1 /*
2 * Copyright (c) 1997, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/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/convertnode.hpp"
35 #include "opto/loopnode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/matcher.hpp"
38 #include "opto/memnode.hpp"
39 #include "opto/mulnode.hpp"
40 #include "opto/narrowptrnode.hpp"
41 #include "opto/phaseX.hpp"
42 #include "opto/regmask.hpp"
43
44 // Portions of code courtesy of Clifford Click
45
46 // Optimization - Graph Style
47
48 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
49
50 //=============================================================================
51 uint MemNode::size_of() const { return sizeof(*this); }
52
53 const TypePtr *MemNode::adr_type() const {
54 Node* adr = in(Address);
55 const TypePtr* cross_check = NULL;
56 DEBUG_ONLY(cross_check = _adr_type);
57 return calculate_adr_type(adr->bottom_type(), cross_check);
58 }
59
60 #ifndef PRODUCT
61 void MemNode::dump_spec(outputStream *st) const {
62 if (in(Address) == NULL) return; // node is dead
63 #ifndef ASSERT
64 // fake the missing field
65 const TypePtr* _adr_type = NULL;
66 if (in(Address) != NULL)
67 _adr_type = in(Address)->bottom_type()->isa_ptr();
68 #endif
69 dump_adr_type(this, _adr_type, st);
70
71 Compile* C = Compile::current();
72 if( C->alias_type(_adr_type)->is_volatile() )
73 st->print(" Volatile!");
74 }
75
76 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
77 st->print(" @");
78 if (adr_type == NULL) {
79 st->print("NULL");
80 } else {
81 adr_type->dump_on(st);
82 Compile* C = Compile::current();
83 Compile::AliasType* atp = NULL;
84 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
85 if (atp == NULL)
86 st->print(", idx=?\?;");
87 else if (atp->index() == Compile::AliasIdxBot)
88 st->print(", idx=Bot;");
89 else if (atp->index() == Compile::AliasIdxTop)
90 st->print(", idx=Top;");
91 else if (atp->index() == Compile::AliasIdxRaw)
92 st->print(", idx=Raw;");
93 else {
94 ciField* field = atp->field();
95 if (field) {
96 st->print(", name=");
97 field->print_name_on(st);
98 }
99 st->print(", idx=%d;", atp->index());
100 }
101 }
102 }
103
104 extern void print_alias_types();
105
106 #endif
107
108 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
109 assert((t_oop != NULL), "sanity");
110 bool is_instance = t_oop->is_known_instance_field();
111 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
112 (load != NULL) && load->is_Load() &&
113 (phase->is_IterGVN() != NULL);
114 if (!(is_instance || is_boxed_value_load))
115 return mchain; // don't try to optimize non-instance types
116 uint instance_id = t_oop->instance_id();
117 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
118 Node *prev = NULL;
119 Node *result = mchain;
120 while (prev != result) {
121 prev = result;
122 if (result == start_mem)
123 break; // hit one of our sentinels
124 // skip over a call which does not affect this memory slice
125 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
126 Node *proj_in = result->in(0);
127 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
128 break; // hit one of our sentinels
129 } else if (proj_in->is_Call()) {
130 CallNode *call = proj_in->as_Call();
131 if (!call->may_modify(t_oop, phase)) { // returns false for instances
132 result = call->in(TypeFunc::Memory);
133 }
134 } else if (proj_in->is_Initialize()) {
135 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
136 // Stop if this is the initialization for the object instance which
137 // which contains this memory slice, otherwise skip over it.
138 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
139 break;
140 }
141 if (is_instance) {
142 result = proj_in->in(TypeFunc::Memory);
143 } else if (is_boxed_value_load) {
144 Node* klass = alloc->in(AllocateNode::KlassNode);
145 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
146 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
147 result = proj_in->in(TypeFunc::Memory); // not related allocation
148 }
149 }
150 } else if (proj_in->is_MemBar()) {
151 result = proj_in->in(TypeFunc::Memory);
152 } else {
153 assert(false, "unexpected projection");
154 }
155 } else if (result->is_ClearArray()) {
156 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
157 // Can not bypass initialization of the instance
158 // we are looking for.
159 break;
160 }
161 // Otherwise skip it (the call updated 'result' value).
162 } else if (result->is_MergeMem()) {
163 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
164 }
165 }
166 return result;
167 }
168
169 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
170 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
171 if (t_oop == NULL)
172 return mchain; // don't try to optimize non-oop types
173 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
174 bool is_instance = t_oop->is_known_instance_field();
175 PhaseIterGVN *igvn = phase->is_IterGVN();
176 if (is_instance && igvn != NULL && result->is_Phi()) {
177 PhiNode *mphi = result->as_Phi();
178 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
179 const TypePtr *t = mphi->adr_type();
180 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
181 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
182 t->is_oopptr()->cast_to_exactness(true)
183 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
184 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
185 // clone the Phi with our address type
186 result = mphi->split_out_instance(t_adr, igvn);
187 } else {
188 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
189 }
190 }
191 return result;
192 }
193
194 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
195 uint alias_idx = phase->C->get_alias_index(tp);
196 Node *mem = mmem;
197 #ifdef ASSERT
198 {
199 // Check that current type is consistent with the alias index used during graph construction
200 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
201 bool consistent = adr_check == NULL || adr_check->empty() ||
202 phase->C->must_alias(adr_check, alias_idx );
203 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
204 if( !consistent && adr_check != NULL && !adr_check->empty() &&
205 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
206 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
207 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
208 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
209 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
210 // don't assert if it is dead code.
211 consistent = true;
212 }
213 if( !consistent ) {
214 st->print("alias_idx==%d, adr_check==", alias_idx);
215 if( adr_check == NULL ) {
216 st->print("NULL");
217 } else {
218 adr_check->dump();
219 }
220 st->cr();
221 print_alias_types();
222 assert(consistent, "adr_check must match alias idx");
223 }
224 }
225 #endif
226 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
227 // means an array I have not precisely typed yet. Do not do any
228 // alias stuff with it any time soon.
229 const TypeOopPtr *toop = tp->isa_oopptr();
230 if( tp->base() != Type::AnyPtr &&
231 !(toop &&
232 toop->klass() != NULL &&
233 toop->klass()->is_java_lang_Object() &&
234 toop->offset() == Type::OffsetBot) ) {
235 // compress paths and change unreachable cycles to TOP
236 // If not, we can update the input infinitely along a MergeMem cycle
237 // Equivalent code in PhiNode::Ideal
238 Node* m = phase->transform(mmem);
239 // If transformed to a MergeMem, get the desired slice
240 // Otherwise the returned node represents memory for every slice
241 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
242 // Update input if it is progress over what we have now
243 }
244 return mem;
245 }
246
247 //--------------------------Ideal_common---------------------------------------
248 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
249 // Unhook non-raw memories from complete (macro-expanded) initializations.
250 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
251 // If our control input is a dead region, kill all below the region
252 Node *ctl = in(MemNode::Control);
253 if (ctl && remove_dead_region(phase, can_reshape))
254 return this;
255 ctl = in(MemNode::Control);
256 // Don't bother trying to transform a dead node
257 if (ctl && ctl->is_top()) return NodeSentinel;
258
259 PhaseIterGVN *igvn = phase->is_IterGVN();
260 // Wait if control on the worklist.
261 if (ctl && can_reshape && igvn != NULL) {
262 Node* bol = NULL;
263 Node* cmp = NULL;
264 if (ctl->in(0)->is_If()) {
265 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
266 bol = ctl->in(0)->in(1);
267 if (bol->is_Bool())
268 cmp = ctl->in(0)->in(1)->in(1);
269 }
270 if (igvn->_worklist.member(ctl) ||
271 (bol != NULL && igvn->_worklist.member(bol)) ||
272 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
273 // This control path may be dead.
274 // Delay this memory node transformation until the control is processed.
275 phase->is_IterGVN()->_worklist.push(this);
276 return NodeSentinel; // caller will return NULL
277 }
278 }
279 // Ignore if memory is dead, or self-loop
280 Node *mem = in(MemNode::Memory);
281 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
282 assert(mem != this, "dead loop in MemNode::Ideal");
283
284 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
285 // This memory slice may be dead.
286 // Delay this mem node transformation until the memory is processed.
287 phase->is_IterGVN()->_worklist.push(this);
288 return NodeSentinel; // caller will return NULL
289 }
290
291 Node *address = in(MemNode::Address);
292 const Type *t_adr = phase->type(address);
293 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
294
295 if (can_reshape && igvn != NULL &&
296 (igvn->_worklist.member(address) ||
297 igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) {
298 // The address's base and type may change when the address is processed.
299 // Delay this mem node transformation until the address is processed.
300 phase->is_IterGVN()->_worklist.push(this);
301 return NodeSentinel; // caller will return NULL
302 }
303
304 // Do NOT remove or optimize the next lines: ensure a new alias index
305 // is allocated for an oop pointer type before Escape Analysis.
306 // Note: C++ will not remove it since the call has side effect.
307 if (t_adr->isa_oopptr()) {
308 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
309 }
310
311 Node* base = NULL;
312 if (address->is_AddP()) {
313 base = address->in(AddPNode::Base);
314 }
315 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
316 !t_adr->isa_rawptr()) {
317 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
318 // Skip this node optimization if its address has TOP base.
319 return NodeSentinel; // caller will return NULL
320 }
321
322 // Avoid independent memory operations
323 Node* old_mem = mem;
324
325 // The code which unhooks non-raw memories from complete (macro-expanded)
326 // initializations was removed. After macro-expansion all stores catched
327 // by Initialize node became raw stores and there is no information
328 // which memory slices they modify. So it is unsafe to move any memory
329 // operation above these stores. Also in most cases hooked non-raw memories
330 // were already unhooked by using information from detect_ptr_independence()
331 // and find_previous_store().
332
333 if (mem->is_MergeMem()) {
334 MergeMemNode* mmem = mem->as_MergeMem();
335 const TypePtr *tp = t_adr->is_ptr();
336
337 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
338 }
339
340 if (mem != old_mem) {
341 set_req(MemNode::Memory, mem);
342 if (can_reshape && old_mem->outcnt() == 0) {
343 igvn->_worklist.push(old_mem);
344 }
345 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
346 return this;
347 }
348
349 // let the subclass continue analyzing...
350 return NULL;
351 }
352
353 // Helper function for proving some simple control dominations.
354 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
355 // Already assumes that 'dom' is available at 'sub', and that 'sub'
356 // is not a constant (dominated by the method's StartNode).
357 // Used by MemNode::find_previous_store to prove that the
358 // control input of a memory operation predates (dominates)
359 // an allocation it wants to look past.
360 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
361 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
362 return false; // Conservative answer for dead code
363
364 // Check 'dom'. Skip Proj and CatchProj nodes.
365 dom = dom->find_exact_control(dom);
366 if (dom == NULL || dom->is_top())
367 return false; // Conservative answer for dead code
368
369 if (dom == sub) {
370 // For the case when, for example, 'sub' is Initialize and the original
371 // 'dom' is Proj node of the 'sub'.
372 return false;
373 }
374
375 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
376 return true;
377
378 // 'dom' dominates 'sub' if its control edge and control edges
379 // of all its inputs dominate or equal to sub's control edge.
380
381 // Currently 'sub' is either Allocate, Initialize or Start nodes.
382 // Or Region for the check in LoadNode::Ideal();
383 // 'sub' should have sub->in(0) != NULL.
384 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
385 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
386
387 // Get control edge of 'sub'.
388 Node* orig_sub = sub;
389 sub = sub->find_exact_control(sub->in(0));
390 if (sub == NULL || sub->is_top())
391 return false; // Conservative answer for dead code
392
393 assert(sub->is_CFG(), "expecting control");
394
395 if (sub == dom)
396 return true;
397
398 if (sub->is_Start() || sub->is_Root())
399 return false;
400
401 {
402 // Check all control edges of 'dom'.
403
404 ResourceMark rm;
405 Arena* arena = Thread::current()->resource_area();
406 Node_List nlist(arena);
407 Unique_Node_List dom_list(arena);
408
409 dom_list.push(dom);
410 bool only_dominating_controls = false;
411
412 for (uint next = 0; next < dom_list.size(); next++) {
413 Node* n = dom_list.at(next);
414 if (n == orig_sub)
415 return false; // One of dom's inputs dominated by sub.
416 if (!n->is_CFG() && n->pinned()) {
417 // Check only own control edge for pinned non-control nodes.
418 n = n->find_exact_control(n->in(0));
419 if (n == NULL || n->is_top())
420 return false; // Conservative answer for dead code
421 assert(n->is_CFG(), "expecting control");
422 dom_list.push(n);
423 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
424 only_dominating_controls = true;
425 } else if (n->is_CFG()) {
426 if (n->dominates(sub, nlist))
427 only_dominating_controls = true;
428 else
429 return false;
430 } else {
431 // First, own control edge.
432 Node* m = n->find_exact_control(n->in(0));
433 if (m != NULL) {
434 if (m->is_top())
435 return false; // Conservative answer for dead code
436 dom_list.push(m);
437 }
438 // Now, the rest of edges.
