1 /*
2 * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2024, Alibaba Group Holding Limited. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "precompiled.hpp"
27 #include "classfile/javaClasses.hpp"
28 #include "compiler/compileLog.hpp"
29 #include "gc/shared/barrierSet.hpp"
30 #include "gc/shared/c2/barrierSetC2.hpp"
31 #include "gc/shared/tlab_globals.hpp"
32 #include "memory/allocation.inline.hpp"
33 #include "memory/resourceArea.hpp"
34 #include "oops/objArrayKlass.hpp"
35 #include "opto/addnode.hpp"
36 #include "opto/arraycopynode.hpp"
37 #include "opto/cfgnode.hpp"
38 #include "opto/regalloc.hpp"
39 #include "opto/compile.hpp"
40 #include "opto/connode.hpp"
41 #include "opto/convertnode.hpp"
42 #include "opto/loopnode.hpp"
43 #include "opto/machnode.hpp"
44 #include "opto/matcher.hpp"
45 #include "opto/memnode.hpp"
46 #include "opto/mulnode.hpp"
47 #include "opto/narrowptrnode.hpp"
48 #include "opto/phaseX.hpp"
49 #include "opto/regmask.hpp"
50 #include "opto/rootnode.hpp"
51 #include "opto/vectornode.hpp"
52 #include "utilities/align.hpp"
53 #include "utilities/copy.hpp"
54 #include "utilities/macros.hpp"
55 #include "utilities/powerOfTwo.hpp"
56 #include "utilities/vmError.hpp"
57
58 // Portions of code courtesy of Clifford Click
59
60 // Optimization - Graph Style
61
62 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
63
64 //=============================================================================
65 uint MemNode::size_of() const { return sizeof(*this); }
66
67 const TypePtr *MemNode::adr_type() const {
68 Node* adr = in(Address);
69 if (adr == nullptr) return nullptr; // node is dead
70 const TypePtr* cross_check = nullptr;
71 DEBUG_ONLY(cross_check = _adr_type);
72 return calculate_adr_type(adr->bottom_type(), cross_check);
73 }
74
75 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
76 if (adr != nullptr) {
77 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
78 return true;
79 }
80 }
81 return false;
82 }
83
84 #ifndef PRODUCT
85 void MemNode::dump_spec(outputStream *st) const {
86 if (in(Address) == nullptr) return; // node is dead
87 #ifndef ASSERT
88 // fake the missing field
89 const TypePtr* _adr_type = nullptr;
90 if (in(Address) != nullptr)
91 _adr_type = in(Address)->bottom_type()->isa_ptr();
92 #endif
93 dump_adr_type(this, _adr_type, st);
94
95 Compile* C = Compile::current();
96 if (C->alias_type(_adr_type)->is_volatile()) {
97 st->print(" Volatile!");
98 }
99 if (_unaligned_access) {
100 st->print(" unaligned");
101 }
102 if (_mismatched_access) {
103 st->print(" mismatched");
104 }
105 if (_unsafe_access) {
106 st->print(" unsafe");
107 }
108 }
109
110 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
111 st->print(" @");
112 if (adr_type == nullptr) {
113 st->print("null");
114 } else {
115 adr_type->dump_on(st);
116 Compile* C = Compile::current();
117 Compile::AliasType* atp = nullptr;
118 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
119 if (atp == nullptr)
120 st->print(", idx=?\?;");
121 else if (atp->index() == Compile::AliasIdxBot)
122 st->print(", idx=Bot;");
123 else if (atp->index() == Compile::AliasIdxTop)
124 st->print(", idx=Top;");
125 else if (atp->index() == Compile::AliasIdxRaw)
126 st->print(", idx=Raw;");
127 else {
128 ciField* field = atp->field();
129 if (field) {
130 st->print(", name=");
131 field->print_name_on(st);
132 }
133 st->print(", idx=%d;", atp->index());
134 }
135 }
136 }
137
138 extern void print_alias_types();
139
140 #endif
141
142 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
143 assert((t_oop != nullptr), "sanity");
144 bool is_instance = t_oop->is_known_instance_field();
145 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
146 (load != nullptr) && load->is_Load() &&
147 (phase->is_IterGVN() != nullptr);
148 if (!(is_instance || is_boxed_value_load))
149 return mchain; // don't try to optimize non-instance types
150 uint instance_id = t_oop->instance_id();
151 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
152 Node *prev = nullptr;
153 Node *result = mchain;
154 while (prev != result) {
155 prev = result;
156 if (result == start_mem)
157 break; // hit one of our sentinels
158 // skip over a call which does not affect this memory slice
159 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
160 Node *proj_in = result->in(0);
161 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
162 break; // hit one of our sentinels
163 } else if (proj_in->is_Call()) {
164 // ArrayCopyNodes processed here as well
165 CallNode *call = proj_in->as_Call();
166 if (!call->may_modify(t_oop, phase)) { // returns false for instances
167 result = call->in(TypeFunc::Memory);
168 }
169 } else if (proj_in->is_Initialize()) {
170 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
171 // Stop if this is the initialization for the object instance which
172 // contains this memory slice, otherwise skip over it.
173 if ((alloc == nullptr) || (alloc->_idx == instance_id)) {
174 break;
175 }
176 if (is_instance) {
177 result = proj_in->in(TypeFunc::Memory);
178 } else if (is_boxed_value_load) {
179 Node* klass = alloc->in(AllocateNode::KlassNode);
180 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
181 if (tklass->klass_is_exact() && !tklass->exact_klass()->equals(t_oop->is_instptr()->exact_klass())) {
182 result = proj_in->in(TypeFunc::Memory); // not related allocation
183 }
184 }
185 } else if (proj_in->is_MemBar()) {
186 ArrayCopyNode* ac = nullptr;
187 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
188 break;
189 }
190 result = proj_in->in(TypeFunc::Memory);
191 } else if (proj_in->is_top()) {
192 break; // dead code
193 } else {
194 assert(false, "unexpected projection");
195 }
196 } else if (result->is_ClearArray()) {
197 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
198 // Can not bypass initialization of the instance
199 // we are looking for.
200 break;
201 }
202 // Otherwise skip it (the call updated 'result' value).
203 } else if (result->is_MergeMem()) {
204 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, nullptr, tty);
205 }
206 }
207 return result;
208 }
209
210 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
211 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
212 if (t_oop == nullptr)
213 return mchain; // don't try to optimize non-oop types
214 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
215 bool is_instance = t_oop->is_known_instance_field();
216 PhaseIterGVN *igvn = phase->is_IterGVN();
217 if (is_instance && igvn != nullptr && result->is_Phi()) {
218 PhiNode *mphi = result->as_Phi();
219 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
220 const TypePtr *t = mphi->adr_type();
221 bool do_split = false;
222 // In the following cases, Load memory input can be further optimized based on
223 // its precise address type
224 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ) {
225 do_split = true;
226 } else if (t->isa_oopptr() && !t->is_oopptr()->is_known_instance()) {
227 const TypeOopPtr* mem_t =
228 t->is_oopptr()->cast_to_exactness(true)
229 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
230 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id());
231 if (t_oop->isa_aryptr()) {
232 mem_t = mem_t->is_aryptr()
233 ->cast_to_stable(t_oop->is_aryptr()->is_stable())
234 ->cast_to_size(t_oop->is_aryptr()->size())
235 ->with_offset(t_oop->is_aryptr()->offset())
236 ->is_aryptr();
237 }
238 do_split = mem_t == t_oop;
239 }
240 if (do_split) {
241 // clone the Phi with our address type
242 result = mphi->split_out_instance(t_adr, igvn);
243 } else {
244 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
245 }
246 }
247 return result;
248 }
249
250 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
251 uint alias_idx = phase->C->get_alias_index(tp);
252 Node *mem = mmem;
253 #ifdef ASSERT
254 {
255 // Check that current type is consistent with the alias index used during graph construction
256 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
257 bool consistent = adr_check == nullptr || adr_check->empty() ||
258 phase->C->must_alias(adr_check, alias_idx );
259 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
260 if( !consistent && adr_check != nullptr && !adr_check->empty() &&
261 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
262 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
263 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
264 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
265 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
266 // don't assert if it is dead code.
267 consistent = true;
268 }
269 if( !consistent ) {
270 st->print("alias_idx==%d, adr_check==", alias_idx);
271 if( adr_check == nullptr ) {
272 st->print("null");
273 } else {
274 adr_check->dump();
275 }
276 st->cr();
277 print_alias_types();
278 assert(consistent, "adr_check must match alias idx");
279 }
280 }
281 #endif
282 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
283 // means an array I have not precisely typed yet. Do not do any
284 // alias stuff with it any time soon.
285 const TypeOopPtr *toop = tp->isa_oopptr();
286 if (tp->base() != Type::AnyPtr &&
287 !(toop &&
288 toop->isa_instptr() &&
289 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
290 toop->offset() == Type::OffsetBot)) {
291 // compress paths and change unreachable cycles to TOP
292 // If not, we can update the input infinitely along a MergeMem cycle
293 // Equivalent code in PhiNode::Ideal
294 Node* m = phase->transform(mmem);
295 // If transformed to a MergeMem, get the desired slice
296 // Otherwise the returned node represents memory for every slice
297 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
298 // Update input if it is progress over what we have now
299 }
300 return mem;
301 }
302
303 //--------------------------Ideal_common---------------------------------------
304 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
305 // Unhook non-raw memories from complete (macro-expanded) initializations.
306 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
307 // If our control input is a dead region, kill all below the region
308 Node *ctl = in(MemNode::Control);
309 if (ctl && remove_dead_region(phase, can_reshape))
310 return this;
311 ctl = in(MemNode::Control);
312 // Don't bother trying to transform a dead node
313 if (ctl && ctl->is_top()) return NodeSentinel;
314
315 PhaseIterGVN *igvn = phase->is_IterGVN();
316 // Wait if control on the worklist.
317 if (ctl && can_reshape && igvn != nullptr) {
318 Node* bol = nullptr;
319 Node* cmp = nullptr;
320 if (ctl->in(0)->is_If()) {
321 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
322 bol = ctl->in(0)->in(1);
323 if (bol->is_Bool())
324 cmp = ctl->in(0)->in(1)->in(1);
325 }
326 if (igvn->_worklist.member(ctl) ||
327 (bol != nullptr && igvn->_worklist.member(bol)) ||
328 (cmp != nullptr && igvn->_worklist.member(cmp)) ) {
329 // This control path may be dead.
330 // Delay this memory node transformation until the control is processed.
331 igvn->_worklist.push(this);
332 return NodeSentinel; // caller will return null
333 }
334 }
335 // Ignore if memory is dead, or self-loop
336 Node *mem = in(MemNode::Memory);
337 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return null
338 assert(mem != this, "dead loop in MemNode::Ideal");
339
340 if (can_reshape && igvn != nullptr && igvn->_worklist.member(mem)) {
341 // This memory slice may be dead.
342 // Delay this mem node transformation until the memory is processed.
343 igvn->_worklist.push(this);
344 return NodeSentinel; // caller will return null
345 }
346
347 Node *address = in(MemNode::Address);
348 const Type *t_adr = phase->type(address);
349 if (t_adr == Type::TOP) return NodeSentinel; // caller will return null
350
351 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
352 // Unsafe off-heap access with zero address. Remove access and other control users
353 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
354 assert(ctl != nullptr, "unsafe accesses should be control dependent");
355 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
356 Node* u = ctl->fast_out(i);
357 if (u != ctl) {
358 igvn->rehash_node_delayed(u);
359 int nb = u->replace_edge(ctl, phase->C->top(), igvn);
360 --i, imax -= nb;
361 }
362 }
363 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
364 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
365 phase->C->root()->add_req(halt);
366 return this;
367 }
368
369 if (can_reshape && igvn != nullptr &&
370 (igvn->_worklist.member(address) ||
371 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
372 // The address's base and type may change when the address is processed.
373 // Delay this mem node transformation until the address is processed.
374 igvn->_worklist.push(this);
375 return NodeSentinel; // caller will return null
376 }
377
378 // Do NOT remove or optimize the next lines: ensure a new alias index
379 // is allocated for an oop pointer type before Escape Analysis.
380 // Note: C++ will not remove it since the call has side effect.
381 if (t_adr->isa_oopptr()) {
382 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
383 }
384
385 Node* base = nullptr;
386 if (address->is_AddP()) {
387 base = address->in(AddPNode::Base);
388 }
389 if (base != nullptr && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
390 !t_adr->isa_rawptr()) {
391 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
392 // Skip this node optimization if its address has TOP base.
393 return NodeSentinel; // caller will return null
394 }
395
396 // Avoid independent memory operations
397 Node* old_mem = mem;
398
399 // The code which unhooks non-raw memories from complete (macro-expanded)
400 // initializations was removed. After macro-expansion all stores caught
401 // by Initialize node became raw stores and there is no information
402 // which memory slices they modify. So it is unsafe to move any memory
403 // operation above these stores. Also in most cases hooked non-raw memories
404 // were already unhooked by using information from detect_ptr_independence()
405 // and find_previous_store().
406
407 if (mem->is_MergeMem()) {
408 MergeMemNode* mmem = mem->as_MergeMem();
409 const TypePtr *tp = t_adr->is_ptr();
410
411 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
412 }
413
414 if (mem != old_mem) {
415 set_req_X(MemNode::Memory, mem, phase);
416 if (phase->type(mem) == Type::TOP) return NodeSentinel;
417 return this;
418 }
419
420 // let the subclass continue analyzing...
421 return nullptr;
422 }
423
424 // Helper function for proving some simple control dominations.
425 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
426 // Already assumes that 'dom' is available at 'sub', and that 'sub'
427 // is not a constant (dominated by the method's StartNode).
428 // Used by MemNode::find_previous_store to prove that the
429 // control input of a memory operation predates (dominates)
430 // an allocation it wants to look past.
431 // Returns 'DomResult::Dominate' if all control inputs of 'dom'
432 // dominate 'sub', 'DomResult::NotDominate' if not,
433 // and 'DomResult::EncounteredDeadCode' if we can't decide due to
434 // dead code, but at the end of IGVN, we know the definite result
435 // once the dead code is cleaned up.
436 Node::DomResult MemNode::maybe_all_controls_dominate(Node* dom, Node* sub) {
437 if (dom == nullptr || dom->is_top() || sub == nullptr || sub->is_top()) {
438 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
439 }
440
441 // Check 'dom'. Skip Proj and CatchProj nodes.
442 dom = dom->find_exact_control(dom);
443 if (dom == nullptr || dom->is_top()) {
444 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
445 }
446
447 if (dom == sub) {
448 // For the case when, for example, 'sub' is Initialize and the original
449 // 'dom' is Proj node of the 'sub'.
450 return DomResult::NotDominate;
451 }
452
453 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) {
454 return DomResult::Dominate;
455 }
456
457 // 'dom' dominates 'sub' if its control edge and control edges
458 // of all its inputs dominate or equal to sub's control edge.
459
460 // Currently 'sub' is either Allocate, Initialize or Start nodes.
461 // Or Region for the check in LoadNode::Ideal();
462 // 'sub' should have sub->in(0) != nullptr.
463 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
464 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
465
466 // Get control edge of 'sub'.
467 Node* orig_sub = sub;
468 sub = sub->find_exact_control(sub->in(0));
469 if (sub == nullptr || sub->is_top()) {
470 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
471 }
472
473 assert(sub->is_CFG(), "expecting control");
474
475 if (sub == dom) {
476 return DomResult::Dominate;
477 }
478
479 if (sub->is_Start() || sub->is_Root()) {
480 return DomResult::NotDominate;
481 }
482
483 {
484 // Check all control edges of 'dom'.
485
486 ResourceMark rm;
487 Node_List nlist;
488 Unique_Node_List dom_list;
489
490 dom_list.push(dom);
491 bool only_dominating_controls = false;
492
493 for (uint next = 0; next < dom_list.size(); next++) {
494 Node* n = dom_list.at(next);
495 if (n == orig_sub) {
496 return DomResult::NotDominate; // One of dom's inputs dominated by sub.
497 }
498 if (!n->is_CFG() && n->pinned()) {
499 // Check only own control edge for pinned non-control nodes.
500 n = n->find_exact_control(n->in(0));
501 if (n == nullptr || n->is_top()) {
502 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
503 }
504 assert(n->is_CFG(), "expecting control");
505 dom_list.push(n);
506 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
507 only_dominating_controls = true;
508 } else if (n->is_CFG()) {
509 DomResult dom_result = n->dominates(sub, nlist);
510 if (dom_result == DomResult::Dominate) {
511 only_dominating_controls = true;
512 } else {
513 return dom_result;
514 }
515 } else {
516 // First, own control edge.
517 Node* m = n->find_exact_control(n->in(0));
518 if (m != nullptr) {
519 if (m->is_top()) {
520 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
521 }
522 dom_list.push(m);
523 }
524 // Now, the rest of edges.
525 uint cnt = n->req();
526 for (uint i = 1; i < cnt; i++) {
527 m = n->find_exact_control(n->in(i));
528 if (m == nullptr || m->is_top()) {
529 continue;
530 }
531 dom_list.push(m);
532 }
533 }
534 }
535 return only_dominating_controls ? DomResult::Dominate : DomResult::NotDominate;
536 }
537 }
538
539 //---------------------detect_ptr_independence---------------------------------
540 // Used by MemNode::find_previous_store to prove that two base
541 // pointers are never equal.
542 // The pointers are accompanied by their associated allocations,
543 // if any, which have been previously discovered by the caller.
544 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
545 Node* p2, AllocateNode* a2,
546 PhaseTransform* phase) {
547 // Attempt to prove that these two pointers cannot be aliased.
548 // They may both manifestly be allocations, and they should differ.
549 // Or, if they are not both allocations, they can be distinct constants.
550 // Otherwise, one is an allocation and the other a pre-existing value.
551 if (a1 == nullptr && a2 == nullptr) { // neither an allocation
552 return (p1 != p2) && p1->is_Con() && p2->is_Con();
553 } else if (a1 != nullptr && a2 != nullptr) { // both allocations
554 return (a1 != a2);
555 } else if (a1 != nullptr) { // one allocation a1
556 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
557 return all_controls_dominate(p2, a1);
558 } else { //(a2 != null) // one allocation a2
559 return all_controls_dominate(p1, a2);
560 }
561 return false;
562 }
563
564
565 // Find an arraycopy ac that produces the memory state represented by parameter mem.
566 // Return ac if
567 // (a) can_see_stored_value=true and ac must have set the value for this load or if
568 // (b) can_see_stored_value=false and ac could have set the value for this load or if
569 // (c) can_see_stored_value=false and ac cannot have set the value for this load.
570 // In case (c) change the parameter mem to the memory input of ac to skip it
571 // when searching stored value.
572 // Otherwise return null.
573 Node* LoadNode::find_previous_arraycopy(PhaseValues* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
574 ArrayCopyNode* ac = find_array_copy_clone(ld_alloc, mem);
575 if (ac != nullptr) {
576 Node* ld_addp = in(MemNode::Address);
577 Node* src = ac->in(ArrayCopyNode::Src);
578 const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
579
580 // This is a load from a cloned array. The corresponding arraycopy ac must
581 // have set the value for the load and we can return ac but only if the load
582 // is known to be within bounds. This is checked below.
583 if (ary_t != nullptr && ld_addp->is_AddP()) {
584 Node* ld_offs = ld_addp->in(AddPNode::Offset);
585 BasicType ary_elem = ary_t->elem()->array_element_basic_type();
586 jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
587 jlong elemsize = type2aelembytes(ary_elem);
588
589 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
590 const TypeInt* sizetype = ary_t->size();
591
592 if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
593 // The load is known to be within bounds. It receives its value from ac.
594 return ac;
595 }
596 // The load is known to be out-of-bounds.
597 }
598 // The load could be out-of-bounds. It must not be hoisted but must remain
599 // dependent on the runtime range check. This is achieved by returning null.
600 } else if (mem->is_Proj() && mem->in(0) != nullptr && mem->in(0)->is_ArrayCopy()) {
601 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
602
603 if (ac->is_arraycopy_validated() ||
604 ac->is_copyof_validated() ||
605 ac->is_copyofrange_validated()) {
606 Node* ld_addp = in(MemNode::Address);
607 if (ld_addp->is_AddP()) {
608 Node* ld_base = ld_addp->in(AddPNode::Address);
609 Node* ld_offs = ld_addp->in(AddPNode::Offset);
610
611 Node* dest = ac->in(ArrayCopyNode::Dest);
612
613 if (dest == ld_base) {
614 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
615 assert(!ld_offs_t->empty(), "dead reference should be checked already");
616 // Take into account vector or unsafe access size
617 jlong ld_size_in_bytes = (jlong)memory_size();
618 jlong offset_hi = ld_offs_t->_hi + ld_size_in_bytes - 1;
619 offset_hi = MIN2(offset_hi, (jlong)(TypeX::MAX->_hi)); // Take care for overflow in 32-bit VM
620 if (ac->modifies(ld_offs_t->_lo, (intptr_t)offset_hi, phase, can_see_stored_value)) {
621 return ac;
622 }
623 if (!can_see_stored_value) {
624 mem = ac->in(TypeFunc::Memory);
625 return ac;
626 }
627 }
628 }
629 }
630 }
631 return nullptr;
632 }
633
634 ArrayCopyNode* MemNode::find_array_copy_clone(Node* ld_alloc, Node* mem) const {
635 if (mem->is_Proj() && mem->in(0) != nullptr && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
636 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
637 if (ld_alloc != nullptr) {
638 // Check if there is an array copy for a clone
639 Node* mb = mem->in(0);
640 ArrayCopyNode* ac = nullptr;
641 if (mb->in(0) != nullptr && mb->in(0)->is_Proj() &&
642 mb->in(0)->in(0) != nullptr && mb->in(0)->in(0)->is_ArrayCopy()) {
643 ac = mb->in(0)->in(0)->as_ArrayCopy();
644 } else {
645 // Step over GC barrier when ReduceInitialCardMarks is disabled
646 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
647 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
648
649 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
650 ac = control_proj_ac->in(0)->as_ArrayCopy();
651 }
652 }
653
654 if (ac != nullptr && ac->is_clonebasic()) {
655 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest));
656 if (alloc != nullptr && alloc == ld_alloc) {
657 return ac;
658 }
659 }
660 }
661 }
662 return nullptr;
663 }
664
665 // The logic for reordering loads and stores uses four steps:
666 // (a) Walk carefully past stores and initializations which we
667 // can prove are independent of this load.
668 // (b) Observe that the next memory state makes an exact match
669 // with self (load or store), and locate the relevant store.
670 // (c) Ensure that, if we were to wire self directly to the store,
671 // the optimizer would fold it up somehow.
672 // (d) Do the rewiring, and return, depending on some other part of
673 // the optimizer to fold up the load.
674 // This routine handles steps (a) and (b). Steps (c) and (d) are
675 // specific to loads and stores, so they are handled by the callers.
676 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
677 //
678 Node* MemNode::find_previous_store(PhaseValues* phase) {
679 Node* ctrl = in(MemNode::Control);
680 Node* adr = in(MemNode::Address);
681 intptr_t offset = 0;
682 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
683 AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
684
685 if (offset == Type::OffsetBot)
686 return nullptr; // cannot unalias unless there are precise offsets
687
688 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
689 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
690
691 intptr_t size_in_bytes = memory_size();
692
693 Node* mem = in(MemNode::Memory); // start searching here...
694
695 int cnt = 50; // Cycle limiter
696 for (;;) { // While we can dance past unrelated stores...
697 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
698
699 Node* prev = mem;
700 if (mem->is_Store()) {
701 Node* st_adr = mem->in(MemNode::Address);
702 intptr_t st_offset = 0;
703 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
704 if (st_base == nullptr)
705 break; // inscrutable pointer
706
707 // For raw accesses it's not enough to prove that constant offsets don't intersect.
708 // We need the bases to be the equal in order for the offset check to make sense.
709 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
710 break;
711 }
712
713 if (st_offset != offset && st_offset != Type::OffsetBot) {
714 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
715 assert(mem->as_Store()->memory_size() <= MAX_STORE, "");
716 if (st_offset >= offset + size_in_bytes ||
717 st_offset <= offset - MAX_STORE ||
718 st_offset <= offset - mem->as_Store()->memory_size()) {
719 // Success: The offsets are provably independent.
720 // (You may ask, why not just test st_offset != offset and be done?
721 // The answer is that stores of different sizes can co-exist
722 // in the same sequence of RawMem effects. We sometimes initialize
723 // a whole 'tile' of array elements with a single jint or jlong.)
724 mem = mem->in(MemNode::Memory);
725 continue; // (a) advance through independent store memory
726 }
727 }
728 if (st_base != base &&
729 detect_ptr_independence(base, alloc,
730 st_base,
731 AllocateNode::Ideal_allocation(st_base),
732 phase)) {
733 // Success: The bases are provably independent.
734 mem = mem->in(MemNode::Memory);
735 continue; // (a) advance through independent store memory
736 }
737
738 // (b) At this point, if the bases or offsets do not agree, we lose,
739 // since we have not managed to prove 'this' and 'mem' independent.
740 if (st_base == base && st_offset == offset) {
741 return mem; // let caller handle steps (c), (d)
742 }
743
744 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
745 InitializeNode* st_init = mem->in(0)->as_Initialize();
746 AllocateNode* st_alloc = st_init->allocation();
747 if (st_alloc == nullptr) {
748 break; // something degenerated
749 }
750 bool known_identical = false;
751 bool known_independent = false;
752 if (alloc == st_alloc) {
753 known_identical = true;
754 } else if (alloc != nullptr) {
755 known_independent = true;
756 } else if (all_controls_dominate(this, st_alloc)) {
757 known_independent = true;
758 }
759
760 if (known_independent) {
761 // The bases are provably independent: Either they are
762 // manifestly distinct allocations, or else the control
763 // of this load dominates the store's allocation.
