Mon, 25 Feb 2013 14:13:04 +0100
8007294: ReduceFieldZeroing doesn't check for dependent load and can lead to incorrect execution
Summary: InitializeNode::can_capture_store() must check that the captured store doesn't overwrite a memory location that is loaded before the store.
Reviewed-by: kvn
1 /*
2 * Copyright (c) 1997, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "oops/objArrayKlass.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/cfgnode.hpp"
32 #include "opto/compile.hpp"
33 #include "opto/connode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/machnode.hpp"
36 #include "opto/matcher.hpp"
37 #include "opto/memnode.hpp"
38 #include "opto/mulnode.hpp"
39 #include "opto/phaseX.hpp"
40 #include "opto/regmask.hpp"
42 // Portions of code courtesy of Clifford Click
44 // Optimization - Graph Style
46 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
48 //=============================================================================
49 uint MemNode::size_of() const { return sizeof(*this); }
51 const TypePtr *MemNode::adr_type() const {
52 Node* adr = in(Address);
53 const TypePtr* cross_check = NULL;
54 DEBUG_ONLY(cross_check = _adr_type);
55 return calculate_adr_type(adr->bottom_type(), cross_check);
56 }
58 #ifndef PRODUCT
59 void MemNode::dump_spec(outputStream *st) const {
60 if (in(Address) == NULL) return; // node is dead
61 #ifndef ASSERT
62 // fake the missing field
63 const TypePtr* _adr_type = NULL;
64 if (in(Address) != NULL)
65 _adr_type = in(Address)->bottom_type()->isa_ptr();
66 #endif
67 dump_adr_type(this, _adr_type, st);
69 Compile* C = Compile::current();
70 if( C->alias_type(_adr_type)->is_volatile() )
71 st->print(" Volatile!");
72 }
74 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
75 st->print(" @");
76 if (adr_type == NULL) {
77 st->print("NULL");
78 } else {
79 adr_type->dump_on(st);
80 Compile* C = Compile::current();
81 Compile::AliasType* atp = NULL;
82 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
83 if (atp == NULL)
84 st->print(", idx=?\?;");
85 else if (atp->index() == Compile::AliasIdxBot)
86 st->print(", idx=Bot;");
87 else if (atp->index() == Compile::AliasIdxTop)
88 st->print(", idx=Top;");
89 else if (atp->index() == Compile::AliasIdxRaw)
90 st->print(", idx=Raw;");
91 else {
92 ciField* field = atp->field();
93 if (field) {
94 st->print(", name=");
95 field->print_name_on(st);
96 }
97 st->print(", idx=%d;", atp->index());
98 }
99 }
100 }
102 extern void print_alias_types();
104 #endif
106 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
107 const TypeOopPtr *tinst = t_adr->isa_oopptr();
108 if (tinst == NULL || !tinst->is_known_instance_field())
109 return mchain; // don't try to optimize non-instance types
110 uint instance_id = tinst->instance_id();
111 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
112 Node *prev = NULL;
113 Node *result = mchain;
114 while (prev != result) {
115 prev = result;
116 if (result == start_mem)
117 break; // hit one of our sentinels
118 // skip over a call which does not affect this memory slice
119 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
120 Node *proj_in = result->in(0);
121 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
122 break; // hit one of our sentinels
123 } else if (proj_in->is_Call()) {
124 CallNode *call = proj_in->as_Call();
125 if (!call->may_modify(t_adr, phase)) {
126 result = call->in(TypeFunc::Memory);
127 }
128 } else if (proj_in->is_Initialize()) {
129 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
130 // Stop if this is the initialization for the object instance which
131 // which contains this memory slice, otherwise skip over it.
132 if (alloc != NULL && alloc->_idx != instance_id) {
133 result = proj_in->in(TypeFunc::Memory);
134 }
135 } else if (proj_in->is_MemBar()) {
136 result = proj_in->in(TypeFunc::Memory);
137 } else {
138 assert(false, "unexpected projection");
139 }
140 } else if (result->is_ClearArray()) {
141 if (!ClearArrayNode::step_through(&result, instance_id, phase)) {
142 // Can not bypass initialization of the instance
143 // we are looking for.
144 break;
145 }
146 // Otherwise skip it (the call updated 'result' value).
147 } else if (result->is_MergeMem()) {
148 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
149 }
150 }
151 return result;
152 }
154 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
155 const TypeOopPtr *t_oop = t_adr->isa_oopptr();
156 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field();
157 PhaseIterGVN *igvn = phase->is_IterGVN();
158 Node *result = mchain;
159 result = optimize_simple_memory_chain(result, t_adr, phase);
160 if (is_instance && igvn != NULL && result->is_Phi()) {
161 PhiNode *mphi = result->as_Phi();
162 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
163 const TypePtr *t = mphi->adr_type();
164 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
165 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
166 t->is_oopptr()->cast_to_exactness(true)
167 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
168 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
169 // clone the Phi with our address type
170 result = mphi->split_out_instance(t_adr, igvn);
171 } else {
172 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
173 }
174 }
175 return result;
176 }
178 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
179 uint alias_idx = phase->C->get_alias_index(tp);
180 Node *mem = mmem;
181 #ifdef ASSERT
182 {
183 // Check that current type is consistent with the alias index used during graph construction
184 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
185 bool consistent = adr_check == NULL || adr_check->empty() ||
186 phase->C->must_alias(adr_check, alias_idx );
187 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
188 if( !consistent && adr_check != NULL && !adr_check->empty() &&
189 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
190 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
191 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
192 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
193 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
194 // don't assert if it is dead code.
195 consistent = true;
196 }
197 if( !consistent ) {
198 st->print("alias_idx==%d, adr_check==", alias_idx);
199 if( adr_check == NULL ) {
200 st->print("NULL");
201 } else {
202 adr_check->dump();
203 }
204 st->cr();
205 print_alias_types();
206 assert(consistent, "adr_check must match alias idx");
207 }
208 }
209 #endif
210 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
211 // means an array I have not precisely typed yet. Do not do any
212 // alias stuff with it any time soon.
213 const TypeOopPtr *toop = tp->isa_oopptr();
214 if( tp->base() != Type::AnyPtr &&
215 !(toop &&
216 toop->klass() != NULL &&
217 toop->klass()->is_java_lang_Object() &&
218 toop->offset() == Type::OffsetBot) ) {
219 // compress paths and change unreachable cycles to TOP
220 // If not, we can update the input infinitely along a MergeMem cycle
221 // Equivalent code in PhiNode::Ideal
222 Node* m = phase->transform(mmem);
223 // If transformed to a MergeMem, get the desired slice
224 // Otherwise the returned node represents memory for every slice
225 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
226 // Update input if it is progress over what we have now
227 }
228 return mem;
229 }
231 //--------------------------Ideal_common---------------------------------------
232 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
233 // Unhook non-raw memories from complete (macro-expanded) initializations.
234 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
235 // If our control input is a dead region, kill all below the region
236 Node *ctl = in(MemNode::Control);
237 if (ctl && remove_dead_region(phase, can_reshape))
238 return this;
239 ctl = in(MemNode::Control);
240 // Don't bother trying to transform a dead node
241 if( ctl && ctl->is_top() ) return NodeSentinel;
243 PhaseIterGVN *igvn = phase->is_IterGVN();
244 // Wait if control on the worklist.
245 if (ctl && can_reshape && igvn != NULL) {
246 Node* bol = NULL;
247 Node* cmp = NULL;
248 if (ctl->in(0)->is_If()) {
249 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
250 bol = ctl->in(0)->in(1);
251 if (bol->is_Bool())
252 cmp = ctl->in(0)->in(1)->in(1);
253 }
254 if (igvn->_worklist.member(ctl) ||
255 (bol != NULL && igvn->_worklist.member(bol)) ||
256 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
257 // This control path may be dead.
258 // Delay this memory node transformation until the control is processed.
259 phase->is_IterGVN()->_worklist.push(this);
260 return NodeSentinel; // caller will return NULL
261 }
262 }
263 // Ignore if memory is dead, or self-loop
264 Node *mem = in(MemNode::Memory);
265 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
266 assert( mem != this, "dead loop in MemNode::Ideal" );
268 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
269 // This memory slice may be dead.
270 // Delay this mem node transformation until the memory is processed.
271 phase->is_IterGVN()->_worklist.push(this);
272 return NodeSentinel; // caller will return NULL
273 }
275 Node *address = in(MemNode::Address);
276 const Type *t_adr = phase->type( address );
277 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
279 if( can_reshape && igvn != NULL &&
280 (igvn->_worklist.member(address) ||
281 igvn->_worklist.size() > 0 && (phase->type(address) != adr_type())) ) {
282 // The address's base and type may change when the address is processed.
283 // Delay this mem node transformation until the address is processed.
284 phase->is_IterGVN()->_worklist.push(this);
285 return NodeSentinel; // caller will return NULL
286 }
288 // Do NOT remove or optimize the next lines: ensure a new alias index
289 // is allocated for an oop pointer type before Escape Analysis.
290 // Note: C++ will not remove it since the call has side effect.
291 if ( t_adr->isa_oopptr() ) {
292 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
293 }
295 #ifdef ASSERT
296 Node* base = NULL;
297 if (address->is_AddP())
298 base = address->in(AddPNode::Base);
299 assert(base == NULL || t_adr->isa_rawptr() ||
300 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?");
301 #endif
303 // Avoid independent memory operations
304 Node* old_mem = mem;
306 // The code which unhooks non-raw memories from complete (macro-expanded)
307 // initializations was removed. After macro-expansion all stores catched
308 // by Initialize node became raw stores and there is no information
309 // which memory slices they modify. So it is unsafe to move any memory
310 // operation above these stores. Also in most cases hooked non-raw memories
311 // were already unhooked by using information from detect_ptr_independence()
312 // and find_previous_store().
314 if (mem->is_MergeMem()) {
315 MergeMemNode* mmem = mem->as_MergeMem();
316 const TypePtr *tp = t_adr->is_ptr();
318 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
319 }
321 if (mem != old_mem) {
322 set_req(MemNode::Memory, mem);
323 if (can_reshape && old_mem->outcnt() == 0) {
324 igvn->_worklist.push(old_mem);
325 }
326 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
327 return this;
328 }
330 // let the subclass continue analyzing...
331 return NULL;
332 }
334 // Helper function for proving some simple control dominations.
335 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
336 // Already assumes that 'dom' is available at 'sub', and that 'sub'
337 // is not a constant (dominated by the method's StartNode).
338 // Used by MemNode::find_previous_store to prove that the
339 // control input of a memory operation predates (dominates)
340 // an allocation it wants to look past.
341 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
342 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
343 return false; // Conservative answer for dead code
345 // Check 'dom'. Skip Proj and CatchProj nodes.
346 dom = dom->find_exact_control(dom);
347 if (dom == NULL || dom->is_top())
348 return false; // Conservative answer for dead code
350 if (dom == sub) {
351 // For the case when, for example, 'sub' is Initialize and the original
352 // 'dom' is Proj node of the 'sub'.
353 return false;
354 }
356 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
357 return true;
359 // 'dom' dominates 'sub' if its control edge and control edges
360 // of all its inputs dominate or equal to sub's control edge.
362 // Currently 'sub' is either Allocate, Initialize or Start nodes.
363 // Or Region for the check in LoadNode::Ideal();
364 // 'sub' should have sub->in(0) != NULL.
365 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
366 sub->is_Region(), "expecting only these nodes");
368 // Get control edge of 'sub'.
369 Node* orig_sub = sub;
370 sub = sub->find_exact_control(sub->in(0));
371 if (sub == NULL || sub->is_top())
372 return false; // Conservative answer for dead code
374 assert(sub->is_CFG(), "expecting control");
376 if (sub == dom)
377 return true;
379 if (sub->is_Start() || sub->is_Root())
380 return false;
382 {
383 // Check all control edges of 'dom'.
385 ResourceMark rm;
386 Arena* arena = Thread::current()->resource_area();
387 Node_List nlist(arena);
388 Unique_Node_List dom_list(arena);
390 dom_list.push(dom);
391 bool only_dominating_controls = false;
393 for (uint next = 0; next < dom_list.size(); next++) {
394 Node* n = dom_list.at(next);
395 if (n == orig_sub)
396 return false; // One of dom's inputs dominated by sub.
397 if (!n->is_CFG() && n->pinned()) {
398 // Check only own control edge for pinned non-control nodes.
399 n = n->find_exact_control(n->in(0));
400 if (n == NULL || n->is_top())
401 return false; // Conservative answer for dead code
402 assert(n->is_CFG(), "expecting control");
403 dom_list.push(n);
404 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
405 only_dominating_controls = true;
406 } else if (n->is_CFG()) {
407 if (n->dominates(sub, nlist))
408 only_dominating_controls = true;
409 else
410 return false;
411 } else {
412 // First, own control edge.
413 Node* m = n->find_exact_control(n->in(0));
414 if (m != NULL) {
415 if (m->is_top())
416 return false; // Conservative answer for dead code
417 dom_list.push(m);
418 }
419 // Now, the rest of edges.
420 uint cnt = n->req();
421 for (uint i = 1; i < cnt; i++) {
422 m = n->find_exact_control(n->in(i));
423 if (m == NULL || m->is_top())
424 continue;
425 dom_list.push(m);
426 }
427 }
428 }
429 return only_dominating_controls;
430 }
431 }
433 //---------------------detect_ptr_independence---------------------------------
434 // Used by MemNode::find_previous_store to prove that two base
435 // pointers are never equal.
436 // The pointers are accompanied by their associated allocations,
437 // if any, which have been previously discovered by the caller.
438 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
439 Node* p2, AllocateNode* a2,
440 PhaseTransform* phase) {
441 // Attempt to prove that these two pointers cannot be aliased.
442 // They may both manifestly be allocations, and they should differ.
443 // Or, if they are not both allocations, they can be distinct constants.
444 // Otherwise, one is an allocation and the other a pre-existing value.
445 if (a1 == NULL && a2 == NULL) { // neither an allocation
446 return (p1 != p2) && p1->is_Con() && p2->is_Con();
447 } else if (a1 != NULL && a2 != NULL) { // both allocations
448 return (a1 != a2);
449 } else if (a1 != NULL) { // one allocation a1
450 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
451 return all_controls_dominate(p2, a1);
452 } else { //(a2 != NULL) // one allocation a2
453 return all_controls_dominate(p1, a2);
454 }
455 return false;
456 }
459 // The logic for reordering loads and stores uses four steps:
460 // (a) Walk carefully past stores and initializations which we
461 // can prove are independent of this load.
462 // (b) Observe that the next memory state makes an exact match
463 // with self (load or store), and locate the relevant store.
464 // (c) Ensure that, if we were to wire self directly to the store,
465 // the optimizer would fold it up somehow.
466 // (d) Do the rewiring, and return, depending on some other part of
467 // the optimizer to fold up the load.
468 // This routine handles steps (a) and (b). Steps (c) and (d) are
469 // specific to loads and stores, so they are handled by the callers.
470 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
471 //
472 Node* MemNode::find_previous_store(PhaseTransform* phase) {
473 Node* ctrl = in(MemNode::Control);
474 Node* adr = in(MemNode::Address);
475 intptr_t offset = 0;
476 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
477 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
479 if (offset == Type::OffsetBot)
480 return NULL; // cannot unalias unless there are precise offsets
482 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
484 intptr_t size_in_bytes = memory_size();
486 Node* mem = in(MemNode::Memory); // start searching here...
488 int cnt = 50; // Cycle limiter
489 for (;;) { // While we can dance past unrelated stores...
490 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
492 if (mem->is_Store()) {
493 Node* st_adr = mem->in(MemNode::Address);
494 intptr_t st_offset = 0;
495 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
496 if (st_base == NULL)
497 break; // inscrutable pointer
498 if (st_offset != offset && st_offset != Type::OffsetBot) {
499 const int MAX_STORE = BytesPerLong;
500 if (st_offset >= offset + size_in_bytes ||
501 st_offset <= offset - MAX_STORE ||
502 st_offset <= offset - mem->as_Store()->memory_size()) {
503 // Success: The offsets are provably independent.
504 // (You may ask, why not just test st_offset != offset and be done?
505 // The answer is that stores of different sizes can co-exist
506 // in the same sequence of RawMem effects. We sometimes initialize
507 // a whole 'tile' of array elements with a single jint or jlong.)
508 mem = mem->in(MemNode::Memory);
509 continue; // (a) advance through independent store memory
510 }
511 }
512 if (st_base != base &&
513 detect_ptr_independence(base, alloc,
514 st_base,
515 AllocateNode::Ideal_allocation(st_base, phase),
516 phase)) {
517 // Success: The bases are provably independent.
518 mem = mem->in(MemNode::Memory);
519 continue; // (a) advance through independent store memory
520 }
522 // (b) At this point, if the bases or offsets do not agree, we lose,
523 // since we have not managed to prove 'this' and 'mem' independent.
524 if (st_base == base && st_offset == offset) {
525 return mem; // let caller handle steps (c), (d)
526 }
528 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
529 InitializeNode* st_init = mem->in(0)->as_Initialize();
530 AllocateNode* st_alloc = st_init->allocation();
531 if (st_alloc == NULL)
532 break; // something degenerated
533 bool known_identical = false;
534 bool known_independent = false;
535 if (alloc == st_alloc)
536 known_identical = true;
537 else if (alloc != NULL)
538 known_independent = true;
539 else if (all_controls_dominate(this, st_alloc))
540 known_independent = true;
542 if (known_independent) {
543 // The bases are provably independent: Either they are
544 // manifestly distinct allocations, or else the control
545 // of this load dominates the store's allocation.
546 int alias_idx = phase->C->get_alias_index(adr_type());
547 if (alias_idx == Compile::AliasIdxRaw) {
548 mem = st_alloc->in(TypeFunc::Memory);
549 } else {
550 mem = st_init->memory(alias_idx);
551 }
552 continue; // (a) advance through independent store memory
553 }
555 // (b) at this point, if we are not looking at a store initializing
556 // the same allocation we are loading from, we lose.
557 if (known_identical) {
558 // From caller, can_see_stored_value will consult find_captured_store.
559 return mem; // let caller handle steps (c), (d)
560 }
562 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
563 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
564 if (mem->is_Proj() && mem->in(0)->is_Call()) {
565 CallNode *call = mem->in(0)->as_Call();
566 if (!call->may_modify(addr_t, phase)) {
567 mem = call->in(TypeFunc::Memory);
568 continue; // (a) advance through independent call memory
569 }
570 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
571 mem = mem->in(0)->in(TypeFunc::Memory);
572 continue; // (a) advance through independent MemBar memory
573 } else if (mem->is_ClearArray()) {
574 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
575 // (the call updated 'mem' value)
576 continue; // (a) advance through independent allocation memory
577 } else {
578 // Can not bypass initialization of the instance
579 // we are looking for.
580 return mem;
581 }
582 } else if (mem->is_MergeMem()) {
583 int alias_idx = phase->C->get_alias_index(adr_type());
584 mem = mem->as_MergeMem()->memory_at(alias_idx);
585 continue; // (a) advance through independent MergeMem memory
586 }
587 }
589 // Unless there is an explicit 'continue', we must bail out here,
590 // because 'mem' is an inscrutable memory state (e.g., a call).
