1 /*
2 * Copyright (c) 1997, 2024, 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->isa_aryptr()) {
231 mem_t = mem_t->is_aryptr()
232 ->cast_to_stable(t_oop->is_aryptr()->is_stable())
233 ->cast_to_size(t_oop->is_aryptr()->size())
234 ->with_offset(t_oop->is_aryptr()->offset())
235 ->is_aryptr();
236 }
237 do_split = mem_t == t_oop;
238 }
239 if (do_split) {
240 // clone the Phi with our address type
241 result = mphi->split_out_instance(t_adr, igvn);
242 } else {
243 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
244 }
245 }
246 return result;
247 }
248
249 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
250 uint alias_idx = phase->C->get_alias_index(tp);
251 Node *mem = mmem;
252 #ifdef ASSERT
253 {
254 // Check that current type is consistent with the alias index used during graph construction
255 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
256 bool consistent = adr_check == nullptr || adr_check->empty() ||
257 phase->C->must_alias(adr_check, alias_idx );
258 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
259 if( !consistent && adr_check != nullptr && !adr_check->empty() &&
260 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
261 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
262 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
263 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
264 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
265 // don't assert if it is dead code.
266 consistent = true;
267 }
268 if( !consistent ) {
269 st->print("alias_idx==%d, adr_check==", alias_idx);
270 if( adr_check == nullptr ) {
271 st->print("null");
272 } else {
273 adr_check->dump();
274 }
275 st->cr();
276 print_alias_types();
277 assert(consistent, "adr_check must match alias idx");
278 }
279 }
280 #endif
281 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
282 // means an array I have not precisely typed yet. Do not do any
283 // alias stuff with it any time soon.
284 const TypeOopPtr *toop = tp->isa_oopptr();
285 if (tp->base() != Type::AnyPtr &&
286 !(toop &&
287 toop->isa_instptr() &&
288 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
289 toop->offset() == Type::OffsetBot)) {
290 // compress paths and change unreachable cycles to TOP
291 // If not, we can update the input infinitely along a MergeMem cycle
292 // Equivalent code in PhiNode::Ideal
293 Node* m = phase->transform(mmem);
294 // If transformed to a MergeMem, get the desired slice
295 // Otherwise the returned node represents memory for every slice
296 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
297 // Update input if it is progress over what we have now
298 }
299 return mem;
300 }
301
302 //--------------------------Ideal_common---------------------------------------
303 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
304 // Unhook non-raw memories from complete (macro-expanded) initializations.
305 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
306 // If our control input is a dead region, kill all below the region
307 Node *ctl = in(MemNode::Control);
308 if (ctl && remove_dead_region(phase, can_reshape))
309 return this;
310 ctl = in(MemNode::Control);
311 // Don't bother trying to transform a dead node
312 if (ctl && ctl->is_top()) return NodeSentinel;
313
314 PhaseIterGVN *igvn = phase->is_IterGVN();
315 // Wait if control on the worklist.
316 if (ctl && can_reshape && igvn != nullptr) {
317 Node* bol = nullptr;
318 Node* cmp = nullptr;
319 if (ctl->in(0)->is_If()) {
320 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
321 bol = ctl->in(0)->in(1);
322 if (bol->is_Bool())
323 cmp = ctl->in(0)->in(1)->in(1);
324 }
325 if (igvn->_worklist.member(ctl) ||
326 (bol != nullptr && igvn->_worklist.member(bol)) ||
327 (cmp != nullptr && igvn->_worklist.member(cmp)) ) {
328 // This control path may be dead.
329 // Delay this memory node transformation until the control is processed.
330 igvn->_worklist.push(this);
331 return NodeSentinel; // caller will return null
332 }
333 }
334 // Ignore if memory is dead, or self-loop
335 Node *mem = in(MemNode::Memory);
336 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return null
337 assert(mem != this, "dead loop in MemNode::Ideal");
338
339 if (can_reshape && igvn != nullptr && igvn->_worklist.member(mem)) {
340 // This memory slice may be dead.
341 // Delay this mem node transformation until the memory is processed.
342 igvn->_worklist.push(this);
343 return NodeSentinel; // caller will return null
344 }
345
346 Node *address = in(MemNode::Address);
347 const Type *t_adr = phase->type(address);
348 if (t_adr == Type::TOP) return NodeSentinel; // caller will return null
349
350 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
351 // Unsafe off-heap access with zero address. Remove access and other control users
352 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
353 assert(ctl != nullptr, "unsafe accesses should be control dependent");
354 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
355 Node* u = ctl->fast_out(i);
356 if (u != ctl) {
357 igvn->rehash_node_delayed(u);
358 int nb = u->replace_edge(ctl, phase->C->top(), igvn);
359 --i, imax -= nb;
360 }
361 }
362 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
363 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
364 phase->C->root()->add_req(halt);
365 return this;
366 }
367
368 if (can_reshape && igvn != nullptr &&
369 (igvn->_worklist.member(address) ||
370 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
371 // The address's base and type may change when the address is processed.
372 // Delay this mem node transformation until the address is processed.
373 igvn->_worklist.push(this);
374 return NodeSentinel; // caller will return null
375 }
376
377 // Do NOT remove or optimize the next lines: ensure a new alias index
378 // is allocated for an oop pointer type before Escape Analysis.
379 // Note: C++ will not remove it since the call has side effect.
380 if (t_adr->isa_oopptr()) {
381 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
382 }
383
384 Node* base = nullptr;
385 if (address->is_AddP()) {
386 base = address->in(AddPNode::Base);
387 }
388 if (base != nullptr && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
389 !t_adr->isa_rawptr()) {
390 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
391 // Skip this node optimization if its address has TOP base.
392 return NodeSentinel; // caller will return null
393 }
394
395 // Avoid independent memory operations
396 Node* old_mem = mem;
397
398 // The code which unhooks non-raw memories from complete (macro-expanded)
399 // initializations was removed. After macro-expansion all stores caught
400 // by Initialize node became raw stores and there is no information
401 // which memory slices they modify. So it is unsafe to move any memory
402 // operation above these stores. Also in most cases hooked non-raw memories
403 // were already unhooked by using information from detect_ptr_independence()
404 // and find_previous_store().
405
406 if (mem->is_MergeMem()) {
407 MergeMemNode* mmem = mem->as_MergeMem();
408 const TypePtr *tp = t_adr->is_ptr();
409
410 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
411 }
412
413 if (mem != old_mem) {
414 set_req_X(MemNode::Memory, mem, phase);
415 if (phase->type(mem) == Type::TOP) return NodeSentinel;
416 return this;
417 }
418
419 // let the subclass continue analyzing...
420 return nullptr;
421 }
422
423 // Helper function for proving some simple control dominations.
424 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
425 // Already assumes that 'dom' is available at 'sub', and that 'sub'
426 // is not a constant (dominated by the method's StartNode).
427 // Used by MemNode::find_previous_store to prove that the
428 // control input of a memory operation predates (dominates)
429 // an allocation it wants to look past.
430 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
431 if (dom == nullptr || dom->is_top() || sub == nullptr || sub->is_top())
432 return false; // Conservative answer for dead code
433
434 // Check 'dom'. Skip Proj and CatchProj nodes.
435 dom = dom->find_exact_control(dom);
436 if (dom == nullptr || dom->is_top())
437 return false; // Conservative answer for dead code
438
439 if (dom == sub) {
440 // For the case when, for example, 'sub' is Initialize and the original
441 // 'dom' is Proj node of the 'sub'.
442 return false;
443 }
444
445 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
446 return true;
447
448 // 'dom' dominates 'sub' if its control edge and control edges
449 // of all its inputs dominate or equal to sub's control edge.
450
451 // Currently 'sub' is either Allocate, Initialize or Start nodes.
452 // Or Region for the check in LoadNode::Ideal();
453 // 'sub' should have sub->in(0) != nullptr.
454 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
455 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
456
457 // Get control edge of 'sub'.
458 Node* orig_sub = sub;
459 sub = sub->find_exact_control(sub->in(0));
460 if (sub == nullptr || sub->is_top())
461 return false; // Conservative answer for dead code
462
463 assert(sub->is_CFG(), "expecting control");
464
465 if (sub == dom)
466 return true;
467
468 if (sub->is_Start() || sub->is_Root())
469 return false;
470
471 {
472 // Check all control edges of 'dom'.
473
474 ResourceMark rm;
475 Node_List nlist;
476 Unique_Node_List dom_list;
477
478 dom_list.push(dom);
479 bool only_dominating_controls = false;
480
481 for (uint next = 0; next < dom_list.size(); next++) {
482 Node* n = dom_list.at(next);
483 if (n == orig_sub)
484 return false; // One of dom's inputs dominated by sub.
485 if (!n->is_CFG() && n->pinned()) {
486 // Check only own control edge for pinned non-control nodes.
487 n = n->find_exact_control(n->in(0));
488 if (n == nullptr || n->is_top())
489 return false; // Conservative answer for dead code
490 assert(n->is_CFG(), "expecting control");
491 dom_list.push(n);
492 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
493 only_dominating_controls = true;
494 } else if (n->is_CFG()) {
495 if (n->dominates(sub, nlist))
496 only_dominating_controls = true;
497 else
498 return false;
499 } else {
500 // First, own control edge.
501 Node* m = n->find_exact_control(n->in(0));
502 if (m != nullptr) {
503 if (m->is_top())
504 return false; // Conservative answer for dead code
505 dom_list.push(m);
506 }
507 // Now, the rest of edges.
508 uint cnt = n->req();
509 for (uint i = 1; i < cnt; i++) {
510 m = n->find_exact_control(n->in(i));
511 if (m == nullptr || m->is_top())
512 continue;
513 dom_list.push(m);
514 }
515 }
516 }
517 return only_dominating_controls;
518 }
519 }
520
521 //---------------------detect_ptr_independence---------------------------------
522 // Used by MemNode::find_previous_store to prove that two base
523 // pointers are never equal.
524 // The pointers are accompanied by their associated allocations,
525 // if any, which have been previously discovered by the caller.
526 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
527 Node* p2, AllocateNode* a2,
528 PhaseTransform* phase) {
529 // Attempt to prove that these two pointers cannot be aliased.
530 // They may both manifestly be allocations, and they should differ.
531 // Or, if they are not both allocations, they can be distinct constants.
532 // Otherwise, one is an allocation and the other a pre-existing value.
533 if (a1 == nullptr && a2 == nullptr) { // neither an allocation
534 return (p1 != p2) && p1->is_Con() && p2->is_Con();
535 } else if (a1 != nullptr && a2 != nullptr) { // both allocations
536 return (a1 != a2);
537 } else if (a1 != nullptr) { // one allocation a1
538 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
539 return all_controls_dominate(p2, a1);
540 } else { //(a2 != null) // one allocation a2
541 return all_controls_dominate(p1, a2);
542 }
543 return false;
544 }
545
546
547 // Find an arraycopy ac that produces the memory state represented by parameter mem.
548 // Return ac if
549 // (a) can_see_stored_value=true and ac must have set the value for this load or if
550 // (b) can_see_stored_value=false and ac could have set the value for this load or if
551 // (c) can_see_stored_value=false and ac cannot have set the value for this load.
552 // In case (c) change the parameter mem to the memory input of ac to skip it
553 // when searching stored value.
554 // Otherwise return null.
555 Node* LoadNode::find_previous_arraycopy(PhaseValues* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
556 ArrayCopyNode* ac = find_array_copy_clone(ld_alloc, mem);
557 if (ac != nullptr) {
558 Node* ld_addp = in(MemNode::Address);
559 Node* src = ac->in(ArrayCopyNode::Src);
560 const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
561
562 // This is a load from a cloned array. The corresponding arraycopy ac must
563 // have set the value for the load and we can return ac but only if the load
564 // is known to be within bounds. This is checked below.
565 if (ary_t != nullptr && ld_addp->is_AddP()) {
566 Node* ld_offs = ld_addp->in(AddPNode::Offset);
567 BasicType ary_elem = ary_t->elem()->array_element_basic_type();
568 jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
569 jlong elemsize = type2aelembytes(ary_elem);
570
571 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
572 const TypeInt* sizetype = ary_t->size();
573
574 if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
575 // The load is known to be within bounds. It receives its value from ac.
576 return ac;
577 }
578 // The load is known to be out-of-bounds.
579 }
580 // The load could be out-of-bounds. It must not be hoisted but must remain
581 // dependent on the runtime range check. This is achieved by returning null.
582 } else if (mem->is_Proj() && mem->in(0) != nullptr && mem->in(0)->is_ArrayCopy()) {
583 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
584
585 if (ac->is_arraycopy_validated() ||
586 ac->is_copyof_validated() ||
587 ac->is_copyofrange_validated()) {
588 Node* ld_addp = in(MemNode::Address);
589 if (ld_addp->is_AddP()) {
590 Node* ld_base = ld_addp->in(AddPNode::Address);
591 Node* ld_offs = ld_addp->in(AddPNode::Offset);
592
593 Node* dest = ac->in(ArrayCopyNode::Dest);
594
595 if (dest == ld_base) {
596 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
597 assert(!ld_offs_t->empty(), "dead reference should be checked already");
598 // Take into account vector or unsafe access size
599 jlong ld_size_in_bytes = (jlong)memory_size();
600 jlong offset_hi = ld_offs_t->_hi + ld_size_in_bytes - 1;
601 offset_hi = MIN2(offset_hi, (jlong)(TypeX::MAX->_hi)); // Take care for overflow in 32-bit VM
602 if (ac->modifies(ld_offs_t->_lo, (intptr_t)offset_hi, phase, can_see_stored_value)) {
603 return ac;
604 }
605 if (!can_see_stored_value) {
606 mem = ac->in(TypeFunc::Memory);
607 return ac;
608 }
609 }
610 }
611 }
612 }
613 return nullptr;
614 }
615
616 ArrayCopyNode* MemNode::find_array_copy_clone(Node* ld_alloc, Node* mem) const {
617 if (mem->is_Proj() && mem->in(0) != nullptr && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
618 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
619 if (ld_alloc != nullptr) {
620 // Check if there is an array copy for a clone
621 Node* mb = mem->in(0);
622 ArrayCopyNode* ac = nullptr;
623 if (mb->in(0) != nullptr && mb->in(0)->is_Proj() &&
624 mb->in(0)->in(0) != nullptr && mb->in(0)->in(0)->is_ArrayCopy()) {
625 ac = mb->in(0)->in(0)->as_ArrayCopy();
626 } else {
627 // Step over GC barrier when ReduceInitialCardMarks is disabled
628 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
629 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
630
631 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
632 ac = control_proj_ac->in(0)->as_ArrayCopy();
633 }
634 }
635
636 if (ac != nullptr && ac->is_clonebasic()) {
637 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest));
638 if (alloc != nullptr && alloc == ld_alloc) {
639 return ac;
640 }
641 }
642 }
643 }
644 return nullptr;
645 }
646
647 // The logic for reordering loads and stores uses four steps:
648 // (a) Walk carefully past stores and initializations which we
649 // can prove are independent of this load.
650 // (b) Observe that the next memory state makes an exact match
651 // with self (load or store), and locate the relevant store.
652 // (c) Ensure that, if we were to wire self directly to the store,
653 // the optimizer would fold it up somehow.
654 // (d) Do the rewiring, and return, depending on some other part of
655 // the optimizer to fold up the load.
656 // This routine handles steps (a) and (b). Steps (c) and (d) are
657 // specific to loads and stores, so they are handled by the callers.
658 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
659 //
660 Node* MemNode::find_previous_store(PhaseValues* phase) {
661 Node* ctrl = in(MemNode::Control);
662 Node* adr = in(MemNode::Address);
663 intptr_t offset = 0;
664 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
665 AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
666
667 if (offset == Type::OffsetBot)
668 return nullptr; // cannot unalias unless there are precise offsets
669
670 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
671 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
672
673 intptr_t size_in_bytes = memory_size();
674
675 Node* mem = in(MemNode::Memory); // start searching here...
676
677 int cnt = 50; // Cycle limiter
678 for (;;) { // While we can dance past unrelated stores...
679 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
680
681 Node* prev = mem;
682 if (mem->is_Store()) {
683 Node* st_adr = mem->in(MemNode::Address);
684 intptr_t st_offset = 0;
685 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
686 if (st_base == nullptr)
687 break; // inscrutable pointer
688
689 // For raw accesses it's not enough to prove that constant offsets don't intersect.
690 // We need the bases to be the equal in order for the offset check to make sense.
691 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
692 break;
693 }
694
695 if (st_offset != offset && st_offset != Type::OffsetBot) {
696 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
697 assert(mem->as_Store()->memory_size() <= MAX_STORE, "");
698 if (st_offset >= offset + size_in_bytes ||
699 st_offset <= offset - MAX_STORE ||
700 st_offset <= offset - mem->as_Store()->memory_size()) {
701 // Success: The offsets are provably independent.
702 // (You may ask, why not just test st_offset != offset and be done?
703 // The answer is that stores of different sizes can co-exist
704 // in the same sequence of RawMem effects. We sometimes initialize
705 // a whole 'tile' of array elements with a single jint or jlong.)
706 mem = mem->in(MemNode::Memory);
707 continue; // (a) advance through independent store memory
708 }
709 }
710 if (st_base != base &&
711 detect_ptr_independence(base, alloc,
712 st_base,
713 AllocateNode::Ideal_allocation(st_base),
714 phase)) {
715 // Success: The bases are provably independent.
716 mem = mem->in(MemNode::Memory);
717 continue; // (a) advance through independent store memory
718 }
719
720 // (b) At this point, if the bases or offsets do not agree, we lose,
721 // since we have not managed to prove 'this' and 'mem' independent.
722 if (st_base == base && st_offset == offset) {
723 return mem; // let caller handle steps (c), (d)
724 }
725
726 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
727 InitializeNode* st_init = mem->in(0)->as_Initialize();
728 AllocateNode* st_alloc = st_init->allocation();
729 if (st_alloc == nullptr)
730 break; // something degenerated
731 bool known_identical = false;
732 bool known_independent = false;
733 if (alloc == st_alloc)
734 known_identical = true;
735 else if (alloc != nullptr)
736 known_independent = true;
737 else if (all_controls_dominate(this, st_alloc))
738 known_independent = true;
739
740 if (known_independent) {
741 // The bases are provably independent: Either they are
742 // manifestly distinct allocations, or else the control
743 // of this load dominates the store's allocation.
744 int alias_idx = phase->C->get_alias_index(adr_type());
745 if (alias_idx == Compile::AliasIdxRaw) {
746 mem = st_alloc->in(TypeFunc::Memory);
747 } else {
748 mem = st_init->memory(alias_idx);
749 }
750 continue; // (a) advance through independent store memory
751 }
752
753 // (b) at this point, if we are not looking at a store initializing
754 // the same allocation we are loading from, we lose.
755 if (known_identical) {
756 // From caller, can_see_stored_value will consult find_captured_store.
757 return mem; // let caller handle steps (c), (d)
758 }
759
760 } else if (find_previous_arraycopy(phase, alloc, mem, false) != nullptr) {
761 if (prev != mem) {
762 // Found an arraycopy but it doesn't affect that load
763 continue;
764 }
765 // Found an arraycopy that may affect that load
766 return mem;
767 } else if (addr_t != nullptr && addr_t->is_known_instance_field()) {
768 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
769 if (mem->is_Proj() && mem->in(0)->is_Call()) {
770 // ArrayCopyNodes processed here as well.
771 CallNode *call = mem->in(0)->as_Call();
772 if (!call->may_modify(addr_t, phase)) {
773 mem = call->in(TypeFunc::Memory);
774 continue; // (a) advance through independent call memory
775 }
776 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
777 ArrayCopyNode* ac = nullptr;
778 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
779 break;
780 }
781 mem = mem->in(0)->in(TypeFunc::Memory);
782 continue; // (a) advance through independent MemBar memory
783 } else if (mem->is_ClearArray()) {
784 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
785 // (the call updated 'mem' value)
786 continue; // (a) advance through independent allocation memory
787 } else {
788 // Can not bypass initialization of the instance
789 // we are looking for.
790 return mem;
791 }
792 } else if (mem->is_MergeMem()) {
793 int alias_idx = phase->C->get_alias_index(adr_type());
794 mem = mem->as_MergeMem()->memory_at(alias_idx);
795 continue; // (a) advance through independent MergeMem memory
796 }
797 }
798
799 // Unless there is an explicit 'continue', we must bail out here,
800 // because 'mem' is an inscrutable memory state (e.g., a call).
801 break;
802 }
803
804 return nullptr; // bail out
805 }
806
807 //----------------------calculate_adr_type-------------------------------------
808 // Helper function. Notices when the given type of address hits top or bottom.
809 // Also, asserts a cross-check of the type against the expected address type.
810 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
811 if (t == Type::TOP) return nullptr; // does not touch memory any more?