439 uint cnt = n->req();
440 for (uint i = 1; i < cnt; i++) {
441 m = n->find_exact_control(n->in(i));
442 if (m == NULL || m->is_top())
443 continue;
444 dom_list.push(m);
445 }
446 }
447 }
448 return only_dominating_controls;
449 }
450 }
451
452 //---------------------detect_ptr_independence---------------------------------
453 // Used by MemNode::find_previous_store to prove that two base
454 // pointers are never equal.
455 // The pointers are accompanied by their associated allocations,
456 // if any, which have been previously discovered by the caller.
457 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
458 Node* p2, AllocateNode* a2,
459 PhaseTransform* phase) {
460 // Attempt to prove that these two pointers cannot be aliased.
461 // They may both manifestly be allocations, and they should differ.
462 // Or, if they are not both allocations, they can be distinct constants.
463 // Otherwise, one is an allocation and the other a pre-existing value.
464 if (a1 == NULL && a2 == NULL) { // neither an allocation
465 return (p1 != p2) && p1->is_Con() && p2->is_Con();
466 } else if (a1 != NULL && a2 != NULL) { // both allocations
467 return (a1 != a2);
468 } else if (a1 != NULL) { // one allocation a1
469 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
470 return all_controls_dominate(p2, a1);
471 } else { //(a2 != NULL) // one allocation a2
472 return all_controls_dominate(p1, a2);
473 }
474 return false;
475 }
476
477
478 // The logic for reordering loads and stores uses four steps:
479 // (a) Walk carefully past stores and initializations which we
480 // can prove are independent of this load.
481 // (b) Observe that the next memory state makes an exact match
482 // with self (load or store), and locate the relevant store.
483 // (c) Ensure that, if we were to wire self directly to the store,
484 // the optimizer would fold it up somehow.
485 // (d) Do the rewiring, and return, depending on some other part of
486 // the optimizer to fold up the load.
487 // This routine handles steps (a) and (b). Steps (c) and (d) are
488 // specific to loads and stores, so they are handled by the callers.
489 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
490 //
491 Node* MemNode::find_previous_store(PhaseTransform* phase) {
492 Node* ctrl = in(MemNode::Control);
493 Node* adr = in(MemNode::Address);
494 intptr_t offset = 0;
495 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
496 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
497
498 if (offset == Type::OffsetBot)
499 return NULL; // cannot unalias unless there are precise offsets
500
501 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
502
503 intptr_t size_in_bytes = memory_size();
504
505 Node* mem = in(MemNode::Memory); // start searching here...
506
507 int cnt = 50; // Cycle limiter
508 for (;;) { // While we can dance past unrelated stores...
509 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
510
511 if (mem->is_Store()) {
512 Node* st_adr = mem->in(MemNode::Address);
513 intptr_t st_offset = 0;
514 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
515 if (st_base == NULL)
516 break; // inscrutable pointer
517 if (st_offset != offset && st_offset != Type::OffsetBot) {
518 const int MAX_STORE = BytesPerLong;
519 if (st_offset >= offset + size_in_bytes ||
520 st_offset <= offset - MAX_STORE ||
521 st_offset <= offset - mem->as_Store()->memory_size()) {
522 // Success: The offsets are provably independent.
523 // (You may ask, why not just test st_offset != offset and be done?
524 // The answer is that stores of different sizes can co-exist
525 // in the same sequence of RawMem effects. We sometimes initialize
526 // a whole 'tile' of array elements with a single jint or jlong.)
527 mem = mem->in(MemNode::Memory);
528 continue; // (a) advance through independent store memory
529 }
530 }
531 if (st_base != base &&
532 detect_ptr_independence(base, alloc,
533 st_base,
534 AllocateNode::Ideal_allocation(st_base, phase),
535 phase)) {
536 // Success: The bases are provably independent.
537 mem = mem->in(MemNode::Memory);
538 continue; // (a) advance through independent store memory
539 }
540
541 // (b) At this point, if the bases or offsets do not agree, we lose,
542 // since we have not managed to prove 'this' and 'mem' independent.
543 if (st_base == base && st_offset == offset) {
544 return mem; // let caller handle steps (c), (d)
545 }
546
547 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
548 InitializeNode* st_init = mem->in(0)->as_Initialize();
549 AllocateNode* st_alloc = st_init->allocation();
550 if (st_alloc == NULL)
551 break; // something degenerated
552 bool known_identical = false;
553 bool known_independent = false;
554 if (alloc == st_alloc)
555 known_identical = true;
556 else if (alloc != NULL)
557 known_independent = true;
558 else if (all_controls_dominate(this, st_alloc))
559 known_independent = true;
560
561 if (known_independent) {
562 // The bases are provably independent: Either they are
563 // manifestly distinct allocations, or else the control
564 // of this load dominates the store's allocation.
565 int alias_idx = phase->C->get_alias_index(adr_type());
566 if (alias_idx == Compile::AliasIdxRaw) {
567 mem = st_alloc->in(TypeFunc::Memory);
568 } else {
569 mem = st_init->memory(alias_idx);
570 }
571 continue; // (a) advance through independent store memory
572 }
573
574 // (b) at this point, if we are not looking at a store initializing
575 // the same allocation we are loading from, we lose.
576 if (known_identical) {
577 // From caller, can_see_stored_value will consult find_captured_store.
578 return mem; // let caller handle steps (c), (d)
579 }
580
581 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
582 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
583 if (mem->is_Proj() && mem->in(0)->is_Call()) {
584 CallNode *call = mem->in(0)->as_Call();
585 if (!call->may_modify(addr_t, phase)) {
586 mem = call->in(TypeFunc::Memory);
587 continue; // (a) advance through independent call memory
588 }
589 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
590 mem = mem->in(0)->in(TypeFunc::Memory);
591 continue; // (a) advance through independent MemBar memory
592 } else if (mem->is_ClearArray()) {
593 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
594 // (the call updated 'mem' value)
595 continue; // (a) advance through independent allocation memory
596 } else {
597 // Can not bypass initialization of the instance
598 // we are looking for.
599 return mem;
600 }
601 } else if (mem->is_MergeMem()) {
602 int alias_idx = phase->C->get_alias_index(adr_type());
603 mem = mem->as_MergeMem()->memory_at(alias_idx);
604 continue; // (a) advance through independent MergeMem memory
605 }
606 }
607
608 // Unless there is an explicit 'continue', we must bail out here,
609 // because 'mem' is an inscrutable memory state (e.g., a call).
610 break;
611 }
612
613 return NULL; // bail out
614 }
615
616 //----------------------calculate_adr_type-------------------------------------
617 // Helper function. Notices when the given type of address hits top or bottom.
618 // Also, asserts a cross-check of the type against the expected address type.
619 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
620 if (t == Type::TOP) return NULL; // does not touch memory any more?
621 #ifdef PRODUCT
622 cross_check = NULL;
623 #else
624 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
625 #endif
626 const TypePtr* tp = t->isa_ptr();
627 if (tp == NULL) {
628 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
629 return TypePtr::BOTTOM; // touches lots of memory
630 } else {
631 #ifdef ASSERT
632 // %%%% [phh] We don't check the alias index if cross_check is
633 // TypeRawPtr::BOTTOM. Needs to be investigated.
634 if (cross_check != NULL &&
635 cross_check != TypePtr::BOTTOM &&
636 cross_check != TypeRawPtr::BOTTOM) {
637 // Recheck the alias index, to see if it has changed (due to a bug).
638 Compile* C = Compile::current();
639 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
640 "must stay in the original alias category");
641 // The type of the address must be contained in the adr_type,
642 // disregarding "null"-ness.
643 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
644 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
645 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
646 "real address must not escape from expected memory type");
647 }
648 #endif
649 return tp;
650 }
651 }
652
653 //------------------------adr_phi_is_loop_invariant----------------------------
654 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
655 // loop is loop invariant. Make a quick traversal of Phi and associated
656 // CastPP nodes, looking to see if they are a closed group within the loop.
657 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
658 // The idea is that the phi-nest must boil down to only CastPP nodes
659 // with the same data. This implies that any path into the loop already
660 // includes such a CastPP, and so the original cast, whatever its input,
661 // must be covered by an equivalent cast, with an earlier control input.
662 ResourceMark rm;
663
664 // The loop entry input of the phi should be the unique dominating
665 // node for every Phi/CastPP in the loop.
666 Unique_Node_List closure;
667 closure.push(adr_phi->in(LoopNode::EntryControl));
668
669 // Add the phi node and the cast to the worklist.
670 Unique_Node_List worklist;
671 worklist.push(adr_phi);
672 if( cast != NULL ){
673 if( !cast->is_ConstraintCast() ) return false;
674 worklist.push(cast);
675 }
676
677 // Begin recursive walk of phi nodes.
678 while( worklist.size() ){
679 // Take a node off the worklist
680 Node *n = worklist.pop();
681 if( !closure.member(n) ){
682 // Add it to the closure.
683 closure.push(n);
684 // Make a sanity check to ensure we don't waste too much time here.
685 if( closure.size() > 20) return false;
686 // This node is OK if:
687 // - it is a cast of an identical value
688 // - or it is a phi node (then we add its inputs to the worklist)
689 // Otherwise, the node is not OK, and we presume the cast is not invariant
690 if( n->is_ConstraintCast() ){
691 worklist.push(n->in(1));
692 } else if( n->is_Phi() ) {
693 for( uint i = 1; i < n->req(); i++ ) {
694 worklist.push(n->in(i));
695 }
696 } else {
697 return false;
698 }
699 }
700 }
701
702 // Quit when the worklist is empty, and we've found no offending nodes.
703 return true;
704 }
705
706 //------------------------------Ideal_DU_postCCP-------------------------------
707 // Find any cast-away of null-ness and keep its control. Null cast-aways are
708 // going away in this pass and we need to make this memory op depend on the
709 // gating null check.
710 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
711 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
712 }
713
714 // I tried to leave the CastPP's in. This makes the graph more accurate in
715 // some sense; we get to keep around the knowledge that an oop is not-null
716 // after some test. Alas, the CastPP's interfere with GVN (some values are
717 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
718 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
719 // some of the more trivial cases in the optimizer. Removing more useless
720 // Phi's started allowing Loads to illegally float above null checks. I gave
721 // up on this approach. CNC 10/20/2000
722 // This static method may be called not from MemNode (EncodePNode calls it).
723 // Only the control edge of the node 'n' might be updated.
724 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
725 Node *skipped_cast = NULL;
726 // Need a null check? Regular static accesses do not because they are
727 // from constant addresses. Array ops are gated by the range check (which
728 // always includes a NULL check). Just check field ops.
729 if( n->in(MemNode::Control) == NULL ) {
730 // Scan upwards for the highest location we can place this memory op.
731 while( true ) {
732 switch( adr->Opcode() ) {
733
734 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
735 adr = adr->in(AddPNode::Base);
736 continue;
737
738 case Op_DecodeN: // No change to NULL-ness, so peek thru
739 case Op_DecodeNKlass:
740 adr = adr->in(1);
741 continue;
742
743 case Op_EncodeP:
744 case Op_EncodePKlass:
745 // EncodeP node's control edge could be set by this method
746 // when EncodeP node depends on CastPP node.
747 //
748 // Use its control edge for memory op because EncodeP may go away
749 // later when it is folded with following or preceding DecodeN node.
750 if (adr->in(0) == NULL) {
751 // Keep looking for cast nodes.
752 adr = adr->in(1);
753 continue;
754 }
755 ccp->hash_delete(n);
756 n->set_req(MemNode::Control, adr->in(0));
757 ccp->hash_insert(n);
758 return n;
759
760 case Op_CastPP:
761 // If the CastPP is useless, just peek on through it.
762 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
763 // Remember the cast that we've peeked though. If we peek
764 // through more than one, then we end up remembering the highest
765 // one, that is, if in a loop, the one closest to the top.
766 skipped_cast = adr;
767 adr = adr->in(1);
768 continue;
769 }
770 // CastPP is going away in this pass! We need this memory op to be
771 // control-dependent on the test that is guarding the CastPP.
772 ccp->hash_delete(n);
773 n->set_req(MemNode::Control, adr->in(0));
774 ccp->hash_insert(n);
775 return n;
776
777 case Op_Phi:
778 // Attempt to float above a Phi to some dominating point.
779 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
780 // If we've already peeked through a Cast (which could have set the
781 // control), we can't float above a Phi, because the skipped Cast
782 // may not be loop invariant.
783 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
784 adr = adr->in(1);
785 continue;
786 }
787 }
788
789 // Intentional fallthrough!
790
791 // No obvious dominating point. The mem op is pinned below the Phi
792 // by the Phi itself. If the Phi goes away (no true value is merged)
793 // then the mem op can float, but not indefinitely. It must be pinned
794 // behind the controls leading to the Phi.
795 case Op_CheckCastPP:
796 // These usually stick around to change address type, however a
797 // useless one can be elided and we still need to pick up a control edge
798 if (adr->in(0) == NULL) {
799 // This CheckCastPP node has NO control and is likely useless. But we
800 // need check further up the ancestor chain for a control input to keep
801 // the node in place. 4959717.
802 skipped_cast = adr;
803 adr = adr->in(1);
804 continue;
805 }
806 ccp->hash_delete(n);
807 n->set_req(MemNode::Control, adr->in(0));
808 ccp->hash_insert(n);
809 return n;
810
811 // List of "safe" opcodes; those that implicitly block the memory
812 // op below any null check.