591 break;
592 }
594 return NULL; // bail out
595 }
597 //----------------------calculate_adr_type-------------------------------------
598 // Helper function. Notices when the given type of address hits top or bottom.
599 // Also, asserts a cross-check of the type against the expected address type.
600 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
601 if (t == Type::TOP) return NULL; // does not touch memory any more?
602 #ifdef PRODUCT
603 cross_check = NULL;
604 #else
605 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
606 #endif
607 const TypePtr* tp = t->isa_ptr();
608 if (tp == NULL) {
609 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
610 return TypePtr::BOTTOM; // touches lots of memory
611 } else {
612 #ifdef ASSERT
613 // %%%% [phh] We don't check the alias index if cross_check is
614 // TypeRawPtr::BOTTOM. Needs to be investigated.
615 if (cross_check != NULL &&
616 cross_check != TypePtr::BOTTOM &&
617 cross_check != TypeRawPtr::BOTTOM) {
618 // Recheck the alias index, to see if it has changed (due to a bug).
619 Compile* C = Compile::current();
620 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
621 "must stay in the original alias category");
622 // The type of the address must be contained in the adr_type,
623 // disregarding "null"-ness.
624 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
625 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
626 assert(cross_check->meet(tp_notnull) == cross_check,
627 "real address must not escape from expected memory type");
628 }
629 #endif
630 return tp;
631 }
632 }
634 //------------------------adr_phi_is_loop_invariant----------------------------
635 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
636 // loop is loop invariant. Make a quick traversal of Phi and associated
637 // CastPP nodes, looking to see if they are a closed group within the loop.
638 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
639 // The idea is that the phi-nest must boil down to only CastPP nodes
640 // with the same data. This implies that any path into the loop already
641 // includes such a CastPP, and so the original cast, whatever its input,
642 // must be covered by an equivalent cast, with an earlier control input.
643 ResourceMark rm;
645 // The loop entry input of the phi should be the unique dominating
646 // node for every Phi/CastPP in the loop.
647 Unique_Node_List closure;
648 closure.push(adr_phi->in(LoopNode::EntryControl));
650 // Add the phi node and the cast to the worklist.
651 Unique_Node_List worklist;
652 worklist.push(adr_phi);
653 if( cast != NULL ){
654 if( !cast->is_ConstraintCast() ) return false;
655 worklist.push(cast);
656 }
658 // Begin recursive walk of phi nodes.
659 while( worklist.size() ){
660 // Take a node off the worklist
661 Node *n = worklist.pop();
662 if( !closure.member(n) ){
663 // Add it to the closure.
664 closure.push(n);
665 // Make a sanity check to ensure we don't waste too much time here.
666 if( closure.size() > 20) return false;
667 // This node is OK if:
668 // - it is a cast of an identical value
669 // - or it is a phi node (then we add its inputs to the worklist)
670 // Otherwise, the node is not OK, and we presume the cast is not invariant
671 if( n->is_ConstraintCast() ){
672 worklist.push(n->in(1));
673 } else if( n->is_Phi() ) {
674 for( uint i = 1; i < n->req(); i++ ) {
675 worklist.push(n->in(i));
676 }
677 } else {
678 return false;
679 }
680 }
681 }
683 // Quit when the worklist is empty, and we've found no offending nodes.
684 return true;
685 }
687 //------------------------------Ideal_DU_postCCP-------------------------------
688 // Find any cast-away of null-ness and keep its control. Null cast-aways are
689 // going away in this pass and we need to make this memory op depend on the
690 // gating null check.
691 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
692 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
693 }
695 // I tried to leave the CastPP's in. This makes the graph more accurate in
696 // some sense; we get to keep around the knowledge that an oop is not-null
697 // after some test. Alas, the CastPP's interfere with GVN (some values are
698 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
699 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
700 // some of the more trivial cases in the optimizer. Removing more useless
701 // Phi's started allowing Loads to illegally float above null checks. I gave
702 // up on this approach. CNC 10/20/2000
703 // This static method may be called not from MemNode (EncodePNode calls it).
704 // Only the control edge of the node 'n' might be updated.
705 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
706 Node *skipped_cast = NULL;
707 // Need a null check? Regular static accesses do not because they are
708 // from constant addresses. Array ops are gated by the range check (which
709 // always includes a NULL check). Just check field ops.
710 if( n->in(MemNode::Control) == NULL ) {
711 // Scan upwards for the highest location we can place this memory op.
712 while( true ) {
713 switch( adr->Opcode() ) {
715 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
716 adr = adr->in(AddPNode::Base);
717 continue;
719 case Op_DecodeN: // No change to NULL-ness, so peek thru
720 case Op_DecodeNKlass:
721 adr = adr->in(1);
722 continue;
724 case Op_EncodeP:
725 case Op_EncodePKlass:
726 // EncodeP node's control edge could be set by this method
727 // when EncodeP node depends on CastPP node.
728 //
729 // Use its control edge for memory op because EncodeP may go away
730 // later when it is folded with following or preceding DecodeN node.
731 if (adr->in(0) == NULL) {
732 // Keep looking for cast nodes.
733 adr = adr->in(1);
734 continue;
735 }
736 ccp->hash_delete(n);
737 n->set_req(MemNode::Control, adr->in(0));
738 ccp->hash_insert(n);
739 return n;
741 case Op_CastPP:
742 // If the CastPP is useless, just peek on through it.
743 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
744 // Remember the cast that we've peeked though. If we peek
745 // through more than one, then we end up remembering the highest
746 // one, that is, if in a loop, the one closest to the top.
747 skipped_cast = adr;
748 adr = adr->in(1);
749 continue;
750 }
751 // CastPP is going away in this pass! We need this memory op to be
752 // control-dependent on the test that is guarding the CastPP.
753 ccp->hash_delete(n);
754 n->set_req(MemNode::Control, adr->in(0));
755 ccp->hash_insert(n);
756 return n;
758 case Op_Phi:
759 // Attempt to float above a Phi to some dominating point.
760 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
761 // If we've already peeked through a Cast (which could have set the
762 // control), we can't float above a Phi, because the skipped Cast
763 // may not be loop invariant.
764 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
765 adr = adr->in(1);
766 continue;
767 }
768 }
770 // Intentional fallthrough!
772 // No obvious dominating point. The mem op is pinned below the Phi
773 // by the Phi itself. If the Phi goes away (no true value is merged)
774 // then the mem op can float, but not indefinitely. It must be pinned
775 // behind the controls leading to the Phi.
776 case Op_CheckCastPP:
777 // These usually stick around to change address type, however a
778 // useless one can be elided and we still need to pick up a control edge
779 if (adr->in(0) == NULL) {
780 // This CheckCastPP node has NO control and is likely useless. But we
781 // need check further up the ancestor chain for a control input to keep
782 // the node in place. 4959717.
783 skipped_cast = adr;
784 adr = adr->in(1);
785 continue;
786 }
787 ccp->hash_delete(n);
788 n->set_req(MemNode::Control, adr->in(0));
789 ccp->hash_insert(n);
790 return n;
792 // List of "safe" opcodes; those that implicitly block the memory
793 // op below any null check.
794 case Op_CastX2P: // no null checks on native pointers
795 case Op_Parm: // 'this' pointer is not null
796 case Op_LoadP: // Loading from within a klass
797 case Op_LoadN: // Loading from within a klass
798 case Op_LoadKlass: // Loading from within a klass
799 case Op_LoadNKlass: // Loading from within a klass
800 case Op_ConP: // Loading from a klass
801 case Op_ConN: // Loading from a klass
802 case Op_ConNKlass: // Loading from a klass
803 case Op_CreateEx: // Sucking up the guts of an exception oop
804 case Op_Con: // Reading from TLS
805 case Op_CMoveP: // CMoveP is pinned
806 case Op_CMoveN: // CMoveN is pinned
807 break; // No progress
809 case Op_Proj: // Direct call to an allocation routine
810 case Op_SCMemProj: // Memory state from store conditional ops
811 #ifdef ASSERT
812 {
813 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
814 const Node* call = adr->in(0);
815 if (call->is_CallJava()) {
816 const CallJavaNode* call_java = call->as_CallJava();
817 const TypeTuple *r = call_java->tf()->range();
818 assert(r->cnt() > TypeFunc::Parms, "must return value");
819 const Type* ret_type = r->field_at(TypeFunc::Parms);
820 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
821 // We further presume that this is one of
822 // new_instance_Java, new_array_Java, or
823 // the like, but do not assert for this.
824 } else if (call->is_Allocate()) {
825 // similar case to new_instance_Java, etc.
826 } else if (!call->is_CallLeaf()) {
827 // Projections from fetch_oop (OSR) are allowed as well.
828 ShouldNotReachHere();
829 }
830 }
831 #endif
832 break;
833 default:
834 ShouldNotReachHere();
835 }
836 break;
837 }
838 }
840 return NULL; // No progress
841 }
844 //=============================================================================
845 uint LoadNode::size_of() const { return sizeof(*this); }
846 uint LoadNode::cmp( const Node &n ) const
847 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
848 const Type *LoadNode::bottom_type() const { return _type; }
849 uint LoadNode::ideal_reg() const {
850 return _type->ideal_reg();
851 }
853 #ifndef PRODUCT
854 void LoadNode::dump_spec(outputStream *st) const {
855 MemNode::dump_spec(st);
856 if( !Verbose && !WizardMode ) {
857 // standard dump does this in Verbose and WizardMode
858 st->print(" #"); _type->dump_on(st);
859 }
860 }
861 #endif
863 #ifdef ASSERT
864 //----------------------------is_immutable_value-------------------------------
865 // Helper function to allow a raw load without control edge for some cases
866 bool LoadNode::is_immutable_value(Node* adr) {
867 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
868 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
869 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
870 in_bytes(JavaThread::osthread_offset())));
871 }
872 #endif
874 //----------------------------LoadNode::make-----------------------------------
875 // Polymorphic factory method:
876 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
877 Compile* C = gvn.C;
879 // sanity check the alias category against the created node type
880 assert(!(adr_type->isa_oopptr() &&
881 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
882 "use LoadKlassNode instead");
883 assert(!(adr_type->isa_aryptr() &&
884 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
885 "use LoadRangeNode instead");
886 // Check control edge of raw loads
887 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
888 // oop will be recorded in oop map if load crosses safepoint
889 rt->isa_oopptr() || is_immutable_value(adr),
890 "raw memory operations should have control edge");
891 switch (bt) {
892 case T_BOOLEAN: return new (C) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() );
893 case T_BYTE: return new (C) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() );
894 case T_INT: return new (C) LoadINode (ctl, mem, adr, adr_type, rt->is_int() );
895 case T_CHAR: return new (C) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() );
896 case T_SHORT: return new (C) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() );
897 case T_LONG: return new (C) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() );
898 case T_FLOAT: return new (C) LoadFNode (ctl, mem, adr, adr_type, rt );
899 case T_DOUBLE: return new (C) LoadDNode (ctl, mem, adr, adr_type, rt );
900 case T_ADDRESS: return new (C) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() );
901 case T_OBJECT:
902 #ifdef _LP64
903 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
904 Node* load = gvn.transform(new (C) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));
905 return new (C) DecodeNNode(load, load->bottom_type()->make_ptr());
906 } else
907 #endif
908 {
909 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
910 return new (C) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
911 }
912 }
913 ShouldNotReachHere();
914 return (LoadNode*)NULL;
915 }
917 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
918 bool require_atomic = true;
919 return new (C) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
920 }
925 //------------------------------hash-------------------------------------------
926 uint LoadNode::hash() const {
927 // unroll addition of interesting fields
928 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
929 }
931 //---------------------------can_see_stored_value------------------------------
932 // This routine exists to make sure this set of tests is done the same
933 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
934 // will change the graph shape in a way which makes memory alive twice at the
935 // same time (uses the Oracle model of aliasing), then some
936 // LoadXNode::Identity will fold things back to the equivalence-class model
937 // of aliasing.
938 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
939 Node* ld_adr = in(MemNode::Address);
941 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
942 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
943 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
944 atp->field() != NULL && !atp->field()->is_volatile()) {
945 uint alias_idx = atp->index();
946 bool final = atp->field()->is_final();
947 Node* result = NULL;
948 Node* current = st;
949 // Skip through chains of MemBarNodes checking the MergeMems for
950 // new states for the slice of this load. Stop once any other
951 // kind of node is encountered. Loads from final memory can skip
952 // through any kind of MemBar but normal loads shouldn't skip
953 // through MemBarAcquire since the could allow them to move out of
954 // a synchronized region.
955 while (current->is_Proj()) {
956 int opc = current->in(0)->Opcode();
957 if ((final && (opc == Op_MemBarAcquire || opc == Op_MemBarAcquireLock)) ||
958 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder ||
959 opc == Op_MemBarReleaseLock) {
960 Node* mem = current->in(0)->in(TypeFunc::Memory);
961 if (mem->is_MergeMem()) {
962 MergeMemNode* merge = mem->as_MergeMem();
963 Node* new_st = merge->memory_at(alias_idx);
964 if (new_st == merge->base_memory()) {
965 // Keep searching
966 current = merge->base_memory();
967 continue;
968 }
969 // Save the new memory state for the slice and fall through
970 // to exit.
971 result = new_st;
972 }
973 }
974 break;
975 }
976 if (result != NULL) {
977 st = result;
978 }
979 }
982 // Loop around twice in the case Load -> Initialize -> Store.
983 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
984 for (int trip = 0; trip <= 1; trip++) {
986 if (st->is_Store()) {
987 Node* st_adr = st->in(MemNode::Address);
988 if (!phase->eqv(st_adr, ld_adr)) {
989 // Try harder before giving up... Match raw and non-raw pointers.
990 intptr_t st_off = 0;
991 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
992 if (alloc == NULL) return NULL;
993 intptr_t ld_off = 0;
994 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
995 if (alloc != allo2) return NULL;
996 if (ld_off != st_off) return NULL;
997 // At this point we have proven something like this setup:
998 // A = Allocate(...)
999 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
1000 // S = StoreQ(, AddP(, A.Parm , #Off), V)
1001 // (Actually, we haven't yet proven the Q's are the same.)
1002 // In other words, we are loading from a casted version of
1003 // the same pointer-and-offset that we stored to.
1004 // Thus, we are able to replace L by V.
1005 }
1006 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1007 if (store_Opcode() != st->Opcode())
1008 return NULL;
1009 return st->in(MemNode::ValueIn);
1010 }
1012 intptr_t offset = 0; // scratch
1014 // A load from a freshly-created object always returns zero.
1015 // (This can happen after LoadNode::Ideal resets the load's memory input
1016 // to find_captured_store, which returned InitializeNode::zero_memory.)
1017 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1018 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
1019 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
1020 // return a zero value for the load's basic type
1021 // (This is one of the few places where a generic PhaseTransform
1022 // can create new nodes. Think of it as lazily manifesting
1023 // virtually pre-existing constants.)
1024 return phase->zerocon(memory_type());
1025 }
1027 // A load from an initialization barrier can match a captured store.
1028 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1029 InitializeNode* init = st->in(0)->as_Initialize();
1030 AllocateNode* alloc = init->allocation();
1031 if (alloc != NULL &&
1032 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
1033 // examine a captured store value
1034 st = init->find_captured_store(offset, memory_size(), phase);
1035 if (st != NULL)
1036 continue; // take one more trip around
1037 }
1038 }
1040 break;
1041 }
1043 return NULL;
1044 }
1046 //----------------------is_instance_field_load_with_local_phi------------------
1047 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1048 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
1049 in(MemNode::Address)->is_AddP() ) {
1050 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
1051 // Only instances.
1052 if( t_oop != NULL && t_oop->is_known_instance_field() &&
1053 t_oop->offset() != Type::OffsetBot &&
1054 t_oop->offset() != Type::OffsetTop) {
1055 return true;
1056 }
1057 }
1058 return false;
1059 }
1061 //------------------------------Identity---------------------------------------
1062 // Loads are identity if previous store is to same address
1063 Node *LoadNode::Identity( PhaseTransform *phase ) {
1064 // If the previous store-maker is the right kind of Store, and the store is
1065 // to the same address, then we are equal to the value stored.
1066 Node* mem = in(MemNode::Memory);
1067 Node* value = can_see_stored_value(mem, phase);
1068 if( value ) {
1069 // byte, short & char stores truncate naturally.
1070 // A load has to load the truncated value which requires
1071 // some sort of masking operation and that requires an
1072 // Ideal call instead of an Identity call.
1073 if (memory_size() < BytesPerInt) {
1074 // If the input to the store does not fit with the load's result type,
1075 // it must be truncated via an Ideal call.
1076 if (!phase->type(value)->higher_equal(phase->type(this)))
1077 return this;
1078 }
1079 // (This works even when value is a Con, but LoadNode::Value
1080 // usually runs first, producing the singleton type of the Con.)
1081 return value;
1082 }
1084 // Search for an existing data phi which was generated before for the same
1085 // instance's field to avoid infinite generation of phis in a loop.
1086 Node *region = mem->in(0);
1087 if (is_instance_field_load_with_local_phi(region)) {
1088 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
1089 int this_index = phase->C->get_alias_index(addr_t);
1090 int this_offset = addr_t->offset();
1091 int this_id = addr_t->is_oopptr()->instance_id();
1092 const Type* this_type = bottom_type();
1093 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1094 Node* phi = region->fast_out(i);
1095 if (phi->is_Phi() && phi != mem &&
1096 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
1097 return phi;
1098 }
1099 }
1100 }
1102 return this;
1103 }
1106 // Returns true if the AliasType refers to the field that holds the
1107 // cached box array. Currently only handles the IntegerCache case.
1108 static bool is_autobox_cache(Compile::AliasType* atp) {
1109 if (atp != NULL && atp->field() != NULL) {
1110 ciField* field = atp->field();
1111 ciSymbol* klass = field->holder()->name();
1112 if (field->name() == ciSymbol::cache_field_name() &&
1113 field->holder()->uses_default_loader() &&
1114 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
1115 return true;
1116 }
1117 }
1118 return false;
1119 }
1121 // Fetch the base value in the autobox array
1122 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
1123 if (atp != NULL && atp->field() != NULL) {
1124 ciField* field = atp->field();
1125 ciSymbol* klass = field->holder()->name();
1126 if (field->name() == ciSymbol::cache_field_name() &&
1127 field->holder()->uses_default_loader() &&
1128 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
1129 assert(field->is_constant(), "what?");
1130 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
1131 // Fetch the box object at the base of the array and get its value
1132 ciInstance* box = array->obj_at(0)->as_instance();
1133 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1134 if (ik->nof_nonstatic_fields() == 1) {
1135 // This should be true nonstatic_field_at requires calling
1136 // nof_nonstatic_fields so check it anyway
1137 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1138 cache_offset = c.as_int();
1139 }
1140 return true;
1141 }
1142 }
1143 return false;
1144 }
1146 // Returns true if the AliasType refers to the value field of an
1147 // autobox object. Currently only handles Integer.
1148 static bool is_autobox_object(Compile::AliasType* atp) {
1149 if (atp != NULL && atp->field() != NULL) {
1150 ciField* field = atp->field();
1151 ciSymbol* klass = field->holder()->name();
1152 if (field->name() == ciSymbol::value_name() &&
1153 field->holder()->uses_default_loader() &&
1154 klass == ciSymbol::java_lang_Integer()) {
1155 return true;
1156 }
1157 }
1158 return false;
1159 }
1162 // We're loading from an object which has autobox behaviour.
