(); break;
4160 case T_LONG: con = val->is_long()->get_con(); break;
4161 case T_FLOAT: con = jint_cast(val->getf()); break;
4162 case T_DOUBLE: con = jlong_cast(val->getd()); break;
4163 default: continue; //skip (odd store type)
4164 }
4165
4166 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
4167 st->Opcode() == Op_StoreL) {
4168 continue; // This StoreL is already optimal.
4169 }
4170
4171 // Store down the constant.
4172 store_constant(tiles, num_tiles, st_off, st_size, con);
4173
4174 intptr_t j = st_off >> LogBytesPerLong;
4175
4176 if (type == T_INT && st_size == BytesPerInt
4177 && (st_off & BytesPerInt) == BytesPerInt) {
4178 jlong lcon = tiles[j];
4179 if (!Matcher::isSimpleConstant64(lcon) &&
4180 st->Opcode() == Op_StoreI) {
4181 // This StoreI is already optimal by itself.
4182 jint* intcon = (jint*) &tiles[j];
4183 intcon[1] = 0; // undo the store_constant()
4184
4185 // If the previous store is also optimal by itself, back up and
4186 // undo the action of the previous loop iteration... if we can.
4187 // But if we can't, just let the previous half take care of itself.
4188 st = nodes[j];
4189 st_off -= BytesPerInt;
4190 con = intcon[0];
4191 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) {
4192 assert(st_off >= header_size, "still ignoring header");
4193 assert(get_store_offset(st, phase) == st_off, "must be");
4194 assert(in(i-1) == zmem, "must be");
4195 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
4196 assert(con == tcon->is_int()->get_con(), "must be");
4197 // Undo the effects of the previous loop trip, which swallowed st:
4198 intcon[0] = 0; // undo store_constant()
4199 set_req(i-1, st); // undo set_req(i, zmem)
4200 nodes[j] = nullptr; // undo nodes[j] = st
4201 --old_subword; // undo ++old_subword
4202 }
4203 continue; // This StoreI is already optimal.
4204 }
4205 }
4206
4207 // This store is not needed.
4208 set_req(i, zmem);
4209 nodes[j] = st; // record for the moment
4210 if (st_size < BytesPerLong) // something has changed
4211 ++old_subword; // includes int/float, but who's counting...
4212 else ++old_long;
4213 }
4214
4215 if ((old_subword + old_long) == 0)
4216 return; // nothing more to do
4217
4218 //// Pass B: Convert any non-zero tiles into optimal constant stores.
4219 // Be sure to insert them before overlapping non-constant stores.
4220 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
4221 for (int j = 0; j < num_tiles; j++) {
4222 jlong con = tiles[j];
4223 jlong init = inits[j];
4224 if (con == 0) continue;
4225 jint con0, con1; // split the constant, address-wise
4226 jint init0, init1; // split the init map, address-wise
4227 { union { jlong con; jint intcon[2]; } u;
4228 u.con = con;
4229 con0 = u.intcon[0];
4230 con1 = u.intcon[1];
4231 u.con = init;
4232 init0 = u.intcon[0];
4233 init1 = u.intcon[1];
4234 }
4235
4236 Node* old = nodes[j];
4237 assert(old != nullptr, "need the prior store");
4238 intptr_t offset = (j * BytesPerLong);
4239
4240 bool split = !Matcher::isSimpleConstant64(con);
4241
4242 if (offset < header_size) {
4243 assert(offset + BytesPerInt >= header_size, "second int counts");
4244 assert(*(jint*)&tiles[j] == 0, "junk in header");
4245 split = true; // only the second word counts
4246 // Example: int a[] = { 42 ... }
4247 } else if (con0 == 0 && init0 == -1) {
4248 split = true; // first word is covered by full inits
4249 // Example: int a[] = { ... foo(), 42 ... }
4250 } else if (con1 == 0 && init1 == -1) {
4251 split = true; // second word is covered by full inits
4252 // Example: int a[] = { ... 42, foo() ... }
4253 }
4254
4255 // Here's a case where init0 is neither 0 nor -1:
4256 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
4257 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
4258 // In this case the tile is not split; it is (jlong)42.