764 int alias_idx = phase->C->get_alias_index(adr_type());
765 if (alias_idx == Compile::AliasIdxRaw) {
766 mem = st_alloc->in(TypeFunc::Memory);
767 } else {
768 mem = st_init->memory(alias_idx);
769 }
770 continue; // (a) advance through independent store memory
771 }
772
773 // (b) at this point, if we are not looking at a store initializing
774 // the same allocation we are loading from, we lose.
775 if (known_identical) {
776 // From caller, can_see_stored_value will consult find_captured_store.
777 return mem; // let caller handle steps (c), (d)
778 }
779
780 } else if (find_previous_arraycopy(phase, alloc, mem, false) != nullptr) {
781 if (prev != mem) {
782 // Found an arraycopy but it doesn't affect that load
783 continue;
784 }
785 // Found an arraycopy that may affect that load
786 return mem;
787 } else if (addr_t != nullptr && addr_t->is_known_instance_field()) {
788 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
789 if (mem->is_Proj() && mem->in(0)->is_Call()) {
790 // ArrayCopyNodes processed here as well.
791 CallNode *call = mem->in(0)->as_Call();
792 if (!call->may_modify(addr_t, phase)) {
793 mem = call->in(TypeFunc::Memory);
794 continue; // (a) advance through independent call memory
795 }
796 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
797 ArrayCopyNode* ac = nullptr;
798 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
799 break;
800 }
801 mem = mem->in(0)->in(TypeFunc::Memory);
802 continue; // (a) advance through independent MemBar memory
803 } else if (mem->is_ClearArray()) {
804 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
805 // (the call updated 'mem' value)
806 continue; // (a) advance through independent allocation memory
807 } else {
808 // Can not bypass initialization of the instance
809 // we are looking for.
810 return mem;
811 }
812 } else if (mem->is_MergeMem()) {
813 int alias_idx = phase->C->get_alias_index(adr_type());
814 mem = mem->as_MergeMem()->memory_at(alias_idx);
815 continue; // (a) advance through independent MergeMem memory
816 }
817 }
818
819 // Unless there is an explicit 'continue', we must bail out here,
820 // because 'mem' is an inscrutable memory state (e.g., a call).
821 break;
822 }
823
824 return nullptr; // bail out
825 }
826
827 //----------------------calculate_adr_type-------------------------------------
828 // Helper function. Notices when the given type of address hits top or bottom.
829 // Also, asserts a cross-check of the type against the expected address type.
830 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
831 if (t == Type::TOP) return nullptr; // does not touch memory any more?
832 #ifdef ASSERT
833 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = nullptr;
834 #endif
835 const TypePtr* tp = t->isa_ptr();
836 if (tp == nullptr) {
837 assert(cross_check == nullptr || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
838 return TypePtr::BOTTOM; // touches lots of memory
839 } else {
840 #ifdef ASSERT
841 // %%%% [phh] We don't check the alias index if cross_check is
842 // TypeRawPtr::BOTTOM. Needs to be investigated.
843 if (cross_check != nullptr &&
844 cross_check != TypePtr::BOTTOM &&
845 cross_check != TypeRawPtr::BOTTOM) {
846 // Recheck the alias index, to see if it has changed (due to a bug).
847 Compile* C = Compile::current();
848 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
849 "must stay in the original alias category");
850 // The type of the address must be contained in the adr_type,
851 // disregarding "null"-ness.
852 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
853 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
854 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
855 "real address must not escape from expected memory type");
856 }
857 #endif
858 return tp;
859 }
860 }
861
862 uint8_t MemNode::barrier_data(const Node* n) {
863 if (n->is_LoadStore()) {
864 return n->as_LoadStore()->barrier_data();
865 } else if (n->is_Mem()) {
866 return n->as_Mem()->barrier_data();
867 }
868 return 0;
869 }
870
871 //=============================================================================
872 // Should LoadNode::Ideal() attempt to remove control edges?
873 bool LoadNode::can_remove_control() const {
874 return !has_pinned_control_dependency();
875 }
876 uint LoadNode::size_of() const { return sizeof(*this); }
877 bool LoadNode::cmp(const Node &n) const {
878 LoadNode& load = (LoadNode &)n;
879 return Type::equals(_type, load._type) &&
880 _control_dependency == load._control_dependency &&
881 _mo == load._mo;
882 }
883 const Type *LoadNode::bottom_type() const { return _type; }
884 uint LoadNode::ideal_reg() const {
885 return _type->ideal_reg();
886 }
887
888 #ifndef PRODUCT
889 void LoadNode::dump_spec(outputStream *st) const {
890 MemNode::dump_spec(st);
891 if( !Verbose && !WizardMode ) {
892 // standard dump does this in Verbose and WizardMode
893 st->print(" #"); _type->dump_on(st);
894 }
895 if (!depends_only_on_test()) {
896 st->print(" (does not depend only on test, ");
897 if (control_dependency() == UnknownControl) {
898 st->print("unknown control");
899 } else if (control_dependency() == Pinned) {
900 st->print("pinned");
901 } else if (adr_type() == TypeRawPtr::BOTTOM) {
902 st->print("raw access");
903 } else {
904 st->print("unknown reason");
905 }
906 st->print(")");
907 }
908 }
909 #endif
910
911 #ifdef ASSERT
912 //----------------------------is_immutable_value-------------------------------
913 // Helper function to allow a raw load without control edge for some cases
914 bool LoadNode::is_immutable_value(Node* adr) {
915 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
916 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) {
917
918 jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1);
919 int offsets[] = {
920 in_bytes(JavaThread::osthread_offset()),
921 in_bytes(JavaThread::threadObj_offset()),
922 in_bytes(JavaThread::vthread_offset()),
923 in_bytes(JavaThread::scopedValueCache_offset()),
924 };
925
926 for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) {
927 if (offset == offsets[i]) {
928 return true;
929 }
930 }
931 }
932
933 return false;
934 }
935 #endif
936
937 //----------------------------LoadNode::make-----------------------------------
938 // Polymorphic factory method:
939 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo,
940 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
941 Compile* C = gvn.C;
942
943 // sanity check the alias category against the created node type
944 assert(!(adr_type->isa_oopptr() &&
945 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
946 "use LoadKlassNode instead");
947 assert(!(adr_type->isa_aryptr() &&
948 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
949 "use LoadRangeNode instead");
950 // Check control edge of raw loads
951 assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
952 // oop will be recorded in oop map if load crosses safepoint
953 rt->isa_oopptr() || is_immutable_value(adr),
954 "raw memory operations should have control edge");
955 LoadNode* load = nullptr;
956 switch (bt) {
957 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
958 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
959 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
960 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
961 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
962 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break;
963 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
964 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break;
965 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
966 case T_OBJECT:
967 case T_NARROWOOP:
968 #ifdef _LP64
969 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
970 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
971 } else
972 #endif
973 {
974 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
975 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
976 }
977 break;
978 default:
979 ShouldNotReachHere();
980 break;
981 }
982 assert(load != nullptr, "LoadNode should have been created");
983 if (unaligned) {
984 load->set_unaligned_access();
985 }
986 if (mismatched) {
987 load->set_mismatched_access();
988 }
989 if (unsafe) {
990 load->set_unsafe_access();
991 }
992 load->set_barrier_data(barrier_data);
993 if (load->Opcode() == Op_LoadN) {
994 Node* ld = gvn.transform(load);
995 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
996 }
997
998 return load;
999 }
1000
1001 //------------------------------hash-------------------------------------------
1002 uint LoadNode::hash() const {
1003 // unroll addition of interesting fields
1004 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
1005 }
1006
1007 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
1008 if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) {
1009 bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile();
1010 bool is_stable_ary = FoldStableValues &&
1011 (tp != nullptr) && (tp->isa_aryptr() != nullptr) &&
1012 tp->isa_aryptr()->is_stable();
1013
1014 return (eliminate_boxing && non_volatile) || is_stable_ary;
1015 }
1016
1017 return false;
1018 }
1019
1020 LoadNode* LoadNode::pin_array_access_node() const {
1021 const TypePtr* adr_type = this->adr_type();
1022 if (adr_type != nullptr && adr_type->isa_aryptr()) {
1023 return clone_pinned();
1024 }
1025 return nullptr;
1026 }
1027
1028 // Is the value loaded previously stored by an arraycopy? If so return
1029 // a load node that reads from the source array so we may be able to
1030 // optimize out the ArrayCopy node later.
1031 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
1032 Node* ld_adr = in(MemNode::Address);
1033 intptr_t ld_off = 0;
1034 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
1035 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
1036 if (ac != nullptr) {
1037 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
1038
1039 Node* mem = ac->in(TypeFunc::Memory);
1040 Node* ctl = ac->in(0);
1041 Node* src = ac->in(ArrayCopyNode::Src);
1042
1043 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
1044 return nullptr;
1045 }
1046
1047 // load depends on the tests that validate the arraycopy
1048 LoadNode* ld = clone_pinned();
1049 Node* addp = in(MemNode::Address)->clone();
1050 if (ac->as_ArrayCopy()->is_clonebasic()) {
1051 assert(ld_alloc != nullptr, "need an alloc");
1052 assert(addp->is_AddP(), "address must be addp");
1053 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1054 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1055 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1056 addp->set_req(AddPNode::Base, src);
1057 addp->set_req(AddPNode::Address, src);
1058 } else {
1059 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1060 ac->as_ArrayCopy()->is_copyof_validated() ||
1061 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1062 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1063 addp->set_req(AddPNode::Base, src);
1064 addp->set_req(AddPNode::Address, src);
1065
1066 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1067 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
1068 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
1069
1070 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
1071 uint shift = exact_log2(type2aelembytes(ary_elem));
1072
1073 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1074 #ifdef _LP64
1075 diff = phase->transform(new ConvI2LNode(diff));
1076 #endif
1077 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1078
1079 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1080 addp->set_req(AddPNode::Offset, offset);
1081 }
1082 addp = phase->transform(addp);
1083 #ifdef ASSERT
1084 const TypePtr* adr_type = phase->type(addp)->is_ptr();
1085 ld->_adr_type = adr_type;
1086 #endif
1087 ld->set_req(MemNode::Address, addp);
1088 ld->set_req(0, ctl);
1089 ld->set_req(MemNode::Memory, mem);
1090 return ld;
1091 }
1092 return nullptr;
1093 }
1094
1095
1096 //---------------------------can_see_stored_value------------------------------
1097 // This routine exists to make sure this set of tests is done the same
1098 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
1099 // will change the graph shape in a way which makes memory alive twice at the
1100 // same time (uses the Oracle model of aliasing), then some
1101 // LoadXNode::Identity will fold things back to the equivalence-class model
1102 // of aliasing.
1103 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const {
1104 Node* ld_adr = in(MemNode::Address);
1105 intptr_t ld_off = 0;
1106 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1107 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base);
1108 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1109 Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr;
1110 // This is more general than load from boxing objects.
1111 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1112 uint alias_idx = atp->index();
1113 Node* result = nullptr;
1114 Node* current = st;
1115 // Skip through chains of MemBarNodes checking the MergeMems for
1116 // new states for the slice of this load. Stop once any other
1117 // kind of node is encountered. Loads from final memory can skip
1118 // through any kind of MemBar but normal loads shouldn't skip
1119 // through MemBarAcquire since the could allow them to move out of
1120 // a synchronized region. It is not safe to step over MemBarCPUOrder,
1121 // because alias info above them may be inaccurate (e.g., due to
1122 // mixed/mismatched unsafe accesses).
1123 bool is_final_mem = !atp->is_rewritable();
1124 while (current->is_Proj()) {
1125 int opc = current->in(0)->Opcode();
1126 if ((is_final_mem && (opc == Op_MemBarAcquire ||
1127 opc == Op_MemBarAcquireLock ||
1128 opc == Op_LoadFence)) ||
1129 opc == Op_MemBarRelease ||
1130 opc == Op_StoreFence ||
1131 opc == Op_MemBarReleaseLock ||
1132 opc == Op_MemBarStoreStore ||
1133 opc == Op_StoreStoreFence) {
1134 Node* mem = current->in(0)->in(TypeFunc::Memory);
1135 if (mem->is_MergeMem()) {
1136 MergeMemNode* merge = mem->as_MergeMem();
1137 Node* new_st = merge->memory_at(alias_idx);
1138 if (new_st == merge->base_memory()) {
1139 // Keep searching
1140 current = new_st;
1141 continue;
1142 }
1143 // Save the new memory state for the slice and fall through
1144 // to exit.
1145 result = new_st;
1146 }
1147 }
1148 break;
1149 }
1150 if (result != nullptr) {
1151 st = result;
1152 }
1153 }
1154
1155 // Loop around twice in the case Load -> Initialize -> Store.
1156 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1157 for (int trip = 0; trip <= 1; trip++) {
1158
1159 if (st->is_Store()) {
1160 Node* st_adr = st->in(MemNode::Address);
1161 if (st_adr != ld_adr) {
1162 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1163 intptr_t st_off = 0;
1164 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1165 if (ld_base == nullptr) return nullptr;
1166 if (st_base == nullptr) return nullptr;
1167 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr;
1168 if (ld_off != st_off) return nullptr;
1169 if (ld_off == Type::OffsetBot) return nullptr;
1170 // Same base, same offset.
1171 // Possible improvement for arrays: check index value instead of absolute offset.
1172
1173 // At this point we have proven something like this setup:
1174 // B = << base >>
1175 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1176 // S = StoreQ(AddP( B , #Off), V)
1177 // (Actually, we haven't yet proven the Q's are the same.)
1178 // In other words, we are loading from a casted version of
1179 // the same pointer-and-offset that we stored to.
1180 // Casted version may carry a dependency and it is respected.
1181 // Thus, we are able to replace L by V.
1182 }
1183 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1184 if (store_Opcode() != st->Opcode()) {
1185 return nullptr;
1186 }
1187 // LoadVector/StoreVector needs additional check to ensure the types match.
1188 if (st->is_StoreVector()) {
1189 const TypeVect* in_vt = st->as_StoreVector()->vect_type();
1190 const TypeVect* out_vt = as_LoadVector()->vect_type();
1191 if (in_vt != out_vt) {
1192 return nullptr;
1193 }
1194 }
1195 return st->in(MemNode::ValueIn);
1196 }
1197
1198 // A load from a freshly-created object always returns zero.
1199 // (This can happen after LoadNode::Ideal resets the load's memory input
1200 // to find_captured_store, which returned InitializeNode::zero_memory.)
1201 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1202 (st->in(0) == ld_alloc) &&
1203 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1204 // return a zero value for the load's basic type
1205 // (This is one of the few places where a generic PhaseTransform
1206 // can create new nodes. Think of it as lazily manifesting
1207 // virtually pre-existing constants.)
1208 if (memory_type() != T_VOID) {
1209 if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) {
1210 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1211 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1212 // by the ArrayCopyNode.
1213 return phase->zerocon(memory_type());
1214 }
1215 } else {
1216 // TODO: materialize all-zero vector constant
1217 assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1218 }
1219 }
1220
1221 // A load from an initialization barrier can match a captured store.
1222 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1223 InitializeNode* init = st->in(0)->as_Initialize();
1224 AllocateNode* alloc = init->allocation();
1225 if ((alloc != nullptr) && (alloc == ld_alloc)) {
1226 // examine a captured store value
1227 st = init->find_captured_store(ld_off, memory_size(), phase);
1228 if (st != nullptr) {
1229 continue; // take one more trip around
1230 }
1231 }
1232 }
1233
1234 // Load boxed value from result of valueOf() call is input parameter.
1235 if (this->is_Load() && ld_adr->is_AddP() &&
1236 (tp != nullptr) && tp->is_ptr_to_boxed_value()) {
1237 intptr_t ignore = 0;
1238 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1239 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1240 base = bs->step_over_gc_barrier(base);
1241 if (base != nullptr && base->is_Proj() &&
1242 base->as_Proj()->_con == TypeFunc::Parms &&
1243 base->in(0)->is_CallStaticJava() &&
1244 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1245 return base->in(0)->in(TypeFunc::Parms);
1246 }
1247 }
1248
1249 break;
1250 }
1251
1252 return nullptr;
1253 }
1254
1255 //----------------------is_instance_field_load_with_local_phi------------------
1256 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1257 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1258 in(Address)->is_AddP() ) {
1259 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1260 // Only instances and boxed values.
1261 if( t_oop != nullptr &&
1262 (t_oop->is_ptr_to_boxed_value() ||
1263 t_oop->is_known_instance_field()) &&
1264 t_oop->offset() != Type::OffsetBot &&
1265 t_oop->offset() != Type::OffsetTop) {
1266 return true;
1267 }
1268 }
1269 return false;
1270 }
1271
1272 //------------------------------Identity---------------------------------------
1273 // Loads are identity if previous store is to same address
1274 Node* LoadNode::Identity(PhaseGVN* phase) {
1275 // If the previous store-maker is the right kind of Store, and the store is
1276 // to the same address, then we are equal to the value stored.
1277 Node* mem = in(Memory);
1278 Node* value = can_see_stored_value(mem, phase);
1279 if( value ) {
1280 // byte, short & char stores truncate naturally.
1281 // A load has to load the truncated value which requires
1282 // some sort of masking operation and that requires an
1283 // Ideal call instead of an Identity call.
1284 if (memory_size() < BytesPerInt) {
1285 // If the input to the store does not fit with the load's result type,
1286 // it must be truncated via an Ideal call.
1287 if (!phase->type(value)->higher_equal(phase->type(this)))
1288 return this;
1289 }
1290 // (This works even when value is a Con, but LoadNode::Value
1291 // usually runs first, producing the singleton type of the Con.)
1292 if (!has_pinned_control_dependency() || value->is_Con()) {
1293 return value;
1294 } else {
1295 return this;
1296 }
1297 }
1298
1299 if (has_pinned_control_dependency()) {
1300 return this;
1301 }
1302 // Search for an existing data phi which was generated before for the same
1303 // instance's field to avoid infinite generation of phis in a loop.
1304 Node *region = mem->in(0);
1305 if (is_instance_field_load_with_local_phi(region)) {
1306 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1307 int this_index = phase->C->get_alias_index(addr_t);
1308 int this_offset = addr_t->offset();
1309 int this_iid = addr_t->instance_id();
1310 if (!addr_t->is_known_instance() &&
1311 addr_t->is_ptr_to_boxed_value()) {
1312 // Use _idx of address base (could be Phi node) for boxed values.
1313 intptr_t ignore = 0;
1314 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1315 if (base == nullptr) {
1316 return this;
1317 }
1318 this_iid = base->_idx;
1319 }
1320 const Type* this_type = bottom_type();
1321 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1322 Node* phi = region->fast_out(i);
1323 if (phi->is_Phi() && phi != mem &&
1324 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1325 return phi;
1326 }
1327 }
1328 }
1329
1330 return this;
1331 }
1332
1333 // Construct an equivalent unsigned load.
1334 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1335 BasicType bt = T_ILLEGAL;
1336 const Type* rt = nullptr;
1337 switch (Opcode()) {
1338 case Op_LoadUB: return this;
1339 case Op_LoadUS: return this;
1340 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1341 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1342 default:
1343 assert(false, "no unsigned 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 // Construct an equivalent signed load.
1352 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1353 BasicType bt = T_ILLEGAL;
1354 const Type* rt = nullptr;
1355 switch (Opcode()) {
1356 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1357 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1358 case Op_LoadB: // fall through
1359 case Op_LoadS: // fall through
1360 case Op_LoadI: // fall through
1361 case Op_LoadL: return this;
1362 default:
1363 assert(false, "no signed variant: %s", Name());
1364 return nullptr;
1365 }
1366 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1367 raw_adr_type(), rt, bt, _mo, _control_dependency,
1368 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1369 }
1370
1371 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1372 BasicType bt = rt->basic_type();
1373 switch (Opcode()) {
1374 case Op_LoadI: return (bt == T_FLOAT);
1375 case Op_LoadL: return (bt == T_DOUBLE);
1376 case Op_LoadF: return (bt == T_INT);
1377 case Op_LoadD: return (bt == T_LONG);
1378
1379 default: return false;
1380 }
1381 }
1382
1383 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1384 BasicType bt = rt->basic_type();
1385 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1386 bool is_mismatched = is_mismatched_access();
1387 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1388 if (raw_type == nullptr) {
1389 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1390 }
1391 const int op = Opcode();
1392 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) ||
1393 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access());
1394 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1395 raw_adr_type(), rt, bt, _mo, _control_dependency,
1396 require_atomic_access, is_unaligned_access(), is_mismatched);
1397 }
1398
1399 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1400 BasicType bt = vt->basic_type();
1401 switch (Opcode()) {
1402 case Op_StoreI: return (bt == T_FLOAT);
1403 case Op_StoreL: return (bt == T_DOUBLE);
1404 case Op_StoreF: return (bt == T_INT);
1405 case Op_StoreD: return (bt == T_LONG);
1406
1407 default: return false;
1408 }
1409 }
1410
1411 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1412 BasicType bt = vt->basic_type();
1413 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1414 const int op = Opcode();
1415 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) ||
1416 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access());
1417 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1418 raw_adr_type(), val, bt, _mo, require_atomic_access);
1419
1420 bool is_mismatched = is_mismatched_access();
1421 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1422 if (raw_type == nullptr) {
1423 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1424 }
1425 if (is_mismatched) {
1426 st->set_mismatched_access();
1427 }
1428 return st;
1429 }
1430
1431 // We're loading from an object which has autobox behaviour.
1432 // If this object is result of a valueOf call we'll have a phi
1433 // merging a newly allocated object and a load from the cache.
1434 // We want to replace this load with the original incoming
1435 // argument to the valueOf call.
1436 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1437 assert(igvn->C->eliminate_boxing(), "sanity");
1438 intptr_t ignore = 0;
1439 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1440 if ((base == nullptr) || base->is_Phi()) {
1441 // Push the loads from the phi that comes from valueOf up
1442 // through it to allow elimination of the loads and the recovery
1443 // of the original value. It is done in split_through_phi().
1444 return nullptr;
1445 } else if (base->is_Load() ||
1446 (base->is_DecodeN() && base->in(1)->is_Load())) {
1447 // Eliminate the load of boxed value for integer types from the cache
1448 // array by deriving the value from the index into the array.
1449 // Capture the offset of the load and then reverse the computation.
1450
1451 // Get LoadN node which loads a boxing object from 'cache' array.
1452 if (base->is_DecodeN()) {
1453 base = base->in(1);
1454 }
1455 if (!base->in(Address)->is_AddP()) {
1456 return nullptr; // Complex address
1457 }
1458 AddPNode* address = base->in(Address)->as_AddP();
1459 Node* cache_base = address->in(AddPNode::Base);
1460 if ((cache_base != nullptr) && cache_base->is_DecodeN()) {
1461 // Get ConP node which is static 'cache' field.
1462 cache_base = cache_base->in(1);
1463 }
1464 if ((cache_base != nullptr) && cache_base->is_Con()) {
1465 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1466 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
1467 Node* elements[4];
1468 int shift = exact_log2(type2aelembytes(T_OBJECT));
1469 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1470 if (count > 0 && elements[0]->is_Con() &&
1471 (count == 1 ||
1472 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1473 elements[1]->in(2) == igvn->intcon(shift)))) {
1474 ciObjArray* array = base_type->const_oop()->as_obj_array();
1475 // Fetch the box object cache[0] at the base of the array and get its value
1476 ciInstance* box = array->obj_at(0)->as_instance();
1477 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1478 assert(ik->is_box_klass(), "sanity");
1479 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1480 if (ik->nof_nonstatic_fields() == 1) {
1481 // This should be true nonstatic_field_at requires calling
1482 // nof_nonstatic_fields so check it anyway
1483 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1484 BasicType bt = c.basic_type();
1485 // Only integer types have boxing cache.
1486 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1487 bt == T_BYTE || bt == T_SHORT ||
1488 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1489 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1490 if (cache_low != (int)cache_low) {
1491 return nullptr; // should not happen since cache is array indexed by value
1492 }
1493 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1494 if (offset != (int)offset) {
1495 return nullptr; // should not happen since cache is array indexed by value
1496 }
1497 // Add up all the offsets making of the address of the load
1498 Node* result = elements[0];
1499 for (int i = 1; i < count; i++) {
1500 result = igvn->transform(new AddXNode(result, elements[i]));
1501 }
1502 // Remove the constant offset from the address and then
1503 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1504 // remove the scaling of the offset to recover the original index.
1505 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1506 // Peel the shift off directly but wrap it in a dummy node
1507 // since Ideal can't return existing nodes
1508 igvn->_worklist.push(result); // remove dead node later
1509 result = new RShiftXNode(result->in(1), igvn->intcon(0));
1510 } else if (result->is_Add() && result->in(2)->is_Con() &&
1511 result->in(1)->Opcode() == Op_LShiftX &&
1512 result->in(1)->in(2) == igvn->intcon(shift)) {
1513 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1514 // but for boxing cache access we know that X<<Z will not overflow
1515 // (there is range check) so we do this optimizatrion by hand here.
1516 igvn->_worklist.push(result); // remove dead node later
1517 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1518 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1519 } else {
1520 result = new RShiftXNode(result, igvn->intcon(shift));
1521 }
1522 #ifdef _LP64
1523 if (bt != T_LONG) {
1524 result = new ConvL2INode(igvn->transform(result));
1525 }
1526 #else
1527 if (bt == T_LONG) {
1528 result = new ConvI2LNode(igvn->transform(result));
1529 }
1530 #endif
1531 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1532 // Need to preserve unboxing load type if it is unsigned.
1533 switch(this->Opcode()) {
1534 case Op_LoadUB:
1535 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1536 break;
1537 case Op_LoadUS:
1538 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1539 break;
1540 }
1541 return result;
1542 }
1543 }
1544 }
1545 }
1546 }
1547 return nullptr;
1548 }
1549
1550 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1551 Node* region = phi->in(0);
1552 if (region == nullptr) {
1553 return false; // Wait stable graph
1554 }
1555 uint cnt = phi->req();
1556 for (uint i = 1; i < cnt; i++) {
1557 Node* rc = region->in(i);
1558 if (rc == nullptr || phase->type(rc) == Type::TOP)
1559 return false; // Wait stable graph
1560 Node* in = phi->in(i);
1561 if (in == nullptr || phase->type(in) == Type::TOP)
1562 return false; // Wait stable graph
1563 }
1564 return true;
1565 }
1566
1567 //------------------------------split_through_phi------------------------------
1568 // Check whether a call to 'split_through_phi' would split this load through the
1569 // Phi *base*. This method is essentially a copy of the validations performed
1570 // by 'split_through_phi'. The first use of this method was in EA code as part
1571 // of simplification of allocation merges.