1163 // If this object is result of a valueOf call we'll have a phi
1164 // merging a newly allocated object and a load from the cache.
1165 // We want to replace this load with the original incoming
1166 // argument to the valueOf call.
1167 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1168 Node* base = in(Address)->in(AddPNode::Base);
1169 if (base->is_Phi() && base->req() == 3) {
1170 AllocateNode* allocation = NULL;
1171 int allocation_index = -1;
1172 int load_index = -1;
1173 for (uint i = 1; i < base->req(); i++) {
1174 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
1175 if (allocation != NULL) {
1176 allocation_index = i;
1177 load_index = 3 - allocation_index;
1178 break;
1179 }
1180 }
1181 bool has_load = ( allocation != NULL &&
1182 (base->in(load_index)->is_Load() ||
1183 base->in(load_index)->is_DecodeN() &&
1184 base->in(load_index)->in(1)->is_Load()) );
1185 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
1186 // Push the loads from the phi that comes from valueOf up
1187 // through it to allow elimination of the loads and the recovery
1188 // of the original value.
1189 Node* mem_phi = in(Memory);
1190 Node* offset = in(Address)->in(AddPNode::Offset);
1191 Node* region = base->in(0);
1193 Node* in1 = clone();
1194 Node* in1_addr = in1->in(Address)->clone();
1195 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
1196 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
1197 in1_addr->set_req(AddPNode::Offset, offset);
1198 in1->set_req(0, region->in(allocation_index));
1199 in1->set_req(Address, in1_addr);
1200 in1->set_req(Memory, mem_phi->in(allocation_index));
1202 Node* in2 = clone();
1203 Node* in2_addr = in2->in(Address)->clone();
1204 in2_addr->set_req(AddPNode::Base, base->in(load_index));
1205 in2_addr->set_req(AddPNode::Address, base->in(load_index));
1206 in2_addr->set_req(AddPNode::Offset, offset);
1207 in2->set_req(0, region->in(load_index));
1208 in2->set_req(Address, in2_addr);
1209 in2->set_req(Memory, mem_phi->in(load_index));
1211 in1_addr = phase->transform(in1_addr);
1212 in1 = phase->transform(in1);
1213 in2_addr = phase->transform(in2_addr);
1214 in2 = phase->transform(in2);
1216 PhiNode* result = PhiNode::make_blank(region, this);
1217 result->set_req(allocation_index, in1);
1218 result->set_req(load_index, in2);
1219 return result;
1220 }
1221 } else if (base->is_Load() ||
1222 base->is_DecodeN() && base->in(1)->is_Load()) {
1223 if (base->is_DecodeN()) {
1224 // Get LoadN node which loads cached Integer object
1225 base = base->in(1);
1226 }
1227 // Eliminate the load of Integer.value for integers from the cache
1228 // array by deriving the value from the index into the array.
1229 // Capture the offset of the load and then reverse the computation.
1230 Node* load_base = base->in(Address)->in(AddPNode::Base);
1231 if (load_base->is_DecodeN()) {
1232 // Get LoadN node which loads IntegerCache.cache field
1233 load_base = load_base->in(1);
1234 }
1235 if (load_base != NULL) {
1236 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
1237 intptr_t cache_offset;
1238 int shift = -1;
1239 Node* cache = NULL;
1240 if (is_autobox_cache(atp)) {
1241 shift = exact_log2(type2aelembytes(T_OBJECT));
1242 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
1243 }
1244 if (cache != NULL && base->in(Address)->is_AddP()) {
1245 Node* elements[4];
1246 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
1247 int cache_low;
1248 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
1249 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
1250 // Add up all the offsets making of the address of the load
1251 Node* result = elements[0];
1252 for (int i = 1; i < count; i++) {
1253 result = phase->transform(new (phase->C) AddXNode(result, elements[i]));
1254 }
1255 // Remove the constant offset from the address and then
1256 // remove the scaling of the offset to recover the original index.
1257 result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-offset)));
1258 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1259 // Peel the shift off directly but wrap it in a dummy node
1260 // since Ideal can't return existing nodes
1261 result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0));
1262 } else {
1263 result = new (phase->C) RShiftXNode(result, phase->intcon(shift));
1264 }
1265 #ifdef _LP64
1266 result = new (phase->C) ConvL2INode(phase->transform(result));
1267 #endif
1268 return result;
1269 }
1270 }
1271 }
1272 }
1273 return NULL;
1274 }
1276 //------------------------------split_through_phi------------------------------
1277 // Split instance field load through Phi.
1278 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1279 Node* mem = in(MemNode::Memory);
1280 Node* address = in(MemNode::Address);
1281 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1282 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1284 assert(mem->is_Phi() && (t_oop != NULL) &&
1285 t_oop->is_known_instance_field(), "invalide conditions");
1287 Node *region = mem->in(0);
1288 if (region == NULL) {
1289 return NULL; // Wait stable graph
1290 }
1291 uint cnt = mem->req();
1292 for (uint i = 1; i < cnt; i++) {
1293 Node* rc = region->in(i);
1294 if (rc == NULL || phase->type(rc) == Type::TOP)
1295 return NULL; // Wait stable graph
1296 Node *in = mem->in(i);
1297 if (in == NULL) {
1298 return NULL; // Wait stable graph
1299 }
1300 }
1301 // Check for loop invariant.
1302 if (cnt == 3) {
1303 for (uint i = 1; i < cnt; i++) {
1304 Node *in = mem->in(i);
1305 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
1306 if (m == mem) {
1307 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
1308 return this;
1309 }
1310 }
1311 }
1312 // Split through Phi (see original code in loopopts.cpp).
1313 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
1315 // Do nothing here if Identity will find a value
1316 // (to avoid infinite chain of value phis generation).
1317 if (!phase->eqv(this, this->Identity(phase)))
1318 return NULL;
1320 // Skip the split if the region dominates some control edge of the address.
1321 if (!MemNode::all_controls_dominate(address, region))
1322 return NULL;
1324 const Type* this_type = this->bottom_type();
1325 int this_index = phase->C->get_alias_index(addr_t);
1326 int this_offset = addr_t->offset();
1327 int this_iid = addr_t->is_oopptr()->instance_id();
1328 PhaseIterGVN *igvn = phase->is_IterGVN();
1329 Node *phi = new (igvn->C) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1330 for (uint i = 1; i < region->req(); i++) {
1331 Node *x;
1332 Node* the_clone = NULL;
1333 if (region->in(i) == phase->C->top()) {
1334 x = phase->C->top(); // Dead path? Use a dead data op
1335 } else {
1336 x = this->clone(); // Else clone up the data op
1337 the_clone = x; // Remember for possible deletion.
1338 // Alter data node to use pre-phi inputs
1339 if (this->in(0) == region) {
1340 x->set_req(0, region->in(i));
1341 } else {
1342 x->set_req(0, NULL);
1343 }
1344 for (uint j = 1; j < this->req(); j++) {
1345 Node *in = this->in(j);
1346 if (in->is_Phi() && in->in(0) == region)
1347 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
1348 }
1349 }
1350 // Check for a 'win' on some paths
1351 const Type *t = x->Value(igvn);
1353 bool singleton = t->singleton();
1355 // See comments in PhaseIdealLoop::split_thru_phi().
1356 if (singleton && t == Type::TOP) {
1357 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1358 }
1360 if (singleton) {
1361 x = igvn->makecon(t);
1362 } else {
1363 // We now call Identity to try to simplify the cloned node.
1364 // Note that some Identity methods call phase->type(this).
1365 // Make sure that the type array is big enough for
1366 // our new node, even though we may throw the node away.
1367 // (This tweaking with igvn only works because x is a new node.)
1368 igvn->set_type(x, t);
1369 // If x is a TypeNode, capture any more-precise type permanently into Node
1370 // otherwise it will be not updated during igvn->transform since
1371 // igvn->type(x) is set to x->Value() already.
1372 x->raise_bottom_type(t);
1373 Node *y = x->Identity(igvn);
1374 if (y != x) {
1375 x = y;
1376 } else {
1377 y = igvn->hash_find(x);
1378 if (y) {
1379 x = y;
1380 } else {
1381 // Else x is a new node we are keeping
1382 // We do not need register_new_node_with_optimizer
1383 // because set_type has already been called.
1384 igvn->_worklist.push(x);
1385 }
1386 }
1387 }
1388 if (x != the_clone && the_clone != NULL)
1389 igvn->remove_dead_node(the_clone);
1390 phi->set_req(i, x);
1391 }
1392 // Record Phi
1393 igvn->register_new_node_with_optimizer(phi);
1394 return phi;
1395 }
1397 //------------------------------Ideal------------------------------------------
1398 // If the load is from Field memory and the pointer is non-null, we can
1399 // zero out the control input.
1400 // If the offset is constant and the base is an object allocation,
1401 // try to hook me up to the exact initializing store.
1402 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1403 Node* p = MemNode::Ideal_common(phase, can_reshape);
1404 if (p) return (p == NodeSentinel) ? NULL : p;
1406 Node* ctrl = in(MemNode::Control);
1407 Node* address = in(MemNode::Address);
1409 // Skip up past a SafePoint control. Cannot do this for Stores because
1410 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1411 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1412 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1413 ctrl = ctrl->in(0);
1414 set_req(MemNode::Control,ctrl);
1415 }
1417 intptr_t ignore = 0;
1418 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1419 if (base != NULL
1420 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1421 // Check for useless control edge in some common special cases
1422 if (in(MemNode::Control) != NULL
1423 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1424 && all_controls_dominate(base, phase->C->start())) {
1425 // A method-invariant, non-null address (constant or 'this' argument).
1426 set_req(MemNode::Control, NULL);
1427 }
1429 if (EliminateAutoBox && can_reshape) {
1430 assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null");
1431 Compile::AliasType* atp = phase->C->alias_type(adr_type());
1432 if (is_autobox_object(atp)) {
1433 Node* result = eliminate_autobox(phase);
1434 if (result != NULL) return result;
1435 }
1436 }
1437 }
1439 Node* mem = in(MemNode::Memory);
1440 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1442 if (addr_t != NULL) {
1443 // try to optimize our memory input
1444 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
1445 if (opt_mem != mem) {
1446 set_req(MemNode::Memory, opt_mem);
1447 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1448 return this;
1449 }
1450 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1451 if (can_reshape && opt_mem->is_Phi() &&
1452 (t_oop != NULL) && t_oop->is_known_instance_field()) {
1453 PhaseIterGVN *igvn = phase->is_IterGVN();
1454 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1455 // Delay this transformation until memory Phi is processed.
1456 phase->is_IterGVN()->_worklist.push(this);
1457 return NULL;
1458 }
1459 // Split instance field load through Phi.
1460 Node* result = split_through_phi(phase);
1461 if (result != NULL) return result;
1462 }
1463 }
1465 // Check for prior store with a different base or offset; make Load
1466 // independent. Skip through any number of them. Bail out if the stores
1467 // are in an endless dead cycle and report no progress. This is a key
1468 // transform for Reflection. However, if after skipping through the Stores
1469 // we can't then fold up against a prior store do NOT do the transform as
1470 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1471 // array memory alive twice: once for the hoisted Load and again after the
1472 // bypassed Store. This situation only works if EVERYBODY who does
1473 // anti-dependence work knows how to bypass. I.e. we need all
1474 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1475 // the alias index stuff. So instead, peek through Stores and IFF we can
1476 // fold up, do so.
1477 Node* prev_mem = find_previous_store(phase);
1478 // Steps (a), (b): Walk past independent stores to find an exact match.
1479 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1480 // (c) See if we can fold up on the spot, but don't fold up here.
1481 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1482 // just return a prior value, which is done by Identity calls.
1483 if (can_see_stored_value(prev_mem, phase)) {
1484 // Make ready for step (d):
1485 set_req(MemNode::Memory, prev_mem);
1486 return this;
1487 }
1488 }
1490 return NULL; // No further progress
1491 }
1493 // Helper to recognize certain Klass fields which are invariant across
1494 // some group of array types (e.g., int[] or all T[] where T < Object).
1495 const Type*
1496 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1497 ciKlass* klass) const {
1498 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1499 // The field is Klass::_modifier_flags. Return its (constant) value.
1500 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1501 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1502 return TypeInt::make(klass->modifier_flags());
1503 }
1504 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1505 // The field is Klass::_access_flags. Return its (constant) value.
1506 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1507 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1508 return TypeInt::make(klass->access_flags());
1509 }
1510 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1511 // The field is Klass::_layout_helper. Return its constant value if known.
1512 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1513 return TypeInt::make(klass->layout_helper());
1514 }
1516 // No match.
1517 return NULL;
1518 }
1520 //------------------------------Value-----------------------------------------
1521 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1522 // Either input is TOP ==> the result is TOP
1523 Node* mem = in(MemNode::Memory);
1524 const Type *t1 = phase->type(mem);
1525 if (t1 == Type::TOP) return Type::TOP;
1526 Node* adr = in(MemNode::Address);
1527 const TypePtr* tp = phase->type(adr)->isa_ptr();
1528 if (tp == NULL || tp->empty()) return Type::TOP;
1529 int off = tp->offset();
1530 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1531 Compile* C = phase->C;
1533 // Try to guess loaded type from pointer type
1534 if (tp->base() == Type::AryPtr) {
1535 const Type *t = tp->is_aryptr()->elem();
1536 // Don't do this for integer types. There is only potential profit if
1537 // the element type t is lower than _type; that is, for int types, if _type is
1538 // more restrictive than t. This only happens here if one is short and the other
1539 // char (both 16 bits), and in those cases we've made an intentional decision
1540 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1541 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1542 //
1543 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1544 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1545 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1546 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1547 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1548 // In fact, that could have been the original type of p1, and p1 could have
1549 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1550 // expression (LShiftL quux 3) independently optimized to the constant 8.
1551 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1552 && (_type->isa_vect() == NULL)
1553 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1554 // t might actually be lower than _type, if _type is a unique
1555 // concrete subclass of abstract class t.
1556 // Make sure the reference is not into the header, by comparing
1557 // the offset against the offset of the start of the array's data.
1558 // Different array types begin at slightly different offsets (12 vs. 16).
1559 // We choose T_BYTE as an example base type that is least restrictive
1560 // as to alignment, which will therefore produce the smallest
1561 // possible base offset.
1562 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1563 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
1564 const Type* jt = t->join(_type);
1565 // In any case, do not allow the join, per se, to empty out the type.
1566 if (jt->empty() && !t->empty()) {
1567 // This can happen if a interface-typed array narrows to a class type.
1568 jt = _type;
1569 }
1571 if (EliminateAutoBox && adr->is_AddP()) {
1572 // The pointers in the autobox arrays are always non-null
1573 Node* base = adr->in(AddPNode::Base);
1574 if (base != NULL &&
1575 !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) {
1576 Compile::AliasType* atp = C->alias_type(base->adr_type());
1577 if (is_autobox_cache(atp)) {
1578 return jt->join(TypePtr::NOTNULL)->is_ptr();
1579 }
1580 }
1581 }
1582 return jt;
1583 }
1584 }
1585 } else if (tp->base() == Type::InstPtr) {
1586 ciEnv* env = C->env();
1587 const TypeInstPtr* tinst = tp->is_instptr();
1588 ciKlass* klass = tinst->klass();
1589 assert( off != Type::OffsetBot ||
1590 // arrays can be cast to Objects
1591 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1592 // unsafe field access may not have a constant offset
1593 C->has_unsafe_access(),
1594 "Field accesses must be precise" );
1595 // For oop loads, we expect the _type to be precise
1596 if (klass == env->String_klass() &&
1597 adr->is_AddP() && off != Type::OffsetBot) {
1598 // For constant Strings treat the final fields as compile time constants.
1599 Node* base = adr->in(AddPNode::Base);
1600 const TypeOopPtr* t = phase->type(base)->isa_oopptr();
1601 if (t != NULL && t->singleton()) {
1602 ciField* field = env->String_klass()->get_field_by_offset(off, false);
1603 if (field != NULL && field->is_final()) {
1604 ciObject* string = t->const_oop();
1605 ciConstant constant = string->as_instance()->field_value(field);
1606 if (constant.basic_type() == T_INT) {
1607 return TypeInt::make(constant.as_int());
1608 } else if (constant.basic_type() == T_ARRAY) {
1609 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1610 return TypeNarrowOop::make_from_constant(constant.as_object(), true);
1611 } else {
1612 return TypeOopPtr::make_from_constant(constant.as_object(), true);
1613 }
1614 }
1615 }
1616 }
1617 }
1618 // Optimizations for constant objects
1619 ciObject* const_oop = tinst->const_oop();
1620 if (const_oop != NULL) {
1621 // For constant CallSites treat the target field as a compile time constant.
1622 if (const_oop->is_call_site()) {
1623 ciCallSite* call_site = const_oop->as_call_site();
1624 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);
1625 if (field != NULL && field->is_call_site_target()) {
1626 ciMethodHandle* target = call_site->get_target();
1627 if (target != NULL) { // just in case
1628 ciConstant constant(T_OBJECT, target);
1629 const Type* t;
1630 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1631 t = TypeNarrowOop::make_from_constant(constant.as_object(), true);
1632 } else {
1633 t = TypeOopPtr::make_from_constant(constant.as_object(), true);
1634 }
1635 // Add a dependence for invalidation of the optimization.
1636 if (!call_site->is_constant_call_site()) {
1637 C->dependencies()->assert_call_site_target_value(call_site, target);
1638 }
1639 return t;
1640 }
1641 }
1642 }
1643 }
1644 } else if (tp->base() == Type::KlassPtr) {
1645 assert( off != Type::OffsetBot ||
1646 // arrays can be cast to Objects
1647 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1648 // also allow array-loading from the primary supertype
1649 // array during subtype checks
1650 Opcode() == Op_LoadKlass,
1651 "Field accesses must be precise" );
1652 // For klass/static loads, we expect the _type to be precise
1653 }
1655 const TypeKlassPtr *tkls = tp->isa_klassptr();
1656 if (tkls != NULL && !StressReflectiveCode) {
1657 ciKlass* klass = tkls->klass();
1658 if (klass->is_loaded() && tkls->klass_is_exact()) {
1659 // We are loading a field from a Klass metaobject whose identity
1660 // is known at compile time (the type is "exact" or "precise").
1661 // Check for fields we know are maintained as constants by the VM.
1662 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1663 // The field is Klass::_super_check_offset. Return its (constant) value.
1664 // (Folds up type checking code.)
1665 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1666 return TypeInt::make(klass->super_check_offset());
1667 }
1668 // Compute index into primary_supers array
1669 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1670 // Check for overflowing; use unsigned compare to handle the negative case.
1671 if( depth < ciKlass::primary_super_limit() ) {
1672 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1673 // (Folds up type checking code.)
1674 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1675 ciKlass *ss = klass->super_of_depth(depth);
1676 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1677 }
1678 const Type* aift = load_array_final_field(tkls, klass);
1679 if (aift != NULL) return aift;
1680 if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset())
1681 && klass->is_array_klass()) {
1682 // The field is ArrayKlass::_component_mirror. Return its (constant) value.