4259 // The big tile is stored down, and then the foo() value is inserted.
4260 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
4261
4262 Node* ctl = old->in(MemNode::Control);
4263 Node* adr = make_raw_address(offset, phase);
4264 const TypePtr* atp = TypeRawPtr::BOTTOM;
4265
4266 // One or two coalesced stores to plop down.
4267 Node* st[2];
4268 intptr_t off[2];
4269 int nst = 0;
4270 if (!split) {
4271 ++new_long;
4272 off[nst] = offset;
4273 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4274 phase->longcon(con), T_LONG, MemNode::unordered);
4275 } else {
4276 // Omit either if it is a zero.
4277 if (con0 != 0) {
4278 ++new_int;
4279 off[nst] = offset;
4280 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4281 phase->intcon(con0), T_INT, MemNode::unordered);
4282 }
4283 if (con1 != 0) {
4284 ++new_int;
4285 offset += BytesPerInt;
4286 adr = make_raw_address(offset, phase);
4287 off[nst] = offset;
4288 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4289 phase->intcon(con1), T_INT, MemNode::unordered);
4290 }
4291 }
4292
4293 // Insert second store first, then the first before the second.
4294 // Insert each one just before any overlapping non-constant stores.
4295 while (nst > 0) {
4296 Node* st1 = st[--nst];
4297 C->copy_node_notes_to(st1, old);
4298 st1 = phase->transform(st1);
4299 offset = off[nst];
4300 assert(offset >= header_size, "do not smash header");
4301 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4302 guarantee(ins_idx != 0, "must re-insert constant store");
4303 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
4304 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4305 set_req(--ins_idx, st1);
4306 else
4307 ins_req(ins_idx, st1);
4308 }
4309 }
4310
4311 if (PrintCompilation && WizardMode)
4312 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4313 old_subword, old_long, new_int, new_long);
4314 if (C->log() != nullptr)
4315 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4316 old_subword, old_long, new_int, new_long);
4317
4318 // Clean up any remaining occurrences of zmem:
4319 remove_extra_zeroes();
4320 }
4321
4322 // Explore forward from in(start) to find the first fully initialized
4323 // word, and return its offset. Skip groups of subword stores which
4324 // together initialize full words. If in(start) is itself part of a
4325 // fully initialized word, return the offset of in(start). If there
4326 // are no following full-word stores, or if something is fishy, return
4327 // a negative value.
4328 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4329 int int_map = 0;
4330 intptr_t int_map_off = 0;
4331 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
4332
4333 for (uint i = start, limit = req(); i < limit; i++) {
4334 Node* st = in(i);
4335
4336 intptr_t st_off = get_store_offset(st, phase);
4337 if (st_off < 0) break; // return conservative answer
4338
4339 int st_size = st->as_Store()->memory_size();
4340 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4341 return st_off; // we found a complete word init
4342 }
4343
4344 // update the map:
4345
4346 intptr_t this_int_off = align_down(st_off, BytesPerInt);
4347 if (this_int_off != int_map_off) {
4348 // reset the map:
4349 int_map = 0;
4350 int_map_off = this_int_off;
4351 }
4352
4353 int subword_off = st_off - this_int_off;
4354 int_map |= right_n_bits(st_size) << subword_off;
4355 if ((int_map & FULL_MAP) == FULL_MAP) {
4356 return this_int_off; // we found a complete word init
4357 }
4358
4359 // Did this store hit or cross the word boundary?
4360 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4361 if (next_int_off == this_int_off + BytesPerInt) {
4362 // We passed the current int, without fully initializing it.
4363 int_map_off = next_int_off;
4364 int_map >>= BytesPerInt;
4365 } else if (next_int_off > this_int_off + BytesPerInt) {
4366 // We passed the current and next int.
4367 return this_int_off + BytesPerInt;
4368 }
4369 }
4370
4371 return -1;
4372 }
4373
4374
4375 // Called when the associated AllocateNode is expanded into CFG.
4376 // At this point, we may perform additional optimizations.
4377 // Linearize the stores by ascending offset, to make memory
4378 // activity as coherent as possible.