1572 // Some differences from original method (split_through_phi):
1573 // - If base->is_CastPP(): base = base->in(1)
1574 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) {
1575 Node* mem = in(Memory);
1576 Node* address = in(Address);
1577 intptr_t ignore = 0;
1578 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1579
1580 if (base->is_CastPP()) {
1581 base = base->in(1);
1582 }
1583
1584 if (req() > 3 || base == nullptr || !base->is_Phi()) {
1585 return false;
1586 }
1587
1588 if (!mem->is_Phi()) {
1589 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1590 return false;
1591 }
1592 } else if (base->in(0) != mem->in(0)) {
1593 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1594 return false;
1595 }
1596 }
1597
1598 return true;
1599 }
1600
1601 //------------------------------split_through_phi------------------------------
1602 // Split instance or boxed field load through Phi.
1603 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) {
1604 if (req() > 3) {
1605 assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs");
1606 // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account
1607 return nullptr;
1608 }
1609 Node* mem = in(Memory);
1610 Node* address = in(Address);
1611 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1612
1613 assert((t_oop != nullptr) &&
1614 (ignore_missing_instance_id ||
1615 t_oop->is_known_instance_field() ||
1616 t_oop->is_ptr_to_boxed_value()), "invalid conditions");
1617
1618 Compile* C = phase->C;
1619 intptr_t ignore = 0;
1620 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1621 bool base_is_phi = (base != nullptr) && base->is_Phi();
1622 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1623 (base != nullptr) && (base == address->in(AddPNode::Base)) &&
1624 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1625
1626 if (!((mem->is_Phi() || base_is_phi) &&
1627 (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) {
1628 return nullptr; // Neither memory or base are Phi
1629 }
1630
1631 if (mem->is_Phi()) {
1632 if (!stable_phi(mem->as_Phi(), phase)) {
1633 return nullptr; // Wait stable graph
1634 }
1635 uint cnt = mem->req();
1636 // Check for loop invariant memory.
1637 if (cnt == 3) {
1638 for (uint i = 1; i < cnt; i++) {
1639 Node* in = mem->in(i);
1640 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1641 if (m == mem) {
1642 if (i == 1) {
1643 // if the first edge was a loop, check second edge too.
1644 // If both are replaceable - we are in an infinite loop
1645 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1646 if (n == mem) {
1647 break;
1648 }
1649 }
1650 set_req(Memory, mem->in(cnt - i));
1651 return this; // made change
1652 }
1653 }
1654 }
1655 }
1656 if (base_is_phi) {
1657 if (!stable_phi(base->as_Phi(), phase)) {
1658 return nullptr; // Wait stable graph
1659 }
1660 uint cnt = base->req();
1661 // Check for loop invariant memory.
1662 if (cnt == 3) {
1663 for (uint i = 1; i < cnt; i++) {
1664 if (base->in(i) == base) {
1665 return nullptr; // Wait stable graph
1666 }
1667 }
1668 }
1669 }
1670
1671 // Split through Phi (see original code in loopopts.cpp).
1672 assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type");
1673
1674 // Do nothing here if Identity will find a value
1675 // (to avoid infinite chain of value phis generation).
1676 if (this != Identity(phase)) {
1677 return nullptr;
1678 }
1679
1680 // Select Region to split through.
1681 Node* region;
1682 DomResult dom_result = DomResult::Dominate;
1683 if (!base_is_phi) {
1684 assert(mem->is_Phi(), "sanity");
1685 region = mem->in(0);
1686 // Skip if the region dominates some control edge of the address.
1687 // We will check `dom_result` later.
1688 dom_result = MemNode::maybe_all_controls_dominate(address, region);
1689 } else if (!mem->is_Phi()) {
1690 assert(base_is_phi, "sanity");
1691 region = base->in(0);
1692 // Skip if the region dominates some control edge of the memory.
1693 // We will check `dom_result` later.
1694 dom_result = MemNode::maybe_all_controls_dominate(mem, region);
1695 } else if (base->in(0) != mem->in(0)) {
1696 assert(base_is_phi && mem->is_Phi(), "sanity");
1697 dom_result = MemNode::maybe_all_controls_dominate(mem, base->in(0));
1698 if (dom_result == DomResult::Dominate) {
1699 region = base->in(0);
1700 } else {
1701 dom_result = MemNode::maybe_all_controls_dominate(address, mem->in(0));
1702 if (dom_result == DomResult::Dominate) {
1703 region = mem->in(0);
1704 }
1705 // Otherwise we encountered a complex graph.
1706 }
1707 } else {
1708 assert(base->in(0) == mem->in(0), "sanity");
1709 region = mem->in(0);
1710 }
1711
1712 PhaseIterGVN* igvn = phase->is_IterGVN();
1713 if (dom_result != DomResult::Dominate) {
1714 if (dom_result == DomResult::EncounteredDeadCode) {
1715 // There is some dead code which eventually will be removed in IGVN.
1716 // Once this is the case, we get an unambiguous dominance result.
1717 // Push the node to the worklist again until the dead code is removed.
1718 igvn->_worklist.push(this);
1719 }
1720 return nullptr;
1721 }
1722
1723 Node* phi = nullptr;
1724 const Type* this_type = this->bottom_type();
1725 if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) {
1726 int this_index = C->get_alias_index(t_oop);
1727 int this_offset = t_oop->offset();
1728 int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx;
1729 phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset);
1730 } else if (ignore_missing_instance_id) {
1731 phi = new PhiNode(region, this_type, nullptr, mem->_idx);
1732 } else {
1733 return nullptr;
1734 }
1735
1736 for (uint i = 1; i < region->req(); i++) {
1737 Node* x;
1738 Node* the_clone = nullptr;
1739 Node* in = region->in(i);
1740 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1741 in != nullptr && in->is_OuterStripMinedLoop()) {
1742 // No node should go in the outer strip mined loop
1743 in = in->in(LoopNode::EntryControl);
1744 }
1745 if (in == nullptr || in == C->top()) {
1746 x = C->top(); // Dead path? Use a dead data op
1747 } else {
1748 x = this->clone(); // Else clone up the data op
1749 the_clone = x; // Remember for possible deletion.
1750 // Alter data node to use pre-phi inputs
1751 if (this->in(0) == region) {
1752 x->set_req(0, in);
1753 } else {
1754 x->set_req(0, nullptr);
1755 }
1756 if (mem->is_Phi() && (mem->in(0) == region)) {
1757 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1758 }
1759 if (address->is_Phi() && address->in(0) == region) {
1760 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1761 }
1762 if (base_is_phi && (base->in(0) == region)) {
1763 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1764 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1765 x->set_req(Address, adr_x);
1766 }
1767 }
1768 // Check for a 'win' on some paths
1769 const Type *t = x->Value(igvn);
1770
1771 bool singleton = t->singleton();
1772
1773 // See comments in PhaseIdealLoop::split_thru_phi().
1774 if (singleton && t == Type::TOP) {
1775 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1776 }
1777
1778 if (singleton) {
1779 x = igvn->makecon(t);
1780 } else {
1781 // We now call Identity to try to simplify the cloned node.
1782 // Note that some Identity methods call phase->type(this).
1783 // Make sure that the type array is big enough for
1784 // our new node, even though we may throw the node away.
1785 // (This tweaking with igvn only works because x is a new node.)
1786 igvn->set_type(x, t);
1787 // If x is a TypeNode, capture any more-precise type permanently into Node
1788 // otherwise it will be not updated during igvn->transform since
1789 // igvn->type(x) is set to x->Value() already.
1790 x->raise_bottom_type(t);
1791 Node* y = x->Identity(igvn);
1792 if (y != x) {
1793 x = y;
1794 } else {
1795 y = igvn->hash_find_insert(x);
1796 if (y) {
1797 x = y;
1798 } else {
1799 // Else x is a new node we are keeping
1800 // We do not need register_new_node_with_optimizer
1801 // because set_type has already been called.
1802 igvn->_worklist.push(x);
1803 }
1804 }
1805 }
1806 if (x != the_clone && the_clone != nullptr) {
1807 igvn->remove_dead_node(the_clone);
1808 }
1809 phi->set_req(i, x);
1810 }
1811 // Record Phi
1812 igvn->register_new_node_with_optimizer(phi);
1813 return phi;
1814 }
1815
1816 AllocateNode* LoadNode::is_new_object_mark_load() const {
1817 if (Opcode() == Op_LoadX) {
1818 Node* address = in(MemNode::Address);
1819 AllocateNode* alloc = AllocateNode::Ideal_allocation(address);
1820 Node* mem = in(MemNode::Memory);
1821 if (alloc != nullptr && mem->is_Proj() &&
1822 mem->in(0) != nullptr &&
1823 mem->in(0) == alloc->initialization() &&
1824 alloc->initialization()->proj_out_or_null(0) != nullptr) {
1825 return alloc;
1826 }
1827 }
1828 return nullptr;
1829 }
1830
1831
1832 //------------------------------Ideal------------------------------------------
1833 // If the load is from Field memory and the pointer is non-null, it might be possible to
1834 // zero out the control input.
1835 // If the offset is constant and the base is an object allocation,
1836 // try to hook me up to the exact initializing store.
1837 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1838 if (has_pinned_control_dependency()) {
1839 return nullptr;
1840 }
1841 Node* p = MemNode::Ideal_common(phase, can_reshape);
1842 if (p) return (p == NodeSentinel) ? nullptr : p;
1843
1844 Node* ctrl = in(MemNode::Control);
1845 Node* address = in(MemNode::Address);
1846 bool progress = false;
1847
1848 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1849 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1850
1851 // Skip up past a SafePoint control. Cannot do this for Stores because
1852 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1853 if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint &&
1854 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1855 !addr_mark &&
1856 (depends_only_on_test() || has_unknown_control_dependency())) {
1857 ctrl = ctrl->in(0);
1858 set_req(MemNode::Control,ctrl);
1859 progress = true;
1860 }
1861
1862 intptr_t ignore = 0;
1863 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1864 if (base != nullptr
1865 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1866 // Check for useless control edge in some common special cases
1867 if (in(MemNode::Control) != nullptr
1868 && can_remove_control()
1869 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1870 && all_controls_dominate(base, phase->C->start())) {
1871 // A method-invariant, non-null address (constant or 'this' argument).
1872 set_req(MemNode::Control, nullptr);
1873 progress = true;
1874 }
1875 }
1876
1877 Node* mem = in(MemNode::Memory);
1878 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1879
1880 if (can_reshape && (addr_t != nullptr)) {
1881 // try to optimize our memory input
1882 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1883 if (opt_mem != mem) {
1884 set_req_X(MemNode::Memory, opt_mem, phase);
1885 if (phase->type( opt_mem ) == Type::TOP) return nullptr;
1886 return this;
1887 }
1888 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1889 if ((t_oop != nullptr) &&
1890 (t_oop->is_known_instance_field() ||
1891 t_oop->is_ptr_to_boxed_value())) {
1892 PhaseIterGVN *igvn = phase->is_IterGVN();
1893 assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true");
1894 if (igvn->_worklist.member(opt_mem)) {
1895 // Delay this transformation until memory Phi is processed.
1896 igvn->_worklist.push(this);
1897 return nullptr;
1898 }
1899 // Split instance field load through Phi.
1900 Node* result = split_through_phi(phase);
1901 if (result != nullptr) return result;
1902
1903 if (t_oop->is_ptr_to_boxed_value()) {
1904 Node* result = eliminate_autobox(igvn);
1905 if (result != nullptr) return result;
1906 }
1907 }
1908 }
1909
1910 // Is there a dominating load that loads the same value? Leave
1911 // anything that is not a load of a field/array element (like
1912 // barriers etc.) alone
1913 if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) {
1914 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1915 Node *use = mem->fast_out(i);
1916 if (use != this &&
1917 use->Opcode() == Opcode() &&
1918 use->in(0) != nullptr &&
1919 use->in(0) != in(0) &&
1920 use->in(Address) == in(Address)) {
1921 Node* ctl = in(0);
1922 for (int i = 0; i < 10 && ctl != nullptr; i++) {
1923 ctl = IfNode::up_one_dom(ctl);
1924 if (ctl == use->in(0)) {
1925 set_req(0, use->in(0));
1926 return this;
1927 }
1928 }
1929 }
1930 }
1931 }
1932
1933 // Check for prior store with a different base or offset; make Load
1934 // independent. Skip through any number of them. Bail out if the stores
1935 // are in an endless dead cycle and report no progress. This is a key
1936 // transform for Reflection. However, if after skipping through the Stores
1937 // we can't then fold up against a prior store do NOT do the transform as
1938 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1939 // array memory alive twice: once for the hoisted Load and again after the
1940 // bypassed Store. This situation only works if EVERYBODY who does
1941 // anti-dependence work knows how to bypass. I.e. we need all
1942 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1943 // the alias index stuff. So instead, peek through Stores and IFF we can
1944 // fold up, do so.
1945 Node* prev_mem = find_previous_store(phase);
1946 if (prev_mem != nullptr) {
1947 Node* value = can_see_arraycopy_value(prev_mem, phase);
1948 if (value != nullptr) {
1949 return value;
1950 }
1951 }
1952 // Steps (a), (b): Walk past independent stores to find an exact match.
1953 if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) {
1954 // (c) See if we can fold up on the spot, but don't fold up here.
1955 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1956 // just return a prior value, which is done by Identity calls.
1957 if (can_see_stored_value(prev_mem, phase)) {
1958 // Make ready for step (d):
1959 set_req_X(MemNode::Memory, prev_mem, phase);
1960 return this;
1961 }
1962 }
1963
1964 return progress ? this : nullptr;
1965 }
1966
1967 // Helper to recognize certain Klass fields which are invariant across
1968 // some group of array types (e.g., int[] or all T[] where T < Object).
1969 const Type*
1970 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1971 ciKlass* klass) const {
1972 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1973 // The field is Klass::_modifier_flags. Return its (constant) value.
1974 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1975 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1976 return TypeInt::make(klass->modifier_flags());
1977 }
1978 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1979 // The field is Klass::_access_flags. Return its (constant) value.
1980 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1981 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1982 return TypeInt::make(klass->access_flags());
1983 }
1984 if (tkls->offset() == in_bytes(Klass::misc_flags_offset())) {
1985 // The field is Klass::_misc_flags. Return its (constant) value.
1986 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1987 assert(this->Opcode() == Op_LoadUB, "must load an unsigned byte from _misc_flags");
1988 return TypeInt::make(klass->misc_flags());
1989 }
1990 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1991 // The field is Klass::_layout_helper. Return its constant value if known.
1992 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1993 return TypeInt::make(klass->layout_helper());
1994 }
1995
1996 // No match.
1997 return nullptr;
1998 }
1999
2000 //------------------------------Value-----------------------------------------
2001 const Type* LoadNode::Value(PhaseGVN* phase) const {
2002 // Either input is TOP ==> the result is TOP
2003 Node* mem = in(MemNode::Memory);
2004 const Type *t1 = phase->type(mem);
2005 if (t1 == Type::TOP) return Type::TOP;
2006 Node* adr = in(MemNode::Address);
2007 const TypePtr* tp = phase->type(adr)->isa_ptr();
2008 if (tp == nullptr || tp->empty()) return Type::TOP;
2009 int off = tp->offset();
2010 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
2011 Compile* C = phase->C;
2012
2013 // Try to guess loaded type from pointer type
2014 if (tp->isa_aryptr()) {
2015 const TypeAryPtr* ary = tp->is_aryptr();
2016 const Type* t = ary->elem();
2017
2018 // Determine whether the reference is beyond the header or not, by comparing
2019 // the offset against the offset of the start of the array's data.
2020 // Different array types begin at slightly different offsets (12 vs. 16).
2021 // We choose T_BYTE as an example base type that is least restrictive
2022 // as to alignment, which will therefore produce the smallest
2023 // possible base offset.
2024 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2025 const bool off_beyond_header = (off >= min_base_off);
2026
2027 // Try to constant-fold a stable array element.
2028 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
2029 // Make sure the reference is not into the header and the offset is constant
2030 ciObject* aobj = ary->const_oop();
2031 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
2032 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
2033 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
2034 stable_dimension,
2035 memory_type(), is_unsigned());
2036 if (con_type != nullptr) {
2037 return con_type;
2038 }
2039 }
2040 }
2041
2042 // Don't do this for integer types. There is only potential profit if
2043 // the element type t is lower than _type; that is, for int types, if _type is
2044 // more restrictive than t. This only happens here if one is short and the other
2045 // char (both 16 bits), and in those cases we've made an intentional decision
2046 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2047 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2048 //
2049 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2050 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
2051 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2052 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
2053 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2054 // In fact, that could have been the original type of p1, and p1 could have
2055 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2056 // expression (LShiftL quux 3) independently optimized to the constant 8.
2057 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2058 && (_type->isa_vect() == nullptr)
2059 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2060 // t might actually be lower than _type, if _type is a unique
2061 // concrete subclass of abstract class t.
2062 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
2063 const Type* jt = t->join_speculative(_type);
2064 // In any case, do not allow the join, per se, to empty out the type.
2065 if (jt->empty() && !t->empty()) {
2066 // This can happen if a interface-typed array narrows to a class type.
2067 jt = _type;
2068 }
2069 #ifdef ASSERT
2070 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2071 // The pointers in the autobox arrays are always non-null
2072 Node* base = adr->in(AddPNode::Base);
2073 if ((base != nullptr) && base->is_DecodeN()) {
2074 // Get LoadN node which loads IntegerCache.cache field
2075 base = base->in(1);
2076 }
2077 if ((base != nullptr) && base->is_Con()) {
2078 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2079 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2080 // It could be narrow oop
2081 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2082 }
2083 }
2084 }
2085 #endif
2086 return jt;
2087 }
2088 }
2089 } else if (tp->base() == Type::InstPtr) {
2090 assert( off != Type::OffsetBot ||
2091 // arrays can be cast to Objects
2092 !tp->isa_instptr() ||
2093 tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
2094 // unsafe field access may not have a constant offset
2095 C->has_unsafe_access(),
2096 "Field accesses must be precise" );
2097 // For oop loads, we expect the _type to be precise.
2098
2099 // Optimize loads from constant fields.
2100 const TypeInstPtr* tinst = tp->is_instptr();
2101 ciObject* const_oop = tinst->const_oop();
2102 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2103 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
2104 if (con_type != nullptr) {
2105 return con_type;
2106 }
2107 }
2108 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2109 assert(off != Type::OffsetBot ||
2110 !tp->isa_instklassptr() ||
2111 // arrays can be cast to Objects
2112 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2113 // also allow array-loading from the primary supertype
2114 // array during subtype checks
2115 Opcode() == Op_LoadKlass,
2116 "Field accesses must be precise");
2117 // For klass/static loads, we expect the _type to be precise
2118 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2119 /* With mirrors being an indirect in the Klass*
2120 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2121 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2122 *
2123 * So check the type and klass of the node before the LoadP.
2124 */
2125 Node* adr2 = adr->in(MemNode::Address);
2126 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2127 if (tkls != nullptr && !StressReflectiveCode) {
2128 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2129 ciKlass* klass = tkls->exact_klass();
2130 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2131 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2132 return TypeInstPtr::make(klass->java_mirror());
2133 }
2134 }
2135 }
2136
2137 const TypeKlassPtr *tkls = tp->isa_klassptr();
2138 if (tkls != nullptr) {
2139 if (tkls->is_loaded() && tkls->klass_is_exact()) {
2140 ciKlass* klass = tkls->exact_klass();
2141 // We are loading a field from a Klass metaobject whose identity
2142 // is known at compile time (the type is "exact" or "precise").
2143 // Check for fields we know are maintained as constants by the VM.
2144 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2145 // The field is Klass::_super_check_offset. Return its (constant) value.
2146 // (Folds up type checking code.)
2147 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2148 return TypeInt::make(klass->super_check_offset());
2149 }
2150 // Compute index into primary_supers array
2151 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2152 // Check for overflowing; use unsigned compare to handle the negative case.
2153 if( depth < ciKlass::primary_super_limit() ) {
2154 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2155 // (Folds up type checking code.)
2156 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2157 ciKlass *ss = klass->super_of_depth(depth);
2158 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2159 }
2160 const Type* aift = load_array_final_field(tkls, klass);
2161 if (aift != nullptr) return aift;
2162 }
2163
2164 // We can still check if we are loading from the primary_supers array at a
2165 // shallow enough depth. Even though the klass is not exact, entries less
2166 // than or equal to its super depth are correct.
2167 if (tkls->is_loaded()) {
2168 ciKlass* klass = nullptr;
2169 if (tkls->isa_instklassptr()) {
2170 klass = tkls->is_instklassptr()->instance_klass();
2171 } else {
2172 int dims;
2173 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims);
2174 if (inner->isa_instklassptr()) {
2175 klass = inner->is_instklassptr()->instance_klass();
2176 klass = ciObjArrayKlass::make(klass, dims);
2177 }
2178 }
2179 if (klass != nullptr) {
2180 // Compute index into primary_supers array
2181 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2182 // Check for overflowing; use unsigned compare to handle the negative case.
2183 if (depth < ciKlass::primary_super_limit() &&
2184 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case
2185 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2186 // (Folds up type checking code.)
2187 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2188 ciKlass *ss = klass->super_of_depth(depth);
2189 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2190 }
2191 }
2192 }
2193
2194 // If the type is enough to determine that the thing is not an array,
2195 // we can give the layout_helper a positive interval type.
2196 // This will help short-circuit some reflective code.
2197 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) &&
2198 tkls->isa_instklassptr() && // not directly typed as an array
2199 !tkls->is_instklassptr()->instance_klass()->is_java_lang_Object() // not the supertype of all T[] and specifically not Serializable & Cloneable
2200 ) {
2201 // Note: When interfaces are reliable, we can narrow the interface
2202 // test to (klass != Serializable && klass != Cloneable).
2203 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2204 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2205 // The key property of this type is that it folds up tests
2206 // for array-ness, since it proves that the layout_helper is positive.
2207 // Thus, a generic value like the basic object layout helper works fine.
2208 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2209 }
2210 }
2211
2212 // If we are loading from a freshly-allocated object, produce a zero,
2213 // if the load is provably beyond the header of the object.
2214 // (Also allow a variable load from a fresh array to produce zero.)
2215 const TypeOopPtr *tinst = tp->isa_oopptr();
2216 bool is_instance = (tinst != nullptr) && tinst->is_known_instance_field();
2217 bool is_boxed_value = (tinst != nullptr) && tinst->is_ptr_to_boxed_value();
2218 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
2219 Node* value = can_see_stored_value(mem,phase);
2220 if (value != nullptr && value->is_Con()) {
2221 assert(value->bottom_type()->higher_equal(_type),"sanity");
2222 return value->bottom_type();
2223 }
2224 }
2225
2226 bool is_vect = (_type->isa_vect() != nullptr);
2227 if (is_instance && !is_vect) {
2228 // If we have an instance type and our memory input is the
2229 // programs's initial memory state, there is no matching store,
2230 // so just return a zero of the appropriate type -
2231 // except if it is vectorized - then we have no zero constant.
2232 Node *mem = in(MemNode::Memory);
2233 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2234 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2235 return Type::get_zero_type(_type->basic_type());
2236 }
2237 }
2238
2239 Node* alloc = is_new_object_mark_load();
2240 if (alloc != nullptr) {
2241 return TypeX::make(markWord::prototype().value());
2242 }
2243
2244 return _type;
2245 }
2246
2247 //------------------------------match_edge-------------------------------------
2248 // Do we Match on this edge index or not? Match only the address.
2249 uint LoadNode::match_edge(uint idx) const {
2250 return idx == MemNode::Address;
2251 }
2252
2253 //--------------------------LoadBNode::Ideal--------------------------------------
2254 //
2255 // If the previous store is to the same address as this load,
2256 // and the value stored was larger than a byte, replace this load
2257 // with the value stored truncated to a byte. If no truncation is
2258 // needed, the replacement is done in LoadNode::Identity().
2259 //
2260 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2261 Node* mem = in(MemNode::Memory);
2262 Node* value = can_see_stored_value(mem,phase);
2263 if (value != nullptr) {
2264 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2265 if (narrow != value) {
2266 return narrow;
2267 }
2268 }
2269 // Identity call will handle the case where truncation is not needed.
2270 return LoadNode::Ideal(phase, can_reshape);
2271 }
2272
2273 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2274 Node* mem = in(MemNode::Memory);
2275 Node* value = can_see_stored_value(mem,phase);
2276 if (value != nullptr && value->is_Con() &&
2277 !value->bottom_type()->higher_equal(_type)) {
2278 // If the input to the store does not fit with the load's result type,
2279 // it must be truncated. We can't delay until Ideal call since
2280 // a singleton Value is needed for split_thru_phi optimization.
2281 int con = value->get_int();
2282 return TypeInt::make((con << 24) >> 24);
2283 }
2284 return LoadNode::Value(phase);
2285 }
2286
2287 //--------------------------LoadUBNode::Ideal-------------------------------------
2288 //
2289 // If the previous store is to the same address as this load,
2290 // and the value stored was larger than a byte, replace this load
2291 // with the value stored truncated to a byte. If no truncation is
2292 // needed, the replacement is done in LoadNode::Identity().
2293 //
2294 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2295 Node* mem = in(MemNode::Memory);
2296 Node* value = can_see_stored_value(mem, phase);
2297 if (value != nullptr) {
2298 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2299 if (narrow != value) {
2300 return narrow;
2301 }
2302 }
2303 // Identity call will handle the case where truncation is not needed.