1683 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1684 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1685 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1686 }
1687 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1688 // The field is Klass::_java_mirror. Return its (constant) value.
1689 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1690 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1691 return TypeInstPtr::make(klass->java_mirror());
1692 }
1693 }
1695 // We can still check if we are loading from the primary_supers array at a
1696 // shallow enough depth. Even though the klass is not exact, entries less
1697 // than or equal to its super depth are correct.
1698 if (klass->is_loaded() ) {
1699 ciType *inner = klass;
1700 while( inner->is_obj_array_klass() )
1701 inner = inner->as_obj_array_klass()->base_element_type();
1702 if( inner->is_instance_klass() &&
1703 !inner->as_instance_klass()->flags().is_interface() ) {
1704 // Compute index into primary_supers array
1705 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1706 // Check for overflowing; use unsigned compare to handle the negative case.
1707 if( depth < ciKlass::primary_super_limit() &&
1708 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1709 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1710 // (Folds up type checking code.)
1711 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1712 ciKlass *ss = klass->super_of_depth(depth);
1713 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1714 }
1715 }
1716 }
1718 // If the type is enough to determine that the thing is not an array,
1719 // we can give the layout_helper a positive interval type.
1720 // This will help short-circuit some reflective code.
1721 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1722 && !klass->is_array_klass() // not directly typed as an array
1723 && !klass->is_interface() // specifically not Serializable & Cloneable
1724 && !klass->is_java_lang_Object() // not the supertype of all T[]
1725 ) {
1726 // Note: When interfaces are reliable, we can narrow the interface
1727 // test to (klass != Serializable && klass != Cloneable).
1728 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1729 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1730 // The key property of this type is that it folds up tests
1731 // for array-ness, since it proves that the layout_helper is positive.
1732 // Thus, a generic value like the basic object layout helper works fine.
1733 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1734 }
1735 }
1737 // If we are loading from a freshly-allocated object, produce a zero,
1738 // if the load is provably beyond the header of the object.
1739 // (Also allow a variable load from a fresh array to produce zero.)
1740 const TypeOopPtr *tinst = tp->isa_oopptr();
1741 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1742 if (ReduceFieldZeroing || is_instance) {
1743 Node* value = can_see_stored_value(mem,phase);
1744 if (value != NULL && value->is_Con()) {
1745 assert(value->bottom_type()->higher_equal(_type),"sanity");
1746 return value->bottom_type();
1747 }
1748 }
1750 if (is_instance) {
1751 // If we have an instance type and our memory input is the
1752 // programs's initial memory state, there is no matching store,
1753 // so just return a zero of the appropriate type
1754 Node *mem = in(MemNode::Memory);
1755 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1756 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1757 return Type::get_zero_type(_type->basic_type());
1758 }
1759 }
1760 return _type;
1761 }
1763 //------------------------------match_edge-------------------------------------
1764 // Do we Match on this edge index or not? Match only the address.
1765 uint LoadNode::match_edge(uint idx) const {
1766 return idx == MemNode::Address;
1767 }
1769 //--------------------------LoadBNode::Ideal--------------------------------------
1770 //
1771 // If the previous store is to the same address as this load,
1772 // and the value stored was larger than a byte, replace this load
1773 // with the value stored truncated to a byte. If no truncation is
1774 // needed, the replacement is done in LoadNode::Identity().
1775 //
1776 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1777 Node* mem = in(MemNode::Memory);
1778 Node* value = can_see_stored_value(mem,phase);
1779 if( value && !phase->type(value)->higher_equal( _type ) ) {
1780 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) );
1781 return new (phase->C) RShiftINode(result, phase->intcon(24));
1782 }
1783 // Identity call will handle the case where truncation is not needed.
1784 return LoadNode::Ideal(phase, can_reshape);
1785 }
1787 const Type* LoadBNode::Value(PhaseTransform *phase) const {
1788 Node* mem = in(MemNode::Memory);
1789 Node* value = can_see_stored_value(mem,phase);
1790 if (value != NULL && value->is_Con() &&
1791 !value->bottom_type()->higher_equal(_type)) {
1792 // If the input to the store does not fit with the load's result type,
1793 // it must be truncated. We can't delay until Ideal call since
1794 // a singleton Value is needed for split_thru_phi optimization.
1795 int con = value->get_int();
1796 return TypeInt::make((con << 24) >> 24);
1797 }
1798 return LoadNode::Value(phase);
1799 }
1801 //--------------------------LoadUBNode::Ideal-------------------------------------
1802 //
1803 // If the previous store is to the same address as this load,
1804 // and the value stored was larger than a byte, replace this load
1805 // with the value stored truncated to a byte. If no truncation is
1806 // needed, the replacement is done in LoadNode::Identity().
1807 //
1808 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1809 Node* mem = in(MemNode::Memory);
1810 Node* value = can_see_stored_value(mem, phase);
1811 if (value && !phase->type(value)->higher_equal(_type))
1812 return new (phase->C) AndINode(value, phase->intcon(0xFF));
1813 // Identity call will handle the case where truncation is not needed.
1814 return LoadNode::Ideal(phase, can_reshape);
1815 }
1817 const Type* LoadUBNode::Value(PhaseTransform *phase) const {
1818 Node* mem = in(MemNode::Memory);
1819 Node* value = can_see_stored_value(mem,phase);
1820 if (value != NULL && value->is_Con() &&
1821 !value->bottom_type()->higher_equal(_type)) {
1822 // If the input to the store does not fit with the load's result type,
1823 // it must be truncated. We can't delay until Ideal call since
1824 // a singleton Value is needed for split_thru_phi optimization.
1825 int con = value->get_int();
1826 return TypeInt::make(con & 0xFF);
1827 }
1828 return LoadNode::Value(phase);
1829 }
1831 //--------------------------LoadUSNode::Ideal-------------------------------------
1832 //
1833 // If the previous store is to the same address as this load,
1834 // and the value stored was larger than a char, replace this load
1835 // with the value stored truncated to a char. If no truncation is
1836 // needed, the replacement is done in LoadNode::Identity().
1837 //
1838 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1839 Node* mem = in(MemNode::Memory);
1840 Node* value = can_see_stored_value(mem,phase);
1841 if( value && !phase->type(value)->higher_equal( _type ) )
1842 return new (phase->C) AndINode(value,phase->intcon(0xFFFF));
1843 // Identity call will handle the case where truncation is not needed.
1844 return LoadNode::Ideal(phase, can_reshape);
1845 }
1847 const Type* LoadUSNode::Value(PhaseTransform *phase) const {
1848 Node* mem = in(MemNode::Memory);
1849 Node* value = can_see_stored_value(mem,phase);
1850 if (value != NULL && value->is_Con() &&
1851 !value->bottom_type()->higher_equal(_type)) {
1852 // If the input to the store does not fit with the load's result type,
1853 // it must be truncated. We can't delay until Ideal call since
1854 // a singleton Value is needed for split_thru_phi optimization.
1855 int con = value->get_int();
1856 return TypeInt::make(con & 0xFFFF);
1857 }
1858 return LoadNode::Value(phase);
1859 }
1861 //--------------------------LoadSNode::Ideal--------------------------------------
1862 //
1863 // If the previous store is to the same address as this load,
1864 // and the value stored was larger than a short, replace this load
1865 // with the value stored truncated to a short. If no truncation is
1866 // needed, the replacement is done in LoadNode::Identity().
1867 //
1868 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1869 Node* mem = in(MemNode::Memory);
1870 Node* value = can_see_stored_value(mem,phase);
1871 if( value && !phase->type(value)->higher_equal( _type ) ) {
1872 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) );
1873 return new (phase->C) RShiftINode(result, phase->intcon(16));
1874 }
1875 // Identity call will handle the case where truncation is not needed.
1876 return LoadNode::Ideal(phase, can_reshape);
1877 }
1879 const Type* LoadSNode::Value(PhaseTransform *phase) const {
1880 Node* mem = in(MemNode::Memory);
1881 Node* value = can_see_stored_value(mem,phase);
1882 if (value != NULL && value->is_Con() &&
1883 !value->bottom_type()->higher_equal(_type)) {
1884 // If the input to the store does not fit with the load's result type,
1885 // it must be truncated. We can't delay until Ideal call since
1886 // a singleton Value is needed for split_thru_phi optimization.
1887 int con = value->get_int();
1888 return TypeInt::make((con << 16) >> 16);
1889 }
1890 return LoadNode::Value(phase);
1891 }
1893 //=============================================================================
1894 //----------------------------LoadKlassNode::make------------------------------
1895 // Polymorphic factory method:
1896 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
1897 Compile* C = gvn.C;
1898 Node *ctl = NULL;
1899 // sanity check the alias category against the created node type
1900 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
1901 assert(adr_type != NULL, "expecting TypeKlassPtr");
1902 #ifdef _LP64
1903 if (adr_type->is_ptr_to_narrowklass()) {
1904 assert(UseCompressedKlassPointers, "no compressed klasses");
1905 Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass()));
1906 return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
1907 }
1908 #endif
1909 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
1910 return new (C) LoadKlassNode(ctl, mem, adr, at, tk);
1911 }
1913 //------------------------------Value------------------------------------------
1914 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1915 return klass_value_common(phase);
1916 }
1918 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
1919 // Either input is TOP ==> the result is TOP
1920 const Type *t1 = phase->type( in(MemNode::Memory) );
1921 if (t1 == Type::TOP) return Type::TOP;
1922 Node *adr = in(MemNode::Address);
1923 const Type *t2 = phase->type( adr );
1924 if (t2 == Type::TOP) return Type::TOP;
1925 const TypePtr *tp = t2->is_ptr();
1926 if (TypePtr::above_centerline(tp->ptr()) ||
1927 tp->ptr() == TypePtr::Null) return Type::TOP;
1929 // Return a more precise klass, if possible
1930 const TypeInstPtr *tinst = tp->isa_instptr();
1931 if (tinst != NULL) {
1932 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1933 int offset = tinst->offset();
1934 if (ik == phase->C->env()->Class_klass()
1935 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1936 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1937 // We are loading a special hidden field from a Class mirror object,
1938 // the field which points to the VM's Klass metaobject.
1939 ciType* t = tinst->java_mirror_type();
1940 // java_mirror_type returns non-null for compile-time Class constants.
1941 if (t != NULL) {
1942 // constant oop => constant klass
1943 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1944 return TypeKlassPtr::make(ciArrayKlass::make(t));
1945 }
1946 if (!t->is_klass()) {
1947 // a primitive Class (e.g., int.class) has NULL for a klass field
1948 return TypePtr::NULL_PTR;
1949 }
1950 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1951 return TypeKlassPtr::make(t->as_klass());
1952 }
1953 // non-constant mirror, so we can't tell what's going on
1954 }
1955 if( !ik->is_loaded() )
1956 return _type; // Bail out if not loaded
1957 if (offset == oopDesc::klass_offset_in_bytes()) {
1958 if (tinst->klass_is_exact()) {
1959 return TypeKlassPtr::make(ik);
1960 }
1961 // See if we can become precise: no subklasses and no interface
1962 // (Note: We need to support verified interfaces.)
1963 if (!ik->is_interface() && !ik->has_subklass()) {
1964 //assert(!UseExactTypes, "this code should be useless with exact types");
1965 // Add a dependence; if any subclass added we need to recompile
1966 if (!ik->is_final()) {
1967 // %%% should use stronger assert_unique_concrete_subtype instead
1968 phase->C->dependencies()->assert_leaf_type(ik);
1969 }
1970 // Return precise klass
1971 return TypeKlassPtr::make(ik);
1972 }
1974 // Return root of possible klass
1975 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1976 }
1977 }
1979 // Check for loading klass from an array
1980 const TypeAryPtr *tary = tp->isa_aryptr();
1981 if( tary != NULL ) {
1982 ciKlass *tary_klass = tary->klass();
1983 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1984 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1985 if (tary->klass_is_exact()) {
1986 return TypeKlassPtr::make(tary_klass);
1987 }
1988 ciArrayKlass *ak = tary->klass()->as_array_klass();
1989 // If the klass is an object array, we defer the question to the
1990 // array component klass.
1991 if( ak->is_obj_array_klass() ) {
1992 assert( ak->is_loaded(), "" );
1993 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1994 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1995 ciInstanceKlass* ik = base_k->as_instance_klass();
1996 // See if we can become precise: no subklasses and no interface
1997 if (!ik->is_interface() && !ik->has_subklass()) {
1998 //assert(!UseExactTypes, "this code should be useless with exact types");
1999 // Add a dependence; if any subclass added we need to recompile
2000 if (!ik->is_final()) {
2001 phase->C->dependencies()->assert_leaf_type(ik);
2002 }
2003 // Return precise array klass
2004 return TypeKlassPtr::make(ak);
2005 }
2006 }
2007 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2008 } else { // Found a type-array?
2009 //assert(!UseExactTypes, "this code should be useless with exact types");
2010 assert( ak->is_type_array_klass(), "" );
2011 return TypeKlassPtr::make(ak); // These are always precise
2012 }
2013 }
2014 }
2016 // Check for loading klass from an array klass
2017 const TypeKlassPtr *tkls = tp->isa_klassptr();
2018 if (tkls != NULL && !StressReflectiveCode) {
2019 ciKlass* klass = tkls->klass();
2020 if( !klass->is_loaded() )
2021 return _type; // Bail out if not loaded
2022 if( klass->is_obj_array_klass() &&
2023 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2024 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2025 // // Always returning precise element type is incorrect,
2026 // // e.g., element type could be object and array may contain strings
2027 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2029 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2030 // according to the element type's subclassing.
2031 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2032 }
2033 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2034 tkls->offset() == in_bytes(Klass::super_offset())) {
2035 ciKlass* sup = klass->as_instance_klass()->super();
2036 // The field is Klass::_super. Return its (constant) value.
2037 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2038 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2039 }
2040 }
2042 // Bailout case
2043 return LoadNode::Value(phase);
2044 }
2046 //------------------------------Identity---------------------------------------
2047 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2048 // Also feed through the klass in Allocate(...klass...)._klass.
2049 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
2050 return klass_identity_common(phase);
2051 }
2053 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
2054 Node* x = LoadNode::Identity(phase);
2055 if (x != this) return x;
2057 // Take apart the address into an oop and and offset.
2058 // Return 'this' if we cannot.
2059 Node* adr = in(MemNode::Address);
2060 intptr_t offset = 0;
2061 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2062 if (base == NULL) return this;
2063 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2064 if (toop == NULL) return this;
2066 // We can fetch the klass directly through an AllocateNode.
2067 // This works even if the klass is not constant (clone or newArray).
2068 if (offset == oopDesc::klass_offset_in_bytes()) {
2069 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2070 if (allocated_klass != NULL) {
2071 return allocated_klass;
2072 }
2073 }
2075 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2076 // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass.
2077 // See inline_native_Class_query for occurrences of these patterns.
2078 // Java Example: x.getClass().isAssignableFrom(y)
2079 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
2080 //
2081 // This improves reflective code, often making the Class
2082 // mirror go completely dead. (Current exception: Class
2083 // mirrors may appear in debug info, but we could clean them out by
2084 // introducing a new debug info operator for Klass*.java_mirror).
2085 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2086 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2087 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2088 // We are loading a special hidden field from a Class mirror,
2089 // the field which points to its Klass or ArrayKlass metaobject.
2090 if (base->is_Load()) {
2091 Node* adr2 = base->in(MemNode::Address);
2092 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2093 if (tkls != NULL && !tkls->empty()
2094 && (tkls->klass()->is_instance_klass() ||
2095 tkls->klass()->is_array_klass())
2096 && adr2->is_AddP()
2097 ) {
2098 int mirror_field = in_bytes(Klass::java_mirror_offset());
2099 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2100 mirror_field = in_bytes(ArrayKlass::component_mirror_offset());
2101 }
2102 if (tkls->offset() == mirror_field) {
2103 return adr2->in(AddPNode::Base);
2104 }
2105 }
2106 }
2107 }
2109 return this;
2110 }
2113 //------------------------------Value------------------------------------------
2114 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
2115 const Type *t = klass_value_common(phase);
2116 if (t == Type::TOP)
2117 return t;
2119 return t->make_narrowklass();
2120 }
2122 //------------------------------Identity---------------------------------------
2123 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2124 // Also feed through the klass in Allocate(...klass...)._klass.
2125 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
2126 Node *x = klass_identity_common(phase);
2128 const Type *t = phase->type( x );
2129 if( t == Type::TOP ) return x;
2130 if( t->isa_narrowklass()) return x;
2131 assert (!t->isa_narrowoop(), "no narrow oop here");
2133 return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass()));
2134 }
2136 //------------------------------Value-----------------------------------------
2137 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
2138 // Either input is TOP ==> the result is TOP
2139 const Type *t1 = phase->type( in(MemNode::Memory) );
2140 if( t1 == Type::TOP ) return Type::TOP;
2141 Node *adr = in(MemNode::Address);
2142 const Type *t2 = phase->type( adr );
2143 if( t2 == Type::TOP ) return Type::TOP;
2144 const TypePtr *tp = t2->is_ptr();
2145 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2146 const TypeAryPtr *tap = tp->isa_aryptr();
2147 if( !tap ) return _type;
2148 return tap->size();
2149 }
2151 //-------------------------------Ideal---------------------------------------
2152 // Feed through the length in AllocateArray(...length...)._length.
2153 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2154 Node* p = MemNode::Ideal_common(phase, can_reshape);
2155 if (p) return (p == NodeSentinel) ? NULL : p;
2157 // Take apart the address into an oop and and offset.
2158 // Return 'this' if we cannot.
2159 Node* adr = in(MemNode::Address);
2160 intptr_t offset = 0;
2161 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2162 if (base == NULL) return NULL;
2163 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2164 if (tary == NULL) return NULL;
2166 // We can fetch the length directly through an AllocateArrayNode.
2167 // This works even if the length is not constant (clone or newArray).
2168 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2169 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2170 if (alloc != NULL) {
2171 Node* allocated_length = alloc->Ideal_length();
2172 Node* len = alloc->make_ideal_length(tary, phase);
2173 if (allocated_length != len) {
2174 // New CastII improves on this.
2175 return len;
2176 }
2177 }
2178 }
2180 return NULL;
2181 }
2183 //------------------------------Identity---------------------------------------
2184 // Feed through the length in AllocateArray(...length...)._length.
2185 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
2186 Node* x = LoadINode::Identity(phase);
2187 if (x != this) return x;
2189 // Take apart the address into an oop and and offset.
2190 // Return 'this' if we cannot.
2191 Node* adr = in(MemNode::Address);
2192 intptr_t offset = 0;
2193 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2194 if (base == NULL) return this;
2195 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2196 if (tary == NULL) return this;
2198 // We can fetch the length directly through an AllocateArrayNode.
2199 // This works even if the length is not constant (clone or newArray).
2200 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2201 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2202 if (alloc != NULL) {
2203 Node* allocated_length = alloc->Ideal_length();
2204 // Do not allow make_ideal_length to allocate a CastII node.
2205 Node* len = alloc->make_ideal_length(tary, phase, false);
2206 if (allocated_length == len) {
2207 // Return allocated_length only if it would not be improved by a CastII.