812 #ifdef ASSERT
813 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = nullptr;
814 #endif
815 const TypePtr* tp = t->isa_ptr();
816 if (tp == nullptr) {
817 assert(cross_check == nullptr || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
818 return TypePtr::BOTTOM; // touches lots of memory
819 } else {
820 #ifdef ASSERT
821 // %%%% [phh] We don't check the alias index if cross_check is
822 // TypeRawPtr::BOTTOM. Needs to be investigated.
823 if (cross_check != nullptr &&
824 cross_check != TypePtr::BOTTOM &&
825 cross_check != TypeRawPtr::BOTTOM) {
826 // Recheck the alias index, to see if it has changed (due to a bug).
827 Compile* C = Compile::current();
828 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
829 "must stay in the original alias category");
830 // The type of the address must be contained in the adr_type,
831 // disregarding "null"-ness.
832 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
833 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
834 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
835 "real address must not escape from expected memory type");
836 }
837 #endif
838 return tp;
839 }
840 }
841
842 uint8_t MemNode::barrier_data(const Node* n) {
843 if (n->is_LoadStore()) {
844 return n->as_LoadStore()->barrier_data();
845 } else if (n->is_Mem()) {
846 return n->as_Mem()->barrier_data();
847 }
848 return 0;
849 }
850
851 //=============================================================================
852 // Should LoadNode::Ideal() attempt to remove control edges?
853 bool LoadNode::can_remove_control() const {
854 return !has_pinned_control_dependency();
855 }
856 uint LoadNode::size_of() const { return sizeof(*this); }
857 bool LoadNode::cmp(const Node &n) const {
858 LoadNode& load = (LoadNode &)n;
859 return !Type::cmp(_type, load._type) &&
860 _control_dependency == load._control_dependency &&
861 _mo == load._mo;
862 }
863 const Type *LoadNode::bottom_type() const { return _type; }
864 uint LoadNode::ideal_reg() const {
865 return _type->ideal_reg();
866 }
867
868 #ifndef PRODUCT
869 void LoadNode::dump_spec(outputStream *st) const {
870 MemNode::dump_spec(st);
871 if( !Verbose && !WizardMode ) {
872 // standard dump does this in Verbose and WizardMode
873 st->print(" #"); _type->dump_on(st);
874 }
875 if (!depends_only_on_test()) {
876 st->print(" (does not depend only on test, ");
877 if (control_dependency() == UnknownControl) {
878 st->print("unknown control");
879 } else if (control_dependency() == Pinned) {
880 st->print("pinned");
881 } else if (adr_type() == TypeRawPtr::BOTTOM) {
882 st->print("raw access");
883 } else {
884 st->print("unknown reason");
885 }
886 st->print(")");
887 }
888 }
889 #endif
890
891 #ifdef ASSERT
892 //----------------------------is_immutable_value-------------------------------
893 // Helper function to allow a raw load without control edge for some cases
894 bool LoadNode::is_immutable_value(Node* adr) {
895 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
896 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) {
897
898 jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1);
899 int offsets[] = {
900 in_bytes(JavaThread::osthread_offset()),
901 in_bytes(JavaThread::threadObj_offset()),
902 in_bytes(JavaThread::vthread_offset()),
903 in_bytes(JavaThread::scopedValueCache_offset()),
904 };
905
906 for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) {
907 if (offset == offsets[i]) {
908 return true;
909 }
910 }
911 }
912
913 return false;
914 }
915 #endif
916
917 //----------------------------LoadNode::make-----------------------------------
918 // Polymorphic factory method:
919 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo,
920 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
921 Compile* C = gvn.C;
922
923 // sanity check the alias category against the created node type
924 assert(!(adr_type->isa_oopptr() &&
925 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
926 "use LoadKlassNode instead");
927 assert(!(adr_type->isa_aryptr() &&
928 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
929 "use LoadRangeNode instead");
930 // Check control edge of raw loads
931 assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
932 // oop will be recorded in oop map if load crosses safepoint
933 rt->isa_oopptr() || is_immutable_value(adr),
934 "raw memory operations should have control edge");
935 LoadNode* load = nullptr;
936 switch (bt) {
937 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
938 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
939 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
940 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
941 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
942 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break;
943 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
944 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break;
945 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
946 case T_OBJECT:
947 case T_NARROWOOP:
948 #ifdef _LP64
949 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
950 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
951 } else
952 #endif
953 {
954 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
955 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
956 }
957 break;
958 default:
959 ShouldNotReachHere();
960 break;
961 }
962 assert(load != nullptr, "LoadNode should have been created");
963 if (unaligned) {
964 load->set_unaligned_access();
965 }
966 if (mismatched) {
967 load->set_mismatched_access();
968 }
969 if (unsafe) {
970 load->set_unsafe_access();
971 }
972 load->set_barrier_data(barrier_data);
973 if (load->Opcode() == Op_LoadN) {
974 Node* ld = gvn.transform(load);
975 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
976 }
977
978 return load;
979 }
980
981 //------------------------------hash-------------------------------------------
982 uint LoadNode::hash() const {
983 // unroll addition of interesting fields
984 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
985 }
986
987 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
988 if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) {
989 bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile();
990 bool is_stable_ary = FoldStableValues &&
991 (tp != nullptr) && (tp->isa_aryptr() != nullptr) &&
992 tp->isa_aryptr()->is_stable();
993
994 return (eliminate_boxing && non_volatile) || is_stable_ary;
995 }
996
997 return false;
998 }
999
1000 LoadNode* LoadNode::pin_array_access_node() const {
1001 const TypePtr* adr_type = this->adr_type();
1002 if (adr_type != nullptr && adr_type->isa_aryptr()) {
1003 return clone_pinned();
1004 }
1005 return nullptr;
1006 }
1007
1008 // Is the value loaded previously stored by an arraycopy? If so return
1009 // a load node that reads from the source array so we may be able to
1010 // optimize out the ArrayCopy node later.
1011 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
1012 Node* ld_adr = in(MemNode::Address);
1013 intptr_t ld_off = 0;
1014 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
1015 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
1016 if (ac != nullptr) {
1017 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
1018
1019 Node* mem = ac->in(TypeFunc::Memory);
1020 Node* ctl = ac->in(0);
1021 Node* src = ac->in(ArrayCopyNode::Src);
1022
1023 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
1024 return nullptr;
1025 }
1026
1027 // load depends on the tests that validate the arraycopy
1028 LoadNode* ld = clone_pinned();
1029 Node* addp = in(MemNode::Address)->clone();
1030 if (ac->as_ArrayCopy()->is_clonebasic()) {
1031 assert(ld_alloc != nullptr, "need an alloc");
1032 assert(addp->is_AddP(), "address must be addp");
1033 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1034 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1035 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1036 addp->set_req(AddPNode::Base, src);
1037 addp->set_req(AddPNode::Address, src);
1038 } else {
1039 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1040 ac->as_ArrayCopy()->is_copyof_validated() ||
1041 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1042 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1043 addp->set_req(AddPNode::Base, src);
1044 addp->set_req(AddPNode::Address, src);
1045
1046 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1047 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
1048 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
1049
1050 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
1051 uint shift = exact_log2(type2aelembytes(ary_elem));
1052
1053 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1054 #ifdef _LP64
1055 diff = phase->transform(new ConvI2LNode(diff));
1056 #endif
1057 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1058
1059 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1060 addp->set_req(AddPNode::Offset, offset);
1061 }
1062 addp = phase->transform(addp);
1063 #ifdef ASSERT
1064 const TypePtr* adr_type = phase->type(addp)->is_ptr();
1065 ld->_adr_type = adr_type;
1066 #endif
1067 ld->set_req(MemNode::Address, addp);
1068 ld->set_req(0, ctl);
1069 ld->set_req(MemNode::Memory, mem);
1070 return ld;
1071 }
1072 return nullptr;
1073 }
1074
1075
1076 //---------------------------can_see_stored_value------------------------------
1077 // This routine exists to make sure this set of tests is done the same
1078 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
1079 // will change the graph shape in a way which makes memory alive twice at the
1080 // same time (uses the Oracle model of aliasing), then some
1081 // LoadXNode::Identity will fold things back to the equivalence-class model
1082 // of aliasing.
1083 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const {
1084 Node* ld_adr = in(MemNode::Address);
1085 intptr_t ld_off = 0;
1086 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1087 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base);
1088 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1089 Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr;
1090 // This is more general than load from boxing objects.
1091 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1092 uint alias_idx = atp->index();
1093 Node* result = nullptr;
1094 Node* current = st;
1095 // Skip through chains of MemBarNodes checking the MergeMems for
1096 // new states for the slice of this load. Stop once any other
1097 // kind of node is encountered. Loads from final memory can skip
1098 // through any kind of MemBar but normal loads shouldn't skip
1099 // through MemBarAcquire since the could allow them to move out of
1100 // a synchronized region. It is not safe to step over MemBarCPUOrder,
1101 // because alias info above them may be inaccurate (e.g., due to
1102 // mixed/mismatched unsafe accesses).
1103 bool is_final_mem = !atp->is_rewritable();
1104 while (current->is_Proj()) {
1105 int opc = current->in(0)->Opcode();
1106 if ((is_final_mem && (opc == Op_MemBarAcquire ||
1107 opc == Op_MemBarAcquireLock ||
1108 opc == Op_LoadFence)) ||
1109 opc == Op_MemBarRelease ||
1110 opc == Op_StoreFence ||
1111 opc == Op_MemBarReleaseLock ||
1112 opc == Op_MemBarStoreStore ||
1113 opc == Op_StoreStoreFence) {
1114 Node* mem = current->in(0)->in(TypeFunc::Memory);
1115 if (mem->is_MergeMem()) {
1116 MergeMemNode* merge = mem->as_MergeMem();
1117 Node* new_st = merge->memory_at(alias_idx);
1118 if (new_st == merge->base_memory()) {
1119 // Keep searching
1120 current = new_st;
1121 continue;
1122 }
1123 // Save the new memory state for the slice and fall through
1124 // to exit.
1125 result = new_st;
1126 }
1127 }
1128 break;
1129 }
1130 if (result != nullptr) {
1131 st = result;
1132 }
1133 }
1134
1135 // Loop around twice in the case Load -> Initialize -> Store.
1136 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1137 for (int trip = 0; trip <= 1; trip++) {
1138
1139 if (st->is_Store()) {
1140 Node* st_adr = st->in(MemNode::Address);
1141 if (st_adr != ld_adr) {
1142 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1143 intptr_t st_off = 0;
1144 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1145 if (ld_base == nullptr) return nullptr;
1146 if (st_base == nullptr) return nullptr;
1147 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr;
1148 if (ld_off != st_off) return nullptr;
1149 if (ld_off == Type::OffsetBot) return nullptr;
1150 // Same base, same offset.
1151 // Possible improvement for arrays: check index value instead of absolute offset.
1152
1153 // At this point we have proven something like this setup:
1154 // B = << base >>
1155 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1156 // S = StoreQ(AddP( B , #Off), V)
1157 // (Actually, we haven't yet proven the Q's are the same.)
1158 // In other words, we are loading from a casted version of
1159 // the same pointer-and-offset that we stored to.
1160 // Casted version may carry a dependency and it is respected.
1161 // Thus, we are able to replace L by V.
1162 }
1163 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1164 if (store_Opcode() != st->Opcode()) {
1165 return nullptr;
1166 }
1167 // LoadVector/StoreVector needs additional check to ensure the types match.
1168 if (st->is_StoreVector()) {
1169 const TypeVect* in_vt = st->as_StoreVector()->vect_type();
1170 const TypeVect* out_vt = as_LoadVector()->vect_type();
1171 if (in_vt != out_vt) {
1172 return nullptr;
1173 }
1174 }
1175 return st->in(MemNode::ValueIn);
1176 }
1177
1178 // A load from a freshly-created object always returns zero.
1179 // (This can happen after LoadNode::Ideal resets the load's memory input
1180 // to find_captured_store, which returned InitializeNode::zero_memory.)
1181 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1182 (st->in(0) == ld_alloc) &&
1183 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1184 // return a zero value for the load's basic type
1185 // (This is one of the few places where a generic PhaseTransform
1186 // can create new nodes. Think of it as lazily manifesting
1187 // virtually pre-existing constants.)
1188 if (memory_type() != T_VOID) {
1189 if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) {
1190 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1191 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1192 // by the ArrayCopyNode.
1193 return phase->zerocon(memory_type());
1194 }
1195 } else {
1196 // TODO: materialize all-zero vector constant
1197 assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1198 }
1199 }
1200
1201 // A load from an initialization barrier can match a captured store.
1202 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1203 InitializeNode* init = st->in(0)->as_Initialize();
1204 AllocateNode* alloc = init->allocation();
1205 if ((alloc != nullptr) && (alloc == ld_alloc)) {
1206 // examine a captured store value
1207 st = init->find_captured_store(ld_off, memory_size(), phase);
1208 if (st != nullptr) {
1209 continue; // take one more trip around
1210 }
1211 }
1212 }
1213
1214 // Load boxed value from result of valueOf() call is input parameter.
1215 if (this->is_Load() && ld_adr->is_AddP() &&
1216 (tp != nullptr) && tp->is_ptr_to_boxed_value()) {
1217 intptr_t ignore = 0;
1218 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1219 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1220 base = bs->step_over_gc_barrier(base);
1221 if (base != nullptr && base->is_Proj() &&
1222 base->as_Proj()->_con == TypeFunc::Parms &&
1223 base->in(0)->is_CallStaticJava() &&
1224 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1225 return base->in(0)->in(TypeFunc::Parms);
1226 }
1227 }
1228
1229 break;
1230 }
1231
1232 return nullptr;
1233 }
1234
1235 //----------------------is_instance_field_load_with_local_phi------------------
1236 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1237 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1238 in(Address)->is_AddP() ) {
1239 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1240 // Only instances and boxed values.
1241 if( t_oop != nullptr &&
1242 (t_oop->is_ptr_to_boxed_value() ||
1243 t_oop->is_known_instance_field()) &&
1244 t_oop->offset() != Type::OffsetBot &&
1245 t_oop->offset() != Type::OffsetTop) {
1246 return true;
1247 }
1248 }
1249 return false;
1250 }
1251
1252 //------------------------------Identity---------------------------------------
1253 // Loads are identity if previous store is to same address
1254 Node* LoadNode::Identity(PhaseGVN* phase) {
1255 // If the previous store-maker is the right kind of Store, and the store is
1256 // to the same address, then we are equal to the value stored.
1257 Node* mem = in(Memory);
1258 Node* value = can_see_stored_value(mem, phase);
1259 if( value ) {
1260 // byte, short & char stores truncate naturally.
1261 // A load has to load the truncated value which requires
1262 // some sort of masking operation and that requires an
1263 // Ideal call instead of an Identity call.
1264 if (memory_size() < BytesPerInt) {
1265 // If the input to the store does not fit with the load's result type,
1266 // it must be truncated via an Ideal call.
1267 if (!phase->type(value)->higher_equal(phase->type(this)))
1268 return this;
1269 }
1270 // (This works even when value is a Con, but LoadNode::Value
1271 // usually runs first, producing the singleton type of the Con.)
1272 if (!has_pinned_control_dependency() || value->is_Con()) {
1273 return value;
1274 } else {
1275 return this;
1276 }
1277 }
1278
1279 if (has_pinned_control_dependency()) {
1280 return this;
1281 }
1282 // Search for an existing data phi which was generated before for the same
1283 // instance's field to avoid infinite generation of phis in a loop.
1284 Node *region = mem->in(0);
1285 if (is_instance_field_load_with_local_phi(region)) {
1286 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1287 int this_index = phase->C->get_alias_index(addr_t);
1288 int this_offset = addr_t->offset();
1289 int this_iid = addr_t->instance_id();
1290 if (!addr_t->is_known_instance() &&
1291 addr_t->is_ptr_to_boxed_value()) {
1292 // Use _idx of address base (could be Phi node) for boxed values.
1293 intptr_t ignore = 0;
1294 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1295 if (base == nullptr) {
1296 return this;
1297 }
1298 this_iid = base->_idx;
1299 }
1300 const Type* this_type = bottom_type();
1301 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1302 Node* phi = region->fast_out(i);
1303 if (phi->is_Phi() && phi != mem &&
1304 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1305 return phi;
1306 }
1307 }
1308 }
1309
1310 return this;
1311 }
1312
1313 // Construct an equivalent unsigned load.
1314 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1315 BasicType bt = T_ILLEGAL;
1316 const Type* rt = nullptr;
1317 switch (Opcode()) {
1318 case Op_LoadUB: return this;
1319 case Op_LoadUS: return this;
1320 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1321 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1322 default:
1323 assert(false, "no unsigned variant: %s", Name());
1324 return nullptr;
1325 }
1326 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1327 raw_adr_type(), rt, bt, _mo, _control_dependency,
1328 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1329 }
1330
1331 // Construct an equivalent signed load.
1332 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1333 BasicType bt = T_ILLEGAL;
1334 const Type* rt = nullptr;
1335 switch (Opcode()) {
1336 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1337 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1338 case Op_LoadB: // fall through
1339 case Op_LoadS: // fall through
1340 case Op_LoadI: // fall through
1341 case Op_LoadL: return this;
1342 default:
1343 assert(false, "no signed variant: %s", Name());
1344 return nullptr;
1345 }
1346 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1347 raw_adr_type(), rt, bt, _mo, _control_dependency,
1348 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1349 }
1350
1351 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1352 BasicType bt = rt->basic_type();
1353 switch (Opcode()) {
1354 case Op_LoadI: return (bt == T_FLOAT);
1355 case Op_LoadL: return (bt == T_DOUBLE);
1356 case Op_LoadF: return (bt == T_INT);
1357 case Op_LoadD: return (bt == T_LONG);
1358
1359 default: return false;
1360 }
1361 }
1362
1363 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1364 BasicType bt = rt->basic_type();
1365 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1366 bool is_mismatched = is_mismatched_access();
1367 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1368 if (raw_type == nullptr) {
1369 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1370 }
1371 const int op = Opcode();
1372 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) ||
1373 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access());
1374 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1375 raw_adr_type(), rt, bt, _mo, _control_dependency,
1376 require_atomic_access, is_unaligned_access(), is_mismatched);
1377 }
1378
1379 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1380 BasicType bt = vt->basic_type();
1381 switch (Opcode()) {
1382 case Op_StoreI: return (bt == T_FLOAT);
1383 case Op_StoreL: return (bt == T_DOUBLE);
1384 case Op_StoreF: return (bt == T_INT);
1385 case Op_StoreD: return (bt == T_LONG);
1386
1387 default: return false;
1388 }
1389 }
1390
1391 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1392 BasicType bt = vt->basic_type();
1393 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1394 const int op = Opcode();
1395 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) ||
1396 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access());
1397 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1398 raw_adr_type(), val, bt, _mo, require_atomic_access);
1399
1400 bool is_mismatched = is_mismatched_access();
1401 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1402 if (raw_type == nullptr) {
1403 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1404 }
1405 if (is_mismatched) {
1406 st->set_mismatched_access();
1407 }
1408 return st;
1409 }
1410
1411 // We're loading from an object which has autobox behaviour.
1412 // If this object is result of a valueOf call we'll have a phi
1413 // merging a newly allocated object and a load from the cache.
1414 // We want to replace this load with the original incoming
1415 // argument to the valueOf call.
1416 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1417 assert(igvn->C->eliminate_boxing(), "sanity");
1418 intptr_t ignore = 0;
1419 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1420 if ((base == nullptr) || base->is_Phi()) {
1421 // Push the loads from the phi that comes from valueOf up
1422 // through it to allow elimination of the loads and the recovery
1423 // of the original value. It is done in split_through_phi().
1424 return nullptr;
1425 } else if (base->is_Load() ||
1426 (base->is_DecodeN() && base->in(1)->is_Load())) {
1427 // Eliminate the load of boxed value for integer types from the cache
1428 // array by deriving the value from the index into the array.
1429 // Capture the offset of the load and then reverse the computation.
1430
1431 // Get LoadN node which loads a boxing object from 'cache' array.
1432 if (base->is_DecodeN()) {
1433 base = base->in(1);
1434 }
1435 if (!base->in(Address)->is_AddP()) {
1436 return nullptr; // Complex address
1437 }
1438 AddPNode* address = base->in(Address)->as_AddP();
1439 Node* cache_base = address->in(AddPNode::Base);
1440 if ((cache_base != nullptr) && cache_base->is_DecodeN()) {
1441 // Get ConP node which is static 'cache' field.
1442 cache_base = cache_base->in(1);
1443 }
1444 if ((cache_base != nullptr) && cache_base->is_Con()) {
1445 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1446 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
1447 Node* elements[4];
1448 int shift = exact_log2(type2aelembytes(T_OBJECT));
1449 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1450 if (count > 0 && elements[0]->is_Con() &&
1451 (count == 1 ||
1452 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1453 elements[1]->in(2) == igvn->intcon(shift)))) {
1454 ciObjArray* array = base_type->const_oop()->as_obj_array();
1455 // Fetch the box object cache[0] at the base of the array and get its value
1456 ciInstance* box = array->obj_at(0)->as_instance();
1457 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1458 assert(ik->is_box_klass(), "sanity");
1459 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1460 if (ik->nof_nonstatic_fields() == 1) {
1461 // This should be true nonstatic_field_at requires calling
1462 // nof_nonstatic_fields so check it anyway
1463 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1464 BasicType bt = c.basic_type();
1465 // Only integer types have boxing cache.
