1 /*
2 * Copyright (c) 1997, 2025, 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 "ci/ciFlatArrayKlass.hpp"
27 #include "classfile/javaClasses.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/c2/barrierSetC2.hpp"
32 #include "gc/shared/tlab_globals.hpp"
33 #include "memory/allocation.inline.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "oops/objArrayKlass.hpp"
36 #include "opto/addnode.hpp"
37 #include "opto/arraycopynode.hpp"
38 #include "opto/cfgnode.hpp"
39 #include "opto/regalloc.hpp"
40 #include "opto/compile.hpp"
41 #include "opto/connode.hpp"
42 #include "opto/convertnode.hpp"
43 #include "opto/inlinetypenode.hpp"
44 #include "opto/loopnode.hpp"
45 #include "opto/machnode.hpp"
46 #include "opto/matcher.hpp"
47 #include "opto/memnode.hpp"
48 #include "opto/mempointer.hpp"
49 #include "opto/mulnode.hpp"
50 #include "opto/narrowptrnode.hpp"
51 #include "opto/phaseX.hpp"
52 #include "opto/regmask.hpp"
53 #include "opto/rootnode.hpp"
54 #include "opto/traceMergeStoresTag.hpp"
55 #include "opto/vectornode.hpp"
56 #include "utilities/align.hpp"
57 #include "utilities/copy.hpp"
58 #include "utilities/macros.hpp"
59 #include "utilities/powerOfTwo.hpp"
60 #include "utilities/vmError.hpp"
61
62 // Portions of code courtesy of Clifford Click
63
64 // Optimization - Graph Style
65
66 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
67
68 //=============================================================================
69 uint MemNode::size_of() const { return sizeof(*this); }
70
71 const TypePtr *MemNode::adr_type() const {
72 Node* adr = in(Address);
73 if (adr == nullptr) return nullptr; // node is dead
74 const TypePtr* cross_check = nullptr;
75 DEBUG_ONLY(cross_check = _adr_type);
76 return calculate_adr_type(adr->bottom_type(), cross_check);
77 }
78
79 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
80 if (adr != nullptr) {
81 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
82 return true;
83 }
84 }
85 return false;
86 }
87
88 #ifndef PRODUCT
89 void MemNode::dump_spec(outputStream *st) const {
90 if (in(Address) == nullptr) return; // node is dead
91 #ifndef ASSERT
92 // fake the missing field
93 const TypePtr* _adr_type = nullptr;
94 if (in(Address) != nullptr)
95 _adr_type = in(Address)->bottom_type()->isa_ptr();
96 #endif
97 dump_adr_type(this, _adr_type, st);
98
99 Compile* C = Compile::current();
100 if (C->alias_type(_adr_type)->is_volatile()) {
101 st->print(" Volatile!");
102 }
103 if (_unaligned_access) {
104 st->print(" unaligned");
105 }
106 if (_mismatched_access) {
107 st->print(" mismatched");
108 }
109 if (_unsafe_access) {
110 st->print(" unsafe");
111 }
112 }
113
114 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
115 st->print(" @");
116 if (adr_type == nullptr) {
117 st->print("null");
118 } else {
119 adr_type->dump_on(st);
120 Compile* C = Compile::current();
121 Compile::AliasType* atp = nullptr;
122 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
123 if (atp == nullptr)
124 st->print(", idx=?\?;");
125 else if (atp->index() == Compile::AliasIdxBot)
126 st->print(", idx=Bot;");
127 else if (atp->index() == Compile::AliasIdxTop)
128 st->print(", idx=Top;");
129 else if (atp->index() == Compile::AliasIdxRaw)
130 st->print(", idx=Raw;");
131 else {
132 ciField* field = atp->field();
133 if (field) {
134 st->print(", name=");
135 field->print_name_on(st);
136 }
137 st->print(", idx=%d;", atp->index());
138 }
139 }
140 }
141
142 extern void print_alias_types();
143
144 #endif
145
146 // Find the memory output corresponding to the fall-through path of a call
147 static Node* find_call_fallthrough_mem_output(CallNode* call) {
148 ResourceMark rm;
149 CallProjections* projs = call->extract_projections(false, false);
150 Node* res = projs->fallthrough_memproj;
151 assert(res != nullptr, "must have a fallthrough mem output");
152 return res;
153 }
154
155 // Try to find a better memory input for a load from a strict final field
156 static Node* try_optimize_strict_final_load_memory(PhaseGVN* phase, Node* adr, ProjNode*& base_local) {
157 intptr_t offset = 0;
158 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
159 if (base == nullptr) {
160 return nullptr;
161 }
162
163 Node* base_uncasted = base->uncast();
164 if (base_uncasted->is_Proj()) {
165 MultiNode* multi = base_uncasted->in(0)->as_Multi();
166 if (multi->is_Allocate()) {
167 base_local = base_uncasted->as_Proj();
168 return nullptr;
169 } else if (multi->is_Call()) {
170 // The oop is returned from a call, the memory can be the fallthrough output of the call
171 return find_call_fallthrough_mem_output(multi->as_Call());
172 } else if (multi->is_Start()) {
173 // The oop is a parameter
174 if (phase->C->method()->is_object_constructor() && base_uncasted->as_Proj()->_con == TypeFunc::Parms) {
175 // The receiver of a constructor is similar to the result of an AllocateNode
176 base_local = base_uncasted->as_Proj();
177 return nullptr;
178 } else {
179 // Use the start memory otherwise
180 return multi->proj_out(TypeFunc::Memory);
181 }
182 }
183 }
184
185 return nullptr;
186 }
187
188 // Whether a call can modify a strict final field, given that the object is allocated inside the
189 // current compilation unit, or is the first parameter when the compilation root is a constructor.
190 // This is equivalent to asking whether 'call' is a constructor invocation and the class declaring
191 // the target method is a subclass of the class declaring 'field'.
192 static bool call_can_modify_local_object(ciField* field, CallNode* call) {
193 if (!call->is_CallJava()) {
194 return false;
195 }
196
197 ciMethod* target = call->as_CallJava()->method();
198 if (target == nullptr || !target->is_object_constructor()) {
199 return false;
200 }
201
202 // If 'field' is declared in a class that is a subclass of the one declaring the constructor,
203 // then the field is set inside the constructor, else the field must be set before the
204 // constructor invocation. E.g. A field Super.x will be set during the execution of Sub::<init>,
205 // while a field Sub.y must be set before Super::<init> is invoked.
206 // We can try to be more heroic and decide if the receiver of the constructor invocation is the
207 // object from which we are loading from. This, however, may be problematic as deciding if 2
208 // nodes are definitely different may not be trivial, especially if the graph is not canonical.
209 // As a result, it is made more conservative for now.
210 assert(call->req() > TypeFunc::Parms, "constructor must have at least 1 argument");
211 return target->holder()->is_subclass_of(field->holder());
212 }
213
214 Node* MemNode::optimize_simple_memory_chain(Node* mchain, const TypeOopPtr* t_oop, Node* load, PhaseGVN* phase) {
215 assert(t_oop != nullptr, "sanity");
216 bool is_instance = t_oop->is_known_instance_field();
217
218 ciField* field = phase->C->alias_type(t_oop)->field();
219 bool is_strict_final_load = false;
220
221 // After macro expansion, an allocation may become a call, changing the memory input to the
222 // memory output of that call would be illegal. As a result, disallow this transformation after
223 // macro expansion.
224 if (phase->is_IterGVN() && phase->C->allow_macro_nodes() && load != nullptr && load->is_Load() && !load->as_Load()->is_mismatched_access()) {
225 if (EnableValhalla) {
226 if (field != nullptr && (field->holder()->is_inlinetype() || field->holder()->is_abstract_value_klass())) {
227 is_strict_final_load = true;
228 }
229 #ifdef ASSERT
230 if (t_oop->is_inlinetypeptr() && t_oop->inline_klass()->contains_field_offset(t_oop->offset())) {
231 assert(is_strict_final_load, "sanity check for basic cases");
232 }
233 #endif
234 } else {
235 is_strict_final_load = field != nullptr && t_oop->is_ptr_to_boxed_value();
236 }
237 }
238
239 if (!is_instance && !is_strict_final_load) {
240 return mchain;
241 }
242
243 Node* result = mchain;
244 ProjNode* base_local = nullptr;
245
246 if (is_strict_final_load) {
247 Node* adr = load->in(MemNode::Address);
248 assert(phase->type(adr) == t_oop, "inconsistent type");
249 Node* tmp = try_optimize_strict_final_load_memory(phase, adr, base_local);
250 if (tmp != nullptr) {
251 result = tmp;
252 }
253 }
254
255 uint instance_id = t_oop->instance_id();
256 Node* start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
257 Node* prev = nullptr;
258 while (prev != result) {
259 prev = result;
260 if (result == start_mem) {
261 // start_mem is the earliest memory possible
262 break;
263 }
264
265 // skip over a call which does not affect this memory slice
266 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
267 Node* proj_in = result->in(0);
268 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
269 // This is the allocation that creates the object from which we are loading from
270 break;
271 } else if (proj_in->is_Call()) {
272 // ArrayCopyNodes processed here as well
273 CallNode* call = proj_in->as_Call();
274 if (!call->may_modify(t_oop, phase)) {
275 result = call->in(TypeFunc::Memory);
276 } else if (is_strict_final_load && base_local != nullptr && !call_can_modify_local_object(field, call)) {
277 result = call->in(TypeFunc::Memory);
278 }
279 } else if (proj_in->is_Initialize()) {
280 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
281 // Stop if this is the initialization for the object instance which
282 // contains this memory slice, otherwise skip over it.
283 if ((alloc == nullptr) || (alloc->_idx == instance_id)) {
284 break;
285 }
286 if (is_instance) {
287 result = proj_in->in(TypeFunc::Memory);
288 } else if (is_strict_final_load) {
289 Node* klass = alloc->in(AllocateNode::KlassNode);
290 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
291 if (tklass->klass_is_exact() && !tklass->exact_klass()->equals(t_oop->is_instptr()->exact_klass())) {
292 // Allocation of another type, must be another object
293 result = proj_in->in(TypeFunc::Memory);
294 } else if (base_local != nullptr && (base_local->is_Parm() || base_local->in(0) != alloc)) {
295 // Allocation of another object
296 result = proj_in->in(TypeFunc::Memory);
297 }
298 }
299 } else if (proj_in->is_MemBar()) {
300 ArrayCopyNode* ac = nullptr;
301 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
302 break;
303 }
304 result = proj_in->in(TypeFunc::Memory);
305 } else if (proj_in->is_top()) {
306 break; // dead code
307 } else {
308 assert(false, "unexpected projection");
309 }
310 } else if (result->is_ClearArray()) {
311 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
312 // Can not bypass initialization of the instance
313 // we are looking for.
314 break;
315 }
316 // Otherwise skip it (the call updated 'result' value).
317 } else if (result->is_MergeMem()) {
318 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, nullptr, tty);
319 }
320 }
321 return result;
322 }
323
324 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
325 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
326 if (t_oop == nullptr)
327 return mchain; // don't try to optimize non-oop types
328 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
329 bool is_instance = t_oop->is_known_instance_field();
330 PhaseIterGVN *igvn = phase->is_IterGVN();
331 if (is_instance && igvn != nullptr && result->is_Phi()) {
332 PhiNode *mphi = result->as_Phi();
333 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
334 const TypePtr *t = mphi->adr_type();
335 bool do_split = false;
336 // In the following cases, Load memory input can be further optimized based on
337 // its precise address type
338 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ) {
339 do_split = true;
340 } else if (t->isa_oopptr() && !t->is_oopptr()->is_known_instance()) {
341 const TypeOopPtr* mem_t =
342 t->is_oopptr()->cast_to_exactness(true)
343 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
344 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id());
345 if (t_oop->isa_aryptr()) {
346 mem_t = mem_t->is_aryptr()
347 ->cast_to_stable(t_oop->is_aryptr()->is_stable())
348 ->cast_to_size(t_oop->is_aryptr()->size())
349 ->cast_to_not_flat(t_oop->is_aryptr()->is_not_flat())
350 ->cast_to_not_null_free(t_oop->is_aryptr()->is_not_null_free())
351 ->with_offset(t_oop->is_aryptr()->offset())
352 ->is_aryptr();
353 }
354 do_split = mem_t == t_oop;
355 }
356 if (do_split) {
357 // clone the Phi with our address type
358 result = mphi->split_out_instance(t_adr, igvn);
359 } else {
360 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
361 }
362 }
363 return result;
364 }
365
366 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
367 uint alias_idx = phase->C->get_alias_index(tp);
368 Node *mem = mmem;
369 #ifdef ASSERT
370 {
371 // Check that current type is consistent with the alias index used during graph construction
372 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
373 bool consistent = adr_check == nullptr || adr_check->empty() ||
374 phase->C->must_alias(adr_check, alias_idx );
375 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
376 if( !consistent && adr_check != nullptr && !adr_check->empty() &&
377 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
378 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
379 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
380 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
381 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
382 // don't assert if it is dead code.
383 consistent = true;
384 }
385 if( !consistent ) {
386 st->print("alias_idx==%d, adr_check==", alias_idx);
387 if( adr_check == nullptr ) {
388 st->print("null");
389 } else {
390 adr_check->dump();
391 }
392 st->cr();
393 print_alias_types();
394 assert(consistent, "adr_check must match alias idx");
395 }
396 }
397 #endif
398 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
399 // means an array I have not precisely typed yet. Do not do any
400 // alias stuff with it any time soon.
401 const TypeOopPtr *toop = tp->isa_oopptr();
402 if (tp->base() != Type::AnyPtr &&
403 !(toop &&
404 toop->isa_instptr() &&
405 toop->is_instptr()->instance_klass()->is_java_lang_Object() &&
406 toop->offset() == Type::OffsetBot)) {
407 // IGVN _delay_transform may be set to true and if that is the case and mmem
408 // is already a registered node then the validation inside transform will
409 // complain.
410 Node* m = mmem;
411 PhaseIterGVN* igvn = phase->is_IterGVN();
412 if (igvn == nullptr || !igvn->delay_transform()) {
413 // compress paths and change unreachable cycles to TOP
414 // If not, we can update the input infinitely along a MergeMem cycle
415 // Equivalent code in PhiNode::Ideal
416 m = phase->transform(mmem);
417 }
418 // If transformed to a MergeMem, get the desired slice
419 // Otherwise the returned node represents memory for every slice
420 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
421 // Update input if it is progress over what we have now
422 }
423 return mem;
424 }
425
426 //--------------------------Ideal_common---------------------------------------
427 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
428 // Unhook non-raw memories from complete (macro-expanded) initializations.
429 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
430 // If our control input is a dead region, kill all below the region
431 Node *ctl = in(MemNode::Control);
432 if (ctl && remove_dead_region(phase, can_reshape))
433 return this;
434 ctl = in(MemNode::Control);
435 // Don't bother trying to transform a dead node
436 if (ctl && ctl->is_top()) return NodeSentinel;
437
438 PhaseIterGVN *igvn = phase->is_IterGVN();
439 // Wait if control on the worklist.
440 if (ctl && can_reshape && igvn != nullptr) {
441 Node* bol = nullptr;
442 Node* cmp = nullptr;
443 if (ctl->in(0)->is_If()) {
444 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
445 bol = ctl->in(0)->in(1);
446 if (bol->is_Bool())
447 cmp = ctl->in(0)->in(1)->in(1);
448 }
449 if (igvn->_worklist.member(ctl) ||
450 (bol != nullptr && igvn->_worklist.member(bol)) ||
451 (cmp != nullptr && igvn->_worklist.member(cmp)) ) {
452 // This control path may be dead.
453 // Delay this memory node transformation until the control is processed.
454 igvn->_worklist.push(this);
455 return NodeSentinel; // caller will return null
456 }
457 }
458 // Ignore if memory is dead, or self-loop
459 Node *mem = in(MemNode::Memory);
460 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return null
461 assert(mem != this, "dead loop in MemNode::Ideal");
462
463 if (can_reshape && igvn != nullptr && igvn->_worklist.member(mem)) {
464 // This memory slice may be dead.
465 // Delay this mem node transformation until the memory is processed.
466 igvn->_worklist.push(this);
467 return NodeSentinel; // caller will return null
468 }
469
470 Node *address = in(MemNode::Address);
471 const Type *t_adr = phase->type(address);
472 if (t_adr == Type::TOP) return NodeSentinel; // caller will return null
473
474 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
475 // Unsafe off-heap access with zero address. Remove access and other control users
476 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
477 assert(ctl != nullptr, "unsafe accesses should be control dependent");
478 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
479 Node* u = ctl->fast_out(i);
480 if (u != ctl) {
481 igvn->rehash_node_delayed(u);
482 int nb = u->replace_edge(ctl, phase->C->top(), igvn);
483 --i, imax -= nb;
484 }
485 }
486 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
487 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
488 phase->C->root()->add_req(halt);
489 return this;
490 }
491
492 if (can_reshape && igvn != nullptr &&
493 (igvn->_worklist.member(address) ||
494 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
495 // The address's base and type may change when the address is processed.
496 // Delay this mem node transformation until the address is processed.
497 igvn->_worklist.push(this);
498 return NodeSentinel; // caller will return null
499 }
500
501 // Do NOT remove or optimize the next lines: ensure a new alias index
502 // is allocated for an oop pointer type before Escape Analysis.
503 // Note: C++ will not remove it since the call has side effect.
504 if (t_adr->isa_oopptr()) {
505 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
506 }
507
508 Node* base = nullptr;
509 if (address->is_AddP()) {
510 base = address->in(AddPNode::Base);
511 }
512 if (base != nullptr && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
513 !t_adr->isa_rawptr()) {
514 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
515 // Skip this node optimization if its address has TOP base.
516 return NodeSentinel; // caller will return null
517 }
518
519 // Avoid independent memory operations
520 Node* old_mem = mem;
521
522 // The code which unhooks non-raw memories from complete (macro-expanded)
523 // initializations was removed. After macro-expansion all stores caught
524 // by Initialize node became raw stores and there is no information
525 // which memory slices they modify. So it is unsafe to move any memory
526 // operation above these stores. Also in most cases hooked non-raw memories
527 // were already unhooked by using information from detect_ptr_independence()
528 // and find_previous_store().
529
530 if (mem->is_MergeMem()) {
531 MergeMemNode* mmem = mem->as_MergeMem();
532 const TypePtr *tp = t_adr->is_ptr();
533
534 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
535 }
536
537 if (mem != old_mem) {
538 set_req_X(MemNode::Memory, mem, phase);
539 if (phase->type(mem) == Type::TOP) return NodeSentinel;
540 return this;
541 }
542
543 // let the subclass continue analyzing...
544 return nullptr;
545 }
546
547 // Helper function for proving some simple control dominations.
548 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
549 // Already assumes that 'dom' is available at 'sub', and that 'sub'
550 // is not a constant (dominated by the method's StartNode).
551 // Used by MemNode::find_previous_store to prove that the
552 // control input of a memory operation predates (dominates)
553 // an allocation it wants to look past.
554 // Returns 'DomResult::Dominate' if all control inputs of 'dom'
555 // dominate 'sub', 'DomResult::NotDominate' if not,
556 // and 'DomResult::EncounteredDeadCode' if we can't decide due to
557 // dead code, but at the end of IGVN, we know the definite result
558 // once the dead code is cleaned up.
559 Node::DomResult MemNode::maybe_all_controls_dominate(Node* dom, Node* sub) {
560 if (dom == nullptr || dom->is_top() || sub == nullptr || sub->is_top()) {
561 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
562 }
563
564 // Check 'dom'. Skip Proj and CatchProj nodes.
565 dom = dom->find_exact_control(dom);
566 if (dom == nullptr || dom->is_top()) {
567 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
568 }
569
570 if (dom == sub) {
571 // For the case when, for example, 'sub' is Initialize and the original
572 // 'dom' is Proj node of the 'sub'.
573 return DomResult::NotDominate;
574 }
575
576 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub) {
577 return DomResult::Dominate;
578 }
579
580 // 'dom' dominates 'sub' if its control edge and control edges
581 // of all its inputs dominate or equal to sub's control edge.
582
583 // Currently 'sub' is either Allocate, Initialize or Start nodes.
584 // Or Region for the check in LoadNode::Ideal();
585 // 'sub' should have sub->in(0) != nullptr.
586 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
587 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
588
589 // Get control edge of 'sub'.
590 Node* orig_sub = sub;
591 sub = sub->find_exact_control(sub->in(0));
592 if (sub == nullptr || sub->is_top()) {
593 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
594 }
595
596 assert(sub->is_CFG(), "expecting control");
597
598 if (sub == dom) {
599 return DomResult::Dominate;
600 }
601
602 if (sub->is_Start() || sub->is_Root()) {
603 return DomResult::NotDominate;
604 }
605
606 {
607 // Check all control edges of 'dom'.
608
609 ResourceMark rm;
610 Node_List nlist;
611 Unique_Node_List dom_list;
612
613 dom_list.push(dom);
614 bool only_dominating_controls = false;
615
616 for (uint next = 0; next < dom_list.size(); next++) {
617 Node* n = dom_list.at(next);
618 if (n == orig_sub) {
619 return DomResult::NotDominate; // One of dom's inputs dominated by sub.
620 }
621 if (!n->is_CFG() && n->pinned()) {
622 // Check only own control edge for pinned non-control nodes.
623 n = n->find_exact_control(n->in(0));
624 if (n == nullptr || n->is_top()) {
625 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
626 }
627 assert(n->is_CFG(), "expecting control");
628 dom_list.push(n);
629 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
630 only_dominating_controls = true;
631 } else if (n->is_CFG()) {
632 DomResult dom_result = n->dominates(sub, nlist);
633 if (dom_result == DomResult::Dominate) {
634 only_dominating_controls = true;
635 } else {
636 return dom_result;
637 }
638 } else {
639 // First, own control edge.
640 Node* m = n->find_exact_control(n->in(0));
641 if (m != nullptr) {
642 if (m->is_top()) {
643 return DomResult::EncounteredDeadCode; // Conservative answer for dead code
644 }
645 dom_list.push(m);
646 }
647 // Now, the rest of edges.
648 uint cnt = n->req();
649 for (uint i = 1; i < cnt; i++) {
650 m = n->find_exact_control(n->in(i));
651 if (m == nullptr || m->is_top()) {
652 continue;
653 }
654 dom_list.push(m);
655 }
656 }
657 }
658 return only_dominating_controls ? DomResult::Dominate : DomResult::NotDominate;
659 }
660 }
661
662 //---------------------detect_ptr_independence---------------------------------
663 // Used by MemNode::find_previous_store to prove that two base
664 // pointers are never equal.
665 // The pointers are accompanied by their associated allocations,
666 // if any, which have been previously discovered by the caller.
667 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
668 Node* p2, AllocateNode* a2,
669 PhaseTransform* phase) {
670 // Attempt to prove that these two pointers cannot be aliased.
671 // They may both manifestly be allocations, and they should differ.
672 // Or, if they are not both allocations, they can be distinct constants.
673 // Otherwise, one is an allocation and the other a pre-existing value.
674 if (a1 == nullptr && a2 == nullptr) { // neither an allocation
675 return (p1 != p2) && p1->is_Con() && p2->is_Con();
676 } else if (a1 != nullptr && a2 != nullptr) { // both allocations
677 return (a1 != a2);
678 } else if (a1 != nullptr) { // one allocation a1
679 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
680 return all_controls_dominate(p2, a1);
681 } else { //(a2 != null) // one allocation a2
682 return all_controls_dominate(p1, a2);
683 }
684 return false;
685 }
686
687
688 // Find an arraycopy ac that produces the memory state represented by parameter mem.
689 // Return ac if
690 // (a) can_see_stored_value=true and ac must have set the value for this load or if
691 // (b) can_see_stored_value=false and ac could have set the value for this load or if
692 // (c) can_see_stored_value=false and ac cannot have set the value for this load.
693 // In case (c) change the parameter mem to the memory input of ac to skip it
694 // when searching stored value.
695 // Otherwise return null.
696 Node* LoadNode::find_previous_arraycopy(PhaseValues* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
697 ArrayCopyNode* ac = find_array_copy_clone(ld_alloc, mem);
698 if (ac != nullptr) {
699 Node* ld_addp = in(MemNode::Address);
700 Node* src = ac->in(ArrayCopyNode::Src);
701 const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
702
703 // This is a load from a cloned array. The corresponding arraycopy ac must
704 // have set the value for the load and we can return ac but only if the load
705 // is known to be within bounds. This is checked below.
706 if (ary_t != nullptr && ld_addp->is_AddP()) {
707 Node* ld_offs = ld_addp->in(AddPNode::Offset);
708 BasicType ary_elem = ary_t->elem()->array_element_basic_type();
709 jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
710 jlong elemsize = type2aelembytes(ary_elem);
711
712 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
713 const TypeInt* sizetype = ary_t->size();
714
715 if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
716 // The load is known to be within bounds. It receives its value from ac.
717 return ac;
718 }
719 // The load is known to be out-of-bounds.
720 }
721 // The load could be out-of-bounds. It must not be hoisted but must remain
722 // dependent on the runtime range check. This is achieved by returning null.
723 } else if (mem->is_Proj() && mem->in(0) != nullptr && mem->in(0)->is_ArrayCopy()) {
724 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
725
726 if (ac->is_arraycopy_validated() ||
727 ac->is_copyof_validated() ||
728 ac->is_copyofrange_validated()) {
729 Node* ld_addp = in(MemNode::Address);
730 if (ld_addp->is_AddP()) {
731 Node* ld_base = ld_addp->in(AddPNode::Address);
732 Node* ld_offs = ld_addp->in(AddPNode::Offset);
733
734 Node* dest = ac->in(ArrayCopyNode::Dest);
735
736 if (dest == ld_base) {
737 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
738 assert(!ld_offs_t->empty(), "dead reference should be checked already");
739 // Take into account vector or unsafe access size
740 jlong ld_size_in_bytes = (jlong)memory_size();
741 jlong offset_hi = ld_offs_t->_hi + ld_size_in_bytes - 1;
742 offset_hi = MIN2(offset_hi, (jlong)(TypeX::MAX->_hi)); // Take care for overflow in 32-bit VM
743 if (ac->modifies(ld_offs_t->_lo, (intptr_t)offset_hi, phase, can_see_stored_value)) {
744 return ac;
745 }
746 if (!can_see_stored_value) {
747 mem = ac->in(TypeFunc::Memory);
748 return ac;
749 }
750 }
751 }
752 }
753 }
754 return nullptr;
755 }
756
757 ArrayCopyNode* MemNode::find_array_copy_clone(Node* ld_alloc, Node* mem) const {
758 if (mem->is_Proj() && mem->in(0) != nullptr && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
759 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
760 if (ld_alloc != nullptr) {
761 // Check if there is an array copy for a clone
762 Node* mb = mem->in(0);
763 ArrayCopyNode* ac = nullptr;
764 if (mb->in(0) != nullptr && mb->in(0)->is_Proj() &&
765 mb->in(0)->in(0) != nullptr && mb->in(0)->in(0)->is_ArrayCopy()) {
766 ac = mb->in(0)->in(0)->as_ArrayCopy();
767 } else {
768 // Step over GC barrier when ReduceInitialCardMarks is disabled
769 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
770 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
771
772 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
773 ac = control_proj_ac->in(0)->as_ArrayCopy();
774 }
775 }
776
777 if (ac != nullptr && ac->is_clonebasic()) {
778 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest));
779 if (alloc != nullptr && alloc == ld_alloc) {
780 return ac;
781 }
782 }
783 }
784 }
785 return nullptr;
786 }
787
788 // The logic for reordering loads and stores uses four steps:
789 // (a) Walk carefully past stores and initializations which we
790 // can prove are independent of this load.
791 // (b) Observe that the next memory state makes an exact match
792 // with self (load or store), and locate the relevant store.
793 // (c) Ensure that, if we were to wire self directly to the store,
794 // the optimizer would fold it up somehow.
795 // (d) Do the rewiring, and return, depending on some other part of
796 // the optimizer to fold up the load.
797 // This routine handles steps (a) and (b). Steps (c) and (d) are
798 // specific to loads and stores, so they are handled by the callers.
799 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
800 //
801 Node* MemNode::find_previous_store(PhaseValues* phase) {
802 Node* ctrl = in(MemNode::Control);
803 Node* adr = in(MemNode::Address);
804 intptr_t offset = 0;
805 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
806 AllocateNode* alloc = AllocateNode::Ideal_allocation(base);
807
808 if (offset == Type::OffsetBot)
809 return nullptr; // cannot unalias unless there are precise offsets
810
811 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
812 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
813
814 intptr_t size_in_bytes = memory_size();
815
816 Node* mem = in(MemNode::Memory); // start searching here...
817
818 int cnt = 50; // Cycle limiter
819 for (;;) { // While we can dance past unrelated stores...