813 case Op_CastX2P: // no null checks on native pointers
814 case Op_Parm: // 'this' pointer is not null
815 case Op_LoadP: // Loading from within a klass
816 case Op_LoadN: // Loading from within a klass
817 case Op_LoadKlass: // Loading from within a klass
818 case Op_LoadNKlass: // Loading from within a klass
819 case Op_ConP: // Loading from a klass
820 case Op_ConN: // Loading from a klass
821 case Op_ConNKlass: // Loading from a klass
822 case Op_CreateEx: // Sucking up the guts of an exception oop
823 case Op_Con: // Reading from TLS
824 case Op_CMoveP: // CMoveP is pinned
825 case Op_CMoveN: // CMoveN is pinned
826 break; // No progress
827
828 case Op_Proj: // Direct call to an allocation routine
829 case Op_SCMemProj: // Memory state from store conditional ops
830 #ifdef ASSERT
831 {
832 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
833 const Node* call = adr->in(0);
834 if (call->is_CallJava()) {
835 const CallJavaNode* call_java = call->as_CallJava();
836 const TypeTuple *r = call_java->tf()->range();
837 assert(r->cnt() > TypeFunc::Parms, "must return value");
838 const Type* ret_type = r->field_at(TypeFunc::Parms);
839 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
840 // We further presume that this is one of
841 // new_instance_Java, new_array_Java, or
842 // the like, but do not assert for this.
843 } else if (call->is_Allocate()) {
844 // similar case to new_instance_Java, etc.
845 } else if (!call->is_CallLeaf()) {
846 // Projections from fetch_oop (OSR) are allowed as well.
847 ShouldNotReachHere();
848 }
849 }
850 #endif
851 break;
852 default:
853 ShouldNotReachHere();
854 }
855 break;
856 }
857 }
858
859 return NULL; // No progress
860 }
861
862
863 //=============================================================================
864 uint LoadNode::size_of() const { return sizeof(*this); }
865 uint LoadNode::cmp( const Node &n ) const
866 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
867 const Type *LoadNode::bottom_type() const { return _type; }
868 uint LoadNode::ideal_reg() const {
869 return _type->ideal_reg();
870 }
871
872 #ifndef PRODUCT
873 void LoadNode::dump_spec(outputStream *st) const {
874 MemNode::dump_spec(st);
875 if( !Verbose && !WizardMode ) {
876 // standard dump does this in Verbose and WizardMode
877 st->print(" #"); _type->dump_on(st);
878 }
879 }
880 #endif
881
882 #ifdef ASSERT
883 //----------------------------is_immutable_value-------------------------------
884 // Helper function to allow a raw load without control edge for some cases
885 bool LoadNode::is_immutable_value(Node* adr) {
886 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
887 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
888 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
889 in_bytes(JavaThread::osthread_offset())));
890 }
891 #endif
892
893 //----------------------------LoadNode::make-----------------------------------
894 // Polymorphic factory method:
895 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo) {
896 Compile* C = gvn.C;
897
898 // sanity check the alias category against the created node type
899 assert(!(adr_type->isa_oopptr() &&
900 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
901 "use LoadKlassNode instead");
902 assert(!(adr_type->isa_aryptr() &&
903 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
904 "use LoadRangeNode instead");
905 // Check control edge of raw loads
906 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
907 // oop will be recorded in oop map if load crosses safepoint
908 rt->isa_oopptr() || is_immutable_value(adr),
909 "raw memory operations should have control edge");
910 switch (bt) {
911 case T_BOOLEAN: return new (C) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo);
912 case T_BYTE: return new (C) LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo);
913 case T_INT: return new (C) LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo);
914 case T_CHAR: return new (C) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo);
915 case T_SHORT: return new (C) LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo);
916 case T_LONG: return new (C) LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo);
917 case T_FLOAT: return new (C) LoadFNode (ctl, mem, adr, adr_type, rt, mo);
918 case T_DOUBLE: return new (C) LoadDNode (ctl, mem, adr, adr_type, rt, mo);
919 case T_ADDRESS: return new (C) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo);
920 case T_OBJECT:
921 #ifdef _LP64
922 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
923 Node* load = gvn.transform(new (C) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo));
924 return new (C) DecodeNNode(load, load->bottom_type()->make_ptr());
925 } else
926 #endif
927 {
928 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
929 return new (C) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr(), mo);
930 }
931 }
932 ShouldNotReachHere();
933 return (LoadNode*)NULL;
934 }
935
936 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo) {
937 bool require_atomic = true;
938 return new (C) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, require_atomic);
939 }
940
941
942
943
944 //------------------------------hash-------------------------------------------
945 uint LoadNode::hash() const {
946 // unroll addition of interesting fields
947 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
948 }
949
950 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
951 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
952 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
953 bool is_stable_ary = FoldStableValues &&
954 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
955 tp->isa_aryptr()->is_stable();
956
957 return (eliminate_boxing && non_volatile) || is_stable_ary;
958 }
959
960 return false;
961 }
962
963 //---------------------------can_see_stored_value------------------------------
964 // This routine exists to make sure this set of tests is done the same
965 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
966 // will change the graph shape in a way which makes memory alive twice at the
967 // same time (uses the Oracle model of aliasing), then some
968 // LoadXNode::Identity will fold things back to the equivalence-class model
969 // of aliasing.
970 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
971 Node* ld_adr = in(MemNode::Address);
972 intptr_t ld_off = 0;
973 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
974 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
975 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
976 // This is more general than load from boxing objects.
977 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
978 uint alias_idx = atp->index();
979 bool final = !atp->is_rewritable();
980 Node* result = NULL;
981 Node* current = st;
982 // Skip through chains of MemBarNodes checking the MergeMems for
983 // new states for the slice of this load. Stop once any other
984 // kind of node is encountered. Loads from final memory can skip
985 // through any kind of MemBar but normal loads shouldn't skip
986 // through MemBarAcquire since the could allow them to move out of
987 // a synchronized region.
988 while (current->is_Proj()) {
989 int opc = current->in(0)->Opcode();
990 if ((final && (opc == Op_MemBarAcquire ||
991 opc == Op_MemBarAcquireLock ||
992 opc == Op_LoadFence)) ||
993 opc == Op_MemBarRelease ||
994 opc == Op_StoreFence ||
995 opc == Op_MemBarReleaseLock ||
996 opc == Op_MemBarCPUOrder) {
997 Node* mem = current->in(0)->in(TypeFunc::Memory);
998 if (mem->is_MergeMem()) {
999 MergeMemNode* merge = mem->as_MergeMem();
1000 Node* new_st = merge->memory_at(alias_idx);
1001 if (new_st == merge->base_memory()) {
1002 // Keep searching
1003 current = new_st;
1004 continue;
1005 }
1006 // Save the new memory state for the slice and fall through
1007 // to exit.
1008 result = new_st;
1009 }
1010 }
1011 break;
1012 }
1013 if (result != NULL) {
1014 st = result;
1015 }
1016 }
1017
1018 // Loop around twice in the case Load -> Initialize -> Store.
1019 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1020 for (int trip = 0; trip <= 1; trip++) {
1021
1022 if (st->is_Store()) {
1023 Node* st_adr = st->in(MemNode::Address);
1024 if (!phase->eqv(st_adr, ld_adr)) {
1025 // Try harder before giving up... Match raw and non-raw pointers.
1026 intptr_t st_off = 0;
1027 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
1028 if (alloc == NULL) return NULL;
1029 if (alloc != ld_alloc) return NULL;
1030 if (ld_off != st_off) return NULL;
1031 // At this point we have proven something like this setup:
1032 // A = Allocate(...)
1033 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
1034 // S = StoreQ(, AddP(, A.Parm , #Off), V)
1035 // (Actually, we haven't yet proven the Q's are the same.)
1036 // In other words, we are loading from a casted version of
1037 // the same pointer-and-offset that we stored to.
1038 // Thus, we are able to replace L by V.
1039 }
1040 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1041 if (store_Opcode() != st->Opcode())
1042 return NULL;
1043 return st->in(MemNode::ValueIn);
1044 }
1045
1046 // A load from a freshly-created object always returns zero.
1047 // (This can happen after LoadNode::Ideal resets the load's memory input
1048 // to find_captured_store, which returned InitializeNode::zero_memory.)
1049 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1050 (st->in(0) == ld_alloc) &&
1051 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1052 // return a zero value for the load's basic type
1053 // (This is one of the few places where a generic PhaseTransform
1054 // can create new nodes. Think of it as lazily manifesting
1055 // virtually pre-existing constants.)
1056 return phase->zerocon(memory_type());
1057 }
1058
1059 // A load from an initialization barrier can match a captured store.
1060 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1061 InitializeNode* init = st->in(0)->as_Initialize();
1062 AllocateNode* alloc = init->allocation();
1063 if ((alloc != NULL) && (alloc == ld_alloc)) {
1064 // examine a captured store value
1065 st = init->find_captured_store(ld_off, memory_size(), phase);
1066 if (st != NULL)
1067 continue; // take one more trip around
1068 }
1069 }
1070
1071 // Load boxed value from result of valueOf() call is input parameter.
1072 if (this->is_Load() && ld_adr->is_AddP() &&
1073 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1074 intptr_t ignore = 0;
1075 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1076 if (base != NULL && base->is_Proj() &&
1077 base->as_Proj()->_con == TypeFunc::Parms &&
1078 base->in(0)->is_CallStaticJava() &&
1079 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1080 return base->in(0)->in(TypeFunc::Parms);
1081 }
1082 }
1083
1084 break;
1085 }
1086
1087 return NULL;
1088 }
1089
1090 //----------------------is_instance_field_load_with_local_phi------------------
1091 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1092 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1093 in(Address)->is_AddP() ) {
1094 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1095 // Only instances and boxed values.
1096 if( t_oop != NULL &&
1097 (t_oop->is_ptr_to_boxed_value() ||
1098 t_oop->is_known_instance_field()) &&
1099 t_oop->offset() != Type::OffsetBot &&
1100 t_oop->offset() != Type::OffsetTop) {
1101 return true;
1102 }
1103 }
1104 return false;
1105 }
1106
1107 //------------------------------Identity---------------------------------------
1108 // Loads are identity if previous store is to same address
1109 Node *LoadNode::Identity( PhaseTransform *phase ) {
1110 // If the previous store-maker is the right kind of Store, and the store is
1111 // to the same address, then we are equal to the value stored.
1112 Node* mem = in(Memory);
1113 Node* value = can_see_stored_value(mem, phase);
1114 if( value ) {
1115 // byte, short & char stores truncate naturally.
1116 // A load has to load the truncated value which requires
1117 // some sort of masking operation and that requires an
1118 // Ideal call instead of an Identity call.
1119 if (memory_size() < BytesPerInt) {
1120 // If the input to the store does not fit with the load's result type,
1121 // it must be truncated via an Ideal call.
1122 if (!phase->type(value)->higher_equal(phase->type(this)))
1123 return this;
1124 }
1125 // (This works even when value is a Con, but LoadNode::Value
1126 // usually runs first, producing the singleton type of the Con.)
1127 return value;
1128 }
1129
1130 // Search for an existing data phi which was generated before for the same
1131 // instance's field to avoid infinite generation of phis in a loop.
1132 Node *region = mem->in(0);
1133 if (is_instance_field_load_with_local_phi(region)) {
1134 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1135 int this_index = phase->C->get_alias_index(addr_t);
1136 int this_offset = addr_t->offset();
1137 int this_iid = addr_t->instance_id();
1138 if (!addr_t->is_known_instance() &&
1139 addr_t->is_ptr_to_boxed_value()) {
1140 // Use _idx of address base (could be Phi node) for boxed values.
1141 intptr_t ignore = 0;
1142 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1143 this_iid = base->_idx;
1144 }
1145 const Type* this_type = bottom_type();
1146 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1147 Node* phi = region->fast_out(i);
1148 if (phi->is_Phi() && phi != mem &&
1149 phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) {
1150 return phi;
1151 }
1152 }
1153 }
1154
1155 return this;
1156 }
1157
1158 // We're loading from an object which has autobox behaviour.
1159 // If this object is result of a valueOf call we'll have a phi
1160 // merging a newly allocated object and a load from the cache.
1161 // We want to replace this load with the original incoming
1162 // argument to the valueOf call.
1163 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1164 assert(phase->C->eliminate_boxing(), "sanity");
1165 intptr_t ignore = 0;
1166 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1167 if ((base == NULL) || base->is_Phi()) {
1168 // Push the loads from the phi that comes from valueOf up
1169 // through it to allow elimination of the loads and the recovery
1170 // of the original value. It is done in split_through_phi().
1171 return NULL;
1172 } else if (base->is_Load() ||
1173 base->is_DecodeN() && base->in(1)->is_Load()) {
1174 // Eliminate the load of boxed value for integer types from the cache
1175 // array by deriving the value from the index into the array.
1176 // Capture the offset of the load and then reverse the computation.
1177
1178 // Get LoadN node which loads a boxing object from 'cache' array.
1179 if (base->is_DecodeN()) {
1180 base = base->in(1);
1181 }
1182 if (!base->in(Address)->is_AddP()) {
1183 return NULL; // Complex address
1184 }
1185 AddPNode* address = base->in(Address)->as_AddP();
1186 Node* cache_base = address->in(AddPNode::Base);
1187 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1188 // Get ConP node which is static 'cache' field.