1466 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1467 bt == T_BYTE || bt == T_SHORT ||
1468 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1469 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1470 if (cache_low != (int)cache_low) {
1471 return nullptr; // should not happen since cache is array indexed by value
1472 }
1473 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1474 if (offset != (int)offset) {
1475 return nullptr; // should not happen since cache is array indexed by value
1476 }
1477 // Add up all the offsets making of the address of the load
1478 Node* result = elements[0];
1479 for (int i = 1; i < count; i++) {
1480 result = igvn->transform(new AddXNode(result, elements[i]));
1481 }
1482 // Remove the constant offset from the address and then
1483 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1484 // remove the scaling of the offset to recover the original index.
1485 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1486 // Peel the shift off directly but wrap it in a dummy node
1487 // since Ideal can't return existing nodes
1488 igvn->_worklist.push(result); // remove dead node later
1489 result = new RShiftXNode(result->in(1), igvn->intcon(0));
1490 } else if (result->is_Add() && result->in(2)->is_Con() &&
1491 result->in(1)->Opcode() == Op_LShiftX &&
1492 result->in(1)->in(2) == igvn->intcon(shift)) {
1493 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1494 // but for boxing cache access we know that X<<Z will not overflow
1495 // (there is range check) so we do this optimizatrion by hand here.
1496 igvn->_worklist.push(result); // remove dead node later
1497 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1498 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1499 } else {
1500 result = new RShiftXNode(result, igvn->intcon(shift));
1501 }
1502 #ifdef _LP64
1503 if (bt != T_LONG) {
1504 result = new ConvL2INode(igvn->transform(result));
1505 }
1506 #else
1507 if (bt == T_LONG) {
1508 result = new ConvI2LNode(igvn->transform(result));
1509 }
1510 #endif
1511 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1512 // Need to preserve unboxing load type if it is unsigned.
1513 switch(this->Opcode()) {
1514 case Op_LoadUB:
1515 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1516 break;
1517 case Op_LoadUS:
1518 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1519 break;
1520 }
1521 return result;
1522 }
1523 }
1524 }
1525 }
1526 }
1527 return nullptr;
1528 }
1529
1530 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1531 Node* region = phi->in(0);
1532 if (region == nullptr) {
1533 return false; // Wait stable graph
1534 }
1535 uint cnt = phi->req();
1536 for (uint i = 1; i < cnt; i++) {
1537 Node* rc = region->in(i);
1538 if (rc == nullptr || phase->type(rc) == Type::TOP)
1539 return false; // Wait stable graph
1540 Node* in = phi->in(i);
1541 if (in == nullptr || phase->type(in) == Type::TOP)
1542 return false; // Wait stable graph
1543 }
1544 return true;
1545 }
1546
1547 //------------------------------split_through_phi------------------------------
1548 // Check whether a call to 'split_through_phi' would split this load through the
1549 // Phi *base*. This method is essentially a copy of the validations performed
1550 // by 'split_through_phi'. The first use of this method was in EA code as part
1551 // of simplification of allocation merges.
1552 // Some differences from original method (split_through_phi):
1553 // - If base->is_CastPP(): base = base->in(1)
1554 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) {
1555 Node* mem = in(Memory);
1556 Node* address = in(Address);
1557 intptr_t ignore = 0;
1558 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1559
1560 if (base->is_CastPP()) {
1561 base = base->in(1);
1562 }
1563
1564 if (req() > 3 || base == nullptr || !base->is_Phi()) {
1565 return false;
1566 }
1567
1568 if (!mem->is_Phi()) {
1569 if (!MemNode::all_controls_dominate(mem, base->in(0)))
1570 return false;
1571 } else if (base->in(0) != mem->in(0)) {
1572 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1573 return false;
1574 }
1575 }
1576
1577 return true;
1578 }
1579
1580 //------------------------------split_through_phi------------------------------
1581 // Split instance or boxed field load through Phi.
1582 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) {
1583 if (req() > 3) {
1584 assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs");
1585 // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account
1586 return nullptr;
1587 }
1588 Node* mem = in(Memory);
1589 Node* address = in(Address);
1590 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1591
1592 assert((t_oop != nullptr) &&
1593 (ignore_missing_instance_id ||
1594 t_oop->is_known_instance_field() ||
1595 t_oop->is_ptr_to_boxed_value()), "invalid conditions");
1596
1597 Compile* C = phase->C;
1598 intptr_t ignore = 0;
1599 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1600 bool base_is_phi = (base != nullptr) && base->is_Phi();
1601 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1602 (base != nullptr) && (base == address->in(AddPNode::Base)) &&
1603 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1604
1605 if (!((mem->is_Phi() || base_is_phi) &&
1606 (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) {
1607 return nullptr; // Neither memory or base are Phi
1608 }
1609
1610 if (mem->is_Phi()) {
1611 if (!stable_phi(mem->as_Phi(), phase)) {
1612 return nullptr; // Wait stable graph
1613 }
1614 uint cnt = mem->req();
1615 // Check for loop invariant memory.
1616 if (cnt == 3) {
1617 for (uint i = 1; i < cnt; i++) {
1618 Node* in = mem->in(i);
1619 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1620 if (m == mem) {
1621 if (i == 1) {
1622 // if the first edge was a loop, check second edge too.
1623 // If both are replaceable - we are in an infinite loop
1624 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1625 if (n == mem) {
1626 break;
1627 }
1628 }
1629 set_req(Memory, mem->in(cnt - i));
1630 return this; // made change
1631 }
1632 }
1633 }
1634 }
1635 if (base_is_phi) {
1636 if (!stable_phi(base->as_Phi(), phase)) {
1637 return nullptr; // Wait stable graph
1638 }
1639 uint cnt = base->req();
1640 // Check for loop invariant memory.
1641 if (cnt == 3) {
1642 for (uint i = 1; i < cnt; i++) {
1643 if (base->in(i) == base) {
1644 return nullptr; // Wait stable graph
1645 }
1646 }
1647 }
1648 }
1649
1650 // Split through Phi (see original code in loopopts.cpp).
1651 assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type");
1652
1653 // Do nothing here if Identity will find a value
1654 // (to avoid infinite chain of value phis generation).
1655 if (this != Identity(phase)) {
1656 return nullptr;
1657 }
1658
1659 // Select Region to split through.
1660 Node* region;
1661 if (!base_is_phi) {
1662 assert(mem->is_Phi(), "sanity");
1663 region = mem->in(0);
1664 // Skip if the region dominates some control edge of the address.
1665 if (!MemNode::all_controls_dominate(address, region))
1666 return nullptr;
1667 } else if (!mem->is_Phi()) {
1668 assert(base_is_phi, "sanity");
1669 region = base->in(0);
1670 // Skip if the region dominates some control edge of the memory.
1671 if (!MemNode::all_controls_dominate(mem, region))
1672 return nullptr;
1673 } else if (base->in(0) != mem->in(0)) {
1674 assert(base_is_phi && mem->is_Phi(), "sanity");
1675 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1676 region = base->in(0);
1677 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1678 region = mem->in(0);
1679 } else {
1680 return nullptr; // complex graph
1681 }
1682 } else {
1683 assert(base->in(0) == mem->in(0), "sanity");
1684 region = mem->in(0);
1685 }
1686
1687 Node* phi = nullptr;
1688 const Type* this_type = this->bottom_type();
1689 PhaseIterGVN* igvn = phase->is_IterGVN();
1690 if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) {
1691 int this_index = C->get_alias_index(t_oop);
1692 int this_offset = t_oop->offset();
1693 int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx;
1694 phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset);
1695 } else if (ignore_missing_instance_id) {
1696 phi = new PhiNode(region, this_type, nullptr, mem->_idx);
1697 } else {
1698 return nullptr;
1699 }
1700
1701 for (uint i = 1; i < region->req(); i++) {
1702 Node* x;
1703 Node* the_clone = nullptr;
1704 Node* in = region->in(i);
1705 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1706 in != nullptr && in->is_OuterStripMinedLoop()) {
1707 // No node should go in the outer strip mined loop
1708 in = in->in(LoopNode::EntryControl);
1709 }
1710 if (in == nullptr || in == C->top()) {
1711 x = C->top(); // Dead path? Use a dead data op
1712 } else {
1713 x = this->clone(); // Else clone up the data op
1714 the_clone = x; // Remember for possible deletion.
1715 // Alter data node to use pre-phi inputs
1716 if (this->in(0) == region) {
1717 x->set_req(0, in);
1718 } else {
1719 x->set_req(0, nullptr);
1720 }
1721 if (mem->is_Phi() && (mem->in(0) == region)) {
1722 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1723 }
1724 if (address->is_Phi() && address->in(0) == region) {
1725 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1726 }
1727 if (base_is_phi && (base->in(0) == region)) {
1728 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1729 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1730 x->set_req(Address, adr_x);
1731 }
1732 }
1733 // Check for a 'win' on some paths
1734 const Type *t = x->Value(igvn);
1735
1736 bool singleton = t->singleton();
1737
1738 // See comments in PhaseIdealLoop::split_thru_phi().
1739 if (singleton && t == Type::TOP) {
1740 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1741 }
1742
1743 if (singleton) {
1744 x = igvn->makecon(t);
1745 } else {
1746 // We now call Identity to try to simplify the cloned node.
1747 // Note that some Identity methods call phase->type(this).
1748 // Make sure that the type array is big enough for
1749 // our new node, even though we may throw the node away.
1750 // (This tweaking with igvn only works because x is a new node.)
1751 igvn->set_type(x, t);
1752 // If x is a TypeNode, capture any more-precise type permanently into Node
1753 // otherwise it will be not updated during igvn->transform since
1754 // igvn->type(x) is set to x->Value() already.
1755 x->raise_bottom_type(t);
1756 Node* y = x->Identity(igvn);
1757 if (y != x) {
1758 x = y;
1759 } else {
1760 y = igvn->hash_find_insert(x);
1761 if (y) {
1762 x = y;
1763 } else {
1764 // Else x is a new node we are keeping
1765 // We do not need register_new_node_with_optimizer
1766 // because set_type has already been called.
1767 igvn->_worklist.push(x);
1768 }
1769 }
1770 }
1771 if (x != the_clone && the_clone != nullptr) {
1772 igvn->remove_dead_node(the_clone);
1773 }
1774 phi->set_req(i, x);
1775 }
1776 // Record Phi
1777 igvn->register_new_node_with_optimizer(phi);
1778 return phi;
1779 }
1780
1781 AllocateNode* LoadNode::is_new_object_mark_load() const {
1782 if (Opcode() == Op_LoadX) {
1783 Node* address = in(MemNode::Address);
1784 AllocateNode* alloc = AllocateNode::Ideal_allocation(address);
1785 Node* mem = in(MemNode::Memory);
1786 if (alloc != nullptr && mem->is_Proj() &&
1787 mem->in(0) != nullptr &&
1788 mem->in(0) == alloc->initialization() &&
1789 alloc->initialization()->proj_out_or_null(0) != nullptr) {
1790 return alloc;
1791 }
1792 }
1793 return nullptr;
1794 }
1795
1796
1797 //------------------------------Ideal------------------------------------------
1798 // If the load is from Field memory and the pointer is non-null, it might be possible to
1799 // zero out the control input.
1800 // If the offset is constant and the base is an object allocation,
1801 // try to hook me up to the exact initializing store.
1802 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1803 if (has_pinned_control_dependency()) {
1804 return nullptr;
1805 }
1806 Node* p = MemNode::Ideal_common(phase, can_reshape);
1807 if (p) return (p == NodeSentinel) ? nullptr : p;
1808
1809 Node* ctrl = in(MemNode::Control);
1810 Node* address = in(MemNode::Address);
1811 bool progress = false;
1812
1813 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1814 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1815
1816 // Skip up past a SafePoint control. Cannot do this for Stores because
1817 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1818 if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint &&
1819 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1820 !addr_mark &&
1821 (depends_only_on_test() || has_unknown_control_dependency())) {
1822 ctrl = ctrl->in(0);
1823 set_req(MemNode::Control,ctrl);
1824 progress = true;
1825 }
1826
1827 intptr_t ignore = 0;
1828 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1829 if (base != nullptr
1830 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1831 // Check for useless control edge in some common special cases
1832 if (in(MemNode::Control) != nullptr
1833 && can_remove_control()
1834 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1835 && all_controls_dominate(base, phase->C->start())) {
1836 // A method-invariant, non-null address (constant or 'this' argument).
1837 set_req(MemNode::Control, nullptr);
1838 progress = true;
1839 }
1840 }
1841
1842 Node* mem = in(MemNode::Memory);
1843 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1844
1845 if (can_reshape && (addr_t != nullptr)) {
1846 // try to optimize our memory input
1847 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1848 if (opt_mem != mem) {
1849 set_req_X(MemNode::Memory, opt_mem, phase);
1850 if (phase->type( opt_mem ) == Type::TOP) return nullptr;
1851 return this;
1852 }
1853 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1854 if ((t_oop != nullptr) &&
1855 (t_oop->is_known_instance_field() ||
1856 t_oop->is_ptr_to_boxed_value())) {
1857 PhaseIterGVN *igvn = phase->is_IterGVN();
1858 assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true");
1859 if (igvn->_worklist.member(opt_mem)) {
1860 // Delay this transformation until memory Phi is processed.
1861 igvn->_worklist.push(this);
1862 return nullptr;
1863 }
1864 // Split instance field load through Phi.
1865 Node* result = split_through_phi(phase);
1866 if (result != nullptr) return result;
1867
1868 if (t_oop->is_ptr_to_boxed_value()) {
1869 Node* result = eliminate_autobox(igvn);
1870 if (result != nullptr) return result;
1871 }
1872 }
1873 }
1874
1875 // Is there a dominating load that loads the same value? Leave
1876 // anything that is not a load of a field/array element (like
1877 // barriers etc.) alone
1878 if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) {
1879 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1880 Node *use = mem->fast_out(i);
1881 if (use != this &&
1882 use->Opcode() == Opcode() &&
1883 use->in(0) != nullptr &&
1884 use->in(0) != in(0) &&
1885 use->in(Address) == in(Address)) {
1886 Node* ctl = in(0);
1887 for (int i = 0; i < 10 && ctl != nullptr; i++) {
1888 ctl = IfNode::up_one_dom(ctl);
1889 if (ctl == use->in(0)) {
1890 set_req(0, use->in(0));
1891 return this;
1892 }
1893 }
1894 }
1895 }
1896 }
1897
1898 // Check for prior store with a different base or offset; make Load
1899 // independent. Skip through any number of them. Bail out if the stores
1900 // are in an endless dead cycle and report no progress. This is a key
1901 // transform for Reflection. However, if after skipping through the Stores
1902 // we can't then fold up against a prior store do NOT do the transform as
1903 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1904 // array memory alive twice: once for the hoisted Load and again after the
1905 // bypassed Store. This situation only works if EVERYBODY who does
1906 // anti-dependence work knows how to bypass. I.e. we need all
1907 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1908 // the alias index stuff. So instead, peek through Stores and IFF we can
1909 // fold up, do so.
1910 Node* prev_mem = find_previous_store(phase);
1911 if (prev_mem != nullptr) {
1912 Node* value = can_see_arraycopy_value(prev_mem, phase);
1913 if (value != nullptr) {
1914 return value;
1915 }
1916 }
1917 // Steps (a), (b): Walk past independent stores to find an exact match.
1918 if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) {
1919 // (c) See if we can fold up on the spot, but don't fold up here.
1920 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1921 // just return a prior value, which is done by Identity calls.
1922 if (can_see_stored_value(prev_mem, phase)) {
1923 // Make ready for step (d):
1924 set_req_X(MemNode::Memory, prev_mem, phase);
1925 return this;
1926 }
1927 }
1928
1929 return progress ? this : nullptr;
1930 }
1931
1932 // Helper to recognize certain Klass fields which are invariant across
1933 // some group of array types (e.g., int[] or all T[] where T < Object).
1934 const Type*
1935 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1936 ciKlass* klass) const {
1937 if (UseCompactObjectHeaders) {
1938 if (tkls->offset() == in_bytes(Klass::prototype_header_offset())) {
1939 // The field is Klass::_prototype_header. Return its (constant) value.
1940 assert(this->Opcode() == Op_LoadX, "must load a proper type from _prototype_header");
1941 return TypeX::make(klass->prototype_header());
1942 }
1943 }
1944 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1945 // The field is Klass::_modifier_flags. Return its (constant) value.
1946 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1947 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1948 return TypeInt::make(klass->modifier_flags());
1949 }
1950 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1951 // The field is Klass::_access_flags. Return its (constant) value.
1952 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1953 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1954 return TypeInt::make(klass->access_flags());
1955 }
1956 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1957 // The field is Klass::_layout_helper. Return its constant value if known.
1958 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1959 return TypeInt::make(klass->layout_helper());
1960 }
1961
1962 // No match.
1963 return nullptr;
1964 }
1965
1966 //------------------------------Value-----------------------------------------
1967 const Type* LoadNode::Value(PhaseGVN* phase) const {
1968 // Either input is TOP ==> the result is TOP
1969 Node* mem = in(MemNode::Memory);
1970 const Type *t1 = phase->type(mem);
1971 if (t1 == Type::TOP) return Type::TOP;
1972 Node* adr = in(MemNode::Address);
1973 const TypePtr* tp = phase->type(adr)->isa_ptr();
1974 if (tp == nullptr || tp->empty()) return Type::TOP;
1975 int off = tp->offset();
1976 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1977 Compile* C = phase->C;
1978
1979 // Try to guess loaded type from pointer type
1980 if (tp->isa_aryptr()) {
1981 const TypeAryPtr* ary = tp->is_aryptr();
1982 const Type* t = ary->elem();
1983
1984 // Determine whether the reference is beyond the header or not, by comparing
1985 // the offset against the offset of the start of the array's data.
1986 // Different array types begin at slightly different offsets (12 vs. 16).
1987 // We choose T_BYTE as an example base type that is least restrictive
1988 // as to alignment, which will therefore produce the smallest
1989 // possible base offset.
1990 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1991 const bool off_beyond_header = (off >= min_base_off);
1992
1993 // Try to constant-fold a stable array element.
1994 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1995 // Make sure the reference is not into the header and the offset is constant
1996 ciObject* aobj = ary->const_oop();
1997 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1998 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1999 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
2000 stable_dimension,
2001 memory_type(), is_unsigned());
2002 if (con_type != nullptr) {
2003 return con_type;
2004 }
2005 }
2006 }
2007
2008 // Don't do this for integer types. There is only potential profit if
2009 // the element type t is lower than _type; that is, for int types, if _type is
2010 // more restrictive than t. This only happens here if one is short and the other
2011 // char (both 16 bits), and in those cases we've made an intentional decision
2012 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2013 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2014 //
2015 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2016 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
2017 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2018 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
2019 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2020 // In fact, that could have been the original type of p1, and p1 could have
2021 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2022 // expression (LShiftL quux 3) independently optimized to the constant 8.
2023 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2024 && (_type->isa_vect() == nullptr)
2025 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2026 // t might actually be lower than _type, if _type is a unique
2027 // concrete subclass of abstract class t.
2028 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
2029 const Type* jt = t->join_speculative(_type);
2030 // In any case, do not allow the join, per se, to empty out the type.
2031 if (jt->empty() && !t->empty()) {
2032 // This can happen if a interface-typed array narrows to a class type.
2033 jt = _type;
2034 }
2035 #ifdef ASSERT
2036 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2037 // The pointers in the autobox arrays are always non-null
2038 Node* base = adr->in(AddPNode::Base);
2039 if ((base != nullptr) && base->is_DecodeN()) {
2040 // Get LoadN node which loads IntegerCache.cache field
2041 base = base->in(1);
2042 }
2043 if ((base != nullptr) && base->is_Con()) {
2044 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2045 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2046 // It could be narrow oop
2047 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2048 }
2049 }
2050 }
2051 #endif
2052 return jt;
2053 }
2054 }
2055 } else if (tp->base() == Type::InstPtr) {
2056 assert( off != Type::OffsetBot ||
2057 // arrays can be cast to Objects
2058 !tp->isa_instptr() ||
2059 tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
2060 // unsafe field access may not have a constant offset
2061 C->has_unsafe_access(),
2062 "Field accesses must be precise" );
2063 // For oop loads, we expect the _type to be precise.
2064
2065 // Optimize loads from constant fields.
