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