1189 cache_base = cache_base->in(1);
1190 }
1191 if ((cache_base != NULL) && cache_base->is_Con()) {
1192 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1193 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1194 Node* elements[4];
1195 int shift = exact_log2(type2aelembytes(T_OBJECT));
1196 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1197 if ((count > 0) && elements[0]->is_Con() &&
1198 ((count == 1) ||
1199 (count == 2) && elements[1]->Opcode() == Op_LShiftX &&
1200 elements[1]->in(2) == phase->intcon(shift))) {
1201 ciObjArray* array = base_type->const_oop()->as_obj_array();
1202 // Fetch the box object cache[0] at the base of the array and get its value
1203 ciInstance* box = array->obj_at(0)->as_instance();
1204 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1205 assert(ik->is_box_klass(), "sanity");
1206 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1207 if (ik->nof_nonstatic_fields() == 1) {
1208 // This should be true nonstatic_field_at requires calling
1209 // nof_nonstatic_fields so check it anyway
1210 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1211 BasicType bt = c.basic_type();
1212 // Only integer types have boxing cache.
1213 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1214 bt == T_BYTE || bt == T_SHORT ||
1215 bt == T_INT || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt)));
1216 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1217 if (cache_low != (int)cache_low) {
1218 return NULL; // should not happen since cache is array indexed by value
1219 }
1220 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1221 if (offset != (int)offset) {
1222 return NULL; // should not happen since cache is array indexed by value
1223 }
1224 // Add up all the offsets making of the address of the load
1225 Node* result = elements[0];
1226 for (int i = 1; i < count; i++) {
1227 result = phase->transform(new (phase->C) AddXNode(result, elements[i]));
1228 }
1229 // Remove the constant offset from the address and then
1230 result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-(int)offset)));
1231 // remove the scaling of the offset to recover the original index.
1232 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1233 // Peel the shift off directly but wrap it in a dummy node
1234 // since Ideal can't return existing nodes
1235 result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0));
1236 } else if (result->is_Add() && result->in(2)->is_Con() &&
1237 result->in(1)->Opcode() == Op_LShiftX &&
1238 result->in(1)->in(2) == phase->intcon(shift)) {
1239 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1240 // but for boxing cache access we know that X<<Z will not overflow
1241 // (there is range check) so we do this optimizatrion by hand here.
1242 Node* add_con = new (phase->C) RShiftXNode(result->in(2), phase->intcon(shift));
1243 result = new (phase->C) AddXNode(result->in(1)->in(1), phase->transform(add_con));
1244 } else {
1245 result = new (phase->C) RShiftXNode(result, phase->intcon(shift));
1246 }
1247 #ifdef _LP64
1248 if (bt != T_LONG) {
1249 result = new (phase->C) ConvL2INode(phase->transform(result));
1250 }
1251 #else
1252 if (bt == T_LONG) {
1253 result = new (phase->C) ConvI2LNode(phase->transform(result));
1254 }
1255 #endif
1256 return result;
1257 }
1258 }
1259 }
1260 }
1261 }
1262 return NULL;
1263 }
1264
1265 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1266 Node* region = phi->in(0);
1267 if (region == NULL) {
1268 return false; // Wait stable graph
1269 }
1270 uint cnt = phi->req();
1271 for (uint i = 1; i < cnt; i++) {
1272 Node* rc = region->in(i);
1273 if (rc == NULL || phase->type(rc) == Type::TOP)
1274 return false; // Wait stable graph
1275 Node* in = phi->in(i);
1276 if (in == NULL || phase->type(in) == Type::TOP)
1277 return false; // Wait stable graph
1278 }
1279 return true;
1280 }
1281 //------------------------------split_through_phi------------------------------
1282 // Split instance or boxed field load through Phi.
1283 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1284 Node* mem = in(Memory);
1285 Node* address = in(Address);
1286 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1287
1288 assert((t_oop != NULL) &&
1289 (t_oop->is_known_instance_field() ||
1290 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1291
1292 Compile* C = phase->C;
1293 intptr_t ignore = 0;
1294 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1295 bool base_is_phi = (base != NULL) && base->is_Phi();
1296 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1297 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1298 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1299
1300 if (!((mem->is_Phi() || base_is_phi) &&
1301 (load_boxed_values || t_oop->is_known_instance_field()))) {
1302 return NULL; // memory is not Phi
1303 }
1304
1305 if (mem->is_Phi()) {
1306 if (!stable_phi(mem->as_Phi(), phase)) {
1307 return NULL; // Wait stable graph
1308 }
1309 uint cnt = mem->req();
1310 // Check for loop invariant memory.
1311 if (cnt == 3) {
1312 for (uint i = 1; i < cnt; i++) {
1313 Node* in = mem->in(i);
1314 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1315 if (m == mem) {
1316 set_req(Memory, mem->in(cnt - i));
1317 return this; // made change
1318 }
1319 }
1320 }
1321 }
1322 if (base_is_phi) {
1323 if (!stable_phi(base->as_Phi(), phase)) {
1324 return NULL; // Wait stable graph
1325 }
1326 uint cnt = base->req();
1327 // Check for loop invariant memory.
1328 if (cnt == 3) {
1329 for (uint i = 1; i < cnt; i++) {
1330 if (base->in(i) == base) {
1331 return NULL; // Wait stable graph
1332 }
1333 }
1334 }
1335 }
1336
1337 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1338
1339 // Split through Phi (see original code in loopopts.cpp).
1340 assert(C->have_alias_type(t_oop), "instance should have alias type");
1341
1342 // Do nothing here if Identity will find a value
1343 // (to avoid infinite chain of value phis generation).
1344 if (!phase->eqv(this, this->Identity(phase)))
1345 return NULL;
1346
1347 // Select Region to split through.
1348 Node* region;
1349 if (!base_is_phi) {
1350 assert(mem->is_Phi(), "sanity");
1351 region = mem->in(0);
1352 // Skip if the region dominates some control edge of the address.
1353 if (!MemNode::all_controls_dominate(address, region))
1354 return NULL;
1355 } else if (!mem->is_Phi()) {
1356 assert(base_is_phi, "sanity");
1357 region = base->in(0);
1358 // Skip if the region dominates some control edge of the memory.
1359 if (!MemNode::all_controls_dominate(mem, region))
1360 return NULL;
1361 } else if (base->in(0) != mem->in(0)) {
1362 assert(base_is_phi && mem->is_Phi(), "sanity");
1363 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1364 region = base->in(0);
1365 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1366 region = mem->in(0);
1367 } else {
1368 return NULL; // complex graph
1369 }
1370 } else {
1371 assert(base->in(0) == mem->in(0), "sanity");
1372 region = mem->in(0);
1373 }
1374
1375 const Type* this_type = this->bottom_type();
1376 int this_index = C->get_alias_index(t_oop);
1377 int this_offset = t_oop->offset();
1378 int this_iid = t_oop->instance_id();
1379 if (!t_oop->is_known_instance() && load_boxed_values) {
1380 // Use _idx of address base for boxed values.
1381 this_iid = base->_idx;
1382 }
1383 PhaseIterGVN* igvn = phase->is_IterGVN();
1384 Node* phi = new (C) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1385 for (uint i = 1; i < region->req(); i++) {
1386 Node* x;
1387 Node* the_clone = NULL;
1388 if (region->in(i) == C->top()) {
1389 x = C->top(); // Dead path? Use a dead data op
1390 } else {
1391 x = this->clone(); // Else clone up the data op
1392 the_clone = x; // Remember for possible deletion.
1393 // Alter data node to use pre-phi inputs
1394 if (this->in(0) == region) {
1395 x->set_req(0, region->in(i));
1396 } else {
1397 x->set_req(0, NULL);
1398 }
1399 if (mem->is_Phi() && (mem->in(0) == region)) {
1400 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1401 }
1402 if (address->is_Phi() && address->in(0) == region) {
1403 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1404 }
1405 if (base_is_phi && (base->in(0) == region)) {
1406 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1407 Node* adr_x = phase->transform(new (C) AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1408 x->set_req(Address, adr_x);
1409 }
1410 }
1411 // Check for a 'win' on some paths
1412 const Type *t = x->Value(igvn);
1413
1414 bool singleton = t->singleton();
1415
1416 // See comments in PhaseIdealLoop::split_thru_phi().
1417 if (singleton && t == Type::TOP) {
1418 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1419 }
1420
1421 if (singleton) {
1422 x = igvn->makecon(t);
1423 } else {
1424 // We now call Identity to try to simplify the cloned node.
1425 // Note that some Identity methods call phase->type(this).
1426 // Make sure that the type array is big enough for
1427 // our new node, even though we may throw the node away.
1428 // (This tweaking with igvn only works because x is a new node.)
1429 igvn->set_type(x, t);
1430 // If x is a TypeNode, capture any more-precise type permanently into Node
1431 // otherwise it will be not updated during igvn->transform since
1432 // igvn->type(x) is set to x->Value() already.
1433 x->raise_bottom_type(t);
1434 Node *y = x->Identity(igvn);
1435 if (y != x) {
1436 x = y;
1437 } else {
1438 y = igvn->hash_find_insert(x);
1439 if (y) {
1440 x = y;
1441 } else {
1442 // Else x is a new node we are keeping
1443 // We do not need register_new_node_with_optimizer
1444 // because set_type has already been called.
1445 igvn->_worklist.push(x);
1446 }
1447 }
1448 }
1449 if (x != the_clone && the_clone != NULL) {
1450 igvn->remove_dead_node(the_clone);
1451 }
1452 phi->set_req(i, x);
1453 }
1454 // Record Phi
1455 igvn->register_new_node_with_optimizer(phi);
1456 return phi;
1457 }
1458
1459 //------------------------------Ideal------------------------------------------
1460 // If the load is from Field memory and the pointer is non-null, we can
1461 // zero out the control input.
1462 // If the offset is constant and the base is an object allocation,
1463 // try to hook me up to the exact initializing store.
1464 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1465 Node* p = MemNode::Ideal_common(phase, can_reshape);
1466 if (p) return (p == NodeSentinel) ? NULL : p;
1467
1468 Node* ctrl = in(MemNode::Control);
1469 Node* address = in(MemNode::Address);
1470
1471 // Skip up past a SafePoint control. Cannot do this for Stores because
1472 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1473 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1474 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1475 ctrl = ctrl->in(0);
1476 set_req(MemNode::Control,ctrl);
1477 }
1478
1479 intptr_t ignore = 0;
1480 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1481 if (base != NULL
1482 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1483 // Check for useless control edge in some common special cases
1484 if (in(MemNode::Control) != NULL
1485 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1486 && all_controls_dominate(base, phase->C->start())) {
1487 // A method-invariant, non-null address (constant or 'this' argument).
1488 set_req(MemNode::Control, NULL);
1489 }
1490 }
1491
1492 Node* mem = in(MemNode::Memory);
1493 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1494
1495 if (can_reshape && (addr_t != NULL)) {
1496 // try to optimize our memory input
1497 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1498 if (opt_mem != mem) {
1499 set_req(MemNode::Memory, opt_mem);
1500 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1501 return this;
1502 }
1503 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1504 if ((t_oop != NULL) &&
1505 (t_oop->is_known_instance_field() ||
1506 t_oop->is_ptr_to_boxed_value())) {
1507 PhaseIterGVN *igvn = phase->is_IterGVN();
1508 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1509 // Delay this transformation until memory Phi is processed.
1510 phase->is_IterGVN()->_worklist.push(this);
1511 return NULL;
1512 }
1513 // Split instance field load through Phi.
1514 Node* result = split_through_phi(phase);
1515 if (result != NULL) return result;
1516
1517 if (t_oop->is_ptr_to_boxed_value()) {
1518 Node* result = eliminate_autobox(phase);
1519 if (result != NULL) return result;
1520 }
1521 }
1522 }
1523
1524 // Check for prior store with a different base or offset; make Load
1525 // independent. Skip through any number of them. Bail out if the stores
1526 // are in an endless dead cycle and report no progress. This is a key
1527 // transform for Reflection. However, if after skipping through the Stores
1528 // we can't then fold up against a prior store do NOT do the transform as
1529 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1530 // array memory alive twice: once for the hoisted Load and again after the
1531 // bypassed Store. This situation only works if EVERYBODY who does
1532 // anti-dependence work knows how to bypass. I.e. we need all
1533 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1534 // the alias index stuff. So instead, peek through Stores and IFF we can
1535 // fold up, do so.
1536 Node* prev_mem = find_previous_store(phase);
1537 // Steps (a), (b): Walk past independent stores to find an exact match.
1538 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1539 // (c) See if we can fold up on the spot, but don't fold up here.
1540 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1541 // just return a prior value, which is done by Identity calls.
1542 if (can_see_stored_value(prev_mem, phase)) {
1543 // Make ready for step (d):
1544 set_req(MemNode::Memory, prev_mem);
1545 return this;
1546 }
1547 }
1548
1549 return NULL; // No further progress
1550 }
1551
1552 // Helper to recognize certain Klass fields which are invariant across
1553 // some group of array types (e.g., int[] or all T[] where T < Object).
1554 const Type*
1555 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1556 ciKlass* klass) const {
1557 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1558 // The field is Klass::_modifier_flags. Return its (constant) value.