2066 const TypeInstPtr* tinst = tp->is_instptr();
2067 ciObject* const_oop = tinst->const_oop();
2068 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2069 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
2070 if (con_type != nullptr) {
2071 return con_type;
2072 }
2073 }
2074 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2075 assert(off != Type::OffsetBot ||
2076 !tp->isa_instklassptr() ||
2077 // arrays can be cast to Objects
2078 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2079 // also allow array-loading from the primary supertype
2080 // array during subtype checks
2081 Opcode() == Op_LoadKlass,
2082 "Field accesses must be precise");
2083 // For klass/static loads, we expect the _type to be precise
2084 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2085 /* With mirrors being an indirect in the Klass*
2086 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2087 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2088 *
2089 * So check the type and klass of the node before the LoadP.
2090 */
2091 Node* adr2 = adr->in(MemNode::Address);
2092 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2093 if (tkls != nullptr && !StressReflectiveCode) {
2094 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2095 ciKlass* klass = tkls->exact_klass();
2096 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2097 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2098 return TypeInstPtr::make(klass->java_mirror());
2099 }
2100 }
2101 }
2102
2103 const TypeKlassPtr *tkls = tp->isa_klassptr();
2104 if (tkls != nullptr) {
2105 if (tkls->is_loaded() && tkls->klass_is_exact()) {
2106 ciKlass* klass = tkls->exact_klass();
2107 // We are loading a field from a Klass metaobject whose identity
2108 // is known at compile time (the type is "exact" or "precise").
2109 // Check for fields we know are maintained as constants by the VM.
2110 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2111 // The field is Klass::_super_check_offset. Return its (constant) value.
2112 // (Folds up type checking code.)
2113 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2114 return TypeInt::make(klass->super_check_offset());
2115 }
2116 if (UseCompactObjectHeaders) {
2117 if (tkls->offset() == in_bytes(Klass::prototype_header_offset())) {
2118 // The field is Klass::_prototype_header. Return its (constant) value.
2119 assert(this->Opcode() == Op_LoadX, "must load a proper type from _prototype_header");
2120 return TypeX::make(klass->prototype_header());
2121 }
2122 }
2123 // Compute index into primary_supers array
2124 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2125 // Check for overflowing; use unsigned compare to handle the negative case.
2126 if( depth < ciKlass::primary_super_limit() ) {
2127 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2128 // (Folds up type checking code.)
2129 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2130 ciKlass *ss = klass->super_of_depth(depth);
2131 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2132 }
2133 const Type* aift = load_array_final_field(tkls, klass);
2134 if (aift != nullptr) return aift;
2135 }
2136
2137 // We can still check if we are loading from the primary_supers array at a
2138 // shallow enough depth. Even though the klass is not exact, entries less
2139 // than or equal to its super depth are correct.
2140 if (tkls->is_loaded()) {
2141 ciKlass* klass = nullptr;
2142 if (tkls->isa_instklassptr()) {
2143 klass = tkls->is_instklassptr()->instance_klass();
2144 } else {
2145 int dims;
2146 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims);
2147 if (inner->isa_instklassptr()) {
2148 klass = inner->is_instklassptr()->instance_klass();
2149 klass = ciObjArrayKlass::make(klass, dims);
2150 }
2151 }
2152 if (klass != nullptr) {
2153 // Compute index into primary_supers array
2154 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2155 // Check for overflowing; use unsigned compare to handle the negative case.
2156 if (depth < ciKlass::primary_super_limit() &&
2157 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case
2158 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2159 // (Folds up type checking code.)
2160 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2161 ciKlass *ss = klass->super_of_depth(depth);
2162 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2163 }
2164 }
2165 }
2166
2167 // If the type is enough to determine that the thing is not an array,
2168 // we can give the layout_helper a positive interval type.
2169 // This will help short-circuit some reflective code.
2170 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) &&
2171 tkls->isa_instklassptr() && // not directly typed as an array
2172 !tkls->is_instklassptr()->instance_klass()->is_java_lang_Object() // not the supertype of all T[] and specifically not Serializable & Cloneable
2173 ) {
2174 // Note: When interfaces are reliable, we can narrow the interface
2175 // test to (klass != Serializable && klass != Cloneable).
2176 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2177 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2178 // The key property of this type is that it folds up tests
2179 // for array-ness, since it proves that the layout_helper is positive.
2180 // Thus, a generic value like the basic object layout helper works fine.
2181 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2182 }
2183 }
2184
2185 // If we are loading from a freshly-allocated object, produce a zero,
2186 // if the load is provably beyond the header of the object.
2187 // (Also allow a variable load from a fresh array to produce zero.)
2188 const TypeOopPtr *tinst = tp->isa_oopptr();
2189 bool is_instance = (tinst != nullptr) && tinst->is_known_instance_field();
2190 bool is_boxed_value = (tinst != nullptr) && tinst->is_ptr_to_boxed_value();
2191 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
2192 Node* value = can_see_stored_value(mem,phase);
2193 if (value != nullptr && value->is_Con()) {
2194 assert(value->bottom_type()->higher_equal(_type),"sanity");
2195 return value->bottom_type();
2196 }
2197 }
2198
2199 bool is_vect = (_type->isa_vect() != nullptr);
2200 if (is_instance && !is_vect) {
2201 // If we have an instance type and our memory input is the
2202 // programs's initial memory state, there is no matching store,
2203 // so just return a zero of the appropriate type -
2204 // except if it is vectorized - then we have no zero constant.
2205 Node *mem = in(MemNode::Memory);
2206 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2207 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2208 return Type::get_zero_type(_type->basic_type());
2209 }
2210 }
2211
2212 Node* alloc = is_new_object_mark_load();
2213 if (!UseCompactObjectHeaders && alloc != nullptr) {
2214 return TypeX::make(markWord::prototype().value());
2215 }
2216
2217 return _type;
2218 }
2219
2220 //------------------------------match_edge-------------------------------------
2221 // Do we Match on this edge index or not? Match only the address.
2222 uint LoadNode::match_edge(uint idx) const {
2223 return idx == MemNode::Address;
2224 }
2225
2226 //--------------------------LoadBNode::Ideal--------------------------------------
2227 //
2228 // If the previous store is to the same address as this load,
2229 // and the value stored was larger than a byte, replace this load
2230 // with the value stored truncated to a byte. If no truncation is
2231 // needed, the replacement is done in LoadNode::Identity().
2232 //
2233 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2234 Node* mem = in(MemNode::Memory);
2235 Node* value = can_see_stored_value(mem,phase);
2236 if (value != nullptr) {
2237 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2238 if (narrow != value) {
2239 return narrow;
2240 }
2241 }
2242 // Identity call will handle the case where truncation is not needed.
2243 return LoadNode::Ideal(phase, can_reshape);
2244 }
2245
2246 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2247 Node* mem = in(MemNode::Memory);
2248 Node* value = can_see_stored_value(mem,phase);
2249 if (value != nullptr && value->is_Con() &&
2250 !value->bottom_type()->higher_equal(_type)) {
2251 // If the input to the store does not fit with the load's result type,
2252 // it must be truncated. We can't delay until Ideal call since
2253 // a singleton Value is needed for split_thru_phi optimization.
2254 int con = value->get_int();
2255 return TypeInt::make((con << 24) >> 24);
2256 }
2257 return LoadNode::Value(phase);
2258 }
2259
2260 //--------------------------LoadUBNode::Ideal-------------------------------------
2261 //
2262 // If the previous store is to the same address as this load,
2263 // and the value stored was larger than a byte, replace this load
2264 // with the value stored truncated to a byte. If no truncation is
2265 // needed, the replacement is done in LoadNode::Identity().
2266 //
2267 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2268 Node* mem = in(MemNode::Memory);
2269 Node* value = can_see_stored_value(mem, phase);
2270 if (value != nullptr) {
2271 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2272 if (narrow != value) {
2273 return narrow;
2274 }
2275 }
2276 // Identity call will handle the case where truncation is not needed.
2277 return LoadNode::Ideal(phase, can_reshape);
2278 }
2279
2280 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2281 Node* mem = in(MemNode::Memory);
2282 Node* value = can_see_stored_value(mem,phase);
2283 if (value != nullptr && value->is_Con() &&
2284 !value->bottom_type()->higher_equal(_type)) {
2285 // If the input to the store does not fit with the load's result type,
2286 // it must be truncated. We can't delay until Ideal call since
2287 // a singleton Value is needed for split_thru_phi optimization.
2288 int con = value->get_int();
2289 return TypeInt::make(con & 0xFF);
2290 }
2291 return LoadNode::Value(phase);
2292 }
2293
2294 //--------------------------LoadUSNode::Ideal-------------------------------------
2295 //
2296 // If the previous store is to the same address as this load,
2297 // and the value stored was larger than a char, replace this load
2298 // with the value stored truncated to a char. If no truncation is
2299 // needed, the replacement is done in LoadNode::Identity().
2300 //
2301 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2302 Node* mem = in(MemNode::Memory);
2303 Node* value = can_see_stored_value(mem,phase);
2304 if (value != nullptr) {
2305 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2306 if (narrow != value) {
2307 return narrow;
2308 }
2309 }
2310 // Identity call will handle the case where truncation is not needed.
2311 return LoadNode::Ideal(phase, can_reshape);
2312 }
2313
2314 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2315 Node* mem = in(MemNode::Memory);
2316 Node* value = can_see_stored_value(mem,phase);
2317 if (value != nullptr && value->is_Con() &&
2318 !value->bottom_type()->higher_equal(_type)) {
2319 // If the input to the store does not fit with the load's result type,
2320 // it must be truncated. We can't delay until Ideal call since
2321 // a singleton Value is needed for split_thru_phi optimization.
2322 int con = value->get_int();
2323 return TypeInt::make(con & 0xFFFF);
2324 }
2325 return LoadNode::Value(phase);
2326 }
2327
2328 //--------------------------LoadSNode::Ideal--------------------------------------
2329 //
2330 // If the previous store is to the same address as this load,
2331 // and the value stored was larger than a short, replace this load
2332 // with the value stored truncated to a short. If no truncation is
2333 // needed, the replacement is done in LoadNode::Identity().
2334 //
2335 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2336 Node* mem = in(MemNode::Memory);
2337 Node* value = can_see_stored_value(mem,phase);
2338 if (value != nullptr) {
2339 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2340 if (narrow != value) {
2341 return narrow;
2342 }
2343 }
2344 // Identity call will handle the case where truncation is not needed.
2345 return LoadNode::Ideal(phase, can_reshape);
2346 }
2347
2348 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2349 Node* mem = in(MemNode::Memory);
2350 Node* value = can_see_stored_value(mem,phase);
2351 if (value != nullptr && value->is_Con() &&
2352 !value->bottom_type()->higher_equal(_type)) {
2353 // If the input to the store does not fit with the load's result type,
2354 // it must be truncated. We can't delay until Ideal call since
2355 // a singleton Value is needed for split_thru_phi optimization.
2356 int con = value->get_int();
2357 return TypeInt::make((con << 16) >> 16);
2358 }
2359 return LoadNode::Value(phase);
2360 }
2361
2362 //=============================================================================
2363 //----------------------------LoadKlassNode::make------------------------------
2364 // Polymorphic factory method:
2365 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2366 // sanity check the alias category against the created node type
2367 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2368 assert(adr_type != nullptr, "expecting TypeKlassPtr");
2369 #ifdef _LP64
2370 if (adr_type->is_ptr_to_narrowklass()) {
2371 assert(UseCompressedClassPointers, "no compressed klasses");
2372 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2373 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2374 }
2375 #endif
2376 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2377 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2378 }
2379
2380 //------------------------------Value------------------------------------------
2381 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2382 return klass_value_common(phase);
2383 }
2384
2385 // In most cases, LoadKlassNode does not have the control input set. If the control
2386 // input is set, it must not be removed (by LoadNode::Ideal()).
2387 bool LoadKlassNode::can_remove_control() const {
2388 return false;
2389 }
2390
2391 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2392 // Either input is TOP ==> the result is TOP
2393 const Type *t1 = phase->type( in(MemNode::Memory) );
2394 if (t1 == Type::TOP) return Type::TOP;
2395 Node *adr = in(MemNode::Address);
2396 const Type *t2 = phase->type( adr );
2397 if (t2 == Type::TOP) return Type::TOP;
2398 const TypePtr *tp = t2->is_ptr();
2399 if (TypePtr::above_centerline(tp->ptr()) ||
2400 tp->ptr() == TypePtr::Null) return Type::TOP;
2401
2402 // Return a more precise klass, if possible
2403 const TypeInstPtr *tinst = tp->isa_instptr();
2404 if (tinst != nullptr) {
2405 ciInstanceKlass* ik = tinst->instance_klass();
2406 int offset = tinst->offset();
2407 if (ik == phase->C->env()->Class_klass()
2408 && (offset == java_lang_Class::klass_offset() ||
2409 offset == java_lang_Class::array_klass_offset())) {
2410 // We are loading a special hidden field from a Class mirror object,
2411 // the field which points to the VM's Klass metaobject.
2412 ciType* t = tinst->java_mirror_type();
2413 // java_mirror_type returns non-null for compile-time Class constants.
2414 if (t != nullptr) {
2415 // constant oop => constant klass
2416 if (offset == java_lang_Class::array_klass_offset()) {
2417 if (t->is_void()) {
2418 // We cannot create a void array. Since void is a primitive type return null
2419 // klass. Users of this result need to do a null check on the returned klass.
2420 return TypePtr::NULL_PTR;
2421 }
2422 return TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);
2423 }
2424 if (!t->is_klass()) {
2425 // a primitive Class (e.g., int.class) has null for a klass field
2426 return TypePtr::NULL_PTR;
2427 }
2428 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2429 return TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);
2430 }
2431 // non-constant mirror, so we can't tell what's going on
2432 }
2433 if (!tinst->is_loaded())
2434 return _type; // Bail out if not loaded
2435 if (offset == oopDesc::klass_offset_in_bytes()) {
2436 return tinst->as_klass_type(true);
2437 }
2438 }
2439
2440 // Check for loading klass from an array
2441 const TypeAryPtr *tary = tp->isa_aryptr();
2442 if (tary != nullptr &&
2443 tary->offset() == oopDesc::klass_offset_in_bytes()) {
2444 return tary->as_klass_type(true);
2445 }
2446
2447 // Check for loading klass from an array klass
2448 const TypeKlassPtr *tkls = tp->isa_klassptr();
2449 if (tkls != nullptr && !StressReflectiveCode) {
2450 if (!tkls->is_loaded())
2451 return _type; // Bail out if not loaded
2452 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2453 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2454 // // Always returning precise element type is incorrect,
2455 // // e.g., element type could be object and array may contain strings
2456 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2457
2458 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2459 // according to the element type's subclassing.
2460 return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2461 }
2462 if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() &&
2463 tkls->offset() == in_bytes(Klass::super_offset())) {
2464 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super();
2465 // The field is Klass::_super. Return its (constant) value.
2466 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2467 return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR;
2468 }
2469 }
2470
2471 // Bailout case
2472 return LoadNode::Value(phase);
2473 }
2474
2475 //------------------------------Identity---------------------------------------
2476 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2477 // Also feed through the klass in Allocate(...klass...)._klass.
2478 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2479 return klass_identity_common(phase);
2480 }
2481
2482 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2483 Node* x = LoadNode::Identity(phase);
2484 if (x != this) return x;
2485
2486 // Take apart the address into an oop and offset.
2487 // Return 'this' if we cannot.
2488 Node* adr = in(MemNode::Address);
2489 intptr_t offset = 0;
2490 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2491 if (base == nullptr) return this;
2492 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2493 if (toop == nullptr) return this;
2494
2495 // Step over potential GC barrier for OopHandle resolve
2496 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2497 if (bs->is_gc_barrier_node(base)) {
2498 base = bs->step_over_gc_barrier(base);
2499 }
2500
2501 // We can fetch the klass directly through an AllocateNode.
2502 // This works even if the klass is not constant (clone or newArray).
2503 if (offset == oopDesc::klass_offset_in_bytes()) {
2504 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2505 if (allocated_klass != nullptr) {
2506 return allocated_klass;
2507 }
2508 }
2509
2510 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2511 // See inline_native_Class_query for occurrences of these patterns.
2512 // Java Example: x.getClass().isAssignableFrom(y)
2513 //
2514 // This improves reflective code, often making the Class
2515 // mirror go completely dead. (Current exception: Class
2516 // mirrors may appear in debug info, but we could clean them out by
2517 // introducing a new debug info operator for Klass.java_mirror).
2518
2519 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass()
2520 && offset == java_lang_Class::klass_offset()) {
2521 if (base->is_Load()) {
2522 Node* base2 = base->in(MemNode::Address);
2523 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2524 Node* adr2 = base2->in(MemNode::Address);
2525 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2526 if (tkls != nullptr && !tkls->empty()
2527 && (tkls->isa_instklassptr() || tkls->isa_aryklassptr())
2528 && adr2->is_AddP()
2529 ) {
2530 int mirror_field = in_bytes(Klass::java_mirror_offset());
2531 if (tkls->offset() == mirror_field) {
2532 return adr2->in(AddPNode::Base);
2533 }
2534 }
2535 }
2536 }
2537 }
2538
2539 return this;
2540 }
2541
2542 LoadNode* LoadNode::clone_pinned() const {
2543 LoadNode* ld = clone()->as_Load();
2544 ld->_control_dependency = UnknownControl;
2545 return ld;
2546 }
2547
2548
2549 //------------------------------Value------------------------------------------
2550 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2551 const Type *t = klass_value_common(phase);
2552 if (t == Type::TOP)
2553 return t;
2554
2555 return t->make_narrowklass();
2556 }
2557
2558 //------------------------------Identity---------------------------------------
2559 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2560 // Also feed through the klass in Allocate(...klass...)._klass.
2561 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2562 Node *x = klass_identity_common(phase);
2563
2564 const Type *t = phase->type( x );
2565 if( t == Type::TOP ) return x;
2566 if( t->isa_narrowklass()) return x;
2567 assert (!t->isa_narrowoop(), "no narrow oop here");
2568
2569 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2570 }
2571
2572 //------------------------------Value-----------------------------------------
2573 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2574 // Either input is TOP ==> the result is TOP
2575 const Type *t1 = phase->type( in(MemNode::Memory) );
2576 if( t1 == Type::TOP ) return Type::TOP;
2577 Node *adr = in(MemNode::Address);
2578 const Type *t2 = phase->type( adr );
2579 if( t2 == Type::TOP ) return Type::TOP;
2580 const TypePtr *tp = t2->is_ptr();
2581 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2582 const TypeAryPtr *tap = tp->isa_aryptr();
2583 if( !tap ) return _type;
2584 return tap->size();
2585 }
2586
2587 //-------------------------------Ideal---------------------------------------
2588 // Feed through the length in AllocateArray(...length...)._length.
2589 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2590 Node* p = MemNode::Ideal_common(phase, can_reshape);
2591 if (p) return (p == NodeSentinel) ? nullptr : p;
2592
2593 // Take apart the address into an oop and offset.
2594 // Return 'this' if we cannot.
2595 Node* adr = in(MemNode::Address);
2596 intptr_t offset = 0;
2597 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2598 if (base == nullptr) return nullptr;
2599 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2600 if (tary == nullptr) return nullptr;
2601
2602 // We can fetch the length directly through an AllocateArrayNode.
2603 // This works even if the length is not constant (clone or newArray).
2604 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2605 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2606 if (alloc != nullptr) {
2607 Node* allocated_length = alloc->Ideal_length();
2608 Node* len = alloc->make_ideal_length(tary, phase);
2609 if (allocated_length != len) {
2610 // New CastII improves on this.
2611 return len;
2612 }
2613 }
2614 }
2615
2616 return nullptr;
2617 }
2618
2619 //------------------------------Identity---------------------------------------
2620 // Feed through the length in AllocateArray(...length...)._length.
2621 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2622 Node* x = LoadINode::Identity(phase);
2623 if (x != this) return x;
2624
2625 // Take apart the address into an oop and offset.
2626 // Return 'this' if we cannot.
2627 Node* adr = in(MemNode::Address);
2628 intptr_t offset = 0;
2629 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2630 if (base == nullptr) return this;
2631 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2632 if (tary == nullptr) return this;
2633
2634 // We can fetch the length directly through an AllocateArrayNode.
2635 // This works even if the length is not constant (clone or newArray).
2636 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2637 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2638 if (alloc != nullptr) {
2639 Node* allocated_length = alloc->Ideal_length();
2640 // Do not allow make_ideal_length to allocate a CastII node.
2641 Node* len = alloc->make_ideal_length(tary, phase, false);
2642 if (allocated_length == len) {
2643 // Return allocated_length only if it would not be improved by a CastII.