820 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
821
822 Node* prev = mem;
823 if (mem->is_Store()) {
824 Node* st_adr = mem->in(MemNode::Address);
825 intptr_t st_offset = 0;
826 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
827 if (st_base == nullptr)
828 break; // inscrutable pointer
829
830 // For raw accesses it's not enough to prove that constant offsets don't intersect.
831 // We need the bases to be the equal in order for the offset check to make sense.
832 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
833 break;
834 }
835
836 if (st_offset != offset && st_offset != Type::OffsetBot) {
837 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
838 assert(mem->as_Store()->memory_size() <= MAX_STORE, "");
839 if (st_offset >= offset + size_in_bytes ||
840 st_offset <= offset - MAX_STORE ||
841 st_offset <= offset - mem->as_Store()->memory_size()) {
842 // Success: The offsets are provably independent.
843 // (You may ask, why not just test st_offset != offset and be done?
844 // The answer is that stores of different sizes can co-exist
845 // in the same sequence of RawMem effects. We sometimes initialize
846 // a whole 'tile' of array elements with a single jint or jlong.)
847 mem = mem->in(MemNode::Memory);
848 continue; // (a) advance through independent store memory
849 }
850 }
851 if (st_base != base &&
852 detect_ptr_independence(base, alloc,
853 st_base,
854 AllocateNode::Ideal_allocation(st_base),
855 phase)) {
856 // Success: The bases are provably independent.
857 mem = mem->in(MemNode::Memory);
858 continue; // (a) advance through independent store memory
859 }
860
861 // (b) At this point, if the bases or offsets do not agree, we lose,
862 // since we have not managed to prove 'this' and 'mem' independent.
863 if (st_base == base && st_offset == offset) {
864 return mem; // let caller handle steps (c), (d)
865 }
866
867 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
868 InitializeNode* st_init = mem->in(0)->as_Initialize();
869 AllocateNode* st_alloc = st_init->allocation();
870 if (st_alloc == nullptr) {
871 break; // something degenerated
872 }
873 bool known_identical = false;
874 bool known_independent = false;
875 if (alloc == st_alloc) {
876 known_identical = true;
877 } else if (alloc != nullptr) {
878 known_independent = true;
879 } else if (all_controls_dominate(this, st_alloc)) {
880 known_independent = true;
881 }
882
883 if (known_independent) {
884 // The bases are provably independent: Either they are
885 // manifestly distinct allocations, or else the control
886 // of this load dominates the store's allocation.
887 int alias_idx = phase->C->get_alias_index(adr_type());
888 if (alias_idx == Compile::AliasIdxRaw) {
889 mem = st_alloc->in(TypeFunc::Memory);
890 } else {
891 mem = st_init->memory(alias_idx);
892 }
893 continue; // (a) advance through independent store memory
894 }
895
896 // (b) at this point, if we are not looking at a store initializing
897 // the same allocation we are loading from, we lose.
898 if (known_identical) {
899 // From caller, can_see_stored_value will consult find_captured_store.
900 return mem; // let caller handle steps (c), (d)
901 }
902
903 } else if (find_previous_arraycopy(phase, alloc, mem, false) != nullptr) {
904 if (prev != mem) {
905 // Found an arraycopy but it doesn't affect that load
906 continue;
907 }
908 // Found an arraycopy that may affect that load
909 return mem;
910 } else if (addr_t != nullptr && addr_t->is_known_instance_field()) {
911 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
912 if (mem->is_Proj() && mem->in(0)->is_Call()) {
913 // ArrayCopyNodes processed here as well.
914 CallNode *call = mem->in(0)->as_Call();
915 if (!call->may_modify(addr_t, phase)) {
916 mem = call->in(TypeFunc::Memory);
917 continue; // (a) advance through independent call memory
918 }
919 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
920 ArrayCopyNode* ac = nullptr;
921 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
922 break;
923 }
924 mem = mem->in(0)->in(TypeFunc::Memory);
925 continue; // (a) advance through independent MemBar memory
926 } else if (mem->is_ClearArray()) {
927 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
928 // (the call updated 'mem' value)
929 continue; // (a) advance through independent allocation memory
930 } else {
931 // Can not bypass initialization of the instance
932 // we are looking for.
933 return mem;
934 }
935 } else if (mem->is_MergeMem()) {
936 int alias_idx = phase->C->get_alias_index(adr_type());
937 mem = mem->as_MergeMem()->memory_at(alias_idx);
938 continue; // (a) advance through independent MergeMem memory
939 }
940 }
941
942 // Unless there is an explicit 'continue', we must bail out here,
943 // because 'mem' is an inscrutable memory state (e.g., a call).
944 break;
945 }
946
947 return nullptr; // bail out
948 }
949
950 //----------------------calculate_adr_type-------------------------------------
951 // Helper function. Notices when the given type of address hits top or bottom.
952 // Also, asserts a cross-check of the type against the expected address type.
953 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
954 if (t == Type::TOP) return nullptr; // does not touch memory any more?
955 #ifdef ASSERT
956 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = nullptr;
957 #endif
958 const TypePtr* tp = t->isa_ptr();
959 if (tp == nullptr) {
960 assert(cross_check == nullptr || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
961 return TypePtr::BOTTOM; // touches lots of memory
962 } else {
963 #ifdef ASSERT
964 // %%%% [phh] We don't check the alias index if cross_check is
965 // TypeRawPtr::BOTTOM. Needs to be investigated.
966 if (cross_check != nullptr &&
967 cross_check != TypePtr::BOTTOM &&
968 cross_check != TypeRawPtr::BOTTOM) {
969 // Recheck the alias index, to see if it has changed (due to a bug).
970 Compile* C = Compile::current();
971 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
972 "must stay in the original alias category");
973 // The type of the address must be contained in the adr_type,
974 // disregarding "null"-ness.
975 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
976 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
977 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
978 "real address must not escape from expected memory type");
979 }
980 #endif
981 return tp;
982 }
983 }
984
985 uint8_t MemNode::barrier_data(const Node* n) {
986 if (n->is_LoadStore()) {
987 return n->as_LoadStore()->barrier_data();
988 } else if (n->is_Mem()) {
989 return n->as_Mem()->barrier_data();
990 }
991 return 0;
992 }
993
994 //=============================================================================
995 // Should LoadNode::Ideal() attempt to remove control edges?
996 bool LoadNode::can_remove_control() const {
997 return !has_pinned_control_dependency();
998 }
999 uint LoadNode::size_of() const { return sizeof(*this); }
1000 bool LoadNode::cmp(const Node &n) const {
1001 LoadNode& load = (LoadNode &)n;
1002 return Type::equals(_type, load._type) &&
1003 _control_dependency == load._control_dependency &&
1004 _mo == load._mo;
1005 }
1006 const Type *LoadNode::bottom_type() const { return _type; }
1007 uint LoadNode::ideal_reg() const {
1008 return _type->ideal_reg();
1009 }
1010
1011 #ifndef PRODUCT
1012 void LoadNode::dump_spec(outputStream *st) const {
1013 MemNode::dump_spec(st);
1014 if( !Verbose && !WizardMode ) {
1015 // standard dump does this in Verbose and WizardMode
1016 st->print(" #"); _type->dump_on(st);
1017 }
1018 if (!depends_only_on_test()) {
1019 st->print(" (does not depend only on test, ");
1020 if (control_dependency() == UnknownControl) {
1021 st->print("unknown control");
1022 } else if (control_dependency() == Pinned) {
1023 st->print("pinned");
1024 } else if (adr_type() == TypeRawPtr::BOTTOM) {
1025 st->print("raw access");
1026 } else {
1027 st->print("unknown reason");
1028 }
1029 st->print(")");
1030 }
1031 }
1032 #endif
1033
1034 #ifdef ASSERT
1035 //----------------------------is_immutable_value-------------------------------
1036 // Helper function to allow a raw load without control edge for some cases
1037 bool LoadNode::is_immutable_value(Node* adr) {
1038 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
1039 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal) {
1040
1041 jlong offset = adr->in(AddPNode::Offset)->find_intptr_t_con(-1);
1042 int offsets[] = {
1043 in_bytes(JavaThread::osthread_offset()),
1044 in_bytes(JavaThread::threadObj_offset()),
1045 in_bytes(JavaThread::vthread_offset()),
1046 in_bytes(JavaThread::scopedValueCache_offset()),
1047 };
1048
1049 for (size_t i = 0; i < sizeof offsets / sizeof offsets[0]; i++) {
1050 if (offset == offsets[i]) {
1051 return true;
1052 }
1053 }
1054 }
1055
1056 return false;
1057 }
1058 #endif
1059
1060 //----------------------------LoadNode::make-----------------------------------
1061 // Polymorphic factory method:
1062 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo,
1063 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
1064 Compile* C = gvn.C;
1065
1066 // sanity check the alias category against the created node type
1067 assert(!(adr_type->isa_oopptr() &&
1068 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
1069 "use LoadKlassNode instead");
1070 assert(!(adr_type->isa_aryptr() &&
1071 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
1072 "use LoadRangeNode instead");
1073 // Check control edge of raw loads
1074 assert( ctl != nullptr || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
1075 // oop will be recorded in oop map if load crosses safepoint
1076 rt->isa_oopptr() || is_immutable_value(adr),
1077 "raw memory operations should have control edge");
1078 LoadNode* load = nullptr;
1079 switch (bt) {
1080 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1081 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1082 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1083 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1084 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
1085 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break;
1086 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
1087 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break;
1088 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
1089 case T_OBJECT:
1090 case T_NARROWOOP:
1091 #ifdef _LP64
1092 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1093 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
1094 } else
1095 #endif
1096 {
1097 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
1098 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
1099 }
1100 break;
1101 default:
1102 ShouldNotReachHere();
1103 break;
1104 }
1105 assert(load != nullptr, "LoadNode should have been created");
1106 if (unaligned) {
1107 load->set_unaligned_access();
1108 }
1109 if (mismatched) {
1110 load->set_mismatched_access();
1111 }
1112 if (unsafe) {
1113 load->set_unsafe_access();
1114 }
1115 load->set_barrier_data(barrier_data);
1116 if (load->Opcode() == Op_LoadN) {
1117 Node* ld = gvn.transform(load);
1118 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
1119 }
1120
1121 return load;
1122 }
1123
1124 //------------------------------hash-------------------------------------------
1125 uint LoadNode::hash() const {
1126 // unroll addition of interesting fields
1127 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
1128 }
1129
1130 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
1131 if ((atp != nullptr) && (atp->index() >= Compile::AliasIdxRaw)) {
1132 bool non_volatile = (atp->field() != nullptr) && !atp->field()->is_volatile();
1133 bool is_stable_ary = FoldStableValues &&
1134 (tp != nullptr) && (tp->isa_aryptr() != nullptr) &&
1135 tp->isa_aryptr()->is_stable();
1136
1137 return (eliminate_boxing && non_volatile) || is_stable_ary || tp->is_inlinetypeptr();
1138 }
1139
1140 return false;
1141 }
1142
1143 LoadNode* LoadNode::pin_array_access_node() const {
1144 const TypePtr* adr_type = this->adr_type();
1145 if (adr_type != nullptr && adr_type->isa_aryptr()) {
1146 return clone_pinned();
1147 }
1148 return nullptr;
1149 }
1150
1151 // Is the value loaded previously stored by an arraycopy? If so return
1152 // a load node that reads from the source array so we may be able to
1153 // optimize out the ArrayCopy node later.
1154 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
1155 Node* ld_adr = in(MemNode::Address);
1156 intptr_t ld_off = 0;
1157 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
1158 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
1159 if (ac != nullptr) {
1160 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
1161
1162 Node* mem = ac->in(TypeFunc::Memory);
1163 Node* ctl = ac->in(0);
1164 Node* src = ac->in(ArrayCopyNode::Src);
1165
1166 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
1167 return nullptr;
1168 }
1169
1170 // load depends on the tests that validate the arraycopy
1171 LoadNode* ld = clone_pinned();
1172 Node* addp = in(MemNode::Address)->clone();
1173 if (ac->as_ArrayCopy()->is_clonebasic()) {
1174 assert(ld_alloc != nullptr, "need an alloc");
1175 assert(addp->is_AddP(), "address must be addp");
1176 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1177 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1178 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
1179 addp->set_req(AddPNode::Base, src);
1180 addp->set_req(AddPNode::Address, src);
1181 } else {
1182 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1183 ac->as_ArrayCopy()->is_copyof_validated() ||
1184 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1185 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1186 addp->set_req(AddPNode::Base, src);
1187 addp->set_req(AddPNode::Address, src);
1188
1189 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1190 BasicType ary_elem = ary_t->isa_aryptr()->elem()->array_element_basic_type();
1191 if (is_reference_type(ary_elem, true)) ary_elem = T_OBJECT;
1192
1193 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
1194 uint shift = ary_t->is_flat() ? ary_t->flat_log_elem_size() : exact_log2(type2aelembytes(ary_elem));
1195
1196 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1197 #ifdef _LP64
1198 diff = phase->transform(new ConvI2LNode(diff));
1199 #endif
1200 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1201
1202 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1203 addp->set_req(AddPNode::Offset, offset);
1204 }
1205 addp = phase->transform(addp);
1206 #ifdef ASSERT
1207 const TypePtr* adr_type = phase->type(addp)->is_ptr();
1208 ld->_adr_type = adr_type;
1209 #endif
1210 ld->set_req(MemNode::Address, addp);
1211 ld->set_req(0, ctl);
1212 ld->set_req(MemNode::Memory, mem);
1213 return ld;
1214 }
1215 return nullptr;
1216 }
1217
1218 static Node* see_through_inline_type(PhaseValues* phase, const MemNode* load, Node* base, int offset) {
1219 if (!load->is_mismatched_access() && base != nullptr && base->is_InlineType() && offset > oopDesc::klass_offset_in_bytes()) {
1220 InlineTypeNode* vt = base->as_InlineType();
1221 Node* value = vt->field_value_by_offset(offset, true);
1222 assert(value != nullptr, "must see some value");
1223 return value;
1224 }
1225
1226 return nullptr;
1227 }
1228
1229 //---------------------------can_see_stored_value------------------------------
1230 // This routine exists to make sure this set of tests is done the same
1231 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
1232 // will change the graph shape in a way which makes memory alive twice at the
1233 // same time (uses the Oracle model of aliasing), then some
1234 // LoadXNode::Identity will fold things back to the equivalence-class model
1235 // of aliasing.
1236 Node* MemNode::can_see_stored_value(Node* st, PhaseValues* phase) const {
1237 Node* ld_adr = in(MemNode::Address);
1238 intptr_t ld_off = 0;
1239 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1240 // Try to see through an InlineTypeNode
1241 // LoadN is special because the input is not compressed
1242 if (Opcode() != Op_LoadN) {
1243 Node* value = see_through_inline_type(phase, this, ld_base, ld_off);
1244 if (value != nullptr) {
1245 return value;
1246 }
1247 }
1248
1249 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base);
1250 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1251 Compile::AliasType* atp = (tp != nullptr) ? phase->C->alias_type(tp) : nullptr;
1252 // This is more general than load from boxing objects.
1253 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1254 uint alias_idx = atp->index();
1255 Node* result = nullptr;
1256 Node* current = st;
1257 // Skip through chains of MemBarNodes checking the MergeMems for
1258 // new states for the slice of this load. Stop once any other
1259 // kind of node is encountered. Loads from final memory can skip
1260 // through any kind of MemBar but normal loads shouldn't skip
1261 // through MemBarAcquire since the could allow them to move out of
1262 // a synchronized region. It is not safe to step over MemBarCPUOrder,
1263 // because alias info above them may be inaccurate (e.g., due to
1264 // mixed/mismatched unsafe accesses).
1265 bool is_final_mem = !atp->is_rewritable();
1266 while (current->is_Proj()) {
1267 int opc = current->in(0)->Opcode();
1268 if ((is_final_mem && (opc == Op_MemBarAcquire ||
1269 opc == Op_MemBarAcquireLock ||
1270 opc == Op_LoadFence)) ||
1271 opc == Op_MemBarRelease ||
1272 opc == Op_StoreFence ||
1273 opc == Op_MemBarReleaseLock ||
1274 opc == Op_MemBarStoreStore ||
1275 opc == Op_StoreStoreFence) {
1276 Node* mem = current->in(0)->in(TypeFunc::Memory);
1277 if (mem->is_MergeMem()) {
1278 MergeMemNode* merge = mem->as_MergeMem();
1279 Node* new_st = merge->memory_at(alias_idx);
1280 if (new_st == merge->base_memory()) {
1281 // Keep searching
1282 current = new_st;
1283 continue;
1284 }
1285 // Save the new memory state for the slice and fall through
1286 // to exit.
1287 result = new_st;
1288 }
1289 }
1290 break;
1291 }
1292 if (result != nullptr) {
1293 st = result;
1294 }
1295 }
1296
1297 // Loop around twice in the case Load -> Initialize -> Store.
1298 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1299 for (int trip = 0; trip <= 1; trip++) {
1300
1301 if (st->is_Store()) {
1302 Node* st_adr = st->in(MemNode::Address);
1303 if (st_adr != ld_adr) {
1304 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1305 intptr_t st_off = 0;
1306 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1307 if (ld_base == nullptr) return nullptr;
1308 if (st_base == nullptr) return nullptr;
1309 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return nullptr;
1310 if (ld_off != st_off) return nullptr;
1311 if (ld_off == Type::OffsetBot) return nullptr;
1312 // Same base, same offset.
1313 // Possible improvement for arrays: check index value instead of absolute offset.
1314
1315 // At this point we have proven something like this setup:
1316 // B = << base >>
1317 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1318 // S = StoreQ(AddP( B , #Off), V)
1319 // (Actually, we haven't yet proven the Q's are the same.)
1320 // In other words, we are loading from a casted version of
1321 // the same pointer-and-offset that we stored to.
1322 // Casted version may carry a dependency and it is respected.
1323 // Thus, we are able to replace L by V.
1324 }
1325 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1326 if (store_Opcode() != st->Opcode()) {
1327 return nullptr;
1328 }
1329 // LoadVector/StoreVector needs additional check to ensure the types match.
1330 if (st->is_StoreVector()) {
1331 const TypeVect* in_vt = st->as_StoreVector()->vect_type();
1332 const TypeVect* out_vt = is_Load() ? as_LoadVector()->vect_type() : as_StoreVector()->vect_type();
1333 if (in_vt != out_vt) {
1334 return nullptr;
1335 }
1336 }
1337 return st->in(MemNode::ValueIn);
1338 }
1339
1340 // A load from a freshly-created object always returns zero.
1341 // (This can happen after LoadNode::Ideal resets the load's memory input
1342 // to find_captured_store, which returned InitializeNode::zero_memory.)
1343 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1344 (st->in(0) == ld_alloc) &&
1345 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1346 // return a zero value for the load's basic type
1347 // (This is one of the few places where a generic PhaseTransform
1348 // can create new nodes. Think of it as lazily manifesting
1349 // virtually pre-existing constants.)
1350 Node* init_value = ld_alloc->in(AllocateNode::InitValue);
1351 if (init_value != nullptr) {
1352 // TODO 8350865 Is this correct for non-all-zero init values? Don't we need field_value_by_offset?
1353 return init_value;
1354 }
1355 assert(ld_alloc->in(AllocateNode::RawInitValue) == nullptr, "init value may not be null");
1356 if (memory_type() != T_VOID) {
1357 if (ReduceBulkZeroing || find_array_copy_clone(ld_alloc, in(MemNode::Memory)) == nullptr) {
1358 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1359 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1360 // by the ArrayCopyNode.
1361 return phase->zerocon(memory_type());
1362 }
1363 } else {
1364 // TODO: materialize all-zero vector constant
1365 assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1366 }
1367 }
1368
1369 // A load from an initialization barrier can match a captured store.
1370 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1371 InitializeNode* init = st->in(0)->as_Initialize();
1372 AllocateNode* alloc = init->allocation();
1373 if ((alloc != nullptr) && (alloc == ld_alloc)) {
1374 // examine a captured store value
1375 st = init->find_captured_store(ld_off, memory_size(), phase);
1376 if (st != nullptr) {
1377 continue; // take one more trip around
1378 }
1379 }
1380 }
1381
1382 // Load boxed value from result of valueOf() call is input parameter.
1383 if (this->is_Load() && ld_adr->is_AddP() &&
1384 (tp != nullptr) && tp->is_ptr_to_boxed_value()) {
1385 intptr_t ignore = 0;
1386 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1387 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1388 base = bs->step_over_gc_barrier(base);
1389 if (base != nullptr && base->is_Proj() &&
1390 base->as_Proj()->_con == TypeFunc::Parms &&
1391 base->in(0)->is_CallStaticJava() &&
1392 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1393 return base->in(0)->in(TypeFunc::Parms);
1394 }
1395 }
1396
1397 break;
1398 }
1399
1400 return nullptr;
1401 }
1402
1403 //----------------------is_instance_field_load_with_local_phi------------------
1404 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1405 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1406 in(Address)->is_AddP() ) {
1407 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1408 // Only instances and boxed values.
1409 if( t_oop != nullptr &&
1410 (t_oop->is_ptr_to_boxed_value() ||
1411 t_oop->is_known_instance_field()) &&
1412 t_oop->offset() != Type::OffsetBot &&
1413 t_oop->offset() != Type::OffsetTop) {
1414 return true;
1415 }
1416 }
1417 return false;
1418 }
1419
1420 //------------------------------Identity---------------------------------------
1421 // Loads are identity if previous store is to same address
1422 Node* LoadNode::Identity(PhaseGVN* phase) {
1423 // If the previous store-maker is the right kind of Store, and the store is
1424 // to the same address, then we are equal to the value stored.
1425 Node* mem = in(Memory);
1426 Node* value = can_see_stored_value(mem, phase);
1427 if( value ) {
1428 // byte, short & char stores truncate naturally.
1429 // A load has to load the truncated value which requires
1430 // some sort of masking operation and that requires an
1431 // Ideal call instead of an Identity call.
1432 if (memory_size() < BytesPerInt) {
1433 // If the input to the store does not fit with the load's result type,
1434 // it must be truncated via an Ideal call.
1435 if (!phase->type(value)->higher_equal(phase->type(this)))
1436 return this;
1437 }
1438 // (This works even when value is a Con, but LoadNode::Value
1439 // usually runs first, producing the singleton type of the Con.)
1440 if (!has_pinned_control_dependency() || value->is_Con()) {
1441 return value;
1442 } else {
1443 return this;
1444 }
1445 }
1446
1447 if (has_pinned_control_dependency()) {
1448 return this;
1449 }
1450 // Search for an existing data phi which was generated before for the same
1451 // instance's field to avoid infinite generation of phis in a loop.
1452 Node *region = mem->in(0);
1453 if (is_instance_field_load_with_local_phi(region)) {
1454 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1455 int this_index = phase->C->get_alias_index(addr_t);
1456 int this_offset = addr_t->offset();
1457 int this_iid = addr_t->instance_id();
1458 if (!addr_t->is_known_instance() &&
1459 addr_t->is_ptr_to_boxed_value()) {
1460 // Use _idx of address base (could be Phi node) for boxed values.
1461 intptr_t ignore = 0;
1462 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1463 if (base == nullptr) {
1464 return this;
1465 }
1466 this_iid = base->_idx;
1467 }
1468 const Type* this_type = bottom_type();
1469 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1470 Node* phi = region->fast_out(i);
1471 if (phi->is_Phi() && phi != mem &&
1472 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1473 return phi;
1474 }
1475 }
1476 }
1477
1478 return this;
1479 }
1480
1481 // Construct an equivalent unsigned load.
1482 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1483 BasicType bt = T_ILLEGAL;
1484 const Type* rt = nullptr;
1485 switch (Opcode()) {
1486 case Op_LoadUB: return this;
1487 case Op_LoadUS: return this;
1488 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1489 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1490 default:
1491 assert(false, "no unsigned variant: %s", Name());
1492 return nullptr;
1493 }
1494 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1495 raw_adr_type(), rt, bt, _mo, _control_dependency,
1496 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1497 }
1498
1499 // Construct an equivalent signed load.
1500 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1501 BasicType bt = T_ILLEGAL;
1502 const Type* rt = nullptr;
1503 switch (Opcode()) {
1504 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1505 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1506 case Op_LoadB: // fall through
1507 case Op_LoadS: // fall through
1508 case Op_LoadI: // fall through
1509 case Op_LoadL: return this;
1510 default:
1511 assert(false, "no signed variant: %s", Name());
1512 return nullptr;
1513 }
1514 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1515 raw_adr_type(), rt, bt, _mo, _control_dependency,
1516 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1517 }
1518
1519 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1520 BasicType bt = rt->basic_type();
1521 switch (Opcode()) {
1522 case Op_LoadI: return (bt == T_FLOAT);
1523 case Op_LoadL: return (bt == T_DOUBLE);
1524 case Op_LoadF: return (bt == T_INT);
1525 case Op_LoadD: return (bt == T_LONG);
1526
1527 default: return false;
1528 }
1529 }
1530
1531 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1532 BasicType bt = rt->basic_type();
1533 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1534 bool is_mismatched = is_mismatched_access();
1535 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1536 if (raw_type == nullptr) {
1537 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1538 }
1539 const int op = Opcode();
1540 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) ||
1541 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access());
1542 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1543 raw_adr_type(), rt, bt, _mo, _control_dependency,
1544 require_atomic_access, is_unaligned_access(), is_mismatched);
1545 }
1546
1547 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1548 BasicType bt = vt->basic_type();
1549 switch (Opcode()) {
1550 case Op_StoreI: return (bt == T_FLOAT);
1551 case Op_StoreL: return (bt == T_DOUBLE);
1552 case Op_StoreF: return (bt == T_INT);
1553 case Op_StoreD: return (bt == T_LONG);
1554
1555 default: return false;
1556 }
1557 }
1558
1559 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1560 BasicType bt = vt->basic_type();
1561 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1562 const int op = Opcode();
1563 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) ||
1564 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access());
1565 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1566 raw_adr_type(), val, bt, _mo, require_atomic_access);
1567
1568 bool is_mismatched = is_mismatched_access();
1569 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1570 if (raw_type == nullptr) {
1571 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1572 }
1573 if (is_mismatched) {
1574 st->set_mismatched_access();
1575 }
1576 return st;
1577 }
1578
1579 // We're loading from an object which has autobox behaviour.
1580 // If this object is result of a valueOf call we'll have a phi
1581 // merging a newly allocated object and a load from the cache.
1582 // We want to replace this load with the original incoming
1583 // argument to the valueOf call.
1584 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1585 assert(igvn->C->eliminate_boxing(), "sanity");
1586 intptr_t ignore = 0;
1587 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1588 if ((base == nullptr) || base->is_Phi()) {
1589 // Push the loads from the phi that comes from valueOf up
1590 // through it to allow elimination of the loads and the recovery
1591 // of the original value. It is done in split_through_phi().
1592 return nullptr;
1593 } else if (base->is_Load() ||
1594 (base->is_DecodeN() && base->in(1)->is_Load())) {
1595 // Eliminate the load of boxed value for integer types from the cache
1596 // array by deriving the value from the index into the array.
1597 // Capture the offset of the load and then reverse the computation.
1598
1599 // Get LoadN node which loads a boxing object from 'cache' array.
1600 if (base->is_DecodeN()) {
1601 base = base->in(1);
1602 }
1603 if (!base->in(Address)->is_AddP()) {
1604 return nullptr; // Complex address
1605 }
1606 AddPNode* address = base->in(Address)->as_AddP();
1607 Node* cache_base = address->in(AddPNode::Base);
1608 if ((cache_base != nullptr) && cache_base->is_DecodeN()) {
1609 // Get ConP node which is static 'cache' field.
1610 cache_base = cache_base->in(1);
1611 }
1612 if ((cache_base != nullptr) && cache_base->is_Con()) {
1613 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1614 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
1615 Node* elements[4];
1616 int shift = exact_log2(type2aelembytes(T_OBJECT));
1617 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1618 if (count > 0 && elements[0]->is_Con() &&
1619 (count == 1 ||
1620 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1621 elements[1]->in(2) == igvn->intcon(shift)))) {
1622 ciObjArray* array = base_type->const_oop()->as_obj_array();
1623 // Fetch the box object cache[0] at the base of the array and get its value
1624 ciInstance* box = array->obj_at(0)->as_instance();
1625 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1626 assert(ik->is_box_klass(), "sanity");
1627 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1628 if (ik->nof_nonstatic_fields() == 1) {
1629 // This should be true nonstatic_field_at requires calling
1630 // nof_nonstatic_fields so check it anyway
1631 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1632 BasicType bt = c.basic_type();
1633 // Only integer types have boxing cache.
1634 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1635 bt == T_BYTE || bt == T_SHORT ||
1636 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1637 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1638 if (cache_low != (int)cache_low) {
1639 return nullptr; // should not happen since cache is array indexed by value
1640 }
1641 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1642 if (offset != (int)offset) {
1643 return nullptr; // should not happen since cache is array indexed by value
1644 }
1645 // Add up all the offsets making of the address of the load
1646 Node* result = elements[0];
1647 for (int i = 1; i < count; i++) {
1648 result = igvn->transform(new AddXNode(result, elements[i]));
1649 }
1650 // Remove the constant offset from the address and then
1651 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1652 // remove the scaling of the offset to recover the original index.