2208 return allocated_length;
2209 }
2210 }
2211 }
2213 return this;
2215 }
2217 //=============================================================================
2218 //---------------------------StoreNode::make-----------------------------------
2219 // Polymorphic factory method:
2220 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
2221 Compile* C = gvn.C;
2222 assert( C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2223 ctl != NULL, "raw memory operations should have control edge");
2225 switch (bt) {
2226 case T_BOOLEAN:
2227 case T_BYTE: return new (C) StoreBNode(ctl, mem, adr, adr_type, val);
2228 case T_INT: return new (C) StoreINode(ctl, mem, adr, adr_type, val);
2229 case T_CHAR:
2230 case T_SHORT: return new (C) StoreCNode(ctl, mem, adr, adr_type, val);
2231 case T_LONG: return new (C) StoreLNode(ctl, mem, adr, adr_type, val);
2232 case T_FLOAT: return new (C) StoreFNode(ctl, mem, adr, adr_type, val);
2233 case T_DOUBLE: return new (C) StoreDNode(ctl, mem, adr, adr_type, val);
2234 case T_METADATA:
2235 case T_ADDRESS:
2236 case T_OBJECT:
2237 #ifdef _LP64
2238 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2239 val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop()));
2240 return new (C) StoreNNode(ctl, mem, adr, adr_type, val);
2241 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2242 (UseCompressedKlassPointers && val->bottom_type()->isa_klassptr() &&
2243 adr->bottom_type()->isa_rawptr())) {
2244 val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2245 return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val);
2246 }
2247 #endif
2248 {
2249 return new (C) StorePNode(ctl, mem, adr, adr_type, val);
2250 }
2251 }
2252 ShouldNotReachHere();
2253 return (StoreNode*)NULL;
2254 }
2256 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
2257 bool require_atomic = true;
2258 return new (C) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
2259 }
2262 //--------------------------bottom_type----------------------------------------
2263 const Type *StoreNode::bottom_type() const {
2264 return Type::MEMORY;
2265 }
2267 //------------------------------hash-------------------------------------------
2268 uint StoreNode::hash() const {
2269 // unroll addition of interesting fields
2270 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2272 // Since they are not commoned, do not hash them:
2273 return NO_HASH;
2274 }
2276 //------------------------------Ideal------------------------------------------
2277 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2278 // When a store immediately follows a relevant allocation/initialization,
2279 // try to capture it into the initialization, or hoist it above.
2280 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2281 Node* p = MemNode::Ideal_common(phase, can_reshape);
2282 if (p) return (p == NodeSentinel) ? NULL : p;
2284 Node* mem = in(MemNode::Memory);
2285 Node* address = in(MemNode::Address);
2287 // Back-to-back stores to same address? Fold em up. Generally
2288 // unsafe if I have intervening uses... Also disallowed for StoreCM
2289 // since they must follow each StoreP operation. Redundant StoreCMs
2290 // are eliminated just before matching in final_graph_reshape.
2291 if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) &&
2292 mem->Opcode() != Op_StoreCM) {
2293 // Looking at a dead closed cycle of memory?
2294 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2296 assert(Opcode() == mem->Opcode() ||
2297 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
2298 "no mismatched stores, except on raw memory");
2300 if (mem->outcnt() == 1 && // check for intervening uses
2301 mem->as_Store()->memory_size() <= this->memory_size()) {
2302 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
2303 // For example, 'mem' might be the final state at a conditional return.
2304 // Or, 'mem' might be used by some node which is live at the same time
2305 // 'this' is live, which might be unschedulable. So, require exactly
2306 // ONE user, the 'this' store, until such time as we clone 'mem' for
2307 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
2308 if (can_reshape) { // (%%% is this an anachronism?)
2309 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
2310 phase->is_IterGVN());
2311 } else {
2312 // It's OK to do this in the parser, since DU info is always accurate,
2313 // and the parser always refers to nodes via SafePointNode maps.
2314 set_req(MemNode::Memory, mem->in(MemNode::Memory));
2315 }
2316 return this;
2317 }
2318 }
2320 // Capture an unaliased, unconditional, simple store into an initializer.
2321 // Or, if it is independent of the allocation, hoist it above the allocation.
2322 if (ReduceFieldZeroing && /*can_reshape &&*/
2323 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2324 InitializeNode* init = mem->in(0)->as_Initialize();
2325 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2326 if (offset > 0) {
2327 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2328 // If the InitializeNode captured me, it made a raw copy of me,
2329 // and I need to disappear.
2330 if (moved != NULL) {
2331 // %%% hack to ensure that Ideal returns a new node:
2332 mem = MergeMemNode::make(phase->C, mem);
2333 return mem; // fold me away
2334 }
2335 }
2336 }
2338 return NULL; // No further progress
2339 }
2341 //------------------------------Value-----------------------------------------
2342 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2343 // Either input is TOP ==> the result is TOP
2344 const Type *t1 = phase->type( in(MemNode::Memory) );
2345 if( t1 == Type::TOP ) return Type::TOP;
2346 const Type *t2 = phase->type( in(MemNode::Address) );
2347 if( t2 == Type::TOP ) return Type::TOP;
2348 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2349 if( t3 == Type::TOP ) return Type::TOP;
2350 return Type::MEMORY;
2351 }
2353 //------------------------------Identity---------------------------------------
2354 // Remove redundant stores:
2355 // Store(m, p, Load(m, p)) changes to m.
2356 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2357 Node *StoreNode::Identity( PhaseTransform *phase ) {
2358 Node* mem = in(MemNode::Memory);
2359 Node* adr = in(MemNode::Address);
2360 Node* val = in(MemNode::ValueIn);
2362 // Load then Store? Then the Store is useless
2363 if (val->is_Load() &&
2364 val->in(MemNode::Address)->eqv_uncast(adr) &&
2365 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2366 val->as_Load()->store_Opcode() == Opcode()) {
2367 return mem;
2368 }
2370 // Two stores in a row of the same value?
2371 if (mem->is_Store() &&
2372 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2373 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2374 mem->Opcode() == Opcode()) {
2375 return mem;
2376 }
2378 // Store of zero anywhere into a freshly-allocated object?
2379 // Then the store is useless.
2380 // (It must already have been captured by the InitializeNode.)
2381 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2382 // a newly allocated object is already all-zeroes everywhere
2383 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2384 return mem;
2385 }
2387 // the store may also apply to zero-bits in an earlier object
2388 Node* prev_mem = find_previous_store(phase);
2389 // Steps (a), (b): Walk past independent stores to find an exact match.
2390 if (prev_mem != NULL) {
2391 Node* prev_val = can_see_stored_value(prev_mem, phase);
2392 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2393 // prev_val and val might differ by a cast; it would be good
2394 // to keep the more informative of the two.
2395 return mem;
2396 }
2397 }
2398 }
2400 return this;
2401 }
2403 //------------------------------match_edge-------------------------------------
2404 // Do we Match on this edge index or not? Match only memory & value
2405 uint StoreNode::match_edge(uint idx) const {
2406 return idx == MemNode::Address || idx == MemNode::ValueIn;
2407 }
2409 //------------------------------cmp--------------------------------------------
2410 // Do not common stores up together. They generally have to be split
2411 // back up anyways, so do not bother.
2412 uint StoreNode::cmp( const Node &n ) const {
2413 return (&n == this); // Always fail except on self
2414 }
2416 //------------------------------Ideal_masked_input-----------------------------
2417 // Check for a useless mask before a partial-word store
2418 // (StoreB ... (AndI valIn conIa) )
2419 // If (conIa & mask == mask) this simplifies to
2420 // (StoreB ... (valIn) )
2421 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2422 Node *val = in(MemNode::ValueIn);
2423 if( val->Opcode() == Op_AndI ) {
2424 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2425 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2426 set_req(MemNode::ValueIn, val->in(1));
2427 return this;
2428 }
2429 }
2430 return NULL;
2431 }
2434 //------------------------------Ideal_sign_extended_input----------------------
2435 // Check for useless sign-extension before a partial-word store
2436 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2437 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2438 // (StoreB ... (valIn) )
2439 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2440 Node *val = in(MemNode::ValueIn);
2441 if( val->Opcode() == Op_RShiftI ) {
2442 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2443 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2444 Node *shl = val->in(1);
2445 if( shl->Opcode() == Op_LShiftI ) {
2446 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2447 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2448 set_req(MemNode::ValueIn, shl->in(1));
2449 return this;
2450 }
2451 }
2452 }
2453 }
2454 return NULL;
2455 }
2457 //------------------------------value_never_loaded-----------------------------------
2458 // Determine whether there are any possible loads of the value stored.
2459 // For simplicity, we actually check if there are any loads from the
2460 // address stored to, not just for loads of the value stored by this node.
2461 //
2462 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2463 Node *adr = in(Address);
2464 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2465 if (adr_oop == NULL)
2466 return false;
2467 if (!adr_oop->is_known_instance_field())
2468 return false; // if not a distinct instance, there may be aliases of the address
2469 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2470 Node *use = adr->fast_out(i);
2471 int opc = use->Opcode();
2472 if (use->is_Load() || use->is_LoadStore()) {
2473 return false;
2474 }
2475 }
2476 return true;
2477 }
2479 //=============================================================================
2480 //------------------------------Ideal------------------------------------------
2481 // If the store is from an AND mask that leaves the low bits untouched, then
2482 // we can skip the AND operation. If the store is from a sign-extension
2483 // (a left shift, then right shift) we can skip both.
2484 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2485 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2486 if( progress != NULL ) return progress;
2488 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2489 if( progress != NULL ) return progress;
2491 // Finally check the default case
2492 return StoreNode::Ideal(phase, can_reshape);
2493 }
2495 //=============================================================================
2496 //------------------------------Ideal------------------------------------------
2497 // If the store is from an AND mask that leaves the low bits untouched, then
2498 // we can skip the AND operation
2499 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2500 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2501 if( progress != NULL ) return progress;
2503 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2504 if( progress != NULL ) return progress;
2506 // Finally check the default case
2507 return StoreNode::Ideal(phase, can_reshape);
2508 }
2510 //=============================================================================
2511 //------------------------------Identity---------------------------------------
2512 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2513 // No need to card mark when storing a null ptr
2514 Node* my_store = in(MemNode::OopStore);
2515 if (my_store->is_Store()) {
2516 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2517 if( t1 == TypePtr::NULL_PTR ) {
2518 return in(MemNode::Memory);
2519 }
2520 }
2521 return this;
2522 }
2524 //=============================================================================
2525 //------------------------------Ideal---------------------------------------
2526 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2527 Node* progress = StoreNode::Ideal(phase, can_reshape);
2528 if (progress != NULL) return progress;
2530 Node* my_store = in(MemNode::OopStore);
2531 if (my_store->is_MergeMem()) {
2532 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2533 set_req(MemNode::OopStore, mem);
2534 return this;
2535 }
2537 return NULL;
2538 }
2540 //------------------------------Value-----------------------------------------
2541 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2542 // Either input is TOP ==> the result is TOP
2543 const Type *t = phase->type( in(MemNode::Memory) );
2544 if( t == Type::TOP ) return Type::TOP;
2545 t = phase->type( in(MemNode::Address) );
2546 if( t == Type::TOP ) return Type::TOP;
2547 t = phase->type( in(MemNode::ValueIn) );
2548 if( t == Type::TOP ) return Type::TOP;
2549 // If extra input is TOP ==> the result is TOP
2550 t = phase->type( in(MemNode::OopStore) );
2551 if( t == Type::TOP ) return Type::TOP;
2553 return StoreNode::Value( phase );
2554 }
2557 //=============================================================================
2558 //----------------------------------SCMemProjNode------------------------------
2559 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2560 {
2561 return bottom_type();
2562 }
2564 //=============================================================================
2565 //----------------------------------LoadStoreNode------------------------------
2566 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2567 : Node(required),
2568 _type(rt),
2569 _adr_type(at)
2570 {
2571 init_req(MemNode::Control, c );
2572 init_req(MemNode::Memory , mem);
2573 init_req(MemNode::Address, adr);
2574 init_req(MemNode::ValueIn, val);
2575 init_class_id(Class_LoadStore);
2576 }
2578 uint LoadStoreNode::ideal_reg() const {
2579 return _type->ideal_reg();
2580 }
2582 bool LoadStoreNode::result_not_used() const {
2583 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2584 Node *x = fast_out(i);
2585 if (x->Opcode() == Op_SCMemProj) continue;
2586 return false;
2587 }
2588 return true;
2589 }
2591 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2593 //=============================================================================
2594 //----------------------------------LoadStoreConditionalNode--------------------
2595 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2596 init_req(ExpectedIn, ex );
2597 }
2599 //=============================================================================
2600 //-------------------------------adr_type--------------------------------------
2601 // Do we Match on this edge index or not? Do not match memory
2602 const TypePtr* ClearArrayNode::adr_type() const {
2603 Node *adr = in(3);
2604 return MemNode::calculate_adr_type(adr->bottom_type());
2605 }
2607 //------------------------------match_edge-------------------------------------
2608 // Do we Match on this edge index or not? Do not match memory
2609 uint ClearArrayNode::match_edge(uint idx) const {
2610 return idx > 1;
2611 }
2613 //------------------------------Identity---------------------------------------
2614 // Clearing a zero length array does nothing
2615 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2616 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2617 }
2619 //------------------------------Idealize---------------------------------------
2620 // Clearing a short array is faster with stores
2621 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2622 const int unit = BytesPerLong;
2623 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2624 if (!t) return NULL;
2625 if (!t->is_con()) return NULL;
2626 intptr_t raw_count = t->get_con();
2627 intptr_t size = raw_count;
2628 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2629 // Clearing nothing uses the Identity call.
2630 // Negative clears are possible on dead ClearArrays
2631 // (see jck test stmt114.stmt11402.val).
2632 if (size <= 0 || size % unit != 0) return NULL;
2633 intptr_t count = size / unit;
2634 // Length too long; use fast hardware clear
2635 if (size > Matcher::init_array_short_size) return NULL;
2636 Node *mem = in(1);
2637 if( phase->type(mem)==Type::TOP ) return NULL;
2638 Node *adr = in(3);
2639 const Type* at = phase->type(adr);
2640 if( at==Type::TOP ) return NULL;
2641 const TypePtr* atp = at->isa_ptr();
2642 // adjust atp to be the correct array element address type
2643 if (atp == NULL) atp = TypePtr::BOTTOM;
2644 else atp = atp->add_offset(Type::OffsetBot);
2645 // Get base for derived pointer purposes
2646 if( adr->Opcode() != Op_AddP ) Unimplemented();
2647 Node *base = adr->in(1);
2649 Node *zero = phase->makecon(TypeLong::ZERO);
2650 Node *off = phase->MakeConX(BytesPerLong);
2651 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero);
2652 count--;
2653 while( count-- ) {
2654 mem = phase->transform(mem);
2655 adr = phase->transform(new (phase->C) AddPNode(base,adr,off));
2656 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero);
2657 }
2658 return mem;
2659 }
2661 //----------------------------step_through----------------------------------
2662 // Return allocation input memory edge if it is different instance
2663 // or itself if it is the one we are looking for.
2664 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2665 Node* n = *np;
2666 assert(n->is_ClearArray(), "sanity");
2667 intptr_t offset;
2668 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2669 // This method is called only before Allocate nodes are expanded during
2670 // macro nodes expansion. Before that ClearArray nodes are only generated
2671 // in LibraryCallKit::generate_arraycopy() which follows allocations.
2672 assert(alloc != NULL, "should have allocation");
2673 if (alloc->_idx == instance_id) {
2674 // Can not bypass initialization of the instance we are looking for.
2675 return false;
2676 }
2677 // Otherwise skip it.
2678 InitializeNode* init = alloc->initialization();
2679 if (init != NULL)
2680 *np = init->in(TypeFunc::Memory);
2681 else
2682 *np = alloc->in(TypeFunc::Memory);
2683 return true;
2684 }
2686 //----------------------------clear_memory-------------------------------------
2687 // Generate code to initialize object storage to zero.
2688 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2689 intptr_t start_offset,
2690 Node* end_offset,
2691 PhaseGVN* phase) {
2692 Compile* C = phase->C;
2693 intptr_t offset = start_offset;
2695 int unit = BytesPerLong;
2696 if ((offset % unit) != 0) {
2697 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(offset));
2698 adr = phase->transform(adr);
2699 const TypePtr* atp = TypeRawPtr::BOTTOM;
2700 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2701 mem = phase->transform(mem);
2702 offset += BytesPerInt;
2703 }
2704 assert((offset % unit) == 0, "");
2706 // Initialize the remaining stuff, if any, with a ClearArray.
2707 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2708 }
2710 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2711 Node* start_offset,
2712 Node* end_offset,
2713 PhaseGVN* phase) {
2714 if (start_offset == end_offset) {
2715 // nothing to do
2716 return mem;
2717 }
2719 Compile* C = phase->C;
2720 int unit = BytesPerLong;
2721 Node* zbase = start_offset;
2722 Node* zend = end_offset;
2724 // Scale to the unit required by the CPU:
2725 if (!Matcher::init_array_count_is_in_bytes) {
2726 Node* shift = phase->intcon(exact_log2(unit));
2727 zbase = phase->transform( new(C) URShiftXNode(zbase, shift) );
2728 zend = phase->transform( new(C) URShiftXNode(zend, shift) );
2729 }
2731 // Bulk clear double-words
2732 Node* zsize = phase->transform( new(C) SubXNode(zend, zbase) );
2733 Node* adr = phase->transform( new(C) AddPNode(dest, dest, start_offset) );
2734 mem = new (C) ClearArrayNode(ctl, mem, zsize, adr);
2735 return phase->transform(mem);
2736 }
2738 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2739 intptr_t start_offset,
2740 intptr_t end_offset,
2741 PhaseGVN* phase) {
2742 if (start_offset == end_offset) {
2743 // nothing to do
2744 return mem;
2745 }
2747 Compile* C = phase->C;
2748 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2749 intptr_t done_offset = end_offset;
2750 if ((done_offset % BytesPerLong) != 0) {
2751 done_offset -= BytesPerInt;
2752 }
2753 if (done_offset > start_offset) {
2754 mem = clear_memory(ctl, mem, dest,
2755 start_offset, phase->MakeConX(done_offset), phase);
2756 }
2757 if (done_offset < end_offset) { // emit the final 32-bit store
2758 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(done_offset));
2759 adr = phase->transform(adr);
2760 const TypePtr* atp = TypeRawPtr::BOTTOM;
2761 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2762 mem = phase->transform(mem);
2763 done_offset += BytesPerInt;
2764 }
2765 assert(done_offset == end_offset, "");
2766 return mem;
2767 }
2769 //=============================================================================
2770 // Do not match memory edge.