2644 return allocated_length;
2645 }
2646 }
2647 }
2648
2649 return this;
2650
2651 }
2652
2653 //=============================================================================
2654 //---------------------------StoreNode::make-----------------------------------
2655 // Polymorphic factory method:
2656 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) {
2657 assert((mo == unordered || mo == release), "unexpected");
2658 Compile* C = gvn.C;
2659 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2660 ctl != nullptr, "raw memory operations should have control edge");
2661
2662 switch (bt) {
2663 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2664 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2665 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2666 case T_CHAR:
2667 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2668 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2669 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2670 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2671 case T_METADATA:
2672 case T_ADDRESS:
2673 case T_OBJECT:
2674 #ifdef _LP64
2675 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2676 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2677 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2678 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2679 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2680 adr->bottom_type()->isa_rawptr())) {
2681 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2682 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2683 }
2684 #endif
2685 {
2686 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2687 }
2688 default:
2689 ShouldNotReachHere();
2690 return (StoreNode*)nullptr;
2691 }
2692 }
2693
2694 //--------------------------bottom_type----------------------------------------
2695 const Type *StoreNode::bottom_type() const {
2696 return Type::MEMORY;
2697 }
2698
2699 //------------------------------hash-------------------------------------------
2700 uint StoreNode::hash() const {
2701 // unroll addition of interesting fields
2702 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2703
2704 // Since they are not commoned, do not hash them:
2705 return NO_HASH;
2706 }
2707
2708 //------------------------------Ideal------------------------------------------
2709 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2710 // When a store immediately follows a relevant allocation/initialization,
2711 // try to capture it into the initialization, or hoist it above.
2712 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2713 Node* p = MemNode::Ideal_common(phase, can_reshape);
2714 if (p) return (p == NodeSentinel) ? nullptr : p;
2715
2716 Node* mem = in(MemNode::Memory);
2717 Node* address = in(MemNode::Address);
2718 Node* value = in(MemNode::ValueIn);
2719 // Back-to-back stores to same address? Fold em up. Generally
2720 // unsafe if I have intervening uses... Also disallowed for StoreCM
2721 // since they must follow each StoreP operation. Redundant StoreCMs
2722 // are eliminated just before matching in final_graph_reshape.
2723 {
2724 Node* st = mem;
2725 // If Store 'st' has more than one use, we cannot fold 'st' away.
2726 // For example, 'st' might be the final state at a conditional
2727 // return. Or, 'st' might be used by some node which is live at
2728 // the same time 'st' is live, which might be unschedulable. So,
2729 // require exactly ONE user until such time as we clone 'mem' for
2730 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2731 // true).
2732 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2733 // Looking at a dead closed cycle of memory?
2734 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2735 assert(Opcode() == st->Opcode() ||
2736 st->Opcode() == Op_StoreVector ||
2737 Opcode() == Op_StoreVector ||
2738 st->Opcode() == Op_StoreVectorScatter ||
2739 Opcode() == Op_StoreVectorScatter ||
2740 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2741 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2742 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2743 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2744 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2745
2746 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2747 st->as_Store()->memory_size() <= this->memory_size()) {
2748 Node* use = st->raw_out(0);
2749 if (phase->is_IterGVN()) {
2750 phase->is_IterGVN()->rehash_node_delayed(use);
2751 }
2752 // It's OK to do this in the parser, since DU info is always accurate,
2753 // and the parser always refers to nodes via SafePointNode maps.
2754 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
2755 return this;
2756 }
2757 st = st->in(MemNode::Memory);
2758 }
2759 }
2760
2761
2762 // Capture an unaliased, unconditional, simple store into an initializer.
2763 // Or, if it is independent of the allocation, hoist it above the allocation.
2764 if (ReduceFieldZeroing && /*can_reshape &&*/
2765 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2766 InitializeNode* init = mem->in(0)->as_Initialize();
2767 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2768 if (offset > 0) {
2769 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2770 // If the InitializeNode captured me, it made a raw copy of me,
2771 // and I need to disappear.
2772 if (moved != nullptr) {
2773 // %%% hack to ensure that Ideal returns a new node:
2774 mem = MergeMemNode::make(mem);
2775 return mem; // fold me away
2776 }
2777 }
2778 }
2779
2780 // Fold reinterpret cast into memory operation:
2781 // StoreX mem (MoveY2X v) => StoreY mem v
2782 if (value->is_Move()) {
2783 const Type* vt = value->in(1)->bottom_type();
2784 if (has_reinterpret_variant(vt)) {
2785 if (phase->C->post_loop_opts_phase()) {
2786 return convert_to_reinterpret_store(*phase, value->in(1), vt);
2787 } else {
2788 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
2789 }
2790 }
2791 }
2792
2793 return nullptr; // No further progress
2794 }
2795
2796 //------------------------------Value-----------------------------------------
2797 const Type* StoreNode::Value(PhaseGVN* phase) const {
2798 // Either input is TOP ==> the result is TOP
2799 const Type *t1 = phase->type( in(MemNode::Memory) );
2800 if( t1 == Type::TOP ) return Type::TOP;
2801 const Type *t2 = phase->type( in(MemNode::Address) );
2802 if( t2 == Type::TOP ) return Type::TOP;
2803 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2804 if( t3 == Type::TOP ) return Type::TOP;
2805 return Type::MEMORY;
2806 }
2807
2808 //------------------------------Identity---------------------------------------
2809 // Remove redundant stores:
2810 // Store(m, p, Load(m, p)) changes to m.
2811 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2812 Node* StoreNode::Identity(PhaseGVN* phase) {
2813 Node* mem = in(MemNode::Memory);
2814 Node* adr = in(MemNode::Address);
2815 Node* val = in(MemNode::ValueIn);
2816
2817 Node* result = this;
2818
2819 // Load then Store? Then the Store is useless
2820 if (val->is_Load() &&
2821 val->in(MemNode::Address)->eqv_uncast(adr) &&
2822 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2823 val->as_Load()->store_Opcode() == Opcode()) {
2824 result = mem;
2825 }
2826
2827 // Two stores in a row of the same value?
2828 if (result == this &&
2829 mem->is_Store() &&
2830 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2831 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2832 mem->Opcode() == Opcode()) {
2833 result = mem;
2834 }
2835
2836 // Store of zero anywhere into a freshly-allocated object?
2837 // Then the store is useless.
2838 // (It must already have been captured by the InitializeNode.)
2839 if (result == this &&
2840 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2841 // a newly allocated object is already all-zeroes everywhere
2842 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2843 result = mem;
2844 }
2845
2846 if (result == this) {
2847 // the store may also apply to zero-bits in an earlier object
2848 Node* prev_mem = find_previous_store(phase);
2849 // Steps (a), (b): Walk past independent stores to find an exact match.
2850 if (prev_mem != nullptr) {
2851 Node* prev_val = can_see_stored_value(prev_mem, phase);
2852 if (prev_val != nullptr && prev_val == val) {
2853 // prev_val and val might differ by a cast; it would be good
2854 // to keep the more informative of the two.
2855 result = mem;
2856 }
2857 }
2858 }
2859 }
2860
2861 PhaseIterGVN* igvn = phase->is_IterGVN();
2862 if (result != this && igvn != nullptr) {
2863 MemBarNode* trailing = trailing_membar();
2864 if (trailing != nullptr) {
2865 #ifdef ASSERT
2866 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2867 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects");
2868 #endif
2869 trailing->remove(igvn);
2870 }
2871 }
2872
2873 return result;
2874 }
2875
2876 //------------------------------match_edge-------------------------------------
2877 // Do we Match on this edge index or not? Match only memory & value
2878 uint StoreNode::match_edge(uint idx) const {
2879 return idx == MemNode::Address || idx == MemNode::ValueIn;
2880 }
2881
2882 //------------------------------cmp--------------------------------------------
2883 // Do not common stores up together. They generally have to be split
2884 // back up anyways, so do not bother.
2885 bool StoreNode::cmp( const Node &n ) const {
2886 return (&n == this); // Always fail except on self
2887 }
2888
2889 //------------------------------Ideal_masked_input-----------------------------
2890 // Check for a useless mask before a partial-word store
2891 // (StoreB ... (AndI valIn conIa) )
2892 // If (conIa & mask == mask) this simplifies to
2893 // (StoreB ... (valIn) )
2894 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2895 Node *val = in(MemNode::ValueIn);
2896 if( val->Opcode() == Op_AndI ) {
2897 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2898 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2899 set_req_X(MemNode::ValueIn, val->in(1), phase);
2900 return this;
2901 }
2902 }
2903 return nullptr;
2904 }
2905
2906
2907 //------------------------------Ideal_sign_extended_input----------------------
2908 // Check for useless sign-extension before a partial-word store
2909 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2910 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2911 // (StoreB ... (valIn) )
2912 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2913 Node *val = in(MemNode::ValueIn);
2914 if( val->Opcode() == Op_RShiftI ) {
2915 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2916 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2917 Node *shl = val->in(1);
2918 if( shl->Opcode() == Op_LShiftI ) {
2919 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2920 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2921 set_req_X(MemNode::ValueIn, shl->in(1), phase);
2922 return this;
2923 }
2924 }
2925 }
2926 }
2927 return nullptr;
2928 }
2929
2930 //------------------------------value_never_loaded-----------------------------------
2931 // Determine whether there are any possible loads of the value stored.
2932 // For simplicity, we actually check if there are any loads from the
2933 // address stored to, not just for loads of the value stored by this node.
2934 //
2935 bool StoreNode::value_never_loaded(PhaseValues* phase) const {
2936 Node *adr = in(Address);
2937 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2938 if (adr_oop == nullptr)
2939 return false;
2940 if (!adr_oop->is_known_instance_field())
2941 return false; // if not a distinct instance, there may be aliases of the address
2942 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2943 Node *use = adr->fast_out(i);
2944 if (use->is_Load() || use->is_LoadStore()) {
2945 return false;
2946 }
2947 }
2948 return true;
2949 }
2950
2951 MemBarNode* StoreNode::trailing_membar() const {
2952 if (is_release()) {
2953 MemBarNode* trailing_mb = nullptr;
2954 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2955 Node* u = fast_out(i);
2956 if (u->is_MemBar()) {
2957 if (u->as_MemBar()->trailing_store()) {
2958 assert(u->Opcode() == Op_MemBarVolatile, "");
2959 assert(trailing_mb == nullptr, "only one");
2960 trailing_mb = u->as_MemBar();
2961 #ifdef ASSERT
2962 Node* leading = u->as_MemBar()->leading_membar();
2963 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2964 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2965 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2966 #endif
2967 } else {
2968 assert(u->as_MemBar()->standalone(), "");
2969 }
2970 }
2971 }
2972 return trailing_mb;
2973 }
2974 return nullptr;
2975 }
2976
2977
2978 //=============================================================================
2979 //------------------------------Ideal------------------------------------------
2980 // If the store is from an AND mask that leaves the low bits untouched, then
2981 // we can skip the AND operation. If the store is from a sign-extension
2982 // (a left shift, then right shift) we can skip both.
2983 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2984 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2985 if( progress != nullptr ) return progress;
2986
2987 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2988 if( progress != nullptr ) return progress;
2989
2990 // Finally check the default case
2991 return StoreNode::Ideal(phase, can_reshape);
2992 }
2993
2994 //=============================================================================
2995 //------------------------------Ideal------------------------------------------
2996 // If the store is from an AND mask that leaves the low bits untouched, then
2997 // we can skip the AND operation
2998 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2999 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
3000 if( progress != nullptr ) return progress;
3001
3002 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
3003 if( progress != nullptr ) return progress;
3004
3005 // Finally check the default case
3006 return StoreNode::Ideal(phase, can_reshape);
3007 }
3008
3009 //=============================================================================
3010 //------------------------------Identity---------------------------------------
3011 Node* StoreCMNode::Identity(PhaseGVN* phase) {
3012 // No need to card mark when storing a null ptr
3013 Node* my_store = in(MemNode::OopStore);
3014 if (my_store->is_Store()) {
3015 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
3016 if( t1 == TypePtr::NULL_PTR ) {
3017 return in(MemNode::Memory);
3018 }
3019 }
3020 return this;
3021 }
3022
3023 //=============================================================================
3024 //------------------------------Ideal---------------------------------------
3025 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
3026 Node* progress = StoreNode::Ideal(phase, can_reshape);
3027 if (progress != nullptr) return progress;
3028
3029 Node* my_store = in(MemNode::OopStore);
3030 if (my_store->is_MergeMem()) {
3031 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
3032 set_req_X(MemNode::OopStore, mem, phase);
3033 return this;
3034 }
3035
3036 return nullptr;
3037 }
3038
3039 //------------------------------Value-----------------------------------------
3040 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
3041 // Either input is TOP ==> the result is TOP (checked in StoreNode::Value).
3042 // If extra input is TOP ==> the result is TOP
3043 const Type* t = phase->type(in(MemNode::OopStore));
3044 if (t == Type::TOP) {
3045 return Type::TOP;
3046 }
3047 return StoreNode::Value(phase);
3048 }
3049
3050
3051 //=============================================================================
3052 //----------------------------------SCMemProjNode------------------------------
3053 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
3054 {
3055 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
3056 return Type::TOP;
3057 }
3058 return bottom_type();
3059 }
3060
3061 //=============================================================================
3062 //----------------------------------LoadStoreNode------------------------------
3063 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
3064 : Node(required),
3065 _type(rt),
3066 _adr_type(at),
3067 _barrier_data(0)
3068 {
3069 init_req(MemNode::Control, c );
3070 init_req(MemNode::Memory , mem);
3071 init_req(MemNode::Address, adr);
3072 init_req(MemNode::ValueIn, val);
3073 init_class_id(Class_LoadStore);
3074 }
3075
3076 //------------------------------Value-----------------------------------------
3077 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
3078 // Either input is TOP ==> the result is TOP
3079 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
3080 return Type::TOP;
3081 }
3082 const Type* t = phase->type(in(MemNode::Memory));
3083 if (t == Type::TOP) {
3084 return Type::TOP;
3085 }
3086 t = phase->type(in(MemNode::Address));
3087 if (t == Type::TOP) {
3088 return Type::TOP;
3089 }
3090 t = phase->type(in(MemNode::ValueIn));
3091 if (t == Type::TOP) {
3092 return Type::TOP;
3093 }
3094 return bottom_type();
3095 }
3096
3097 uint LoadStoreNode::ideal_reg() const {
3098 return _type->ideal_reg();
3099 }
3100
3101 // This method conservatively checks if the result of a LoadStoreNode is
3102 // used, that is, if it returns true, then it is definitely the case that
3103 // the result of the node is not needed.
3104 // For example, GetAndAdd can be matched into a lock_add instead of a
3105 // lock_xadd if the result of LoadStoreNode::result_not_used() is true
3106 bool LoadStoreNode::result_not_used() const {
3107 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3108 Node *x = fast_out(i);
3109 if (x->Opcode() == Op_SCMemProj) {
3110 continue;
3111 }
3112 if (x->bottom_type() == TypeTuple::MEMBAR &&
3113 !x->is_Call() &&
3114 x->Opcode() != Op_Blackhole) {
3115 continue;
3116 }
3117 return false;
3118 }
3119 return true;
3120 }
3121
3122 MemBarNode* LoadStoreNode::trailing_membar() const {
3123 MemBarNode* trailing = nullptr;
3124 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3125 Node* u = fast_out(i);
3126 if (u->is_MemBar()) {
3127 if (u->as_MemBar()->trailing_load_store()) {
3128 assert(u->Opcode() == Op_MemBarAcquire, "");
3129 assert(trailing == nullptr, "only one");
3130 trailing = u->as_MemBar();
3131 #ifdef ASSERT
3132 Node* leading = trailing->leading_membar();
3133 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3134 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
3135 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
3136 #endif
3137 } else {
3138 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
3139 }
3140 }
3141 }
3142
3143 return trailing;
3144 }
3145
3146 uint LoadStoreNode::size_of() const { return sizeof(*this); }
3147
3148 //=============================================================================
3149 //----------------------------------LoadStoreConditionalNode--------------------
3150 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) {
3151 init_req(ExpectedIn, ex );
3152 }
3153
3154 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
3155 // Either input is TOP ==> the result is TOP
3156 const Type* t = phase->type(in(ExpectedIn));
3157 if (t == Type::TOP) {
3158 return Type::TOP;
3159 }
3160 return LoadStoreNode::Value(phase);
3161 }
3162
3163 //=============================================================================
3164 //-------------------------------adr_type--------------------------------------
3165 const TypePtr* ClearArrayNode::adr_type() const {
3166 Node *adr = in(3);
3167 if (adr == nullptr) return nullptr; // node is dead
3168 return MemNode::calculate_adr_type(adr->bottom_type());
3169 }
3170
3171 //------------------------------match_edge-------------------------------------
3172 // Do we Match on this edge index or not? Do not match memory
3173 uint ClearArrayNode::match_edge(uint idx) const {
3174 return idx > 1;
3175 }
3176
3177 //------------------------------Identity---------------------------------------
3178 // Clearing a zero length array does nothing
3179 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
3180 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
3181 }
3182
3183 //------------------------------Idealize---------------------------------------
3184 // Clearing a short array is faster with stores
3185 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3186 // Already know this is a large node, do not try to ideal it
3187 if (_is_large) return nullptr;
3188
3189 const int unit = BytesPerLong;
3190 const TypeX* t = phase->type(in(2))->isa_intptr_t();
3191 if (!t) return nullptr;
3192 if (!t->is_con()) return nullptr;
3193 intptr_t raw_count = t->get_con();
3194 intptr_t size = raw_count;
3195 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
3196 // Clearing nothing uses the Identity call.
3197 // Negative clears are possible on dead ClearArrays
3198 // (see jck test stmt114.stmt11402.val).
3199 if (size <= 0 || size % unit != 0) return nullptr;
3200 intptr_t count = size / unit;
3201 // Length too long; communicate this to matchers and assemblers.
3202 // Assemblers are responsible to produce fast hardware clears for it.
3203 if (size > InitArrayShortSize) {
3204 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
3205 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
3206 return nullptr;
3207 }
3208 if (!IdealizeClearArrayNode) return nullptr;
3209 Node *mem = in(1);
3210 if( phase->type(mem)==Type::TOP ) return nullptr;
3211 Node *adr = in(3);
3212 const Type* at = phase->type(adr);
3213 if( at==Type::TOP ) return nullptr;
3214 const TypePtr* atp = at->isa_ptr();
3215 // adjust atp to be the correct array element address type
3216 if (atp == nullptr) atp = TypePtr::BOTTOM;
3217 else atp = atp->add_offset(Type::OffsetBot);
3218 // Get base for derived pointer purposes
3219 if( adr->Opcode() != Op_AddP ) Unimplemented();
3220 Node *base = adr->in(1);
3221
3222 Node *zero = phase->makecon(TypeLong::ZERO);
3223 Node *off = phase->MakeConX(BytesPerLong);
3224 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3225 count--;
3226 while( count-- ) {
3227 mem = phase->transform(mem);
3228 adr = phase->transform(new AddPNode(base,adr,off));
3229 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3230 }
3231 return mem;
3232 }
3233
3234 //----------------------------step_through----------------------------------
3235 // Return allocation input memory edge if it is different instance
3236 // or itself if it is the one we are looking for.
3237 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
3238 Node* n = *np;
3239 assert(n->is_ClearArray(), "sanity");
3240 intptr_t offset;
3241 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
3242 // This method is called only before Allocate nodes are expanded
3243 // during macro nodes expansion. Before that ClearArray nodes are
3244 // only generated in PhaseMacroExpand::generate_arraycopy() (before
3245 // Allocate nodes are expanded) which follows allocations.
3246 assert(alloc != nullptr, "should have allocation");
3247 if (alloc->_idx == instance_id) {
3248 // Can not bypass initialization of the instance we are looking for.
3249 return false;
3250 }
3251 // Otherwise skip it.
3252 InitializeNode* init = alloc->initialization();
3253 if (init != nullptr)
3254 *np = init->in(TypeFunc::Memory);
3255 else
3256 *np = alloc->in(TypeFunc::Memory);
3257 return true;
3258 }
3259
3260 //----------------------------clear_memory-------------------------------------
3261 // Generate code to initialize object storage to zero.
3262 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3263 intptr_t start_offset,
3264 Node* end_offset,
3265 PhaseGVN* phase) {
3266 intptr_t offset = start_offset;
3267
3268 int unit = BytesPerLong;
3269 if ((offset % unit) != 0) {
3270 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
3271 adr = phase->transform(adr);
3272 const TypePtr* atp = TypeRawPtr::BOTTOM;
3273 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3274 mem = phase->transform(mem);
3275 offset += BytesPerInt;
3276 }
3277 assert((offset % unit) == 0, "");
3278
3279 // Initialize the remaining stuff, if any, with a ClearArray.