1559 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1560 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1561 return TypeInt::make(klass->modifier_flags());
1562 }
1563 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1564 // The field is Klass::_access_flags. Return its (constant) value.
1565 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1566 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1567 return TypeInt::make(klass->access_flags());
1568 }
1569 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1570 // The field is Klass::_layout_helper. Return its constant value if known.
1571 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1572 return TypeInt::make(klass->layout_helper());
1573 }
1574
1575 // No match.
1576 return NULL;
1577 }
1578
1579 // Try to constant-fold a stable array element.
1580 static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) {
1581 assert(ary->const_oop(), "array should be constant");
1582 assert(ary->is_stable(), "array should be stable");
1583
1584 // Decode the results of GraphKit::array_element_address.
1585 ciArray* aobj = ary->const_oop()->as_array();
1586 ciConstant con = aobj->element_value_by_offset(off);
1587
1588 if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) {
1589 const Type* con_type = Type::make_from_constant(con);
1590 if (con_type != NULL) {
1591 if (con_type->isa_aryptr()) {
1592 // Join with the array element type, in case it is also stable.
1593 int dim = ary->stable_dimension();
1594 con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1);
1595 }
1596 if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) {
1597 con_type = con_type->make_narrowoop();
1598 }
1599 #ifndef PRODUCT
1600 if (TraceIterativeGVN) {
1601 tty->print("FoldStableValues: array element [off=%d]: con_type=", off);
1602 con_type->dump(); tty->cr();
1603 }
1604 #endif //PRODUCT
1605 return con_type;
1606 }
1607 }
1608 return NULL;
1609 }
1610
1611 //------------------------------Value-----------------------------------------
1612 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1613 // Either input is TOP ==> the result is TOP
1614 Node* mem = in(MemNode::Memory);
1615 const Type *t1 = phase->type(mem);
1616 if (t1 == Type::TOP) return Type::TOP;
1617 Node* adr = in(MemNode::Address);
1618 const TypePtr* tp = phase->type(adr)->isa_ptr();
1619 if (tp == NULL || tp->empty()) return Type::TOP;
1620 int off = tp->offset();
1621 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1622 Compile* C = phase->C;
1623
1624 // Try to guess loaded type from pointer type
1625 if (tp->isa_aryptr()) {
1626 const TypeAryPtr* ary = tp->is_aryptr();
1627 const Type* t = ary->elem();
1628
1629 // Determine whether the reference is beyond the header or not, by comparing
1630 // the offset against the offset of the start of the array's data.
1631 // Different array types begin at slightly different offsets (12 vs. 16).
1632 // We choose T_BYTE as an example base type that is least restrictive
1633 // as to alignment, which will therefore produce the smallest
1634 // possible base offset.
1635 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1636 const bool off_beyond_header = ((uint)off >= (uint)min_base_off);
1637
1638 // Try to constant-fold a stable array element.
1639 if (FoldStableValues && ary->is_stable() && ary->const_oop() != NULL) {
1640 // Make sure the reference is not into the header and the offset is constant
1641 if (off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1642 const Type* con_type = fold_stable_ary_elem(ary, off, memory_type());
1643 if (con_type != NULL) {
1644 return con_type;
1645 }
1646 }
1647 }
1648
1649 // Don't do this for integer types. There is only potential profit if
1650 // the element type t is lower than _type; that is, for int types, if _type is
1651 // more restrictive than t. This only happens here if one is short and the other
1652 // char (both 16 bits), and in those cases we've made an intentional decision
1653 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1654 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1655 //
1656 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1657 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1658 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1659 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1660 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1661 // In fact, that could have been the original type of p1, and p1 could have
1662 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1663 // expression (LShiftL quux 3) independently optimized to the constant 8.
1664 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1665 && (_type->isa_vect() == NULL)
1666 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1667 // t might actually be lower than _type, if _type is a unique
1668 // concrete subclass of abstract class t.
1669 if (off_beyond_header) { // is the offset beyond the header?
1670 const Type* jt = t->join_speculative(_type);
1671 // In any case, do not allow the join, per se, to empty out the type.
1672 if (jt->empty() && !t->empty()) {
1673 // This can happen if a interface-typed array narrows to a class type.
1674 jt = _type;
1675 }
1676 #ifdef ASSERT
1677 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1678 // The pointers in the autobox arrays are always non-null
1679 Node* base = adr->in(AddPNode::Base);
1680 if ((base != NULL) && base->is_DecodeN()) {
1681 // Get LoadN node which loads IntegerCache.cache field
1682 base = base->in(1);
1683 }
1684 if ((base != NULL) && base->is_Con()) {
1685 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1686 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1687 // It could be narrow oop
1688 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1689 }
1690 }
1691 }
1692 #endif
1693 return jt;
1694 }
1695 }
1696 } else if (tp->base() == Type::InstPtr) {
1697 ciEnv* env = C->env();
1698 const TypeInstPtr* tinst = tp->is_instptr();
1699 ciKlass* klass = tinst->klass();
1700 assert( off != Type::OffsetBot ||
1701 // arrays can be cast to Objects
1702 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1703 // unsafe field access may not have a constant offset
1704 C->has_unsafe_access(),
1705 "Field accesses must be precise" );
1706 // For oop loads, we expect the _type to be precise
1707 if (klass == env->String_klass() &&
1708 adr->is_AddP() && off != Type::OffsetBot) {
1709 // For constant Strings treat the final fields as compile time constants.
1710 Node* base = adr->in(AddPNode::Base);
1711 const TypeOopPtr* t = phase->type(base)->isa_oopptr();
1712 if (t != NULL && t->singleton()) {
1713 ciField* field = env->String_klass()->get_field_by_offset(off, false);
1714 if (field != NULL && field->is_final()) {
1715 ciObject* string = t->const_oop();
1716 ciConstant constant = string->as_instance()->field_value(field);
1717 if (constant.basic_type() == T_INT) {
1718 return TypeInt::make(constant.as_int());
1719 } else if (constant.basic_type() == T_ARRAY) {
1720 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1721 return TypeNarrowOop::make_from_constant(constant.as_object(), true);
1722 } else {
1723 return TypeOopPtr::make_from_constant(constant.as_object(), true);
1724 }
1725 }
1726 }
1727 }
1728 }
1729 // Optimizations for constant objects
1730 ciObject* const_oop = tinst->const_oop();
1731 if (const_oop != NULL) {
1732 // For constant Boxed value treat the target field as a compile time constant.
1733 if (tinst->is_ptr_to_boxed_value()) {
1734 return tinst->get_const_boxed_value();
1735 } else
1736 // For constant CallSites treat the target field as a compile time constant.
1737 if (const_oop->is_call_site()) {
1738 ciCallSite* call_site = const_oop->as_call_site();
1739 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);
1740 if (field != NULL && field->is_call_site_target()) {
1741 ciMethodHandle* target = call_site->get_target();
1742 if (target != NULL) { // just in case
1743 ciConstant constant(T_OBJECT, target);
1744 const Type* t;
1745 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1746 t = TypeNarrowOop::make_from_constant(constant.as_object(), true);
1747 } else {
1748 t = TypeOopPtr::make_from_constant(constant.as_object(), true);
1749 }
1750 // Add a dependence for invalidation of the optimization.
1751 if (!call_site->is_constant_call_site()) {
1752 C->dependencies()->assert_call_site_target_value(call_site, target);
1753 }
1754 return t;
1755 }
1756 }
1757 }
1758 }
1759 } else if (tp->base() == Type::KlassPtr) {
1760 assert( off != Type::OffsetBot ||
1761 // arrays can be cast to Objects
1762 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1763 // also allow array-loading from the primary supertype
1764 // array during subtype checks
1765 Opcode() == Op_LoadKlass,
1766 "Field accesses must be precise" );
1767 // For klass/static loads, we expect the _type to be precise
1768 }
1769
1770 const TypeKlassPtr *tkls = tp->isa_klassptr();
1771 if (tkls != NULL && !StressReflectiveCode) {
1772 ciKlass* klass = tkls->klass();
1773 if (klass->is_loaded() && tkls->klass_is_exact()) {
1774 // We are loading a field from a Klass metaobject whose identity
1775 // is known at compile time (the type is "exact" or "precise").
1776 // Check for fields we know are maintained as constants by the VM.
1777 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1778 // The field is Klass::_super_check_offset. Return its (constant) value.
1779 // (Folds up type checking code.)
1780 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1781 return TypeInt::make(klass->super_check_offset());
1782 }
1783 // Compute index into primary_supers array
1784 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1785 // Check for overflowing; use unsigned compare to handle the negative case.
1786 if( depth < ciKlass::primary_super_limit() ) {
1787 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1788 // (Folds up type checking code.)
1789 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1790 ciKlass *ss = klass->super_of_depth(depth);
1791 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1792 }
1793 const Type* aift = load_array_final_field(tkls, klass);
1794 if (aift != NULL) return aift;
1795 if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset())
1796 && klass->is_array_klass()) {
1797 // The field is ArrayKlass::_component_mirror. Return its (constant) value.
1798 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1799 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1800 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1801 }
1802 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1803 // The field is Klass::_java_mirror. Return its (constant) value.
1804 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1805 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1806 return TypeInstPtr::make(klass->java_mirror());
1807 }
1808 }
1809
1810 // We can still check if we are loading from the primary_supers array at a
1811 // shallow enough depth. Even though the klass is not exact, entries less
1812 // than or equal to its super depth are correct.
1813 if (klass->is_loaded() ) {
1814 ciType *inner = klass;
1815 while( inner->is_obj_array_klass() )
1816 inner = inner->as_obj_array_klass()->base_element_type();
1817 if( inner->is_instance_klass() &&
1818 !inner->as_instance_klass()->flags().is_interface() ) {
1819 // Compute index into primary_supers array
1820 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1821 // Check for overflowing; use unsigned compare to handle the negative case.
1822 if( depth < ciKlass::primary_super_limit() &&
1823 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1824 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1825 // (Folds up type checking code.)
1826 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1827 ciKlass *ss = klass->super_of_depth(depth);
1828 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1829 }
1830 }
1831 }
1832
1833 // If the type is enough to determine that the thing is not an array,
1834 // we can give the layout_helper a positive interval type.
1835 // This will help short-circuit some reflective code.
1836 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1837 && !klass->is_array_klass() // not directly typed as an array
1838 && !klass->is_interface() // specifically not Serializable & Cloneable
1839 && !klass->is_java_lang_Object() // not the supertype of all T[]
1840 ) {
1841 // Note: When interfaces are reliable, we can narrow the interface
1842 // test to (klass != Serializable && klass != Cloneable).
1843 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1844 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1845 // The key property of this type is that it folds up tests
1846 // for array-ness, since it proves that the layout_helper is positive.
1847 // Thus, a generic value like the basic object layout helper works fine.
1848 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1849 }
1850 }
1851
1852 // If we are loading from a freshly-allocated object, produce a zero,
1853 // if the load is provably beyond the header of the object.
1854 // (Also allow a variable load from a fresh array to produce zero.)
1855 const TypeOopPtr *tinst = tp->isa_oopptr();
1856 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1857 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1858 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1859 Node* value = can_see_stored_value(mem,phase);
1860 if (value != NULL && value->is_Con()) {
1861 assert(value->bottom_type()->higher_equal(_type),"sanity");
1862 return value->bottom_type();
1863 }
1864 }
1865
1866 if (is_instance) {
1867 // If we have an instance type and our memory input is the
1868 // programs's initial memory state, there is no matching store,
1869 // so just return a zero of the appropriate type
1870 Node *mem = in(MemNode::Memory);
1871 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1872 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1873 return Type::get_zero_type(_type->basic_type());
1874 }
1875 }
1876 return _type;
1877 }
1878
1879 //------------------------------match_edge-------------------------------------
1880 // Do we Match on this edge index or not? Match only the address.
1881 uint LoadNode::match_edge(uint idx) const {
1882 return idx == MemNode::Address;
1883 }
1884
1885 //--------------------------LoadBNode::Ideal--------------------------------------
1886 //
1887 // If the previous store is to the same address as this load,
1888 // and the value stored was larger than a byte, replace this load
1889 // with the value stored truncated to a byte. If no truncation is
1890 // needed, the replacement is done in LoadNode::Identity().
1891 //
1892 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1893 Node* mem = in(MemNode::Memory);
1894 Node* value = can_see_stored_value(mem,phase);
1895 if( value && !phase->type(value)->higher_equal( _type ) ) {
1896 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) );
1897 return new (phase->C) RShiftINode(result, phase->intcon(24));
1898 }
1899 // Identity call will handle the case where truncation is not needed.
1900 return LoadNode::Ideal(phase, can_reshape);
1901 }
1902
1903 const Type* LoadBNode::Value(PhaseTransform *phase) const {
1904 Node* mem = in(MemNode::Memory);
1905 Node* value = can_see_stored_value(mem,phase);
1906 if (value != NULL && value->is_Con() &&
1907 !value->bottom_type()->higher_equal(_type)) {
1908 // If the input to the store does not fit with the load's result type,
1909 // it must be truncated. We can't delay until Ideal call since
1910 // a singleton Value is needed for split_thru_phi optimization.