4379 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4380 intptr_t header_size,
4381 Node* size_in_bytes,
4382 PhaseIterGVN* phase) {
4383 assert(!is_complete(), "not already complete");
4384 assert(stores_are_sane(phase), "");
4385 assert(allocation() != nullptr, "must be present");
4386
4387 remove_extra_zeroes();
4388
4389 if (ReduceFieldZeroing || ReduceBulkZeroing)
4390 // reduce instruction count for common initialization patterns
4391 coalesce_subword_stores(header_size, size_in_bytes, phase);
4392
4393 Node* zmem = zero_memory(); // initially zero memory state
4394 Node* inits = zmem; // accumulating a linearized chain of inits
4395 #ifdef ASSERT
4396 intptr_t first_offset = allocation()->minimum_header_size();
4397 intptr_t last_init_off = first_offset; // previous init offset
4398 intptr_t last_init_end = first_offset; // previous init offset+size
4399 intptr_t last_tile_end = first_offset; // previous tile offset+size
4400 #endif
4401 intptr_t zeroes_done = header_size;
4402
4403 bool do_zeroing = true; // we might give up if inits are very sparse
4404 int big_init_gaps = 0; // how many large gaps have we seen?
4405
4406 if (UseTLAB && ZeroTLAB) do_zeroing = false;
4407 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
4408
4409 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4410 Node* st = in(i);
4411 intptr_t st_off = get_store_offset(st, phase);
4412 if (st_off < 0)
4413 break; // unknown junk in the inits
4414 if (st->in(MemNode::Memory) != zmem)
4415 break; // complicated store chains somehow in list
4416
4417 int st_size = st->as_Store()->memory_size();
4418 intptr_t next_init_off = st_off + st_size;
4419
4420 if (do_zeroing && zeroes_done < next_init_off) {
4421 // See if this store needs a zero before it or under it.
4422 intptr_t zeroes_needed = st_off;
4423
4424 if (st_size < BytesPerInt) {
4425 // Look for subword stores which only partially initialize words.
4426 // If we find some, we must lay down some word-level zeroes first,
4427 // underneath the subword stores.
4428 //
4429 // Examples:
4430 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
4431 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
4432 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
4433 //
4434 // Note: coalesce_subword_stores may have already done this,
4435 // if it was prompted by constant non-zero subword initializers.
4436 // But this case can still arise with non-constant stores.
4437
4438 intptr_t next_full_store = find_next_fullword_store(i, phase);
4439
4440 // In the examples above:
4441 // in(i) p q r s x y z
4442 // st_off 12 13 14 15 12 13 14
4443 // st_size 1 1 1 1 1 1 1
4444 // next_full_s. 12 16 16 16 16 16 16
4445 // z's_done 12 16 16 16 12 16 12
4446 // z's_needed 12 16 16 16 16 16 16
4447 // zsize 0 0 0 0 4 0 4
4448 if (next_full_store < 0) {
4449 // Conservative tack: Zero to end of current word.
4450 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4451 } else {
4452 // Zero to beginning of next fully initialized word.
4453 // Or, don't zero at all, if we are already in that word.
4454 assert(next_full_store >= zeroes_needed, "must go forward");
4455 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4456 zeroes_needed = next_full_store;
4457 }
4458 }
4459
4460 if (zeroes_needed > zeroes_done) {
4461 intptr_t zsize = zeroes_needed - zeroes_done;
4462 // Do some incremental zeroing on rawmem, in parallel with inits.
4463 zeroes_done = align_down(zeroes_done, BytesPerInt);
4464 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4465 zeroes_done, zeroes_needed,
4466 phase);
4467 zeroes_done = zeroes_needed;
4468 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4469 do_zeroing = false; // leave the hole, next time
4470 }
4471 }
4472
4473 // Collect the store and move on:
4474 phase->replace_input_of(st, MemNode::Memory, inits);
4475 inits = st; // put it on the linearized chain
4476 set_req(i, zmem); // unhook from previous position
4477
4478 if (zeroes_done == st_off)
4479 zeroes_done = next_init_off;
4480
4481 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4482
4483 #ifdef ASSERT
4484 // Various order invariants. Weaker than stores_are_sane because
4485 // a large constant tile can be filled in by smaller non-constant stores.