3280 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
3281 }
3282
3283 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3284 Node* start_offset,
3285 Node* end_offset,
3286 PhaseGVN* phase) {
3287 if (start_offset == end_offset) {
3288 // nothing to do
3289 return mem;
3290 }
3291
3292 int unit = BytesPerLong;
3293 Node* zbase = start_offset;
3294 Node* zend = end_offset;
3295
3296 // Scale to the unit required by the CPU:
3297 if (!Matcher::init_array_count_is_in_bytes) {
3298 Node* shift = phase->intcon(exact_log2(unit));
3299 zbase = phase->transform(new URShiftXNode(zbase, shift) );
3300 zend = phase->transform(new URShiftXNode(zend, shift) );
3301 }
3302
3303 // Bulk clear double-words
3304 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
3305 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
3306 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
3307 return phase->transform(mem);
3308 }
3309
3310 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3311 intptr_t start_offset,
3312 intptr_t end_offset,
3313 PhaseGVN* phase) {
3314 if (start_offset == end_offset) {
3315 // nothing to do
3316 return mem;
3317 }
3318
3319 assert((end_offset % BytesPerInt) == 0, "odd end offset");
3320 intptr_t done_offset = end_offset;
3321 if ((done_offset % BytesPerLong) != 0) {
3322 done_offset -= BytesPerInt;
3323 }
3324 if (done_offset > start_offset) {
3325 mem = clear_memory(ctl, mem, dest,
3326 start_offset, phase->MakeConX(done_offset), phase);
3327 }
3328 if (done_offset < end_offset) { // emit the final 32-bit store
3329 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3330 adr = phase->transform(adr);
3331 const TypePtr* atp = TypeRawPtr::BOTTOM;
3332 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3333 mem = phase->transform(mem);
3334 done_offset += BytesPerInt;
3335 }
3336 assert(done_offset == end_offset, "");
3337 return mem;
3338 }
3339
3340 //=============================================================================
3341 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3342 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
3343 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3344 #ifdef ASSERT
3345 , _pair_idx(0)
3346 #endif
3347 {
3348 init_class_id(Class_MemBar);
3349 Node* top = C->top();
3350 init_req(TypeFunc::I_O,top);
3351 init_req(TypeFunc::FramePtr,top);
3352 init_req(TypeFunc::ReturnAdr,top);
3353 if (precedent != nullptr)
3354 init_req(TypeFunc::Parms, precedent);
3355 }
3356
3357 //------------------------------cmp--------------------------------------------
3358 uint MemBarNode::hash() const { return NO_HASH; }
3359 bool MemBarNode::cmp( const Node &n ) const {
3360 return (&n == this); // Always fail except on self
3361 }
3362
3363 //------------------------------make-------------------------------------------
3364 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3365 switch (opcode) {
3366 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
3367 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
3368 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
3369 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
3370 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
3371 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn);
3372 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3373 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3374 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
3375 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
3376 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
3377 case Op_Initialize: return new InitializeNode(C, atp, pn);
3378 default: ShouldNotReachHere(); return nullptr;
3379 }
3380 }
3381
3382 void MemBarNode::remove(PhaseIterGVN *igvn) {
3383 if (outcnt() != 2) {
3384 assert(Opcode() == Op_Initialize, "Only seen when there are no use of init memory");
3385 assert(outcnt() == 1, "Only control then");
3386 }
3387 if (trailing_store() || trailing_load_store()) {
3388 MemBarNode* leading = leading_membar();
3389 if (leading != nullptr) {
3390 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3391 leading->remove(igvn);
3392 }
3393 }
3394 if (proj_out_or_null(TypeFunc::Memory) != nullptr) {
3395 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3396 }
3397 if (proj_out_or_null(TypeFunc::Control) != nullptr) {
3398 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3399 }
3400 }
3401
3402 //------------------------------Ideal------------------------------------------
3403 // Return a node which is more "ideal" than the current node. Strip out
3404 // control copies
3405 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3406 if (remove_dead_region(phase, can_reshape)) return this;
3407 // Don't bother trying to transform a dead node
3408 if (in(0) && in(0)->is_top()) {
3409 return nullptr;
3410 }
3411
3412 bool progress = false;
3413 // Eliminate volatile MemBars for scalar replaced objects.
3414 if (can_reshape && req() == (Precedent+1)) {
3415 bool eliminate = false;
3416 int opc = Opcode();
3417 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3418 // Volatile field loads and stores.
3419 Node* my_mem = in(MemBarNode::Precedent);
3420 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3421 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3422 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3423 // replace this Precedent (decodeN) with the Load instead.
3424 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3425 Node* load_node = my_mem->in(1);
3426 set_req(MemBarNode::Precedent, load_node);
3427 phase->is_IterGVN()->_worklist.push(my_mem);
3428 my_mem = load_node;
3429 } else {
3430 assert(my_mem->unique_out() == this, "sanity");
3431 assert(!trailing_load_store(), "load store node can't be eliminated");
3432 del_req(Precedent);
3433 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3434 my_mem = nullptr;
3435 }
3436 progress = true;
3437 }
3438 if (my_mem != nullptr && my_mem->is_Mem()) {
3439 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3440 // Check for scalar replaced object reference.
3441 if( t_oop != nullptr && t_oop->is_known_instance_field() &&
3442 t_oop->offset() != Type::OffsetBot &&
3443 t_oop->offset() != Type::OffsetTop) {
3444 eliminate = true;
3445 }
3446 }
3447 } else if (opc == Op_MemBarRelease) {
3448 // Final field stores.
3449 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent));
3450 if ((alloc != nullptr) && alloc->is_Allocate() &&
3451 alloc->as_Allocate()->does_not_escape_thread()) {
3452 // The allocated object does not escape.
3453 eliminate = true;
3454 }
3455 }
3456 if (eliminate) {
3457 // Replace MemBar projections by its inputs.
3458 PhaseIterGVN* igvn = phase->is_IterGVN();
3459 remove(igvn);
3460 // Must return either the original node (now dead) or a new node
3461 // (Do not return a top here, since that would break the uniqueness of top.)
3462 return new ConINode(TypeInt::ZERO);
3463 }
3464 }
3465 return progress ? this : nullptr;
3466 }
3467
3468 //------------------------------Value------------------------------------------
3469 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3470 if( !in(0) ) return Type::TOP;
3471 if( phase->type(in(0)) == Type::TOP )
3472 return Type::TOP;
3473 return TypeTuple::MEMBAR;
3474 }
3475
3476 //------------------------------match------------------------------------------
3477 // Construct projections for memory.
3478 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3479 switch (proj->_con) {
3480 case TypeFunc::Control:
3481 case TypeFunc::Memory:
3482 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3483 }
3484 ShouldNotReachHere();
3485 return nullptr;
3486 }
3487
3488 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3489 trailing->_kind = TrailingStore;
3490 leading->_kind = LeadingStore;
3491 #ifdef ASSERT
3492 trailing->_pair_idx = leading->_idx;
3493 leading->_pair_idx = leading->_idx;
3494 #endif
3495 }
3496
3497 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3498 trailing->_kind = TrailingLoadStore;
3499 leading->_kind = LeadingLoadStore;
3500 #ifdef ASSERT
3501 trailing->_pair_idx = leading->_idx;
3502 leading->_pair_idx = leading->_idx;
3503 #endif
3504 }
3505
3506 MemBarNode* MemBarNode::trailing_membar() const {
3507 ResourceMark rm;
3508 Node* trailing = (Node*)this;
3509 VectorSet seen;
3510 Node_Stack multis(0);
3511 do {
3512 Node* c = trailing;
3513 uint i = 0;
3514 do {
3515 trailing = nullptr;
3516 for (; i < c->outcnt(); i++) {
3517 Node* next = c->raw_out(i);
3518 if (next != c && next->is_CFG()) {
3519 if (c->is_MultiBranch()) {
3520 if (multis.node() == c) {
3521 multis.set_index(i+1);
3522 } else {
3523 multis.push(c, i+1);
3524 }
3525 }
3526 trailing = next;
3527 break;
3528 }
3529 }
3530 if (trailing != nullptr && !seen.test_set(trailing->_idx)) {
3531 break;
3532 }
3533 while (multis.size() > 0) {
3534 c = multis.node();
3535 i = multis.index();
3536 if (i < c->req()) {
3537 break;
3538 }
3539 multis.pop();
3540 }
3541 } while (multis.size() > 0);
3542 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
3543
3544 MemBarNode* mb = trailing->as_MemBar();
3545 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3546 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3547 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3548 return mb;
3549 }
3550
3551 MemBarNode* MemBarNode::leading_membar() const {
3552 ResourceMark rm;
3553 VectorSet seen;
3554 Node_Stack regions(0);
3555 Node* leading = in(0);
3556 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3557 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) {
3558 leading = nullptr;
3559 while (regions.size() > 0 && leading == nullptr) {
3560 Node* r = regions.node();
3561 uint i = regions.index();
3562 if (i < r->req()) {
3563 leading = r->in(i);
3564 regions.set_index(i+1);
3565 } else {
3566 regions.pop();
3567 }
3568 }
3569 if (leading == nullptr) {
3570 assert(regions.size() == 0, "all paths should have been tried");
3571 return nullptr;
3572 }
3573 }
3574 if (leading->is_Region()) {
3575 regions.push(leading, 2);
3576 leading = leading->in(1);
3577 } else {
3578 leading = leading->in(0);
3579 }
3580 }
3581 #ifdef ASSERT
3582 Unique_Node_List wq;
3583 wq.push((Node*)this);
3584 uint found = 0;
3585 for (uint i = 0; i < wq.size(); i++) {
3586 Node* n = wq.at(i);
3587 if (n->is_Region()) {
3588 for (uint j = 1; j < n->req(); j++) {
3589 Node* in = n->in(j);
3590 if (in != nullptr && !in->is_top()) {
3591 wq.push(in);
3592 }
3593 }
3594 } else {
3595 if (n->is_MemBar() && n->as_MemBar()->leading()) {
3596 assert(n == leading, "consistency check failed");
3597 found++;
3598 } else {
3599 Node* in = n->in(0);
3600 if (in != nullptr && !in->is_top()) {
3601 wq.push(in);
3602 }
3603 }
3604 }
3605 }
3606 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed");
3607 #endif
3608 if (leading == nullptr) {
3609 return nullptr;
3610 }
3611 MemBarNode* mb = leading->as_MemBar();
3612 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3613 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3614 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3615 return mb;
3616 }
3617
3618
3619 //===========================InitializeNode====================================
3620 // SUMMARY:
3621 // This node acts as a memory barrier on raw memory, after some raw stores.
3622 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3623 // The Initialize can 'capture' suitably constrained stores as raw inits.
3624 // It can coalesce related raw stores into larger units (called 'tiles').
3625 // It can avoid zeroing new storage for memory units which have raw inits.
3626 // At macro-expansion, it is marked 'complete', and does not optimize further.
3627 //
3628 // EXAMPLE:
3629 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3630 // ctl = incoming control; mem* = incoming memory
3631 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3632 // First allocate uninitialized memory and fill in the header:
3633 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3634 // ctl := alloc.Control; mem* := alloc.Memory*
3635 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3636 // Then initialize to zero the non-header parts of the raw memory block:
3637 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3638 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3639 // After the initialize node executes, the object is ready for service:
3640 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3641 // Suppose its body is immediately initialized as {1,2}:
3642 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3643 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3644 // mem.SLICE(#short[*]) := store2
3645 //
3646 // DETAILS:
3647 // An InitializeNode collects and isolates object initialization after
3648 // an AllocateNode and before the next possible safepoint. As a
3649 // memory barrier (MemBarNode), it keeps critical stores from drifting
3650 // down past any safepoint or any publication of the allocation.
3651 // Before this barrier, a newly-allocated object may have uninitialized bits.
3652 // After this barrier, it may be treated as a real oop, and GC is allowed.
3653 //
3654 // The semantics of the InitializeNode include an implicit zeroing of
3655 // the new object from object header to the end of the object.
3656 // (The object header and end are determined by the AllocateNode.)
3657 //
3658 // Certain stores may be added as direct inputs to the InitializeNode.
3659 // These stores must update raw memory, and they must be to addresses
3660 // derived from the raw address produced by AllocateNode, and with
3661 // a constant offset. They must be ordered by increasing offset.
3662 // The first one is at in(RawStores), the last at in(req()-1).
3663 // Unlike most memory operations, they are not linked in a chain,
3664 // but are displayed in parallel as users of the rawmem output of
3665 // the allocation.
3666 //
3667 // (See comments in InitializeNode::capture_store, which continue
3668 // the example given above.)
3669 //
3670 // When the associated Allocate is macro-expanded, the InitializeNode
3671 // may be rewritten to optimize collected stores. A ClearArrayNode
3672 // may also be created at that point to represent any required zeroing.
3673 // The InitializeNode is then marked 'complete', prohibiting further
3674 // capturing of nearby memory operations.
3675 //
3676 // During macro-expansion, all captured initializations which store
3677 // constant values of 32 bits or smaller are coalesced (if advantageous)
3678 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3679 // initialized in fewer memory operations. Memory words which are
3680 // covered by neither tiles nor non-constant stores are pre-zeroed
3681 // by explicit stores of zero. (The code shape happens to do all
3682 // zeroing first, then all other stores, with both sequences occurring
3683 // in order of ascending offsets.)
3684 //
3685 // Alternatively, code may be inserted between an AllocateNode and its
3686 // InitializeNode, to perform arbitrary initialization of the new object.
3687 // E.g., the object copying intrinsics insert complex data transfers here.
3688 // The initialization must then be marked as 'complete' disable the
3689 // built-in zeroing semantics and the collection of initializing stores.
3690 //
3691 // While an InitializeNode is incomplete, reads from the memory state
3692 // produced by it are optimizable if they match the control edge and
3693 // new oop address associated with the allocation/initialization.
3694 // They return a stored value (if the offset matches) or else zero.
3695 // A write to the memory state, if it matches control and address,
3696 // and if it is to a constant offset, may be 'captured' by the
3697 // InitializeNode. It is cloned as a raw memory operation and rewired
3698 // inside the initialization, to the raw oop produced by the allocation.
3699 // Operations on addresses which are provably distinct (e.g., to
3700 // other AllocateNodes) are allowed to bypass the initialization.
3701 //
3702 // The effect of all this is to consolidate object initialization
3703 // (both arrays and non-arrays, both piecewise and bulk) into a
3704 // single location, where it can be optimized as a unit.
3705 //
3706 // Only stores with an offset less than TrackedInitializationLimit words
3707 // will be considered for capture by an InitializeNode. This puts a
3708 // reasonable limit on the complexity of optimized initializations.
3709
3710 //---------------------------InitializeNode------------------------------------
3711 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3712 : MemBarNode(C, adr_type, rawoop),
3713 _is_complete(Incomplete), _does_not_escape(false)
3714 {
3715 init_class_id(Class_Initialize);
3716
3717 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3718 assert(in(RawAddress) == rawoop, "proper init");
3719 // Note: allocation() can be null, for secondary initialization barriers
3720 }
3721
3722 // Since this node is not matched, it will be processed by the
3723 // register allocator. Declare that there are no constraints
3724 // on the allocation of the RawAddress edge.
3725 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3726 // This edge should be set to top, by the set_complete. But be conservative.
3727 if (idx == InitializeNode::RawAddress)
3728 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3729 return RegMask::Empty;
3730 }
3731
3732 Node* InitializeNode::memory(uint alias_idx) {
3733 Node* mem = in(Memory);
3734 if (mem->is_MergeMem()) {
3735 return mem->as_MergeMem()->memory_at(alias_idx);
3736 } else {
3737 // incoming raw memory is not split
3738 return mem;
3739 }
3740 }
3741
3742 bool InitializeNode::is_non_zero() {
3743 if (is_complete()) return false;
3744 remove_extra_zeroes();
3745 return (req() > RawStores);
3746 }
3747
3748 void InitializeNode::set_complete(PhaseGVN* phase) {
3749 assert(!is_complete(), "caller responsibility");
3750 _is_complete = Complete;
3751
3752 // After this node is complete, it contains a bunch of
3753 // raw-memory initializations. There is no need for
3754 // it to have anything to do with non-raw memory effects.
3755 // Therefore, tell all non-raw users to re-optimize themselves,
3756 // after skipping the memory effects of this initialization.
3757 PhaseIterGVN* igvn = phase->is_IterGVN();
3758 if (igvn) igvn->add_users_to_worklist(this);
3759 }
3760
3761 // convenience function
3762 // return false if the init contains any stores already
3763 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3764 InitializeNode* init = initialization();
3765 if (init == nullptr || init->is_complete()) return false;
3766 init->remove_extra_zeroes();
3767 // for now, if this allocation has already collected any inits, bail:
3768 if (init->is_non_zero()) return false;
3769 init->set_complete(phase);
3770 return true;
3771 }
3772
3773 void InitializeNode::remove_extra_zeroes() {
3774 if (req() == RawStores) return;
3775 Node* zmem = zero_memory();
3776 uint fill = RawStores;
3777 for (uint i = fill; i < req(); i++) {
3778 Node* n = in(i);
3779 if (n->is_top() || n == zmem) continue; // skip
3780 if (fill < i) set_req(fill, n); // compact
3781 ++fill;
3782 }
3783 // delete any empty spaces created:
3784 while (fill < req()) {
3785 del_req(fill);
3786 }
3787 }
3788
3789 // Helper for remembering which stores go with which offsets.
3790 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) {
3791 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3792 intptr_t offset = -1;
3793 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3794 phase, offset);
3795 if (base == nullptr) return -1; // something is dead,
3796 if (offset < 0) return -1; // dead, dead
3797 return offset;
3798 }
3799
3800 // Helper for proving that an initialization expression is
3801 // "simple enough" to be folded into an object initialization.
3802 // Attempts to prove that a store's initial value 'n' can be captured
3803 // within the initialization without creating a vicious cycle, such as:
3804 // { Foo p = new Foo(); p.next = p; }
3805 // True for constants and parameters and small combinations thereof.
3806 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
3807 ResourceMark rm;
3808 Unique_Node_List worklist;
3809 worklist.push(value);
3810
3811 uint complexity_limit = 20;
3812 for (uint j = 0; j < worklist.size(); j++) {
3813 if (j >= complexity_limit) {
3814 return false; // Bail out if processed too many nodes
3815 }
3816
3817 Node* n = worklist.at(j);
3818 if (n == nullptr) continue; // (can this really happen?)
3819 if (n->is_Proj()) n = n->in(0);
3820 if (n == this) return false; // found a cycle
3821 if (n->is_Con()) continue;
3822 if (n->is_Start()) continue; // params, etc., are OK
3823 if (n->is_Root()) continue; // even better
3824
3825 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
3826 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
3827 continue;
3828 }
3829
3830 Node* ctl = n->in(0);
3831 if (ctl != nullptr && !ctl->is_top()) {
3832 if (ctl->is_Proj()) ctl = ctl->in(0);
3833 if (ctl == this) return false;
3834
3835 // If we already know that the enclosing memory op is pinned right after
3836 // the init, then any control flow that the store has picked up
3837 // must have preceded the init, or else be equal to the init.
3838 // Even after loop optimizations (which might change control edges)
3839 // a store is never pinned *before* the availability of its inputs.
3840 if (!MemNode::all_controls_dominate(n, this))
3841 return false; // failed to prove a good control
3842 }
3843
3844 // Check data edges for possible dependencies on 'this'.
3845 for (uint i = 1; i < n->req(); i++) {
3846 Node* m = n->in(i);
3847 if (m == nullptr || m == n || m->is_top()) continue;
3848
3849 // Only process data inputs once
3850 worklist.push(m);
3851 }
3852 }
3853
3854 return true;
3855 }
3856
3857 // Here are all the checks a Store must pass before it can be moved into
3858 // an initialization. Returns zero if a check fails.
3859 // On success, returns the (constant) offset to which the store applies,
3860 // within the initialized memory.
3861 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
3862 const int FAIL = 0;
3863 if (st->req() != MemNode::ValueIn + 1)
3864 return FAIL; // an inscrutable StoreNode (card mark?)
3865 Node* ctl = st->in(MemNode::Control);
3866 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this))
3867 return FAIL; // must be unconditional after the initialization
3868 Node* mem = st->in(MemNode::Memory);
3869 if (!(mem->is_Proj() && mem->in(0) == this))
3870 return FAIL; // must not be preceded by other stores
3871 Node* adr = st->in(MemNode::Address);
3872 intptr_t offset;
3873 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3874 if (alloc == nullptr)
3875 return FAIL; // inscrutable address
3876 if (alloc != allocation())
3877 return FAIL; // wrong allocation! (store needs to float up)
3878 int size_in_bytes = st->memory_size();
3879 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
3880 return FAIL; // mismatched access
3881 }
3882 Node* val = st->in(MemNode::ValueIn);
3883
3884 if (!detect_init_independence(val, phase))
3885 return FAIL; // stored value must be 'simple enough'
3886
3887 // The Store can be captured only if nothing after the allocation
3888 // and before the Store is using the memory location that the store
3889 // overwrites.
3890 bool failed = false;
3891 // If is_complete_with_arraycopy() is true the shape of the graph is
3892 // well defined and is safe so no need for extra checks.
3893 if (!is_complete_with_arraycopy()) {
3894 // We are going to look at each use of the memory state following
3895 // the allocation to make sure nothing reads the memory that the
3896 // Store writes.