1653 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1654 // Peel the shift off directly but wrap it in a dummy node
1655 // since Ideal can't return existing nodes
1656 igvn->_worklist.push(result); // remove dead node later
1657 result = new RShiftXNode(result->in(1), igvn->intcon(0));
1658 } else if (result->is_Add() && result->in(2)->is_Con() &&
1659 result->in(1)->Opcode() == Op_LShiftX &&
1660 result->in(1)->in(2) == igvn->intcon(shift)) {
1661 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1662 // but for boxing cache access we know that X<<Z will not overflow
1663 // (there is range check) so we do this optimizatrion by hand here.
1664 igvn->_worklist.push(result); // remove dead node later
1665 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1666 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1667 } else {
1668 result = new RShiftXNode(result, igvn->intcon(shift));
1669 }
1670 #ifdef _LP64
1671 if (bt != T_LONG) {
1672 result = new ConvL2INode(igvn->transform(result));
1673 }
1674 #else
1675 if (bt == T_LONG) {
1676 result = new ConvI2LNode(igvn->transform(result));
1677 }
1678 #endif
1679 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1680 // Need to preserve unboxing load type if it is unsigned.
1681 switch(this->Opcode()) {
1682 case Op_LoadUB:
1683 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1684 break;
1685 case Op_LoadUS:
1686 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1687 break;
1688 }
1689 return result;
1690 }
1691 }
1692 }
1693 }
1694 }
1695 return nullptr;
1696 }
1697
1698 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1699 Node* region = phi->in(0);
1700 if (region == nullptr) {
1701 return false; // Wait stable graph
1702 }
1703 uint cnt = phi->req();
1704 for (uint i = 1; i < cnt; i++) {
1705 Node* rc = region->in(i);
1706 if (rc == nullptr || phase->type(rc) == Type::TOP)
1707 return false; // Wait stable graph
1708 Node* in = phi->in(i);
1709 if (in == nullptr || phase->type(in) == Type::TOP)
1710 return false; // Wait stable graph
1711 }
1712 return true;
1713 }
1714
1715 //------------------------------split_through_phi------------------------------
1716 // Check whether a call to 'split_through_phi' would split this load through the
1717 // Phi *base*. This method is essentially a copy of the validations performed
1718 // by 'split_through_phi'. The first use of this method was in EA code as part
1719 // of simplification of allocation merges.
1720 // Some differences from original method (split_through_phi):
1721 // - If base->is_CastPP(): base = base->in(1)
1722 bool LoadNode::can_split_through_phi_base(PhaseGVN* phase) {
1723 Node* mem = in(Memory);
1724 Node* address = in(Address);
1725 intptr_t ignore = 0;
1726 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1727
1728 if (base->is_CastPP()) {
1729 base = base->in(1);
1730 }
1731
1732 if (req() > 3 || base == nullptr || !base->is_Phi()) {
1733 return false;
1734 }
1735
1736 if (!mem->is_Phi()) {
1737 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1738 return false;
1739 }
1740 } else if (base->in(0) != mem->in(0)) {
1741 if (!MemNode::all_controls_dominate(mem, base->in(0))) {
1742 return false;
1743 }
1744 }
1745
1746 return true;
1747 }
1748
1749 //------------------------------split_through_phi------------------------------
1750 // Split instance or boxed field load through Phi.
1751 Node* LoadNode::split_through_phi(PhaseGVN* phase, bool ignore_missing_instance_id) {
1752 if (req() > 3) {
1753 assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs");
1754 // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account
1755 return nullptr;
1756 }
1757 Node* mem = in(Memory);
1758 Node* address = in(Address);
1759 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1760
1761 assert((t_oop != nullptr) &&
1762 (ignore_missing_instance_id ||
1763 t_oop->is_known_instance_field() ||
1764 t_oop->is_ptr_to_boxed_value()), "invalid conditions");
1765
1766 Compile* C = phase->C;
1767 intptr_t ignore = 0;
1768 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1769 bool base_is_phi = (base != nullptr) && base->is_Phi();
1770 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1771 (base != nullptr) && (base == address->in(AddPNode::Base)) &&
1772 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1773
1774 if (!((mem->is_Phi() || base_is_phi) &&
1775 (ignore_missing_instance_id || load_boxed_values || t_oop->is_known_instance_field()))) {
1776 return nullptr; // Neither memory or base are Phi
1777 }
1778
1779 if (mem->is_Phi()) {
1780 if (!stable_phi(mem->as_Phi(), phase)) {
1781 return nullptr; // Wait stable graph
1782 }
1783 uint cnt = mem->req();
1784 // Check for loop invariant memory.
1785 if (cnt == 3) {
1786 for (uint i = 1; i < cnt; i++) {
1787 Node* in = mem->in(i);
1788 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1789 if (m == mem) {
1790 if (i == 1) {
1791 // if the first edge was a loop, check second edge too.
1792 // If both are replaceable - we are in an infinite loop
1793 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1794 if (n == mem) {
1795 break;
1796 }
1797 }
1798 set_req(Memory, mem->in(cnt - i));
1799 return this; // made change
1800 }
1801 }
1802 }
1803 }
1804 if (base_is_phi) {
1805 if (!stable_phi(base->as_Phi(), phase)) {
1806 return nullptr; // Wait stable graph
1807 }
1808 uint cnt = base->req();
1809 // Check for loop invariant memory.
1810 if (cnt == 3) {
1811 for (uint i = 1; i < cnt; i++) {
1812 if (base->in(i) == base) {
1813 return nullptr; // Wait stable graph
1814 }
1815 }
1816 }
1817 }
1818
1819 // Split through Phi (see original code in loopopts.cpp).
1820 assert(ignore_missing_instance_id || C->have_alias_type(t_oop), "instance should have alias type");
1821
1822 // Do nothing here if Identity will find a value
1823 // (to avoid infinite chain of value phis generation).
1824 if (this != Identity(phase)) {
1825 return nullptr;
1826 }
1827
1828 // Select Region to split through.
1829 Node* region;
1830 DomResult dom_result = DomResult::Dominate;
1831 if (!base_is_phi) {
1832 assert(mem->is_Phi(), "sanity");
1833 region = mem->in(0);
1834 // Skip if the region dominates some control edge of the address.
1835 // We will check `dom_result` later.
1836 dom_result = MemNode::maybe_all_controls_dominate(address, region);
1837 } else if (!mem->is_Phi()) {
1838 assert(base_is_phi, "sanity");
1839 region = base->in(0);
1840 // Skip if the region dominates some control edge of the memory.
1841 // We will check `dom_result` later.
1842 dom_result = MemNode::maybe_all_controls_dominate(mem, region);
1843 } else if (base->in(0) != mem->in(0)) {
1844 assert(base_is_phi && mem->is_Phi(), "sanity");
1845 dom_result = MemNode::maybe_all_controls_dominate(mem, base->in(0));
1846 if (dom_result == DomResult::Dominate) {
1847 region = base->in(0);
1848 } else {
1849 dom_result = MemNode::maybe_all_controls_dominate(address, mem->in(0));
1850 if (dom_result == DomResult::Dominate) {
1851 region = mem->in(0);
1852 }
1853 // Otherwise we encountered a complex graph.
1854 }
1855 } else {
1856 assert(base->in(0) == mem->in(0), "sanity");
1857 region = mem->in(0);
1858 }
1859
1860 PhaseIterGVN* igvn = phase->is_IterGVN();
1861 if (dom_result != DomResult::Dominate) {
1862 if (dom_result == DomResult::EncounteredDeadCode) {
1863 // There is some dead code which eventually will be removed in IGVN.
1864 // Once this is the case, we get an unambiguous dominance result.
1865 // Push the node to the worklist again until the dead code is removed.
1866 igvn->_worklist.push(this);
1867 }
1868 return nullptr;
1869 }
1870
1871 Node* phi = nullptr;
1872 const Type* this_type = this->bottom_type();
1873 if (t_oop != nullptr && (t_oop->is_known_instance_field() || load_boxed_values)) {
1874 int this_index = C->get_alias_index(t_oop);
1875 int this_offset = t_oop->offset();
1876 int this_iid = t_oop->is_known_instance_field() ? t_oop->instance_id() : base->_idx;
1877 phi = new PhiNode(region, this_type, nullptr, mem->_idx, this_iid, this_index, this_offset);
1878 } else if (ignore_missing_instance_id) {
1879 phi = new PhiNode(region, this_type, nullptr, mem->_idx);
1880 } else {
1881 return nullptr;
1882 }
1883
1884 for (uint i = 1; i < region->req(); i++) {
1885 Node* x;
1886 Node* the_clone = nullptr;
1887 Node* in = region->in(i);
1888 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1889 in != nullptr && in->is_OuterStripMinedLoop()) {
1890 // No node should go in the outer strip mined loop
1891 in = in->in(LoopNode::EntryControl);
1892 }
1893 if (in == nullptr || in == C->top()) {
1894 x = C->top(); // Dead path? Use a dead data op
1895 } else {
1896 x = this->clone(); // Else clone up the data op
1897 the_clone = x; // Remember for possible deletion.
1898 // Alter data node to use pre-phi inputs
1899 if (this->in(0) == region) {
1900 x->set_req(0, in);
1901 } else {
1902 x->set_req(0, nullptr);
1903 }
1904 if (mem->is_Phi() && (mem->in(0) == region)) {
1905 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1906 }
1907 if (address->is_Phi() && address->in(0) == region) {
1908 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1909 }
1910 if (base_is_phi && (base->in(0) == region)) {
1911 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1912 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1913 x->set_req(Address, adr_x);
1914 }
1915 }
1916 // Check for a 'win' on some paths
1917 const Type *t = x->Value(igvn);
1918
1919 bool singleton = t->singleton();
1920
1921 // See comments in PhaseIdealLoop::split_thru_phi().
1922 if (singleton && t == Type::TOP) {
1923 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1924 }
1925
1926 if (singleton) {
1927 x = igvn->makecon(t);
1928 } else {
1929 // We now call Identity to try to simplify the cloned node.
1930 // Note that some Identity methods call phase->type(this).
1931 // Make sure that the type array is big enough for
1932 // our new node, even though we may throw the node away.
1933 // (This tweaking with igvn only works because x is a new node.)
1934 igvn->set_type(x, t);
1935 // If x is a TypeNode, capture any more-precise type permanently into Node
1936 // otherwise it will be not updated during igvn->transform since
1937 // igvn->type(x) is set to x->Value() already.
1938 x->raise_bottom_type(t);
1939 Node* y = x->Identity(igvn);
1940 if (y != x) {
1941 x = y;
1942 } else {
1943 y = igvn->hash_find_insert(x);
1944 if (y) {
1945 x = y;
1946 } else {
1947 // Else x is a new node we are keeping
1948 // We do not need register_new_node_with_optimizer
1949 // because set_type has already been called.
1950 igvn->_worklist.push(x);
1951 }
1952 }
1953 }
1954 if (x != the_clone && the_clone != nullptr) {
1955 igvn->remove_dead_node(the_clone);
1956 }
1957 phi->set_req(i, x);
1958 }
1959 // Record Phi
1960 igvn->register_new_node_with_optimizer(phi);
1961 return phi;
1962 }
1963
1964 AllocateNode* LoadNode::is_new_object_mark_load() const {
1965 if (Opcode() == Op_LoadX) {
1966 Node* address = in(MemNode::Address);
1967 AllocateNode* alloc = AllocateNode::Ideal_allocation(address);
1968 Node* mem = in(MemNode::Memory);
1969 if (alloc != nullptr && mem->is_Proj() &&
1970 mem->in(0) != nullptr &&
1971 mem->in(0) == alloc->initialization() &&
1972 alloc->initialization()->proj_out_or_null(0) != nullptr) {
1973 return alloc;
1974 }
1975 }
1976 return nullptr;
1977 }
1978
1979
1980 //------------------------------Ideal------------------------------------------
1981 // If the load is from Field memory and the pointer is non-null, it might be possible to
1982 // zero out the control input.
1983 // If the offset is constant and the base is an object allocation,
1984 // try to hook me up to the exact initializing store.
1985 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1986 if (has_pinned_control_dependency()) {
1987 return nullptr;
1988 }
1989 Node* p = MemNode::Ideal_common(phase, can_reshape);
1990 if (p) return (p == NodeSentinel) ? nullptr : p;
1991
1992 Node* ctrl = in(MemNode::Control);
1993 Node* address = in(MemNode::Address);
1994 bool progress = false;
1995
1996 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1997 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1998
1999 // Skip up past a SafePoint control. Cannot do this for Stores because
2000 // pointer stores & cardmarks must stay on the same side of a SafePoint.
2001 if( ctrl != nullptr && ctrl->Opcode() == Op_SafePoint &&
2002 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
2003 !addr_mark &&
2004 (depends_only_on_test() || has_unknown_control_dependency())) {
2005 ctrl = ctrl->in(0);
2006 set_req(MemNode::Control,ctrl);
2007 progress = true;
2008 }
2009
2010 intptr_t ignore = 0;
2011 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
2012 if (base != nullptr
2013 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
2014 // Check for useless control edge in some common special cases
2015 if (in(MemNode::Control) != nullptr
2016 && !(phase->type(address)->is_inlinetypeptr() && is_mismatched_access())
2017 && can_remove_control()
2018 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
2019 && all_controls_dominate(base, phase->C->start())) {
2020 // A method-invariant, non-null address (constant or 'this' argument).
2021 set_req(MemNode::Control, nullptr);
2022 progress = true;
2023 }
2024 }
2025
2026 Node* mem = in(MemNode::Memory);
2027 const TypePtr *addr_t = phase->type(address)->isa_ptr();
2028
2029 if (can_reshape && (addr_t != nullptr)) {
2030 // try to optimize our memory input
2031 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
2032 if (opt_mem != mem) {
2033 set_req_X(MemNode::Memory, opt_mem, phase);
2034 if (phase->type( opt_mem ) == Type::TOP) return nullptr;
2035 return this;
2036 }
2037 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
2038 if ((t_oop != nullptr) &&
2039 (t_oop->is_known_instance_field() ||
2040 t_oop->is_ptr_to_boxed_value())) {
2041 PhaseIterGVN *igvn = phase->is_IterGVN();
2042 assert(igvn != nullptr, "must be PhaseIterGVN when can_reshape is true");
2043 if (igvn->_worklist.member(opt_mem)) {
2044 // Delay this transformation until memory Phi is processed.
2045 igvn->_worklist.push(this);
2046 return nullptr;
2047 }
2048 // Split instance field load through Phi.
2049 Node* result = split_through_phi(phase);
2050 if (result != nullptr) return result;
2051
2052 if (t_oop->is_ptr_to_boxed_value()) {
2053 Node* result = eliminate_autobox(igvn);
2054 if (result != nullptr) return result;
2055 }
2056 }
2057 }
2058
2059 // Is there a dominating load that loads the same value? Leave
2060 // anything that is not a load of a field/array element (like
2061 // barriers etc.) alone
2062 if (in(0) != nullptr && !adr_type()->isa_rawptr() && can_reshape) {
2063 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
2064 Node *use = mem->fast_out(i);
2065 if (use != this &&
2066 use->Opcode() == Opcode() &&
2067 use->in(0) != nullptr &&
2068 use->in(0) != in(0) &&
2069 use->in(Address) == in(Address)) {
2070 Node* ctl = in(0);
2071 for (int i = 0; i < 10 && ctl != nullptr; i++) {
2072 ctl = IfNode::up_one_dom(ctl);
2073 if (ctl == use->in(0)) {
2074 set_req(0, use->in(0));
2075 return this;
2076 }
2077 }
2078 }
2079 }
2080 }
2081
2082 // Check for prior store with a different base or offset; make Load
2083 // independent. Skip through any number of them. Bail out if the stores
2084 // are in an endless dead cycle and report no progress. This is a key
2085 // transform for Reflection. However, if after skipping through the Stores
2086 // we can't then fold up against a prior store do NOT do the transform as
2087 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
2088 // array memory alive twice: once for the hoisted Load and again after the
2089 // bypassed Store. This situation only works if EVERYBODY who does
2090 // anti-dependence work knows how to bypass. I.e. we need all
2091 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
2092 // the alias index stuff. So instead, peek through Stores and IFF we can
2093 // fold up, do so.
2094 Node* prev_mem = find_previous_store(phase);
2095 if (prev_mem != nullptr) {
2096 Node* value = can_see_arraycopy_value(prev_mem, phase);
2097 if (value != nullptr) {
2098 return value;
2099 }
2100 }
2101 // Steps (a), (b): Walk past independent stores to find an exact match.
2102 if (prev_mem != nullptr && prev_mem != in(MemNode::Memory)) {
2103 // (c) See if we can fold up on the spot, but don't fold up here.
2104 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
2105 // just return a prior value, which is done by Identity calls.
2106 if (can_see_stored_value(prev_mem, phase)) {
2107 // Make ready for step (d):
2108 set_req_X(MemNode::Memory, prev_mem, phase);
2109 return this;
2110 }
2111 }
2112
2113 if (progress) {
2114 return this;
2115 }
2116
2117 if (!can_reshape) {
2118 phase->record_for_igvn(this);
2119 }
2120 return nullptr;
2121 }
2122
2123 // Helper to recognize certain Klass fields which are invariant across
2124 // some group of array types (e.g., int[] or all T[] where T < Object).
2125 const Type*
2126 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
2127 ciKlass* klass) const {
2128 assert(!UseCompactObjectHeaders || tkls->offset() != in_bytes(Klass::prototype_header_offset()),
2129 "must not happen");
2130 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
2131 // The field is Klass::_access_flags. Return its (constant) value.
2132 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
2133 assert(Opcode() == Op_LoadUS, "must load an unsigned short from _access_flags");
2134 return TypeInt::make(klass->access_flags());
2135 }
2136 if (tkls->offset() == in_bytes(Klass::misc_flags_offset())) {
2137 // The field is Klass::_misc_flags. Return its (constant) value.
2138 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
2139 assert(Opcode() == Op_LoadUB, "must load an unsigned byte from _misc_flags");
2140 return TypeInt::make(klass->misc_flags());
2141 }
2142 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
2143 // The field is Klass::_layout_helper. Return its constant value if known.
2144 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2145 return TypeInt::make(klass->layout_helper());
2146 }
2147
2148 // No match.
2149 return nullptr;
2150 }
2151
2152 //------------------------------Value-----------------------------------------
2153 const Type* LoadNode::Value(PhaseGVN* phase) const {
2154 // Either input is TOP ==> the result is TOP
2155 Node* mem = in(MemNode::Memory);
2156 const Type *t1 = phase->type(mem);
2157 if (t1 == Type::TOP) return Type::TOP;
2158 Node* adr = in(MemNode::Address);
2159 const TypePtr* tp = phase->type(adr)->isa_ptr();
2160 if (tp == nullptr || tp->empty()) return Type::TOP;
2161 int off = tp->offset();
2162 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
2163 Compile* C = phase->C;
2164
2165 // If we are loading from a freshly-allocated object, produce a zero,
2166 // if the load is provably beyond the header of the object.
2167 // (Also allow a variable load from a fresh array to produce zero.)
2168 const TypeOopPtr* tinst = tp->isa_oopptr();
2169 bool is_instance = (tinst != nullptr) && tinst->is_known_instance_field();
2170 Node* value = can_see_stored_value(mem, phase);
2171 if (value != nullptr && value->is_Con()) {
2172 assert(value->bottom_type()->higher_equal(_type), "sanity");
2173 return value->bottom_type();
2174 }
2175
2176 // Try to guess loaded type from pointer type
2177 if (tp->isa_aryptr()) {
2178 const TypeAryPtr* ary = tp->is_aryptr();
2179 const Type* t = ary->elem();
2180
2181 // Determine whether the reference is beyond the header or not, by comparing
2182 // the offset against the offset of the start of the array's data.
2183 // Different array types begin at slightly different offsets (12 vs. 16).
2184 // We choose T_BYTE as an example base type that is least restrictive
2185 // as to alignment, which will therefore produce the smallest
2186 // possible base offset.
2187 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
2188 const bool off_beyond_header = (off >= min_base_off);
2189
2190 // Try to constant-fold a stable array element.
2191 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
2192 // Make sure the reference is not into the header and the offset is constant
2193 ciObject* aobj = ary->const_oop();
2194 if (aobj != nullptr && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
2195 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
2196 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
2197 stable_dimension,
2198 memory_type(), is_unsigned());
2199 if (con_type != nullptr) {
2200 return con_type;
2201 }
2202 }
2203 }
2204
2205 // Don't do this for integer types. There is only potential profit if
2206 // the element type t is lower than _type; that is, for int types, if _type is
2207 // more restrictive than t. This only happens here if one is short and the other
2208 // char (both 16 bits), and in those cases we've made an intentional decision
2209 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
2210 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
2211 //
2212 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
2213 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
2214 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
2215 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
2216 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
2217 // In fact, that could have been the original type of p1, and p1 could have
2218 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
2219 // expression (LShiftL quux 3) independently optimized to the constant 8.
2220 if ((t->isa_int() == nullptr) && (t->isa_long() == nullptr)
2221 && (_type->isa_vect() == nullptr)
2222 && !ary->is_flat()
2223 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
2224 // t might actually be lower than _type, if _type is a unique
2225 // concrete subclass of abstract class t.
2226 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
2227 const Type* jt = t->join_speculative(_type);
2228 // In any case, do not allow the join, per se, to empty out the type.
2229 if (jt->empty() && !t->empty()) {
2230 // This can happen if a interface-typed array narrows to a class type.
2231 jt = _type;
2232 }
2233 #ifdef ASSERT
2234 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
2235 // The pointers in the autobox arrays are always non-null
2236 Node* base = adr->in(AddPNode::Base);
2237 if ((base != nullptr) && base->is_DecodeN()) {
2238 // Get LoadN node which loads IntegerCache.cache field
2239 base = base->in(1);
2240 }
2241 if ((base != nullptr) && base->is_Con()) {
2242 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
2243 if ((base_type != nullptr) && base_type->is_autobox_cache()) {
2244 // It could be narrow oop
2245 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
2246 }
2247 }
2248 }
2249 #endif
2250 return jt;
2251 }
2252 }
2253 } else if (tp->base() == Type::InstPtr) {
2254 assert( off != Type::OffsetBot ||
2255 // arrays can be cast to Objects
2256 !tp->isa_instptr() ||
2257 tp->is_instptr()->instance_klass()->is_java_lang_Object() ||
2258 // Default value load
2259 tp->is_instptr()->instance_klass() == ciEnv::current()->Class_klass() ||
2260 // unsafe field access may not have a constant offset
2261 C->has_unsafe_access(),
2262 "Field accesses must be precise" );
2263 // For oop loads, we expect the _type to be precise.
2264
2265 const TypeInstPtr* tinst = tp->is_instptr();
2266 BasicType bt = memory_type();
2267
2268 // Optimize loads from constant fields.
2269 ciObject* const_oop = tinst->const_oop();
2270 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != nullptr && const_oop->is_instance()) {
2271 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt);
2272 if (con_type != nullptr) {
2273 return con_type;
2274 }
2275 }
2276 } else if (tp->base() == Type::KlassPtr || tp->base() == Type::InstKlassPtr || tp->base() == Type::AryKlassPtr) {
2277 assert(off != Type::OffsetBot ||
2278 !tp->isa_instklassptr() ||
2279 // arrays can be cast to Objects
2280 tp->isa_instklassptr()->instance_klass()->is_java_lang_Object() ||
2281 // also allow array-loading from the primary supertype
2282 // array during subtype checks
2283 Opcode() == Op_LoadKlass,
2284 "Field accesses must be precise");
2285 // For klass/static loads, we expect the _type to be precise
2286 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
2287 /* With mirrors being an indirect in the Klass*
2288 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
2289 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
2290 *
2291 * So check the type and klass of the node before the LoadP.
2292 */
2293 Node* adr2 = adr->in(MemNode::Address);
2294 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2295 if (tkls != nullptr && !StressReflectiveCode) {
2296 if (tkls->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2297 ciKlass* klass = tkls->exact_klass();
2298 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2299 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2300 return TypeInstPtr::make(klass->java_mirror());
2301 }
2302 }
2303 }
2304
2305 const TypeKlassPtr *tkls = tp->isa_klassptr();
2306 if (tkls != nullptr) {
2307 if (tkls->is_loaded() && tkls->klass_is_exact()) {
2308 ciKlass* klass = tkls->exact_klass();
2309 // We are loading a field from a Klass metaobject whose identity
2310 // is known at compile time (the type is "exact" or "precise").
2311 // Check for fields we know are maintained as constants by the VM.
2312 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2313 // The field is Klass::_super_check_offset. Return its (constant) value.
2314 // (Folds up type checking code.)
2315 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2316 return TypeInt::make(klass->super_check_offset());
2317 }
2318 if (UseCompactObjectHeaders) { // TODO: Should EnableValhalla also take this path ?
2319 if (tkls->offset() == in_bytes(Klass::prototype_header_offset())) {
2320 // The field is Klass::_prototype_header. Return its (constant) value.
2321 assert(this->Opcode() == Op_LoadX, "must load a proper type from _prototype_header");
2322 return TypeX::make(klass->prototype_header());
2323 }
2324 }
2325 // Compute index into primary_supers array
2326 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2327 // Check for overflowing; use unsigned compare to handle the negative case.
2328 if( depth < ciKlass::primary_super_limit() ) {
2329 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2330 // (Folds up type checking code.)
2331 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2332 ciKlass *ss = klass->super_of_depth(depth);
2333 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2334 }
2335 const Type* aift = load_array_final_field(tkls, klass);
2336 if (aift != nullptr) return aift;
2337 }
2338
2339 // We can still check if we are loading from the primary_supers array at a
2340 // shallow enough depth. Even though the klass is not exact, entries less
2341 // than or equal to its super depth are correct.
2342 if (tkls->is_loaded()) {
2343 ciKlass* klass = nullptr;
2344 if (tkls->isa_instklassptr()) {
2345 klass = tkls->is_instklassptr()->instance_klass();
2346 } else {
2347 int dims;
2348 const Type* inner = tkls->is_aryklassptr()->base_element_type(dims);
2349 if (inner->isa_instklassptr()) {
2350 klass = inner->is_instklassptr()->instance_klass();
2351 klass = ciObjArrayKlass::make(klass, dims);
2352 }
2353 }
2354 if (klass != nullptr) {
2355 // Compute index into primary_supers array
2356 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2357 // Check for overflowing; use unsigned compare to handle the negative case.
2358 if (depth < ciKlass::primary_super_limit() &&
2359 depth <= klass->super_depth()) { // allow self-depth checks to handle self-check case
2360 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2361 // (Folds up type checking code.)
2362 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2363 ciKlass *ss = klass->super_of_depth(depth);
2364 return ss ? TypeKlassPtr::make(ss, Type::trust_interfaces) : TypePtr::NULL_PTR;
2365 }
2366 }
2367 }
2368
2369 // If the type is enough to determine that the thing is not an array,
2370 // we can give the layout_helper a positive interval type.
2371 // This will help short-circuit some reflective code.
2372 if (tkls->offset() == in_bytes(Klass::layout_helper_offset()) &&
2373 tkls->isa_instklassptr() && // not directly typed as an array
2374 !tkls->is_instklassptr()->might_be_an_array() // not the supertype of all T[] (java.lang.Object) or has an interface that is not Serializable or Cloneable
2375 ) {
2376 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2377 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2378 // The key property of this type is that it folds up tests
2379 // for array-ness, since it proves that the layout_helper is positive.
2380 // Thus, a generic value like the basic object layout helper works fine.
2381 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2382 }
2383 }
2384
2385 bool is_vect = (_type->isa_vect() != nullptr);
2386 if (is_instance && !is_vect) {
2387 // If we have an instance type and our memory input is the
2388 // programs's initial memory state, there is no matching store,
2389 // so just return a zero of the appropriate type -
2390 // except if it is vectorized - then we have no zero constant.
2391 Node *mem = in(MemNode::Memory);
2392 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2393 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2394 // TODO 8350865 This is needed for flat array accesses, somehow the memory of the loads bypasses the intrinsic
2395 // Run TestArrays.test6 in Scenario4, we need more tests for this. TestBasicFunctionality::test20 also needs this.