4486 assert(st_off >= last_init_off, "inits do not reverse");
4487 last_init_off = st_off;
4488 const Type* val = nullptr;
4489 if (st_size >= BytesPerInt &&
4490 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4491 (int)val->basic_type() < (int)T_OBJECT) {
4492 assert(st_off >= last_tile_end, "tiles do not overlap");
4493 assert(st_off >= last_init_end, "tiles do not overwrite inits");
4494 last_tile_end = MAX2(last_tile_end, next_init_off);
4495 } else {
4496 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4497 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4498 assert(st_off >= last_init_end, "inits do not overlap");
4499 last_init_end = next_init_off; // it's a non-tile
4500 }
4501 #endif //ASSERT
4502 }
4503
4504 remove_extra_zeroes(); // clear out all the zmems left over
4505 add_req(inits);
4506
4507 if (!(UseTLAB && ZeroTLAB)) {
4508 // If anything remains to be zeroed, zero it all now.
4509 zeroes_done = align_down(zeroes_done, BytesPerInt);
4510 // if it is the last unused 4 bytes of an instance, forget about it
4511 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4512 if (zeroes_done + BytesPerLong >= size_limit) {
4513 AllocateNode* alloc = allocation();
4514 assert(alloc != nullptr, "must be present");
4515 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
4516 Node* klass_node = alloc->in(AllocateNode::KlassNode);
4517 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4518 if (zeroes_done == k->layout_helper())
4519 zeroes_done = size_limit;
4520 }
4521 }
4522 if (zeroes_done < size_limit) {
4523 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4524 zeroes_done, size_in_bytes, phase);
4525 }
4526 }
4527
4528 set_complete(phase);
4529 return rawmem;
4530 }
4531
4532
4533 #ifdef ASSERT
4534 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4535 if (is_complete())
4536 return true; // stores could be anything at this point
4537 assert(allocation() != nullptr, "must be present");
4538 intptr_t last_off = allocation()->minimum_header_size();
4539 for (uint i = InitializeNode::RawStores; i < req(); i++) {
4540 Node* st = in(i);
4541 intptr_t st_off = get_store_offset(st, phase);
4542 if (st_off < 0) continue; // ignore dead garbage
4543 if (last_off > st_off) {
4544 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4545 this->dump(2);
4546 assert(false, "ascending store offsets");
4547 return false;
4548 }
4549 last_off = st_off + st->as_Store()->memory_size();
4550 }
4551 return true;
4552 }
4553 #endif //ASSERT
4554
4555
4556
4557
4558 //============================MergeMemNode=====================================
4559 //
4560 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4561 // contributing store or call operations. Each contributor provides the memory
4562 // state for a particular "alias type" (see Compile::alias_type). For example,
4563 // if a MergeMem has an input X for alias category #6, then any memory reference
4564 // to alias category #6 may use X as its memory state input, as an exact equivalent
4565 // to using the MergeMem as a whole.
4566 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4567 //
4568 // (Here, the <N> notation gives the index of the relevant adr_type.)
4569 //
4570 // In one special case (and more cases in the future), alias categories overlap.
4571 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4572 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4573 // it is exactly equivalent to that state W:
4574 // MergeMem(<Bot>: W) <==> W
4575 //
4576 // Usually, the merge has more than one input. In that case, where inputs
4577 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4578 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4579 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4580 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4581 //
4582 // A merge can take a "wide" memory state as one of its narrow inputs.
4583 // This simply means that the merge observes out only the relevant parts of
4584 // the wide input. That is, wide memory states arriving at narrow merge inputs
4585 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4586 //
4587 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4588 // and that memory slices "leak through":
4589 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4590 //
4591 // But, in such a cascade, repeated memory slices can "block the leak":
4592 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4593 //
4594 // In the last example, Y is not part of the combined memory state of the
4595 // outermost MergeMem. The system must, of course, prevent unschedulable
4596 // memory states from arising, so you can be sure that the state Y is somehow
4597 // a precursor to state Y'.