2771 uint StrIntrinsicNode::match_edge(uint idx) const {
2772 return idx == 2 || idx == 3;
2773 }
2775 //------------------------------Ideal------------------------------------------
2776 // Return a node which is more "ideal" than the current node. Strip out
2777 // control copies
2778 Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2779 if (remove_dead_region(phase, can_reshape)) return this;
2780 // Don't bother trying to transform a dead node
2781 if (in(0) && in(0)->is_top()) return NULL;
2783 if (can_reshape) {
2784 Node* mem = phase->transform(in(MemNode::Memory));
2785 // If transformed to a MergeMem, get the desired slice
2786 uint alias_idx = phase->C->get_alias_index(adr_type());
2787 mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem;
2788 if (mem != in(MemNode::Memory)) {
2789 set_req(MemNode::Memory, mem);
2790 return this;
2791 }
2792 }
2793 return NULL;
2794 }
2796 //------------------------------Value------------------------------------------
2797 const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const {
2798 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP;
2799 return bottom_type();
2800 }
2802 //=============================================================================
2803 //------------------------------match_edge-------------------------------------
2804 // Do not match memory edge
2805 uint EncodeISOArrayNode::match_edge(uint idx) const {
2806 return idx == 2 || idx == 3; // EncodeISOArray src (Binary dst len)
2807 }
2809 //------------------------------Ideal------------------------------------------
2810 // Return a node which is more "ideal" than the current node. Strip out
2811 // control copies
2812 Node *EncodeISOArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2813 return remove_dead_region(phase, can_reshape) ? this : NULL;
2814 }
2816 //------------------------------Value------------------------------------------
2817 const Type *EncodeISOArrayNode::Value(PhaseTransform *phase) const {
2818 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP;
2819 return bottom_type();
2820 }
2822 //=============================================================================
2823 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2824 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2825 _adr_type(C->get_adr_type(alias_idx))
2826 {
2827 init_class_id(Class_MemBar);
2828 Node* top = C->top();
2829 init_req(TypeFunc::I_O,top);
2830 init_req(TypeFunc::FramePtr,top);
2831 init_req(TypeFunc::ReturnAdr,top);
2832 if (precedent != NULL)
2833 init_req(TypeFunc::Parms, precedent);
2834 }
2836 //------------------------------cmp--------------------------------------------
2837 uint MemBarNode::hash() const { return NO_HASH; }
2838 uint MemBarNode::cmp( const Node &n ) const {
2839 return (&n == this); // Always fail except on self
2840 }
2842 //------------------------------make-------------------------------------------
2843 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2844 switch (opcode) {
2845 case Op_MemBarAcquire: return new(C) MemBarAcquireNode(C, atp, pn);
2846 case Op_MemBarRelease: return new(C) MemBarReleaseNode(C, atp, pn);
2847 case Op_MemBarAcquireLock: return new(C) MemBarAcquireLockNode(C, atp, pn);
2848 case Op_MemBarReleaseLock: return new(C) MemBarReleaseLockNode(C, atp, pn);
2849 case Op_MemBarVolatile: return new(C) MemBarVolatileNode(C, atp, pn);
2850 case Op_MemBarCPUOrder: return new(C) MemBarCPUOrderNode(C, atp, pn);
2851 case Op_Initialize: return new(C) InitializeNode(C, atp, pn);
2852 case Op_MemBarStoreStore: return new(C) MemBarStoreStoreNode(C, atp, pn);
2853 default: ShouldNotReachHere(); return NULL;
2854 }
2855 }
2857 //------------------------------Ideal------------------------------------------
2858 // Return a node which is more "ideal" than the current node. Strip out
2859 // control copies
2860 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2861 if (remove_dead_region(phase, can_reshape)) return this;
2862 // Don't bother trying to transform a dead node
2863 if (in(0) && in(0)->is_top()) return NULL;
2865 // Eliminate volatile MemBars for scalar replaced objects.
2866 if (can_reshape && req() == (Precedent+1) &&
2867 (Opcode() == Op_MemBarAcquire || Opcode() == Op_MemBarVolatile)) {
2868 // Volatile field loads and stores.
2869 Node* my_mem = in(MemBarNode::Precedent);
2870 if (my_mem != NULL && my_mem->is_Mem()) {
2871 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
2872 // Check for scalar replaced object reference.
2873 if( t_oop != NULL && t_oop->is_known_instance_field() &&
2874 t_oop->offset() != Type::OffsetBot &&
2875 t_oop->offset() != Type::OffsetTop) {
2876 // Replace MemBar projections by its inputs.
2877 PhaseIterGVN* igvn = phase->is_IterGVN();
2878 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
2879 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
2880 // Must return either the original node (now dead) or a new node
2881 // (Do not return a top here, since that would break the uniqueness of top.)
2882 return new (phase->C) ConINode(TypeInt::ZERO);
2883 }
2884 }
2885 }
2886 return NULL;
2887 }
2889 //------------------------------Value------------------------------------------
2890 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2891 if( !in(0) ) return Type::TOP;
2892 if( phase->type(in(0)) == Type::TOP )
2893 return Type::TOP;
2894 return TypeTuple::MEMBAR;
2895 }
2897 //------------------------------match------------------------------------------
2898 // Construct projections for memory.
2899 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2900 switch (proj->_con) {
2901 case TypeFunc::Control:
2902 case TypeFunc::Memory:
2903 return new (m->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2904 }
2905 ShouldNotReachHere();
2906 return NULL;
2907 }
2909 //===========================InitializeNode====================================
2910 // SUMMARY:
2911 // This node acts as a memory barrier on raw memory, after some raw stores.
2912 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2913 // The Initialize can 'capture' suitably constrained stores as raw inits.
2914 // It can coalesce related raw stores into larger units (called 'tiles').
2915 // It can avoid zeroing new storage for memory units which have raw inits.
2916 // At macro-expansion, it is marked 'complete', and does not optimize further.
2917 //
2918 // EXAMPLE:
2919 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2920 // ctl = incoming control; mem* = incoming memory
2921 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
2922 // First allocate uninitialized memory and fill in the header:
2923 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
2924 // ctl := alloc.Control; mem* := alloc.Memory*
2925 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
2926 // Then initialize to zero the non-header parts of the raw memory block:
2927 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2928 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2929 // After the initialize node executes, the object is ready for service:
2930 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2931 // Suppose its body is immediately initialized as {1,2}:
2932 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2933 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2934 // mem.SLICE(#short[*]) := store2
2935 //
2936 // DETAILS:
2937 // An InitializeNode collects and isolates object initialization after
2938 // an AllocateNode and before the next possible safepoint. As a
2939 // memory barrier (MemBarNode), it keeps critical stores from drifting
2940 // down past any safepoint or any publication of the allocation.
2941 // Before this barrier, a newly-allocated object may have uninitialized bits.
2942 // After this barrier, it may be treated as a real oop, and GC is allowed.
2943 //
2944 // The semantics of the InitializeNode include an implicit zeroing of
2945 // the new object from object header to the end of the object.
2946 // (The object header and end are determined by the AllocateNode.)
2947 //
2948 // Certain stores may be added as direct inputs to the InitializeNode.
2949 // These stores must update raw memory, and they must be to addresses
2950 // derived from the raw address produced by AllocateNode, and with
2951 // a constant offset. They must be ordered by increasing offset.
2952 // The first one is at in(RawStores), the last at in(req()-1).
2953 // Unlike most memory operations, they are not linked in a chain,
2954 // but are displayed in parallel as users of the rawmem output of
2955 // the allocation.
2956 //
2957 // (See comments in InitializeNode::capture_store, which continue
2958 // the example given above.)
2959 //
2960 // When the associated Allocate is macro-expanded, the InitializeNode
2961 // may be rewritten to optimize collected stores. A ClearArrayNode
2962 // may also be created at that point to represent any required zeroing.
2963 // The InitializeNode is then marked 'complete', prohibiting further
2964 // capturing of nearby memory operations.
2965 //
2966 // During macro-expansion, all captured initializations which store
2967 // constant values of 32 bits or smaller are coalesced (if advantageous)
2968 // into larger 'tiles' 32 or 64 bits. This allows an object to be
2969 // initialized in fewer memory operations. Memory words which are
2970 // covered by neither tiles nor non-constant stores are pre-zeroed
2971 // by explicit stores of zero. (The code shape happens to do all
2972 // zeroing first, then all other stores, with both sequences occurring
2973 // in order of ascending offsets.)
2974 //
2975 // Alternatively, code may be inserted between an AllocateNode and its
2976 // InitializeNode, to perform arbitrary initialization of the new object.
2977 // E.g., the object copying intrinsics insert complex data transfers here.
2978 // The initialization must then be marked as 'complete' disable the
2979 // built-in zeroing semantics and the collection of initializing stores.
2980 //
2981 // While an InitializeNode is incomplete, reads from the memory state
2982 // produced by it are optimizable if they match the control edge and
2983 // new oop address associated with the allocation/initialization.
2984 // They return a stored value (if the offset matches) or else zero.
2985 // A write to the memory state, if it matches control and address,
2986 // and if it is to a constant offset, may be 'captured' by the
2987 // InitializeNode. It is cloned as a raw memory operation and rewired
2988 // inside the initialization, to the raw oop produced by the allocation.
2989 // Operations on addresses which are provably distinct (e.g., to
2990 // other AllocateNodes) are allowed to bypass the initialization.
2991 //
2992 // The effect of all this is to consolidate object initialization
2993 // (both arrays and non-arrays, both piecewise and bulk) into a
2994 // single location, where it can be optimized as a unit.
2995 //
2996 // Only stores with an offset less than TrackedInitializationLimit words
2997 // will be considered for capture by an InitializeNode. This puts a
2998 // reasonable limit on the complexity of optimized initializations.
3000 //---------------------------InitializeNode------------------------------------
3001 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3002 : _is_complete(Incomplete), _does_not_escape(false),
3003 MemBarNode(C, adr_type, rawoop)
3004 {
3005 init_class_id(Class_Initialize);
3007 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3008 assert(in(RawAddress) == rawoop, "proper init");
3009 // Note: allocation() can be NULL, for secondary initialization barriers
3010 }
3012 // Since this node is not matched, it will be processed by the
3013 // register allocator. Declare that there are no constraints
3014 // on the allocation of the RawAddress edge.
3015 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3016 // This edge should be set to top, by the set_complete. But be conservative.
3017 if (idx == InitializeNode::RawAddress)
3018 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3019 return RegMask::Empty;
3020 }
3022 Node* InitializeNode::memory(uint alias_idx) {
3023 Node* mem = in(Memory);
3024 if (mem->is_MergeMem()) {
3025 return mem->as_MergeMem()->memory_at(alias_idx);
3026 } else {
3027 // incoming raw memory is not split
3028 return mem;
3029 }
3030 }
3032 bool InitializeNode::is_non_zero() {
3033 if (is_complete()) return false;
3034 remove_extra_zeroes();
3035 return (req() > RawStores);
3036 }
3038 void InitializeNode::set_complete(PhaseGVN* phase) {
3039 assert(!is_complete(), "caller responsibility");
3040 _is_complete = Complete;
3042 // After this node is complete, it contains a bunch of
3043 // raw-memory initializations. There is no need for
3044 // it to have anything to do with non-raw memory effects.
3045 // Therefore, tell all non-raw users to re-optimize themselves,
3046 // after skipping the memory effects of this initialization.
3047 PhaseIterGVN* igvn = phase->is_IterGVN();
3048 if (igvn) igvn->add_users_to_worklist(this);
3049 }
3051 // convenience function
3052 // return false if the init contains any stores already
3053 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3054 InitializeNode* init = initialization();
3055 if (init == NULL || init->is_complete()) return false;
3056 init->remove_extra_zeroes();
3057 // for now, if this allocation has already collected any inits, bail:
3058 if (init->is_non_zero()) return false;
3059 init->set_complete(phase);
3060 return true;
3061 }
3063 void InitializeNode::remove_extra_zeroes() {
3064 if (req() == RawStores) return;
3065 Node* zmem = zero_memory();
3066 uint fill = RawStores;
3067 for (uint i = fill; i < req(); i++) {
3068 Node* n = in(i);
3069 if (n->is_top() || n == zmem) continue; // skip
3070 if (fill < i) set_req(fill, n); // compact
3071 ++fill;
3072 }
3073 // delete any empty spaces created:
3074 while (fill < req()) {
3075 del_req(fill);
3076 }
3077 }
3079 // Helper for remembering which stores go with which offsets.
3080 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3081 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3082 intptr_t offset = -1;
3083 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3084 phase, offset);
3085 if (base == NULL) return -1; // something is dead,
3086 if (offset < 0) return -1; // dead, dead
3087 return offset;
3088 }
3090 // Helper for proving that an initialization expression is
3091 // "simple enough" to be folded into an object initialization.
3092 // Attempts to prove that a store's initial value 'n' can be captured
3093 // within the initialization without creating a vicious cycle, such as:
3094 // { Foo p = new Foo(); p.next = p; }
3095 // True for constants and parameters and small combinations thereof.
3096 bool InitializeNode::detect_init_independence(Node* n,
3097 bool st_is_pinned,
3098 int& count) {
3099 if (n == NULL) return true; // (can this really happen?)
3100 if (n->is_Proj()) n = n->in(0);
3101 if (n == this) return false; // found a cycle
3102 if (n->is_Con()) return true;
3103 if (n->is_Start()) return true; // params, etc., are OK
3104 if (n->is_Root()) return true; // even better
3106 Node* ctl = n->in(0);
3107 if (ctl != NULL && !ctl->is_top()) {
3108 if (ctl->is_Proj()) ctl = ctl->in(0);
3109 if (ctl == this) return false;
3111 // If we already know that the enclosing memory op is pinned right after
3112 // the init, then any control flow that the store has picked up
3113 // must have preceded the init, or else be equal to the init.
3114 // Even after loop optimizations (which might change control edges)
3115 // a store is never pinned *before* the availability of its inputs.
3116 if (!MemNode::all_controls_dominate(n, this))
3117 return false; // failed to prove a good control
3119 }
3121 // Check data edges for possible dependencies on 'this'.
3122 if ((count += 1) > 20) return false; // complexity limit
3123 for (uint i = 1; i < n->req(); i++) {
3124 Node* m = n->in(i);
3125 if (m == NULL || m == n || m->is_top()) continue;
3126 uint first_i = n->find_edge(m);
3127 if (i != first_i) continue; // process duplicate edge just once
3128 if (!detect_init_independence(m, st_is_pinned, count)) {
3129 return false;
3130 }
3131 }
3133 return true;
3134 }
3136 // Here are all the checks a Store must pass before it can be moved into
3137 // an initialization. Returns zero if a check fails.
3138 // On success, returns the (constant) offset to which the store applies,
3139 // within the initialized memory.
3140 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3141 const int FAIL = 0;
3142 if (st->req() != MemNode::ValueIn + 1)
3143 return FAIL; // an inscrutable StoreNode (card mark?)
3144 Node* ctl = st->in(MemNode::Control);
3145 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3146 return FAIL; // must be unconditional after the initialization
3147 Node* mem = st->in(MemNode::Memory);
3148 if (!(mem->is_Proj() && mem->in(0) == this))
3149 return FAIL; // must not be preceded by other stores
3150 Node* adr = st->in(MemNode::Address);
3151 intptr_t offset;
3152 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3153 if (alloc == NULL)
3154 return FAIL; // inscrutable address
3155 if (alloc != allocation())
3156 return FAIL; // wrong allocation! (store needs to float up)
3157 Node* val = st->in(MemNode::ValueIn);
3158 int complexity_count = 0;
3159 if (!detect_init_independence(val, true, complexity_count))
3160 return FAIL; // stored value must be 'simple enough'
3162 // The Store can be captured only if nothing after the allocation
3163 // and before the Store is using the memory location that the store
3164 // overwrites.
3165 bool failed = false;
3166 // If is_complete_with_arraycopy() is true the shape of the graph is
3167 // well defined and is safe so no need for extra checks.
3168 if (!is_complete_with_arraycopy()) {
3169 // We are going to look at each use of the memory state following
3170 // the allocation to make sure nothing reads the memory that the
3171 // Store writes.
3172 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3173 int alias_idx = phase->C->get_alias_index(t_adr);
3174 ResourceMark rm;
3175 Unique_Node_List mems;
3176 mems.push(mem);
3177 Node* unique_merge = NULL;
3178 for (uint next = 0; next < mems.size(); ++next) {
3179 Node *m = mems.at(next);
3180 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3181 Node *n = m->fast_out(j);
3182 if (n->outcnt() == 0) {
3183 continue;
3184 }
3185 if (n == st) {
3186 continue;
3187 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3188 // If the control of this use is different from the control
3189 // of the Store which is right after the InitializeNode then
3190 // this node cannot be between the InitializeNode and the
3191 // Store.
3192 continue;
3193 } else if (n->is_MergeMem()) {
3194 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3195 // We can hit a MergeMemNode (that will likely go away
3196 // later) that is a direct use of the memory state
3197 // following the InitializeNode on the same slice as the
3198 // store node that we'd like to capture. We need to check
3199 // the uses of the MergeMemNode.
3200 mems.push(n);
3201 }
3202 } else if (n->is_Mem()) {
3203 Node* other_adr = n->in(MemNode::Address);
3204 if (other_adr == adr) {
3205 failed = true;
3206 break;
3207 } else {
3208 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3209 if (other_t_adr != NULL) {
3210 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3211 if (other_alias_idx == alias_idx) {
3212 // A load from the same memory slice as the store right
3213 // after the InitializeNode. We check the control of the
3214 // object/array that is loaded from. If it's the same as
3215 // the store control then we cannot capture the store.
3216 assert(!n->is_Store(), "2 stores to same slice on same control?");
3217 Node* base = other_adr;
3218 assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name()));
3219 base = base->in(AddPNode::Base);
3220 if (base != NULL) {
3221 base = base->uncast();
3222 if (base->is_Proj() && base->in(0) == alloc) {
3223 failed = true;
3224 break;
3225 }
3226 }
3227 }
3228 }
3229 }
3230 } else {
3231 failed = true;
3232 break;
3233 }
3234 }
3235 }
3236 }
3237 if (failed) {
3238 if (!can_reshape) {
3239 // We decided we couldn't capture the store during parsing. We
3240 // should try again during the next IGVN once the graph is
3241 // cleaner.
3242 phase->C->record_for_igvn(st);
3243 }
3244 return FAIL;
3245 }
3247 return offset; // success
3248 }
3250 // Find the captured store in(i) which corresponds to the range
3251 // [start..start+size) in the initialized object.
3252 // If there is one, return its index i. If there isn't, return the
3253 // negative of the index where it should be inserted.
3254 // Return 0 if the queried range overlaps an initialization boundary
3255 // or if dead code is encountered.
3256 // If size_in_bytes is zero, do not bother with overlap checks.