1911 int con = value->get_int();
1912 return TypeInt::make((con << 24) >> 24);
1913 }
1914 return LoadNode::Value(phase);
1915 }
1916
1917 //--------------------------LoadUBNode::Ideal-------------------------------------
1918 //
1919 // If the previous store is to the same address as this load,
1920 // and the value stored was larger than a byte, replace this load
1921 // with the value stored truncated to a byte. If no truncation is
1922 // needed, the replacement is done in LoadNode::Identity().
1923 //
1924 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1925 Node* mem = in(MemNode::Memory);
1926 Node* value = can_see_stored_value(mem, phase);
1927 if (value && !phase->type(value)->higher_equal(_type))
1928 return new (phase->C) AndINode(value, phase->intcon(0xFF));
1929 // Identity call will handle the case where truncation is not needed.
1930 return LoadNode::Ideal(phase, can_reshape);
1931 }
1932
1933 const Type* LoadUBNode::Value(PhaseTransform *phase) const {
1934 Node* mem = in(MemNode::Memory);
1935 Node* value = can_see_stored_value(mem,phase);
1936 if (value != NULL && value->is_Con() &&
1937 !value->bottom_type()->higher_equal(_type)) {
1938 // If the input to the store does not fit with the load's result type,
1939 // it must be truncated. We can't delay until Ideal call since
1940 // a singleton Value is needed for split_thru_phi optimization.
1941 int con = value->get_int();
1942 return TypeInt::make(con & 0xFF);
1943 }
1944 return LoadNode::Value(phase);
1945 }
1946
1947 //--------------------------LoadUSNode::Ideal-------------------------------------
1948 //
1949 // If the previous store is to the same address as this load,
1950 // and the value stored was larger than a char, replace this load
1951 // with the value stored truncated to a char. If no truncation is
1952 // needed, the replacement is done in LoadNode::Identity().
1953 //
1954 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1955 Node* mem = in(MemNode::Memory);
1956 Node* value = can_see_stored_value(mem,phase);
1957 if( value && !phase->type(value)->higher_equal( _type ) )
1958 return new (phase->C) AndINode(value,phase->intcon(0xFFFF));
1959 // Identity call will handle the case where truncation is not needed.
1960 return LoadNode::Ideal(phase, can_reshape);
1961 }
1962
1963 const Type* LoadUSNode::Value(PhaseTransform *phase) const {
1964 Node* mem = in(MemNode::Memory);
1965 Node* value = can_see_stored_value(mem,phase);
1966 if (value != NULL && value->is_Con() &&
1967 !value->bottom_type()->higher_equal(_type)) {
1968 // If the input to the store does not fit with the load's result type,
1969 // it must be truncated. We can't delay until Ideal call since
1970 // a singleton Value is needed for split_thru_phi optimization.
1971 int con = value->get_int();
1972 return TypeInt::make(con & 0xFFFF);
1973 }
1974 return LoadNode::Value(phase);
1975 }
1976
1977 //--------------------------LoadSNode::Ideal--------------------------------------
1978 //
1979 // If the previous store is to the same address as this load,
1980 // and the value stored was larger than a short, replace this load
1981 // with the value stored truncated to a short. If no truncation is
1982 // needed, the replacement is done in LoadNode::Identity().
1983 //
1984 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1985 Node* mem = in(MemNode::Memory);
1986 Node* value = can_see_stored_value(mem,phase);
1987 if( value && !phase->type(value)->higher_equal( _type ) ) {
1988 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) );
1989 return new (phase->C) RShiftINode(result, phase->intcon(16));
1990 }
1991 // Identity call will handle the case where truncation is not needed.
1992 return LoadNode::Ideal(phase, can_reshape);
1993 }
1994
1995 const Type* LoadSNode::Value(PhaseTransform *phase) const {
1996 Node* mem = in(MemNode::Memory);
1997 Node* value = can_see_stored_value(mem,phase);
1998 if (value != NULL && value->is_Con() &&
1999 !value->bottom_type()->higher_equal(_type)) {
2000 // If the input to the store does not fit with the load's result type,
2001 // it must be truncated. We can't delay until Ideal call since
2002 // a singleton Value is needed for split_thru_phi optimization.
2003 int con = value->get_int();
2004 return TypeInt::make((con << 16) >> 16);
2005 }
2006 return LoadNode::Value(phase);
2007 }
2008
2009 //=============================================================================
2010 //----------------------------LoadKlassNode::make------------------------------
2011 // Polymorphic factory method:
2012 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
2013 Compile* C = gvn.C;
2014 Node *ctl = NULL;
2015 // sanity check the alias category against the created node type
2016 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2017 assert(adr_type != NULL, "expecting TypeKlassPtr");
2018 #ifdef _LP64
2019 if (adr_type->is_ptr_to_narrowklass()) {
2020 assert(UseCompressedClassPointers, "no compressed klasses");
2021 Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2022 return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2023 }
2024 #endif
2025 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2026 return new (C) LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2027 }
2028
2029 //------------------------------Value------------------------------------------
2030 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
2031 return klass_value_common(phase);
2032 }
2033
2034 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
2035 // Either input is TOP ==> the result is TOP
2036 const Type *t1 = phase->type( in(MemNode::Memory) );
2037 if (t1 == Type::TOP) return Type::TOP;
2038 Node *adr = in(MemNode::Address);
2039 const Type *t2 = phase->type( adr );
2040 if (t2 == Type::TOP) return Type::TOP;
2041 const TypePtr *tp = t2->is_ptr();
2042 if (TypePtr::above_centerline(tp->ptr()) ||
2043 tp->ptr() == TypePtr::Null) return Type::TOP;
2044
2045 // Return a more precise klass, if possible
2046 const TypeInstPtr *tinst = tp->isa_instptr();
2047 if (tinst != NULL) {
2048 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2049 int offset = tinst->offset();
2050 if (ik == phase->C->env()->Class_klass()
2051 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2052 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2053 // We are loading a special hidden field from a Class mirror object,
2054 // the field which points to the VM's Klass metaobject.
2055 ciType* t = tinst->java_mirror_type();
2056 // java_mirror_type returns non-null for compile-time Class constants.
2057 if (t != NULL) {
2058 // constant oop => constant klass
2059 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2060 if (t->is_void()) {
2061 // We cannot create a void array. Since void is a primitive type return null
2062 // klass. Users of this result need to do a null check on the returned klass.
2063 return TypePtr::NULL_PTR;
2064 }
2065 return TypeKlassPtr::make(ciArrayKlass::make(t));
2066 }
2067 if (!t->is_klass()) {
2068 // a primitive Class (e.g., int.class) has NULL for a klass field
2069 return TypePtr::NULL_PTR;
2070 }
2071 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2072 return TypeKlassPtr::make(t->as_klass());
2073 }
2074 // non-constant mirror, so we can't tell what's going on
2075 }
2076 if( !ik->is_loaded() )
2077 return _type; // Bail out if not loaded
2078 if (offset == oopDesc::klass_offset_in_bytes()) {
2079 if (tinst->klass_is_exact()) {
2080 return TypeKlassPtr::make(ik);
2081 }
2082 // See if we can become precise: no subklasses and no interface
2083 // (Note: We need to support verified interfaces.)
2084 if (!ik->is_interface() && !ik->has_subklass()) {
2085 //assert(!UseExactTypes, "this code should be useless with exact types");
2086 // Add a dependence; if any subclass added we need to recompile
2087 if (!ik->is_final()) {
2088 // %%% should use stronger assert_unique_concrete_subtype instead
2089 phase->C->dependencies()->assert_leaf_type(ik);
2090 }
2091 // Return precise klass
2092 return TypeKlassPtr::make(ik);
2093 }
2094
2095 // Return root of possible klass
2096 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2097 }
2098 }
2099
2100 // Check for loading klass from an array
2101 const TypeAryPtr *tary = tp->isa_aryptr();
2102 if( tary != NULL ) {
2103 ciKlass *tary_klass = tary->klass();
2104 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2105 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2106 if (tary->klass_is_exact()) {
2107 return TypeKlassPtr::make(tary_klass);
2108 }
2109 ciArrayKlass *ak = tary->klass()->as_array_klass();
2110 // If the klass is an object array, we defer the question to the
2111 // array component klass.
2112 if( ak->is_obj_array_klass() ) {
2113 assert( ak->is_loaded(), "" );
2114 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2115 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2116 ciInstanceKlass* ik = base_k->as_instance_klass();
2117 // See if we can become precise: no subklasses and no interface
2118 if (!ik->is_interface() && !ik->has_subklass()) {
2119 //assert(!UseExactTypes, "this code should be useless with exact types");
2120 // Add a dependence; if any subclass added we need to recompile
2121 if (!ik->is_final()) {
2122 phase->C->dependencies()->assert_leaf_type(ik);
2123 }
2124 // Return precise array klass
2125 return TypeKlassPtr::make(ak);
2126 }
2127 }
2128 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2129 } else { // Found a type-array?
2130 //assert(!UseExactTypes, "this code should be useless with exact types");
2131 assert( ak->is_type_array_klass(), "" );
2132 return TypeKlassPtr::make(ak); // These are always precise
2133 }
2134 }
2135 }
2136
2137 // Check for loading klass from an array klass
2138 const TypeKlassPtr *tkls = tp->isa_klassptr();
2139 if (tkls != NULL && !StressReflectiveCode) {
2140 ciKlass* klass = tkls->klass();
2141 if( !klass->is_loaded() )
2142 return _type; // Bail out if not loaded
2143 if( klass->is_obj_array_klass() &&
2144 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2145 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2146 // // Always returning precise element type is incorrect,
2147 // // e.g., element type could be object and array may contain strings
2148 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2149
2150 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2151 // according to the element type's subclassing.
2152 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2153 }
2154 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2155 tkls->offset() == in_bytes(Klass::super_offset())) {
2156 ciKlass* sup = klass->as_instance_klass()->super();
2157 // The field is Klass::_super. Return its (constant) value.
2158 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2159 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2160 }
2161 }
2162
2163 // Bailout case
2164 return LoadNode::Value(phase);
2165 }
2166
2167 //------------------------------Identity---------------------------------------
2168 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2169 // Also feed through the klass in Allocate(...klass...)._klass.
2170 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
2171 return klass_identity_common(phase);
2172 }
2173
2174 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
2175 Node* x = LoadNode::Identity(phase);
2176 if (x != this) return x;
2177
2178 // Take apart the address into an oop and and offset.
2179 // Return 'this' if we cannot.
2180 Node* adr = in(MemNode::Address);
2181 intptr_t offset = 0;
2182 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2183 if (base == NULL) return this;
2184 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2185 if (toop == NULL) return this;
2186
2187 // We can fetch the klass directly through an AllocateNode.
2188 // This works even if the klass is not constant (clone or newArray).
2189 if (offset == oopDesc::klass_offset_in_bytes()) {
2190 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2191 if (allocated_klass != NULL) {
2192 return allocated_klass;
2193 }
2194 }
2195
2196 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2197 // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass.
2198 // See inline_native_Class_query for occurrences of these patterns.
2199 // Java Example: x.getClass().isAssignableFrom(y)
2200 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
2201 //
2202 // This improves reflective code, often making the Class
2203 // mirror go completely dead. (Current exception: Class
2204 // mirrors may appear in debug info, but we could clean them out by
2205 // introducing a new debug info operator for Klass*.java_mirror).
2206 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2207 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2208 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2209 // We are loading a special hidden field from a Class mirror,
2210 // the field which points to its Klass or ArrayKlass metaobject.
2211 if (base->is_Load()) {
2212 Node* adr2 = base->in(MemNode::Address);
2213 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2214 if (tkls != NULL && !tkls->empty()
2215 && (tkls->klass()->is_instance_klass() ||
2216 tkls->klass()->is_array_klass())
2217 && adr2->is_AddP()
2218 ) {
2219 int mirror_field = in_bytes(Klass::java_mirror_offset());
2220 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2221 mirror_field = in_bytes(ArrayKlass::component_mirror_offset());
2222 }
2223 if (tkls->offset() == mirror_field) {
2224 return adr2->in(AddPNode::Base);
2225 }
2226 }
2227 }
2228 }
2229
2230 return this;
2231 }
2232
2233
2234 //------------------------------Value------------------------------------------
2235 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
2236 const Type *t = klass_value_common(phase);
2237 if (t == Type::TOP)
2238 return t;
2239
2240 return t->make_narrowklass();
2241 }
2242
2243 //------------------------------Identity---------------------------------------
2244 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2245 // Also feed through the klass in Allocate(...klass...)._klass.
2246 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
2247 Node *x = klass_identity_common(phase);
2248
2249 const Type *t = phase->type( x );
2250 if( t == Type::TOP ) return x;
2251 if( t->isa_narrowklass()) return x;
2252 assert (!t->isa_narrowoop(), "no narrow oop here");
2253
2254 return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass()));
2255 }
2256
2257 //------------------------------Value-----------------------------------------
2258 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
2259 // Either input is TOP ==> the result is TOP
2260 const Type *t1 = phase->type( in(MemNode::Memory) );
2261 if( t1 == Type::TOP ) return Type::TOP;
2262 Node *adr = in(MemNode::Address);
2263 const Type *t2 = phase->type( adr );
2264 if( t2 == Type::TOP ) return Type::TOP;
2265 const TypePtr *tp = t2->is_ptr();
2266 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2267 const TypeAryPtr *tap = tp->isa_aryptr();
2268 if( !tap ) return _type;
2269 return tap->size();
2270 }
2271
2272 //-------------------------------Ideal---------------------------------------
2273 // Feed through the length in AllocateArray(...length...)._length.