4598 //
4599 //
4600 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4601 // of each MergeMemNode array are exactly the numerical alias indexes, including
4602 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4603 // Compile::alias_type (and kin) produce and manage these indexes.
4604 //
4605 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4606 // (Note that this provides quick access to the top node inside MergeMem methods,
4607 // without the need to reach out via TLS to Compile::current.)
4608 //
4609 // As a consequence of what was just described, a MergeMem that represents a full
4610 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4611 // containing all alias categories.
4612 //
4613 // MergeMem nodes never (?) have control inputs, so in(0) is null.
4614 //
4615 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4616 // a memory state for the alias type <N>, or else the top node, meaning that
4617 // there is no particular input for that alias type. Note that the length of
4618 // a MergeMem is variable, and may be extended at any time to accommodate new
4619 // memory states at larger alias indexes. When merges grow, they are of course
4620 // filled with "top" in the unused in() positions.
4621 //
4622 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4623 // (Top was chosen because it works smoothly with passes like GCM.)
4624 //
4625 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4626 // the type of random VM bits like TLS references.) Since it is always the
4627 // first non-Bot memory slice, some low-level loops use it to initialize an
4628 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4629 //
4630 //
4631 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4632 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4633 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4634 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4635 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4636 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4637 //
4638 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4639 // really that different from the other memory inputs. An abbreviation called
4640 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4641 //
4642 //
4643 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4644 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4645 // that "emerges though" the base memory will be marked as excluding the alias types
4646 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4647 //
4648 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4649 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4650 //
4651 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4652 // (It is currently unimplemented.) As you can see, the resulting merge is
4653 // actually a disjoint union of memory states, rather than an overlay.
4654 //
4655
4656 //------------------------------MergeMemNode-----------------------------------
4657 Node* MergeMemNode::make_empty_memory() {
4658 Node* empty_memory = (Node*) Compile::current()->top();
4659 assert(empty_memory->is_top(), "correct sentinel identity");
4660 return empty_memory;
4661 }
4662
4663 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4664 init_class_id(Class_MergeMem);
4665 // all inputs are nullified in Node::Node(int)
4666 // set_input(0, nullptr); // no control input
4667
4668 // Initialize the edges uniformly to top, for starters.
4669 Node* empty_mem = make_empty_memory();
4670 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4671 init_req(i,empty_mem);
4672 }
4673 assert(empty_memory() == empty_mem, "");
4674
4675 if( new_base != nullptr && new_base->is_MergeMem() ) {
4676 MergeMemNode* mdef = new_base->as_MergeMem();
4677 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4678 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4679 mms.set_memory(mms.memory2());
4680 }
4681 assert(base_memory() == mdef->base_memory(), "");
4682 } else {
4683 set_base_memory(new_base);
4684 }
4685 }
4686
4687 // Make a new, untransformed MergeMem with the same base as 'mem'.
4688 // If mem is itself a MergeMem, populate the result with the same edges.
4689 MergeMemNode* MergeMemNode::make(Node* mem) {
4690 return new MergeMemNode(mem);
4691 }
4692
4693 //------------------------------cmp--------------------------------------------
4694 uint MergeMemNode::hash() const { return NO_HASH; }
4695 bool MergeMemNode::cmp( const Node &n ) const {
4696 return (&n == this); // Always fail except on self
4697 }
4698
4699 //------------------------------Identity---------------------------------------
4700 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4701 // Identity if this merge point does not record any interesting memory
4702 // disambiguations.
4703 Node* base_mem = base_memory();
4704 Node* empty_mem = empty_memory();
4705 if (base_mem != empty_mem) { // Memory path is not dead?
4706 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4707 Node* mem = in(i);
4708 if (mem != empty_mem && mem != base_mem) {
4709 return this; // Many memory splits; no change
4710 }
4711 }
4712 }
4713 return base_mem; // No memory splits; ID on the one true input
4714 }
4715
4716 //------------------------------Ideal------------------------------------------
4717 // This method is invoked recursively on chains of MergeMem nodes
4718 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4719 // Remove chain'd MergeMems
4720 //
4721 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4722 // relative to the "in(Bot)". Since we are patching both at the same time,
4723 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4724 // but rewrite each "in(i)" relative to the new "in(Bot)".