2396 if (tp->isa_aryptr() && tp->is_aryptr()->is_flat() && !UseFieldFlattening) {
2397 return _type;
2398 }
2399 return Type::get_zero_type(_type->basic_type());
2400 }
2401 }
2402 if (!UseCompactObjectHeaders) {
2403 Node* alloc = is_new_object_mark_load();
2404 if (alloc != nullptr) {
2405 if (EnableValhalla) {
2406 // The mark word may contain property bits (inline, flat, null-free)
2407 Node* klass_node = alloc->in(AllocateNode::KlassNode);
2408 const TypeKlassPtr* tkls = phase->type(klass_node)->isa_klassptr();
2409 if (tkls != nullptr && tkls->is_loaded() && tkls->klass_is_exact()) {
2410 return TypeX::make(tkls->exact_klass()->prototype_header());
2411 }
2412 } else {
2413 return TypeX::make(markWord::prototype().value());
2414 }
2415 }
2416 }
2417
2418 return _type;
2419 }
2420
2421 //------------------------------match_edge-------------------------------------
2422 // Do we Match on this edge index or not? Match only the address.
2423 uint LoadNode::match_edge(uint idx) const {
2424 return idx == MemNode::Address;
2425 }
2426
2427 //--------------------------LoadBNode::Ideal--------------------------------------
2428 //
2429 // If the previous store is to the same address as this load,
2430 // and the value stored was larger than a byte, replace this load
2431 // with the value stored truncated to a byte. If no truncation is
2432 // needed, the replacement is done in LoadNode::Identity().
2433 //
2434 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2435 Node* mem = in(MemNode::Memory);
2436 Node* value = can_see_stored_value(mem,phase);
2437 if (value != nullptr) {
2438 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2439 if (narrow != value) {
2440 return narrow;
2441 }
2442 }
2443 // Identity call will handle the case where truncation is not needed.
2444 return LoadNode::Ideal(phase, can_reshape);
2445 }
2446
2447 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2448 Node* mem = in(MemNode::Memory);
2449 Node* value = can_see_stored_value(mem,phase);
2450 if (value != nullptr && value->is_Con() &&
2451 !value->bottom_type()->higher_equal(_type)) {
2452 // If the input to the store does not fit with the load's result type,
2453 // it must be truncated. We can't delay until Ideal call since
2454 // a singleton Value is needed for split_thru_phi optimization.
2455 int con = value->get_int();
2456 return TypeInt::make((con << 24) >> 24);
2457 }
2458 return LoadNode::Value(phase);
2459 }
2460
2461 //--------------------------LoadUBNode::Ideal-------------------------------------
2462 //
2463 // If the previous store is to the same address as this load,
2464 // and the value stored was larger than a byte, replace this load
2465 // with the value stored truncated to a byte. If no truncation is
2466 // needed, the replacement is done in LoadNode::Identity().
2467 //
2468 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2469 Node* mem = in(MemNode::Memory);
2470 Node* value = can_see_stored_value(mem, phase);
2471 if (value != nullptr) {
2472 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2473 if (narrow != value) {
2474 return narrow;
2475 }
2476 }
2477 // Identity call will handle the case where truncation is not needed.
2478 return LoadNode::Ideal(phase, can_reshape);
2479 }
2480
2481 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2482 Node* mem = in(MemNode::Memory);
2483 Node* value = can_see_stored_value(mem,phase);
2484 if (value != nullptr && value->is_Con() &&
2485 !value->bottom_type()->higher_equal(_type)) {
2486 // If the input to the store does not fit with the load's result type,
2487 // it must be truncated. We can't delay until Ideal call since
2488 // a singleton Value is needed for split_thru_phi optimization.
2489 int con = value->get_int();
2490 return TypeInt::make(con & 0xFF);
2491 }
2492 return LoadNode::Value(phase);
2493 }
2494
2495 //--------------------------LoadUSNode::Ideal-------------------------------------
2496 //
2497 // If the previous store is to the same address as this load,
2498 // and the value stored was larger than a char, replace this load
2499 // with the value stored truncated to a char. If no truncation is
2500 // needed, the replacement is done in LoadNode::Identity().
2501 //
2502 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2503 Node* mem = in(MemNode::Memory);
2504 Node* value = can_see_stored_value(mem,phase);
2505 if (value != nullptr) {
2506 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2507 if (narrow != value) {
2508 return narrow;
2509 }
2510 }
2511 // Identity call will handle the case where truncation is not needed.
2512 return LoadNode::Ideal(phase, can_reshape);
2513 }
2514
2515 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2516 Node* mem = in(MemNode::Memory);
2517 Node* value = can_see_stored_value(mem,phase);
2518 if (value != nullptr && value->is_Con() &&
2519 !value->bottom_type()->higher_equal(_type)) {
2520 // If the input to the store does not fit with the load's result type,
2521 // it must be truncated. We can't delay until Ideal call since
2522 // a singleton Value is needed for split_thru_phi optimization.
2523 int con = value->get_int();
2524 return TypeInt::make(con & 0xFFFF);
2525 }
2526 return LoadNode::Value(phase);
2527 }
2528
2529 //--------------------------LoadSNode::Ideal--------------------------------------
2530 //
2531 // If the previous store is to the same address as this load,
2532 // and the value stored was larger than a short, replace this load
2533 // with the value stored truncated to a short. If no truncation is
2534 // needed, the replacement is done in LoadNode::Identity().
2535 //
2536 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2537 Node* mem = in(MemNode::Memory);
2538 Node* value = can_see_stored_value(mem,phase);
2539 if (value != nullptr) {
2540 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2541 if (narrow != value) {
2542 return narrow;
2543 }
2544 }
2545 // Identity call will handle the case where truncation is not needed.
2546 return LoadNode::Ideal(phase, can_reshape);
2547 }
2548
2549 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2550 Node* mem = in(MemNode::Memory);
2551 Node* value = can_see_stored_value(mem,phase);
2552 if (value != nullptr && value->is_Con() &&
2553 !value->bottom_type()->higher_equal(_type)) {
2554 // If the input to the store does not fit with the load's result type,
2555 // it must be truncated. We can't delay until Ideal call since
2556 // a singleton Value is needed for split_thru_phi optimization.
2557 int con = value->get_int();
2558 return TypeInt::make((con << 16) >> 16);
2559 }
2560 return LoadNode::Value(phase);
2561 }
2562
2563 Node* LoadNNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2564 // Loading from an InlineType, find the input and make an EncodeP
2565 Node* addr = in(Address);
2566 intptr_t offset;
2567 Node* base = AddPNode::Ideal_base_and_offset(addr, phase, offset);
2568 Node* value = see_through_inline_type(phase, this, base, offset);
2569 if (value != nullptr) {
2570 return new EncodePNode(value, type());
2571 }
2572
2573 return LoadNode::Ideal(phase, can_reshape);
2574 }
2575
2576 //=============================================================================
2577 //----------------------------LoadKlassNode::make------------------------------
2578 // Polymorphic factory method:
2579 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk, bool fold_for_arrays) {
2580 // sanity check the alias category against the created node type
2581 const TypePtr* adr_type = adr->bottom_type()->isa_ptr();
2582 assert(adr_type != nullptr, "expecting TypeKlassPtr");
2583 #ifdef _LP64
2584 if (adr_type->is_ptr_to_narrowklass()) {
2585 assert(UseCompressedClassPointers, "no compressed klasses");
2586 Node* load_klass = gvn.transform(new LoadNKlassNode(mem, adr, at, tk->make_narrowklass(), MemNode::unordered, fold_for_arrays));
2587 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2588 }
2589 #endif
2590 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2591 return new LoadKlassNode(mem, adr, at, tk, MemNode::unordered, fold_for_arrays);
2592 }
2593
2594 //------------------------------Value------------------------------------------
2595 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2596 return klass_value_common(phase, _fold_for_arrays);
2597 }
2598
2599 const Type* LoadNode::klass_value_common(PhaseGVN* phase, bool fold_for_arrays) const {
2600 // Either input is TOP ==> the result is TOP
2601 const Type *t1 = phase->type( in(MemNode::Memory) );
2602 if (t1 == Type::TOP) return Type::TOP;
2603 Node *adr = in(MemNode::Address);
2604 const Type *t2 = phase->type( adr );
2605 if (t2 == Type::TOP) return Type::TOP;
2606 const TypePtr *tp = t2->is_ptr();
2607 if (TypePtr::above_centerline(tp->ptr()) ||
2608 tp->ptr() == TypePtr::Null) return Type::TOP;
2609
2610 // Return a more precise klass, if possible
2611 const TypeInstPtr *tinst = tp->isa_instptr();
2612 if (tinst != nullptr) {
2613 ciInstanceKlass* ik = tinst->instance_klass();
2614 int offset = tinst->offset();
2615 if (ik == phase->C->env()->Class_klass()
2616 && (offset == java_lang_Class::klass_offset() ||
2617 offset == java_lang_Class::array_klass_offset())) {
2618 // We are loading a special hidden field from a Class mirror object,
2619 // the field which points to the VM's Klass metaobject.
2620 bool is_null_free_array = false;
2621 ciType* t = tinst->java_mirror_type(&is_null_free_array);
2622 // java_mirror_type returns non-null for compile-time Class constants.
2623 if (t != nullptr) {
2624 // constant oop => constant klass
2625 if (offset == java_lang_Class::array_klass_offset()) {
2626 if (t->is_void()) {
2627 // We cannot create a void array. Since void is a primitive type return null
2628 // klass. Users of this result need to do a null check on the returned klass.
2629 return TypePtr::NULL_PTR;
2630 }
2631 const TypeKlassPtr* tklass = TypeKlassPtr::make(ciArrayKlass::make(t), Type::trust_interfaces);
2632 if (is_null_free_array) {
2633 tklass = tklass->is_aryklassptr()->cast_to_null_free();
2634 }
2635 return tklass;
2636 }
2637 if (!t->is_klass()) {
2638 // a primitive Class (e.g., int.class) has null for a klass field
2639 return TypePtr::NULL_PTR;
2640 }
2641 // Fold up the load of the hidden field
2642 const TypeKlassPtr* tklass = TypeKlassPtr::make(t->as_klass(), Type::trust_interfaces);
2643 if (is_null_free_array) {
2644 tklass = tklass->is_aryklassptr()->cast_to_null_free();
2645 }
2646 return tklass;
2647 }
2648 // non-constant mirror, so we can't tell what's going on
2649 }
2650 if (!tinst->is_loaded())
2651 return _type; // Bail out if not loaded
2652 if (offset == oopDesc::klass_offset_in_bytes()) {
2653 return tinst->as_klass_type(true);
2654 }
2655 }
2656
2657 // Check for loading klass from an array
2658 const TypeAryPtr* tary = tp->isa_aryptr();
2659 if (tary != nullptr && fold_for_arrays &&
2660 tary->offset() == oopDesc::klass_offset_in_bytes()) {
2661 return tary->as_klass_type(true);
2662 }
2663
2664 // Check for loading klass from an array klass
2665 const TypeKlassPtr *tkls = tp->isa_klassptr();
2666 if (tkls != nullptr && !StressReflectiveCode) {
2667 if (!tkls->is_loaded())
2668 return _type; // Bail out if not loaded
2669 if (tkls->isa_aryklassptr() && tkls->is_aryklassptr()->elem()->isa_klassptr() &&
2670 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2671 // // Always returning precise element type is incorrect,
2672 // // e.g., element type could be object and array may contain strings
2673 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2674
2675 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2676 // according to the element type's subclassing.
2677 return tkls->is_aryklassptr()->elem()->isa_klassptr()->cast_to_exactness(tkls->klass_is_exact());
2678 }
2679 if (tkls->isa_instklassptr() != nullptr && tkls->klass_is_exact() &&
2680 tkls->offset() == in_bytes(Klass::super_offset())) {
2681 ciKlass* sup = tkls->is_instklassptr()->instance_klass()->super();
2682 // The field is Klass::_super. Return its (constant) value.
2683 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2684 return sup ? TypeKlassPtr::make(sup, Type::trust_interfaces) : TypePtr::NULL_PTR;
2685 }
2686 }
2687
2688 if (tkls != nullptr && !UseSecondarySupersCache
2689 && tkls->offset() == in_bytes(Klass::secondary_super_cache_offset())) {
2690 // Treat Klass::_secondary_super_cache as a constant when the cache is disabled.
2691 return TypePtr::NULL_PTR;
2692 }
2693
2694 // Bailout case
2695 return LoadNode::Value(phase);
2696 }
2697
2698 //------------------------------Identity---------------------------------------
2699 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2700 // Also feed through the klass in Allocate(...klass...)._klass.
2701 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2702 return klass_identity_common(phase);
2703 }
2704
2705 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2706 Node* x = LoadNode::Identity(phase);
2707 if (x != this) return x;
2708
2709 // Take apart the address into an oop and offset.
2710 // Return 'this' if we cannot.
2711 Node* adr = in(MemNode::Address);
2712 intptr_t offset = 0;
2713 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2714 if (base == nullptr) return this;
2715 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2716 if (toop == nullptr) return this;
2717
2718 // Step over potential GC barrier for OopHandle resolve
2719 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2720 if (bs->is_gc_barrier_node(base)) {
2721 base = bs->step_over_gc_barrier(base);
2722 }
2723
2724 // We can fetch the klass directly through an AllocateNode.
2725 // This works even if the klass is not constant (clone or newArray).
2726 if (offset == oopDesc::klass_offset_in_bytes()) {
2727 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2728 if (allocated_klass != nullptr) {
2729 return allocated_klass;
2730 }
2731 }
2732
2733 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2734 // See inline_native_Class_query for occurrences of these patterns.
2735 // Java Example: x.getClass().isAssignableFrom(y)
2736 //
2737 // This improves reflective code, often making the Class
2738 // mirror go completely dead. (Current exception: Class
2739 // mirrors may appear in debug info, but we could clean them out by
2740 // introducing a new debug info operator for Klass.java_mirror).
2741
2742 if (toop->isa_instptr() && toop->is_instptr()->instance_klass() == phase->C->env()->Class_klass()
2743 && offset == java_lang_Class::klass_offset()) {
2744 if (base->is_Load()) {
2745 Node* base2 = base->in(MemNode::Address);
2746 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2747 Node* adr2 = base2->in(MemNode::Address);
2748 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2749 if (tkls != nullptr && !tkls->empty()
2750 && (tkls->isa_instklassptr() || tkls->isa_aryklassptr())
2751 && adr2->is_AddP()
2752 ) {
2753 int mirror_field = in_bytes(Klass::java_mirror_offset());
2754 if (tkls->offset() == mirror_field) {
2755 return adr2->in(AddPNode::Base);
2756 }
2757 }
2758 }
2759 }
2760 }
2761
2762 return this;
2763 }
2764
2765 LoadNode* LoadNode::clone_pinned() const {
2766 LoadNode* ld = clone()->as_Load();
2767 ld->_control_dependency = UnknownControl;
2768 return ld;
2769 }
2770
2771
2772 //------------------------------Value------------------------------------------
2773 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2774 const Type *t = klass_value_common(phase, _fold_for_arrays);
2775 if (t == Type::TOP)
2776 return t;
2777
2778 return t->make_narrowklass();
2779 }
2780
2781 //------------------------------Identity---------------------------------------
2782 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2783 // Also feed through the klass in Allocate(...klass...)._klass.
2784 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2785 Node *x = klass_identity_common(phase);
2786
2787 const Type *t = phase->type( x );
2788 if( t == Type::TOP ) return x;
2789 if( t->isa_narrowklass()) return x;
2790 assert (!t->isa_narrowoop(), "no narrow oop here");
2791
2792 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2793 }
2794
2795 //------------------------------Value-----------------------------------------
2796 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2797 // Either input is TOP ==> the result is TOP
2798 const Type *t1 = phase->type( in(MemNode::Memory) );
2799 if( t1 == Type::TOP ) return Type::TOP;
2800 Node *adr = in(MemNode::Address);
2801 const Type *t2 = phase->type( adr );
2802 if( t2 == Type::TOP ) return Type::TOP;
2803 const TypePtr *tp = t2->is_ptr();
2804 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2805 const TypeAryPtr *tap = tp->isa_aryptr();
2806 if( !tap ) return _type;
2807 return tap->size();
2808 }
2809
2810 //-------------------------------Ideal---------------------------------------
2811 // Feed through the length in AllocateArray(...length...)._length.
2812 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2813 Node* p = MemNode::Ideal_common(phase, can_reshape);
2814 if (p) return (p == NodeSentinel) ? nullptr : p;
2815
2816 // Take apart the address into an oop and offset.
2817 // Return 'this' if we cannot.
2818 Node* adr = in(MemNode::Address);
2819 intptr_t offset = 0;
2820 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2821 if (base == nullptr) return nullptr;
2822 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2823 if (tary == nullptr) return nullptr;
2824
2825 // We can fetch the length directly through an AllocateArrayNode.
2826 // This works even if the length is not constant (clone or newArray).
2827 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2828 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2829 if (alloc != nullptr) {
2830 Node* allocated_length = alloc->Ideal_length();
2831 Node* len = alloc->make_ideal_length(tary, phase);
2832 if (allocated_length != len) {
2833 // New CastII improves on this.
2834 return len;
2835 }
2836 }
2837 }
2838
2839 return nullptr;
2840 }
2841
2842 //------------------------------Identity---------------------------------------
2843 // Feed through the length in AllocateArray(...length...)._length.
2844 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2845 Node* x = LoadINode::Identity(phase);
2846 if (x != this) return x;
2847
2848 // Take apart the address into an oop and offset.
2849 // Return 'this' if we cannot.
2850 Node* adr = in(MemNode::Address);
2851 intptr_t offset = 0;
2852 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2853 if (base == nullptr) return this;
2854 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2855 if (tary == nullptr) return this;
2856
2857 // We can fetch the length directly through an AllocateArrayNode.
2858 // This works even if the length is not constant (clone or newArray).
2859 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2860 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base);
2861 if (alloc != nullptr) {
2862 Node* allocated_length = alloc->Ideal_length();
2863 // Do not allow make_ideal_length to allocate a CastII node.
2864 Node* len = alloc->make_ideal_length(tary, phase, false);
2865 if (allocated_length == len) {
2866 // Return allocated_length only if it would not be improved by a CastII.
2867 return allocated_length;
2868 }
2869 }
2870 }
2871
2872 return this;
2873
2874 }
2875
2876 //=============================================================================
2877 //---------------------------StoreNode::make-----------------------------------
2878 // Polymorphic factory method:
2879 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) {
2880 assert((mo == unordered || mo == release), "unexpected");
2881 Compile* C = gvn.C;
2882 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2883 ctl != nullptr, "raw memory operations should have control edge");
2884
2885 switch (bt) {
2886 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2887 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2888 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2889 case T_CHAR:
2890 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2891 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2892 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2893 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2894 case T_METADATA:
2895 case T_ADDRESS:
2896 case T_OBJECT:
2897 #ifdef _LP64
2898 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2899 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2900 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2901 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2902 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2903 adr->bottom_type()->isa_rawptr())) {
2904 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2905 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2906 }
2907 #endif
2908 {
2909 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2910 }
2911 default:
2912 ShouldNotReachHere();
2913 return (StoreNode*)nullptr;
2914 }
2915 }
2916
2917 //--------------------------bottom_type----------------------------------------
2918 const Type *StoreNode::bottom_type() const {
2919 return Type::MEMORY;
2920 }
2921
2922 //------------------------------hash-------------------------------------------
2923 uint StoreNode::hash() const {
2924 // unroll addition of interesting fields
2925 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2926
2927 // Since they are not commoned, do not hash them:
2928 return NO_HASH;
2929 }
2930
2931 // Link together multiple stores (B/S/C/I) into a longer one.
2932 //
2933 // Example: _store = StoreB[i+3]
2934 //
2935 // RangeCheck[i+0] RangeCheck[i+0]
2936 // StoreB[i+0]
2937 // RangeCheck[i+3] RangeCheck[i+3]
2938 // StoreB[i+1] --> pass: fail:
2939 // StoreB[i+2] StoreI[i+0] StoreB[i+0]
2940 // StoreB[i+3]
2941 //
2942 // The 4 StoreB are merged into a single StoreI node. We have to be careful with RangeCheck[i+1]: before
2943 // the optimization, if this RangeCheck[i+1] fails, then we execute only StoreB[i+0], and then trap. After
2944 // the optimization, the new StoreI[i+0] is on the passing path of RangeCheck[i+3], and StoreB[i+0] on the
2945 // failing path.
2946 //
2947 // Note: For normal array stores, every store at first has a RangeCheck. But they can be removed with:
2948 // - RCE (RangeCheck Elimination): the RangeChecks in the loop are hoisted out and before the loop,
2949 // and possibly no RangeChecks remain between the stores.
2950 // - RangeCheck smearing: the earlier RangeChecks are adjusted such that they cover later RangeChecks,
2951 // and those later RangeChecks can be removed. Example:
2952 //
2953 // RangeCheck[i+0] RangeCheck[i+0] <- before first store
2954 // StoreB[i+0] StoreB[i+0] <- first store
2955 // RangeCheck[i+1] --> smeared --> RangeCheck[i+3] <- only RC between first and last store
2956 // StoreB[i+1] StoreB[i+1] <- second store
2957 // RangeCheck[i+2] --> removed
2958 // StoreB[i+2] StoreB[i+2]
2959 // RangeCheck[i+3] --> removed
2960 // StoreB[i+3] StoreB[i+3] <- last store
2961 //
2962 // Thus, it is a common pattern that between the first and last store in a chain
2963 // of adjacent stores there remains exactly one RangeCheck, located between the
2964 // first and the second store (e.g. RangeCheck[i+3]).
2965 //
2966 class MergePrimitiveStores : public StackObj {
2967 private:
2968 PhaseGVN* const _phase;
2969 StoreNode* const _store;
2970 // State machine with initial state Unknown
2971 // Allowed transitions:
2972 // Unknown -> Const
2973 // Unknown -> Platform
2974 // Unknown -> Reverse
2975 // Unknown -> NotAdjacent
2976 // Const -> Const
2977 // Const -> NotAdjacent
2978 // Platform -> Platform
2979 // Platform -> NotAdjacent
2980 // Reverse -> Reverse
2981 // Reverse -> NotAdjacent
2982 // NotAdjacent -> NotAdjacent
2983 enum ValueOrder : uint8_t {
2984 Unknown, // Initial state
2985 Const, // Input values are const
2986 Platform, // Platform order
2987 Reverse, // Reverse platform order
2988 NotAdjacent // Not adjacent
2989 };
2990 ValueOrder _value_order;
2991
2992 NOT_PRODUCT( const CHeapBitMap &_trace_tags; )
2993
2994 public:
2995 MergePrimitiveStores(PhaseGVN* phase, StoreNode* store) :
2996 _phase(phase), _store(store), _value_order(ValueOrder::Unknown)
2997 NOT_PRODUCT( COMMA _trace_tags(Compile::current()->directive()->trace_merge_stores_tags()) )
2998 {}
2999
3000 StoreNode* run();
3001
3002 private:
3003 bool is_compatible_store(const StoreNode* other_store) const;
3004 bool is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const;
3005 bool is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const;
3006 static bool is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase);
3007 enum CFGStatus { SuccessNoRangeCheck, SuccessWithRangeCheck, Failure };
3008 static CFGStatus cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store);
3009
3010 class Status {
3011 private:
3012 StoreNode* _found_store;
3013 bool _found_range_check;
3014
3015 Status(StoreNode* found_store, bool found_range_check)
3016 : _found_store(found_store), _found_range_check(found_range_check) {}
3017
3018 public:
3019 StoreNode* found_store() const { return _found_store; }
3020 bool found_range_check() const { return _found_range_check; }
3021 static Status make_failure() { return Status(nullptr, false); }
3022
3023 static Status make(StoreNode* found_store, const CFGStatus cfg_status) {
3024 if (cfg_status == CFGStatus::Failure) {
3025 return Status::make_failure();
3026 }
3027 return Status(found_store, cfg_status == CFGStatus::SuccessWithRangeCheck);
3028 }
3029
3030 #ifndef PRODUCT
3031 void print_on(outputStream* st) const {
3032 if (_found_store == nullptr) {
3033 st->print_cr("None");
3034 } else {
3035 st->print_cr("Found[%d %s, %s]", _found_store->_idx, _found_store->Name(),
3036 _found_range_check ? "RC" : "no-RC");
3037 }
3038 }
3039 #endif
3040 };
3041
3042 enum ValueOrder find_adjacent_input_value_order(const Node* n1, const Node* n2, const int memory_size) const;
3043 Status find_adjacent_use_store(const StoreNode* def_store) const;
3044 Status find_adjacent_def_store(const StoreNode* use_store) const;
3045 Status find_use_store(const StoreNode* def_store) const;
3046 Status find_def_store(const StoreNode* use_store) const;
3047 Status find_use_store_unidirectional(const StoreNode* def_store) const;
3048 Status find_def_store_unidirectional(const StoreNode* use_store) const;
3049
3050 void collect_merge_list(Node_List& merge_list) const;
3051 Node* make_merged_input_value(const Node_List& merge_list);
3052 StoreNode* make_merged_store(const Node_List& merge_list, Node* merged_input_value);
3053
3054 #ifndef PRODUCT
3055 // Access to TraceMergeStores tags
3056 bool is_trace(TraceMergeStores::Tag tag) const {
3057 return _trace_tags.at(tag);
3058 }
3059
3060 bool is_trace_basic() const {
3061 return is_trace(TraceMergeStores::Tag::BASIC);
3062 }
3063
3064 bool is_trace_pointer_parsing() const {
3065 return is_trace(TraceMergeStores::Tag::POINTER_PARSING);
3066 }
3067
3068 bool is_trace_pointer_aliasing() const {
3069 return is_trace(TraceMergeStores::Tag::POINTER_ALIASING);
3070 }
3071
3072 bool is_trace_pointer_adjacency() const {
3073 return is_trace(TraceMergeStores::Tag::POINTER_ADJACENCY);
3074 }
3075
3076 bool is_trace_success() const {
3077 return is_trace(TraceMergeStores::Tag::SUCCESS);
3078 }
3079 #endif
3080
3081 NOT_PRODUCT( void trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const; )
3082 };
3083
3084 StoreNode* MergePrimitiveStores::run() {
3085 // Check for B/S/C/I
3086 int opc = _store->Opcode();
3087 if (opc != Op_StoreB && opc != Op_StoreC && opc != Op_StoreI) {
3088 return nullptr;
3089 }
3090
3091 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] MergePrimitiveStores::run: "); _store->dump(); })
3092
3093 // The _store must be the "last" store in a chain. If we find a use we could merge with
3094 // then that use or a store further down is the "last" store.
3095 Status status_use = find_adjacent_use_store(_store);
3096 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect no use: "); status_use.print_on(tty); })
3097 if (status_use.found_store() != nullptr) {
3098 return nullptr;
3099 }
3100
3101 // Check if we can merge with at least one def, so that we have at least 2 stores to merge.
3102 Status status_def = find_adjacent_def_store(_store);
3103 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] expect def: "); status_def.print_on(tty); })
3104 Node* def_store = status_def.found_store();
3105 if (def_store == nullptr) {
3106 return nullptr;
3107 }
3108
3109 // Initialize value order
3110 _value_order = find_adjacent_input_value_order(def_store->in(MemNode::ValueIn),
3111 _store->in(MemNode::ValueIn),
3112 _store->memory_size());
3113 assert(_value_order != ValueOrder::NotAdjacent && _value_order != ValueOrder::Unknown, "Order should be checked");
3114
3115 ResourceMark rm;
3116 Node_List merge_list;
3117 collect_merge_list(merge_list);
3118
3119 Node* merged_input_value = make_merged_input_value(merge_list);
3120 if (merged_input_value == nullptr) { return nullptr; }
3121
3122 StoreNode* merged_store = make_merged_store(merge_list, merged_input_value);
3123
3124 NOT_PRODUCT( if (is_trace_success()) { trace(merge_list, merged_input_value, merged_store); } )
3125
3126 return merged_store;
3127 }
3128
3129 // Check compatibility between _store and other_store.