2274 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2275 Node* p = MemNode::Ideal_common(phase, can_reshape);
2276 if (p) return (p == NodeSentinel) ? NULL : p;
2277
2278 // Take apart the address into an oop and and offset.
2279 // Return 'this' if we cannot.
2280 Node* adr = in(MemNode::Address);
2281 intptr_t offset = 0;
2282 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2283 if (base == NULL) return NULL;
2284 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2285 if (tary == NULL) return NULL;
2286
2287 // We can fetch the length directly through an AllocateArrayNode.
2288 // This works even if the length is not constant (clone or newArray).
2289 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2290 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2291 if (alloc != NULL) {
2292 Node* allocated_length = alloc->Ideal_length();
2293 Node* len = alloc->make_ideal_length(tary, phase);
2294 if (allocated_length != len) {
2295 // New CastII improves on this.
2296 return len;
2297 }
2298 }
2299 }
2300
2301 return NULL;
2302 }
2303
2304 //------------------------------Identity---------------------------------------
2305 // Feed through the length in AllocateArray(...length...)._length.
2306 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
2307 Node* x = LoadINode::Identity(phase);
2308 if (x != this) return x;
2309
2310 // Take apart the address into an oop and and offset.
2311 // Return 'this' if we cannot.
2312 Node* adr = in(MemNode::Address);
2313 intptr_t offset = 0;
2314 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2315 if (base == NULL) return this;
2316 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2317 if (tary == NULL) return this;
2318
2319 // We can fetch the length directly through an AllocateArrayNode.
2320 // This works even if the length is not constant (clone or newArray).
2321 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2322 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2323 if (alloc != NULL) {
2324 Node* allocated_length = alloc->Ideal_length();
2325 // Do not allow make_ideal_length to allocate a CastII node.
2326 Node* len = alloc->make_ideal_length(tary, phase, false);
2327 if (allocated_length == len) {
2328 // Return allocated_length only if it would not be improved by a CastII.
2329 return allocated_length;
2330 }
2331 }
2332 }
2333
2334 return this;
2335
2336 }
2337
2338 //=============================================================================
2339 //---------------------------StoreNode::make-----------------------------------
2340 // Polymorphic factory method:
2341 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2342 assert((mo == unordered || mo == release), "unexpected");
2343 Compile* C = gvn.C;
2344 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2345 ctl != NULL, "raw memory operations should have control edge");
2346
2347 switch (bt) {
2348 case T_BOOLEAN:
2349 case T_BYTE: return new (C) StoreBNode(ctl, mem, adr, adr_type, val, mo);
2350 case T_INT: return new (C) StoreINode(ctl, mem, adr, adr_type, val, mo);
2351 case T_CHAR:
2352 case T_SHORT: return new (C) StoreCNode(ctl, mem, adr, adr_type, val, mo);
2353 case T_LONG: return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo);
2354 case T_FLOAT: return new (C) StoreFNode(ctl, mem, adr, adr_type, val, mo);
2355 case T_DOUBLE: return new (C) StoreDNode(ctl, mem, adr, adr_type, val, mo);
2356 case T_METADATA:
2357 case T_ADDRESS:
2358 case T_OBJECT:
2359 #ifdef _LP64
2360 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2361 val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop()));
2362 return new (C) StoreNNode(ctl, mem, adr, adr_type, val, mo);
2363 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2364 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2365 adr->bottom_type()->isa_rawptr())) {
2366 val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2367 return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2368 }
2369 #endif
2370 {
2371 return new (C) StorePNode(ctl, mem, adr, adr_type, val, mo);
2372 }
2373 }
2374 ShouldNotReachHere();
2375 return (StoreNode*)NULL;
2376 }
2377
2378 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2379 bool require_atomic = true;
2380 return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2381 }
2382
2383
2384 //--------------------------bottom_type----------------------------------------
2385 const Type *StoreNode::bottom_type() const {
2386 return Type::MEMORY;
2387 }
2388
2389 //------------------------------hash-------------------------------------------
2390 uint StoreNode::hash() const {
2391 // unroll addition of interesting fields
2392 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2393
2394 // Since they are not commoned, do not hash them:
2395 return NO_HASH;
2396 }
2397
2398 //------------------------------Ideal------------------------------------------
2399 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2400 // When a store immediately follows a relevant allocation/initialization,
2401 // try to capture it into the initialization, or hoist it above.
2402 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2403 Node* p = MemNode::Ideal_common(phase, can_reshape);
2404 if (p) return (p == NodeSentinel) ? NULL : p;
2405
2406 Node* mem = in(MemNode::Memory);
2407 Node* address = in(MemNode::Address);
2408
2409 // Back-to-back stores to same address? Fold em up. Generally
2410 // unsafe if I have intervening uses... Also disallowed for StoreCM
2411 // since they must follow each StoreP operation. Redundant StoreCMs
2412 // are eliminated just before matching in final_graph_reshape.
2413 if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) &&
2414 mem->Opcode() != Op_StoreCM) {
2415 // Looking at a dead closed cycle of memory?
2416 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2417
2418 assert(Opcode() == mem->Opcode() ||
2419 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
2420 "no mismatched stores, except on raw memory");
2421
2422 if (mem->outcnt() == 1 && // check for intervening uses
2423 mem->as_Store()->memory_size() <= this->memory_size()) {
2424 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
2425 // For example, 'mem' might be the final state at a conditional return.
2426 // Or, 'mem' might be used by some node which is live at the same time
2427 // 'this' is live, which might be unschedulable. So, require exactly
2428 // ONE user, the 'this' store, until such time as we clone 'mem' for
2429 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
2430 if (can_reshape) { // (%%% is this an anachronism?)
2431 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
2432 phase->is_IterGVN());
2433 } else {
2434 // It's OK to do this in the parser, since DU info is always accurate,
2435 // and the parser always refers to nodes via SafePointNode maps.
2436 set_req(MemNode::Memory, mem->in(MemNode::Memory));
2437 }
2438 return this;
2439 }
2440 }
2441
2442 // Capture an unaliased, unconditional, simple store into an initializer.
2443 // Or, if it is independent of the allocation, hoist it above the allocation.
2444 if (ReduceFieldZeroing && /*can_reshape &&*/
2445 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2446 InitializeNode* init = mem->in(0)->as_Initialize();
2447 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2448 if (offset > 0) {
2449 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2450 // If the InitializeNode captured me, it made a raw copy of me,
2451 // and I need to disappear.
2452 if (moved != NULL) {
2453 // %%% hack to ensure that Ideal returns a new node:
2454 mem = MergeMemNode::make(phase->C, mem);
2455 return mem; // fold me away
2456 }
2457 }
2458 }
2459
2460 return NULL; // No further progress
2461 }
2462
2463 //------------------------------Value-----------------------------------------
2464 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2465 // Either input is TOP ==> the result is TOP
2466 const Type *t1 = phase->type( in(MemNode::Memory) );
2467 if( t1 == Type::TOP ) return Type::TOP;
2468 const Type *t2 = phase->type( in(MemNode::Address) );
2469 if( t2 == Type::TOP ) return Type::TOP;
2470 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2471 if( t3 == Type::TOP ) return Type::TOP;
2472 return Type::MEMORY;
2473 }
2474
2475 //------------------------------Identity---------------------------------------
2476 // Remove redundant stores:
2477 // Store(m, p, Load(m, p)) changes to m.
2478 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2479 Node *StoreNode::Identity( PhaseTransform *phase ) {
2480 Node* mem = in(MemNode::Memory);
2481 Node* adr = in(MemNode::Address);
2482 Node* val = in(MemNode::ValueIn);
2483
2484 // Load then Store? Then the Store is useless
2485 if (val->is_Load() &&
2486 val->in(MemNode::Address)->eqv_uncast(adr) &&
2487 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2488 val->as_Load()->store_Opcode() == Opcode()) {
2489 return mem;
2490 }
2491
2492 // Two stores in a row of the same value?
2493 if (mem->is_Store() &&
2494 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2495 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2496 mem->Opcode() == Opcode()) {
2497 return mem;
2498 }
2499
2500 // Store of zero anywhere into a freshly-allocated object?
2501 // Then the store is useless.
2502 // (It must already have been captured by the InitializeNode.)
2503 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2504 // a newly allocated object is already all-zeroes everywhere
2505 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2506 return mem;
2507 }
2508
2509 // the store may also apply to zero-bits in an earlier object
2510 Node* prev_mem = find_previous_store(phase);
2511 // Steps (a), (b): Walk past independent stores to find an exact match.
2512 if (prev_mem != NULL) {
2513 Node* prev_val = can_see_stored_value(prev_mem, phase);
2514 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2515 // prev_val and val might differ by a cast; it would be good
2516 // to keep the more informative of the two.
2517 return mem;
2518 }
2519 }
2520 }
2521
2522 return this;
2523 }
2524
2525 //------------------------------match_edge-------------------------------------
2526 // Do we Match on this edge index or not? Match only memory & value
2527 uint StoreNode::match_edge(uint idx) const {
2528 return idx == MemNode::Address || idx == MemNode::ValueIn;
2529 }
2530
2531 //------------------------------cmp--------------------------------------------
2532 // Do not common stores up together. They generally have to be split
2533 // back up anyways, so do not bother.
2534 uint StoreNode::cmp( const Node &n ) const {
2535 return (&n == this); // Always fail except on self
2536 }
2537
2538 //------------------------------Ideal_masked_input-----------------------------
2539 // Check for a useless mask before a partial-word store
2540 // (StoreB ... (AndI valIn conIa) )
2541 // If (conIa & mask == mask) this simplifies to
2542 // (StoreB ... (valIn) )
2543 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2544 Node *val = in(MemNode::ValueIn);
2545 if( val->Opcode() == Op_AndI ) {
2546 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2547 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2548 set_req(MemNode::ValueIn, val->in(1));
2549 return this;
2550 }
2551 }
2552 return NULL;
2553 }
2554
2555
2556 //------------------------------Ideal_sign_extended_input----------------------
2557 // Check for useless sign-extension before a partial-word store
2558 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2559 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2560 // (StoreB ... (valIn) )
2561 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2562 Node *val = in(MemNode::ValueIn);
2563 if( val->Opcode() == Op_RShiftI ) {
2564 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2565 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2566 Node *shl = val->in(1);
2567 if( shl->Opcode() == Op_LShiftI ) {
2568 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2569 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2570 set_req(MemNode::ValueIn, shl->in(1));
2571 return this;
2572 }
2573 }
2574 }
2575 }
2576 return NULL;
2577 }
2578
2579 //------------------------------value_never_loaded-----------------------------------
2580 // Determine whether there are any possible loads of the value stored.
2581 // For simplicity, we actually check if there are any loads from the
2582 // address stored to, not just for loads of the value stored by this node.
2583 //
2584 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2585 Node *adr = in(Address);
2586 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2587 if (adr_oop == NULL)
2588 return false;
2589 if (!adr_oop->is_known_instance_field())
2590 return false; // if not a distinct instance, there may be aliases of the address
2591 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2592 Node *use = adr->fast_out(i);
2593 int opc = use->Opcode();
2594 if (use->is_Load() || use->is_LoadStore()) {
2595 return false;
2596 }
2597 }
2598 return true;
2599 }
2600
2601 //=============================================================================
2602 //------------------------------Ideal------------------------------------------
2603 // If the store is from an AND mask that leaves the low bits untouched, then
2604 // we can skip the AND operation. If the store is from a sign-extension
2605 // (a left shift, then right shift) we can skip both.