4725 Node *progress = nullptr;
4726
4727
4728 Node* old_base = base_memory();
4729 Node* empty_mem = empty_memory();
4730 if (old_base == empty_mem)
4731 return nullptr; // Dead memory path.
4732
4733 MergeMemNode* old_mbase;
4734 if (old_base != nullptr && old_base->is_MergeMem())
4735 old_mbase = old_base->as_MergeMem();
4736 else
4737 old_mbase = nullptr;
4738 Node* new_base = old_base;
4739
4740 // simplify stacked MergeMems in base memory
4741 if (old_mbase) new_base = old_mbase->base_memory();
4742
4743 // the base memory might contribute new slices beyond my req()
4744 if (old_mbase) grow_to_match(old_mbase);
4745
4746 // Look carefully at the base node if it is a phi.
4747 PhiNode* phi_base;
4748 if (new_base != NULL && new_base->is_Phi())
4749 phi_base = new_base->as_Phi();
4750 else
4751 phi_base = NULL;
4752
4753 Node* phi_reg = NULL;
4754 uint phi_len = (uint)-1;
4755 if (phi_base != NULL) {
4756 phi_reg = phi_base->region();
4757 phi_len = phi_base->req();
4758 // see if the phi is unfinished
4759 for (uint i = 1; i < phi_len; i++) {
4760 if (phi_base->in(i) == NULL) {
4761 // incomplete phi; do not look at it yet!
4762 phi_reg = NULL;
4763 phi_len = (uint)-1;
4764 break;
4765 }
4766 }
4767 }
4768
4769 // Note: We do not call verify_sparse on entry, because inputs
4770 // can normalize to the base_memory via subsume_node or similar
4771 // mechanisms. This method repairs that damage.
4772
4773 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4774
4775 // Look at each slice.
4776 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4777 Node* old_in = in(i);
4778 // calculate the old memory value
4779 Node* old_mem = old_in;
4780 if (old_mem == empty_mem) old_mem = old_base;
4781 assert(old_mem == memory_at(i), "");
4782
4783 // maybe update (reslice) the old memory value
4784
4785 // simplify stacked MergeMems
4786 Node* new_mem = old_mem;
4787 MergeMemNode* old_mmem;
4788 if (old_mem != nullptr && old_mem->is_MergeMem())
4789 old_mmem = old_mem->as_MergeMem();
4790 else
4791 old_mmem = nullptr;
4792 if (old_mmem == this) {
4793 // This can happen if loops break up and safepoints disappear.
4794 // A merge of BotPtr (default) with a RawPtr memory derived from a
4795 // safepoint can be rewritten to a merge of the same BotPtr with
4796 // the BotPtr phi coming into the loop. If that phi disappears
4797 // also, we can end up with a self-loop of the mergemem.
4798 // In general, if loops degenerate and memory effects disappear,
4799 // a mergemem can be left looking at itself. This simply means
4800 // that the mergemem's default should be used, since there is
4801 // no longer any apparent effect on this slice.
4802 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4803 // from start. Update the input to TOP.
4804 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4805 }
4806 else if (old_mmem != nullptr) {
4807 new_mem = old_mmem->memory_at(i);
4808 }
4809 // else preceding memory was not a MergeMem
4810
4811 // maybe store down a new value
4812 Node* new_in = new_mem;
4813 if (new_in == new_base) new_in = empty_mem;
4814
4815 if (new_in != old_in) {
4816 // Warning: Do not combine this "if" with the previous "if"
4817 // A memory slice might have be be rewritten even if it is semantically
4818 // unchanged, if the base_memory value has changed.
4819 set_req_X(i, new_in, phase);
4820 progress = this; // Report progress
4821 }
4822 }
4823
4824 if (new_base != old_base) {
4825 set_req_X(Compile::AliasIdxBot, new_base, phase);
4826 // Don't use set_base_memory(new_base), because we need to update du.