3897 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3898 int alias_idx = phase->C->get_alias_index(t_adr);
3899 ResourceMark rm;
3900 Unique_Node_List mems;
3901 mems.push(mem);
3902 Node* unique_merge = nullptr;
3903 for (uint next = 0; next < mems.size(); ++next) {
3904 Node *m = mems.at(next);
3905 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3906 Node *n = m->fast_out(j);
3907 if (n->outcnt() == 0) {
3908 continue;
3909 }
3910 if (n == st) {
3911 continue;
3912 } else if (n->in(0) != nullptr && n->in(0) != ctl) {
3913 // If the control of this use is different from the control
3914 // of the Store which is right after the InitializeNode then
3915 // this node cannot be between the InitializeNode and the
3916 // Store.
3917 continue;
3918 } else if (n->is_MergeMem()) {
3919 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3920 // We can hit a MergeMemNode (that will likely go away
3921 // later) that is a direct use of the memory state
3922 // following the InitializeNode on the same slice as the
3923 // store node that we'd like to capture. We need to check
3924 // the uses of the MergeMemNode.
3925 mems.push(n);
3926 }
3927 } else if (n->is_Mem()) {
3928 Node* other_adr = n->in(MemNode::Address);
3929 if (other_adr == adr) {
3930 failed = true;
3931 break;
3932 } else {
3933 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3934 if (other_t_adr != nullptr) {
3935 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3936 if (other_alias_idx == alias_idx) {
3937 // A load from the same memory slice as the store right
3938 // after the InitializeNode. We check the control of the
3939 // object/array that is loaded from. If it's the same as
3940 // the store control then we cannot capture the store.
3941 assert(!n->is_Store(), "2 stores to same slice on same control?");
3942 Node* base = other_adr;
3943 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3944 base = base->in(AddPNode::Base);
3945 if (base != nullptr) {
3946 base = base->uncast();
3947 if (base->is_Proj() && base->in(0) == alloc) {
3948 failed = true;
3949 break;
3950 }
3951 }
3952 }
3953 }
3954 }
3955 } else {
3956 failed = true;
3957 break;
3958 }
3959 }
3960 }
3961 }
3962 if (failed) {
3963 if (!can_reshape) {
3964 // We decided we couldn't capture the store during parsing. We
3965 // should try again during the next IGVN once the graph is
3966 // cleaner.
3967 phase->C->record_for_igvn(st);
3968 }
3969 return FAIL;
3970 }
3971
3972 return offset; // success
3973 }
3974
3975 // Find the captured store in(i) which corresponds to the range
3976 // [start..start+size) in the initialized object.
3977 // If there is one, return its index i. If there isn't, return the
3978 // negative of the index where it should be inserted.
3979 // Return 0 if the queried range overlaps an initialization boundary
3980 // or if dead code is encountered.
3981 // If size_in_bytes is zero, do not bother with overlap checks.
3982 int InitializeNode::captured_store_insertion_point(intptr_t start,
3983 int size_in_bytes,
3984 PhaseValues* phase) {
3985 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
3986
3987 if (is_complete())
3988 return FAIL; // arraycopy got here first; punt
3989
3990 assert(allocation() != nullptr, "must be present");
3991
3992 // no negatives, no header fields:
3993 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3994
3995 // after a certain size, we bail out on tracking all the stores:
3996 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3997 if (start >= ti_limit) return FAIL;
3998
3999 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
4000 if (i >= limit) return -(int)i; // not found; here is where to put it
4001
4002 Node* st = in(i);
4003 intptr_t st_off = get_store_offset(st, phase);
4004 if (st_off < 0) {
4005 if (st != zero_memory()) {
4006 return FAIL; // bail out if there is dead garbage
4007 }
4008 } else if (st_off > start) {
4009 // ...we are done, since stores are ordered
4010 if (st_off < start + size_in_bytes) {
4011 return FAIL; // the next store overlaps
4012 }
4013 return -(int)i; // not found; here is where to put it
4014 } else if (st_off < start) {
4015 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
4016 if (size_in_bytes != 0 &&
4017 start < st_off + MAX_STORE &&
4018 start < st_off + st->as_Store()->memory_size()) {
4019 return FAIL; // the previous store overlaps
4020 }
4021 } else {
4022 if (size_in_bytes != 0 &&
4023 st->as_Store()->memory_size() != size_in_bytes) {
4024 return FAIL; // mismatched store size
4025 }
4026 return i;
4027 }
4028
4029 ++i;
4030 }
4031 }
4032
4033 // Look for a captured store which initializes at the offset 'start'
4034 // with the given size. If there is no such store, and no other
4035 // initialization interferes, then return zero_memory (the memory
4036 // projection of the AllocateNode).
4037 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
4038 PhaseValues* phase) {
4039 assert(stores_are_sane(phase), "");
4040 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4041 if (i == 0) {
4042 return nullptr; // something is dead
4043 } else if (i < 0) {
4044 return zero_memory(); // just primordial zero bits here
4045 } else {
4046 Node* st = in(i); // here is the store at this position
4047 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
4048 return st;
4049 }
4050 }
4051
4052 // Create, as a raw pointer, an address within my new object at 'offset'.
4053 Node* InitializeNode::make_raw_address(intptr_t offset,
4054 PhaseGVN* phase) {
4055 Node* addr = in(RawAddress);
4056 if (offset != 0) {
4057 Compile* C = phase->C;
4058 addr = phase->transform( new AddPNode(C->top(), addr,
4059 phase->MakeConX(offset)) );
4060 }
4061 return addr;
4062 }
4063
4064 // Clone the given store, converting it into a raw store
4065 // initializing a field or element of my new object.
4066 // Caller is responsible for retiring the original store,
4067 // with subsume_node or the like.
4068 //
4069 // From the example above InitializeNode::InitializeNode,
4070 // here are the old stores to be captured:
4071 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4072 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4073 //
4074 // Here is the changed code; note the extra edges on init:
4075 // alloc = (Allocate ...)
4076 // rawoop = alloc.RawAddress
4077 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
4078 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
4079 // init = (Initialize alloc.Control alloc.Memory rawoop
4080 // rawstore1 rawstore2)
4081 //
4082 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
4083 PhaseGVN* phase, bool can_reshape) {
4084 assert(stores_are_sane(phase), "");
4085
4086 if (start < 0) return nullptr;
4087 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
4088
4089 Compile* C = phase->C;
4090 int size_in_bytes = st->memory_size();
4091 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4092 if (i == 0) return nullptr; // bail out
4093 Node* prev_mem = nullptr; // raw memory for the captured store
4094 if (i > 0) {
4095 prev_mem = in(i); // there is a pre-existing store under this one
4096 set_req(i, C->top()); // temporarily disconnect it
4097 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
4098 } else {
4099 i = -i; // no pre-existing store
4100 prev_mem = zero_memory(); // a slice of the newly allocated object
4101 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
4102 set_req(--i, C->top()); // reuse this edge; it has been folded away
4103 else
4104 ins_req(i, C->top()); // build a new edge
4105 }
4106 Node* new_st = st->clone();
4107 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4108 new_st->set_req(MemNode::Control, in(Control));
4109 new_st->set_req(MemNode::Memory, prev_mem);
4110 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
4111 bs->eliminate_gc_barrier_data(new_st);
4112 new_st = phase->transform(new_st);
4113
4114 // At this point, new_st might have swallowed a pre-existing store
4115 // at the same offset, or perhaps new_st might have disappeared,
4116 // if it redundantly stored the same value (or zero to fresh memory).
4117
4118 // In any case, wire it in:
4119 PhaseIterGVN* igvn = phase->is_IterGVN();
4120 if (igvn) {
4121 igvn->rehash_node_delayed(this);
4122 }
4123 set_req(i, new_st);
4124
4125 // The caller may now kill the old guy.
4126 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
4127 assert(check_st == new_st || check_st == nullptr, "must be findable");
4128 assert(!is_complete(), "");
4129 return new_st;
4130 }
4131
4132 static bool store_constant(jlong* tiles, int num_tiles,
4133 intptr_t st_off, int st_size,
4134 jlong con) {
4135 if ((st_off & (st_size-1)) != 0)
4136 return false; // strange store offset (assume size==2**N)
4137 address addr = (address)tiles + st_off;
4138 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
4139 switch (st_size) {
4140 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
4141 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
4142 case sizeof(jint): *(jint*) addr = (jint) con; break;
4143 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
4144 default: return false; // strange store size (detect size!=2**N here)
4145 }
4146 return true; // return success to caller
4147 }
4148
4149 // Coalesce subword constants into int constants and possibly
4150 // into long constants. The goal, if the CPU permits,
4151 // is to initialize the object with a small number of 64-bit tiles.
4152 // Also, convert floating-point constants to bit patterns.
4153 // Non-constants are not relevant to this pass.
4154 //
4155 // In terms of the running example on InitializeNode::InitializeNode
4156 // and InitializeNode::capture_store, here is the transformation
4157 // of rawstore1 and rawstore2 into rawstore12:
4158 // alloc = (Allocate ...)
4159 // rawoop = alloc.RawAddress
4160 // tile12 = 0x00010002
4161 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
4162 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
4163 //
4164 void
4165 InitializeNode::coalesce_subword_stores(intptr_t header_size,
4166 Node* size_in_bytes,
4167 PhaseGVN* phase) {
4168 Compile* C = phase->C;
4169
4170 assert(stores_are_sane(phase), "");
4171 // Note: After this pass, they are not completely sane,
4172 // since there may be some overlaps.
4173
4174 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
4175
4176 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4177 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
4178 size_limit = MIN2(size_limit, ti_limit);
4179 size_limit = align_up(size_limit, BytesPerLong);
4180 int num_tiles = size_limit / BytesPerLong;
4181
4182 // allocate space for the tile map:
4183 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
4184 jlong tiles_buf[small_len];
4185 Node* nodes_buf[small_len];
4186 jlong inits_buf[small_len];
4187 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
4188 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4189 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
4190 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
4191 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
4192 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4193 // tiles: exact bitwise model of all primitive constants
4194 // nodes: last constant-storing node subsumed into the tiles model
4195 // inits: which bytes (in each tile) are touched by any initializations
4196
4197 //// Pass A: Fill in the tile model with any relevant stores.
4198
4199 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
4200 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
4201 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
4202 Node* zmem = zero_memory(); // initially zero memory state
4203 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4204 Node* st = in(i);
4205 intptr_t st_off = get_store_offset(st, phase);
4206
4207 // Figure out the store's offset and constant value:
4208 if (st_off < header_size) continue; //skip (ignore header)
4209 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
4210 int st_size = st->as_Store()->memory_size();
4211 if (st_off + st_size > size_limit) break;
4212
4213 // Record which bytes are touched, whether by constant or not.
4214 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
4215 continue; // skip (strange store size)
4216
4217 const Type* val = phase->type(st->in(MemNode::ValueIn));
4218 if (!val->singleton()) continue; //skip (non-con store)
4219 BasicType type = val->basic_type();
4220
4221 jlong con = 0;
4222 switch (type) {
4223 case T_INT: con = val->is_int()->get_con(); break;
4224 case T_LONG: con = val->is_long()->get_con(); break;
4225 case T_FLOAT: con = jint_cast(val->getf()); break;
4226 case T_DOUBLE: con = jlong_cast(val->getd()); break;
4227 default: continue; //skip (odd store type)
4228 }
4229
4230 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
4231 st->Opcode() == Op_StoreL) {
4232 continue; // This StoreL is already optimal.
4233 }
4234
4235 // Store down the constant.
4236 store_constant(tiles, num_tiles, st_off, st_size, con);
4237
4238 intptr_t j = st_off >> LogBytesPerLong;
4239
4240 if (type == T_INT && st_size == BytesPerInt
4241 && (st_off & BytesPerInt) == BytesPerInt) {
4242 jlong lcon = tiles[j];
4243 if (!Matcher::isSimpleConstant64(lcon) &&
4244 st->Opcode() == Op_StoreI) {
4245 // This StoreI is already optimal by itself.
4246 jint* intcon = (jint*) &tiles[j];
4247 intcon[1] = 0; // undo the store_constant()
4248
4249 // If the previous store is also optimal by itself, back up and
4250 // undo the action of the previous loop iteration... if we can.
4251 // But if we can't, just let the previous half take care of itself.
4252 st = nodes[j];
4253 st_off -= BytesPerInt;
4254 con = intcon[0];
4255 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) {
4256 assert(st_off >= header_size, "still ignoring header");
4257 assert(get_store_offset(st, phase) == st_off, "must be");
4258 assert(in(i-1) == zmem, "must be");
4259 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
4260 assert(con == tcon->is_int()->get_con(), "must be");
4261 // Undo the effects of the previous loop trip, which swallowed st:
4262 intcon[0] = 0; // undo store_constant()
4263 set_req(i-1, st); // undo set_req(i, zmem)
4264 nodes[j] = nullptr; // undo nodes[j] = st
4265 --old_subword; // undo ++old_subword
4266 }
4267 continue; // This StoreI is already optimal.
4268 }
4269 }
4270
4271 // This store is not needed.
4272 set_req(i, zmem);
4273 nodes[j] = st; // record for the moment
4274 if (st_size < BytesPerLong) // something has changed
4275 ++old_subword; // includes int/float, but who's counting...
4276 else ++old_long;
4277 }
4278
4279 if ((old_subword + old_long) == 0)
4280 return; // nothing more to do
4281
4282 //// Pass B: Convert any non-zero tiles into optimal constant stores.
4283 // Be sure to insert them before overlapping non-constant stores.
4284 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
4285 for (int j = 0; j < num_tiles; j++) {
4286 jlong con = tiles[j];
4287 jlong init = inits[j];
4288 if (con == 0) continue;
4289 jint con0, con1; // split the constant, address-wise
4290 jint init0, init1; // split the init map, address-wise
4291 { union { jlong con; jint intcon[2]; } u;
4292 u.con = con;
4293 con0 = u.intcon[0];
4294 con1 = u.intcon[1];
4295 u.con = init;
4296 init0 = u.intcon[0];
4297 init1 = u.intcon[1];
4298 }
4299
4300 Node* old = nodes[j];
4301 assert(old != nullptr, "need the prior store");
4302 intptr_t offset = (j * BytesPerLong);
4303
4304 bool split = !Matcher::isSimpleConstant64(con);
4305
4306 if (offset < header_size) {
4307 assert(offset + BytesPerInt >= header_size, "second int counts");
4308 assert(*(jint*)&tiles[j] == 0, "junk in header");
4309 split = true; // only the second word counts
4310 // Example: int a[] = { 42 ... }
4311 } else if (con0 == 0 && init0 == -1) {
4312 split = true; // first word is covered by full inits
4313 // Example: int a[] = { ... foo(), 42 ... }
4314 } else if (con1 == 0 && init1 == -1) {
4315 split = true; // second word is covered by full inits
4316 // Example: int a[] = { ... 42, foo() ... }
4317 }
4318
4319 // Here's a case where init0 is neither 0 nor -1:
4320 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
4321 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
4322 // In this case the tile is not split; it is (jlong)42.
4323 // The big tile is stored down, and then the foo() value is inserted.
4324 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
4325
4326 Node* ctl = old->in(MemNode::Control);
4327 Node* adr = make_raw_address(offset, phase);
4328 const TypePtr* atp = TypeRawPtr::BOTTOM;
4329
4330 // One or two coalesced stores to plop down.
4331 Node* st[2];
4332 intptr_t off[2];
4333 int nst = 0;
4334 if (!split) {
4335 ++new_long;
4336 off[nst] = offset;
4337 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4338 phase->longcon(con), T_LONG, MemNode::unordered);
4339 } else {
4340 // Omit either if it is a zero.
4341 if (con0 != 0) {
4342 ++new_int;
4343 off[nst] = offset;
4344 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4345 phase->intcon(con0), T_INT, MemNode::unordered);
4346 }
4347 if (con1 != 0) {
4348 ++new_int;
4349 offset += BytesPerInt;
4350 adr = make_raw_address(offset, phase);
4351 off[nst] = offset;
4352 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4353 phase->intcon(con1), T_INT, MemNode::unordered);
4354 }
4355 }
4356
4357 // Insert second store first, then the first before the second.
4358 // Insert each one just before any overlapping non-constant stores.
4359 while (nst > 0) {
4360 Node* st1 = st[--nst];
4361 C->copy_node_notes_to(st1, old);
4362 st1 = phase->transform(st1);
4363 offset = off[nst];
4364 assert(offset >= header_size, "do not smash header");
4365 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4366 guarantee(ins_idx != 0, "must re-insert constant store");
4367 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
4368 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4369 set_req(--ins_idx, st1);
4370 else
4371 ins_req(ins_idx, st1);
4372 }
4373 }
4374
4375 if (PrintCompilation && WizardMode)
4376 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4377 old_subword, old_long, new_int, new_long);
4378 if (C->log() != nullptr)
4379 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4380 old_subword, old_long, new_int, new_long);
4381
4382 // Clean up any remaining occurrences of zmem:
4383 remove_extra_zeroes();
4384 }
4385
4386 // Explore forward from in(start) to find the first fully initialized
4387 // word, and return its offset. Skip groups of subword stores which
4388 // together initialize full words. If in(start) is itself part of a
4389 // fully initialized word, return the offset of in(start). If there
4390 // are no following full-word stores, or if something is fishy, return
4391 // a negative value.
4392 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4393 int int_map = 0;
4394 intptr_t int_map_off = 0;
4395 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
4396
4397 for (uint i = start, limit = req(); i < limit; i++) {
4398 Node* st = in(i);
4399
4400 intptr_t st_off = get_store_offset(st, phase);
4401 if (st_off < 0) break; // return conservative answer
4402
4403 int st_size = st->as_Store()->memory_size();
4404 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4405 return st_off; // we found a complete word init
4406 }
4407
4408 // update the map:
4409
4410 intptr_t this_int_off = align_down(st_off, BytesPerInt);
4411 if (this_int_off != int_map_off) {
4412 // reset the map:
4413 int_map = 0;
4414 int_map_off = this_int_off;
4415 }
4416
4417 int subword_off = st_off - this_int_off;
4418 int_map |= right_n_bits(st_size) << subword_off;
4419 if ((int_map & FULL_MAP) == FULL_MAP) {
4420 return this_int_off; // we found a complete word init
4421 }
4422
4423 // Did this store hit or cross the word boundary?
4424 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4425 if (next_int_off == this_int_off + BytesPerInt) {
4426 // We passed the current int, without fully initializing it.
4427 int_map_off = next_int_off;
4428 int_map >>= BytesPerInt;
4429 } else if (next_int_off > this_int_off + BytesPerInt) {
4430 // We passed the current and next int.
4431 return this_int_off + BytesPerInt;
4432 }
4433 }
4434
4435 return -1;
4436 }
4437
4438
4439 // Called when the associated AllocateNode is expanded into CFG.
4440 // At this point, we may perform additional optimizations.
4441 // Linearize the stores by ascending offset, to make memory
4442 // activity as coherent as possible.
4443 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4444 intptr_t header_size,
4445 Node* size_in_bytes,
4446 PhaseIterGVN* phase) {
4447 assert(!is_complete(), "not already complete");
4448 assert(stores_are_sane(phase), "");
4449 assert(allocation() != nullptr, "must be present");
4450
4451 remove_extra_zeroes();
4452
4453 if (ReduceFieldZeroing || ReduceBulkZeroing)
4454 // reduce instruction count for common initialization patterns
4455 coalesce_subword_stores(header_size, size_in_bytes, phase);
4456
4457 Node* zmem = zero_memory(); // initially zero memory state
4458 Node* inits = zmem; // accumulating a linearized chain of inits
4459 #ifdef ASSERT
4460 intptr_t first_offset = allocation()->minimum_header_size();
4461 intptr_t last_init_off = first_offset; // previous init offset
4462 intptr_t last_init_end = first_offset; // previous init offset+size
4463 intptr_t last_tile_end = first_offset; // previous tile offset+size
4464 #endif
4465 intptr_t zeroes_done = header_size;
4466
4467 bool do_zeroing = true; // we might give up if inits are very sparse
4468 int big_init_gaps = 0; // how many large gaps have we seen?
4469
4470 if (UseTLAB && ZeroTLAB) do_zeroing = false;
4471 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
4472
4473 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4474 Node* st = in(i);
4475 intptr_t st_off = get_store_offset(st, phase);
4476 if (st_off < 0)
4477 break; // unknown junk in the inits
4478 if (st->in(MemNode::Memory) != zmem)
4479 break; // complicated store chains somehow in list
4480
4481 int st_size = st->as_Store()->memory_size();
4482 intptr_t next_init_off = st_off + st_size;
4483
4484 if (do_zeroing && zeroes_done < next_init_off) {
4485 // See if this store needs a zero before it or under it.
4486 intptr_t zeroes_needed = st_off;
4487
4488 if (st_size < BytesPerInt) {
4489 // Look for subword stores which only partially initialize words.
4490 // If we find some, we must lay down some word-level zeroes first,
4491 // underneath the subword stores.
4492 //
4493 // Examples:
4494 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
4495 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
4496 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
4497 //
4498 // Note: coalesce_subword_stores may have already done this,
4499 // if it was prompted by constant non-zero subword initializers.
4500 // But this case can still arise with non-constant stores.