3130 bool MergePrimitiveStores::is_compatible_store(const StoreNode* other_store) const {
3131 int opc = _store->Opcode();
3132 assert(opc == Op_StoreB || opc == Op_StoreC || opc == Op_StoreI, "precondition");
3133
3134 if (other_store == nullptr ||
3135 _store->Opcode() != other_store->Opcode()) {
3136 return false;
3137 }
3138
3139 return true;
3140 }
3141
3142 bool MergePrimitiveStores::is_adjacent_pair(const StoreNode* use_store, const StoreNode* def_store) const {
3143 if (!is_adjacent_input_pair(def_store->in(MemNode::ValueIn),
3144 use_store->in(MemNode::ValueIn),
3145 def_store->memory_size())) {
3146 return false;
3147 }
3148
3149 ResourceMark rm;
3150 #ifndef PRODUCT
3151 const TraceMemPointer trace(is_trace_pointer_parsing(),
3152 is_trace_pointer_aliasing(),
3153 is_trace_pointer_adjacency(),
3154 true);
3155 #endif
3156 const MemPointer pointer_use(use_store NOT_PRODUCT(COMMA trace));
3157 const MemPointer pointer_def(def_store NOT_PRODUCT(COMMA trace));
3158 return pointer_def.is_adjacent_to_and_before(pointer_use);
3159 }
3160
3161 // Check input values n1 and n2 can be merged and return the value order
3162 MergePrimitiveStores::ValueOrder MergePrimitiveStores::find_adjacent_input_value_order(const Node* n1, const Node* n2,
3163 const int memory_size) const {
3164 // Pattern: [n1 = ConI, n2 = ConI]
3165 if (n1->Opcode() == Op_ConI && n2->Opcode() == Op_ConI) {
3166 return ValueOrder::Const;
3167 }
3168
3169 Node const *base_n2;
3170 jint shift_n2;
3171 if (!is_con_RShift(n2, base_n2, shift_n2, _phase)) {
3172 return ValueOrder::NotAdjacent;
3173 }
3174 Node const *base_n1;
3175 jint shift_n1;
3176 if (!is_con_RShift(n1, base_n1, shift_n1, _phase)) {
3177 return ValueOrder::NotAdjacent;
3178 }
3179
3180 int bits_per_store = memory_size * 8;
3181 if (base_n1 != base_n2 ||
3182 abs(shift_n1 - shift_n2) != bits_per_store ||
3183 shift_n1 % bits_per_store != 0) {
3184 // Values are not adjacent
3185 return ValueOrder::NotAdjacent;
3186 }
3187
3188 // Detect value order
3189 #ifdef VM_LITTLE_ENDIAN
3190 return shift_n1 < shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3191 : ValueOrder::Reverse; // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3192 #else
3193 return shift_n1 > shift_n2 ? ValueOrder::Platform // Pattern: [n1 = base >> (shift + memory_size), n2 = base >> shift]
3194 : ValueOrder::Reverse; // Pattern: [n1 = base >> shift, n2 = base >> (shift + memory_size)]
3195 #endif
3196 }
3197
3198 bool MergePrimitiveStores::is_adjacent_input_pair(const Node* n1, const Node* n2, const int memory_size) const {
3199 ValueOrder input_value_order = find_adjacent_input_value_order(n1, n2, memory_size);
3200
3201 switch (input_value_order) {
3202 case ValueOrder::NotAdjacent:
3203 return false;
3204 case ValueOrder::Reverse:
3205 if (memory_size != 1 ||
3206 !Matcher::match_rule_supported(Op_ReverseBytesS) ||
3207 !Matcher::match_rule_supported(Op_ReverseBytesI) ||
3208 !Matcher::match_rule_supported(Op_ReverseBytesL)) {
3209 // ReverseBytes are not supported by platform
3210 return false;
3211 }
3212 // fall-through.
3213 case ValueOrder::Const:
3214 case ValueOrder::Platform:
3215 if (_value_order == ValueOrder::Unknown) {
3216 // Initial state is Unknown, and we find a valid input value order
3217 return true;
3218 }
3219 // The value order can not be changed
3220 return _value_order == input_value_order;
3221 case ValueOrder::Unknown:
3222 default:
3223 ShouldNotReachHere();
3224 }
3225 return false;
3226 }
3227
3228 // Detect pattern: n = base_out >> shift_out
3229 bool MergePrimitiveStores::is_con_RShift(const Node* n, Node const*& base_out, jint& shift_out, PhaseGVN* phase) {
3230 assert(n != nullptr, "precondition");
3231
3232 int opc = n->Opcode();
3233 if (opc == Op_ConvL2I) {
3234 n = n->in(1);
3235 opc = n->Opcode();
3236 }
3237
3238 if ((opc == Op_RShiftI ||
3239 opc == Op_RShiftL ||
3240 opc == Op_URShiftI ||
3241 opc == Op_URShiftL) &&
3242 n->in(2)->is_ConI()) {
3243 base_out = n->in(1);
3244 shift_out = n->in(2)->get_int();
3245 // The shift must be positive:
3246 return shift_out >= 0;
3247 }
3248
3249 if (phase->type(n)->isa_int() != nullptr ||
3250 phase->type(n)->isa_long() != nullptr) {
3251 // (base >> 0)
3252 base_out = n;
3253 shift_out = 0;
3254 return true;
3255 }
3256 return false;
3257 }
3258
3259 // Check if there is nothing between the two stores, except optionally a RangeCheck leading to an uncommon trap.
3260 MergePrimitiveStores::CFGStatus MergePrimitiveStores::cfg_status_for_pair(const StoreNode* use_store, const StoreNode* def_store) {
3261 assert(use_store->in(MemNode::Memory) == def_store, "use-def relationship");
3262
3263 Node* ctrl_use = use_store->in(MemNode::Control);
3264 Node* ctrl_def = def_store->in(MemNode::Control);
3265 if (ctrl_use == nullptr || ctrl_def == nullptr) {
3266 return CFGStatus::Failure;
3267 }
3268
3269 if (ctrl_use == ctrl_def) {
3270 // Same ctrl -> no RangeCheck in between.
3271 // Check: use_store must be the only use of def_store.
3272 if (def_store->outcnt() > 1) {
3273 return CFGStatus::Failure;
3274 }
3275 return CFGStatus::SuccessNoRangeCheck;
3276 }
3277
3278 // Different ctrl -> could have RangeCheck in between.
3279 // Check: 1. def_store only has these uses: use_store and MergeMem for uncommon trap, and
3280 // 2. ctrl separated by RangeCheck.
3281 if (def_store->outcnt() != 2) {
3282 return CFGStatus::Failure; // Cannot have exactly these uses: use_store and MergeMem for uncommon trap.
3283 }
3284 int use_store_out_idx = def_store->raw_out(0) == use_store ? 0 : 1;
3285 Node* merge_mem = def_store->raw_out(1 - use_store_out_idx)->isa_MergeMem();
3286 if (merge_mem == nullptr ||
3287 merge_mem->outcnt() != 1) {
3288 return CFGStatus::Failure; // Does not have MergeMem for uncommon trap.
3289 }
3290 if (!ctrl_use->is_IfProj() ||
3291 !ctrl_use->in(0)->is_RangeCheck() ||
3292 ctrl_use->in(0)->outcnt() != 2) {
3293 return CFGStatus::Failure; // Not RangeCheck.
3294 }
3295 ProjNode* other_proj = ctrl_use->as_IfProj()->other_if_proj();
3296 Node* trap = other_proj->is_uncommon_trap_proj(Deoptimization::Reason_range_check);
3297 if (trap != merge_mem->unique_out() ||
3298 ctrl_use->in(0)->in(0) != ctrl_def) {
3299 return CFGStatus::Failure; // Not RangeCheck with merge_mem leading to uncommon trap.
3300 }
3301
3302 return CFGStatus::SuccessWithRangeCheck;
3303 }
3304
3305 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_use_store(const StoreNode* def_store) const {
3306 Status status_use = find_use_store(def_store);
3307 StoreNode* use_store = status_use.found_store();
3308 if (use_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3309 return Status::make_failure();
3310 }
3311 return status_use;
3312 }
3313
3314 MergePrimitiveStores::Status MergePrimitiveStores::find_adjacent_def_store(const StoreNode* use_store) const {
3315 Status status_def = find_def_store(use_store);
3316 StoreNode* def_store = status_def.found_store();
3317 if (def_store != nullptr && !is_adjacent_pair(use_store, def_store)) {
3318 return Status::make_failure();
3319 }
3320 return status_def;
3321 }
3322
3323 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store(const StoreNode* def_store) const {
3324 Status status_use = find_use_store_unidirectional(def_store);
3325
3326 #ifdef ASSERT
3327 StoreNode* use_store = status_use.found_store();
3328 if (use_store != nullptr) {
3329 Status status_def = find_def_store_unidirectional(use_store);
3330 assert(status_def.found_store() == def_store &&
3331 status_def.found_range_check() == status_use.found_range_check(),
3332 "find_use_store and find_def_store must be symmetric");
3333 }
3334 #endif
3335
3336 return status_use;
3337 }
3338
3339 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store(const StoreNode* use_store) const {
3340 Status status_def = find_def_store_unidirectional(use_store);
3341
3342 #ifdef ASSERT
3343 StoreNode* def_store = status_def.found_store();
3344 if (def_store != nullptr) {
3345 Status status_use = find_use_store_unidirectional(def_store);
3346 assert(status_use.found_store() == use_store &&
3347 status_use.found_range_check() == status_def.found_range_check(),
3348 "find_use_store and find_def_store must be symmetric");
3349 }
3350 #endif
3351
3352 return status_def;
3353 }
3354
3355 MergePrimitiveStores::Status MergePrimitiveStores::find_use_store_unidirectional(const StoreNode* def_store) const {
3356 assert(is_compatible_store(def_store), "precondition: must be compatible with _store");
3357
3358 for (DUIterator_Fast imax, i = def_store->fast_outs(imax); i < imax; i++) {
3359 StoreNode* use_store = def_store->fast_out(i)->isa_Store();
3360 if (is_compatible_store(use_store)) {
3361 return Status::make(use_store, cfg_status_for_pair(use_store, def_store));
3362 }
3363 }
3364
3365 return Status::make_failure();
3366 }
3367
3368 MergePrimitiveStores::Status MergePrimitiveStores::find_def_store_unidirectional(const StoreNode* use_store) const {
3369 assert(is_compatible_store(use_store), "precondition: must be compatible with _store");
3370
3371 StoreNode* def_store = use_store->in(MemNode::Memory)->isa_Store();
3372 if (!is_compatible_store(def_store)) {
3373 return Status::make_failure();
3374 }
3375
3376 return Status::make(def_store, cfg_status_for_pair(use_store, def_store));
3377 }
3378
3379 void MergePrimitiveStores::collect_merge_list(Node_List& merge_list) const {
3380 // The merged store can be at most 8 bytes.
3381 const uint merge_list_max_size = 8 / _store->memory_size();
3382 assert(merge_list_max_size >= 2 &&
3383 merge_list_max_size <= 8 &&
3384 is_power_of_2(merge_list_max_size),
3385 "must be 2, 4 or 8");
3386
3387 // Traverse up the chain of adjacent def stores.
3388 StoreNode* current = _store;
3389 merge_list.push(current);
3390 while (current != nullptr && merge_list.size() < merge_list_max_size) {
3391 Status status = find_adjacent_def_store(current);
3392 NOT_PRODUCT( if (is_trace_basic()) { tty->print("[TraceMergeStores] find def: "); status.print_on(tty); })
3393
3394 current = status.found_store();
3395 if (current != nullptr) {
3396 merge_list.push(current);
3397
3398 // We can have at most one RangeCheck.
3399 if (status.found_range_check()) {
3400 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found RangeCheck, stop traversal."); })
3401 break;
3402 }
3403 }
3404 }
3405
3406 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] found:"); merge_list.dump(); })
3407
3408 // Truncate the merge_list to a power of 2.
3409 const uint pow2size = round_down_power_of_2(merge_list.size());
3410 assert(pow2size >= 2, "must be merging at least 2 stores");
3411 while (merge_list.size() > pow2size) { merge_list.pop(); }
3412
3413 NOT_PRODUCT( if (is_trace_basic()) { tty->print_cr("[TraceMergeStores] truncated:"); merge_list.dump(); })
3414 }
3415
3416 // Merge the input values of the smaller stores to a single larger input value.
3417 Node* MergePrimitiveStores::make_merged_input_value(const Node_List& merge_list) {
3418 int new_memory_size = _store->memory_size() * merge_list.size();
3419 Node* first = merge_list.at(merge_list.size()-1);
3420 Node* merged_input_value = nullptr;
3421 if (_store->in(MemNode::ValueIn)->Opcode() == Op_ConI) {
3422 assert(_value_order == ValueOrder::Const, "must match");
3423 // Pattern: [ConI, ConI, ...] -> new constant
3424 jlong con = 0;
3425 jlong bits_per_store = _store->memory_size() * 8;
3426 jlong mask = (((jlong)1) << bits_per_store) - 1;
3427 for (uint i = 0; i < merge_list.size(); i++) {
3428 jlong con_i = merge_list.at(i)->in(MemNode::ValueIn)->get_int();
3429 #ifdef VM_LITTLE_ENDIAN
3430 con = con << bits_per_store;
3431 con = con | (mask & con_i);
3432 #else // VM_LITTLE_ENDIAN
3433 con_i = (mask & con_i) << (i * bits_per_store);
3434 con = con | con_i;
3435 #endif // VM_LITTLE_ENDIAN
3436 }
3437 merged_input_value = _phase->longcon(con);
3438 } else {
3439 assert(_value_order == ValueOrder::Platform || _value_order == ValueOrder::Reverse, "must match");
3440 // Pattern: [base >> 24, base >> 16, base >> 8, base] -> base
3441 // | |
3442 // _store first
3443 //
3444 Node* hi = _store->in(MemNode::ValueIn);
3445 Node* lo = first->in(MemNode::ValueIn);
3446 #ifndef VM_LITTLE_ENDIAN
3447 // `_store` and `first` are swapped in the diagram above
3448 swap(hi, lo);
3449 #endif // !VM_LITTLE_ENDIAN
3450 if (_value_order == ValueOrder::Reverse) {
3451 swap(hi, lo);
3452 }
3453 Node const* hi_base;
3454 jint hi_shift;
3455 merged_input_value = lo;
3456 bool is_true = is_con_RShift(hi, hi_base, hi_shift, _phase);
3457 assert(is_true, "must detect con RShift");
3458 if (merged_input_value != hi_base && merged_input_value->Opcode() == Op_ConvL2I) {
3459 // look through
3460 merged_input_value = merged_input_value->in(1);
3461 }
3462 if (merged_input_value != hi_base) {
3463 // merged_input_value is not the base
3464 return nullptr;
3465 }
3466 }
3467
3468 if (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size <= 4) {
3469 // Example:
3470 //
3471 // long base = ...;
3472 // a[0] = (byte)(base >> 0);
3473 // a[1] = (byte)(base >> 8);
3474 //
3475 merged_input_value = _phase->transform(new ConvL2INode(merged_input_value));
3476 }
3477
3478 assert((_phase->type(merged_input_value)->isa_int() != nullptr && new_memory_size <= 4) ||
3479 (_phase->type(merged_input_value)->isa_long() != nullptr && new_memory_size == 8),
3480 "merged_input_value is either int or long, and new_memory_size is small enough");
3481
3482 if (_value_order == ValueOrder::Reverse) {
3483 assert(_store->memory_size() == 1, "only implemented for bytes");
3484 if (new_memory_size == 8) {
3485 merged_input_value = _phase->transform(new ReverseBytesLNode(merged_input_value));
3486 } else if (new_memory_size == 4) {
3487 merged_input_value = _phase->transform(new ReverseBytesINode(merged_input_value));
3488 } else {
3489 assert(new_memory_size == 2, "sanity check");
3490 merged_input_value = _phase->transform(new ReverseBytesSNode(merged_input_value));
3491 }
3492 }
3493 return merged_input_value;
3494 }
3495
3496 // //
3497 // first_ctrl first_mem first_adr first_ctrl first_mem first_adr //
3498 // | | | | | | //
3499 // | | | | +---------------+ | //
3500 // | | | | | | | //
3501 // | | +---------+ | | +---------------+ //
3502 // | | | | | | | | //
3503 // +--------------+ | | v1 +------------------------------+ | | v1 //
3504 // | | | | | | | | | | | | //
3505 // RangeCheck first_store RangeCheck | | first_store //
3506 // | | | | | | | //
3507 // last_ctrl | +----> unc_trap last_ctrl | | +----> unc_trap //
3508 // | | ===> | | | //
3509 // +--------------+ | a2 v2 | | | //
3510 // | | | | | | | | //
3511 // | second_store | | | //
3512 // | | | | | [v1 v2 ... vn] //
3513 // ... ... | | | | //
3514 // | | | | | v //
3515 // +--------------+ | an vn +--------------+ | | merged_input_value //
3516 // | | | | | | | | //
3517 // last_store (= _store) merged_store //
3518 // //
3519 StoreNode* MergePrimitiveStores::make_merged_store(const Node_List& merge_list, Node* merged_input_value) {
3520 Node* first_store = merge_list.at(merge_list.size()-1);
3521 Node* last_ctrl = _store->in(MemNode::Control); // after (optional) RangeCheck
3522 Node* first_mem = first_store->in(MemNode::Memory);
3523 Node* first_adr = first_store->in(MemNode::Address);
3524
3525 const TypePtr* new_adr_type = _store->adr_type();
3526
3527 int new_memory_size = _store->memory_size() * merge_list.size();
3528 BasicType bt = T_ILLEGAL;
3529 switch (new_memory_size) {
3530 case 2: bt = T_SHORT; break;
3531 case 4: bt = T_INT; break;
3532 case 8: bt = T_LONG; break;
3533 }
3534
3535 StoreNode* merged_store = StoreNode::make(*_phase, last_ctrl, first_mem, first_adr,
3536 new_adr_type, merged_input_value, bt, MemNode::unordered);
3537
3538 // Marking the store mismatched is sufficient to prevent reordering, since array stores
3539 // are all on the same slice. Hence, we need no barriers.
3540 merged_store->set_mismatched_access();
3541
3542 // Constants above may now also be be packed -> put candidate on worklist
3543 _phase->is_IterGVN()->_worklist.push(first_mem);
3544
3545 return merged_store;
3546 }
3547
3548 #ifndef PRODUCT
3549 void MergePrimitiveStores::trace(const Node_List& merge_list, const Node* merged_input_value, const StoreNode* merged_store) const {
3550 stringStream ss;
3551 ss.print_cr("[TraceMergeStores]: Replace");
3552 for (int i = (int)merge_list.size() - 1; i >= 0; i--) {
3553 merge_list.at(i)->dump("\n", false, &ss);
3554 }
3555 ss.print_cr("[TraceMergeStores]: with");
3556 merged_input_value->dump("\n", false, &ss);
3557 merged_store->dump("\n", false, &ss);
3558 tty->print("%s", ss.as_string());
3559 }
3560 #endif
3561
3562 //------------------------------Ideal------------------------------------------
3563 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
3564 // When a store immediately follows a relevant allocation/initialization,
3565 // try to capture it into the initialization, or hoist it above.
3566 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3567 Node* p = MemNode::Ideal_common(phase, can_reshape);
3568 if (p) return (p == NodeSentinel) ? nullptr : p;
3569
3570 Node* mem = in(MemNode::Memory);
3571 Node* address = in(MemNode::Address);
3572 Node* value = in(MemNode::ValueIn);
3573 // Back-to-back stores to same address? Fold em up. Generally
3574 // unsafe if I have intervening uses...
3575 if (phase->C->get_adr_type(phase->C->get_alias_index(adr_type())) != TypeAryPtr::INLINES) {
3576 Node* st = mem;
3577 // If Store 'st' has more than one use, we cannot fold 'st' away.
3578 // For example, 'st' might be the final state at a conditional
3579 // return. Or, 'st' might be used by some node which is live at
3580 // the same time 'st' is live, which might be unschedulable. So,
3581 // require exactly ONE user until such time as we clone 'mem' for
3582 // each of 'mem's uses (thus making the exactly-1-user-rule hold
3583 // true).
3584 while (st->is_Store() && st->outcnt() == 1) {
3585 // Looking at a dead closed cycle of memory?
3586 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
3587 assert(Opcode() == st->Opcode() ||
3588 st->Opcode() == Op_StoreVector ||
3589 Opcode() == Op_StoreVector ||
3590 st->Opcode() == Op_StoreVectorScatter ||
3591 Opcode() == Op_StoreVectorScatter ||
3592 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
3593 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
3594 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
3595 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) ||
3596 (st->adr_type()->isa_aryptr() && st->adr_type()->is_aryptr()->is_flat()) || // TODO 8343835
3597 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
3598 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
3599
3600 if (st->in(MemNode::Address)->eqv_uncast(address) &&
3601 st->as_Store()->memory_size() <= this->memory_size()) {
3602 Node* use = st->raw_out(0);
3603 if (phase->is_IterGVN()) {
3604 phase->is_IterGVN()->rehash_node_delayed(use);
3605 }
3606 // It's OK to do this in the parser, since DU info is always accurate,
3607 // and the parser always refers to nodes via SafePointNode maps.
3608 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
3609 return this;
3610 }
3611 st = st->in(MemNode::Memory);
3612 }
3613 }
3614
3615
3616 // Capture an unaliased, unconditional, simple store into an initializer.
3617 // Or, if it is independent of the allocation, hoist it above the allocation.
3618 if (ReduceFieldZeroing && /*can_reshape &&*/
3619 mem->is_Proj() && mem->in(0)->is_Initialize()) {
3620 InitializeNode* init = mem->in(0)->as_Initialize();
3621 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
3622 if (offset > 0) {
3623 Node* moved = init->capture_store(this, offset, phase, can_reshape);
3624 // If the InitializeNode captured me, it made a raw copy of me,
3625 // and I need to disappear.
3626 if (moved != nullptr) {
3627 // %%% hack to ensure that Ideal returns a new node:
3628 mem = MergeMemNode::make(mem);
3629 return mem; // fold me away
3630 }
3631 }
3632 }
3633
3634 // Fold reinterpret cast into memory operation:
3635 // StoreX mem (MoveY2X v) => StoreY mem v
3636 if (value->is_Move()) {
3637 const Type* vt = value->in(1)->bottom_type();
3638 if (has_reinterpret_variant(vt)) {
3639 if (phase->C->post_loop_opts_phase()) {
3640 return convert_to_reinterpret_store(*phase, value->in(1), vt);
3641 } else {
3642 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
3643 }
3644 }
3645 }
3646
3647 if (MergeStores && UseUnalignedAccesses) {
3648 if (phase->C->merge_stores_phase()) {
3649 MergePrimitiveStores merge(phase, this);
3650 Node* progress = merge.run();
3651 if (progress != nullptr) { return progress; }
3652 } else {
3653 // We need to wait with merging stores until RangeCheck smearing has removed the RangeChecks during
3654 // the post loops IGVN phase. If we do it earlier, then there may still be some RangeChecks between
3655 // the stores, and we merge the wrong sequence of stores.
3656 // Example:
3657 // StoreI RangeCheck StoreI StoreI RangeCheck StoreI
3658 // Apply MergeStores:
3659 // StoreI RangeCheck [ StoreL ] RangeCheck StoreI
3660 // Remove more RangeChecks:
3661 // StoreI [ StoreL ] StoreI
3662 // But now it would have been better to do this instead:
3663 // [ StoreL ] [ StoreL ]
3664 phase->C->record_for_merge_stores_igvn(this);
3665 }
3666 }
3667
3668 return nullptr; // No further progress
3669 }
3670
3671 //------------------------------Value-----------------------------------------
3672 const Type* StoreNode::Value(PhaseGVN* phase) const {
3673 // Either input is TOP ==> the result is TOP
3674 const Type *t1 = phase->type( in(MemNode::Memory) );
3675 if( t1 == Type::TOP ) return Type::TOP;
3676 const Type *t2 = phase->type( in(MemNode::Address) );
3677 if( t2 == Type::TOP ) return Type::TOP;
3678 const Type *t3 = phase->type( in(MemNode::ValueIn) );
3679 if( t3 == Type::TOP ) return Type::TOP;
3680 return Type::MEMORY;
3681 }
3682
3683 //------------------------------Identity---------------------------------------
3684 // Remove redundant stores:
3685 // Store(m, p, Load(m, p)) changes to m.
3686 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
3687 Node* StoreNode::Identity(PhaseGVN* phase) {
3688 Node* mem = in(MemNode::Memory);
3689 Node* adr = in(MemNode::Address);
3690 Node* val = in(MemNode::ValueIn);
3691
3692 Node* result = this;
3693
3694 // Load then Store? Then the Store is useless
3695 if (val->is_Load() &&
3696 val->in(MemNode::Address)->eqv_uncast(adr) &&
3697 val->in(MemNode::Memory )->eqv_uncast(mem) &&
3698 val->as_Load()->store_Opcode() == Opcode()) {
3699 // Ensure vector type is the same
3700 if (!is_StoreVector() || (mem->is_LoadVector() && as_StoreVector()->vect_type() == mem->as_LoadVector()->vect_type())) {
3701 result = mem;
3702 }
3703 }
3704
3705 // Two stores in a row of the same value?
3706 if (result == this &&
3707 mem->is_Store() &&
3708 mem->in(MemNode::Address)->eqv_uncast(adr) &&
3709 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
3710 mem->Opcode() == Opcode()) {
3711 if (!is_StoreVector()) {
3712 result = mem;
3713 } else {
3714 const StoreVectorNode* store_vector = as_StoreVector();
3715 const StoreVectorNode* mem_vector = mem->as_StoreVector();
3716 const Node* store_indices = store_vector->indices();
3717 const Node* mem_indices = mem_vector->indices();
3718 const Node* store_mask = store_vector->mask();
3719 const Node* mem_mask = mem_vector->mask();
3720 // Ensure types, indices, and masks match
3721 if (store_vector->vect_type() == mem_vector->vect_type() &&
3722 ((store_indices == nullptr) == (mem_indices == nullptr) &&
3723 (store_indices == nullptr || store_indices->eqv_uncast(mem_indices))) &&
3724 ((store_mask == nullptr) == (mem_mask == nullptr) &&
3725 (store_mask == nullptr || store_mask->eqv_uncast(mem_mask)))) {
3726 result = mem;
3727 }
3728 }
3729 }
3730
3731 // Store of zero anywhere into a freshly-allocated object?
3732 // Then the store is useless.
3733 // (It must already have been captured by the InitializeNode.)
3734 if (result == this && ReduceFieldZeroing) {
3735 // a newly allocated object is already all-zeroes everywhere
3736 if (mem->is_Proj() && mem->in(0)->is_Allocate() &&
3737 (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::InitValue) == val)) {
3738 result = mem;
3739 }
3740
3741 if (result == this && phase->type(val)->is_zero_type()) {
3742 // the store may also apply to zero-bits in an earlier object
3743 Node* prev_mem = find_previous_store(phase);
3744 // Steps (a), (b): Walk past independent stores to find an exact match.
3745 if (prev_mem != nullptr) {
3746 Node* prev_val = can_see_stored_value(prev_mem, phase);
3747 if (prev_val != nullptr && prev_val == val) {
3748 // prev_val and val might differ by a cast; it would be good
3749 // to keep the more informative of the two.
3750 result = mem;
3751 }
3752 }
3753 }
3754 }
3755
3756 PhaseIterGVN* igvn = phase->is_IterGVN();
3757 if (result != this && igvn != nullptr) {
3758 MemBarNode* trailing = trailing_membar();
3759 if (trailing != nullptr) {
3760 #ifdef ASSERT
3761 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
3762 assert(t_oop == nullptr || t_oop->is_known_instance_field(), "only for non escaping objects");
3763 #endif
3764 trailing->remove(igvn);
3765 }
3766 }
3767
3768 return result;
3769 }
3770
3771 //------------------------------match_edge-------------------------------------
3772 // Do we Match on this edge index or not? Match only memory & value
3773 uint StoreNode::match_edge(uint idx) const {
3774 return idx == MemNode::Address || idx == MemNode::ValueIn;
3775 }
3776
3777 //------------------------------cmp--------------------------------------------
3778 // Do not common stores up together. They generally have to be split
3779 // back up anyways, so do not bother.
3780 bool StoreNode::cmp( const Node &n ) const {
3781 return (&n == this); // Always fail except on self
3782 }
3783
3784 //------------------------------Ideal_masked_input-----------------------------
3785 // Check for a useless mask before a partial-word store
3786 // (StoreB ... (AndI valIn conIa) )
3787 // If (conIa & mask == mask) this simplifies to
3788 // (StoreB ... (valIn) )
3789 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
3790 Node *val = in(MemNode::ValueIn);
3791 if( val->Opcode() == Op_AndI ) {
3792 const TypeInt *t = phase->type( val->in(2) )->isa_int();
3793 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
3794 set_req_X(MemNode::ValueIn, val->in(1), phase);
3795 return this;
3796 }
3797 }
3798 return nullptr;
3799 }
3800
3801
3802 //------------------------------Ideal_sign_extended_input----------------------
3803 // Check for useless sign-extension before a partial-word store
3804 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
3805 // If (conIL == conIR && conIR <= num_rejected_bits) this simplifies to
3806 // (StoreB ... (v))
3807 // If (conIL > conIR) under some conditions, it can be simplified into
3808 // (StoreB ... (LShiftI _ v (conIL - conIR)))
3809 // This case happens when the value of the store was itself a left shift, that
3810 // gets merged into the inner left shift of the sign-extension. For instance,
3811 // if we have
3812 // array_of_shorts[0] = (short)(v << 2)
3813 // We get a structure such as:
3814 // (StoreB ... (RShiftI _ (LShiftI _ (LShiftI _ v 2) 16) 16))
3815 // that is simplified into
3816 // (StoreB ... (RShiftI _ (LShiftI _ v 18) 16)).
3817 // It is thus useful to handle cases where conIL > conIR. But this simplification
3818 // does not always hold. Let's see in which cases it's valid.