4827 assert(base_memory() == new_base, "");
4828 progress = this;
4829 }
4830
4831 if( base_memory() == this ) {
4832 // a self cycle indicates this memory path is dead
4833 set_req(Compile::AliasIdxBot, empty_mem);
4834 }
4835
4836 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4837 // Recursion must occur after the self cycle check above
4838 if( base_memory()->is_MergeMem() ) {
4839 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4840 Node *m = phase->transform(new_mbase); // Rollup any cycles
4841 if( m != nullptr &&
4842 (m->is_top() ||
4843 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4844 // propagate rollup of dead cycle to self
4845 set_req(Compile::AliasIdxBot, empty_mem);
4846 }
4847 }
4848
4849 if( base_memory() == empty_mem ) {
4850 progress = this;
4851 // Cut inputs during Parse phase only.
4852 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4853 if( !can_reshape ) {
4854 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4855 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4856 }
4857 }
4858 }
4859
4860 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4861 // Check if PhiNode::Ideal's "Split phis through memory merges"
4862 // transform should be attempted. Look for this->phi->this cycle.
4863 uint merge_width = req();
4864 if (merge_width > Compile::AliasIdxRaw) {
4865 PhiNode* phi = base_memory()->as_Phi();
4866 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4867 if (phi->in(i) == this) {
4868 phase->is_IterGVN()->_worklist.push(phi);
4869 break;
4870 }
4871 }
4872 }
4873 }
4874
4875 assert(progress || verify_sparse(), "please, no dups of base");
4876 return progress;
4877 }
4878
4879 //-------------------------set_base_memory-------------------------------------
4880 void MergeMemNode::set_base_memory(Node *new_base) {
4881 Node* empty_mem = empty_memory();
4882 set_req(Compile::AliasIdxBot, new_base);
4883 assert(memory_at(req()) == new_base, "must set default memory");
4884 // Clear out other occurrences of new_base:
4885 if (new_base != empty_mem) {
4886 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4887 if (in(i) == new_base) set_req(i, empty_mem);
4888 }
4889 }
4890 }
4891
4892 //------------------------------out_RegMask------------------------------------
4893 const RegMask &MergeMemNode::out_RegMask() const {
4894 return RegMask::Empty;
4895 }
4896
4897 //------------------------------dump_spec--------------------------------------
4898 #ifndef PRODUCT
4899 void MergeMemNode::dump_spec(outputStream *st) const {
4900 st->print(" {");
4901 Node* base_mem = base_memory();
4902 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4903 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem;
4904 if (mem == base_mem) { st->print(" -"); continue; }
4905 st->print( " N%d:", mem->_idx );
4906 Compile::current()->get_adr_type(i)->dump_on(st);
4907 }
4908 st->print(" }");
4909 }
4910 #endif // !PRODUCT
4911
4912
4913 #ifdef ASSERT
4914 static bool might_be_same(Node* a, Node* b) {
4915 if (a == b) return true;
4916 if (!(a->is_Phi() || b->is_Phi())) return false;
4917 // phis shift around during optimization
4918 return true; // pretty stupid...
4919 }
4920
4921 // verify a narrow slice (either incoming or outgoing)
4922 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4923 if (!VerifyAliases) return; // don't bother to verify unless requested
4924 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
4925 if (Node::in_dump()) return; // muzzle asserts when printing
4926 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4927 assert(n != nullptr, "");
4928 // Elide intervening MergeMem's
4929 while (n->is_MergeMem()) {
4930 n = n->as_MergeMem()->memory_at(alias_idx);
4931 }
4932 Compile* C = Compile::current();
4933 const TypePtr* n_adr_type = n->adr_type();
4934 if (n == m->empty_memory()) {
4935 // Implicit copy of base_memory()
4936 } else if (n_adr_type != TypePtr::BOTTOM) {
4937 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type");
4938 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4939 } else {
4940 // A few places like make_runtime_call "know" that VM calls are narrow,
4941 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4942 bool expected_wide_mem = false;
4943 if (n == m->base_memory()) {
4944 expected_wide_mem = true;
4945 } else if (alias_idx == Compile::AliasIdxRaw ||
4946 n == m->memory_at(Compile::AliasIdxRaw)) {
4947 expected_wide_mem = true;
4948 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4949 // memory can "leak through" calls on channels that
4950 // are write-once. Allow this also.