4501
4502 intptr_t next_full_store = find_next_fullword_store(i, phase);
4503
4504 // In the examples above:
4505 // in(i) p q r s x y z
4506 // st_off 12 13 14 15 12 13 14
4507 // st_size 1 1 1 1 1 1 1
4508 // next_full_s. 12 16 16 16 16 16 16
4509 // z's_done 12 16 16 16 12 16 12
4510 // z's_needed 12 16 16 16 16 16 16
4511 // zsize 0 0 0 0 4 0 4
4512 if (next_full_store < 0) {
4513 // Conservative tack: Zero to end of current word.
4514 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4515 } else {
4516 // Zero to beginning of next fully initialized word.
4517 // Or, don't zero at all, if we are already in that word.
4518 assert(next_full_store >= zeroes_needed, "must go forward");
4519 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4520 zeroes_needed = next_full_store;
4521 }
4522 }
4523
4524 if (zeroes_needed > zeroes_done) {
4525 intptr_t zsize = zeroes_needed - zeroes_done;
4526 // Do some incremental zeroing on rawmem, in parallel with inits.
4527 zeroes_done = align_down(zeroes_done, BytesPerInt);
4528 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4529 zeroes_done, zeroes_needed,
4530 phase);
4531 zeroes_done = zeroes_needed;
4532 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4533 do_zeroing = false; // leave the hole, next time
4534 }
4535 }
4536
4537 // Collect the store and move on:
4538 phase->replace_input_of(st, MemNode::Memory, inits);
4539 inits = st; // put it on the linearized chain
4540 set_req(i, zmem); // unhook from previous position
4541
4542 if (zeroes_done == st_off)
4543 zeroes_done = next_init_off;
4544
4545 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4546
4547 #ifdef ASSERT
4548 // Various order invariants. Weaker than stores_are_sane because
4549 // a large constant tile can be filled in by smaller non-constant stores.
4550 assert(st_off >= last_init_off, "inits do not reverse");
4551 last_init_off = st_off;
4552 const Type* val = nullptr;
4553 if (st_size >= BytesPerInt &&
4554 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4555 (int)val->basic_type() < (int)T_OBJECT) {
4556 assert(st_off >= last_tile_end, "tiles do not overlap");
4557 assert(st_off >= last_init_end, "tiles do not overwrite inits");
4558 last_tile_end = MAX2(last_tile_end, next_init_off);
4559 } else {
4560 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4561 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4562 assert(st_off >= last_init_end, "inits do not overlap");
4563 last_init_end = next_init_off; // it's a non-tile
4564 }
4565 #endif //ASSERT
4566 }
4567
4568 remove_extra_zeroes(); // clear out all the zmems left over
4569 add_req(inits);
4570
4571 if (!(UseTLAB && ZeroTLAB)) {
4572 // If anything remains to be zeroed, zero it all now.
4573 zeroes_done = align_down(zeroes_done, BytesPerInt);
4574 // if it is the last unused 4 bytes of an instance, forget about it
4575 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4576 if (zeroes_done + BytesPerLong >= size_limit) {
4577 AllocateNode* alloc = allocation();
4578 assert(alloc != nullptr, "must be present");
4579 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
4580 Node* klass_node = alloc->in(AllocateNode::KlassNode);
4581 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
4582 if (zeroes_done == k->layout_helper())
4583 zeroes_done = size_limit;
4584 }
4585 }
4586 if (zeroes_done < size_limit) {
4587 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4588 zeroes_done, size_in_bytes, phase);
4589 }
4590 }
4591
4592 set_complete(phase);
4593 return rawmem;
4594 }
4595
4596
4597 #ifdef ASSERT
4598 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
4599 if (is_complete())
4600 return true; // stores could be anything at this point
4601 assert(allocation() != nullptr, "must be present");
4602 intptr_t last_off = allocation()->minimum_header_size();
4603 for (uint i = InitializeNode::RawStores; i < req(); i++) {
4604 Node* st = in(i);
4605 intptr_t st_off = get_store_offset(st, phase);
4606 if (st_off < 0) continue; // ignore dead garbage
4607 if (last_off > st_off) {
4608 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4609 this->dump(2);
4610 assert(false, "ascending store offsets");
4611 return false;
4612 }
4613 last_off = st_off + st->as_Store()->memory_size();
4614 }
4615 return true;
4616 }
4617 #endif //ASSERT
4618
4619
4620
4621
4622 //============================MergeMemNode=====================================
4623 //
4624 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4625 // contributing store or call operations. Each contributor provides the memory
4626 // state for a particular "alias type" (see Compile::alias_type). For example,
4627 // if a MergeMem has an input X for alias category #6, then any memory reference
4628 // to alias category #6 may use X as its memory state input, as an exact equivalent
4629 // to using the MergeMem as a whole.
4630 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4631 //
4632 // (Here, the <N> notation gives the index of the relevant adr_type.)
4633 //
4634 // In one special case (and more cases in the future), alias categories overlap.
4635 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4636 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4637 // it is exactly equivalent to that state W:
4638 // MergeMem(<Bot>: W) <==> W
4639 //
4640 // Usually, the merge has more than one input. In that case, where inputs
4641 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4642 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4643 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4644 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4645 //
4646 // A merge can take a "wide" memory state as one of its narrow inputs.
4647 // This simply means that the merge observes out only the relevant parts of
4648 // the wide input. That is, wide memory states arriving at narrow merge inputs
4649 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4650 //
4651 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4652 // and that memory slices "leak through":
4653 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4654 //
4655 // But, in such a cascade, repeated memory slices can "block the leak":
4656 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4657 //
4658 // In the last example, Y is not part of the combined memory state of the
4659 // outermost MergeMem. The system must, of course, prevent unschedulable
4660 // memory states from arising, so you can be sure that the state Y is somehow
4661 // a precursor to state Y'.
4662 //
4663 //
4664 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4665 // of each MergeMemNode array are exactly the numerical alias indexes, including
4666 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4667 // Compile::alias_type (and kin) produce and manage these indexes.
4668 //
4669 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4670 // (Note that this provides quick access to the top node inside MergeMem methods,
4671 // without the need to reach out via TLS to Compile::current.)
4672 //
4673 // As a consequence of what was just described, a MergeMem that represents a full
4674 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4675 // containing all alias categories.
4676 //
4677 // MergeMem nodes never (?) have control inputs, so in(0) is null.
4678 //
4679 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4680 // a memory state for the alias type <N>, or else the top node, meaning that
4681 // there is no particular input for that alias type. Note that the length of
4682 // a MergeMem is variable, and may be extended at any time to accommodate new
4683 // memory states at larger alias indexes. When merges grow, they are of course
4684 // filled with "top" in the unused in() positions.
4685 //
4686 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4687 // (Top was chosen because it works smoothly with passes like GCM.)
4688 //
4689 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4690 // the type of random VM bits like TLS references.) Since it is always the
4691 // first non-Bot memory slice, some low-level loops use it to initialize an
4692 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4693 //
4694 //
4695 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4696 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4697 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4698 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4699 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4700 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4701 //
4702 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4703 // really that different from the other memory inputs. An abbreviation called
4704 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4705 //
4706 //
4707 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4708 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4709 // that "emerges though" the base memory will be marked as excluding the alias types
4710 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4711 //
4712 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4713 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4714 //
4715 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4716 // (It is currently unimplemented.) As you can see, the resulting merge is
4717 // actually a disjoint union of memory states, rather than an overlay.
4718 //
4719
4720 //------------------------------MergeMemNode-----------------------------------
4721 Node* MergeMemNode::make_empty_memory() {
4722 Node* empty_memory = (Node*) Compile::current()->top();
4723 assert(empty_memory->is_top(), "correct sentinel identity");
4724 return empty_memory;
4725 }
4726
4727 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4728 init_class_id(Class_MergeMem);
4729 // all inputs are nullified in Node::Node(int)
4730 // set_input(0, nullptr); // no control input
4731
4732 // Initialize the edges uniformly to top, for starters.
4733 Node* empty_mem = make_empty_memory();
4734 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4735 init_req(i,empty_mem);
4736 }
4737 assert(empty_memory() == empty_mem, "");
4738
4739 if( new_base != nullptr && new_base->is_MergeMem() ) {
4740 MergeMemNode* mdef = new_base->as_MergeMem();
4741 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4742 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4743 mms.set_memory(mms.memory2());
4744 }
4745 assert(base_memory() == mdef->base_memory(), "");
4746 } else {
4747 set_base_memory(new_base);
4748 }
4749 }
4750
4751 // Make a new, untransformed MergeMem with the same base as 'mem'.
4752 // If mem is itself a MergeMem, populate the result with the same edges.
4753 MergeMemNode* MergeMemNode::make(Node* mem) {
4754 return new MergeMemNode(mem);
4755 }
4756
4757 //------------------------------cmp--------------------------------------------
4758 uint MergeMemNode::hash() const { return NO_HASH; }
4759 bool MergeMemNode::cmp( const Node &n ) const {
4760 return (&n == this); // Always fail except on self
4761 }
4762
4763 //------------------------------Identity---------------------------------------
4764 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4765 // Identity if this merge point does not record any interesting memory
4766 // disambiguations.
4767 Node* base_mem = base_memory();
4768 Node* empty_mem = empty_memory();
4769 if (base_mem != empty_mem) { // Memory path is not dead?
4770 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4771 Node* mem = in(i);
4772 if (mem != empty_mem && mem != base_mem) {
4773 return this; // Many memory splits; no change
4774 }
4775 }
4776 }
4777 return base_mem; // No memory splits; ID on the one true input
4778 }
4779
4780 //------------------------------Ideal------------------------------------------
4781 // This method is invoked recursively on chains of MergeMem nodes
4782 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4783 // Remove chain'd MergeMems
4784 //
4785 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4786 // relative to the "in(Bot)". Since we are patching both at the same time,
4787 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4788 // but rewrite each "in(i)" relative to the new "in(Bot)".
4789 Node *progress = nullptr;
4790
4791
4792 Node* old_base = base_memory();
4793 Node* empty_mem = empty_memory();
4794 if (old_base == empty_mem)
4795 return nullptr; // Dead memory path.
4796
4797 MergeMemNode* old_mbase;
4798 if (old_base != nullptr && old_base->is_MergeMem())
4799 old_mbase = old_base->as_MergeMem();
4800 else
4801 old_mbase = nullptr;
4802 Node* new_base = old_base;
4803
4804 // simplify stacked MergeMems in base memory
4805 if (old_mbase) new_base = old_mbase->base_memory();
4806
4807 // the base memory might contribute new slices beyond my req()
4808 if (old_mbase) grow_to_match(old_mbase);
4809
4810 // Note: We do not call verify_sparse on entry, because inputs
4811 // can normalize to the base_memory via subsume_node or similar
4812 // mechanisms. This method repairs that damage.
4813
4814 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4815
4816 // Look at each slice.
4817 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4818 Node* old_in = in(i);
4819 // calculate the old memory value
4820 Node* old_mem = old_in;
4821 if (old_mem == empty_mem) old_mem = old_base;
4822 assert(old_mem == memory_at(i), "");
4823
4824 // maybe update (reslice) the old memory value
4825
4826 // simplify stacked MergeMems
4827 Node* new_mem = old_mem;
4828 MergeMemNode* old_mmem;
4829 if (old_mem != nullptr && old_mem->is_MergeMem())
4830 old_mmem = old_mem->as_MergeMem();
4831 else
4832 old_mmem = nullptr;
4833 if (old_mmem == this) {
4834 // This can happen if loops break up and safepoints disappear.
4835 // A merge of BotPtr (default) with a RawPtr memory derived from a
4836 // safepoint can be rewritten to a merge of the same BotPtr with
4837 // the BotPtr phi coming into the loop. If that phi disappears
4838 // also, we can end up with a self-loop of the mergemem.
4839 // In general, if loops degenerate and memory effects disappear,
4840 // a mergemem can be left looking at itself. This simply means
4841 // that the mergemem's default should be used, since there is
4842 // no longer any apparent effect on this slice.
4843 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4844 // from start. Update the input to TOP.
4845 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4846 }
4847 else if (old_mmem != nullptr) {
4848 new_mem = old_mmem->memory_at(i);
4849 }
4850 // else preceding memory was not a MergeMem
4851
4852 // maybe store down a new value
4853 Node* new_in = new_mem;
4854 if (new_in == new_base) new_in = empty_mem;
4855
4856 if (new_in != old_in) {
4857 // Warning: Do not combine this "if" with the previous "if"
4858 // A memory slice might have be be rewritten even if it is semantically
4859 // unchanged, if the base_memory value has changed.
4860 set_req_X(i, new_in, phase);
4861 progress = this; // Report progress
4862 }
4863 }
4864
4865 if (new_base != old_base) {
4866 set_req_X(Compile::AliasIdxBot, new_base, phase);
4867 // Don't use set_base_memory(new_base), because we need to update du.
4868 assert(base_memory() == new_base, "");
4869 progress = this;
4870 }
4871
4872 if( base_memory() == this ) {
4873 // a self cycle indicates this memory path is dead
4874 set_req(Compile::AliasIdxBot, empty_mem);
4875 }
4876
4877 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4878 // Recursion must occur after the self cycle check above
4879 if( base_memory()->is_MergeMem() ) {
4880 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4881 Node *m = phase->transform(new_mbase); // Rollup any cycles
4882 if( m != nullptr &&
4883 (m->is_top() ||
4884 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4885 // propagate rollup of dead cycle to self
4886 set_req(Compile::AliasIdxBot, empty_mem);
4887 }
4888 }
4889
4890 if( base_memory() == empty_mem ) {
4891 progress = this;
4892 // Cut inputs during Parse phase only.
4893 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4894 if( !can_reshape ) {
4895 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4896 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4897 }
4898 }
4899 }
4900
4901 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4902 // Check if PhiNode::Ideal's "Split phis through memory merges"
4903 // transform should be attempted. Look for this->phi->this cycle.
4904 uint merge_width = req();
4905 if (merge_width > Compile::AliasIdxRaw) {
4906 PhiNode* phi = base_memory()->as_Phi();
4907 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4908 if (phi->in(i) == this) {
4909 phase->is_IterGVN()->_worklist.push(phi);
4910 break;
4911 }
4912 }
4913 }
4914 }
4915
4916 assert(progress || verify_sparse(), "please, no dups of base");
4917 return progress;
4918 }
4919
4920 //-------------------------set_base_memory-------------------------------------
4921 void MergeMemNode::set_base_memory(Node *new_base) {
4922 Node* empty_mem = empty_memory();
4923 set_req(Compile::AliasIdxBot, new_base);
4924 assert(memory_at(req()) == new_base, "must set default memory");
4925 // Clear out other occurrences of new_base:
4926 if (new_base != empty_mem) {
4927 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4928 if (in(i) == new_base) set_req(i, empty_mem);
4929 }
4930 }
4931 }
4932
4933 //------------------------------out_RegMask------------------------------------
4934 const RegMask &MergeMemNode::out_RegMask() const {
4935 return RegMask::Empty;
4936 }
4937
4938 //------------------------------dump_spec--------------------------------------
4939 #ifndef PRODUCT
4940 void MergeMemNode::dump_spec(outputStream *st) const {
4941 st->print(" {");
4942 Node* base_mem = base_memory();
4943 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4944 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem;
4945 if (mem == base_mem) { st->print(" -"); continue; }
4946 st->print( " N%d:", mem->_idx );
4947 Compile::current()->get_adr_type(i)->dump_on(st);
4948 }
4949 st->print(" }");
4950 }
4951 #endif // !PRODUCT
4952
4953
4954 #ifdef ASSERT
4955 static bool might_be_same(Node* a, Node* b) {
4956 if (a == b) return true;
4957 if (!(a->is_Phi() || b->is_Phi())) return false;
4958 // phis shift around during optimization
4959 return true; // pretty stupid...
4960 }
4961
4962 // verify a narrow slice (either incoming or outgoing)
4963 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4964 if (!VerifyAliases) return; // don't bother to verify unless requested
4965 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
4966 if (Node::in_dump()) return; // muzzle asserts when printing
4967 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4968 assert(n != nullptr, "");
4969 // Elide intervening MergeMem's
4970 while (n->is_MergeMem()) {
4971 n = n->as_MergeMem()->memory_at(alias_idx);
4972 }
4973 Compile* C = Compile::current();
4974 const TypePtr* n_adr_type = n->adr_type();
4975 if (n == m->empty_memory()) {
4976 // Implicit copy of base_memory()
4977 } else if (n_adr_type != TypePtr::BOTTOM) {
4978 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type");
4979 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4980 } else {
4981 // A few places like make_runtime_call "know" that VM calls are narrow,
4982 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4983 bool expected_wide_mem = false;
4984 if (n == m->base_memory()) {
4985 expected_wide_mem = true;
4986 } else if (alias_idx == Compile::AliasIdxRaw ||
4987 n == m->memory_at(Compile::AliasIdxRaw)) {
4988 expected_wide_mem = true;
4989 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4990 // memory can "leak through" calls on channels that
4991 // are write-once. Allow this also.
4992 expected_wide_mem = true;
4993 }
4994 assert(expected_wide_mem, "expected narrow slice replacement");
4995 }
4996 }
4997 #else // !ASSERT
4998 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4999 #endif
5000
5001
5002 //-----------------------------memory_at---------------------------------------
5003 Node* MergeMemNode::memory_at(uint alias_idx) const {
5004 assert(alias_idx >= Compile::AliasIdxRaw ||
5005 (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()),
5006 "must avoid base_memory and AliasIdxTop");
5007
5008 // Otherwise, it is a narrow slice.
5009 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
5010 if (is_empty_memory(n)) {
5011 // the array is sparse; empty slots are the "top" node
5012 n = base_memory();
5013 assert(Node::in_dump()
5014 || n == nullptr || n->bottom_type() == Type::TOP
5015 || n->adr_type() == nullptr // address is TOP
5016 || n->adr_type() == TypePtr::BOTTOM
5017 || n->adr_type() == TypeRawPtr::BOTTOM
5018 || !Compile::current()->do_aliasing(),
5019 "must be a wide memory");
5020 // do_aliasing == false if we are organizing the memory states manually.
5021 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
5022 } else {
5023 // make sure the stored slice is sane
5024 #ifdef ASSERT
5025 if (VMError::is_error_reported() || Node::in_dump()) {
5026 } else if (might_be_same(n, base_memory())) {
5027 // Give it a pass: It is a mostly harmless repetition of the base.
5028 // This can arise normally from node subsumption during optimization.
5029 } else {
5030 verify_memory_slice(this, alias_idx, n);
5031 }
5032 #endif
5033 }
5034 return n;
5035 }
5036
5037 //---------------------------set_memory_at-------------------------------------
5038 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
5039 verify_memory_slice(this, alias_idx, n);
5040 Node* empty_mem = empty_memory();
5041 if (n == base_memory()) n = empty_mem; // collapse default
5042 uint need_req = alias_idx+1;
5043 if (req() < need_req) {
5044 if (n == empty_mem) return; // already the default, so do not grow me
5045 // grow the sparse array
5046 do {
5047 add_req(empty_mem);
5048 } while (req() < need_req);
5049 }
5050 set_req( alias_idx, n );
5051 }
5052
5053
5054
5055 //--------------------------iteration_setup------------------------------------
5056 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
5057 if (other != nullptr) {
5058 grow_to_match(other);
5059 // invariant: the finite support of mm2 is within mm->req()
5060 #ifdef ASSERT
5061 for (uint i = req(); i < other->req(); i++) {
5062 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
5063 }
5064 #endif
5065 }
5066 // Replace spurious copies of base_memory by top.
5067 Node* base_mem = base_memory();
5068 if (base_mem != nullptr && !base_mem->is_top()) {
5069 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
5070 if (in(i) == base_mem)
5071 set_req(i, empty_memory());
5072 }
5073 }
5074 }
5075
5076 //---------------------------grow_to_match-------------------------------------
5077 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
5078 Node* empty_mem = empty_memory();
5079 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
5080 // look for the finite support of the other memory
5081 for (uint i = other->req(); --i >= req(); ) {
5082 if (other->in(i) != empty_mem) {
5083 uint new_len = i+1;
5084 while (req() < new_len) add_req(empty_mem);
5085 break;
5086 }
5087 }
5088 }
5089
5090 //---------------------------verify_sparse-------------------------------------
5091 #ifndef PRODUCT
5092 bool MergeMemNode::verify_sparse() const {
5093 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
5094 Node* base_mem = base_memory();
5095 // The following can happen in degenerate cases, since empty==top.
5096 if (is_empty_memory(base_mem)) return true;
5097 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5098 assert(in(i) != nullptr, "sane slice");
5099 if (in(i) == base_mem) return false; // should have been the sentinel value!
5100 }
5101 return true;
5102 }
5103
5104 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
5105 Node* n;
5106 n = mm->in(idx);
5107 if (mem == n) return true; // might be empty_memory()
5108 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
5109 if (mem == n) return true;
5110 return false;
5111 }
5112 #endif // !PRODUCT