2304 return LoadNode::Ideal(phase, can_reshape);
2305 }
2306
2307 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2308 Node* mem = in(MemNode::Memory);
2309 Node* value = can_see_stored_value(mem,phase);
2310 if (value != nullptr && value->is_Con() &&
2311 !value->bottom_type()->higher_equal(_type)) {
2312 // If the input to the store does not fit with the load's result type,
2313 // it must be truncated. We can't delay until Ideal call since
2314 // a singleton Value is needed for split_thru_phi optimization.
2315 int con = value->get_int();
2316 return TypeInt::make(con & 0xFF);
2317 }
2318 return LoadNode::Value(phase);
2319 }
2320
2321 //--------------------------LoadUSNode::Ideal-------------------------------------
2322 //
2323 // If the previous store is to the same address as this load,
2324 // and the value stored was larger than a char, replace this load
2325 // with the value stored truncated to a char. If no truncation is
2326 // needed, the replacement is done in LoadNode::Identity().
2327 //
2328 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2329 Node* mem = in(MemNode::Memory);
2330 Node* value = can_see_stored_value(mem,phase);
2331 if (value != nullptr) {
2332 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2333 if (narrow != value) {
2334 return narrow;
2335 }
2336 }
2337 // Identity call will handle the case where truncation is not needed.
2338 return LoadNode::Ideal(phase, can_reshape);
2339 }
2340
2341 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2342 Node* mem = in(MemNode::Memory);
2343 Node* value = can_see_stored_value(mem,phase);
2344 if (value != nullptr && value->is_Con() &&
2345 !value->bottom_type()->higher_equal(_type)) {
2346 // If the input to the store does not fit with the load's result type,
2347 // it must be truncated. We can't delay until Ideal call since
2348 // a singleton Value is needed for split_thru_phi optimization.
2349 int con = value->get_int();
2350 return TypeInt::make(con & 0xFFFF);
2351 }
2352 return LoadNode::Value(phase);
2353 }
2354
2355 //--------------------------LoadSNode::Ideal--------------------------------------
2356 //
2357 // If the previous store is to the same address as this load,
2358 // and the value stored was larger than a short, replace this load
2359 // with the value stored truncated to a short. If no truncation is
2360 // needed, the replacement is done in LoadNode::Identity().
2361 //
2362 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2363 Node* mem = in(MemNode::Memory);
2364 Node* value = can_see_stored_value(mem,phase);
2365 if (value != nullptr) {
2366 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2367 if (narrow != value) {
2368 return narrow;
2369 }
2370 }
2371 // Identity call will handle the case where truncation is not needed.
2372 return LoadNode::Ideal(phase, can_reshape);
2373 }
2374
2375 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2376 Node* mem = in(MemNode::Memory);
2377 Node* value = can_see_stored_value(mem,phase);
2378 if (value != nullptr && value->is_Con() &&
2379 !value->bottom_type()->higher_equal(_type)) {
2380 // If the input to the store does not fit with the load's result type,
2381 // it must be truncated. We can't delay until Ideal call since
2382 // a singleton Value is needed for split_thru_phi optimization.
2383 int con = value->get_int();
2384 return TypeInt::make((con << 16) >> 16);
2385 }
2386 return LoadNode::Value(phase);
2387 }
2388
2389 //=============================================================================
2390 //----------------------------LoadKlassNode::make------------------------------
2391 // Polymorphic factory method:
2392 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2393 // sanity check the alias category against the created node type
2394 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2395 assert(adr_type != nullptr, "expecting TypeKlassPtr");
2396 #ifdef _LP64
2397 if (adr_type->is_ptr_to_narrowklass()) {
2398 assert(UseCompressedClassPointers, "no compressed klasses");
2399 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2400 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2401 }
2402 #endif
2403 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2404 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2405 }
2406
2407 //------------------------------Value------------------------------------------
2408 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2409 return klass_value_common(phase);
2410 }
2411
2412 // In most cases, LoadKlassNode does not have the control input set. If the control
2413 // input is set, it must not be removed (by LoadNode::Ideal()).
2414 bool LoadKlassNode::can_remove_control() const {
2415 return false;
2416 }
2417
2418 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2419 // Either input is TOP ==> the result is TOP
2420 const Type *t1 = phase->type( in(MemNode::Memory) );
2421 if (t1 == Type::TOP) return Type::TOP;
2422 Node *adr = in(MemNode::Address);
2423 const Type *t2 = phase->type( adr );
2424 if (t2 == Type::TOP) return Type::TOP;
2425 const TypePtr *tp = t2->is_ptr();
2426 if (TypePtr::above_centerline(tp->ptr()) ||
2427 tp->ptr() == TypePtr::Null) return Type::TOP;
2428
2429 // Return a more precise klass, if possible
2430 const TypeInstPtr *tinst = tp->isa_instptr();
2431 if (tinst != nullptr) {
2432 ciInstanceKlass* ik = tinst->instance_klass();
2433 int offset = tinst->offset();
2434 if (ik == phase->C->env()->Class_klass()
2435 && (offset == java_lang_Class::klass_offset() ||
2436 offset == java_lang_Class::array_klass_offset())) {
2437 // We are loading a special hidden field from a Class mirror object,
2438 // the field which points to the VM's Klass metaobject.
2439 ciType* t = tinst->java_mirror_type();
2440 // java_mirror_type returns non-null for compile-time Class constants.
2441 if (t != nullptr) {
2442 // constant oop => constant klass
2443 if (offset == java_lang_Class::array_klass_offset()) {
2444 if (t->is_void()) {
2445 // We cannot create a void array. Since void is a primitive type return null
2446 // klass. Users of this result need to do a null check on the returned klass.
2447 return TypePtr::NULL_PTR;
2448 }
2449 return TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);
2450 }
2451 if (!t->is_klass()) {
2452 // a primitive Class (e.g., int.class) has null for a klass field
2453 return TypePtr::NULL_PTR;
2454 }
2455 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2456 return TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);
2457 }
2458 // non-constant mirror, so we can't tell what's going on
2459 }
2460 if (!tinst->is_loaded())
2461 return _type; // Bail out if not loaded
2462 if (offset == oopDesc::klass_offset_in_bytes()) {
2463 return tinst->as_klass_type(true);
2464 }
2465 }
2466
2467 // Check for loading klass from an array
2468 const TypeAryPtr *tary = tp->isa_aryptr();
2469 if (tary != nullptr &&
2470 tary->offset() == oopDesc::klass_offset_in_bytes()) {
2471 return tary->as_klass_type(true);
2472 }
2473
2474 // Check for loading klass from an array klass
2475 const TypeKlassPtr *tkls = tp->isa_klassptr();
2476 if (tkls != nullptr && !StressReflectiveCode) {
2477 if (!tkls->is_loaded())
2478 return _type; // Bail out if not loaded
2479 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2480 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2481 // // Always returning precise element type is incorrect,
2482 // // e.g., element type could be object and array may contain strings
2483 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2484
2485 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2486 // according to the element type's subclassing.
2487 return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2488 }
2489 if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() &&
2490 tkls->offset() == in_bytes(Klass::super_offset())) {
2491 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super();
2492 // The field is Klass::_super. Return its (constant) value.
2493 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2494 return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR;
2495 }
2496 }
2497
2498 if (tkls != nullptr && !UseSecondarySupersCache
2499 && tkls->offset() == in_bytes(Klass::secondary_super_cache_offset())) {
2500 // Treat Klass::_secondary_super_cache as a constant when the cache is disabled.
2501 return TypePtr::NULL_PTR;
2502 }
2503
2504 // Bailout case
2505 return LoadNode::Value(phase);
2506 }
2507
2508 //------------------------------Identity---------------------------------------
2509 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2510 // Also feed through the klass in Allocate(...klass...)._klass.
2511 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2512 return klass_identity_common(phase);
2513 }
2514
2515 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2516 Node* x = LoadNode::Identity(phase);
2517 if (x != this) return x;
2518
2519 // Take apart the address into an oop and offset.
2520 // Return 'this' if we cannot.
2521 Node* adr = in(MemNode::Address);
2522 intptr_t offset = 0;
2523 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2524 if (base == nullptr) return this;
2525 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2526 if (toop == nullptr) return this;
2527
2528 // Step over potential GC barrier for OopHandle resolve
2529 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2530 if (bs->is_gc_barrier_node(base)) {
2531 base = bs->step_over_gc_barrier(base);
2532 }
2533
2534 // We can fetch the klass directly through an AllocateNode.
2535 // This works even if the klass is not constant (clone or newArray).
2536 if (offset == oopDesc::klass_offset_in_bytes()) {
2537 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2538 if (allocated_klass != nullptr) {
2539 return allocated_klass;
2540 }
2541 }
2542
2543 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2544 // See inline_native_Class_query for occurrences of these patterns.
2545 // Java Example: x.getClass().isAssignableFrom(y)
2546 //
2547 // This improves reflective code, often making the Class
2548 // mirror go completely dead. (Current exception: Class
2549 // mirrors may appear in debug info, but we could clean them out by
2550 // introducing a new debug info operator for Klass.java_mirror).
2551
2552 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass()
2553 && offset == java_lang_Class::klass_offset()) {
2554 if (base->is_Load()) {
2555 Node* base2 = base->in(MemNode::Address);
2556 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2557 Node* adr2 = base2->in(MemNode::Address);
2558 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2559 if (tkls != nullptr && !tkls->empty()
2560 && (tkls->isa_instklassptr() || tkls->isa_aryklassptr())
2561 && adr2->is_AddP()
2562 ) {
2563 int mirror_field = in_bytes(Klass::java_mirror_offset());
2564 if (tkls->offset() == mirror_field) {
2565 return adr2->in(AddPNode::Base);
2566 }
2567 }
2568 }
2569 }
2570 }
2571
2572 return this;
2573 }
2574
2575 LoadNode* LoadNode::clone_pinned() const {
2576 LoadNode* ld = clone()->as_Load();
2577 ld->_control_dependency = UnknownControl;
2578 return ld;
2579 }
2580
2581
2582 //------------------------------Value------------------------------------------
2583 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2584 const Type *t = klass_value_common(phase);
2585 if (t == Type::TOP)
2586 return t;
2587
2588 return t->make_narrowklass();
2589 }
2590
2591 //------------------------------Identity---------------------------------------
2592 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2593 // Also feed through the klass in Allocate(...klass...)._klass.
2594 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2595 Node *x = klass_identity_common(phase);
2596
2597 const Type *t = phase->type( x );
2598 if( t == Type::TOP ) return x;
2599 if( t->isa_narrowklass()) return x;
2600 assert (!t->isa_narrowoop(), "no narrow oop here");
2601
2602 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2603 }
2604
2605 //------------------------------Value-----------------------------------------
2606 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2607 // Either input is TOP ==> the result is TOP
2608 const Type *t1 = phase->type( in(MemNode::Memory) );
2609 if( t1 == Type::TOP ) return Type::TOP;
2610 Node *adr = in(MemNode::Address);
2611 const Type *t2 = phase->type( adr );
2612 if( t2 == Type::TOP ) return Type::TOP;
2613 const TypePtr *tp = t2->is_ptr();
2614 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2615 const TypeAryPtr *tap = tp->isa_aryptr();
2616 if( !tap ) return _type;
2617 return tap->size();
2618 }
2619
2620 //-------------------------------Ideal---------------------------------------
2621 // Feed through the length in AllocateArray(...length...)._length.
2622 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2623 Node* p = MemNode::Ideal_common(phase, can_reshape);
2624 if (p) return (p == NodeSentinel) ? nullptr : p;
2625
2626 // Take apart the address into an oop and offset.
2627 // Return 'this' if we cannot.
2628 Node* adr = in(MemNode::Address);
2629 intptr_t offset = 0;
2630 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2631 if (base == nullptr) return nullptr;
2632 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2633 if (tary == nullptr) return nullptr;
2634
2635 // We can fetch the length directly through an AllocateArrayNode.
2636 // This works even if the length is not constant (clone or newArray).
2637 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2638 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2639 if (alloc != nullptr) {
2640 Node* allocated_length = alloc->Ideal_length();
2641 Node* len = alloc->make_ideal_length(tary, phase);
2642 if (allocated_length != len) {
2643 // New CastII improves on this.
2644 return len;
2645 }
2646 }
2647 }
2648
2649 return nullptr;
2650 }
2651
2652 //------------------------------Identity---------------------------------------
2653 // Feed through the length in AllocateArray(...length...)._length.
2654 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2655 Node* x = LoadINode::Identity(phase);
2656 if (x != this) return x;
2657
2658 // Take apart the address into an oop and offset.
2659 // Return 'this' if we cannot.
2660 Node* adr = in(MemNode::Address);
2661 intptr_t offset = 0;
2662 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2663 if (base == nullptr) return this;
2664 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2665 if (tary == nullptr) return this;
2666
2667 // We can fetch the length directly through an AllocateArrayNode.
2668 // This works even if the length is not constant (clone or newArray).
2669 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2670 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2671 if (alloc != nullptr) {
2672 Node* allocated_length = alloc->Ideal_length();
2673 // Do not allow make_ideal_length to allocate a CastII node.
2674 Node* len = alloc->make_ideal_length(tary, phase, false);
2675 if (allocated_length == len) {
2676 // Return allocated_length only if it would not be improved by a CastII.
2677 return allocated_length;
2678 }
2679 }
2680 }
2681
2682 return this;
2683
2684 }
2685
2686 //=============================================================================
2687 //---------------------------StoreNode::make-----------------------------------
2688 // Polymorphic factory method:
2689 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) {
2690 assert((mo == unordered || mo == release), "unexpected");
2691 Compile* C = gvn.C;
2692 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2693 ctl != nullptr, "raw memory operations should have control edge");
2694
2695 switch (bt) {
2696 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2697 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2698 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2699 case T_CHAR:
2700 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2701 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2702 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2703 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2704 case T_METADATA:
2705 case T_ADDRESS:
2706 case T_OBJECT:
2707 #ifdef _LP64
2708 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2709 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2710 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2711 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2712 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2713 adr->bottom_type()->isa_rawptr())) {
2714 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2715 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2716 }
2717 #endif
2718 {
2719 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2720 }
2721 default:
2722 ShouldNotReachHere();
2723 return (StoreNode*)nullptr;
2724 }
2725 }
2726
2727 //--------------------------bottom_type----------------------------------------
2728 const Type *StoreNode::bottom_type() const {
2729 return Type::MEMORY;
2730 }
2731
2732 //------------------------------hash-------------------------------------------
2733 uint StoreNode::hash() const {
2734 // unroll addition of interesting fields
2735 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2736
2737 // Since they are not commoned, do not hash them:
2738 return NO_HASH;
2739 }
2740
2741 // Class to parse array pointers, and determine if they are adjacent. We parse the form:
2742 //
2743 // pointer = base
2744 // + constant_offset
2745 // + LShiftL( ConvI2L(int_offset + int_con), int_offset_shift)
2746 // + sum(other_offsets)
2747 //
2748 //
2749 // Note: we accumulate all constant offsets into constant_offset, even the int constant behind
2750 // the "LShiftL(ConvI2L(...))" pattern. We convert "ConvI2L(int_offset + int_con)" to
2751 // "ConvI2L(int_offset) + int_con", which is only safe if we can assume that either all
2752 // compared addresses have an overflow for "int_offset + int_con" or none.
2753 // For loads and stores on arrays, we know that if one overflows and the other not, then
2754 // the two addresses lay almost max_int indices apart, but the maximal array size is
2755 // only about half of that. Therefore, the RangeCheck on at least one of them must have
2756 // failed.
2757 //
2758 // constant_offset += LShiftL( ConvI2L(int_con), int_offset_shift)
2759 //
2760 // pointer = base
2761 // + constant_offset
2762 // + LShiftL( ConvI2L(int_offset), int_offset_shift)
2763 // + sum(other_offsets)
2764 //
2765 class ArrayPointer {
2766 private:
2767 const Node* _pointer; // The final pointer to the position in the array
2768 const Node* _base; // Base address of the array
2769 const jlong _constant_offset; // Sum of collected constant offsets
2770 const Node* _int_offset; // (optional) Offset behind LShiftL and ConvI2L
2771 const GrowableArray<Node*>* _other_offsets; // List of other AddP offsets
2772 const jint _int_offset_shift; // (optional) Shift value for int_offset
2773 const bool _is_valid; // The parsing succeeded
2774
2775 ArrayPointer(const bool is_valid,
2776 const Node* pointer,
2777 const Node* base,
2778 const jlong constant_offset,
2779 const Node* int_offset,
2780 const jint int_offset_shift,
2781 const GrowableArray<Node*>* other_offsets) :
2782 _pointer(pointer),
2783 _base(base),
2784 _constant_offset(constant_offset),
2785 _int_offset(int_offset),
2786 _other_offsets(other_offsets),
2787 _int_offset_shift(int_offset_shift),
2788 _is_valid(is_valid)
2789 {
2790 assert(_pointer != nullptr, "must always have pointer");
2791 assert(is_valid == (_base != nullptr), "have base exactly if valid");
2792 assert(is_valid == (_other_offsets != nullptr), "have other_offsets exactly if valid");
2793 }
2794
2795 static ArrayPointer make_invalid(const Node* pointer) {
2796 return ArrayPointer(false, pointer, nullptr, 0, nullptr, 0, nullptr);
2797 }
2798
2799 static bool parse_int_offset(Node* offset, Node*& int_offset, jint& int_offset_shift) {
2800 // offset = LShiftL( ConvI2L(int_offset), int_offset_shift)
2801 if (offset->Opcode() == Op_LShiftL &&
2802 offset->in(1)->Opcode() == Op_ConvI2L &&
2803 offset->in(2)->Opcode() == Op_ConI) {
2804 int_offset = offset->in(1)->in(1); // LShiftL -> ConvI2L -> int_offset
2805 int_offset_shift = offset->in(2)->get_int(); // LShiftL -> int_offset_shift
2806 return true;
2807 }
2808
2809 // offset = ConvI2L(int_offset) = LShiftL( ConvI2L(int_offset), 0)
2810 if (offset->Opcode() == Op_ConvI2L) {
2811 int_offset = offset->in(1);
2812 int_offset_shift = 0;
2813 return true;
2814 }
2815
2816 // parse failed
2817 return false;
2818 }
2819
2820 public:
2821 // Parse the structure above the pointer
2822 static ArrayPointer make(PhaseGVN* phase, const Node* pointer) {
2823 assert(phase->type(pointer)->isa_aryptr() != nullptr, "must be array pointer");
2824 if (!pointer->is_AddP()) { return ArrayPointer::make_invalid(pointer); }
2825
2826 const Node* base = pointer->in(AddPNode::Base);
2827 if (base == nullptr) { return ArrayPointer::make_invalid(pointer); }
2828
2829 const int search_depth = 5;
2830 Node* offsets[search_depth];
2831 int count = pointer->as_AddP()->unpack_offsets(offsets, search_depth);
2832
2833 // We expect at least a constant each
2834 if (count <= 0) { return ArrayPointer::make_invalid(pointer); }
2835
2836 // We extract the form:
2837 //
2838 // pointer = base
2839 // + constant_offset
2840 // + LShiftL( ConvI2L(int_offset + int_con), int_offset_shift)
2841 // + sum(other_offsets)
2842 //
2843 jlong constant_offset = 0;
2844 Node* int_offset = nullptr;
2845 jint int_offset_shift = 0;
2846 GrowableArray<Node*>* other_offsets = new GrowableArray<Node*>(count);
2847
2848 for (int i = 0; i < count; i++) {
2849 Node* offset = offsets[i];
2850 if (offset->Opcode() == Op_ConI) {
2851 // Constant int offset
2852 constant_offset += offset->get_int();
2853 } else if (offset->Opcode() == Op_ConL) {
2854 // Constant long offset
2855 constant_offset += offset->get_long();
2856 } else if(int_offset == nullptr && parse_int_offset(offset, int_offset, int_offset_shift)) {
2857 // LShiftL( ConvI2L(int_offset), int_offset_shift)
2858 int_offset = int_offset->uncast();
2859 if (int_offset->Opcode() == Op_AddI && int_offset->in(2)->Opcode() == Op_ConI) {
2860 // LShiftL( ConvI2L(int_offset + int_con), int_offset_shift)
2861 constant_offset += ((jlong)int_offset->in(2)->get_int()) << int_offset_shift;
2862 int_offset = int_offset->in(1);
2863 }
2864 } else {
2865 // All others
2866 other_offsets->append(offset);
2867 }
2868 }
2869
2870 return ArrayPointer(true, pointer, base, constant_offset, int_offset, int_offset_shift, other_offsets);
2871 }
2872
2873 bool is_adjacent_to_and_before(const ArrayPointer& other, const jlong data_size) const {
2874 if (!_is_valid || !other._is_valid) { return false; }
2875
2876 // Offset adjacent?
2877 if (this->_constant_offset + data_size != other._constant_offset) { return false; }
2878
2879 // All other components identical?
2880 if (this->_base != other._base ||
2881 this->_int_offset != other._int_offset ||
2882 this->_int_offset_shift != other._int_offset_shift ||
2883 this->_other_offsets->length() != other._other_offsets->length()) {
2884 return false;
2885 }
2886
2887 for (int i = 0; i < this->_other_offsets->length(); i++) {
2888 Node* o1 = this->_other_offsets->at(i);
2889 Node* o2 = other._other_offsets->at(i);
2890 if (o1 != o2) { return false; }
2891 }
2892
2893 return true;
2894 }
2895
2896 #ifndef PRODUCT
2897 void dump() {
2898 if (!_is_valid) {
2899 tty->print("ArrayPointer[%d %s, invalid]", _pointer->_idx, _pointer->Name());
2900 return;
2901 }
2902 tty->print("ArrayPointer[%d %s, base[%d %s] + %lld",
2903 _pointer->_idx, _pointer->Name(),
2904 _base->_idx, _base->Name(),
2905 (long long)_constant_offset);
2906 if (_int_offset != nullptr) {
2907 tty->print(" + I2L[%d %s] << %d",
2908 _int_offset->_idx, _int_offset->Name(), _int_offset_shift);
2909 }
2910 for (int i = 0; i < _other_offsets->length(); i++) {
2911 Node* n = _other_offsets->at(i);
2912 tty->print(" + [%d %s]", n->_idx, n->Name());
2913 }
2914 tty->print_cr("]");
2915 }
2916 #endif
2917 };
2918
2919 // Link together multiple stores (B/S/C/I) into a longer one.
2920 //
2921 // Example: _store = StoreB[i+3]
2922 //
2923 // RangeCheck[i+0] RangeCheck[i+0]
2924 // StoreB[i+0]
2925 // RangeCheck[i+3] RangeCheck[i+3]
2926 // StoreB[i+1] --> pass: fail:
2927 // StoreB[i+2] StoreI[i+0] StoreB[i+0]
2928 // StoreB[i+3]
2929 //
2930 // The 4 StoreB are merged into a single StoreI node. We have to be careful with RangeCheck[i+1]: before
2931 // the optimization, if this RangeCheck[i+1] fails, then we execute only StoreB[i+0], and then trap. After
2932 // the optimization, the new StoreI[i+0] is on the passing path of RangeCheck[i+3], and StoreB[i+0] on the
2933 // failing path.
2934 //
2935 // Note: For normal array stores, every store at first has a RangeCheck. But they can be removed with:
2936 // - RCE (RangeCheck Elimination): the RangeChecks in the loop are hoisted out and before the loop,
2937 // and possibly no RangeChecks remain between the stores.
2938 // - RangeCheck smearing: the earlier RangeChecks are adjusted such that they cover later RangeChecks,
2939 // and those later RangeChecks can be removed. Example:
2940 //
2941 // RangeCheck[i+0] RangeCheck[i+0] <- before first store
2942 // StoreB[i+0] StoreB[i+0] <- first store
2943 // RangeCheck[i+1] --> smeared --> RangeCheck[i+3] <- only RC between first and last store
2944 // StoreB[i+1] StoreB[i+1] <- second store
2945 // RangeCheck[i+2] --> removed
2946 // StoreB[i+2] StoreB[i+2]
2947 // RangeCheck[i+3] --> removed
2948 // StoreB[i+3] StoreB[i+3] <- last store
2949 //
2950 // Thus, it is a common pattern that between the first and last store in a chain
2951 // of adjacent stores there remains exactly one RangeCheck, located between the
2952 // first and the second store (e.g. RangeCheck[i+3]).
2953 //
2954 class MergePrimitiveArrayStores : public StackObj {
2955 private:
2956 PhaseGVN* _phase;
2957 StoreNode* _store;
2958
2959 public:
2960 MergePrimitiveArrayStores(PhaseGVN* phase, StoreNode* store) : _phase(phase), _store(store) {}
2961
2962 StoreNode* run();
2963
2964 private:
2965 bool is_compatible_store(const StoreNode* other_store) const;
2966 bool is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const;
2967 bool is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const;
2968 static bool is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out);
2969 enum CFGStatus { SuccessNoRangeCheck, SuccessWithRangeCheck, Failure };
2970 static CFGStatus cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store);
2971
2972 class Status {
2973 private:
2974 StoreNode* _found_store;
2975 bool _found_range_check;
2976
2977 Status(StoreNode* found_store, bool found_range_check)
2978 : _found_store(found_store), _found_range_check(found_range_check) {}
2979
2980 public:
2981 StoreNode* found_store() const { return _found_store; }
2982 bool found_range_check() const { return _found_range_check; }
2983 static Status make_failure() { return Status(nullptr, false); }
2984
2985 static Status make(StoreNode* found_store, const CFGStatus cfg_status) {
2986 if (cfg_status == CFGStatus::Failure) {
2987 return Status::make_failure();
2988 }
2989 return Status(found_store, cfg_status == CFGStatus::SuccessWithRangeCheck);
2990 }
2991 };
2992
2993 Status find_adjacent_use_store(const StoreNode* def_store) const;
2994 Status find_adjacent_def_store(const StoreNode* use_store) const;
2995 Status find_use_store(const StoreNode* def_store) const;
2996 Status find_def_store(const StoreNode* use_store) const;
2997 Status find_use_store_unidirectional(const StoreNode* def_store) const;
2998 Status find_def_store_unidirectional(const StoreNode* use_store) const;
2999
3000 void collect_merge_list(Node_List& merge_list) const;
3001 Node* make_merged_input_value(const Node_List& merge_list);
3002 StoreNode* make_merged_store(const Node_List& merge_list, Node* merged_input_value);
3003
3004 DEBUG_ONLY( void trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const; )
3005 };
3006
3007 StoreNode* MergePrimitiveArrayStores::run() {
3008 // Check for B/S/C/I
3009 int opc = _store->Opcode();
3010 if (opc != Op_StoreB && opc != Op_StoreC && opc != Op_StoreI) {
3011 return nullptr;
3012 }
3013
3014 // Only merge stores on arrays, and the stores must have the same size as the elements.