4951 expected_wide_mem = true;
4952 }
4953 assert(expected_wide_mem, "expected narrow slice replacement");
4954 }
4955 }
4956 #else // !ASSERT
4957 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4958 #endif
4959
4960
4961 //-----------------------------memory_at---------------------------------------
4962 Node* MergeMemNode::memory_at(uint alias_idx) const {
4963 assert(alias_idx >= Compile::AliasIdxRaw ||
4964 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4965 "must avoid base_memory and AliasIdxTop");
4966
4967 // Otherwise, it is a narrow slice.
4968 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4969 Compile *C = Compile::current();
4970 if (is_empty_memory(n)) {
4971 // the array is sparse; empty slots are the "top" node
4972 n = base_memory();
4973 assert(Node::in_dump()
4974 || n == nullptr || n->bottom_type() == Type::TOP
4975 || n->adr_type() == nullptr // address is TOP
4976 || n->adr_type() == TypePtr::BOTTOM
4977 || n->adr_type() == TypeRawPtr::BOTTOM
4978 || Compile::current()->AliasLevel() == 0,
4979 "must be a wide memory");
4980 // AliasLevel == 0 if we are organizing the memory states manually.
4981 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4982 } else {
4983 // make sure the stored slice is sane
4984 #ifdef ASSERT
4985 if (VMError::is_error_reported() || Node::in_dump()) {
4986 } else if (might_be_same(n, base_memory())) {
4987 // Give it a pass: It is a mostly harmless repetition of the base.
4988 // This can arise normally from node subsumption during optimization.
4989 } else {
4990 verify_memory_slice(this, alias_idx, n);
4991 }
4992 #endif
4993 }
4994 return n;
4995 }
4996
4997 //---------------------------set_memory_at-------------------------------------
4998 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4999 verify_memory_slice(this, alias_idx, n);
5000 Node* empty_mem = empty_memory();
5001 if (n == base_memory()) n = empty_mem; // collapse default
5002 uint need_req = alias_idx+1;
5003 if (req() < need_req) {
5004 if (n == empty_mem) return; // already the default, so do not grow me
5005 // grow the sparse array
5006 do {
5007 add_req(empty_mem);
5008 } while (req() < need_req);
5009 }
5010 set_req( alias_idx, n );
5011 }
5012
5013
5014
5015 //--------------------------iteration_setup------------------------------------
5016 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
5017 if (other != nullptr) {
5018 grow_to_match(other);
5019 // invariant: the finite support of mm2 is within mm->req()
5020 #ifdef ASSERT
5021 for (uint i = req(); i < other->req(); i++) {
5022 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
5023 }
5024 #endif
5025 }
5026 // Replace spurious copies of base_memory by top.
5027 Node* base_mem = base_memory();
5028 if (base_mem != nullptr && !base_mem->is_top()) {
5029 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
5030 if (in(i) == base_mem)
5031 set_req(i, empty_memory());
5032 }
5033 }
5034 }
5035
5036 //---------------------------grow_to_match-------------------------------------
5037 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
5038 Node* empty_mem = empty_memory();
5039 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
5040 // look for the finite support of the other memory
5041 for (uint i = other->req(); --i >= req(); ) {
5042 if (other->in(i) != empty_mem) {
5043 uint new_len = i+1;
5044 while (req() < new_len) add_req(empty_mem);
5045 break;
5046 }
5047 }
5048 }
5049
5050 //---------------------------verify_sparse-------------------------------------
5051 #ifndef PRODUCT
5052 bool MergeMemNode::verify_sparse() const {
5053 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
5054 Node* base_mem = base_memory();
5055 // The following can happen in degenerate cases, since empty==top.
5056 if (is_empty_memory(base_mem)) return true;
5057 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5058 assert(in(i) != nullptr, "sane slice");
5059 if (in(i) == base_mem) return false; // should have been the sentinel value!
5060 }
5061 return true;
5062 }
5063
5064 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
5065 Node* n;
5066 n = mm->in(idx);
5067 if (mem == n) return true; // might be empty_memory()
5068 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
5069 if (mem == n) return true;
5070 return false;
5071 }
5072 #endif // !PRODUCT