3819 //
3820 // Let's assume we have the following 32 bits integer v:
3821 // +----------------------------------+
3822 // | v[0..31] |
3823 // +----------------------------------+
3824 // 31 0
3825 // that will be stuffed in 8 bits byte after a shift left and a shift right of
3826 // potentially different magnitudes.
3827 // We denote num_rejected_bits the number of bits of the discarded part. In this
3828 // case, num_rejected_bits == 24.
3829 //
3830 // Statement (proved further below in case analysis):
3831 // Given:
3832 // - 0 <= conIL < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
3833 // - 0 <= conIR < BitsPerJavaInteger (no wrap in shift, enforced by maskShiftAmount)
3834 // - conIL >= conIR
3835 // - num_rejected_bits >= conIR
3836 // Then this form:
3837 // (RShiftI _ (LShiftI _ v conIL) conIR)
3838 // can be replaced with this form:
3839 // (LShiftI _ v (conIL-conIR))
3840 //
3841 // Note: We only have to show that the non-rejected lowest bits (8 bits for byte)
3842 // have to be correct, as the higher bits are rejected / truncated by the store.
3843 //
3844 // The hypotheses
3845 // 0 <= conIL < BitsPerJavaInteger
3846 // 0 <= conIR < BitsPerJavaInteger
3847 // are ensured by maskShiftAmount (called from ::Ideal of shift nodes). Indeed,
3848 // (v << 31) << 2 must be simplified into 0, not into v << 33 (which is equivalent
3849 // to v << 1).
3850 //
3851 //
3852 // If you don't like case analysis, jump after the conclusion.
3853 // ### Case 1 : conIL == conIR
3854 // ###### Case 1.1: conIL == conIR == num_rejected_bits
3855 // If we do the shift left then right by 24 bits, we get:
3856 // after: v << 24
3857 // +---------+------------------------+
3858 // | v[0..7] | 0 |
3859 // +---------+------------------------+
3860 // 31 24 23 0
3861 // after: (v << 24) >> 24
3862 // +------------------------+---------+
3863 // | sign bit | v[0..7] |
3864 // +------------------------+---------+
3865 // 31 8 7 0
3866 // The non-rejected bits (bits kept by the store, that is the 8 lower bits of the
3867 // result) are the same before and after, so, indeed, simplifying is correct.
3868
3869 // ###### Case 1.2: conIL == conIR < num_rejected_bits
3870 // If we do the shift left then right by 22 bits, we get:
3871 // after: v << 22
3872 // +---------+------------------------+
3873 // | v[0..9] | 0 |
3874 // +---------+------------------------+
3875 // 31 22 21 0
3876 // after: (v << 22) >> 22
3877 // +------------------------+---------+
3878 // | sign bit | v[0..9] |
3879 // +------------------------+---------+
3880 // 31 10 9 0
3881 // The non-rejected bits are the 8 lower bits of v. The bits 8 and 9 of v are still
3882 // present in (v << 22) >> 22 but will be dropped by the store. The simplification is
3883 // still correct.
3884
3885 // ###### But! Case 1.3: conIL == conIR > num_rejected_bits
3886 // If we do the shift left then right by 26 bits, we get:
3887 // after: v << 26
3888 // +---------+------------------------+
3889 // | v[0..5] | 0 |
3890 // +---------+------------------------+
3891 // 31 26 25 0
3892 // after: (v << 26) >> 26
3893 // +------------------------+---------+
3894 // | sign bit | v[0..5] |
3895 // +------------------------+---------+
3896 // 31 6 5 0
3897 // The non-rejected bits are made of
3898 // - 0-5 => the bits 0 to 5 of v
3899 // - 6-7 => the sign bit of v[0..5] (that is v[5])
3900 // Simplifying this as v is not correct.
3901 // The condition conIR <= num_rejected_bits is indeed necessary in Case 1
3902 //
3903 // ### Case 2: conIL > conIR
3904 // ###### Case 2.1: num_rejected_bits == conIR
3905 // We take conIL == 26 for this example.
3906 // after: v << 26
3907 // +---------+------------------------+
3908 // | v[0..5] | 0 |
3909 // +---------+------------------------+
3910 // 31 26 25 0
3911 // after: (v << 26) >> 24
3912 // +------------------+---------+-----+
3913 // | sign bit | v[0..5] | 0 |
3914 // +------------------+---------+-----+
3915 // 31 8 7 2 1 0
3916 // The non-rejected bits are the 8 lower ones of (v << conIL - conIR).
3917 // The bits 6 and 7 of v have been thrown away after the shift left.
3918 // The simplification is still correct.
3919 //
3920 // ###### Case 2.2: num_rejected_bits > conIR.
3921 // Let's say conIL == 26 and conIR == 22.
3922 // after: v << 26
3923 // +---------+------------------------+
3924 // | v[0..5] | 0 |
3925 // +---------+------------------------+
3926 // 31 26 25 0
3927 // after: (v << 26) >> 22
3928 // +------------------+---------+-----+
3929 // | sign bit | v[0..5] | 0 |
3930 // +------------------+---------+-----+
3931 // 31 10 9 4 3 0
3932 // The bits non-rejected by the store are exactly the 8 lower ones of (v << (conIL - conIR)):
3933 // - 0-3 => 0
3934 // - 4-7 => bits 0 to 3 of v
3935 // The simplification is still correct.
3936 // The bits 4 and 5 of v are still present in (v << (conIL - conIR)) but they don't
3937 // matter as they are not in the 8 lower bits: they will be cut out by the store.
3938 //
3939 // ###### But! Case 2.3: num_rejected_bits < conIR.
3940 // Let's see that this case is not as easy to simplify.
3941 // Let's say conIL == 28 and conIR == 26.
3942 // after: v << 28
3943 // +---------+------------------------+
3944 // | v[0..3] | 0 |
3945 // +---------+------------------------+
3946 // 31 28 27 0
3947 // after: (v << 28) >> 26
3948 // +------------------+---------+-----+
3949 // | sign bit | v[0..3] | 0 |
3950 // +------------------+---------+-----+
3951 // 31 6 5 2 1 0
3952 // The non-rejected bits are made of
3953 // - 0-1 => 0
3954 // - 2-5 => the bits 0 to 3 of v
3955 // - 6-7 => the sign bit of v[0..3] (that is v[3])
3956 // Simplifying this as (v << 2) is not correct.
3957 // The condition conIR <= num_rejected_bits is indeed necessary in Case 2.
3958 //
3959 // ### Conclusion:
3960 // Our hypotheses are indeed sufficient:
3961 // - 0 <= conIL < BitsPerJavaInteger
3962 // - 0 <= conIR < BitsPerJavaInteger
3963 // - conIL >= conIR
3964 // - num_rejected_bits >= conIR
3965 //
3966 // ### A rationale without case analysis:
3967 // After the shift left, conIL upper bits of v are discarded and conIL lower bit
3968 // zeroes are added. After the shift right, conIR lower bits of the previous result
3969 // are discarded. If conIL >= conIR, we discard only the zeroes we made up during
3970 // the shift left, but if conIL < conIR, then we discard also lower bits of v. But
3971 // the point of the simplification is to get an expression of the form
3972 // (v << (conIL - conIR)). This expression discard only higher bits of v, thus the
3973 // simplification is not correct if conIL < conIR.
3974 //
3975 // Moreover, after the shift right, the higher bit of (v << conIL) is repeated on the
3976 // conIR higher bits of ((v << conIL) >> conIR), it's the sign-extension. If
3977 // conIR > num_rejected_bits, then at least one artificial copy of this sign bit will
3978 // be in the window of the store. Thus ((v << conIL) >> conIR) is not equivalent to
3979 // (v << (conIL-conIR)) if conIR > num_rejected_bits.
3980 //
3981 // We do not treat the case conIL < conIR here since the point of this function is
3982 // to skip sign-extensions (that is conIL == conIR == num_rejected_bits). The need
3983 // of treating conIL > conIR comes from the cases where the sign-extended value is
3984 // also left-shift expression. Computing the sign-extension of a right-shift expression
3985 // doesn't yield a situation such as
3986 // (StoreB ... (RShiftI _ (LShiftI _ v conIL) conIR))
3987 // where conIL < conIR.
3988 Node* StoreNode::Ideal_sign_extended_input(PhaseGVN* phase, int num_rejected_bits) {
3989 Node* shr = in(MemNode::ValueIn);
3990 if (shr->Opcode() == Op_RShiftI) {
3991 const TypeInt* conIR = phase->type(shr->in(2))->isa_int();
3992 if (conIR != nullptr && conIR->is_con() && conIR->get_con() >= 0 && conIR->get_con() < BitsPerJavaInteger && conIR->get_con() <= num_rejected_bits) {
3993 Node* shl = shr->in(1);
3994 if (shl->Opcode() == Op_LShiftI) {
3995 const TypeInt* conIL = phase->type(shl->in(2))->isa_int();
3996 if (conIL != nullptr && conIL->is_con() && conIL->get_con() >= 0 && conIL->get_con() < BitsPerJavaInteger) {
3997 if (conIL->get_con() == conIR->get_con()) {
3998 set_req_X(MemNode::ValueIn, shl->in(1), phase);
3999 return this;
4000 }
4001 if (conIL->get_con() > conIR->get_con()) {
4002 Node* new_shl = phase->transform(new LShiftINode(shl->in(1), phase->intcon(conIL->get_con() - conIR->get_con())));
4003 set_req_X(MemNode::ValueIn, new_shl, phase);
4004 return this;
4005 }
4006 }
4007 }
4008 }
4009 }
4010 return nullptr;
4011 }
4012
4013 //------------------------------value_never_loaded-----------------------------------
4014 // Determine whether there are any possible loads of the value stored.
4015 // For simplicity, we actually check if there are any loads from the
4016 // address stored to, not just for loads of the value stored by this node.
4017 //
4018 bool StoreNode::value_never_loaded(PhaseValues* phase) const {
4019 Node *adr = in(Address);
4020 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
4021 if (adr_oop == nullptr)
4022 return false;
4023 if (!adr_oop->is_known_instance_field())
4024 return false; // if not a distinct instance, there may be aliases of the address
4025 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
4026 Node *use = adr->fast_out(i);
4027 if (use->is_Load() || use->is_LoadStore()) {
4028 return false;
4029 }
4030 }
4031 return true;
4032 }
4033
4034 MemBarNode* StoreNode::trailing_membar() const {
4035 if (is_release()) {
4036 MemBarNode* trailing_mb = nullptr;
4037 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4038 Node* u = fast_out(i);
4039 if (u->is_MemBar()) {
4040 if (u->as_MemBar()->trailing_store()) {
4041 assert(u->Opcode() == Op_MemBarVolatile, "");
4042 assert(trailing_mb == nullptr, "only one");
4043 trailing_mb = u->as_MemBar();
4044 #ifdef ASSERT
4045 Node* leading = u->as_MemBar()->leading_membar();
4046 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
4047 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
4048 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
4049 #endif
4050 } else {
4051 assert(u->as_MemBar()->standalone(), "");
4052 }
4053 }
4054 }
4055 return trailing_mb;
4056 }
4057 return nullptr;
4058 }
4059
4060
4061 //=============================================================================
4062 //------------------------------Ideal------------------------------------------
4063 // If the store is from an AND mask that leaves the low bits untouched, then
4064 // we can skip the AND operation. If the store is from a sign-extension
4065 // (a left shift, then right shift) we can skip both.
4066 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
4067 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
4068 if( progress != nullptr ) return progress;
4069
4070 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
4071 if( progress != nullptr ) return progress;
4072
4073 // Finally check the default case
4074 return StoreNode::Ideal(phase, can_reshape);
4075 }
4076
4077 //=============================================================================
4078 //------------------------------Ideal------------------------------------------
4079 // If the store is from an AND mask that leaves the low bits untouched, then
4080 // we can skip the AND operation
4081 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
4082 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
4083 if( progress != nullptr ) return progress;
4084
4085 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
4086 if( progress != nullptr ) return progress;
4087
4088 // Finally check the default case
4089 return StoreNode::Ideal(phase, can_reshape);
4090 }
4091
4092 //=============================================================================
4093 //----------------------------------SCMemProjNode------------------------------
4094 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
4095 {
4096 if (in(0) == nullptr || phase->type(in(0)) == Type::TOP) {
4097 return Type::TOP;
4098 }
4099 return bottom_type();
4100 }
4101
4102 //=============================================================================
4103 //----------------------------------LoadStoreNode------------------------------
4104 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
4105 : Node(required),
4106 _type(rt),
4107 _adr_type(at),
4108 _barrier_data(0)
4109 {
4110 init_req(MemNode::Control, c );
4111 init_req(MemNode::Memory , mem);
4112 init_req(MemNode::Address, adr);
4113 init_req(MemNode::ValueIn, val);
4114 init_class_id(Class_LoadStore);
4115 }
4116
4117 //------------------------------Value-----------------------------------------
4118 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
4119 // Either input is TOP ==> the result is TOP
4120 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
4121 return Type::TOP;
4122 }
4123 const Type* t = phase->type(in(MemNode::Memory));
4124 if (t == Type::TOP) {
4125 return Type::TOP;
4126 }
4127 t = phase->type(in(MemNode::Address));
4128 if (t == Type::TOP) {
4129 return Type::TOP;
4130 }
4131 t = phase->type(in(MemNode::ValueIn));
4132 if (t == Type::TOP) {
4133 return Type::TOP;
4134 }
4135 return bottom_type();
4136 }
4137
4138 uint LoadStoreNode::ideal_reg() const {
4139 return _type->ideal_reg();
4140 }
4141
4142 // This method conservatively checks if the result of a LoadStoreNode is
4143 // used, that is, if it returns true, then it is definitely the case that
4144 // the result of the node is not needed.
4145 // For example, GetAndAdd can be matched into a lock_add instead of a
4146 // lock_xadd if the result of LoadStoreNode::result_not_used() is true
4147 bool LoadStoreNode::result_not_used() const {
4148 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4149 Node *x = fast_out(i);
4150 if (x->Opcode() == Op_SCMemProj) {
4151 continue;
4152 }
4153 if (x->bottom_type() == TypeTuple::MEMBAR &&
4154 !x->is_Call() &&
4155 x->Opcode() != Op_Blackhole) {
4156 continue;
4157 }
4158 return false;
4159 }
4160 return true;
4161 }
4162
4163 MemBarNode* LoadStoreNode::trailing_membar() const {
4164 MemBarNode* trailing = nullptr;
4165 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
4166 Node* u = fast_out(i);
4167 if (u->is_MemBar()) {
4168 if (u->as_MemBar()->trailing_load_store()) {
4169 assert(u->Opcode() == Op_MemBarAcquire, "");
4170 assert(trailing == nullptr, "only one");
4171 trailing = u->as_MemBar();
4172 #ifdef ASSERT
4173 Node* leading = trailing->leading_membar();
4174 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
4175 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
4176 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
4177 #endif
4178 } else {
4179 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
4180 }
4181 }
4182 }
4183
4184 return trailing;
4185 }
4186
4187 uint LoadStoreNode::size_of() const { return sizeof(*this); }
4188
4189 //=============================================================================
4190 //----------------------------------LoadStoreConditionalNode--------------------
4191 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, nullptr, TypeInt::BOOL, 5) {
4192 init_req(ExpectedIn, ex );
4193 }
4194
4195 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
4196 // Either input is TOP ==> the result is TOP
4197 const Type* t = phase->type(in(ExpectedIn));
4198 if (t == Type::TOP) {
4199 return Type::TOP;
4200 }
4201 return LoadStoreNode::Value(phase);
4202 }
4203
4204 //=============================================================================
4205 //-------------------------------adr_type--------------------------------------
4206 const TypePtr* ClearArrayNode::adr_type() const {
4207 Node *adr = in(3);
4208 if (adr == nullptr) return nullptr; // node is dead
4209 return MemNode::calculate_adr_type(adr->bottom_type());
4210 }
4211
4212 //------------------------------match_edge-------------------------------------
4213 // Do we Match on this edge index or not? Do not match memory
4214 uint ClearArrayNode::match_edge(uint idx) const {
4215 return idx > 1;
4216 }
4217
4218 //------------------------------Identity---------------------------------------
4219 // Clearing a zero length array does nothing
4220 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
4221 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
4222 }
4223
4224 //------------------------------Idealize---------------------------------------
4225 // Clearing a short array is faster with stores
4226 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4227 // Already know this is a large node, do not try to ideal it
4228 if (_is_large) return nullptr;
4229
4230 const int unit = BytesPerLong;
4231 const TypeX* t = phase->type(in(2))->isa_intptr_t();
4232 if (!t) return nullptr;
4233 if (!t->is_con()) return nullptr;
4234 intptr_t raw_count = t->get_con();
4235 intptr_t size = raw_count;
4236 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
4237 // Clearing nothing uses the Identity call.
4238 // Negative clears are possible on dead ClearArrays
4239 // (see jck test stmt114.stmt11402.val).
4240 if (size <= 0 || size % unit != 0) return nullptr;
4241 intptr_t count = size / unit;
4242 // Length too long; communicate this to matchers and assemblers.
4243 // Assemblers are responsible to produce fast hardware clears for it.
4244 if (size > InitArrayShortSize) {
4245 return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true);
4246 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
4247 return nullptr;
4248 }
4249 if (!IdealizeClearArrayNode) return nullptr;
4250 Node *mem = in(1);
4251 if( phase->type(mem)==Type::TOP ) return nullptr;
4252 Node *adr = in(3);
4253 const Type* at = phase->type(adr);
4254 if( at==Type::TOP ) return nullptr;
4255 const TypePtr* atp = at->isa_ptr();
4256 // adjust atp to be the correct array element address type
4257 if (atp == nullptr) atp = TypePtr::BOTTOM;
4258 else atp = atp->add_offset(Type::OffsetBot);
4259 // Get base for derived pointer purposes
4260 if( adr->Opcode() != Op_AddP ) Unimplemented();
4261 Node *base = adr->in(1);
4262
4263 Node *val = in(4);
4264 Node *off = phase->MakeConX(BytesPerLong);
4265 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
4266 count--;
4267 while( count-- ) {
4268 mem = phase->transform(mem);
4269 adr = phase->transform(new AddPNode(base,adr,off));
4270 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
4271 }
4272 return mem;
4273 }
4274
4275 //----------------------------step_through----------------------------------
4276 // Return allocation input memory edge if it is different instance
4277 // or itself if it is the one we are looking for.
4278 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseValues* phase) {
4279 Node* n = *np;
4280 assert(n->is_ClearArray(), "sanity");
4281 intptr_t offset;
4282 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
4283 // This method is called only before Allocate nodes are expanded
4284 // during macro nodes expansion. Before that ClearArray nodes are
4285 // only generated in PhaseMacroExpand::generate_arraycopy() (before
4286 // Allocate nodes are expanded) which follows allocations.
4287 assert(alloc != nullptr, "should have allocation");
4288 if (alloc->_idx == instance_id) {
4289 // Can not bypass initialization of the instance we are looking for.
4290 return false;
4291 }
4292 // Otherwise skip it.
4293 InitializeNode* init = alloc->initialization();
4294 if (init != nullptr)
4295 *np = init->in(TypeFunc::Memory);
4296 else
4297 *np = alloc->in(TypeFunc::Memory);
4298 return true;
4299 }
4300
4301 //----------------------------clear_memory-------------------------------------
4302 // Generate code to initialize object storage to zero.
4303 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4304 Node* val,
4305 Node* raw_val,
4306 intptr_t start_offset,
4307 Node* end_offset,
4308 PhaseGVN* phase) {
4309 intptr_t offset = start_offset;
4310
4311 int unit = BytesPerLong;
4312 if ((offset % unit) != 0) {
4313 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
4314 adr = phase->transform(adr);
4315 const TypePtr* atp = TypeRawPtr::BOTTOM;
4316 if (val != nullptr) {
4317 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
4318 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
4319 } else {
4320 assert(raw_val == nullptr, "val may not be null");
4321 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4322 }
4323 mem = phase->transform(mem);
4324 offset += BytesPerInt;
4325 }
4326 assert((offset % unit) == 0, "");
4327
4328 // Initialize the remaining stuff, if any, with a ClearArray.
4329 return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, phase);
4330 }
4331
4332 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4333 Node* raw_val,
4334 Node* start_offset,
4335 Node* end_offset,
4336 PhaseGVN* phase) {
4337 if (start_offset == end_offset) {
4338 // nothing to do
4339 return mem;
4340 }
4341
4342 int unit = BytesPerLong;
4343 Node* zbase = start_offset;
4344 Node* zend = end_offset;
4345
4346 // Scale to the unit required by the CPU:
4347 if (!Matcher::init_array_count_is_in_bytes) {
4348 Node* shift = phase->intcon(exact_log2(unit));
4349 zbase = phase->transform(new URShiftXNode(zbase, shift) );
4350 zend = phase->transform(new URShiftXNode(zend, shift) );
4351 }
4352
4353 // Bulk clear double-words
4354 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
4355 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
4356 if (raw_val == nullptr) {
4357 raw_val = phase->MakeConX(0);
4358 }
4359 mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false);
4360 return phase->transform(mem);
4361 }
4362
4363 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
4364 Node* val,
4365 Node* raw_val,
4366 intptr_t start_offset,
4367 intptr_t end_offset,
4368 PhaseGVN* phase) {
4369 if (start_offset == end_offset) {
4370 // nothing to do
4371 return mem;
4372 }
4373
4374 assert((end_offset % BytesPerInt) == 0, "odd end offset");
4375 intptr_t done_offset = end_offset;
4376 if ((done_offset % BytesPerLong) != 0) {
4377 done_offset -= BytesPerInt;
4378 }
4379 if (done_offset > start_offset) {
4380 mem = clear_memory(ctl, mem, dest, val, raw_val,
4381 start_offset, phase->MakeConX(done_offset), phase);
4382 }
4383 if (done_offset < end_offset) { // emit the final 32-bit store
4384 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
4385 adr = phase->transform(adr);
4386 const TypePtr* atp = TypeRawPtr::BOTTOM;
4387 if (val != nullptr) {
4388 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
4389 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
4390 } else {
4391 assert(raw_val == nullptr, "val may not be null");
4392 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
4393 }
4394 mem = phase->transform(mem);
4395 done_offset += BytesPerInt;
4396 }
4397 assert(done_offset == end_offset, "");
4398 return mem;
4399 }
4400
4401 //=============================================================================
4402 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
4403 : MultiNode(TypeFunc::Parms + (precedent == nullptr? 0: 1)),
4404 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
4405 #ifdef ASSERT
4406 , _pair_idx(0)
4407 #endif
4408 {
4409 init_class_id(Class_MemBar);
4410 Node* top = C->top();
4411 init_req(TypeFunc::I_O,top);
4412 init_req(TypeFunc::FramePtr,top);
4413 init_req(TypeFunc::ReturnAdr,top);
4414 if (precedent != nullptr)
4415 init_req(TypeFunc::Parms, precedent);
4416 }
4417
4418 //------------------------------cmp--------------------------------------------
4419 uint MemBarNode::hash() const { return NO_HASH; }
4420 bool MemBarNode::cmp( const Node &n ) const {
4421 return (&n == this); // Always fail except on self
4422 }
4423
4424 //------------------------------make-------------------------------------------
4425 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
4426 switch (opcode) {
4427 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
4428 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
4429 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
4430 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
4431 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
4432 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn);
4433 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
4434 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
4435 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
4436 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
4437 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
4438 case Op_Initialize: return new InitializeNode(C, atp, pn);
4439 default: ShouldNotReachHere(); return nullptr;
4440 }
4441 }
4442
4443 void MemBarNode::remove(PhaseIterGVN *igvn) {
4444 if (outcnt() != 2) {
4445 assert(Opcode() == Op_Initialize, "Only seen when there are no use of init memory");
4446 assert(outcnt() == 1, "Only control then");
4447 }
4448 if (trailing_store() || trailing_load_store()) {
4449 MemBarNode* leading = leading_membar();
4450 if (leading != nullptr) {
4451 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
4452 leading->remove(igvn);
4453 }
4454 }
4455 if (proj_out_or_null(TypeFunc::Memory) != nullptr) {
4456 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
4457 }
4458 if (proj_out_or_null(TypeFunc::Control) != nullptr) {
4459 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
4460 }
4461 }
4462
4463 //------------------------------Ideal------------------------------------------
4464 // Return a node which is more "ideal" than the current node. Strip out
4465 // control copies
4466 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4467 if (remove_dead_region(phase, can_reshape)) return this;
4468 // Don't bother trying to transform a dead node
4469 if (in(0) && in(0)->is_top()) {
4470 return nullptr;
4471 }
4472
4473 bool progress = false;
4474 // Eliminate volatile MemBars for scalar replaced objects.
4475 if (can_reshape && req() == (Precedent+1)) {
4476 bool eliminate = false;
4477 int opc = Opcode();
4478 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
4479 // Volatile field loads and stores.
4480 Node* my_mem = in(MemBarNode::Precedent);
4481 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
4482 if ((my_mem != nullptr) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
4483 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
4484 // replace this Precedent (decodeN) with the Load instead.
4485 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
4486 Node* load_node = my_mem->in(1);
4487 set_req(MemBarNode::Precedent, load_node);
4488 phase->is_IterGVN()->_worklist.push(my_mem);
4489 my_mem = load_node;
4490 } else {
4491 assert(my_mem->unique_out() == this, "sanity");
4492 assert(!trailing_load_store(), "load store node can't be eliminated");
4493 del_req(Precedent);
4494 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
4495 my_mem = nullptr;
4496 }
4497 progress = true;
4498 }
4499 if (my_mem != nullptr && my_mem->is_Mem()) {
4500 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
4501 // Check for scalar replaced object reference.
4502 if( t_oop != nullptr && t_oop->is_known_instance_field() &&
4503 t_oop->offset() != Type::OffsetBot &&
4504 t_oop->offset() != Type::OffsetTop) {
4505 eliminate = true;
4506 }
4507 }
4508 } else if (opc == Op_MemBarRelease || (UseStoreStoreForCtor && opc == Op_MemBarStoreStore)) {
4509 // Final field stores.
4510 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent));
4511 if ((alloc != nullptr) && alloc->is_Allocate() &&
4512 alloc->as_Allocate()->does_not_escape_thread()) {
4513 // The allocated object does not escape.
4514 eliminate = true;
4515 }
4516 }
4517 if (eliminate) {
4518 // Replace MemBar projections by its inputs.
4519 PhaseIterGVN* igvn = phase->is_IterGVN();
4520 remove(igvn);
4521 // Must return either the original node (now dead) or a new node
4522 // (Do not return a top here, since that would break the uniqueness of top.)
4523 return new ConINode(TypeInt::ZERO);
4524 }
4525 }
4526 return progress ? this : nullptr;
4527 }
4528
4529 //------------------------------Value------------------------------------------
4530 const Type* MemBarNode::Value(PhaseGVN* phase) const {
4531 if( !in(0) ) return Type::TOP;
4532 if( phase->type(in(0)) == Type::TOP )
4533 return Type::TOP;
4534 return TypeTuple::MEMBAR;
4535 }
4536
4537 //------------------------------match------------------------------------------
4538 // Construct projections for memory.
