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