3015 const TypePtr* ptr_t = _store->adr_type();
3016 if (ptr_t == nullptr) {
3017 return nullptr;
3018 }
3019 const TypeAryPtr* aryptr_t = ptr_t->isa_aryptr();
3020 if (aryptr_t == nullptr) {
3021 return nullptr;
3022 }
3023 BasicType bt = aryptr_t->elem()->array_element_basic_type();
3024 if (!is_java_primitive(bt) ||
3025 type2aelembytes(bt) != _store->memory_size()) {
3026 return nullptr;
3027 }
3028 if (_store->is_unsafe_access()) {
3029 return nullptr;
3030 }
3031
3032 // The _store must be the "last" store in a chain. If we find a use we could merge with
3033 // then that use or a store further down is the "last" store.
3034 Status status_use = find_adjacent_use_store(_store);
3035 if (status_use.found_store() != nullptr) {
3036 return nullptr;
3037 }
3038
3039 // Check if we can merge with at least one def, so that we have at least 2 stores to merge.
3040 Status status_def = find_adjacent_def_store(_store);
3041 if (status_def.found_store() == nullptr) {
3042 return nullptr;
3043 }
3044
3045 ResourceMark rm;
3046 Node_List merge_list;
3047 collect_merge_list(merge_list);
3048
3049 Node* merged_input_value = make_merged_input_value(merge_list);
3050 if (merged_input_value == nullptr) { return nullptr; }
3051
3052 StoreNode* merged_store = make_merged_store(merge_list, merged_input_value);
3053
3054 DEBUG_ONLY( if(TraceMergeStores) { trace(merge_list, merged_input_value, merged_store); } )
3055
3056 return merged_store;
3057 }
3058
3059 // Check compatibility between _store and other_store.
3060 bool MergePrimitiveArrayStores::is_compatible_store(const StoreNode* other_store) const {
3061 int opc = _store->Opcode();
3062 assert(opc == Op_StoreB || opc == Op_StoreC || opc == Op_StoreI, "precondition");
3063 assert(_store->adr_type()->isa_aryptr() != nullptr, "must be array store");
3064 assert(!_store->is_unsafe_access(), "no unsafe accesses");
3065
3066 if (other_store == nullptr ||
3067 _store->Opcode() != other_store->Opcode() ||
3068 other_store->adr_type() == nullptr ||
3069 other_store->adr_type()->isa_aryptr() == nullptr ||
3070 other_store->is_unsafe_access()) {
3071 return false;
3072 }
3073
3074 // Check that the size of the stores, and the array elements are all the same.
3075 const TypeAryPtr* aryptr_t1 = _store->adr_type()->is_aryptr();
3076 const TypeAryPtr* aryptr_t2 = other_store->adr_type()->is_aryptr();
3077 BasicType aryptr_bt1 = aryptr_t1->elem()->array_element_basic_type();
3078 BasicType aryptr_bt2 = aryptr_t2->elem()->array_element_basic_type();
3079 if (!is_java_primitive(aryptr_bt1) || !is_java_primitive(aryptr_bt2)) {
3080 return false;
3081 }
3082 int size1 = type2aelembytes(aryptr_bt1);
3083 int size2 = type2aelembytes(aryptr_bt2);
3084 if (size1 != size2 ||
3085 size1 != _store->memory_size() ||
3086 _store->memory_size() != other_store->memory_size()) {
3087 return false;
3088 }
3089 return true;
3090 }
3091
3092 bool MergePrimitiveArrayStores::is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const {
3093 if (!is_adjacent_input_pair(def_store->in(MemNode::ValueIn),
3094 use_store->in(MemNode::ValueIn),
3095 def_store->memory_size())) {
3096 return false;
3097 }
3098
3099 ResourceMark rm;
3100 ArrayPointer array_pointer_use = ArrayPointer::make(_phase, use_store->in(MemNode::Address));
3101 ArrayPointer array_pointer_def = ArrayPointer::make(_phase, def_store->in(MemNode::Address));
3102 if (!array_pointer_def.is_adjacent_to_and_before(array_pointer_use, use_store->memory_size())) {
3103 return false;
3104 }
3105
3106 return true;
3107 }
3108
3109 bool MergePrimitiveArrayStores::is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const {
3110 // Pattern: [n1 = ConI, n2 = ConI]
3111 if (n1->Opcode() == Op_ConI) {
3112 return n2->Opcode() == Op_ConI;
3113 }
3114
3115 // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3116 #ifndef VM_LITTLE_ENDIAN
3117 // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3118 // Swapping n1 with n2 gives same pattern as on little endian platforms.
3119 swap(n1, n2);
3120 #endif // !VM_LITTLE_ENDIAN
3121 Node const* base_n2;
3122 jint shift_n2;
3123 if (!is_con_RShift(n2, base_n2, shift_n2)) {
3124 return false;
3125 }
3126 if (n1->Opcode() == Op_ConvL2I) {
3127 // look through
3128 n1 = n1->in(1);
3129 }
3130 Node const* base_n1;
3131 jint shift_n1;
3132 if (n1 == base_n2) {
3133 // n1 = base = base >> 0
3134 base_n1 = n1;
3135 shift_n1 = 0;
3136 } else if (!is_con_RShift(n1, base_n1, shift_n1)) {
3137 return false;
3138 }
3139 int bits_per_store = memory_size * 8;
3140 if (base_n1 != base_n2 ||
3141 shift_n1 + bits_per_store != shift_n2 ||
3142 shift_n1 % bits_per_store != 0) {
3143 return false;
3144 }
3145
3146 // both load from same value with correct shift
3147 return true;
3148 }
3149
3150 // Detect pattern: n = base_out >> shift_out
3151 bool MergePrimitiveArrayStores::is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out) {
3152 assert(n != nullptr, "precondition");
3153
3154 int opc = n->Opcode();
3155 if (opc == Op_ConvL2I) {
3156 n = n->in(1);
3157 opc = n->Opcode();
3158 }
3159
3160 if ((opc == Op_RShiftI ||
3161 opc == Op_RShiftL ||
3162 opc == Op_URShiftI ||
3163 opc == Op_URShiftL) &&
3164 n->in(2)->is_ConI()) {
3165 base_out = n->in(1);
3166 shift_out = n->in(2)->get_int();
3167 // The shift must be positive:
3168 return shift_out >= 0;
3169 }
3170 return false;
3171 }
3172
3173 // Check if there is nothing between the two stores, except optionally a RangeCheck leading to an uncommon trap.
3174 MergePrimitiveArrayStores::CFGStatus MergePrimitiveArrayStores::cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store) {
3175 assert(use_store->in(MemNode::Memory) == def_store, "use-def relationship");
3176
3177 Node* ctrl_use = use_store->in(MemNode::Control);
3178 Node* ctrl_def = def_store->in(MemNode::Control);
3179 if (ctrl_use == nullptr || ctrl_def == nullptr) {
3180 return CFGStatus::Failure;
3181 }
3182
3183 if (ctrl_use == ctrl_def) {
3184 // Same ctrl -> no RangeCheck in between.
3185 // Check: use_store must be the only use of def_store.
3186 if (def_store->outcnt() > 1) {
3187 return CFGStatus::Failure;
3188 }
3189 return CFGStatus::SuccessNoRangeCheck;
3190 }
3191
3192 // Different ctrl -> could have RangeCheck in between.
3193 // Check: 1. def_store only has these uses: use_store and MergeMem for uncommon trap, and
3194 // 2. ctrl separated by RangeCheck.
3195 if (def_store->outcnt() != 2) {
3196 return CFGStatus::Failure; // Cannot have exactly these uses: use_store and MergeMem for uncommon trap.
3197 }
3198 int use_store_out_idx = def_store->raw_out(0) == use_store ? 0 : 1;
3199 Node* merge_mem = def_store->raw_out(1 - use_store_out_idx)->isa_MergeMem();
3200 if (merge_mem == nullptr ||
3201 merge_mem->outcnt() != 1) {
3202 return CFGStatus::Failure; // Does not have MergeMem for uncommon trap.
3203 }
3204 if (!ctrl_use->is_IfProj() ||
3205 !ctrl_use->in(0)->is_RangeCheck() ||
3206 ctrl_use->in(0)->outcnt() != 2) {
3207 return CFGStatus::Failure; // Not RangeCheck.
3208 }
3209 ProjNode* other_proj = ctrl_use->as_IfProj()->other_if_proj();
3210 Node* trap = other_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check);
3211 if (trap != merge_mem->unique_out() ||
3212 ctrl_use->in(0)->in(0) != ctrl_def) {
3213 return CFGStatus::Failure; // Not RangeCheck with merge_mem leading to uncommon trap.
3214 }
3215
3216 return CFGStatus::SuccessWithRangeCheck;
3217 }
3218
3219 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_adjacent_use_store(const StoreNode* def_store) const {
3220 Status status_use = find_use_store(def_store);
3221 StoreNode* use_store = status_use.found_store();
3222 if (use_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3223 return Status::make_failure();
3224 }
3225 return status_use;
3226 }
3227
3228 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_adjacent_def_store(const StoreNode* use_store) const {
3229 Status status_def = find_def_store(use_store);
3230 StoreNode* def_store = status_def.found_store();
3231 if (def_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3232 return Status::make_failure();
3233 }
3234 return status_def;
3235 }
3236
3237 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_use_store(const StoreNode* def_store) const {
3238 Status status_use = find_use_store_unidirectional(def_store);
3239
3240 #ifdef ASSERT
3241 StoreNode* use_store = status_use.found_store();
3242 if (use_store != nullptr) {
3243 Status status_def = find_def_store_unidirectional(use_store);
3244 assert(status_def.found_store() == def_store &&
3245 status_def.found_range_check() == status_use.found_range_check(),
3246 "find_use_store and find_def_store must be symmetric");
3247 }
3248 #endif
3249
3250 return status_use;
3251 }
3252
3253 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_def_store(const StoreNode* use_store) const {
3254 Status status_def = find_def_store_unidirectional(use_store);
3255
3256 #ifdef ASSERT
3257 StoreNode* def_store = status_def.found_store();
3258 if (def_store != nullptr) {
3259 Status status_use = find_use_store_unidirectional(def_store);
3260 assert(status_use.found_store() == use_store &&
3261 status_use.found_range_check() == status_def.found_range_check(),
3262 "find_use_store and find_def_store must be symmetric");
3263 }
3264 #endif
3265
3266 return status_def;
3267 }
3268
3269 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_use_store_unidirectional(const StoreNode* def_store) const {
3270 assert(is_compatible_store(def_store), "precondition: must be compatible with _store");
3271
3272 for (DUIterator_Fast imax, i = def_store->fast_outs(imax); i < imax; i++) {
3273 StoreNode* use_store = def_store->fast_out(i)->isa_Store();
3274 if (is_compatible_store(use_store)) {
3275 return Status::make(use_store, cfg_status_for_pair(use_store, def_store));
3276 }
3277 }
3278
3279 return Status::make_failure();
3280 }
3281
3282 MergePrimitiveArrayStores::Status MergePrimitiveArrayStores::find_def_store_unidirectional(const StoreNode* use_store) const {
3283 assert(is_compatible_store(use_store), "precondition: must be compatible with _store");
3284
3285 StoreNode* def_store = use_store->in(MemNode::Memory)->isa_Store();
3286 if (!is_compatible_store(def_store)) {
3287 return Status::make_failure();
3288 }
3289
3290 return Status::make(def_store, cfg_status_for_pair(use_store, def_store));
3291 }
3292
3293 void MergePrimitiveArrayStores::collect_merge_list(Node_List& merge_list) const {
3294 // The merged store can be at most 8 bytes.
3295 const uint merge_list_max_size = 8 / _store->memory_size();
3296 assert(merge_list_max_size >= 2 &&
3297 merge_list_max_size <= 8 &&
3298 is_power_of_2(merge_list_max_size),
3299 "must be 2, 4 or 8");
3300
3301 // Traverse up the chain of adjacent def stores.
3302 StoreNode* current = _store;
3303 merge_list.push(current);
3304 while (current != nullptr && merge_list.size() < merge_list_max_size) {
3305 Status status = find_adjacent_def_store(current);
3306 current = status.found_store();
3307 if (current != nullptr) {
3308 merge_list.push(current);
3309
3310 // We can have at most one RangeCheck.
3311 if (status.found_range_check()) {
3312 break;
3313 }
3314 }
3315 }
3316
3317 // Truncate the merge_list to a power of 2.
3318 const uint pow2size = round_down_power_of_2(merge_list.size());
3319 assert(pow2size >= 2, "must be merging at least 2 stores");
3320 while (merge_list.size() > pow2size) { merge_list.pop(); }
3321 }
3322
3323 // Merge the input values of the smaller stores to a single larger input value.
3324 Node* MergePrimitiveArrayStores::make_merged_input_value(const Node_List& merge_list) {
3325 int new_memory_size = _store->memory_size() * merge_list.size();
3326 Node* first = merge_list.at(merge_list.size()-1);
3327 Node* merged_input_value = nullptr;
3328 if (_store->in(MemNode::ValueIn)->Opcode() == Op_ConI) {
3329 // Pattern: [ConI, ConI, ...] -> new constant
3330 jlong con = 0;
3331 jlong bits_per_store = _store->memory_size() * 8;
3332 jlong mask = (((jlong)1) << bits_per_store) - 1;
3333 for (uint i = 0; i < merge_list.size(); i++) {
3334 jlong con_i = merge_list.at(i)->in(MemNode::ValueIn)->get_int();
3335 #ifdef VM_LITTLE_ENDIAN
3336 con = con << bits_per_store;
3337 con = con | (mask & con_i);
3338 #else // VM_LITTLE_ENDIAN
3339 con_i = (mask & con_i) << (i * bits_per_store);
3340 con = con | con_i;
3341 #endif // VM_LITTLE_ENDIAN
3342 }
3343 merged_input_value = _phase->longcon(con);
3344 } else {
3345 // Pattern: [base >> 24, base >> 16, base >> 8, base] -> base
3346 // | |
3347 // _store first
3348 //
3349 Node* hi = _store->in(MemNode::ValueIn);
3350 Node* lo = first->in(MemNode::ValueIn);
3351 #ifndef VM_LITTLE_ENDIAN
3352 // `_store` and `first` are swapped in the diagram above
3353 swap(hi, lo);
3354 #endif // !VM_LITTLE_ENDIAN
3355 Node const* hi_base;
3356 jint hi_shift;
3357 merged_input_value = lo;
3358 bool is_true = is_con_RShift(hi, hi_base, hi_shift);
3359 assert(is_true, "must detect con RShift");
3360 if (merged_input_value != hi_base && merged_input_value->Opcode() == Op_ConvL2I) {
3361 // look through
3362 merged_input_value = merged_input_value->in(1);
3363 }
3364 if (merged_input_value != hi_base) {
3365 // merged_input_value is not the base
3366 return nullptr;
3367 }
3368 }
3369
3370 if (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size <= 4) {
3371 // Example:
3372 //
3373 // long base = ...;
3374 // a[0] = (byte)(base >> 0);
3375 // a[1] = (byte)(base >> 8);
3376 //
3377 merged_input_value = _phase->transform(new ConvL2INode(merged_input_value));
3378 }
3379
3380 assert((_phase->type(merged_input_value)->isa_int() != nullptr && new_memory_size <= 4) ||
3381 (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size == 8),
3382 "merged_input_value is either int or long, and new_memory_size is small enough");
3383
3384 return merged_input_value;
3385 }
3386
3387 // //
3388 // first_ctrl first_mem first_adr first_ctrl first_mem first_adr //
3389 // | | | | | | //
3390 // | | | | +---------------+ | //
3391 // | | | | | | | //
3392 // | | +---------+ | | +---------------+ //
3393 // | | | | | | | | //
3394 // +--------------+ | | v1 +------------------------------+ | | v1 //
3395 // | | | | | | | | | | | | //
3396 // RangeCheck first_store RangeCheck | | first_store //
3397 // | | | | | | | //
3398 // last_ctrl | +----> unc_trap last_ctrl | | +----> unc_trap //
3399 // | | ===> | | | //
3400 // +--------------+ | a2 v2 | | | //
3401 // | | | | | | | | //
3402 // | second_store | | | //
3403 // | | | | | [v1 v2 ... vn] //
3404 // ... ... | | | | //
3405 // | | | | | v //
3406 // +--------------+ | an vn +--------------+ | | merged_input_value //
3407 // | | | | | | | | //
3408 // last_store (= _store) merged_store //
3409 // //
3410 StoreNode* MergePrimitiveArrayStores::make_merged_store(const Node_List& merge_list, Node* merged_input_value) {
3411 Node* first_store = merge_list.at(merge_list.size()-1);
3412 Node* last_ctrl = _store->in(MemNode::Control); // after (optional) RangeCheck
3413 Node* first_mem = first_store->in(MemNode::Memory);
3414 Node* first_adr = first_store->in(MemNode::Address);
3415
3416 const TypePtr* new_adr_type = _store->adr_type();
3417
3418 int new_memory_size = _store->memory_size() * merge_list.size();
3419 BasicType bt = T_ILLEGAL;
3420 switch (new_memory_size) {
3421 case 2: bt = T_SHORT; break;
3422 case 4: bt = T_INT; break;
3423 case 8: bt = T_LONG; break;
3424 }
3425
3426 StoreNode* merged_store = StoreNode::make(*_phase, last_ctrl, first_mem, first_adr,
3427 new_adr_type, merged_input_value, bt, MemNode::unordered);
3428
3429 // Marking the store mismatched is sufficient to prevent reordering, since array stores
3430 // are all on the same slice. Hence, we need no barriers.
3431 merged_store->set_mismatched_access();
3432
3433 // Constants above may now also be be packed -> put candidate on worklist
3434 _phase->is_IterGVN()->_worklist.push(first_mem);
3435
3436 return merged_store;
3437 }
3438
3439 #ifdef ASSERT
3440 void MergePrimitiveArrayStores::trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const {
3441 stringStream ss;
3442 ss.print_cr("[TraceMergeStores]: Replace");
3443 for (int i = (int)merge_list.size() - 1; i >= 0; i--) {
3444 merge_list.at(i)->dump("\n", false, &ss);
3445 }
3446 ss.print_cr("[TraceMergeStores]: with");
3447 merged_input_value->dump("\n", false, &ss);
3448 merged_store->dump("\n", false, &ss);
3449 tty->print("%s", ss.as_string());
3450 }
3451 #endif
3452
3453 //------------------------------Ideal------------------------------------------
3454 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
3455 // When a store immediately follows a relevant allocation/initialization,
3456 // try to capture it into the initialization, or hoist it above.
3457 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3458 Node* p = MemNode::Ideal_common(phase, can_reshape);
3459 if (p) return (p == NodeSentinel) ? nullptr : p;
3460
3461 Node* mem = in(MemNode::Memory);
3462 Node* address = in(MemNode::Address);
3463 Node* value = in(MemNode::ValueIn);
3464 // Back-to-back stores to same address? Fold em up. Generally
3465 // unsafe if I have intervening uses.
3466 {
3467 Node* st = mem;
3468 // If Store 'st' has more than one use, we cannot fold 'st' away.
3469 // For example, 'st' might be the final state at a conditional
3470 // return. Or, 'st' might be used by some node which is live at
3471 // the same time 'st' is live, which might be unschedulable. So,
3472 // require exactly ONE user until such time as we clone 'mem' for
3473 // each of 'mem's uses (thus making the exactly-1-user-rule hold
3474 // true).
3475 while (st->is_Store() && st->outcnt() == 1) {
3476 // Looking at a dead closed cycle of memory?
3477 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
3478 assert(Opcode() == st->Opcode() ||
3479 st->Opcode() == Op_StoreVector ||
3480 Opcode() == Op_StoreVector ||
3481 st->Opcode() == Op_StoreVectorScatter ||
3482 Opcode() == Op_StoreVectorScatter ||
3483 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
3484 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
3485 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
3486 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
3487 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
3488
3489 if (st->in(MemNode::Address)->eqv_uncast(address) &&
3490 st->as_Store()->memory_size() <= this->memory_size()) {
3491 Node* use = st->raw_out(0);
3492 if (phase->is_IterGVN()) {
3493 phase->is_IterGVN()->rehash_node_delayed(use);
3494 }
3495 // It's OK to do this in the parser, since DU info is always accurate,
3496 // and the parser always refers to nodes via SafePointNode maps.
3497 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
3498 return this;
3499 }
3500 st = st->in(MemNode::Memory);
3501 }
3502 }
3503
3504
3505 // Capture an unaliased, unconditional, simple store into an initializer.
3506 // Or, if it is independent of the allocation, hoist it above the allocation.
3507 if (ReduceFieldZeroing && /*can_reshape &&*/
3508 mem->is_Proj() && mem->in(0)->is_Initialize()) {
3509 InitializeNode* init = mem->in(0)->as_Initialize();
3510 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
3511 if (offset > 0) {
3512 Node* moved = init->capture_store(this, offset, phase, can_reshape);
3513 // If the InitializeNode captured me, it made a raw copy of me,
3514 // and I need to disappear.
3515 if (moved != nullptr) {
3516 // %%% hack to ensure that Ideal returns a new node:
3517 mem = MergeMemNode::make(mem);
3518 return mem; // fold me away
3519 }
3520 }
3521 }
3522
3523 // Fold reinterpret cast into memory operation:
3524 // StoreX mem (MoveY2X v) => StoreY mem v
3525 if (value->is_Move()) {
3526 const Type* vt = value->in(1)->bottom_type();
3527 if (has_reinterpret_variant(vt)) {
3528 if (phase->C->post_loop_opts_phase()) {
3529 return convert_to_reinterpret_store(*phase, value->in(1), vt);
3530 } else {
3531 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
3532 }
3533 }
3534 }
3535
3536 if (MergeStores && UseUnalignedAccesses) {
3537 if (phase->C->post_loop_opts_phase()) {
3538 MergePrimitiveArrayStores merge(phase, this);
3539 Node* progress = merge.run();
3540 if (progress != nullptr) { return progress; }
3541 } else {
3542 phase->C->record_for_post_loop_opts_igvn(this);
3543 }
3544 }
3545
3546 return nullptr; // No further progress
3547 }
3548
3549 //------------------------------Value-----------------------------------------
3550 const Type* StoreNode::Value(PhaseGVN* phase) const {
3551 // Either input is TOP ==> the result is TOP
3552 const Type *t1 = phase->type( in(MemNode::Memory) );
3553 if( t1 == Type::TOP ) return Type::TOP;
3554 const Type *t2 = phase->type( in(MemNode::Address) );
3555 if( t2 == Type::TOP ) return Type::TOP;
3556 const Type *t3 = phase->type( in(MemNode::ValueIn) );
3557 if( t3 == Type::TOP ) return Type::TOP;
3558 return Type::MEMORY;
3559 }
3560
3561 //------------------------------Identity---------------------------------------
3562 // Remove redundant stores:
3563 // Store(m, p, Load(m, p)) changes to m.
3564 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
3565 Node* StoreNode::Identity(PhaseGVN* phase) {
3566 Node* mem = in(MemNode::Memory);
3567 Node* adr = in(MemNode::Address);
3568 Node* val = in(MemNode::ValueIn);
3569
3570 Node* result = this;
3571
3572 // Load then Store? Then the Store is useless
3573 if (val->is_Load() &&
3574 val->in(MemNode::Address)->eqv_uncast(adr) &&
3575 val->in(MemNode::Memory )->eqv_uncast(mem) &&
3576 val->as_Load()->store_Opcode() == Opcode()) {
3577 // Ensure vector type is the same
3578 if (!is_StoreVector() || (mem->is_LoadVector() && as_StoreVector()->vect_type() == mem->as_LoadVector()->vect_type())) {
3579 result = mem;
3580 }
3581 }
3582
3583 // Two stores in a row of the same value?
3584 if (result == this &&
3585 mem->is_Store() &&
3586 mem->in(MemNode::Address)->eqv_uncast(adr) &&
3587 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
3588 mem->Opcode() == Opcode()) {
3589 if (!is_StoreVector()) {
3590 result = mem;
3591 } else {
3592 const StoreVectorNode* store_vector = as_StoreVector();
3593 const StoreVectorNode* mem_vector = mem->as_StoreVector();
3594 const Node* store_indices = store_vector->indices();
3595 const Node* mem_indices = mem_vector->indices();
3596 const Node* store_mask = store_vector->mask();
3597 const Node* mem_mask = mem_vector->mask();
3598 // Ensure types, indices, and masks match
3599 if (store_vector->vect_type() == mem_vector->vect_type() &&
3600 ((store_indices == nullptr) == (mem_indices == nullptr) &&
3601 (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) &&
3602 ((store_mask == nullptr) == (mem_mask == nullptr) &&
3603 (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) {
3604 result = mem;
3605 }
3606 }
3607 }
3608
3609 // Store of zero anywhere into a freshly-allocated object?
3610 // Then the store is useless.
3611 // (It must already have been captured by the InitializeNode.)
3612 if (result == this &&
3613 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
3614 // a newly allocated object is already all-zeroes everywhere
3615 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
3616 result = mem;
3617 }
3618
3619 if (result == this) {
3620 // the store may also apply to zero-bits in an earlier object
3621 Node* prev_mem = find_previous_store(phase);
3622 // Steps (a), (b): Walk past independent stores to find an exact match.