4539 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) {
4540 switch (proj->_con) {
4541 case TypeFunc::Control:
4542 case TypeFunc::Memory:
4543 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
4544 }
4545 ShouldNotReachHere();
4546 return nullptr;
4547 }
4548
4549 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4550 trailing->_kind = TrailingStore;
4551 leading->_kind = LeadingStore;
4552 #ifdef ASSERT
4553 trailing->_pair_idx = leading->_idx;
4554 leading->_pair_idx = leading->_idx;
4555 #endif
4556 }
4557
4558 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
4559 trailing->_kind = TrailingLoadStore;
4560 leading->_kind = LeadingLoadStore;
4561 #ifdef ASSERT
4562 trailing->_pair_idx = leading->_idx;
4563 leading->_pair_idx = leading->_idx;
4564 #endif
4565 }
4566
4567 MemBarNode* MemBarNode::trailing_membar() const {
4568 ResourceMark rm;
4569 Node* trailing = (Node*)this;
4570 VectorSet seen;
4571 Node_Stack multis(0);
4572 do {
4573 Node* c = trailing;
4574 uint i = 0;
4575 do {
4576 trailing = nullptr;
4577 for (; i < c->outcnt(); i++) {
4578 Node* next = c->raw_out(i);
4579 if (next != c && next->is_CFG()) {
4580 if (c->is_MultiBranch()) {
4581 if (multis.node() == c) {
4582 multis.set_index(i+1);
4583 } else {
4584 multis.push(c, i+1);
4585 }
4586 }
4587 trailing = next;
4588 break;
4589 }
4590 }
4591 if (trailing != nullptr && !seen.test_set(trailing->_idx)) {
4592 break;
4593 }
4594 while (multis.size() > 0) {
4595 c = multis.node();
4596 i = multis.index();
4597 if (i < c->req()) {
4598 break;
4599 }
4600 multis.pop();
4601 }
4602 } while (multis.size() > 0);
4603 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
4604
4605 MemBarNode* mb = trailing->as_MemBar();
4606 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
4607 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
4608 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
4609 return mb;
4610 }
4611
4612 MemBarNode* MemBarNode::leading_membar() const {
4613 ResourceMark rm;
4614 VectorSet seen;
4615 Node_Stack regions(0);
4616 Node* leading = in(0);
4617 while (leading != nullptr && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
4618 while (leading == nullptr || leading->is_top() || seen.test_set(leading->_idx)) {
4619 leading = nullptr;
4620 while (regions.size() > 0 && leading == nullptr) {
4621 Node* r = regions.node();
4622 uint i = regions.index();
4623 if (i < r->req()) {
4624 leading = r->in(i);
4625 regions.set_index(i+1);
4626 } else {
4627 regions.pop();
4628 }
4629 }
4630 if (leading == nullptr) {
4631 assert(regions.size() == 0, "all paths should have been tried");
4632 return nullptr;
4633 }
4634 }
4635 if (leading->is_Region()) {
4636 regions.push(leading, 2);
4637 leading = leading->in(1);
4638 } else {
4639 leading = leading->in(0);
4640 }
4641 }
4642 #ifdef ASSERT
4643 Unique_Node_List wq;
4644 wq.push((Node*)this);
4645 uint found = 0;
4646 for (uint i = 0; i < wq.size(); i++) {
4647 Node* n = wq.at(i);
4648 if (n->is_Region()) {
4649 for (uint j = 1; j < n->req(); j++) {
4650 Node* in = n->in(j);
4651 if (in != nullptr && !in->is_top()) {
4652 wq.push(in);
4653 }
4654 }
4655 } else {
4656 if (n->is_MemBar() && n->as_MemBar()->leading()) {
4657 assert(n == leading, "consistency check failed");
4658 found++;
4659 } else {
4660 Node* in = n->in(0);
4661 if (in != nullptr && !in->is_top()) {
4662 wq.push(in);
4663 }
4664 }
4665 }
4666 }
4667 assert(found == 1 || (found == 0 && leading == nullptr), "consistency check failed");
4668 #endif
4669 if (leading == nullptr) {
4670 return nullptr;
4671 }
4672 MemBarNode* mb = leading->as_MemBar();
4673 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
4674 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
4675 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
4676 return mb;
4677 }
4678
4679
4680 //===========================InitializeNode====================================
4681 // SUMMARY:
4682 // This node acts as a memory barrier on raw memory, after some raw stores.
4683 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
4684 // The Initialize can 'capture' suitably constrained stores as raw inits.
4685 // It can coalesce related raw stores into larger units (called 'tiles').
4686 // It can avoid zeroing new storage for memory units which have raw inits.
4687 // At macro-expansion, it is marked 'complete', and does not optimize further.
4688 //
4689 // EXAMPLE:
4690 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
4691 // ctl = incoming control; mem* = incoming memory
4692 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
4693 // First allocate uninitialized memory and fill in the header:
4694 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
4695 // ctl := alloc.Control; mem* := alloc.Memory*
4696 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
4697 // Then initialize to zero the non-header parts of the raw memory block:
4698 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
4699 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
4700 // After the initialize node executes, the object is ready for service:
4701 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
4702 // Suppose its body is immediately initialized as {1,2}:
4703 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
4704 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
4705 // mem.SLICE(#short[*]) := store2
4706 //
4707 // DETAILS:
4708 // An InitializeNode collects and isolates object initialization after
4709 // an AllocateNode and before the next possible safepoint. As a
4710 // memory barrier (MemBarNode), it keeps critical stores from drifting
4711 // down past any safepoint or any publication of the allocation.
4712 // Before this barrier, a newly-allocated object may have uninitialized bits.
4713 // After this barrier, it may be treated as a real oop, and GC is allowed.
4714 //
4715 // The semantics of the InitializeNode include an implicit zeroing of
4716 // the new object from object header to the end of the object.
4717 // (The object header and end are determined by the AllocateNode.)
4718 //
4719 // Certain stores may be added as direct inputs to the InitializeNode.
4720 // These stores must update raw memory, and they must be to addresses
4721 // derived from the raw address produced by AllocateNode, and with
4722 // a constant offset. They must be ordered by increasing offset.
4723 // The first one is at in(RawStores), the last at in(req()-1).
4724 // Unlike most memory operations, they are not linked in a chain,
4725 // but are displayed in parallel as users of the rawmem output of
4726 // the allocation.
4727 //
4728 // (See comments in InitializeNode::capture_store, which continue
4729 // the example given above.)
4730 //
4731 // When the associated Allocate is macro-expanded, the InitializeNode
4732 // may be rewritten to optimize collected stores. A ClearArrayNode
4733 // may also be created at that point to represent any required zeroing.
4734 // The InitializeNode is then marked 'complete', prohibiting further
4735 // capturing of nearby memory operations.
4736 //
4737 // During macro-expansion, all captured initializations which store
4738 // constant values of 32 bits or smaller are coalesced (if advantageous)
4739 // into larger 'tiles' 32 or 64 bits. This allows an object to be
4740 // initialized in fewer memory operations. Memory words which are
4741 // covered by neither tiles nor non-constant stores are pre-zeroed
4742 // by explicit stores of zero. (The code shape happens to do all
4743 // zeroing first, then all other stores, with both sequences occurring
4744 // in order of ascending offsets.)
4745 //
4746 // Alternatively, code may be inserted between an AllocateNode and its
4747 // InitializeNode, to perform arbitrary initialization of the new object.
4748 // E.g., the object copying intrinsics insert complex data transfers here.
4749 // The initialization must then be marked as 'complete' disable the
4750 // built-in zeroing semantics and the collection of initializing stores.
4751 //
4752 // While an InitializeNode is incomplete, reads from the memory state
4753 // produced by it are optimizable if they match the control edge and
4754 // new oop address associated with the allocation/initialization.
4755 // They return a stored value (if the offset matches) or else zero.
4756 // A write to the memory state, if it matches control and address,
4757 // and if it is to a constant offset, may be 'captured' by the
4758 // InitializeNode. It is cloned as a raw memory operation and rewired
4759 // inside the initialization, to the raw oop produced by the allocation.
4760 // Operations on addresses which are provably distinct (e.g., to
4761 // other AllocateNodes) are allowed to bypass the initialization.
4762 //
4763 // The effect of all this is to consolidate object initialization
4764 // (both arrays and non-arrays, both piecewise and bulk) into a
4765 // single location, where it can be optimized as a unit.
4766 //
4767 // Only stores with an offset less than TrackedInitializationLimit words
4768 // will be considered for capture by an InitializeNode. This puts a
4769 // reasonable limit on the complexity of optimized initializations.
4770
4771 //---------------------------InitializeNode------------------------------------
4772 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
4773 : MemBarNode(C, adr_type, rawoop),
4774 _is_complete(Incomplete), _does_not_escape(false)
4775 {
4776 init_class_id(Class_Initialize);
4777
4778 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
4779 assert(in(RawAddress) == rawoop, "proper init");
4780 // Note: allocation() can be null, for secondary initialization barriers
4781 }
4782
4783 // Since this node is not matched, it will be processed by the
4784 // register allocator. Declare that there are no constraints
4785 // on the allocation of the RawAddress edge.
4786 const RegMask &InitializeNode::in_RegMask(uint idx) const {
4787 // This edge should be set to top, by the set_complete. But be conservative.
4788 if (idx == InitializeNode::RawAddress)
4789 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
4790 return RegMask::Empty;
4791 }
4792
4793 Node* InitializeNode::memory(uint alias_idx) {
4794 Node* mem = in(Memory);
4795 if (mem->is_MergeMem()) {
4796 return mem->as_MergeMem()->memory_at(alias_idx);
4797 } else {
4798 // incoming raw memory is not split
4799 return mem;
4800 }
4801 }
4802
4803 bool InitializeNode::is_non_zero() {
4804 if (is_complete()) return false;
4805 remove_extra_zeroes();
4806 return (req() > RawStores);
4807 }
4808
4809 void InitializeNode::set_complete(PhaseGVN* phase) {
4810 assert(!is_complete(), "caller responsibility");
4811 _is_complete = Complete;
4812
4813 // After this node is complete, it contains a bunch of
4814 // raw-memory initializations. There is no need for
4815 // it to have anything to do with non-raw memory effects.
4816 // Therefore, tell all non-raw users to re-optimize themselves,
4817 // after skipping the memory effects of this initialization.
4818 PhaseIterGVN* igvn = phase->is_IterGVN();
4819 if (igvn) igvn->add_users_to_worklist(this);
4820 }
4821
4822 // convenience function
4823 // return false if the init contains any stores already
4824 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
4825 InitializeNode* init = initialization();
4826 if (init == nullptr || init->is_complete()) {
4827 return false;
4828 }
4829 init->remove_extra_zeroes();
4830 // for now, if this allocation has already collected any inits, bail:
4831 if (init->is_non_zero()) return false;
4832 init->set_complete(phase);
4833 return true;
4834 }
4835
4836 void InitializeNode::remove_extra_zeroes() {
4837 if (req() == RawStores) return;
4838 Node* zmem = zero_memory();
4839 uint fill = RawStores;
4840 for (uint i = fill; i < req(); i++) {
4841 Node* n = in(i);
4842 if (n->is_top() || n == zmem) continue; // skip
4843 if (fill < i) set_req(fill, n); // compact
4844 ++fill;
4845 }
4846 // delete any empty spaces created:
4847 while (fill < req()) {
4848 del_req(fill);
4849 }
4850 }
4851
4852 // Helper for remembering which stores go with which offsets.
4853 intptr_t InitializeNode::get_store_offset(Node* st, PhaseValues* phase) {
4854 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
4855 intptr_t offset = -1;
4856 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
4857 phase, offset);
4858 if (base == nullptr) return -1; // something is dead,
4859 if (offset < 0) return -1; // dead, dead
4860 return offset;
4861 }
4862
4863 // Helper for proving that an initialization expression is
4864 // "simple enough" to be folded into an object initialization.
4865 // Attempts to prove that a store's initial value 'n' can be captured
4866 // within the initialization without creating a vicious cycle, such as:
4867 // { Foo p = new Foo(); p.next = p; }
4868 // True for constants and parameters and small combinations thereof.
4869 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
4870 ResourceMark rm;
4871 Unique_Node_List worklist;
4872 worklist.push(value);
4873
4874 uint complexity_limit = 20;
4875 for (uint j = 0; j < worklist.size(); j++) {
4876 if (j >= complexity_limit) {
4877 return false; // Bail out if processed too many nodes
4878 }
4879
4880 Node* n = worklist.at(j);
4881 if (n == nullptr) continue; // (can this really happen?)
4882 if (n->is_Proj()) n = n->in(0);
4883 if (n == this) return false; // found a cycle
4884 if (n->is_Con()) continue;
4885 if (n->is_Start()) continue; // params, etc., are OK
4886 if (n->is_Root()) continue; // even better
4887
4888 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
4889 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
4890 continue;
4891 }
4892
4893 Node* ctl = n->in(0);
4894 if (ctl != nullptr && !ctl->is_top()) {
4895 if (ctl->is_Proj()) ctl = ctl->in(0);
4896 if (ctl == this) return false;
4897
4898 // If we already know that the enclosing memory op is pinned right after
4899 // the init, then any control flow that the store has picked up
4900 // must have preceded the init, or else be equal to the init.
4901 // Even after loop optimizations (which might change control edges)
4902 // a store is never pinned *before* the availability of its inputs.
4903 if (!MemNode::all_controls_dominate(n, this)) {
4904 return false; // failed to prove a good control
4905 }
4906 }
4907
4908 // Check data edges for possible dependencies on 'this'.
4909 for (uint i = 1; i < n->req(); i++) {
4910 Node* m = n->in(i);
4911 if (m == nullptr || m == n || m->is_top()) continue;
4912
4913 // Only process data inputs once
4914 worklist.push(m);
4915 }
4916 }
4917
4918 return true;
4919 }
4920
4921 // Here are all the checks a Store must pass before it can be moved into
4922 // an initialization. Returns zero if a check fails.
4923 // On success, returns the (constant) offset to which the store applies,
4924 // within the initialized memory.
4925 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
4926 const int FAIL = 0;
4927 if (st->req() != MemNode::ValueIn + 1)
4928 return FAIL; // an inscrutable StoreNode (card mark?)
4929 Node* ctl = st->in(MemNode::Control);
4930 if (!(ctl != nullptr && ctl->is_Proj() && ctl->in(0) == this))
4931 return FAIL; // must be unconditional after the initialization
4932 Node* mem = st->in(MemNode::Memory);
4933 if (!(mem->is_Proj() && mem->in(0) == this))
4934 return FAIL; // must not be preceded by other stores
4935 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4936 if ((st->Opcode() == Op_StoreP || st->Opcode() == Op_StoreN) &&
4937 !bs->can_initialize_object(st)) {
4938 return FAIL;
4939 }
4940 Node* adr = st->in(MemNode::Address);
4941 intptr_t offset;
4942 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
4943 if (alloc == nullptr)
4944 return FAIL; // inscrutable address
4945 if (alloc != allocation())
4946 return FAIL; // wrong allocation! (store needs to float up)
4947 int size_in_bytes = st->memory_size();
4948 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
4949 return FAIL; // mismatched access
4950 }
4951 Node* val = st->in(MemNode::ValueIn);
4952
4953 if (!detect_init_independence(val, phase))
4954 return FAIL; // stored value must be 'simple enough'
4955
4956 // The Store can be captured only if nothing after the allocation
4957 // and before the Store is using the memory location that the store
4958 // overwrites.
4959 bool failed = false;
4960 // If is_complete_with_arraycopy() is true the shape of the graph is
4961 // well defined and is safe so no need for extra checks.
4962 if (!is_complete_with_arraycopy()) {
4963 // We are going to look at each use of the memory state following
4964 // the allocation to make sure nothing reads the memory that the
4965 // Store writes.
4966 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
4967 int alias_idx = phase->C->get_alias_index(t_adr);
4968 ResourceMark rm;
4969 Unique_Node_List mems;
4970 mems.push(mem);
4971 Node* unique_merge = nullptr;
4972 for (uint next = 0; next < mems.size(); ++next) {
4973 Node *m = mems.at(next);
4974 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
4975 Node *n = m->fast_out(j);
4976 if (n->outcnt() == 0) {
4977 continue;
4978 }
4979 if (n == st) {
4980 continue;
4981 } else if (n->in(0) != nullptr && n->in(0) != ctl) {
4982 // If the control of this use is different from the control
4983 // of the Store which is right after the InitializeNode then
4984 // this node cannot be between the InitializeNode and the
4985 // Store.
4986 continue;
4987 } else if (n->is_MergeMem()) {
4988 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
4989 // We can hit a MergeMemNode (that will likely go away
4990 // later) that is a direct use of the memory state
4991 // following the InitializeNode on the same slice as the
4992 // store node that we'd like to capture. We need to check
4993 // the uses of the MergeMemNode.
4994 mems.push(n);
4995 }
4996 } else if (n->is_Mem()) {
4997 Node* other_adr = n->in(MemNode::Address);
4998 if (other_adr == adr) {
4999 failed = true;
5000 break;
5001 } else {
5002 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
5003 if (other_t_adr != nullptr) {
5004 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
5005 if (other_alias_idx == alias_idx) {
5006 // A load from the same memory slice as the store right
5007 // after the InitializeNode. We check the control of the
5008 // object/array that is loaded from. If it's the same as
5009 // the store control then we cannot capture the store.
5010 assert(!n->is_Store(), "2 stores to same slice on same control?");
5011 Node* base = other_adr;
5012 if (base->is_Phi()) {
5013 // In rare case, base may be a PhiNode and it may read
5014 // the same memory slice between InitializeNode and store.
5015 failed = true;
5016 break;
5017 }
5018 assert(base->is_AddP(), "should be addp but is %s", base->Name());
5019 base = base->in(AddPNode::Base);
5020 if (base != nullptr) {
5021 base = base->uncast();
5022 if (base->is_Proj() && base->in(0) == alloc) {
5023 failed = true;
5024 break;
5025 }
5026 }
5027 }
5028 }
5029 }
5030 } else {
5031 failed = true;
5032 break;
5033 }
5034 }
5035 }
5036 }
5037 if (failed) {
5038 if (!can_reshape) {
5039 // We decided we couldn't capture the store during parsing. We
5040 // should try again during the next IGVN once the graph is
5041 // cleaner.
5042 phase->C->record_for_igvn(st);
5043 }
5044 return FAIL;
5045 }
5046
5047 return offset; // success
5048 }
5049
5050 // Find the captured store in(i) which corresponds to the range
5051 // [start..start+size) in the initialized object.
5052 // If there is one, return its index i. If there isn't, return the
5053 // negative of the index where it should be inserted.
5054 // Return 0 if the queried range overlaps an initialization boundary
5055 // or if dead code is encountered.
5056 // If size_in_bytes is zero, do not bother with overlap checks.
5057 int InitializeNode::captured_store_insertion_point(intptr_t start,
5058 int size_in_bytes,
5059 PhaseValues* phase) {
5060 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
5061
5062 if (is_complete())
5063 return FAIL; // arraycopy got here first; punt
5064
5065 assert(allocation() != nullptr, "must be present");
5066
5067 // no negatives, no header fields:
5068 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
5069
5070 // after a certain size, we bail out on tracking all the stores:
5071 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
5072 if (start >= ti_limit) return FAIL;
5073
5074 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
5075 if (i >= limit) return -(int)i; // not found; here is where to put it
5076
5077 Node* st = in(i);
5078 intptr_t st_off = get_store_offset(st, phase);
5079 if (st_off < 0) {
5080 if (st != zero_memory()) {
5081 return FAIL; // bail out if there is dead garbage
5082 }
5083 } else if (st_off > start) {
5084 // ...we are done, since stores are ordered
5085 if (st_off < start + size_in_bytes) {
5086 return FAIL; // the next store overlaps
5087 }
5088 return -(int)i; // not found; here is where to put it
5089 } else if (st_off < start) {
5090 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
5091 if (size_in_bytes != 0 &&
5092 start < st_off + MAX_STORE &&
5093 start < st_off + st->as_Store()->memory_size()) {
5094 return FAIL; // the previous store overlaps
5095 }
5096 } else {
5097 if (size_in_bytes != 0 &&
5098 st->as_Store()->memory_size() != size_in_bytes) {
5099 return FAIL; // mismatched store size
5100 }
5101 return i;
5102 }
5103
5104 ++i;
5105 }
5106 }
5107
5108 // Look for a captured store which initializes at the offset 'start'
5109 // with the given size. If there is no such store, and no other
5110 // initialization interferes, then return zero_memory (the memory
5111 // projection of the AllocateNode).
5112 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
5113 PhaseValues* phase) {
5114 assert(stores_are_sane(phase), "");
5115 int i = captured_store_insertion_point(start, size_in_bytes, phase);
5116 if (i == 0) {
5117 return nullptr; // something is dead
5118 } else if (i < 0) {
5119 return zero_memory(); // just primordial zero bits here
5120 } else {
5121 Node* st = in(i); // here is the store at this position
5122 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
5123 return st;
5124 }
5125 }
5126
5127 // Create, as a raw pointer, an address within my new object at 'offset'.
5128 Node* InitializeNode::make_raw_address(intptr_t offset,
5129 PhaseGVN* phase) {
5130 Node* addr = in(RawAddress);
5131 if (offset != 0) {
5132 Compile* C = phase->C;
5133 addr = phase->transform( new AddPNode(C->top(), addr,
5134 phase->MakeConX(offset)) );
5135 }
5136 return addr;
5137 }
5138
5139 // Clone the given store, converting it into a raw store
5140 // initializing a field or element of my new object.
5141 // Caller is responsible for retiring the original store,
5142 // with subsume_node or the like.
5143 //
5144 // From the example above InitializeNode::InitializeNode,
5145 // here are the old stores to be captured:
5146 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
5147 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
5148 //
5149 // Here is the changed code; note the extra edges on init:
5150 // alloc = (Allocate ...)
5151 // rawoop = alloc.RawAddress
5152 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
5153 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
5154 // init = (Initialize alloc.Control alloc.Memory rawoop
5155 // rawstore1 rawstore2)
5156 //
5157 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
5158 PhaseGVN* phase, bool can_reshape) {
5159 assert(stores_are_sane(phase), "");
5160
5161 if (start < 0) return nullptr;
5162 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
5163
5164 Compile* C = phase->C;
5165 int size_in_bytes = st->memory_size();
5166 int i = captured_store_insertion_point(start, size_in_bytes, phase);
5167 if (i == 0) return nullptr; // bail out
5168 Node* prev_mem = nullptr; // raw memory for the captured store
5169 if (i > 0) {
5170 prev_mem = in(i); // there is a pre-existing store under this one
5171 set_req(i, C->top()); // temporarily disconnect it
5172 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
5173 } else {
5174 i = -i; // no pre-existing store
5175 prev_mem = zero_memory(); // a slice of the newly allocated object
5176 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
5177 set_req(--i, C->top()); // reuse this edge; it has been folded away
5178 else
5179 ins_req(i, C->top()); // build a new edge
5180 }
5181 Node* new_st = st->clone();
5182 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
5183 new_st->set_req(MemNode::Control, in(Control));
5184 new_st->set_req(MemNode::Memory, prev_mem);
5185 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
5186 bs->eliminate_gc_barrier_data(new_st);
5187 new_st = phase->transform(new_st);
5188
5189 // At this point, new_st might have swallowed a pre-existing store
5190 // at the same offset, or perhaps new_st might have disappeared,
5191 // if it redundantly stored the same value (or zero to fresh memory).
5192
5193 // In any case, wire it in:
5194 PhaseIterGVN* igvn = phase->is_IterGVN();
5195 if (igvn) {
5196 igvn->rehash_node_delayed(this);
5197 }
5198 set_req(i, new_st);
5199
5200 // The caller may now kill the old guy.
5201 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
5202 assert(check_st == new_st || check_st == nullptr, "must be findable");
5203 assert(!is_complete(), "");
5204 return new_st;
5205 }
5206
5207 static bool store_constant(jlong* tiles, int num_tiles,
5208 intptr_t st_off, int st_size,
5209 jlong con) {
5210 if ((st_off & (st_size-1)) != 0)
5211 return false; // strange store offset (assume size==2**N)
5212 address addr = (address)tiles + st_off;
5213 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
5214 switch (st_size) {
5215 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
5216 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
5217 case sizeof(jint): *(jint*) addr = (jint) con; break;
5218 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
5219 default: return false; // strange store size (detect size!=2**N here)
5220 }
5221 return true; // return success to caller
5222 }
5223
5224 // Coalesce subword constants into int constants and possibly
5225 // into long constants. The goal, if the CPU permits,
5226 // is to initialize the object with a small number of 64-bit tiles.
5227 // Also, convert floating-point constants to bit patterns.
5228 // Non-constants are not relevant to this pass.
5229 //
5230 // In terms of the running example on InitializeNode::InitializeNode
5231 // and InitializeNode::capture_store, here is the transformation
5232 // of rawstore1 and rawstore2 into rawstore12:
5233 // alloc = (Allocate ...)
5234 // rawoop = alloc.RawAddress
5235 // tile12 = 0x00010002
5236 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
5237 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
5238 //
5239 void
5240 InitializeNode::coalesce_subword_stores(intptr_t header_size,
5241 Node* size_in_bytes,
5242 PhaseGVN* phase) {
5243 Compile* C = phase->C;
5244
5245 assert(stores_are_sane(phase), "");
5246 // Note: After this pass, they are not completely sane,
5247 // since there may be some overlaps.
5248
5249 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
5250
5251 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
5252 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
5253 size_limit = MIN2(size_limit, ti_limit);
5254 size_limit = align_up(size_limit, BytesPerLong);
5255 int num_tiles = size_limit / BytesPerLong;
5256
5257 // allocate space for the tile map:
5258 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
5259 jlong tiles_buf[small_len];
5260 Node* nodes_buf[small_len];
5261 jlong inits_buf[small_len];
5262 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
5263 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5264 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
5265 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
5266 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
5267 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
5268 // tiles: exact bitwise model of all primitive constants
5269 // nodes: last constant-storing node subsumed into the tiles model
5270 // inits: which bytes (in each tile) are touched by any initializations
5271
5272 //// Pass A: Fill in the tile model with any relevant stores.
5273
5274 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
5275 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
5276 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
5277 Node* zmem = zero_memory(); // initially zero memory state
5278 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5279 Node* st = in(i);
5280 intptr_t st_off = get_store_offset(st, phase);
5281
5282 // Figure out the store's offset and constant value:
5283 if (st_off < header_size) continue; //skip (ignore header)
5284 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
5285 int st_size = st->as_Store()->memory_size();
5286 if (st_off + st_size > size_limit) break;
5287
5288 // Record which bytes are touched, whether by constant or not.
5289 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
5290 continue; // skip (strange store size)
5291
5292 const Type* val = phase->type(st->in(MemNode::ValueIn));
5293 if (!val->singleton()) continue; //skip (non-con store)
5294 BasicType type = val->basic_type();
5295
5296 jlong con = 0;
5297 switch (type) {
5298 case T_INT: con = val->is_int()->get_con(); break;
5299 case T_LONG: con = val->is_long()->get_con(); break;
5300 case T_FLOAT: con = jint_cast(val->getf()); break;
5301 case T_DOUBLE: con = jlong_cast(val->getd()); break;
5302 default: continue; //skip (odd store type)
5303 }
5304
5305 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
5306 st->Opcode() == Op_StoreL) {
5307 continue; // This StoreL is already optimal.
5308 }
5309
5310 // Store down the constant.
5311 store_constant(tiles, num_tiles, st_off, st_size, con);
5312
5313 intptr_t j = st_off >> LogBytesPerLong;
5314
5315 if (type == T_INT && st_size == BytesPerInt
5316 && (st_off & BytesPerInt) == BytesPerInt) {
5317 jlong lcon = tiles[j];
5318 if (!Matcher::isSimpleConstant64(lcon) &&
5319 st->Opcode() == Op_StoreI) {
5320 // This StoreI is already optimal by itself.
5321 jint* intcon = (jint*) &tiles[j];
5322 intcon[1] = 0; // undo the store_constant()
5323
5324 // If the previous store is also optimal by itself, back up and
5325 // undo the action of the previous loop iteration... if we can.
5326 // But if we can't, just let the previous half take care of itself.
5327 st = nodes[j];
5328 st_off -= BytesPerInt;
5329 con = intcon[0];
5330 if (con != 0 && st != nullptr && st->Opcode() == Op_StoreI) {
5331 assert(st_off >= header_size, "still ignoring header");
5332 assert(get_store_offset(st, phase) == st_off, "must be");
5333 assert(in(i-1) == zmem, "must be");
5334 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
5335 assert(con == tcon->is_int()->get_con(), "must be");
5336 // Undo the effects of the previous loop trip, which swallowed st:
5337 intcon[0] = 0; // undo store_constant()
5338 set_req(i-1, st); // undo set_req(i, zmem)
5339 nodes[j] = nullptr; // undo nodes[j] = st
5340 --old_subword; // undo ++old_subword
5341 }
5342 continue; // This StoreI is already optimal.
5343 }
5344 }
5345
5346 // This store is not needed.
5347 set_req(i, zmem);
5348 nodes[j] = st; // record for the moment
5349 if (st_size < BytesPerLong) // something has changed
5350 ++old_subword; // includes int/float, but who's counting...
5351 else ++old_long;
5352 }
5353
5354 if ((old_subword + old_long) == 0)
5355 return; // nothing more to do
5356
5357 //// Pass B: Convert any non-zero tiles into optimal constant stores.
5358 // Be sure to insert them before overlapping non-constant stores.
5359 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
5360 for (int j = 0; j < num_tiles; j++) {
5361 jlong con = tiles[j];
5362 jlong init = inits[j];
5363 if (con == 0) continue;
5364 jint con0, con1; // split the constant, address-wise
5365 jint init0, init1; // split the init map, address-wise
5366 { union { jlong con; jint intcon[2]; } u;
5367 u.con = con;
5368 con0 = u.intcon[0];
5369 con1 = u.intcon[1];
5370 u.con = init;
5371 init0 = u.intcon[0];
5372 init1 = u.intcon[1];
5373 }
5374
5375 Node* old = nodes[j];
5376 assert(old != nullptr, "need the prior store");
5377 intptr_t offset = (j * BytesPerLong);
5378
5379 bool split = !Matcher::isSimpleConstant64(con);
5380
5381 if (offset < header_size) {
5382 assert(offset + BytesPerInt >= header_size, "second int counts");
5383 assert(*(jint*)&tiles[j] == 0, "junk in header");
5384 split = true; // only the second word counts
5385 // Example: int a[] = { 42 ... }
5386 } else if (con0 == 0 && init0 == -1) {
5387 split = true; // first word is covered by full inits
5388 // Example: int a[] = { ... foo(), 42 ... }
5389 } else if (con1 == 0 && init1 == -1) {
5390 split = true; // second word is covered by full inits
5391 // Example: int a[] = { ... 42, foo() ... }
5392 }
5393
5394 // Here's a case where init0 is neither 0 nor -1:
5395 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
5396 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
5397 // In this case the tile is not split; it is (jlong)42.