3257 int InitializeNode::captured_store_insertion_point(intptr_t start,
3258 int size_in_bytes,
3259 PhaseTransform* phase) {
3260 const int FAIL = 0, MAX_STORE = BytesPerLong;
3262 if (is_complete())
3263 return FAIL; // arraycopy got here first; punt
3265 assert(allocation() != NULL, "must be present");
3267 // no negatives, no header fields:
3268 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3270 // after a certain size, we bail out on tracking all the stores:
3271 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3272 if (start >= ti_limit) return FAIL;
3274 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3275 if (i >= limit) return -(int)i; // not found; here is where to put it
3277 Node* st = in(i);
3278 intptr_t st_off = get_store_offset(st, phase);
3279 if (st_off < 0) {
3280 if (st != zero_memory()) {
3281 return FAIL; // bail out if there is dead garbage
3282 }
3283 } else if (st_off > start) {
3284 // ...we are done, since stores are ordered
3285 if (st_off < start + size_in_bytes) {
3286 return FAIL; // the next store overlaps
3287 }
3288 return -(int)i; // not found; here is where to put it
3289 } else if (st_off < start) {
3290 if (size_in_bytes != 0 &&
3291 start < st_off + MAX_STORE &&
3292 start < st_off + st->as_Store()->memory_size()) {
3293 return FAIL; // the previous store overlaps
3294 }
3295 } else {
3296 if (size_in_bytes != 0 &&
3297 st->as_Store()->memory_size() != size_in_bytes) {
3298 return FAIL; // mismatched store size
3299 }
3300 return i;
3301 }
3303 ++i;
3304 }
3305 }
3307 // Look for a captured store which initializes at the offset 'start'
3308 // with the given size. If there is no such store, and no other
3309 // initialization interferes, then return zero_memory (the memory
3310 // projection of the AllocateNode).
3311 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3312 PhaseTransform* phase) {
3313 assert(stores_are_sane(phase), "");
3314 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3315 if (i == 0) {
3316 return NULL; // something is dead
3317 } else if (i < 0) {
3318 return zero_memory(); // just primordial zero bits here
3319 } else {
3320 Node* st = in(i); // here is the store at this position
3321 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3322 return st;
3323 }
3324 }
3326 // Create, as a raw pointer, an address within my new object at 'offset'.
3327 Node* InitializeNode::make_raw_address(intptr_t offset,
3328 PhaseTransform* phase) {
3329 Node* addr = in(RawAddress);
3330 if (offset != 0) {
3331 Compile* C = phase->C;
3332 addr = phase->transform( new (C) AddPNode(C->top(), addr,
3333 phase->MakeConX(offset)) );
3334 }
3335 return addr;
3336 }
3338 // Clone the given store, converting it into a raw store
3339 // initializing a field or element of my new object.
3340 // Caller is responsible for retiring the original store,
3341 // with subsume_node or the like.
3342 //
3343 // From the example above InitializeNode::InitializeNode,
3344 // here are the old stores to be captured:
3345 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3346 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3347 //
3348 // Here is the changed code; note the extra edges on init:
3349 // alloc = (Allocate ...)
3350 // rawoop = alloc.RawAddress
3351 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3352 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3353 // init = (Initialize alloc.Control alloc.Memory rawoop
3354 // rawstore1 rawstore2)
3355 //
3356 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3357 PhaseTransform* phase, bool can_reshape) {
3358 assert(stores_are_sane(phase), "");
3360 if (start < 0) return NULL;
3361 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3363 Compile* C = phase->C;
3364 int size_in_bytes = st->memory_size();
3365 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3366 if (i == 0) return NULL; // bail out
3367 Node* prev_mem = NULL; // raw memory for the captured store
3368 if (i > 0) {
3369 prev_mem = in(i); // there is a pre-existing store under this one
3370 set_req(i, C->top()); // temporarily disconnect it
3371 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3372 } else {
3373 i = -i; // no pre-existing store
3374 prev_mem = zero_memory(); // a slice of the newly allocated object
3375 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3376 set_req(--i, C->top()); // reuse this edge; it has been folded away
3377 else
3378 ins_req(i, C->top()); // build a new edge
3379 }
3380 Node* new_st = st->clone();
3381 new_st->set_req(MemNode::Control, in(Control));
3382 new_st->set_req(MemNode::Memory, prev_mem);
3383 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3384 new_st = phase->transform(new_st);
3386 // At this point, new_st might have swallowed a pre-existing store
3387 // at the same offset, or perhaps new_st might have disappeared,
3388 // if it redundantly stored the same value (or zero to fresh memory).
3390 // In any case, wire it in:
3391 set_req(i, new_st);
3393 // The caller may now kill the old guy.
3394 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3395 assert(check_st == new_st || check_st == NULL, "must be findable");
3396 assert(!is_complete(), "");
3397 return new_st;
3398 }
3400 static bool store_constant(jlong* tiles, int num_tiles,
3401 intptr_t st_off, int st_size,
3402 jlong con) {
3403 if ((st_off & (st_size-1)) != 0)
3404 return false; // strange store offset (assume size==2**N)
3405 address addr = (address)tiles + st_off;
3406 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3407 switch (st_size) {
3408 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3409 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3410 case sizeof(jint): *(jint*) addr = (jint) con; break;
3411 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3412 default: return false; // strange store size (detect size!=2**N here)
3413 }
3414 return true; // return success to caller
3415 }
3417 // Coalesce subword constants into int constants and possibly
3418 // into long constants. The goal, if the CPU permits,
3419 // is to initialize the object with a small number of 64-bit tiles.
3420 // Also, convert floating-point constants to bit patterns.
3421 // Non-constants are not relevant to this pass.
3422 //
3423 // In terms of the running example on InitializeNode::InitializeNode
3424 // and InitializeNode::capture_store, here is the transformation
3425 // of rawstore1 and rawstore2 into rawstore12:
3426 // alloc = (Allocate ...)
3427 // rawoop = alloc.RawAddress
3428 // tile12 = 0x00010002
3429 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3430 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3431 //
3432 void
3433 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3434 Node* size_in_bytes,
3435 PhaseGVN* phase) {
3436 Compile* C = phase->C;
3438 assert(stores_are_sane(phase), "");
3439 // Note: After this pass, they are not completely sane,
3440 // since there may be some overlaps.
3442 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3444 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3445 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3446 size_limit = MIN2(size_limit, ti_limit);
3447 size_limit = align_size_up(size_limit, BytesPerLong);
3448 int num_tiles = size_limit / BytesPerLong;
3450 // allocate space for the tile map:
3451 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3452 jlong tiles_buf[small_len];
3453 Node* nodes_buf[small_len];
3454 jlong inits_buf[small_len];
3455 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3456 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3457 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3458 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3459 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3460 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3461 // tiles: exact bitwise model of all primitive constants
3462 // nodes: last constant-storing node subsumed into the tiles model
3463 // inits: which bytes (in each tile) are touched by any initializations
3465 //// Pass A: Fill in the tile model with any relevant stores.
3467 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3468 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3469 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3470 Node* zmem = zero_memory(); // initially zero memory state
3471 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3472 Node* st = in(i);
3473 intptr_t st_off = get_store_offset(st, phase);
3475 // Figure out the store's offset and constant value:
3476 if (st_off < header_size) continue; //skip (ignore header)
3477 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3478 int st_size = st->as_Store()->memory_size();
3479 if (st_off + st_size > size_limit) break;
3481 // Record which bytes are touched, whether by constant or not.
3482 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3483 continue; // skip (strange store size)
3485 const Type* val = phase->type(st->in(MemNode::ValueIn));
3486 if (!val->singleton()) continue; //skip (non-con store)
3487 BasicType type = val->basic_type();
3489 jlong con = 0;
3490 switch (type) {
3491 case T_INT: con = val->is_int()->get_con(); break;
3492 case T_LONG: con = val->is_long()->get_con(); break;
3493 case T_FLOAT: con = jint_cast(val->getf()); break;
3494 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3495 default: continue; //skip (odd store type)
3496 }
3498 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3499 st->Opcode() == Op_StoreL) {
3500 continue; // This StoreL is already optimal.
3501 }
3503 // Store down the constant.
3504 store_constant(tiles, num_tiles, st_off, st_size, con);
3506 intptr_t j = st_off >> LogBytesPerLong;
3508 if (type == T_INT && st_size == BytesPerInt
3509 && (st_off & BytesPerInt) == BytesPerInt) {
3510 jlong lcon = tiles[j];
3511 if (!Matcher::isSimpleConstant64(lcon) &&
3512 st->Opcode() == Op_StoreI) {
3513 // This StoreI is already optimal by itself.
3514 jint* intcon = (jint*) &tiles[j];
3515 intcon[1] = 0; // undo the store_constant()
3517 // If the previous store is also optimal by itself, back up and
3518 // undo the action of the previous loop iteration... if we can.
3519 // But if we can't, just let the previous half take care of itself.
3520 st = nodes[j];
3521 st_off -= BytesPerInt;
3522 con = intcon[0];
3523 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3524 assert(st_off >= header_size, "still ignoring header");
3525 assert(get_store_offset(st, phase) == st_off, "must be");
3526 assert(in(i-1) == zmem, "must be");
3527 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3528 assert(con == tcon->is_int()->get_con(), "must be");
3529 // Undo the effects of the previous loop trip, which swallowed st:
3530 intcon[0] = 0; // undo store_constant()
3531 set_req(i-1, st); // undo set_req(i, zmem)
3532 nodes[j] = NULL; // undo nodes[j] = st
3533 --old_subword; // undo ++old_subword
3534 }
3535 continue; // This StoreI is already optimal.
3536 }
3537 }
3539 // This store is not needed.
3540 set_req(i, zmem);
3541 nodes[j] = st; // record for the moment
3542 if (st_size < BytesPerLong) // something has changed
3543 ++old_subword; // includes int/float, but who's counting...
3544 else ++old_long;
3545 }
3547 if ((old_subword + old_long) == 0)
3548 return; // nothing more to do
3550 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3551 // Be sure to insert them before overlapping non-constant stores.
3552 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3553 for (int j = 0; j < num_tiles; j++) {
3554 jlong con = tiles[j];
3555 jlong init = inits[j];
3556 if (con == 0) continue;
3557 jint con0, con1; // split the constant, address-wise
3558 jint init0, init1; // split the init map, address-wise
3559 { union { jlong con; jint intcon[2]; } u;
3560 u.con = con;
3561 con0 = u.intcon[0];
3562 con1 = u.intcon[1];
3563 u.con = init;
3564 init0 = u.intcon[0];
3565 init1 = u.intcon[1];
3566 }
3568 Node* old = nodes[j];
3569 assert(old != NULL, "need the prior store");
3570 intptr_t offset = (j * BytesPerLong);
3572 bool split = !Matcher::isSimpleConstant64(con);
3574 if (offset < header_size) {
3575 assert(offset + BytesPerInt >= header_size, "second int counts");
3576 assert(*(jint*)&tiles[j] == 0, "junk in header");
3577 split = true; // only the second word counts
3578 // Example: int a[] = { 42 ... }
3579 } else if (con0 == 0 && init0 == -1) {
3580 split = true; // first word is covered by full inits
3581 // Example: int a[] = { ... foo(), 42 ... }
3582 } else if (con1 == 0 && init1 == -1) {
3583 split = true; // second word is covered by full inits
3584 // Example: int a[] = { ... 42, foo() ... }
3585 }
3587 // Here's a case where init0 is neither 0 nor -1:
3588 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3589 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3590 // In this case the tile is not split; it is (jlong)42.
3591 // The big tile is stored down, and then the foo() value is inserted.
3592 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3594 Node* ctl = old->in(MemNode::Control);
3595 Node* adr = make_raw_address(offset, phase);
3596 const TypePtr* atp = TypeRawPtr::BOTTOM;
3598 // One or two coalesced stores to plop down.
3599 Node* st[2];
3600 intptr_t off[2];
3601 int nst = 0;
3602 if (!split) {
3603 ++new_long;
3604 off[nst] = offset;
3605 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3606 phase->longcon(con), T_LONG);
3607 } else {
3608 // Omit either if it is a zero.
3609 if (con0 != 0) {
3610 ++new_int;
3611 off[nst] = offset;
3612 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3613 phase->intcon(con0), T_INT);
3614 }
3615 if (con1 != 0) {
3616 ++new_int;
3617 offset += BytesPerInt;
3618 adr = make_raw_address(offset, phase);
3619 off[nst] = offset;
3620 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3621 phase->intcon(con1), T_INT);
3622 }
3623 }
3625 // Insert second store first, then the first before the second.
3626 // Insert each one just before any overlapping non-constant stores.
3627 while (nst > 0) {
3628 Node* st1 = st[--nst];
3629 C->copy_node_notes_to(st1, old);
3630 st1 = phase->transform(st1);
3631 offset = off[nst];
3632 assert(offset >= header_size, "do not smash header");
3633 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3634 guarantee(ins_idx != 0, "must re-insert constant store");
3635 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3636 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3637 set_req(--ins_idx, st1);
3638 else
3639 ins_req(ins_idx, st1);
3640 }
3641 }
3643 if (PrintCompilation && WizardMode)
3644 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3645 old_subword, old_long, new_int, new_long);
3646 if (C->log() != NULL)
3647 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3648 old_subword, old_long, new_int, new_long);
3650 // Clean up any remaining occurrences of zmem:
3651 remove_extra_zeroes();
3652 }
3654 // Explore forward from in(start) to find the first fully initialized
3655 // word, and return its offset. Skip groups of subword stores which
3656 // together initialize full words. If in(start) is itself part of a
3657 // fully initialized word, return the offset of in(start). If there
3658 // are no following full-word stores, or if something is fishy, return
3659 // a negative value.
3660 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3661 int int_map = 0;
3662 intptr_t int_map_off = 0;
3663 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3665 for (uint i = start, limit = req(); i < limit; i++) {
3666 Node* st = in(i);
3668 intptr_t st_off = get_store_offset(st, phase);
3669 if (st_off < 0) break; // return conservative answer
3671 int st_size = st->as_Store()->memory_size();
3672 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3673 return st_off; // we found a complete word init
3674 }
3676 // update the map:
3678 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3679 if (this_int_off != int_map_off) {
3680 // reset the map:
3681 int_map = 0;
3682 int_map_off = this_int_off;
3683 }
3685 int subword_off = st_off - this_int_off;
3686 int_map |= right_n_bits(st_size) << subword_off;
3687 if ((int_map & FULL_MAP) == FULL_MAP) {
3688 return this_int_off; // we found a complete word init
3689 }
3691 // Did this store hit or cross the word boundary?
3692 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3693 if (next_int_off == this_int_off + BytesPerInt) {
3694 // We passed the current int, without fully initializing it.
3695 int_map_off = next_int_off;
3696 int_map >>= BytesPerInt;
3697 } else if (next_int_off > this_int_off + BytesPerInt) {
3698 // We passed the current and next int.
3699 return this_int_off + BytesPerInt;
3700 }
3701 }
3703 return -1;
3704 }
3707 // Called when the associated AllocateNode is expanded into CFG.
3708 // At this point, we may perform additional optimizations.
3709 // Linearize the stores by ascending offset, to make memory
3710 // activity as coherent as possible.
3711 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3712 intptr_t header_size,
3713 Node* size_in_bytes,
3714 PhaseGVN* phase) {
3715 assert(!is_complete(), "not already complete");
3716 assert(stores_are_sane(phase), "");
3717 assert(allocation() != NULL, "must be present");
3719 remove_extra_zeroes();
3721 if (ReduceFieldZeroing || ReduceBulkZeroing)
3722 // reduce instruction count for common initialization patterns
3723 coalesce_subword_stores(header_size, size_in_bytes, phase);
3725 Node* zmem = zero_memory(); // initially zero memory state
3726 Node* inits = zmem; // accumulating a linearized chain of inits
3727 #ifdef ASSERT
3728 intptr_t first_offset = allocation()->minimum_header_size();
3729 intptr_t last_init_off = first_offset; // previous init offset
3730 intptr_t last_init_end = first_offset; // previous init offset+size
3731 intptr_t last_tile_end = first_offset; // previous tile offset+size
3732 #endif
3733 intptr_t zeroes_done = header_size;
3735 bool do_zeroing = true; // we might give up if inits are very sparse
3736 int big_init_gaps = 0; // how many large gaps have we seen?
3738 if (ZeroTLAB) do_zeroing = false;
3739 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3741 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3742 Node* st = in(i);
3743 intptr_t st_off = get_store_offset(st, phase);
3744 if (st_off < 0)
3745 break; // unknown junk in the inits
3746 if (st->in(MemNode::Memory) != zmem)
3747 break; // complicated store chains somehow in list
3749 int st_size = st->as_Store()->memory_size();
3750 intptr_t next_init_off = st_off + st_size;
3752 if (do_zeroing && zeroes_done < next_init_off) {
3753 // See if this store needs a zero before it or under it.
3754 intptr_t zeroes_needed = st_off;
3756 if (st_size < BytesPerInt) {
3757 // Look for subword stores which only partially initialize words.
3758 // If we find some, we must lay down some word-level zeroes first,
3759 // underneath the subword stores.
3760 //
3761 // Examples:
3762 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3763 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3764 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3765 //
3766 // Note: coalesce_subword_stores may have already done this,
3767 // if it was prompted by constant non-zero subword initializers.
3768 // But this case can still arise with non-constant stores.
3770 intptr_t next_full_store = find_next_fullword_store(i, phase);
3772 // In the examples above:
3773 // in(i) p q r s x y z
3774 // st_off 12 13 14 15 12 13 14
3775 // st_size 1 1 1 1 1 1 1
3776 // next_full_s. 12 16 16 16 16 16 16
3777 // z's_done 12 16 16 16 12 16 12
3778 // z's_needed 12 16 16 16 16 16 16
3779 // zsize 0 0 0 0 4 0 4
3780 if (next_full_store < 0) {
3781 // Conservative tack: Zero to end of current word.
3782 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3783 } else {
3784 // Zero to beginning of next fully initialized word.
3785 // Or, don't zero at all, if we are already in that word.
3786 assert(next_full_store >= zeroes_needed, "must go forward");
3787 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3788 zeroes_needed = next_full_store;
3789 }
3790 }
3792 if (zeroes_needed > zeroes_done) {
3793 intptr_t zsize = zeroes_needed - zeroes_done;
3794 // Do some incremental zeroing on rawmem, in parallel with inits.
3795 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3796 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3797 zeroes_done, zeroes_needed,
3798 phase);
3799 zeroes_done = zeroes_needed;
3800 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3801 do_zeroing = false; // leave the hole, next time
3802 }
3803 }
3805 // Collect the store and move on:
3806 st->set_req(MemNode::Memory, inits);
3807 inits = st; // put it on the linearized chain
3808 set_req(i, zmem); // unhook from previous position
3810 if (zeroes_done == st_off)
3811 zeroes_done = next_init_off;
3813 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3815 #ifdef ASSERT
3816 // Various order invariants. Weaker than stores_are_sane because
3817 // a large constant tile can be filled in by smaller non-constant stores.
3818 assert(st_off >= last_init_off, "inits do not reverse");
3819 last_init_off = st_off;
3820 const Type* val = NULL;
3821 if (st_size >= BytesPerInt &&
3822 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3823 (int)val->basic_type() < (int)T_OBJECT) {
3824 assert(st_off >= last_tile_end, "tiles do not overlap");
3825 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3826 last_tile_end = MAX2(last_tile_end, next_init_off);
3827 } else {
3828 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3829 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3830 assert(st_off >= last_init_end, "inits do not overlap");
3831 last_init_end = next_init_off; // it's a non-tile
3832 }
3833 #endif //ASSERT
3834 }
3836 remove_extra_zeroes(); // clear out all the zmems left over
3837 add_req(inits);
3839 if (!ZeroTLAB) {
3840 // If anything remains to be zeroed, zero it all now.