3623 if (prev_mem != nullptr) {
3624 Node* prev_val = can_see_stored_value(prev_mem, phase);
3625 if (prev_val != nullptr && prev_val == val) {
3626 // prev_val and val might differ by a cast; it would be good
3627 // to keep the more informative of the two.
3628 result = mem;
3629 }
3630 }
3631 }
3632 }
3633
3634 PhaseIterGVN* igvn = phase->is_IterGVN();
3635 if (result != this && igvn != nullptr) {
3636 MemBarNode* trailing = trailing_membar();
3637 if (trailing != nullptr) {
3638 #ifdef ASSERT
3639 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
3640 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects");
3641 #endif
3642 trailing->remove(igvn);
3643 }
3644 }
3645
3646 return result;
3647 }
3648
3649 //------------------------------match_edge-------------------------------------
3650 // Do we Match on this edge index or not? Match only memory & value
3651 uint StoreNode::match_edge(uint idx) const {
3652 return idx == MemNode::Address || idx == MemNode::ValueIn;
3653 }
3654
3655 //------------------------------cmp--------------------------------------------
3656 // Do not common stores up together. They generally have to be split
3657 // back up anyways, so do not bother.
3658 bool StoreNode::cmp( const Node &n ) const {
3659 return (&n == this); // Always fail except on self
3660 }
3661
3662 //------------------------------Ideal_masked_input-----------------------------
3663 // Check for a useless mask before a partial-word store
3664 // (StoreB ... (AndI valIn conIa) )
3665 // If (conIa & mask == mask) this simplifies to
3666 // (StoreB ... (valIn) )
3667 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
3668 Node *val = in(MemNode::ValueIn);
3669 if( val->Opcode() == Op_AndI ) {
3670 const TypeInt *t = phase->type( val->in(2) )->isa_int();
3671 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
3672 set_req_X(MemNode::ValueIn, val->in(1), phase);
3673 return this;
3674 }
3675 }
3676 return nullptr;
3677 }
3678
3679
3680 //------------------------------Ideal_sign_extended_input----------------------
3681 // Check for useless sign-extension before a partial-word store
3682 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
3683 // If (conIL == conIR && conIR <= num_bits) this simplifies to
3684 // (StoreB ... (valIn) )
3685 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
3686 Node *val = in(MemNode::ValueIn);
3687 if( val->Opcode() == Op_RShiftI ) {
3688 const TypeInt *t = phase->type( val->in(2) )->isa_int();
3689 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
3690 Node *shl = val->in(1);
3691 if( shl->Opcode() == Op_LShiftI ) {
3692 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
3693 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
3694 set_req_X(MemNode::ValueIn, shl->in(1), phase);
3695 return this;
3696 }
3697 }
3698 }
3699 }
3700 return nullptr;
3701 }
3702
3703 //------------------------------value_never_loaded-----------------------------------
3704 // Determine whether there are any possible loads of the value stored.
3705 // For simplicity, we actually check if there are any loads from the
3706 // address stored to, not just for loads of the value stored by this node.
3707 //
3708 bool StoreNode::value_never_loaded(PhaseValues* phase) const {
3709 Node *adr = in(Address);
3710 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
3711 if (adr_oop == nullptr)
3712 return false;
3713 if (!adr_oop->is_known_instance_field())
3714 return false; // if not a distinct instance, there may be aliases of the address
3715 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
3716 Node *use = adr->fast_out(i);
3717 if (use->is_Load() || use->is_LoadStore()) {
3718 return false;
3719 }
3720 }
3721 return true;
3722 }
3723
3724 MemBarNode* StoreNode::trailing_membar() const {
3725 if (is_release()) {
3726 MemBarNode* trailing_mb = nullptr;
3727 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3728 Node* u = fast_out(i);
3729 if (u->is_MemBar()) {
3730 if (u->as_MemBar()->trailing_store()) {
3731 assert(u->Opcode() == Op_MemBarVolatile, "");
3732 assert(trailing_mb == nullptr, "only one");
3733 trailing_mb = u->as_MemBar();
3734 #ifdef ASSERT
3735 Node* leading = u->as_MemBar()->leading_membar();
3736 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3737 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
3738 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
3739 #endif
3740 } else {
3741 assert(u->as_MemBar()->standalone(), "");
3742 }
3743 }
3744 }
3745 return trailing_mb;
3746 }
3747 return nullptr;
3748 }
3749
3750
3751 //=============================================================================
3752 //------------------------------Ideal------------------------------------------
3753 // If the store is from an AND mask that leaves the low bits untouched, then
3754 // we can skip the AND operation. If the store is from a sign-extension
3755 // (a left shift, then right shift) we can skip both.
3756 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
3757 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
3758 if( progress != nullptr ) return progress;
3759
3760 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
3761 if( progress != nullptr ) return progress;
3762
3763 // Finally check the default case
3764 return StoreNode::Ideal(phase, can_reshape);
3765 }
3766
3767 //=============================================================================
3768 //------------------------------Ideal------------------------------------------
3769 // If the store is from an AND mask that leaves the low bits untouched, then
3770 // we can skip the AND operation
3771 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
3772 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
3773 if( progress != nullptr ) return progress;
3774
3775 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
3776 if( progress != nullptr ) return progress;
3777
3778 // Finally check the default case
3779 return StoreNode::Ideal(phase, can_reshape);
3780 }
3781
3782 //=============================================================================
3783 //----------------------------------SCMemProjNode------------------------------
3784 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
3785 {
3786 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
3787 return Type::TOP;
3788 }
3789 return bottom_type();
3790 }
3791
3792 //=============================================================================
3793 //----------------------------------LoadStoreNode------------------------------
3794 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
3795 : Node(required),
3796 _type(rt),
3797 _adr_type(at),
3798 _barrier_data(0)
3799 {
3800 init_req(MemNode::Control, c );
3801 init_req(MemNode::Memory , mem);
3802 init_req(MemNode::Address, adr);
3803 init_req(MemNode::ValueIn, val);
3804 init_class_id(Class_LoadStore);
3805 }
3806
3807 //------------------------------Value-----------------------------------------
3808 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
3809 // Either input is TOP ==> the result is TOP
3810 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
3811 return Type::TOP;
3812 }
3813 const Type* t = phase->type(in(MemNode::Memory));
3814 if (t == Type::TOP) {
3815 return Type::TOP;
3816 }
3817 t = phase->type(in(MemNode::Address));
3818 if (t == Type::TOP) {
3819 return Type::TOP;
3820 }
3821 t = phase->type(in(MemNode::ValueIn));
3822 if (t == Type::TOP) {
3823 return Type::TOP;
3824 }
3825 return bottom_type();
3826 }
3827
3828 uint LoadStoreNode::ideal_reg() const {
3829 return _type->ideal_reg();
3830 }
3831
3832 // This method conservatively checks if the result of a LoadStoreNode is
3833 // used, that is, if it returns true, then it is definitely the case that
3834 // the result of the node is not needed.
3835 // For example, GetAndAdd can be matched into a lock_add instead of a
3836 // lock_xadd if the result of LoadStoreNode::result_not_used() is true
3837 bool LoadStoreNode::result_not_used() const {
3838 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3839 Node *x = fast_out(i);
3840 if (x->Opcode() == Op_SCMemProj) {
3841 continue;
3842 }
3843 if (x->bottom_type() == TypeTuple::MEMBAR &&
3844 !x->is_Call() &&
3845 x->Opcode() != Op_Blackhole) {
3846 continue;
3847 }
3848 return false;
3849 }
3850 return true;
3851 }
3852
3853 MemBarNode* LoadStoreNode::trailing_membar() const {
3854 MemBarNode* trailing = nullptr;
3855 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3856 Node* u = fast_out(i);
3857 if (u->is_MemBar()) {
3858 if (u->as_MemBar()->trailing_load_store()) {
3859 assert(u->Opcode() == Op_MemBarAcquire, "");
3860 assert(trailing == nullptr, "only one");
3861 trailing = u->as_MemBar();
3862 #ifdef ASSERT
3863 Node* leading = trailing->leading_membar();
3864 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3865 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
3866 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
3867 #endif
3868 } else {
3869 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
3870 }
3871 }
3872 }
3873
3874 return trailing;
3875 }
3876
3877 uint LoadStoreNode::size_of() const { return sizeof(*this); }
3878
3879 //=============================================================================
3880 //----------------------------------LoadStoreConditionalNode--------------------
3881 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) {
3882 init_req(ExpectedIn, ex );
3883 }
3884
3885 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
3886 // Either input is TOP ==> the result is TOP
3887 const Type* t = phase->type(in(ExpectedIn));
3888 if (t == Type::TOP) {
3889 return Type::TOP;
3890 }
3891 return LoadStoreNode::Value(phase);
3892 }
3893
3894 //=============================================================================
3895 //-------------------------------adr_type--------------------------------------
3896 const TypePtr* ClearArrayNode::adr_type() const {
3897 Node *adr = in(3);
3898 if (adr == nullptr) return nullptr; // node is dead
3899 return MemNode::calculate_adr_type(adr->bottom_type());
3900 }
3901
3902 //------------------------------match_edge-------------------------------------
3903 // Do we Match on this edge index or not? Do not match memory
3904 uint ClearArrayNode::match_edge(uint idx) const {
3905 return idx > 1;
3906 }
3907
3908 //------------------------------Identity---------------------------------------
3909 // Clearing a zero length array does nothing
3910 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
3911 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
3912 }
3913
3914 //------------------------------Idealize---------------------------------------
3915 // Clearing a short array is faster with stores
3916 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3917 // Already know this is a large node, do not try to ideal it
3918 if (_is_large) return nullptr;
3919
3920 const int unit = BytesPerLong;
3921 const TypeX* t = phase->type(in(2))->isa_intptr_t();
3922 if (!t) return nullptr;
3923 if (!t->is_con()) return nullptr;
3924 intptr_t raw_count = t->get_con();
3925 intptr_t size = raw_count;
3926 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
3927 // Clearing nothing uses the Identity call.
3928 // Negative clears are possible on dead ClearArrays
3929 // (see jck test stmt114.stmt11402.val).
3930 if (size <= 0 || size % unit != 0) return nullptr;
3931 intptr_t count = size / unit;
3932 // Length too long; communicate this to matchers and assemblers.
3933 // Assemblers are responsible to produce fast hardware clears for it.
3934 if (size > InitArrayShortSize) {
3935 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
3936 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
3937 return nullptr;
3938 }
3939 if (!IdealizeClearArrayNode) return nullptr;
3940 Node *mem = in(1);
3941 if( phase->type(mem)==Type::TOP ) return nullptr;
3942 Node *adr = in(3);
3943 const Type* at = phase->type(adr);
3944 if( at==Type::TOP ) return nullptr;
3945 const TypePtr* atp = at->isa_ptr();
3946 // adjust atp to be the correct array element address type
3947 if (atp == nullptr) atp = TypePtr::BOTTOM;
3948 else atp = atp->add_offset(Type::OffsetBot);
3949 // Get base for derived pointer purposes
3950 if( adr->Opcode() != Op_AddP ) Unimplemented();
3951 Node *base = adr->in(1);
3952
3953 Node *zero = phase->makecon(TypeLong::ZERO);
3954 Node *off = phase->MakeConX(BytesPerLong);
3955 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3956 count--;
3957 while( count-- ) {
3958 mem = phase->transform(mem);
3959 adr = phase->transform(new AddPNode(base,adr,off));
3960 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3961 }
3962 return mem;
3963 }
3964
3965 //----------------------------step_through----------------------------------
3966 // Return allocation input memory edge if it is different instance
3967 // or itself if it is the one we are looking for.
3968 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
3969 Node* n = *np;
3970 assert(n->is_ClearArray(), "sanity");
3971 intptr_t offset;
3972 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
3973 // This method is called only before Allocate nodes are expanded
3974 // during macro nodes expansion. Before that ClearArray nodes are
3975 // only generated in PhaseMacroExpand::generate_arraycopy() (before
3976 // Allocate nodes are expanded) which follows allocations.
3977 assert(alloc != nullptr, "should have allocation");
3978 if (alloc->_idx == instance_id) {
3979 // Can not bypass initialization of the instance we are looking for.
3980 return false;
3981 }
3982 // Otherwise skip it.
3983 InitializeNode* init = alloc->initialization();
3984 if (init != nullptr)
3985 *np = init->in(TypeFunc::Memory);
3986 else
3987 *np = alloc->in(TypeFunc::Memory);
3988 return true;
3989 }
3990
3991 //----------------------------clear_memory-------------------------------------
3992 // Generate code to initialize object storage to zero.
3993 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3994 intptr_t start_offset,
3995 Node* end_offset,
3996 PhaseGVN* phase) {
3997 intptr_t offset = start_offset;
3998
3999 int unit = BytesPerLong;
4000 if ((offset % unit) != 0) {
4001 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
4002 adr = phase->transform(adr);
4003 const TypePtr* atp = TypeRawPtr::BOTTOM;
4004 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4005 mem = phase->transform(mem);
4006 offset += BytesPerInt;
4007 }
4008 assert((offset % unit) == 0, "");
4009
4010 // Initialize the remaining stuff, if any, with a ClearArray.
4011 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
4012 }
4013
4014 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4015 Node* start_offset,
4016 Node* end_offset,
4017 PhaseGVN* phase) {
4018 if (start_offset == end_offset) {
4019 // nothing to do
4020 return mem;
4021 }
4022
4023 int unit = BytesPerLong;
4024 Node* zbase = start_offset;
4025 Node* zend = end_offset;
4026
4027 // Scale to the unit required by the CPU:
4028 if (!Matcher::init_array_count_is_in_bytes) {
4029 Node* shift = phase->intcon(exact_log2(unit));
4030 zbase = phase->transform(new URShiftXNode(zbase, shift) );
4031 zend = phase->transform(new URShiftXNode(zend, shift) );
4032 }
4033
4034 // Bulk clear double-words
4035 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
4036 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
4037 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
4038 return phase->transform(mem);
4039 }
4040
4041 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4042 intptr_t start_offset,
4043 intptr_t end_offset,
4044 PhaseGVN* phase) {
4045 if (start_offset == end_offset) {
4046 // nothing to do
4047 return mem;
4048 }
4049
4050 assert((end_offset % BytesPerInt) == 0, "odd end offset");
4051 intptr_t done_offset = end_offset;
4052 if ((done_offset % BytesPerLong) != 0) {
4053 done_offset -= BytesPerInt;
4054 }
4055 if (done_offset > start_offset) {
4056 mem = clear_memory(ctl, mem, dest,
4057 start_offset, phase->MakeConX(done_offset), phase);
4058 }
4059 if (done_offset < end_offset) { // emit the final 32-bit store
4060 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
4061 adr = phase->transform(adr);
4062 const TypePtr* atp = TypeRawPtr::BOTTOM;
4063 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4064 mem = phase->transform(mem);
4065 done_offset += BytesPerInt;
4066 }
4067 assert(done_offset == end_offset, "");
4068 return mem;
4069 }
4070
4071 //=============================================================================
4072 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
4073 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
4074 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
4075 #ifdef ASSERT
4076 , _pair_idx(0)
4077 #endif
4078 {
4079 init_class_id(Class_MemBar);
4080 Node* top = C->top();
4081 init_req(TypeFunc::I_O,top);
4082 init_req(TypeFunc::FramePtr,top);
4083 init_req(TypeFunc::ReturnAdr,top);
4084 if (precedent != nullptr)
4085 init_req(TypeFunc::Parms, precedent);
4086 }
4087
4088 //------------------------------cmp--------------------------------------------
4089 uint MemBarNode::hash() const { return NO_HASH; }
4090 bool MemBarNode::cmp( const Node &n ) const {
4091 return (&n == this); // Always fail except on self
4092 }
4093
4094 //------------------------------make-------------------------------------------
4095 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
4096 switch (opcode) {
4097 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
4098 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
4099 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
4100 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
4101 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
4102 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn);
4103 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
4104 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
4105 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
4106 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
4107 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
4108 case Op_Initialize: return new InitializeNode(C, atp, pn);
4109 default: ShouldNotReachHere(); return nullptr;
4110 }
4111 }
4112
4113 void MemBarNode::remove(PhaseIterGVN *igvn) {
4114 if (outcnt() != 2) {
4115 assert(Opcode() == Op_Initialize, "Only seen when there are no use of init memory");
4116 assert(outcnt() == 1, "Only control then");
4117 }
4118 if (trailing_store() || trailing_load_store()) {
4119 MemBarNode* leading = leading_membar();
4120 if (leading != nullptr) {
4121 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
4122 leading->remove(igvn);
4123 }
4124 }
4125 if (proj_out_or_null(TypeFunc::Memory) != nullptr) {
4126 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
4127 }
4128 if (proj_out_or_null(TypeFunc::Control) != nullptr) {
4129 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
4130 }
4131 }
4132
4133 //------------------------------Ideal------------------------------------------
4134 // Return a node which is more "ideal" than the current node. Strip out
4135 // control copies
4136 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4137 if (remove_dead_region(phase, can_reshape)) return this;
4138 // Don't bother trying to transform a dead node
4139 if (in(0) && in(0)->is_top()) {
4140 return nullptr;
4141 }
4142
4143 bool progress = false;
4144 // Eliminate volatile MemBars for scalar replaced objects.
4145 if (can_reshape && req() == (Precedent+1)) {
4146 bool eliminate = false;
4147 int opc = Opcode();
4148 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
4149 // Volatile field loads and stores.
4150 Node* my_mem = in(MemBarNode::Precedent);
4151 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
4152 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
4153 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
4154 // replace this Precedent (decodeN) with the Load instead.
4155 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
4156 Node* load_node = my_mem->in(1);
4157 set_req(MemBarNode::Precedent, load_node);
4158 phase->is_IterGVN()->_worklist.push(my_mem);
4159 my_mem = load_node;
4160 } else {
4161 assert(my_mem->unique_out() == this, "sanity");
4162 assert(!trailing_load_store(), "load store node can't be eliminated");
4163 del_req(Precedent);
4164 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
4165 my_mem = nullptr;
4166 }
4167 progress = true;
4168 }
4169 if (my_mem != nullptr && my_mem->is_Mem()) {
4170 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
4171 // Check for scalar replaced object reference.
4172 if( t_oop != nullptr && t_oop->is_known_instance_field() &&
4173 t_oop->offset() != Type::OffsetBot &&
4174 t_oop->offset() != Type::OffsetTop) {
4175 eliminate = true;
4176 }
4177 }
4178 } else if (opc == Op_MemBarRelease || (UseStoreStoreForCtor && opc == Op_MemBarStoreStore)) {
4179 // Final field stores.
4180 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent));
4181 if ((alloc != nullptr) && alloc->is_Allocate() &&
4182 alloc->as_Allocate()->does_not_escape_thread()) {
4183 // The allocated object does not escape.
4184 eliminate = true;
4185 }
4186 }
4187 if (eliminate) {
4188 // Replace MemBar projections by its inputs.
4189 PhaseIterGVN* igvn = phase->is_IterGVN();
4190 remove(igvn);
4191 // Must return either the original node (now dead) or a new node
4192 // (Do not return a top here, since that would break the uniqueness of top.)
4193 return new ConINode(TypeInt::ZERO);
4194 }
4195 }
4196 return progress ? this : nullptr;
4197 }
4198
4199 //------------------------------Value------------------------------------------
4200 const Type* MemBarNode::Value(PhaseGVN* phase) const {
4201 if( !in(0) ) return Type::TOP;
4202 if( phase->type(in(0)) == Type::TOP )
4203 return Type::TOP;
4204 return TypeTuple::MEMBAR;
4205 }
4206
4207 //------------------------------match------------------------------------------
4208 // Construct projections for memory.
4209 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
4210 switch (proj->_con) {
4211 case TypeFunc::Control:
4212 case TypeFunc::Memory:
4213 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
4214 }
4215 ShouldNotReachHere();
4216 return nullptr;
4217 }
4218
4219 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4220 trailing->_kind = TrailingStore;
4221 leading->_kind = LeadingStore;
4222 #ifdef ASSERT
4223 trailing->_pair_idx = leading->_idx;
4224 leading->_pair_idx = leading->_idx;
4225 #endif
4226 }
4227
4228 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4229 trailing->_kind = TrailingLoadStore;
4230 leading->_kind = LeadingLoadStore;
4231 #ifdef ASSERT
4232 trailing->_pair_idx = leading->_idx;
4233 leading->_pair_idx = leading->_idx;
4234 #endif
4235 }
4236
4237 MemBarNode* MemBarNode::trailing_membar() const {
4238 ResourceMark rm;
4239 Node* trailing = (Node*)this;
4240 VectorSet seen;
4241 Node_Stack multis(0);
4242 do {
4243 Node* c = trailing;
4244 uint i = 0;
4245 do {
4246 trailing = nullptr;
4247 for (; i < c->outcnt(); i++) {
4248 Node* next = c->raw_out(i);
4249 if (next != c && next->is_CFG()) {
4250 if (c->is_MultiBranch()) {
4251 if (multis.node() == c) {
4252 multis.set_index(i+1);
4253 } else {
4254 multis.push(c, i+1);
4255 }
4256 }
4257 trailing = next;
4258 break;
4259 }
4260 }
4261 if (trailing != nullptr && !seen.test_set(trailing->_idx)) {
4262 break;
4263 }
4264 while (multis.size() > 0) {
4265 c = multis.node();
4266 i = multis.index();
4267 if (i < c->req()) {
4268 break;
4269 }
4270 multis.pop();
4271 }
4272 } while (multis.size() > 0);
4273 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
4274
4275 MemBarNode* mb = trailing->as_MemBar();
4276 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
4277 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
4278 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
4279 return mb;
4280 }
4281
4282 MemBarNode* MemBarNode::leading_membar() const {
4283 ResourceMark rm;
4284 VectorSet seen;
4285 Node_Stack regions(0);
4286 Node* leading = in(0);
4287 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
4288 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) {
4289 leading = nullptr;
4290 while (regions.size() > 0 && leading == nullptr) {
4291 Node* r = regions.node();
4292 uint i = regions.index();
4293 if (i < r->req()) {
4294 leading = r->in(i);
4295 regions.set_index(i+1);
4296 } else {
4297 regions.pop();
4298 }
4299 }
4300 if (leading == nullptr) {
4301 assert(regions.size() == 0, "all paths should have been tried");
4302 return nullptr;
4303 }
4304 }
4305 if (leading->is_Region()) {
4306 regions.push(leading, 2);
4307 leading = leading->in(1);
4308 } else {
4309 leading = leading->in(0);
4310 }
4311 }
4312 #ifdef ASSERT
4313 Unique_Node_List wq;
4314 wq.push((Node*)this);
4315 uint found = 0;
4316 for (uint i = 0; i < wq.size(); i++) {
4317 Node* n = wq.at(i);
4318 if (n->is_Region()) {
4319 for (uint j = 1; j < n->req(); j++) {
4320 Node* in = n->in(j);
4321 if (in != nullptr && !in->is_top()) {
4322 wq.push(in);
4323 }
4324 }
4325 } else {
4326 if (n->is_MemBar() && n->as_MemBar()->leading()) {
4327 assert(n == leading, "consistency check failed");
4328 found++;
4329 } else {
4330 Node* in = n->in(0);
4331 if (in != nullptr && !in->is_top()) {
4332 wq.push(in);
4333 }
4334 }
4335 }
4336 }
4337 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed");
4338 #endif
4339 if (leading == nullptr) {
4340 return nullptr;
4341 }
4342 MemBarNode* mb = leading->as_MemBar();
4343 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
4344 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
4345 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
4346 return mb;
4347 }
4348
4349
4350 //===========================InitializeNode====================================
4351 // SUMMARY:
4352 // This node acts as a memory barrier on raw memory, after some raw stores.
4353 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
4354 // The Initialize can 'capture' suitably constrained stores as raw inits.
4355 // It can coalesce related raw stores into larger units (called 'tiles').
4356 // It can avoid zeroing new storage for memory units which have raw inits.
4357 // At macro-expansion, it is marked 'complete', and does not optimize further.
4358 //
4359 // EXAMPLE:
4360 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
4361 // ctl = incoming control; mem* = incoming memory
4362 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
4363 // First allocate uninitialized memory and fill in the header:
4364 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
4365 // ctl := alloc.Control; mem* := alloc.Memory*
4366 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
4367 // Then initialize to zero the non-header parts of the raw memory block:
4368 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
4369 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
4370 // After the initialize node executes, the object is ready for service:
4371 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
4372 // Suppose its body is immediately initialized as {1,2}:
4373 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4374 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4375 // mem.SLICE(#short[*]) := store2
4376 //
4377 // DETAILS:
4378 // An InitializeNode collects and isolates object initialization after
4379 // an AllocateNode and before the next possible safepoint. As a
4380 // memory barrier (MemBarNode), it keeps critical stores from drifting
4381 // down past any safepoint or any publication of the allocation.
4382 // Before this barrier, a newly-allocated object may have uninitialized bits.
4383 // After this barrier, it may be treated as a real oop, and GC is allowed.
4384 //
4385 // The semantics of the InitializeNode include an implicit zeroing of
4386 // the new object from object header to the end of the object.
4387 // (The object header and end are determined by the AllocateNode.)
4388 //
4389 // Certain stores may be added as direct inputs to the InitializeNode.
4390 // These stores must update raw memory, and they must be to addresses
4391 // derived from the raw address produced by AllocateNode, and with
4392 // a constant offset. They must be ordered by increasing offset.
4393 // The first one is at in(RawStores), the last at in(req()-1).
4394 // Unlike most memory operations, they are not linked in a chain,
4395 // but are displayed in parallel as users of the rawmem output of
4396 // the allocation.
4397 //
4398 // (See comments in InitializeNode::capture_store, which continue
4399 // the example given above.)
4400 //
4401 // When the associated Allocate is macro-expanded, the InitializeNode
4402 // may be rewritten to optimize collected stores. A ClearArrayNode
4403 // may also be created at that point to represent any required zeroing.