5398 // The big tile is stored down, and then the foo() value is inserted.
5399 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
5400
5401 Node* ctl = old->in(MemNode::Control);
5402 Node* adr = make_raw_address(offset, phase);
5403 const TypePtr* atp = TypeRawPtr::BOTTOM;
5404
5405 // One or two coalesced stores to plop down.
5406 Node* st[2];
5407 intptr_t off[2];
5408 int nst = 0;
5409 if (!split) {
5410 ++new_long;
5411 off[nst] = offset;
5412 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5413 phase->longcon(con), T_LONG, MemNode::unordered);
5414 } else {
5415 // Omit either if it is a zero.
5416 if (con0 != 0) {
5417 ++new_int;
5418 off[nst] = offset;
5419 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5420 phase->intcon(con0), T_INT, MemNode::unordered);
5421 }
5422 if (con1 != 0) {
5423 ++new_int;
5424 offset += BytesPerInt;
5425 adr = make_raw_address(offset, phase);
5426 off[nst] = offset;
5427 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
5428 phase->intcon(con1), T_INT, MemNode::unordered);
5429 }
5430 }
5431
5432 // Insert second store first, then the first before the second.
5433 // Insert each one just before any overlapping non-constant stores.
5434 while (nst > 0) {
5435 Node* st1 = st[--nst];
5436 C->copy_node_notes_to(st1, old);
5437 st1 = phase->transform(st1);
5438 offset = off[nst];
5439 assert(offset >= header_size, "do not smash header");
5440 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
5441 guarantee(ins_idx != 0, "must re-insert constant store");
5442 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
5443 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
5444 set_req(--ins_idx, st1);
5445 else
5446 ins_req(ins_idx, st1);
5447 }
5448 }
5449
5450 if (PrintCompilation && WizardMode)
5451 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
5452 old_subword, old_long, new_int, new_long);
5453 if (C->log() != nullptr)
5454 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
5455 old_subword, old_long, new_int, new_long);
5456
5457 // Clean up any remaining occurrences of zmem:
5458 remove_extra_zeroes();
5459 }
5460
5461 // Explore forward from in(start) to find the first fully initialized
5462 // word, and return its offset. Skip groups of subword stores which
5463 // together initialize full words. If in(start) is itself part of a
5464 // fully initialized word, return the offset of in(start). If there
5465 // are no following full-word stores, or if something is fishy, return
5466 // a negative value.
5467 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
5468 int int_map = 0;
5469 intptr_t int_map_off = 0;
5470 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
5471
5472 for (uint i = start, limit = req(); i < limit; i++) {
5473 Node* st = in(i);
5474
5475 intptr_t st_off = get_store_offset(st, phase);
5476 if (st_off < 0) break; // return conservative answer
5477
5478 int st_size = st->as_Store()->memory_size();
5479 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
5480 return st_off; // we found a complete word init
5481 }
5482
5483 // update the map:
5484
5485 intptr_t this_int_off = align_down(st_off, BytesPerInt);
5486 if (this_int_off != int_map_off) {
5487 // reset the map:
5488 int_map = 0;
5489 int_map_off = this_int_off;
5490 }
5491
5492 int subword_off = st_off - this_int_off;
5493 int_map |= right_n_bits(st_size) << subword_off;
5494 if ((int_map & FULL_MAP) == FULL_MAP) {
5495 return this_int_off; // we found a complete word init
5496 }
5497
5498 // Did this store hit or cross the word boundary?
5499 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
5500 if (next_int_off == this_int_off + BytesPerInt) {
5501 // We passed the current int, without fully initializing it.
5502 int_map_off = next_int_off;
5503 int_map >>= BytesPerInt;
5504 } else if (next_int_off > this_int_off + BytesPerInt) {
5505 // We passed the current and next int.
5506 return this_int_off + BytesPerInt;
5507 }
5508 }
5509
5510 return -1;
5511 }
5512
5513
5514 // Called when the associated AllocateNode is expanded into CFG.
5515 // At this point, we may perform additional optimizations.
5516 // Linearize the stores by ascending offset, to make memory
5517 // activity as coherent as possible.
5518 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
5519 intptr_t header_size,
5520 Node* size_in_bytes,
5521 PhaseIterGVN* phase) {
5522 assert(!is_complete(), "not already complete");
5523 assert(stores_are_sane(phase), "");
5524 assert(allocation() != nullptr, "must be present");
5525
5526 remove_extra_zeroes();
5527
5528 if (ReduceFieldZeroing || ReduceBulkZeroing)
5529 // reduce instruction count for common initialization patterns
5530 coalesce_subword_stores(header_size, size_in_bytes, phase);
5531
5532 Node* zmem = zero_memory(); // initially zero memory state
5533 Node* inits = zmem; // accumulating a linearized chain of inits
5534 #ifdef ASSERT
5535 intptr_t first_offset = allocation()->minimum_header_size();
5536 intptr_t last_init_off = first_offset; // previous init offset
5537 intptr_t last_init_end = first_offset; // previous init offset+size
5538 intptr_t last_tile_end = first_offset; // previous tile offset+size
5539 #endif
5540 intptr_t zeroes_done = header_size;
5541
5542 bool do_zeroing = true; // we might give up if inits are very sparse
5543 int big_init_gaps = 0; // how many large gaps have we seen?
5544
5545 if (UseTLAB && ZeroTLAB) do_zeroing = false;
5546 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
5547
5548 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
5549 Node* st = in(i);
5550 intptr_t st_off = get_store_offset(st, phase);
5551 if (st_off < 0)
5552 break; // unknown junk in the inits
5553 if (st->in(MemNode::Memory) != zmem)
5554 break; // complicated store chains somehow in list
5555
5556 int st_size = st->as_Store()->memory_size();
5557 intptr_t next_init_off = st_off + st_size;
5558
5559 if (do_zeroing && zeroes_done < next_init_off) {
5560 // See if this store needs a zero before it or under it.
5561 intptr_t zeroes_needed = st_off;
5562
5563 if (st_size < BytesPerInt) {
5564 // Look for subword stores which only partially initialize words.
5565 // If we find some, we must lay down some word-level zeroes first,
5566 // underneath the subword stores.
5567 //
5568 // Examples:
5569 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
5570 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
5571 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
5572 //
5573 // Note: coalesce_subword_stores may have already done this,
5574 // if it was prompted by constant non-zero subword initializers.
5575 // But this case can still arise with non-constant stores.
5576
5577 intptr_t next_full_store = find_next_fullword_store(i, phase);
5578
5579 // In the examples above:
5580 // in(i) p q r s x y z
5581 // st_off 12 13 14 15 12 13 14
5582 // st_size 1 1 1 1 1 1 1
5583 // next_full_s. 12 16 16 16 16 16 16
5584 // z's_done 12 16 16 16 12 16 12
5585 // z's_needed 12 16 16 16 16 16 16
5586 // zsize 0 0 0 0 4 0 4
5587 if (next_full_store < 0) {
5588 // Conservative tack: Zero to end of current word.
5589 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
5590 } else {
5591 // Zero to beginning of next fully initialized word.
5592 // Or, don't zero at all, if we are already in that word.
5593 assert(next_full_store >= zeroes_needed, "must go forward");
5594 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
5595 zeroes_needed = next_full_store;
5596 }
5597 }
5598
5599 if (zeroes_needed > zeroes_done) {
5600 intptr_t zsize = zeroes_needed - zeroes_done;
5601 // Do some incremental zeroing on rawmem, in parallel with inits.
5602 zeroes_done = align_down(zeroes_done, BytesPerInt);
5603 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5604 allocation()->in(AllocateNode::InitValue),
5605 allocation()->in(AllocateNode::RawInitValue),
5606 zeroes_done, zeroes_needed,
5607 phase);
5608 zeroes_done = zeroes_needed;
5609 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
5610 do_zeroing = false; // leave the hole, next time
5611 }
5612 }
5613
5614 // Collect the store and move on:
5615 phase->replace_input_of(st, MemNode::Memory, inits);
5616 inits = st; // put it on the linearized chain
5617 set_req(i, zmem); // unhook from previous position
5618
5619 if (zeroes_done == st_off)
5620 zeroes_done = next_init_off;
5621
5622 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
5623
5624 #ifdef ASSERT
5625 // Various order invariants. Weaker than stores_are_sane because
5626 // a large constant tile can be filled in by smaller non-constant stores.
5627 assert(st_off >= last_init_off, "inits do not reverse");
5628 last_init_off = st_off;
5629 const Type* val = nullptr;
5630 if (st_size >= BytesPerInt &&
5631 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
5632 (int)val->basic_type() < (int)T_OBJECT) {
5633 assert(st_off >= last_tile_end, "tiles do not overlap");
5634 assert(st_off >= last_init_end, "tiles do not overwrite inits");
5635 last_tile_end = MAX2(last_tile_end, next_init_off);
5636 } else {
5637 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
5638 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
5639 assert(st_off >= last_init_end, "inits do not overlap");
5640 last_init_end = next_init_off; // it's a non-tile
5641 }
5642 #endif //ASSERT
5643 }
5644
5645 remove_extra_zeroes(); // clear out all the zmems left over
5646 add_req(inits);
5647
5648 if (!(UseTLAB && ZeroTLAB)) {
5649 // If anything remains to be zeroed, zero it all now.
5650 zeroes_done = align_down(zeroes_done, BytesPerInt);
5651 // if it is the last unused 4 bytes of an instance, forget about it
5652 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
5653 if (zeroes_done + BytesPerLong >= size_limit) {
5654 AllocateNode* alloc = allocation();
5655 assert(alloc != nullptr, "must be present");
5656 if (alloc != nullptr && alloc->Opcode() == Op_Allocate) {
5657 Node* klass_node = alloc->in(AllocateNode::KlassNode);
5658 ciKlass* k = phase->type(klass_node)->is_instklassptr()->instance_klass();
5659 if (zeroes_done == k->layout_helper())
5660 zeroes_done = size_limit;
5661 }
5662 }
5663 if (zeroes_done < size_limit) {
5664 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
5665 allocation()->in(AllocateNode::InitValue),
5666 allocation()->in(AllocateNode::RawInitValue),
5667 zeroes_done, size_in_bytes, phase);
5668 }
5669 }
5670
5671 set_complete(phase);
5672 return rawmem;
5673 }
5674
5675
5676 #ifdef ASSERT
5677 bool InitializeNode::stores_are_sane(PhaseValues* phase) {
5678 if (is_complete())
5679 return true; // stores could be anything at this point
5680 assert(allocation() != nullptr, "must be present");
5681 intptr_t last_off = allocation()->minimum_header_size();
5682 for (uint i = InitializeNode::RawStores; i < req(); i++) {
5683 Node* st = in(i);
5684 intptr_t st_off = get_store_offset(st, phase);
5685 if (st_off < 0) continue; // ignore dead garbage
5686 if (last_off > st_off) {
5687 tty->print_cr("*** bad store offset at %d: %zd > %zd", i, last_off, st_off);
5688 this->dump(2);
5689 assert(false, "ascending store offsets");
5690 return false;
5691 }
5692 last_off = st_off + st->as_Store()->memory_size();
5693 }
5694 return true;
5695 }
5696 #endif //ASSERT
5697
5698
5699
5700
5701 //============================MergeMemNode=====================================
5702 //
5703 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
5704 // contributing store or call operations. Each contributor provides the memory
5705 // state for a particular "alias type" (see Compile::alias_type). For example,
5706 // if a MergeMem has an input X for alias category #6, then any memory reference
5707 // to alias category #6 may use X as its memory state input, as an exact equivalent
5708 // to using the MergeMem as a whole.
5709 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
5710 //
5711 // (Here, the <N> notation gives the index of the relevant adr_type.)
5712 //
5713 // In one special case (and more cases in the future), alias categories overlap.
5714 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
5715 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
5716 // it is exactly equivalent to that state W:
5717 // MergeMem(<Bot>: W) <==> W
5718 //
5719 // Usually, the merge has more than one input. In that case, where inputs
5720 // overlap (i.e., one is Bot), the narrower alias type determines the memory
5721 // state for that type, and the wider alias type (Bot) fills in everywhere else:
5722 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
5723 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
5724 //
5725 // A merge can take a "wide" memory state as one of its narrow inputs.
5726 // This simply means that the merge observes out only the relevant parts of
5727 // the wide input. That is, wide memory states arriving at narrow merge inputs
5728 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
5729 //
5730 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
5731 // and that memory slices "leak through":
5732 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
5733 //
5734 // But, in such a cascade, repeated memory slices can "block the leak":
5735 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
5736 //
5737 // In the last example, Y is not part of the combined memory state of the
5738 // outermost MergeMem. The system must, of course, prevent unschedulable
5739 // memory states from arising, so you can be sure that the state Y is somehow
5740 // a precursor to state Y'.
5741 //
5742 //
5743 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
5744 // of each MergeMemNode array are exactly the numerical alias indexes, including
5745 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
5746 // Compile::alias_type (and kin) produce and manage these indexes.
5747 //
5748 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
5749 // (Note that this provides quick access to the top node inside MergeMem methods,
5750 // without the need to reach out via TLS to Compile::current.)
5751 //
5752 // As a consequence of what was just described, a MergeMem that represents a full
5753 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
5754 // containing all alias categories.
5755 //
5756 // MergeMem nodes never (?) have control inputs, so in(0) is null.
5757 //
5758 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
5759 // a memory state for the alias type <N>, or else the top node, meaning that
5760 // there is no particular input for that alias type. Note that the length of
5761 // a MergeMem is variable, and may be extended at any time to accommodate new
5762 // memory states at larger alias indexes. When merges grow, they are of course
5763 // filled with "top" in the unused in() positions.
5764 //
5765 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
5766 // (Top was chosen because it works smoothly with passes like GCM.)
5767 //
5768 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
5769 // the type of random VM bits like TLS references.) Since it is always the
5770 // first non-Bot memory slice, some low-level loops use it to initialize an
5771 // index variable: for (i = AliasIdxRaw; i < req(); i++).
5772 //
5773 //
5774 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
5775 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
5776 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
5777 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
5778 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
5779 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
5780 //
5781 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
5782 // really that different from the other memory inputs. An abbreviation called
5783 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
5784 //
5785 //
5786 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
5787 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
5788 // that "emerges though" the base memory will be marked as excluding the alias types
5789 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
5790 //
5791 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
5792 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
5793 //
5794 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
5795 // (It is currently unimplemented.) As you can see, the resulting merge is
5796 // actually a disjoint union of memory states, rather than an overlay.
5797 //
5798
5799 //------------------------------MergeMemNode-----------------------------------
5800 Node* MergeMemNode::make_empty_memory() {
5801 Node* empty_memory = (Node*) Compile::current()->top();
5802 assert(empty_memory->is_top(), "correct sentinel identity");
5803 return empty_memory;
5804 }
5805
5806 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
5807 init_class_id(Class_MergeMem);
5808 // all inputs are nullified in Node::Node(int)
5809 // set_input(0, nullptr); // no control input
5810
5811 // Initialize the edges uniformly to top, for starters.
5812 Node* empty_mem = make_empty_memory();
5813 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
5814 init_req(i,empty_mem);
5815 }
5816 assert(empty_memory() == empty_mem, "");
5817
5818 if( new_base != nullptr && new_base->is_MergeMem() ) {
5819 MergeMemNode* mdef = new_base->as_MergeMem();
5820 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
5821 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
5822 mms.set_memory(mms.memory2());
5823 }
5824 assert(base_memory() == mdef->base_memory(), "");
5825 } else {
5826 set_base_memory(new_base);
5827 }
5828 }
5829
5830 // Make a new, untransformed MergeMem with the same base as 'mem'.
5831 // If mem is itself a MergeMem, populate the result with the same edges.
5832 MergeMemNode* MergeMemNode::make(Node* mem) {
5833 return new MergeMemNode(mem);
5834 }
5835
5836 //------------------------------cmp--------------------------------------------
5837 uint MergeMemNode::hash() const { return NO_HASH; }
5838 bool MergeMemNode::cmp( const Node &n ) const {
5839 return (&n == this); // Always fail except on self
5840 }
5841
5842 //------------------------------Identity---------------------------------------
5843 Node* MergeMemNode::Identity(PhaseGVN* phase) {
5844 // Identity if this merge point does not record any interesting memory
5845 // disambiguations.
5846 Node* base_mem = base_memory();
5847 Node* empty_mem = empty_memory();
5848 if (base_mem != empty_mem) { // Memory path is not dead?
5849 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5850 Node* mem = in(i);
5851 if (mem != empty_mem && mem != base_mem) {
5852 return this; // Many memory splits; no change
5853 }
5854 }
5855 }
5856 return base_mem; // No memory splits; ID on the one true input
5857 }
5858
5859 //------------------------------Ideal------------------------------------------
5860 // This method is invoked recursively on chains of MergeMem nodes
5861 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
5862 // Remove chain'd MergeMems
5863 //
5864 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
5865 // relative to the "in(Bot)". Since we are patching both at the same time,
5866 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
5867 // but rewrite each "in(i)" relative to the new "in(Bot)".
5868 Node *progress = nullptr;
5869
5870
5871 Node* old_base = base_memory();
5872 Node* empty_mem = empty_memory();
5873 if (old_base == empty_mem)
5874 return nullptr; // Dead memory path.
5875
5876 MergeMemNode* old_mbase;
5877 if (old_base != nullptr && old_base->is_MergeMem())
5878 old_mbase = old_base->as_MergeMem();
5879 else
5880 old_mbase = nullptr;
5881 Node* new_base = old_base;
5882
5883 // simplify stacked MergeMems in base memory
5884 if (old_mbase) new_base = old_mbase->base_memory();
5885
5886 // the base memory might contribute new slices beyond my req()
5887 if (old_mbase) grow_to_match(old_mbase);
5888
5889 // Note: We do not call verify_sparse on entry, because inputs
5890 // can normalize to the base_memory via subsume_node or similar
5891 // mechanisms. This method repairs that damage.
5892
5893 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
5894
5895 // Look at each slice.
5896 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5897 Node* old_in = in(i);
5898 // calculate the old memory value
5899 Node* old_mem = old_in;
5900 if (old_mem == empty_mem) old_mem = old_base;
5901 assert(old_mem == memory_at(i), "");
5902
5903 // maybe update (reslice) the old memory value
5904
5905 // simplify stacked MergeMems
5906 Node* new_mem = old_mem;
5907 MergeMemNode* old_mmem;
5908 if (old_mem != nullptr && old_mem->is_MergeMem())
5909 old_mmem = old_mem->as_MergeMem();
5910 else
5911 old_mmem = nullptr;
5912 if (old_mmem == this) {
5913 // This can happen if loops break up and safepoints disappear.
5914 // A merge of BotPtr (default) with a RawPtr memory derived from a
5915 // safepoint can be rewritten to a merge of the same BotPtr with
5916 // the BotPtr phi coming into the loop. If that phi disappears
5917 // also, we can end up with a self-loop of the mergemem.
5918 // In general, if loops degenerate and memory effects disappear,
5919 // a mergemem can be left looking at itself. This simply means
5920 // that the mergemem's default should be used, since there is
5921 // no longer any apparent effect on this slice.
5922 // Note: If a memory slice is a MergeMem cycle, it is unreachable
5923 // from start. Update the input to TOP.
5924 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
5925 }
5926 else if (old_mmem != nullptr) {
5927 new_mem = old_mmem->memory_at(i);
5928 }
5929 // else preceding memory was not a MergeMem
5930
5931 // maybe store down a new value
5932 Node* new_in = new_mem;
5933 if (new_in == new_base) new_in = empty_mem;
5934
5935 if (new_in != old_in) {
5936 // Warning: Do not combine this "if" with the previous "if"
5937 // A memory slice might have be be rewritten even if it is semantically
5938 // unchanged, if the base_memory value has changed.
5939 set_req_X(i, new_in, phase);
5940 progress = this; // Report progress
5941 }
5942 }
5943
5944 if (new_base != old_base) {
5945 set_req_X(Compile::AliasIdxBot, new_base, phase);
5946 // Don't use set_base_memory(new_base), because we need to update du.
5947 assert(base_memory() == new_base, "");
5948 progress = this;
5949 }
5950
5951 if( base_memory() == this ) {
5952 // a self cycle indicates this memory path is dead
5953 set_req(Compile::AliasIdxBot, empty_mem);
5954 }
5955
5956 // Resolve external cycles by calling Ideal on a MergeMem base_memory
5957 // Recursion must occur after the self cycle check above
5958 if( base_memory()->is_MergeMem() ) {
5959 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
5960 Node *m = phase->transform(new_mbase); // Rollup any cycles
5961 if( m != nullptr &&
5962 (m->is_top() ||
5963 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
5964 // propagate rollup of dead cycle to self
5965 set_req(Compile::AliasIdxBot, empty_mem);
5966 }
5967 }
5968
5969 if( base_memory() == empty_mem ) {
5970 progress = this;
5971 // Cut inputs during Parse phase only.
5972 // During Optimize phase a dead MergeMem node will be subsumed by Top.
5973 if( !can_reshape ) {
5974 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5975 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
5976 }
5977 }
5978 }
5979
5980 if( !progress && base_memory()->is_Phi() && can_reshape ) {
5981 // Check if PhiNode::Ideal's "Split phis through memory merges"
5982 // transform should be attempted. Look for this->phi->this cycle.
5983 uint merge_width = req();
5984 if (merge_width > Compile::AliasIdxRaw) {
5985 PhiNode* phi = base_memory()->as_Phi();
5986 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
5987 if (phi->in(i) == this) {
5988 phase->is_IterGVN()->_worklist.push(phi);
5989 break;
5990 }
5991 }
5992 }
5993 }
5994
5995 assert(progress || verify_sparse(), "please, no dups of base");
5996 return progress;
5997 }
5998
5999 //-------------------------set_base_memory-------------------------------------
6000 void MergeMemNode::set_base_memory(Node *new_base) {
6001 Node* empty_mem = empty_memory();
6002 set_req(Compile::AliasIdxBot, new_base);
6003 assert(memory_at(req()) == new_base, "must set default memory");
6004 // Clear out other occurrences of new_base:
6005 if (new_base != empty_mem) {
6006 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6007 if (in(i) == new_base) set_req(i, empty_mem);
6008 }
6009 }
6010 }
6011
6012 //------------------------------out_RegMask------------------------------------
6013 const RegMask &MergeMemNode::out_RegMask() const {
6014 return RegMask::Empty;
6015 }
6016
6017 //------------------------------dump_spec--------------------------------------
6018 #ifndef PRODUCT
6019 void MergeMemNode::dump_spec(outputStream *st) const {
6020 st->print(" {");
6021 Node* base_mem = base_memory();
6022 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
6023 Node* mem = (in(i) != nullptr) ? memory_at(i) : base_mem;
6024 if (mem == base_mem) { st->print(" -"); continue; }
6025 st->print( " N%d:", mem->_idx );
6026 Compile::current()->get_adr_type(i)->dump_on(st);
6027 }
6028 st->print(" }");
6029 }
6030 #endif // !PRODUCT
6031
6032
6033 #ifdef ASSERT
6034 static bool might_be_same(Node* a, Node* b) {
6035 if (a == b) return true;
6036 if (!(a->is_Phi() || b->is_Phi())) return false;
6037 // phis shift around during optimization
6038 return true; // pretty stupid...
6039 }
6040
6041 // verify a narrow slice (either incoming or outgoing)
6042 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
6043 if (!VerifyAliases) return; // don't bother to verify unless requested
6044 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
6045 if (Node::in_dump()) return; // muzzle asserts when printing
6046 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
6047 assert(n != nullptr, "");
6048 // Elide intervening MergeMem's
6049 while (n->is_MergeMem()) {
6050 n = n->as_MergeMem()->memory_at(alias_idx);
6051 }
6052 Compile* C = Compile::current();
6053 const TypePtr* n_adr_type = n->adr_type();
6054 if (n == m->empty_memory()) {
6055 // Implicit copy of base_memory()
6056 } else if (n_adr_type != TypePtr::BOTTOM) {
6057 assert(n_adr_type != nullptr, "new memory must have a well-defined adr_type");
6058 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
6059 } else {
6060 // A few places like make_runtime_call "know" that VM calls are narrow,
6061 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
6062 bool expected_wide_mem = false;
6063 if (n == m->base_memory()) {
6064 expected_wide_mem = true;
6065 } else if (alias_idx == Compile::AliasIdxRaw ||
6066 n == m->memory_at(Compile::AliasIdxRaw)) {
6067 expected_wide_mem = true;
6068 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
6069 // memory can "leak through" calls on channels that
6070 // are write-once. Allow this also.
6071 expected_wide_mem = true;
6072 }
6073 assert(expected_wide_mem, "expected narrow slice replacement");
6074 }
6075 }
6076 #else // !ASSERT
6077 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
6078 #endif
6079
6080
6081 //-----------------------------memory_at---------------------------------------
6082 Node* MergeMemNode::memory_at(uint alias_idx) const {
6083 assert(alias_idx >= Compile::AliasIdxRaw ||
6084 (alias_idx == Compile::AliasIdxBot && !Compile::current()->do_aliasing()),
6085 "must avoid base_memory and AliasIdxTop");
6086
6087 // Otherwise, it is a narrow slice.
6088 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
6089 if (is_empty_memory(n)) {
6090 // the array is sparse; empty slots are the "top" node
6091 n = base_memory();
6092 assert(Node::in_dump()
6093 || n == nullptr || n->bottom_type() == Type::TOP
6094 || n->adr_type() == nullptr // address is TOP
6095 || n->adr_type() == TypePtr::BOTTOM
6096 || n->adr_type() == TypeRawPtr::BOTTOM
6097 || !Compile::current()->do_aliasing(),
6098 "must be a wide memory");
6099 // do_aliasing == false if we are organizing the memory states manually.
6100 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
6101 } else {
6102 // make sure the stored slice is sane
6103 #ifdef ASSERT
6104 if (VMError::is_error_reported() || Node::in_dump()) {
6105 } else if (might_be_same(n, base_memory())) {
6106 // Give it a pass: It is a mostly harmless repetition of the base.
6107 // This can arise normally from node subsumption during optimization.
6108 } else {
6109 verify_memory_slice(this, alias_idx, n);
6110 }
6111 #endif
6112 }
6113 return n;
6114 }
6115
6116 //---------------------------set_memory_at-------------------------------------
6117 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
6118 verify_memory_slice(this, alias_idx, n);
6119 Node* empty_mem = empty_memory();
6120 if (n == base_memory()) n = empty_mem; // collapse default
6121 uint need_req = alias_idx+1;
6122 if (req() < need_req) {
6123 if (n == empty_mem) return; // already the default, so do not grow me
6124 // grow the sparse array
6125 do {
6126 add_req(empty_mem);
6127 } while (req() < need_req);
6128 }
6129 set_req( alias_idx, n );
6130 }
6131
6132
6133
6134 //--------------------------iteration_setup------------------------------------
6135 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
6136 if (other != nullptr) {
6137 grow_to_match(other);
6138 // invariant: the finite support of mm2 is within mm->req()
6139 #ifdef ASSERT
6140 for (uint i = req(); i < other->req(); i++) {
6141 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
6142 }
6143 #endif
6144 }
6145 // Replace spurious copies of base_memory by top.
6146 Node* base_mem = base_memory();
6147 if (base_mem != nullptr && !base_mem->is_top()) {
6148 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
6149 if (in(i) == base_mem)
6150 set_req(i, empty_memory());
6151 }
6152 }
6153 }
6154
6155 //---------------------------grow_to_match-------------------------------------
6156 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
6157 Node* empty_mem = empty_memory();
6158 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
6159 // look for the finite support of the other memory
6160 for (uint i = other->req(); --i >= req(); ) {
6161 if (other->in(i) != empty_mem) {
6162 uint new_len = i+1;
6163 while (req() < new_len) add_req(empty_mem);
6164 break;
6165 }
6166 }
6167 }
6168
6169 //---------------------------verify_sparse-------------------------------------
6170 #ifndef PRODUCT
6171 bool MergeMemNode::verify_sparse() const {
6172 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
6173 Node* base_mem = base_memory();
6174 // The following can happen in degenerate cases, since empty==top.
6175 if (is_empty_memory(base_mem)) return true;
6176 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
6177 assert(in(i) != nullptr, "sane slice");
6178 if (in(i) == base_mem) return false; // should have been the sentinel value!
6179 }
6180 return true;
6181 }
6182
6183 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
6184 Node* n;
6185 n = mm->in(idx);
6186 if (mem == n) return true; // might be empty_memory()
6187 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
6188 if (mem == n) return true;
6189 return false;
6190 }
6191 #endif // !PRODUCT