3841 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3842 // if it is the last unused 4 bytes of an instance, forget about it
3843 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3844 if (zeroes_done + BytesPerLong >= size_limit) {
3845 assert(allocation() != NULL, "");
3846 if (allocation()->Opcode() == Op_Allocate) {
3847 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3848 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3849 if (zeroes_done == k->layout_helper())
3850 zeroes_done = size_limit;
3851 }
3852 }
3853 if (zeroes_done < size_limit) {
3854 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3855 zeroes_done, size_in_bytes, phase);
3856 }
3857 }
3859 set_complete(phase);
3860 return rawmem;
3861 }
3864 #ifdef ASSERT
3865 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3866 if (is_complete())
3867 return true; // stores could be anything at this point
3868 assert(allocation() != NULL, "must be present");
3869 intptr_t last_off = allocation()->minimum_header_size();
3870 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3871 Node* st = in(i);
3872 intptr_t st_off = get_store_offset(st, phase);
3873 if (st_off < 0) continue; // ignore dead garbage
3874 if (last_off > st_off) {
3875 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
3876 this->dump(2);
3877 assert(false, "ascending store offsets");
3878 return false;
3879 }
3880 last_off = st_off + st->as_Store()->memory_size();
3881 }
3882 return true;
3883 }
3884 #endif //ASSERT
3889 //============================MergeMemNode=====================================
3890 //
3891 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3892 // contributing store or call operations. Each contributor provides the memory
3893 // state for a particular "alias type" (see Compile::alias_type). For example,
3894 // if a MergeMem has an input X for alias category #6, then any memory reference
3895 // to alias category #6 may use X as its memory state input, as an exact equivalent
3896 // to using the MergeMem as a whole.
3897 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3898 //
3899 // (Here, the <N> notation gives the index of the relevant adr_type.)
3900 //
3901 // In one special case (and more cases in the future), alias categories overlap.
3902 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
3903 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
3904 // it is exactly equivalent to that state W:
3905 // MergeMem(<Bot>: W) <==> W
3906 //
3907 // Usually, the merge has more than one input. In that case, where inputs
3908 // overlap (i.e., one is Bot), the narrower alias type determines the memory
3909 // state for that type, and the wider alias type (Bot) fills in everywhere else:
3910 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
3911 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
3912 //
3913 // A merge can take a "wide" memory state as one of its narrow inputs.
3914 // This simply means that the merge observes out only the relevant parts of
3915 // the wide input. That is, wide memory states arriving at narrow merge inputs
3916 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
3917 //
3918 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
3919 // and that memory slices "leak through":
3920 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
3921 //
3922 // But, in such a cascade, repeated memory slices can "block the leak":
3923 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
3924 //
3925 // In the last example, Y is not part of the combined memory state of the
3926 // outermost MergeMem. The system must, of course, prevent unschedulable
3927 // memory states from arising, so you can be sure that the state Y is somehow
3928 // a precursor to state Y'.
3929 //
3930 //
3931 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
3932 // of each MergeMemNode array are exactly the numerical alias indexes, including
3933 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
3934 // Compile::alias_type (and kin) produce and manage these indexes.
3935 //
3936 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
3937 // (Note that this provides quick access to the top node inside MergeMem methods,
3938 // without the need to reach out via TLS to Compile::current.)
3939 //
3940 // As a consequence of what was just described, a MergeMem that represents a full
3941 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
3942 // containing all alias categories.
3943 //
3944 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
3945 //
3946 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
3947 // a memory state for the alias type <N>, or else the top node, meaning that
3948 // there is no particular input for that alias type. Note that the length of
3949 // a MergeMem is variable, and may be extended at any time to accommodate new
3950 // memory states at larger alias indexes. When merges grow, they are of course
3951 // filled with "top" in the unused in() positions.
3952 //
3953 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
3954 // (Top was chosen because it works smoothly with passes like GCM.)
3955 //
3956 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
3957 // the type of random VM bits like TLS references.) Since it is always the
3958 // first non-Bot memory slice, some low-level loops use it to initialize an
3959 // index variable: for (i = AliasIdxRaw; i < req(); i++).
3960 //
3961 //
3962 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
3963 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
3964 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
3965 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
3966 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
3967 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
3968 //
3969 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
3970 // really that different from the other memory inputs. An abbreviation called
3971 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
3972 //
3973 //
3974 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
3975 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
3976 // that "emerges though" the base memory will be marked as excluding the alias types
3977 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
3978 //
3979 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
3980 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
3981 //
3982 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
3983 // (It is currently unimplemented.) As you can see, the resulting merge is
3984 // actually a disjoint union of memory states, rather than an overlay.
3985 //
3987 //------------------------------MergeMemNode-----------------------------------
3988 Node* MergeMemNode::make_empty_memory() {
3989 Node* empty_memory = (Node*) Compile::current()->top();
3990 assert(empty_memory->is_top(), "correct sentinel identity");
3991 return empty_memory;
3992 }
3994 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3995 init_class_id(Class_MergeMem);
3996 // all inputs are nullified in Node::Node(int)
3997 // set_input(0, NULL); // no control input
3999 // Initialize the edges uniformly to top, for starters.
4000 Node* empty_mem = make_empty_memory();
4001 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4002 init_req(i,empty_mem);
4003 }
4004 assert(empty_memory() == empty_mem, "");
4006 if( new_base != NULL && new_base->is_MergeMem() ) {
4007 MergeMemNode* mdef = new_base->as_MergeMem();
4008 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4009 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4010 mms.set_memory(mms.memory2());
4011 }
4012 assert(base_memory() == mdef->base_memory(), "");
4013 } else {
4014 set_base_memory(new_base);
4015 }
4016 }
4018 // Make a new, untransformed MergeMem with the same base as 'mem'.
4019 // If mem is itself a MergeMem, populate the result with the same edges.
4020 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
4021 return new(C) MergeMemNode(mem);
4022 }
4024 //------------------------------cmp--------------------------------------------
4025 uint MergeMemNode::hash() const { return NO_HASH; }
4026 uint MergeMemNode::cmp( const Node &n ) const {
4027 return (&n == this); // Always fail except on self
4028 }
4030 //------------------------------Identity---------------------------------------
4031 Node* MergeMemNode::Identity(PhaseTransform *phase) {
4032 // Identity if this merge point does not record any interesting memory
4033 // disambiguations.
4034 Node* base_mem = base_memory();
4035 Node* empty_mem = empty_memory();
4036 if (base_mem != empty_mem) { // Memory path is not dead?
4037 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4038 Node* mem = in(i);
4039 if (mem != empty_mem && mem != base_mem) {
4040 return this; // Many memory splits; no change
4041 }
4042 }
4043 }
4044 return base_mem; // No memory splits; ID on the one true input
4045 }
4047 //------------------------------Ideal------------------------------------------
4048 // This method is invoked recursively on chains of MergeMem nodes
4049 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4050 // Remove chain'd MergeMems
4051 //
4052 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4053 // relative to the "in(Bot)". Since we are patching both at the same time,
4054 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4055 // but rewrite each "in(i)" relative to the new "in(Bot)".
4056 Node *progress = NULL;
4059 Node* old_base = base_memory();
4060 Node* empty_mem = empty_memory();
4061 if (old_base == empty_mem)
4062 return NULL; // Dead memory path.
4064 MergeMemNode* old_mbase;
4065 if (old_base != NULL && old_base->is_MergeMem())
4066 old_mbase = old_base->as_MergeMem();
4067 else
4068 old_mbase = NULL;
4069 Node* new_base = old_base;
4071 // simplify stacked MergeMems in base memory
4072 if (old_mbase) new_base = old_mbase->base_memory();
4074 // the base memory might contribute new slices beyond my req()
4075 if (old_mbase) grow_to_match(old_mbase);
4077 // Look carefully at the base node if it is a phi.
4078 PhiNode* phi_base;
4079 if (new_base != NULL && new_base->is_Phi())
4080 phi_base = new_base->as_Phi();
4081 else
4082 phi_base = NULL;
4084 Node* phi_reg = NULL;
4085 uint phi_len = (uint)-1;
4086 if (phi_base != NULL && !phi_base->is_copy()) {
4087 // do not examine phi if degraded to a copy
4088 phi_reg = phi_base->region();
4089 phi_len = phi_base->req();
4090 // see if the phi is unfinished
4091 for (uint i = 1; i < phi_len; i++) {
4092 if (phi_base->in(i) == NULL) {
4093 // incomplete phi; do not look at it yet!
4094 phi_reg = NULL;
4095 phi_len = (uint)-1;
4096 break;
4097 }
4098 }
4099 }
4101 // Note: We do not call verify_sparse on entry, because inputs
4102 // can normalize to the base_memory via subsume_node or similar
4103 // mechanisms. This method repairs that damage.
4105 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4107 // Look at each slice.
4108 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4109 Node* old_in = in(i);
4110 // calculate the old memory value
4111 Node* old_mem = old_in;
4112 if (old_mem == empty_mem) old_mem = old_base;
4113 assert(old_mem == memory_at(i), "");
4115 // maybe update (reslice) the old memory value
4117 // simplify stacked MergeMems
4118 Node* new_mem = old_mem;
4119 MergeMemNode* old_mmem;
4120 if (old_mem != NULL && old_mem->is_MergeMem())
4121 old_mmem = old_mem->as_MergeMem();
4122 else
4123 old_mmem = NULL;
4124 if (old_mmem == this) {
4125 // This can happen if loops break up and safepoints disappear.
4126 // A merge of BotPtr (default) with a RawPtr memory derived from a
4127 // safepoint can be rewritten to a merge of the same BotPtr with
4128 // the BotPtr phi coming into the loop. If that phi disappears
4129 // also, we can end up with a self-loop of the mergemem.
4130 // In general, if loops degenerate and memory effects disappear,
4131 // a mergemem can be left looking at itself. This simply means
4132 // that the mergemem's default should be used, since there is
4133 // no longer any apparent effect on this slice.
4134 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4135 // from start. Update the input to TOP.
4136 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4137 }
4138 else if (old_mmem != NULL) {
4139 new_mem = old_mmem->memory_at(i);
4140 }
4141 // else preceding memory was not a MergeMem
4143 // replace equivalent phis (unfortunately, they do not GVN together)
4144 if (new_mem != NULL && new_mem != new_base &&
4145 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4146 if (new_mem->is_Phi()) {
4147 PhiNode* phi_mem = new_mem->as_Phi();
4148 for (uint i = 1; i < phi_len; i++) {
4149 if (phi_base->in(i) != phi_mem->in(i)) {
4150 phi_mem = NULL;
4151 break;
4152 }
4153 }
4154 if (phi_mem != NULL) {
4155 // equivalent phi nodes; revert to the def
4156 new_mem = new_base;
4157 }
4158 }
4159 }
4161 // maybe store down a new value
4162 Node* new_in = new_mem;
4163 if (new_in == new_base) new_in = empty_mem;
4165 if (new_in != old_in) {
4166 // Warning: Do not combine this "if" with the previous "if"
4167 // A memory slice might have be be rewritten even if it is semantically
4168 // unchanged, if the base_memory value has changed.
4169 set_req(i, new_in);
4170 progress = this; // Report progress
4171 }
4172 }
4174 if (new_base != old_base) {
4175 set_req(Compile::AliasIdxBot, new_base);
4176 // Don't use set_base_memory(new_base), because we need to update du.
4177 assert(base_memory() == new_base, "");
4178 progress = this;
4179 }
4181 if( base_memory() == this ) {
4182 // a self cycle indicates this memory path is dead
4183 set_req(Compile::AliasIdxBot, empty_mem);
4184 }
4186 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4187 // Recursion must occur after the self cycle check above
4188 if( base_memory()->is_MergeMem() ) {
4189 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4190 Node *m = phase->transform(new_mbase); // Rollup any cycles
4191 if( m != NULL && (m->is_top() ||
4192 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
4193 // propagate rollup of dead cycle to self
4194 set_req(Compile::AliasIdxBot, empty_mem);
4195 }
4196 }
4198 if( base_memory() == empty_mem ) {
4199 progress = this;
4200 // Cut inputs during Parse phase only.
4201 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4202 if( !can_reshape ) {
4203 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4204 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4205 }
4206 }
4207 }
4209 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4210 // Check if PhiNode::Ideal's "Split phis through memory merges"
4211 // transform should be attempted. Look for this->phi->this cycle.
4212 uint merge_width = req();
4213 if (merge_width > Compile::AliasIdxRaw) {
4214 PhiNode* phi = base_memory()->as_Phi();
4215 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4216 if (phi->in(i) == this) {
4217 phase->is_IterGVN()->_worklist.push(phi);
4218 break;
4219 }
4220 }
4221 }
4222 }
4224 assert(progress || verify_sparse(), "please, no dups of base");
4225 return progress;
4226 }
4228 //-------------------------set_base_memory-------------------------------------
4229 void MergeMemNode::set_base_memory(Node *new_base) {
4230 Node* empty_mem = empty_memory();
4231 set_req(Compile::AliasIdxBot, new_base);
4232 assert(memory_at(req()) == new_base, "must set default memory");
4233 // Clear out other occurrences of new_base:
4234 if (new_base != empty_mem) {
4235 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4236 if (in(i) == new_base) set_req(i, empty_mem);
4237 }
4238 }
4239 }
4241 //------------------------------out_RegMask------------------------------------
4242 const RegMask &MergeMemNode::out_RegMask() const {
4243 return RegMask::Empty;
4244 }
4246 //------------------------------dump_spec--------------------------------------
4247 #ifndef PRODUCT
4248 void MergeMemNode::dump_spec(outputStream *st) const {
4249 st->print(" {");
4250 Node* base_mem = base_memory();
4251 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4252 Node* mem = memory_at(i);
4253 if (mem == base_mem) { st->print(" -"); continue; }
4254 st->print( " N%d:", mem->_idx );
4255 Compile::current()->get_adr_type(i)->dump_on(st);
4256 }
4257 st->print(" }");
4258 }
4259 #endif // !PRODUCT
4262 #ifdef ASSERT
4263 static bool might_be_same(Node* a, Node* b) {
4264 if (a == b) return true;
4265 if (!(a->is_Phi() || b->is_Phi())) return false;
4266 // phis shift around during optimization
4267 return true; // pretty stupid...
4268 }
4270 // verify a narrow slice (either incoming or outgoing)
4271 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4272 if (!VerifyAliases) return; // don't bother to verify unless requested
4273 if (is_error_reported()) return; // muzzle asserts when debugging an error
4274 if (Node::in_dump()) return; // muzzle asserts when printing
4275 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4276 assert(n != NULL, "");
4277 // Elide intervening MergeMem's
4278 while (n->is_MergeMem()) {
4279 n = n->as_MergeMem()->memory_at(alias_idx);
4280 }
4281 Compile* C = Compile::current();
4282 const TypePtr* n_adr_type = n->adr_type();
4283 if (n == m->empty_memory()) {
4284 // Implicit copy of base_memory()
4285 } else if (n_adr_type != TypePtr::BOTTOM) {
4286 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4287 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4288 } else {
4289 // A few places like make_runtime_call "know" that VM calls are narrow,
4290 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4291 bool expected_wide_mem = false;
4292 if (n == m->base_memory()) {
4293 expected_wide_mem = true;
4294 } else if (alias_idx == Compile::AliasIdxRaw ||
4295 n == m->memory_at(Compile::AliasIdxRaw)) {
4296 expected_wide_mem = true;
4297 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4298 // memory can "leak through" calls on channels that
4299 // are write-once. Allow this also.
4300 expected_wide_mem = true;
4301 }
4302 assert(expected_wide_mem, "expected narrow slice replacement");
4303 }
4304 }
4305 #else // !ASSERT
4306 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
4307 #endif
4310 //-----------------------------memory_at---------------------------------------
4311 Node* MergeMemNode::memory_at(uint alias_idx) const {
4312 assert(alias_idx >= Compile::AliasIdxRaw ||
4313 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4314 "must avoid base_memory and AliasIdxTop");
4316 // Otherwise, it is a narrow slice.
4317 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4318 Compile *C = Compile::current();
4319 if (is_empty_memory(n)) {
4320 // the array is sparse; empty slots are the "top" node
4321 n = base_memory();
4322 assert(Node::in_dump()
4323 || n == NULL || n->bottom_type() == Type::TOP
4324 || n->adr_type() == NULL // address is TOP
4325 || n->adr_type() == TypePtr::BOTTOM
4326 || n->adr_type() == TypeRawPtr::BOTTOM
4327 || Compile::current()->AliasLevel() == 0,
4328 "must be a wide memory");
4329 // AliasLevel == 0 if we are organizing the memory states manually.
4330 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4331 } else {
4332 // make sure the stored slice is sane
4333 #ifdef ASSERT
4334 if (is_error_reported() || Node::in_dump()) {
4335 } else if (might_be_same(n, base_memory())) {
4336 // Give it a pass: It is a mostly harmless repetition of the base.
4337 // This can arise normally from node subsumption during optimization.
4338 } else {
4339 verify_memory_slice(this, alias_idx, n);
4340 }
4341 #endif
4342 }
4343 return n;
4344 }
4346 //---------------------------set_memory_at-------------------------------------
4347 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4348 verify_memory_slice(this, alias_idx, n);
4349 Node* empty_mem = empty_memory();
4350 if (n == base_memory()) n = empty_mem; // collapse default
4351 uint need_req = alias_idx+1;
4352 if (req() < need_req) {
4353 if (n == empty_mem) return; // already the default, so do not grow me
4354 // grow the sparse array
4355 do {
4356 add_req(empty_mem);
4357 } while (req() < need_req);
4358 }
4359 set_req( alias_idx, n );
4360 }
4364 //--------------------------iteration_setup------------------------------------
4365 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4366 if (other != NULL) {
4367 grow_to_match(other);
4368 // invariant: the finite support of mm2 is within mm->req()
4369 #ifdef ASSERT
4370 for (uint i = req(); i < other->req(); i++) {
4371 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4372 }
4373 #endif
4374 }
4375 // Replace spurious copies of base_memory by top.
4376 Node* base_mem = base_memory();
4377 if (base_mem != NULL && !base_mem->is_top()) {
4378 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4379 if (in(i) == base_mem)
4380 set_req(i, empty_memory());
4381 }
4382 }
4383 }
4385 //---------------------------grow_to_match-------------------------------------
4386 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4387 Node* empty_mem = empty_memory();
4388 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4389 // look for the finite support of the other memory
4390 for (uint i = other->req(); --i >= req(); ) {
4391 if (other->in(i) != empty_mem) {
4392 uint new_len = i+1;
4393 while (req() < new_len) add_req(empty_mem);
4394 break;
4395 }
4396 }
4397 }
4399 //---------------------------verify_sparse-------------------------------------
4400 #ifndef PRODUCT
4401 bool MergeMemNode::verify_sparse() const {
4402 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4403 Node* base_mem = base_memory();
4404 // The following can happen in degenerate cases, since empty==top.
4405 if (is_empty_memory(base_mem)) return true;
4406 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4407 assert(in(i) != NULL, "sane slice");
4408 if (in(i) == base_mem) return false; // should have been the sentinel value!
4409 }
4410 return true;
4411 }
4413 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4414 Node* n;
4415 n = mm->in(idx);
4416 if (mem == n) return true; // might be empty_memory()
4417 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4418 if (mem == n) return true;
4419 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4420 if (mem == n) return true;
4421 if (n == NULL) break;
4422 }
4423 return false;
4424 }
4425 #endif // !PRODUCT