4404 // The InitializeNode is then marked 'complete', prohibiting further
4405 // capturing of nearby memory operations.
4406 //
4407 // During macro-expansion, all captured initializations which store
4408 // constant values of 32 bits or smaller are coalesced (if advantageous)
4409 // into larger 'tiles' 32 or 64 bits. This allows an object to be
4410 // initialized in fewer memory operations. Memory words which are
4411 // covered by neither tiles nor non-constant stores are pre-zeroed
4412 // by explicit stores of zero. (The code shape happens to do all
4413 // zeroing first, then all other stores, with both sequences occurring
4414 // in order of ascending offsets.)
4415 //
4416 // Alternatively, code may be inserted between an AllocateNode and its
4417 // InitializeNode, to perform arbitrary initialization of the new object.
4418 // E.g., the object copying intrinsics insert complex data transfers here.
4419 // The initialization must then be marked as 'complete' disable the
4420 // built-in zeroing semantics and the collection of initializing stores.
4421 //
4422 // While an InitializeNode is incomplete, reads from the memory state
4423 // produced by it are optimizable if they match the control edge and
4424 // new oop address associated with the allocation/initialization.
4425 // They return a stored value (if the offset matches) or else zero.
4426 // A write to the memory state, if it matches control and address,
4427 // and if it is to a constant offset, may be 'captured' by the
4428 // InitializeNode. It is cloned as a raw memory operation and rewired
4429 // inside the initialization, to the raw oop produced by the allocation.
4430 // Operations on addresses which are provably distinct (e.g., to
4431 // other AllocateNodes) are allowed to bypass the initialization.
4432 //
4433 // The effect of all this is to consolidate object initialization
4434 // (both arrays and non-arrays, both piecewise and bulk) into a
4435 // single location, where it can be optimized as a unit.
4436 //
4437 // Only stores with an offset less than TrackedInitializationLimit words
4438 // will be considered for capture by an InitializeNode. This puts a
4439 // reasonable limit on the complexity of optimized initializations.
4440
4441 //---------------------------InitializeNode------------------------------------
4442 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
4443 : MemBarNode(C, adr_type, rawoop),
4444 _is_complete(Incomplete), _does_not_escape(false)
4445 {
4446 init_class_id(Class_Initialize);
4447
4448 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
4449 assert(in(RawAddress) == rawoop, "proper init");
4450 // Note: allocation() can be null, for secondary initialization barriers
4451 }
4452
4453 // Since this node is not matched, it will be processed by the
4454 // register allocator. Declare that there are no constraints
4455 // on the allocation of the RawAddress edge.
4456 const RegMask &InitializeNode::in_RegMask(uint idx) const {
4457 // This edge should be set to top, by the set_complete. But be conservative.
4458 if (idx == InitializeNode::RawAddress)
4459 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
4460 return RegMask::Empty;
4461 }
4462
4463 Node* InitializeNode::memory(uint alias_idx) {
4464 Node* mem = in(Memory);
4465 if (mem->is_MergeMem()) {
4466 return mem->as_MergeMem()->memory_at(alias_idx);
4467 } else {
4468 // incoming raw memory is not split
4469 return mem;
4470 }
4471 }
4472
4473 bool InitializeNode::is_non_zero() {
4474 if (is_complete()) return false;
4475 remove_extra_zeroes();
4476 return (req() > RawStores);
4477 }
4478
4479 void InitializeNode::set_complete(PhaseGVN* phase) {
4480 assert(!is_complete(), "caller responsibility");
4481 _is_complete = Complete;
4482
4483 // After this node is complete, it contains a bunch of
4484 // raw-memory initializations. There is no need for
4485 // it to have anything to do with non-raw memory effects.
4486 // Therefore, tell all non-raw users to re-optimize themselves,
4487 // after skipping the memory effects of this initialization.
4488 PhaseIterGVN* igvn = phase->is_IterGVN();
4489 if (igvn) igvn->add_users_to_worklist(this);
4490 }
4491
4492 // convenience function
4493 // return false if the init contains any stores already
4494 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
4495 InitializeNode* init = initialization();
4496 if (init == nullptr || init->is_complete()) return false;
4497 init->remove_extra_zeroes();
4498 // for now, if this allocation has already collected any inits, bail:
4499 if (init->is_non_zero()) return false;
4500 init->set_complete(phase);
4501 return true;
4502 }
4503
4504 void InitializeNode::remove_extra_zeroes() {
4505 if (req() == RawStores) return;
4506 Node* zmem = zero_memory();
4507 uint fill = RawStores;
4508 for (uint i = fill; i < req(); i++) {
4509 Node* n = in(i);
4510 if (n->is_top() || n == zmem) continue; // skip
4511 if (fill < i) set_req(fill, n); // compact
4512 ++fill;
4513 }
4514 // delete any empty spaces created:
4515 while (fill < req()) {
4516 del_req(fill);
4517 }
4518 }
4519
4520 // Helper for remembering which stores go with which offsets.
4521 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) {
4522 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
4523 intptr_t offset = -1;
4524 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
4525 phase, offset);
4526 if (base == nullptr) return -1; // something is dead,
4527 if (offset < 0) return -1; // dead, dead
4528 return offset;
4529 }
4530
4531 // Helper for proving that an initialization expression is
4532 // "simple enough" to be folded into an object initialization.
4533 // Attempts to prove that a store's initial value 'n' can be captured
4534 // within the initialization without creating a vicious cycle, such as:
4535 // { Foo p = new Foo(); p.next = p; }
4536 // True for constants and parameters and small combinations thereof.
4537 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
4538 ResourceMark rm;
4539 Unique_Node_List worklist;
4540 worklist.push(value);
4541
4542 uint complexity_limit = 20;
4543 for (uint j = 0; j < worklist.size(); j++) {
4544 if (j >= complexity_limit) {
4545 return false; // Bail out if processed too many nodes
4546 }
4547
4548 Node* n = worklist.at(j);
4549 if (n == nullptr) continue; // (can this really happen?)
4550 if (n->is_Proj()) n = n->in(0);
4551 if (n == this) return false; // found a cycle
4552 if (n->is_Con()) continue;
4553 if (n->is_Start()) continue; // params, etc., are OK
4554 if (n->is_Root()) continue; // even better
4555
4556 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
4557 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
4558 continue;
4559 }
4560
4561 Node* ctl = n->in(0);
4562 if (ctl != nullptr && !ctl->is_top()) {
4563 if (ctl->is_Proj()) ctl = ctl->in(0);
4564 if (ctl == this) return false;
4565
4566 // If we already know that the enclosing memory op is pinned right after
4567 // the init, then any control flow that the store has picked up
4568 // must have preceded the init, or else be equal to the init.
4569 // Even after loop optimizations (which might change control edges)
4570 // a store is never pinned *before* the availability of its inputs.
4571 if (!MemNode::all_controls_dominate(n, this)) {
4572 return false; // failed to prove a good control
4573 }
4574 }
4575
4576 // Check data edges for possible dependencies on 'this'.
4577 for (uint i = 1; i < n->req(); i++) {
4578 Node* m = n->in(i);
4579 if (m == nullptr || m == n || m->is_top()) continue;
4580
4581 // Only process data inputs once
4582 worklist.push(m);
4583 }
4584 }
4585
4586 return true;
4587 }
4588
4589 // Here are all the checks a Store must pass before it can be moved into
4590 // an initialization. Returns zero if a check fails.
4591 // On success, returns the (constant) offset to which the store applies,
4592 // within the initialized memory.
4593 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
4594 const int FAIL = 0;
4595 if (st->req() != MemNode::ValueIn + 1)
4596 return FAIL; // an inscrutable StoreNode (card mark?)
4597 Node* ctl = st->in(MemNode::Control);
4598 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this))
4599 return FAIL; // must be unconditional after the initialization
4600 Node* mem = st->in(MemNode::Memory);
4601 if (!(mem->is_Proj() && mem->in(0) == this))
4602 return FAIL; // must not be preceded by other stores
4603 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4604 if ((st->Opcode() == Op_StoreP || st->Opcode() == Op_StoreN) &&
4605 !bs->can_initialize_object(st)) {
4606 return FAIL;
4607 }
4608 Node* adr = st->in(MemNode::Address);
4609 intptr_t offset;
4610 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
4611 if (alloc == nullptr)
4612 return FAIL; // inscrutable address
4613 if (alloc != allocation())
4614 return FAIL; // wrong allocation! (store needs to float up)
4615 int size_in_bytes = st->memory_size();
4616 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
4617 return FAIL; // mismatched access
4618 }
4619 Node* val = st->in(MemNode::ValueIn);
4620
4621 if (!detect_init_independence(val, phase))
4622 return FAIL; // stored value must be 'simple enough'
4623
4624 // The Store can be captured only if nothing after the allocation
4625 // and before the Store is using the memory location that the store
4626 // overwrites.
4627 bool failed = false;
4628 // If is_complete_with_arraycopy() is true the shape of the graph is
4629 // well defined and is safe so no need for extra checks.
4630 if (!is_complete_with_arraycopy()) {
4631 // We are going to look at each use of the memory state following
4632 // the allocation to make sure nothing reads the memory that the
4633 // Store writes.
4634 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
4635 int alias_idx = phase->C->get_alias_index(t_adr);
4636 ResourceMark rm;
4637 Unique_Node_List mems;
4638 mems.push(mem);
4639 Node* unique_merge = nullptr;
4640 for (uint next = 0; next < mems.size(); ++next) {
4641 Node *m = mems.at(next);
4642 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
4643 Node *n = m->fast_out(j);
4644 if (n->outcnt() == 0) {
4645 continue;
4646 }
4647 if (n == st) {
4648 continue;
4649 } else if (n->in(0) != nullptr && n->in(0) != ctl) {
4650 // If the control of this use is different from the control
4651 // of the Store which is right after the InitializeNode then
4652 // this node cannot be between the InitializeNode and the
4653 // Store.
4654 continue;
4655 } else if (n->is_MergeMem()) {
4656 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
4657 // We can hit a MergeMemNode (that will likely go away
4658 // later) that is a direct use of the memory state
4659 // following the InitializeNode on the same slice as the
4660 // store node that we'd like to capture. We need to check
4661 // the uses of the MergeMemNode.
4662 mems.push(n);
4663 }
4664 } else if (n->is_Mem()) {
4665 Node* other_adr = n->in(MemNode::Address);
4666 if (other_adr == adr) {
4667 failed = true;
4668 break;
4669 } else {
4670 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
4671 if (other_t_adr != nullptr) {
4672 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
4673 if (other_alias_idx == alias_idx) {
4674 // A load from the same memory slice as the store right
4675 // after the InitializeNode. We check the control of the
4676 // object/array that is loaded from. If it's the same as
4677 // the store control then we cannot capture the store.
4678 assert(!n->is_Store(), "2 stores to same slice on same control?");
4679 Node* base = other_adr;
4680 assert(base->is_AddP(), "should be addp but is %s", base->Name());
4681 base = base->in(AddPNode::Base);
4682 if (base != nullptr) {
4683 base = base->uncast();
4684 if (base->is_Proj() && base->in(0) == alloc) {
4685 failed = true;
4686 break;
4687 }
4688 }
4689 }
4690 }
4691 }
4692 } else {
4693 failed = true;
4694 break;
4695 }
4696 }
4697 }
4698 }
4699 if (failed) {
4700 if (!can_reshape) {
4701 // We decided we couldn't capture the store during parsing. We
4702 // should try again during the next IGVN once the graph is
4703 // cleaner.
4704 phase->C->record_for_igvn(st);
4705 }
4706 return FAIL;
4707 }
4708
4709 return offset; // success
4710 }
4711
4712 // Find the captured store in(i) which corresponds to the range
4713 // [start..start+size) in the initialized object.
4714 // If there is one, return its index i. If there isn't, return the
4715 // negative of the index where it should be inserted.
4716 // Return 0 if the queried range overlaps an initialization boundary
4717 // or if dead code is encountered.
4718 // If size_in_bytes is zero, do not bother with overlap checks.
4719 int InitializeNode::captured_store_insertion_point(intptr_t start,
4720 int size_in_bytes,
4721 PhaseValues* phase) {
4722 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
4723
4724 if (is_complete())
4725 return FAIL; // arraycopy got here first; punt
4726
4727 assert(allocation() != nullptr, "must be present");
4728
4729 // no negatives, no header fields:
4730 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
4731
4732 // after a certain size, we bail out on tracking all the stores:
4733 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4734 if (start >= ti_limit) return FAIL;
4735
4736 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
4737 if (i >= limit) return -(int)i; // not found; here is where to put it
4738
4739 Node* st = in(i);
4740 intptr_t st_off = get_store_offset(st, phase);
4741 if (st_off < 0) {
4742 if (st != zero_memory()) {
4743 return FAIL; // bail out if there is dead garbage
4744 }
4745 } else if (st_off > start) {
4746 // ...we are done, since stores are ordered
4747 if (st_off < start + size_in_bytes) {
4748 return FAIL; // the next store overlaps
4749 }
4750 return -(int)i; // not found; here is where to put it
4751 } else if (st_off < start) {
4752 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
4753 if (size_in_bytes != 0 &&
4754 start < st_off + MAX_STORE &&
4755 start < st_off + st->as_Store()->memory_size()) {
4756 return FAIL; // the previous store overlaps
4757 }
4758 } else {
4759 if (size_in_bytes != 0 &&
4760 st->as_Store()->memory_size() != size_in_bytes) {
4761 return FAIL; // mismatched store size
4762 }
4763 return i;
4764 }
4765
4766 ++i;
4767 }
4768 }
4769
4770 // Look for a captured store which initializes at the offset 'start'
4771 // with the given size. If there is no such store, and no other
4772 // initialization interferes, then return zero_memory (the memory
4773 // projection of the AllocateNode).
4774 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
4775 PhaseValues* phase) {
4776 assert(stores_are_sane(phase), "");
4777 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4778 if (i == 0) {
4779 return nullptr; // something is dead
4780 } else if (i < 0) {
4781 return zero_memory(); // just primordial zero bits here
4782 } else {
4783 Node* st = in(i); // here is the store at this position
4784 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
4785 return st;
4786 }
4787 }
4788
4789 // Create, as a raw pointer, an address within my new object at 'offset'.
4790 Node* InitializeNode::make_raw_address(intptr_t offset,
4791 PhaseGVN* phase) {
4792 Node* addr = in(RawAddress);
4793 if (offset != 0) {
4794 Compile* C = phase->C;
4795 addr = phase->transform( new AddPNode(C->top(), addr,
4796 phase->MakeConX(offset)) );
4797 }
4798 return addr;
4799 }
4800
4801 // Clone the given store, converting it into a raw store
4802 // initializing a field or element of my new object.
4803 // Caller is responsible for retiring the original store,
4804 // with subsume_node or the like.
4805 //
4806 // From the example above InitializeNode::InitializeNode,
4807 // here are the old stores to be captured:
4808 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4809 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4810 //
4811 // Here is the changed code; note the extra edges on init:
4812 // alloc = (Allocate ...)
4813 // rawoop = alloc.RawAddress
4814 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
4815 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
4816 // init = (Initialize alloc.Control alloc.Memory rawoop
4817 // rawstore1 rawstore2)
4818 //
4819 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
4820 PhaseGVN* phase, bool can_reshape) {
4821 assert(stores_are_sane(phase), "");
4822
4823 if (start < 0) return nullptr;
4824 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
4825
4826 Compile* C = phase->C;
4827 int size_in_bytes = st->memory_size();
4828 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4829 if (i == 0) return nullptr; // bail out
4830 Node* prev_mem = nullptr; // raw memory for the captured store
4831 if (i > 0) {
4832 prev_mem = in(i); // there is a pre-existing store under this one
4833 set_req(i, C->top()); // temporarily disconnect it
4834 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
4835 } else {
4836 i = -i; // no pre-existing store
4837 prev_mem = zero_memory(); // a slice of the newly allocated object
4838 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
4839 set_req(--i, C->top()); // reuse this edge; it has been folded away
4840 else
4841 ins_req(i, C->top()); // build a new edge
4842 }
4843 Node* new_st = st->clone();
4844 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4845 new_st->set_req(MemNode::Control, in(Control));
4846 new_st->set_req(MemNode::Memory, prev_mem);
4847 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
4848 bs->eliminate_gc_barrier_data(new_st);
4849 new_st = phase->transform(new_st);
4850
4851 // At this point, new_st might have swallowed a pre-existing store
4852 // at the same offset, or perhaps new_st might have disappeared,
4853 // if it redundantly stored the same value (or zero to fresh memory).
4854
4855 // In any case, wire it in:
4856 PhaseIterGVN* igvn = phase->is_IterGVN();
4857 if (igvn) {
4858 igvn->rehash_node_delayed(this);
4859 }
4860 set_req(i, new_st);
4861
4862 // The caller may now kill the old guy.
4863 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
4864 assert(check_st == new_st || check_st == nullptr, "must be findable");
4865 assert(!is_complete(), "");
4866 return new_st;
4867 }
4868
4869 static bool store_constant(jlong* tiles, int num_tiles,
4870 intptr_t st_off, int st_size,
4871 jlong con) {
4872 if ((st_off & (st_size-1)) != 0)
4873 return false; // strange store offset (assume size==2**N)
4874 address addr = (address)tiles + st_off;
4875 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
4876 switch (st_size) {
4877 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
4878 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
4879 case sizeof(jint): *(jint*) addr = (jint) con; break;
4880 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
4881 default: return false; // strange store size (detect size!=2**N here)
4882 }
4883 return true; // return success to caller
4884 }
4885
4886 // Coalesce subword constants into int constants and possibly
4887 // into long constants. The goal, if the CPU permits,
4888 // is to initialize the object with a small number of 64-bit tiles.
4889 // Also, convert floating-point constants to bit patterns.
4890 // Non-constants are not relevant to this pass.
4891 //
4892 // In terms of the running example on InitializeNode::InitializeNode
4893 // and InitializeNode::capture_store, here is the transformation
4894 // of rawstore1 and rawstore2 into rawstore12:
4895 // alloc = (Allocate ...)
4896 // rawoop = alloc.RawAddress
4897 // tile12 = 0x00010002
4898 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
4899 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
4900 //
4901 void
4902 InitializeNode::coalesce_subword_stores(intptr_t header_size,
4903 Node* size_in_bytes,
4904 PhaseGVN* phase) {
4905 Compile* C = phase->C;
4906
4907 assert(stores_are_sane(phase), "");
4908 // Note: After this pass, they are not completely sane,
4909 // since there may be some overlaps.
4910
4911 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
4912
4913 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4914 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
4915 size_limit = MIN2(size_limit, ti_limit);
4916 size_limit = align_up(size_limit, BytesPerLong);
4917 int num_tiles = size_limit / BytesPerLong;
4918
4919 // allocate space for the tile map:
4920 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
4921 jlong tiles_buf[small_len];
4922 Node* nodes_buf[small_len];
4923 jlong inits_buf[small_len];
4924 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
4925 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4926 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
4927 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
4928 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
4929 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4930 // tiles: exact bitwise model of all primitive constants
4931 // nodes: last constant-storing node subsumed into the tiles model
4932 // inits: which bytes (in each tile) are touched by any initializations
4933
4934 //// Pass A: Fill in the tile model with any relevant stores.
4935
4936 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
4937 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
4938 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
4939 Node* zmem = zero_memory(); // initially zero memory state
4940 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4941 Node* st = in(i);
4942 intptr_t st_off = get_store_offset(st, phase);
4943
4944 // Figure out the store's offset and constant value:
4945 if (st_off < header_size) continue; //skip (ignore header)
4946 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
4947 int st_size = st->as_Store()->memory_size();
4948 if (st_off + st_size > size_limit) break;
4949
4950 // Record which bytes are touched, whether by constant or not.
4951 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
4952 continue; // skip (strange store size)
4953
4954 const Type* val = phase->type(st->in(MemNode::ValueIn));
4955 if (!val->singleton()) continue; //skip (non-con store)
4956 BasicType type = val->basic_type();
4957
4958 jlong con = 0;
4959 switch (type) {
4960 case T_INT: con = val->is_int()->get_con(); break;
4961 case T_LONG: con = val->is_long()->get_con(); break;
4962 case T_FLOAT: con = jint_cast(val->getf()); break;
4963 case T_DOUBLE: con = jlong_cast(val->getd()); break;
4964 default: continue; //skip (odd store type)
4965 }
4966
4967 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
4968 st->Opcode() == Op_StoreL) {
4969 continue; // This StoreL is already optimal.
4970 }
4971
4972 // Store down the constant.
4973 store_constant(tiles, num_tiles, st_off, st_size, con);
4974
4975 intptr_t j = st_off >> LogBytesPerLong;
4976
4977 if (type == T_INT && st_size == BytesPerInt
4978 && (st_off & BytesPerInt) == BytesPerInt) {
4979 jlong lcon = tiles[j];
4980 if (!Matcher::isSimpleConstant64(lcon) &&
4981 st->Opcode() == Op_StoreI) {
4982 // This StoreI is already optimal by itself.
4983 jint* intcon = (jint*) &tiles[j];
4984 intcon[1] = 0; // undo the store_constant()
4985
4986 // If the previous store is also optimal by itself, back up and
4987 // undo the action of the previous loop iteration... if we can.
4988 // But if we can't, just let the previous half take care of itself.
4989 st = nodes[j];
4990 st_off -= BytesPerInt;
4991 con = intcon[0];
4992 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) {
4993 assert(st_off >= header_size, "still ignoring header");
4994 assert(get_store_offset(st, phase) == st_off, "must be");
4995 assert(in(i-1) == zmem, "must be");
4996 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
4997 assert(con == tcon->is_int()->get_con(), "must be");
4998 // Undo the effects of the previous loop trip, which swallowed st:
4999 intcon[0] = 0; // undo store_constant()
5000 set_req(i-1, st); // undo set_req(i, zmem)
5001 nodes[j] = nullptr; // undo nodes[j] = st
5002 --old_subword; // undo ++old_subword
5003 }
5004 continue; // This StoreI is already optimal.
5005 }
5006 }
5007
5008 // This store is not needed.
5009 set_req(i, zmem);
5010 nodes[j] = st; // record for the moment
5011 if (st_size < BytesPerLong) // something has changed
5012 ++old_subword; // includes int/float, but who's counting...
5013 else ++old_long;
5014 }
5015
5016 if ((old_subword + old_long) == 0)
5017 return; // nothing more to do
5018
5019 //// Pass B: Convert any non-zero tiles into optimal constant stores.
5020 // Be sure to insert them before overlapping non-constant stores.
5021 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
5022 for (int j = 0; j < num_tiles; j++) {
5023 jlong con = tiles[j];
5024 jlong init = inits[j];
5025 if (con == 0) continue;
5026 jint con0, con1; // split the constant, address-wise
5027 jint init0, init1; // split the init map, address-wise
5028 { union { jlong con; jint intcon[2]; } u;
5029 u.con = con;
5030 con0 = u.intcon[0];
5031 con1 = u.intcon[1];
5032 u.con = init;
5033 init0 = u.intcon[0];
5034 init1 = u.intcon[1];
5035 }
5036
5037 Node* old = nodes[j];
5038 assert(old != nullptr, "need the prior store");
5039 intptr_t offset = (j * BytesPerLong);
5040
5041 bool split = !Matcher::isSimpleConstant64(con);
5042
5043 if (offset < header_size) {
5044 assert(offset + BytesPerInt >= header_size, "second int counts");
5045 assert(*(jint*)&tiles[j] == 0, "junk in header");
5046 split = true; // only the second word counts
5047 // Example: int a[] = { 42 ... }
5048 } else if (con0 == 0 && init0 == -1) {
5049 split = true; // first word is covered by full inits
5050 // Example: int a[] = { ... foo(), 42 ... }
5051 } else if (con1 == 0 && init1 == -1) {
5052 split = true; // second word is covered by full inits
5053 // Example: int a[] = { ... 42, foo() ... }
5054 }
5055
5056 // Here's a case where init0 is neither 0 nor -1:
5057 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
5058 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
5059 // In this case the tile is not split; it is (jlong)42.
5060 // The big tile is stored down, and then the foo() value is inserted.
5061 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
5062
5063 Node* ctl = old->in(MemNode::Control);
5064 Node* adr = make_raw_address(offset, phase);
5065 const TypePtr* atp = TypeRawPtr::BOTTOM;
5066
5067 // One or two coalesced stores to plop down.
5068 Node* st[2];
5069 intptr_t off[2];
5070 int nst = 0;
5071 if (!split) {
5072 ++new_long;
5073 off[nst] = offset;
5074 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5075 phase->longcon(con), T_LONG, MemNode::unordered);
5076 } else {
5077 // Omit either if it is a zero.
5078 if (con0 != 0) {
5079 ++new_int;
5080 off[nst] = offset;
5081 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5082 phase->intcon(con0), T_INT, MemNode::unordered);
5083 }
5084 if (con1 != 0) {
5085 ++new_int;
5086 offset += BytesPerInt;
5087 adr = make_raw_address(offset, phase);
5088 off[nst] = offset;
5089 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5090 phase->intcon(con1), T_INT, MemNode::unordered);
5091 }
5092 }
5093
5094 // Insert second store first, then the first before the second.
5095 // Insert each one just before any overlapping non-constant stores.
5096 while (nst > 0) {
5097 Node* st1 = st[--nst];
5098 C->copy_node_notes_to(st1, old);
5099 st1 = phase->transform(st1);
5100 offset = off[nst];
5101 assert(offset >= header_size, "do not smash header");
5102 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
5103 guarantee(ins_idx != 0, "must re-insert constant store");
5104 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
5105 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
5106 set_req(--ins_idx, st1);
5107 else
5108 ins_req(ins_idx, st1);
5109 }
5110 }
5111
5112 if (PrintCompilation && WizardMode)
5113 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
5114 old_subword, old_long, new_int, new_long);
5115 if (C->log() != nullptr)
5116 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
5117 old_subword, old_long, new_int, new_long);
5118
5119 // Clean up any remaining occurrences of zmem:
5120 remove_extra_zeroes();
5121 }
5122
5123 // Explore forward from in(start) to find the first fully initialized
5124 // word, and return its offset. Skip groups of subword stores which
5125 // together initialize full words. If in(start) is itself part of a
5126 // fully initialized word, return the offset of in(start). If there
5127 // are no following full-word stores, or if something is fishy, return
5128 // a negative value.
5129 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
5130 int int_map = 0;
5131 intptr_t int_map_off = 0;
5132 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
5133
5134 for (uint i = start, limit = req(); i < limit; i++) {
5135 Node* st = in(i);
5136
5137 intptr_t st_off = get_store_offset(st, phase);
5138 if (st_off < 0) break; // return conservative answer
5139
5140 int st_size = st->as_Store()->memory_size();
5141 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
5142 return st_off; // we found a complete word init
5143 }
5144
5145 // update the map:
5146
5147 intptr_t this_int_off = align_down(st_off, BytesPerInt);
5148 if (this_int_off != int_map_off) {
5149 // reset the map:
5150 int_map = 0;
5151 int_map_off = this_int_off;
5152 }
5153
5154 int subword_off = st_off - this_int_off;
5155 int_map |= right_n_bits(st_size) << subword_off;
5156 if ((int_map & FULL_MAP) == FULL_MAP) {
5157 return this_int_off; // we found a complete word init
5158 }
5159
5160 // Did this store hit or cross the word boundary?
5161 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
5162 if (next_int_off == this_int_off + BytesPerInt) {
5163 // We passed the current int, without fully initializing it.
5164 int_map_off = next_int_off;
5165 int_map >>= BytesPerInt;
5166 } else if (next_int_off > this_int_off + BytesPerInt) {
5167 // We passed the current and next int.
5168 return this_int_off + BytesPerInt;
5169 }
5170 }
5171
5172 return -1;
5173 }
5174
5175
5176 // Called when the associated AllocateNode is expanded into CFG.
5177 // At this point, we may perform additional optimizations.
5178 // Linearize the stores by ascending offset, to make memory
5179 // activity as coherent as possible.
5180 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
5181 intptr_t header_size,
5182 Node* size_in_bytes,
5183 PhaseIterGVN* phase) {
5184 assert(!is_complete(), "not already complete");
5185 assert(stores_are_sane(phase), "");
5186 assert(allocation() != nullptr, "must be present");
5187
5188 remove_extra_zeroes();
5189
5190 if (ReduceFieldZeroing || ReduceBulkZeroing)
5191 // reduce instruction count for common initialization patterns
5192 coalesce_subword_stores(header_size, size_in_bytes, phase);
5193
5194 Node* zmem = zero_memory(); // initially zero memory state
5195 Node* inits = zmem; // accumulating a linearized chain of inits
5196 #ifdef ASSERT
5197 intptr_t first_offset = allocation()->minimum_header_size();
5198 intptr_t last_init_off = first_offset; // previous init offset
5199 intptr_t last_init_end = first_offset; // previous init offset+size
5200 intptr_t last_tile_end = first_offset; // previous tile offset+size
5201 #endif
5202 intptr_t zeroes_done = header_size;
5203
5204 bool do_zeroing = true; // we might give up if inits are very sparse
5205 int big_init_gaps = 0; // how many large gaps have we seen?
5206
5207 if (UseTLAB && ZeroTLAB) do_zeroing = false;
5208 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
5209
5210 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5211 Node* st = in(i);
5212 intptr_t st_off = get_store_offset(st, phase);
5213 if (st_off < 0)
5214 break; // unknown junk in the inits
5215 if (st->in(MemNode::Memory) != zmem)
5216 break; // complicated store chains somehow in list
5217
5218 int st_size = st->as_Store()->memory_size();
5219 intptr_t next_init_off = st_off + st_size;
5220
5221 if (do_zeroing && zeroes_done < next_init_off) {
5222 // See if this store needs a zero before it or under it.
5223 intptr_t zeroes_needed = st_off;
5224
5225 if (st_size < BytesPerInt) {
5226 // Look for subword stores which only partially initialize words.
5227 // If we find some, we must lay down some word-level zeroes first,
5228 // underneath the subword stores.
5229 //
5230 // Examples:
5231 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
5232 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
5233 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
5234 //
5235 // Note: coalesce_subword_stores may have already done this,
5236 // if it was prompted by constant non-zero subword initializers.
5237 // But this case can still arise with non-constant stores.
5238
5239 intptr_t next_full_store = find_next_fullword_store(i, phase);
5240
5241 // In the examples above:
5242 // in(i) p q r s x y z
5243 // st_off 12 13 14 15 12 13 14
5244 // st_size 1 1 1 1 1 1 1
5245 // next_full_s. 12 16 16 16 16 16 16
5246 // z's_done 12 16 16 16 12 16 12
5247 // z's_needed 12 16 16 16 16 16 16
5248 // zsize 0 0 0 0 4 0 4
5249 if (next_full_store < 0) {
5250 // Conservative tack: Zero to end of current word.
5251 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
5252 } else {
5253 // Zero to beginning of next fully initialized word.
5254 // Or, don't zero at all, if we are already in that word.
5255 assert(next_full_store >= zeroes_needed, "must go forward");
5256 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
5257 zeroes_needed = next_full_store;
5258 }
5259 }
5260
5261 if (zeroes_needed > zeroes_done) {
5262 intptr_t zsize = zeroes_needed - zeroes_done;
5263 // Do some incremental zeroing on rawmem, in parallel with inits.
5264 zeroes_done = align_down(zeroes_done, BytesPerInt);
5265 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5266 zeroes_done, zeroes_needed,
5267 phase);
5268 zeroes_done = zeroes_needed;
5269 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
5270 do_zeroing = false; // leave the hole, next time
5271 }
5272 }
5273
5274 // Collect the store and move on:
5275 phase->replace_input_of(st, MemNode::Memory, inits);
5276 inits = st; // put it on the linearized chain
5277 set_req(i, zmem); // unhook from previous position
5278
5279 if (zeroes_done == st_off)
5280 zeroes_done = next_init_off;
5281
5282 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
5283
5284 #ifdef ASSERT
5285 // Various order invariants. Weaker than stores_are_sane because
5286 // a large constant tile can be filled in by smaller non-constant stores.
5287 assert(st_off >= last_init_off, "inits do not reverse");
5288 last_init_off = st_off;
5289 const Type* val = nullptr;
5290 if (st_size >= BytesPerInt &&
5291 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
5292 (int)val->basic_type() < (int)T_OBJECT) {
5293 assert(st_off >= last_tile_end, "tiles do not overlap");
5294 assert(st_off >= last_init_end, "tiles do not overwrite inits");
5295 last_tile_end = MAX2(last_tile_end, next_init_off);
5296 } else {
5297 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
5298 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
5299 assert(st_off >= last_init_end, "inits do not overlap");
5300 last_init_end = next_init_off; // it's a non-tile
5301 }
5302 #endif //ASSERT
5303 }
5304
5305 remove_extra_zeroes(); // clear out all the zmems left over
5306 add_req(inits);
5307
5308 if (!(UseTLAB && ZeroTLAB)) {
5309 // If anything remains to be zeroed, zero it all now.
5310 zeroes_done = align_down(zeroes_done, BytesPerInt);
5311 // if it is the last unused 4 bytes of an instance, forget about it
5312 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
5313 if (zeroes_done + BytesPerLong >= size_limit) {
5314 AllocateNode* alloc = allocation();
5315 assert(alloc != nullptr, "must be present");
5316 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
5317 Node* klass_node = alloc->in(AllocateNode::KlassNode);
5318 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
5319 if (zeroes_done == k->layout_helper())
5320 zeroes_done = size_limit;
5321 }
5322 }
5323 if (zeroes_done < size_limit) {
5324 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5325 zeroes_done, size_in_bytes, phase);
5326 }
5327 }
5328
5329 set_complete(phase);
5330 return rawmem;
5331 }
5332
5333
5334 #ifdef ASSERT
5335 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
5336 if (is_complete())
5337 return true; // stores could be anything at this point
5338 assert(allocation() != nullptr, "must be present");
5339 intptr_t last_off = allocation()->minimum_header_size();
5340 for (uint i = InitializeNode::RawStores; i < req(); i++) {
5341 Node* st = in(i);
5342 intptr_t st_off = get_store_offset(st, phase);
5343 if (st_off < 0) continue; // ignore dead garbage
5344 if (last_off > st_off) {
5345 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
5346 this->dump(2);
5347 assert(false, "ascending store offsets");
5348 return false;
5349 }
5350 last_off = st_off + st->as_Store()->memory_size();
5351 }
5352 return true;
5353 }
5354 #endif //ASSERT
5355
5356
5357
5358
5359 //============================MergeMemNode=====================================
5360 //
5361 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
5362 // contributing store or call operations. Each contributor provides the memory
5363 // state for a particular "alias type" (see Compile::alias_type). For example,
5364 // if a MergeMem has an input X for alias category #6, then any memory reference
5365 // to alias category #6 may use X as its memory state input, as an exact equivalent
5366 // to using the MergeMem as a whole.
5367 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
5368 //
5369 // (Here, the <N> notation gives the index of the relevant adr_type.)
5370 //
5371 // In one special case (and more cases in the future), alias categories overlap.
5372 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
5373 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
5374 // it is exactly equivalent to that state W:
5375 // MergeMem(<Bot>: W) <==> W
5376 //
5377 // Usually, the merge has more than one input. In that case, where inputs
5378 // overlap (i.e., one is Bot), the narrower alias type determines the memory
5379 // state for that type, and the wider alias type (Bot) fills in everywhere else:
5380 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
5381 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
5382 //
5383 // A merge can take a "wide" memory state as one of its narrow inputs.
5384 // This simply means that the merge observes out only the relevant parts of
5385 // the wide input. That is, wide memory states arriving at narrow merge inputs
5386 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
5387 //
5388 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
5389 // and that memory slices "leak through":
5390 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
5391 //
5392 // But, in such a cascade, repeated memory slices can "block the leak":
5393 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
5394 //
5395 // In the last example, Y is not part of the combined memory state of the
5396 // outermost MergeMem. The system must, of course, prevent unschedulable
5397 // memory states from arising, so you can be sure that the state Y is somehow
5398 // a precursor to state Y'.
5399 //
5400 //
5401 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
5402 // of each MergeMemNode array are exactly the numerical alias indexes, including
5403 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
5404 // Compile::alias_type (and kin) produce and manage these indexes.
5405 //
5406 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
5407 // (Note that this provides quick access to the top node inside MergeMem methods,
5408 // without the need to reach out via TLS to Compile::current.)
5409 //
5410 // As a consequence of what was just described, a MergeMem that represents a full
5411 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
5412 // containing all alias categories.
5413 //
5414 // MergeMem nodes never (?) have control inputs, so in(0) is null.
5415 //
5416 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
5417 // a memory state for the alias type <N>, or else the top node, meaning that
5418 // there is no particular input for that alias type. Note that the length of
5419 // a MergeMem is variable, and may be extended at any time to accommodate new
5420 // memory states at larger alias indexes. When merges grow, they are of course
5421 // filled with "top" in the unused in() positions.
5422 //
5423 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
5424 // (Top was chosen because it works smoothly with passes like GCM.)
5425 //
5426 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
5427 // the type of random VM bits like TLS references.) Since it is always the
5428 // first non-Bot memory slice, some low-level loops use it to initialize an
5429 // index variable: for (i = AliasIdxRaw; i < req(); i++).
5430 //
5431 //
5432 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
5433 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
5434 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
5435 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
5436 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
5437 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
5438 //
5439 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
5440 // really that different from the other memory inputs. An abbreviation called
5441 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
5442 //
5443 //
5444 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
5445 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
5446 // that "emerges though" the base memory will be marked as excluding the alias types
5447 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
5448 //
5449 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
5450 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
5451 //
5452 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
5453 // (It is currently unimplemented.) As you can see, the resulting merge is
5454 // actually a disjoint union of memory states, rather than an overlay.
5455 //
5456
5457 //------------------------------MergeMemNode-----------------------------------
5458 Node* MergeMemNode::make_empty_memory() {
5459 Node* empty_memory = (Node*) Compile::current()->top();
5460 assert(empty_memory->is_top(), "correct sentinel identity");
5461 return empty_memory;
5462 }
5463
5464 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
5465 init_class_id(Class_MergeMem);
5466 // all inputs are nullified in Node::Node(int)
5467 // set_input(0, nullptr); // no control input
5468
5469 // Initialize the edges uniformly to top, for starters.
5470 Node* empty_mem = make_empty_memory();
5471 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
5472 init_req(i,empty_mem);
5473 }
5474 assert(empty_memory() == empty_mem, "");
5475
5476 if( new_base != nullptr && new_base->is_MergeMem() ) {
5477 MergeMemNode* mdef = new_base->as_MergeMem();
5478 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
5479 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
5480 mms.set_memory(mms.memory2());
5481 }
5482 assert(base_memory() == mdef->base_memory(), "");
5483 } else {
5484 set_base_memory(new_base);
5485 }
5486 }
5487
5488 // Make a new, untransformed MergeMem with the same base as 'mem'.
5489 // If mem is itself a MergeMem, populate the result with the same edges.
5490 MergeMemNode* MergeMemNode::make(Node* mem) {
5491 return new MergeMemNode(mem);
5492 }
5493
5494 //------------------------------cmp--------------------------------------------
5495 uint MergeMemNode::hash() const { return NO_HASH; }
5496 bool MergeMemNode::cmp( const Node &n ) const {
5497 return (&n == this); // Always fail except on self
5498 }
5499
5500 //------------------------------Identity---------------------------------------
5501 Node* MergeMemNode::Identity(PhaseGVN* phase) {
5502 // Identity if this merge point does not record any interesting memory
5503 // disambiguations.
5504 Node* base_mem = base_memory();
5505 Node* empty_mem = empty_memory();
5506 if (base_mem != empty_mem) { // Memory path is not dead?
5507 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5508 Node* mem = in(i);
5509 if (mem != empty_mem && mem != base_mem) {
5510 return this; // Many memory splits; no change
5511 }
5512 }
5513 }
5514 return base_mem; // No memory splits; ID on the one true input
5515 }
5516
5517 //------------------------------Ideal------------------------------------------
5518 // This method is invoked recursively on chains of MergeMem nodes
5519 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
5520 // Remove chain'd MergeMems
5521 //
5522 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
5523 // relative to the "in(Bot)". Since we are patching both at the same time,
5524 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
5525 // but rewrite each "in(i)" relative to the new "in(Bot)".
5526 Node *progress = nullptr;
5527
5528
5529 Node* old_base = base_memory();
5530 Node* empty_mem = empty_memory();
5531 if (old_base == empty_mem)
5532 return nullptr; // Dead memory path.
5533
5534 MergeMemNode* old_mbase;
5535 if (old_base != nullptr && old_base->is_MergeMem())
5536 old_mbase = old_base->as_MergeMem();
5537 else
5538 old_mbase = nullptr;
5539 Node* new_base = old_base;
5540
5541 // simplify stacked MergeMems in base memory
5542 if (old_mbase) new_base = old_mbase->base_memory();
5543
5544 // the base memory might contribute new slices beyond my req()
5545 if (old_mbase) grow_to_match(old_mbase);
5546
5547 // Note: We do not call verify_sparse on entry, because inputs
5548 // can normalize to the base_memory via subsume_node or similar
5549 // mechanisms. This method repairs that damage.
5550
5551 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
5552
5553 // Look at each slice.
5554 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5555 Node* old_in = in(i);
5556 // calculate the old memory value
5557 Node* old_mem = old_in;
5558 if (old_mem == empty_mem) old_mem = old_base;
5559 assert(old_mem == memory_at(i), "");
5560
5561 // maybe update (reslice) the old memory value
5562
5563 // simplify stacked MergeMems
5564 Node* new_mem = old_mem;
5565 MergeMemNode* old_mmem;
5566 if (old_mem != nullptr && old_mem->is_MergeMem())
5567 old_mmem = old_mem->as_MergeMem();
5568 else
5569 old_mmem = nullptr;
5570 if (old_mmem == this) {
5571 // This can happen if loops break up and safepoints disappear.
5572 // A merge of BotPtr (default) with a RawPtr memory derived from a
5573 // safepoint can be rewritten to a merge of the same BotPtr with
5574 // the BotPtr phi coming into the loop. If that phi disappears
5575 // also, we can end up with a self-loop of the mergemem.
5576 // In general, if loops degenerate and memory effects disappear,
5577 // a mergemem can be left looking at itself. This simply means
5578 // that the mergemem's default should be used, since there is
5579 // no longer any apparent effect on this slice.
5580 // Note: If a memory slice is a MergeMem cycle, it is unreachable
5581 // from start. Update the input to TOP.
5582 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
5583 }
5584 else if (old_mmem != nullptr) {
5585 new_mem = old_mmem->memory_at(i);
5586 }
5587 // else preceding memory was not a MergeMem
5588
5589 // maybe store down a new value
5590 Node* new_in = new_mem;
5591 if (new_in == new_base) new_in = empty_mem;
5592
5593 if (new_in != old_in) {
5594 // Warning: Do not combine this "if" with the previous "if"
5595 // A memory slice might have be be rewritten even if it is semantically
5596 // unchanged, if the base_memory value has changed.
5597 set_req_X(i, new_in, phase);
5598 progress = this; // Report progress
5599 }
5600 }
5601
5602 if (new_base != old_base) {
5603 set_req_X(Compile::AliasIdxBot, new_base, phase);
5604 // Don't use set_base_memory(new_base), because we need to update du.
5605 assert(base_memory() == new_base, "");
5606 progress = this;
5607 }
5608
5609 if( base_memory() == this ) {
5610 // a self cycle indicates this memory path is dead
5611 set_req(Compile::AliasIdxBot, empty_mem);
5612 }
5613
5614 // Resolve external cycles by calling Ideal on a MergeMem base_memory
5615 // Recursion must occur after the self cycle check above
5616 if( base_memory()->is_MergeMem() ) {
5617 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
5618 Node *m = phase->transform(new_mbase); // Rollup any cycles
5619 if( m != nullptr &&
5620 (m->is_top() ||
5621 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
5622 // propagate rollup of dead cycle to self
5623 set_req(Compile::AliasIdxBot, empty_mem);
5624 }
5625 }
5626
5627 if( base_memory() == empty_mem ) {
5628 progress = this;
5629 // Cut inputs during Parse phase only.
5630 // During Optimize phase a dead MergeMem node will be subsumed by Top.
5631 if( !can_reshape ) {
5632 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5633 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
5634 }
5635 }
5636 }
5637
5638 if( !progress && base_memory()->is_Phi() && can_reshape ) {
5639 // Check if PhiNode::Ideal's "Split phis through memory merges"
5640 // transform should be attempted. Look for this->phi->this cycle.
5641 uint merge_width = req();
5642 if (merge_width > Compile::AliasIdxRaw) {
5643 PhiNode* phi = base_memory()->as_Phi();
5644 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
5645 if (phi->in(i) == this) {
5646 phase->is_IterGVN()->_worklist.push(phi);
5647 break;
5648 }
5649 }
5650 }
5651 }
5652
5653 assert(progress || verify_sparse(), "please, no dups of base");
5654 return progress;
5655 }
5656
5657 //-------------------------set_base_memory-------------------------------------
5658 void MergeMemNode::set_base_memory(Node *new_base) {
5659 Node* empty_mem = empty_memory();
5660 set_req(Compile::AliasIdxBot, new_base);
5661 assert(memory_at(req()) == new_base, "must set default memory");
5662 // Clear out other occurrences of new_base:
5663 if (new_base != empty_mem) {
5664 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5665 if (in(i) == new_base) set_req(i, empty_mem);
5666 }
5667 }
5668 }
5669
5670 //------------------------------out_RegMask------------------------------------
5671 const RegMask &MergeMemNode::out_RegMask() const {
5672 return RegMask::Empty;
5673 }
5674
5675 //------------------------------dump_spec--------------------------------------
5676 #ifndef PRODUCT
5677 void MergeMemNode::dump_spec(outputStream *st) const {
5678 st->print(" {");
5679 Node* base_mem = base_memory();
5680 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
5681 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem;
5682 if (mem == base_mem) { st->print(" -"); continue; }
5683 st->print( " N%d:", mem->_idx );
5684 Compile::current()->get_adr_type(i)->dump_on(st);
5685 }
5686 st->print(" }");
5687 }
5688 #endif // !PRODUCT
5689
5690
5691 #ifdef ASSERT
5692 static bool might_be_same(Node* a, Node* b) {
5693 if (a == b) return true;
5694 if (!(a->is_Phi() || b->is_Phi())) return false;
5695 // phis shift around during optimization
5696 return true; // pretty stupid...
5697 }
5698
5699 // verify a narrow slice (either incoming or outgoing)
5700 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
5701 if (!VerifyAliases) return; // don't bother to verify unless requested
5702 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
5703 if (Node::in_dump()) return; // muzzle asserts when printing
5704 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
5705 assert(n != nullptr, "");
5706 // Elide intervening MergeMem's
5707 while (n->is_MergeMem()) {
5708 n = n->as_MergeMem()->memory_at(alias_idx);
5709 }
5710 Compile* C = Compile::current();
5711 const TypePtr* n_adr_type = n->adr_type();
5712 if (n == m->empty_memory()) {
5713 // Implicit copy of base_memory()
5714 } else if (n_adr_type != TypePtr::BOTTOM) {
5715 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type");
5716 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
5717 } else {
5718 // A few places like make_runtime_call "know" that VM calls are narrow,
5719 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
5720 bool expected_wide_mem = false;
5721 if (n == m->base_memory()) {
5722 expected_wide_mem = true;
5723 } else if (alias_idx == Compile::AliasIdxRaw ||
5724 n == m->memory_at(Compile::AliasIdxRaw)) {
5725 expected_wide_mem = true;
5726 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
5727 // memory can "leak through" calls on channels that
5728 // are write-once. Allow this also.
5729 expected_wide_mem = true;
5730 }
5731 assert(expected_wide_mem, "expected narrow slice replacement");
5732 }
5733 }
5734 #else // !ASSERT
5735 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
5736 #endif
5737
5738
5739 //-----------------------------memory_at---------------------------------------
5740 Node* MergeMemNode::memory_at(uint alias_idx) const {
5741 assert(alias_idx >= Compile::AliasIdxRaw ||
5742 (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()),
5743 "must avoid base_memory and AliasIdxTop");
5744
5745 // Otherwise, it is a narrow slice.
5746 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
5747 if (is_empty_memory(n)) {
5748 // the array is sparse; empty slots are the "top" node
5749 n = base_memory();
5750 assert(Node::in_dump()
5751 || n == nullptr || n->bottom_type() == Type::TOP
5752 || n->adr_type() == nullptr // address is TOP
5753 || n->adr_type() == TypePtr::BOTTOM
5754 || n->adr_type() == TypeRawPtr::BOTTOM
5755 || !Compile::current()->do_aliasing(),
5756 "must be a wide memory");
5757 // do_aliasing == false if we are organizing the memory states manually.
5758 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
5759 } else {
5760 // make sure the stored slice is sane
5761 #ifdef ASSERT
5762 if (VMError::is_error_reported() || Node::in_dump()) {
5763 } else if (might_be_same(n, base_memory())) {
5764 // Give it a pass: It is a mostly harmless repetition of the base.
5765 // This can arise normally from node subsumption during optimization.
5766 } else {
5767 verify_memory_slice(this, alias_idx, n);
5768 }
5769 #endif
5770 }
5771 return n;
5772 }
5773
5774 //---------------------------set_memory_at-------------------------------------
5775 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
5776 verify_memory_slice(this, alias_idx, n);
5777 Node* empty_mem = empty_memory();
5778 if (n == base_memory()) n = empty_mem; // collapse default
5779 uint need_req = alias_idx+1;
5780 if (req() < need_req) {
5781 if (n == empty_mem) return; // already the default, so do not grow me
5782 // grow the sparse array
5783 do {
5784 add_req(empty_mem);
5785 } while (req() < need_req);
5786 }
5787 set_req( alias_idx, n );
5788 }
5789
5790
5791
5792 //--------------------------iteration_setup------------------------------------
5793 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
5794 if (other != nullptr) {
5795 grow_to_match(other);
5796 // invariant: the finite support of mm2 is within mm->req()
5797 #ifdef ASSERT
5798 for (uint i = req(); i < other->req(); i++) {
5799 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
5800 }
5801 #endif
5802 }
5803 // Replace spurious copies of base_memory by top.
5804 Node* base_mem = base_memory();
5805 if (base_mem != nullptr && !base_mem->is_top()) {
5806 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
5807 if (in(i) == base_mem)
5808 set_req(i, empty_memory());
5809 }
5810 }
5811 }
5812
5813 //---------------------------grow_to_match-------------------------------------
5814 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
5815 Node* empty_mem = empty_memory();
5816 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
5817 // look for the finite support of the other memory
5818 for (uint i = other->req(); --i >= req(); ) {
5819 if (other->in(i) != empty_mem) {
5820 uint new_len = i+1;
5821 while (req() < new_len) add_req(empty_mem);
5822 break;
5823 }
5824 }
5825 }
5826
5827 //---------------------------verify_sparse-------------------------------------
5828 #ifndef PRODUCT
5829 bool MergeMemNode::verify_sparse() const {
5830 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
5831 Node* base_mem = base_memory();
5832 // The following can happen in degenerate cases, since empty==top.
5833 if (is_empty_memory(base_mem)) return true;
5834 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5835 assert(in(i) != nullptr, "sane slice");
5836 if (in(i) == base_mem) return false; // should have been the sentinel value!
5837 }
5838 return true;
5839 }
5840
5841 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
5842 Node* n;
5843 n = mm->in(idx);
5844 if (mem == n) return true; // might be empty_memory()
5845 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
5846 if (mem == n) return true;
5847 return false;
5848 }
5849 #endif // !PRODUCT