2606 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2607 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2608 if( progress != NULL ) return progress;
2609
2610 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2611 if( progress != NULL ) return progress;
2612
2613 // Finally check the default case
2614 return StoreNode::Ideal(phase, can_reshape);
2615 }
2616
2617 //=============================================================================
2618 //------------------------------Ideal------------------------------------------
2619 // If the store is from an AND mask that leaves the low bits untouched, then
2620 // we can skip the AND operation
2621 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2622 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2623 if( progress != NULL ) return progress;
2624
2625 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2626 if( progress != NULL ) return progress;
2627
2628 // Finally check the default case
2629 return StoreNode::Ideal(phase, can_reshape);
2630 }
2631
2632 //=============================================================================
2633 //------------------------------Identity---------------------------------------
2634 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2635 // No need to card mark when storing a null ptr
2636 Node* my_store = in(MemNode::OopStore);
2637 if (my_store->is_Store()) {
2638 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2639 if( t1 == TypePtr::NULL_PTR ) {
2640 return in(MemNode::Memory);
2641 }
2642 }
2643 return this;
2644 }
2645
2646 //=============================================================================
2647 //------------------------------Ideal---------------------------------------
2648 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2649 Node* progress = StoreNode::Ideal(phase, can_reshape);
2650 if (progress != NULL) return progress;
2651
2652 Node* my_store = in(MemNode::OopStore);
2653 if (my_store->is_MergeMem()) {
2654 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2655 set_req(MemNode::OopStore, mem);
2656 return this;
2657 }
2658
2659 return NULL;
2660 }
2661
2662 //------------------------------Value-----------------------------------------
2663 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2664 // Either input is TOP ==> the result is TOP
2665 const Type *t = phase->type( in(MemNode::Memory) );
2666 if( t == Type::TOP ) return Type::TOP;
2667 t = phase->type( in(MemNode::Address) );
2668 if( t == Type::TOP ) return Type::TOP;
2669 t = phase->type( in(MemNode::ValueIn) );
2670 if( t == Type::TOP ) return Type::TOP;
2671 // If extra input is TOP ==> the result is TOP
2672 t = phase->type( in(MemNode::OopStore) );
2673 if( t == Type::TOP ) return Type::TOP;
2674
2675 return StoreNode::Value( phase );
2676 }
2677
2678
2679 //=============================================================================
2680 //----------------------------------SCMemProjNode------------------------------
2681 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2682 {
2683 return bottom_type();
2684 }
2685
2686 //=============================================================================
2687 //----------------------------------LoadStoreNode------------------------------
2688 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2689 : Node(required),
2690 _type(rt),
2691 _adr_type(at)
2692 {
2693 init_req(MemNode::Control, c );
2694 init_req(MemNode::Memory , mem);
2695 init_req(MemNode::Address, adr);
2696 init_req(MemNode::ValueIn, val);
2697 init_class_id(Class_LoadStore);
2698 }
2699
2700 uint LoadStoreNode::ideal_reg() const {
2701 return _type->ideal_reg();
2702 }
2703
2704 bool LoadStoreNode::result_not_used() const {
2705 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2706 Node *x = fast_out(i);
2707 if (x->Opcode() == Op_SCMemProj) continue;
2708 return false;
2709 }
2710 return true;
2711 }
2712
2713 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2714
2715 //=============================================================================
2716 //----------------------------------LoadStoreConditionalNode--------------------
2717 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2718 init_req(ExpectedIn, ex );
2719 }
2720
2721 //=============================================================================
2722 //-------------------------------adr_type--------------------------------------
2723 // Do we Match on this edge index or not? Do not match memory
2724 const TypePtr* ClearArrayNode::adr_type() const {
2725 Node *adr = in(3);
2726 return MemNode::calculate_adr_type(adr->bottom_type());
2727 }
2728
2729 //------------------------------match_edge-------------------------------------
2730 // Do we Match on this edge index or not? Do not match memory
2731 uint ClearArrayNode::match_edge(uint idx) const {
2732 return idx > 1;
2733 }
2734
2735 //------------------------------Identity---------------------------------------
2736 // Clearing a zero length array does nothing
2737 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2738 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2739 }
2740
2741 //------------------------------Idealize---------------------------------------
2742 // Clearing a short array is faster with stores
2743 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2744 const int unit = BytesPerLong;
2745 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2746 if (!t) return NULL;
2747 if (!t->is_con()) return NULL;
2748 intptr_t raw_count = t->get_con();
2749 intptr_t size = raw_count;
2750 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2751 // Clearing nothing uses the Identity call.
2752 // Negative clears are possible on dead ClearArrays
2753 // (see jck test stmt114.stmt11402.val).
2754 if (size <= 0 || size % unit != 0) return NULL;
2755 intptr_t count = size / unit;
2756 // Length too long; use fast hardware clear
2757 if (size > Matcher::init_array_short_size) return NULL;
2758 Node *mem = in(1);
2759 if( phase->type(mem)==Type::TOP ) return NULL;
2760 Node *adr = in(3);
2761 const Type* at = phase->type(adr);
2762 if( at==Type::TOP ) return NULL;
2763 const TypePtr* atp = at->isa_ptr();
2764 // adjust atp to be the correct array element address type
2765 if (atp == NULL) atp = TypePtr::BOTTOM;
2766 else atp = atp->add_offset(Type::OffsetBot);
2767 // Get base for derived pointer purposes
2768 if( adr->Opcode() != Op_AddP ) Unimplemented();
2769 Node *base = adr->in(1);
2770
2771 Node *zero = phase->makecon(TypeLong::ZERO);
2772 Node *off = phase->MakeConX(BytesPerLong);
2773 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2774 count--;
2775 while( count-- ) {
2776 mem = phase->transform(mem);
2777 adr = phase->transform(new (phase->C) AddPNode(base,adr,off));
2778 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2779 }
2780 return mem;
2781 }
2782
2783 //----------------------------step_through----------------------------------
2784 // Return allocation input memory edge if it is different instance
2785 // or itself if it is the one we are looking for.
2786 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2787 Node* n = *np;
2788 assert(n->is_ClearArray(), "sanity");
2789 intptr_t offset;
2790 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2791 // This method is called only before Allocate nodes are expanded during
2792 // macro nodes expansion. Before that ClearArray nodes are only generated
2793 // in LibraryCallKit::generate_arraycopy() which follows allocations.
2794 assert(alloc != NULL, "should have allocation");
2795 if (alloc->_idx == instance_id) {
2796 // Can not bypass initialization of the instance we are looking for.
2797 return false;
2798 }
2799 // Otherwise skip it.
2800 InitializeNode* init = alloc->initialization();
2801 if (init != NULL)
2802 *np = init->in(TypeFunc::Memory);
2803 else
2804 *np = alloc->in(TypeFunc::Memory);
2805 return true;
2806 }
2807
2808 //----------------------------clear_memory-------------------------------------
2809 // Generate code to initialize object storage to zero.
2810 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2811 intptr_t start_offset,
2812 Node* end_offset,
2813 PhaseGVN* phase) {
2814 Compile* C = phase->C;
2815 intptr_t offset = start_offset;
2816
2817 int unit = BytesPerLong;
2818 if ((offset % unit) != 0) {
2819 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(offset));
2820 adr = phase->transform(adr);
2821 const TypePtr* atp = TypeRawPtr::BOTTOM;
2822 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2823 mem = phase->transform(mem);
2824 offset += BytesPerInt;
2825 }
2826 assert((offset % unit) == 0, "");
2827
2828 // Initialize the remaining stuff, if any, with a ClearArray.
2829 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2830 }
2831
2832 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2833 Node* start_offset,
2834 Node* end_offset,
2835 PhaseGVN* phase) {
2836 if (start_offset == end_offset) {
2837 // nothing to do
2838 return mem;
2839 }
2840
2841 Compile* C = phase->C;
2842 int unit = BytesPerLong;
2843 Node* zbase = start_offset;
2844 Node* zend = end_offset;
2845
2846 // Scale to the unit required by the CPU:
2847 if (!Matcher::init_array_count_is_in_bytes) {
2848 Node* shift = phase->intcon(exact_log2(unit));
2849 zbase = phase->transform( new(C) URShiftXNode(zbase, shift) );
2850 zend = phase->transform( new(C) URShiftXNode(zend, shift) );
2851 }
2852
2853 // Bulk clear double-words
2854 Node* zsize = phase->transform( new(C) SubXNode(zend, zbase) );
2855 Node* adr = phase->transform( new(C) AddPNode(dest, dest, start_offset) );
2856 mem = new (C) ClearArrayNode(ctl, mem, zsize, adr);
2857 return phase->transform(mem);
2858 }
2859
2860 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2861 intptr_t start_offset,
2862 intptr_t end_offset,
2863 PhaseGVN* phase) {
2864 if (start_offset == end_offset) {
2865 // nothing to do
2866 return mem;
2867 }
2868
2869 Compile* C = phase->C;
2870 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2871 intptr_t done_offset = end_offset;
2872 if ((done_offset % BytesPerLong) != 0) {
2873 done_offset -= BytesPerInt;
2874 }
2875 if (done_offset > start_offset) {
2876 mem = clear_memory(ctl, mem, dest,
2877 start_offset, phase->MakeConX(done_offset), phase);
2878 }
2879 if (done_offset < end_offset) { // emit the final 32-bit store
2880 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(done_offset));
2881 adr = phase->transform(adr);
2882 const TypePtr* atp = TypeRawPtr::BOTTOM;
2883 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2884 mem = phase->transform(mem);
2885 done_offset += BytesPerInt;
2886 }
2887 assert(done_offset == end_offset, "");
2888 return mem;
2889 }
2890
2891 //=============================================================================
2892 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2893 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2894 _adr_type(C->get_adr_type(alias_idx))
2895 {
2896 init_class_id(Class_MemBar);
2897 Node* top = C->top();
2898 init_req(TypeFunc::I_O,top);
2899 init_req(TypeFunc::FramePtr,top);
2900 init_req(TypeFunc::ReturnAdr,top);
2901 if (precedent != NULL)
2902 init_req(TypeFunc::Parms, precedent);
2903 }
2904
2905 //------------------------------cmp--------------------------------------------
2906 uint MemBarNode::hash() const { return NO_HASH; }
2907 uint MemBarNode::cmp( const Node &n ) const {
2908 return (&n == this); // Always fail except on self
2909 }
2910
2911 //------------------------------make-------------------------------------------
2912 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2913 switch (opcode) {
2914 case Op_MemBarAcquire: return new(C) MemBarAcquireNode(C, atp, pn);
2915 case Op_LoadFence: return new(C) LoadFenceNode(C, atp, pn);
2916 case Op_MemBarRelease: return new(C) MemBarReleaseNode(C, atp, pn);
2917 case Op_StoreFence: return new(C) StoreFenceNode(C, atp, pn);
2918 case Op_MemBarAcquireLock: return new(C) MemBarAcquireLockNode(C, atp, pn);
2919 case Op_MemBarReleaseLock: return new(C) MemBarReleaseLockNode(C, atp, pn);
2920 case Op_MemBarVolatile: return new(C) MemBarVolatileNode(C, atp, pn);
2921 case Op_MemBarCPUOrder: return new(C) MemBarCPUOrderNode(C, atp, pn);
2922 case Op_Initialize: return new(C) InitializeNode(C, atp, pn);
2923 case Op_MemBarStoreStore: return new(C) MemBarStoreStoreNode(C, atp, pn);
2924 default: ShouldNotReachHere(); return NULL;
2925 }
2926 }
2927
2928 //------------------------------Ideal------------------------------------------
2929 // Return a node which is more "ideal" than the current node. Strip out
2930 // control copies
2931 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2932 if (remove_dead_region(phase, can_reshape)) return this;
2933 // Don't bother trying to transform a dead node
2934 if (in(0) && in(0)->is_top()) {
2935 return NULL;
2936 }
2937
2938 // Eliminate volatile MemBars for scalar replaced objects.
2939 if (can_reshape && req() == (Precedent+1)) {
2940 bool eliminate = false;
2941 int opc = Opcode();
2942 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
2943 // Volatile field loads and stores.
2944 Node* my_mem = in(MemBarNode::Precedent);
2945 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
2946 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
2947 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
2948 // replace this Precedent (decodeN) with the Load instead.
2949 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
2950 Node* load_node = my_mem->in(1);
2951 set_req(MemBarNode::Precedent, load_node);
2952 phase->is_IterGVN()->_worklist.push(my_mem);
2953 my_mem = load_node;
2954 } else {
2955 assert(my_mem->unique_out() == this, "sanity");
2956 del_req(Precedent);
2957 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
2958 my_mem = NULL;
2959 }
2960 }
2961 if (my_mem != NULL && my_mem->is_Mem()) {
2962 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
2963 // Check for scalar replaced object reference.
2964 if( t_oop != NULL && t_oop->is_known_instance_field() &&
2965 t_oop->offset() != Type::OffsetBot &&
2966 t_oop->offset() != Type::OffsetTop) {
2967 eliminate = true;
2968 }
2969 }
2970 } else if (opc == Op_MemBarRelease) {
2971 // Final field stores.
2972 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
2973 if ((alloc != NULL) && alloc->is_Allocate() &&
2974 alloc->as_Allocate()->_is_non_escaping) {
2975 // The allocated object does not escape.
2976 eliminate = true;
2977 }
2978 }
2979 if (eliminate) {
2980 // Replace MemBar projections by its inputs.
2981 PhaseIterGVN* igvn = phase->is_IterGVN();
2982 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
2983 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
2984 // Must return either the original node (now dead) or a new node
2985 // (Do not return a top here, since that would break the uniqueness of top.)
2986 return new (phase->C) ConINode(TypeInt::ZERO);
2987 }
2988 }
2989 return NULL;
2990 }
2991
2992 //------------------------------Value------------------------------------------
2993 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2994 if( !in(0) ) return Type::TOP;
2995 if( phase->type(in(0)) == Type::TOP )
2996 return Type::TOP;
2997 return TypeTuple::MEMBAR;
2998 }
2999
3000 //------------------------------match------------------------------------------
3001 // Construct projections for memory.
3002 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3003 switch (proj->_con) {
3004 case TypeFunc::Control:
3005 case TypeFunc::Memory:
3006 return new (m->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3007 }
3008 ShouldNotReachHere();
3009 return NULL;
3010 }
3011
3012 //===========================InitializeNode====================================
3013 // SUMMARY:
3014 // This node acts as a memory barrier on raw memory, after some raw stores.
3015 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3016 // The Initialize can 'capture' suitably constrained stores as raw inits.
3017 // It can coalesce related raw stores into larger units (called 'tiles').
3018 // It can avoid zeroing new storage for memory units which have raw inits.
3019 // At macro-expansion, it is marked 'complete', and does not optimize further.
3020 //
3021 // EXAMPLE:
3022 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3023 // ctl = incoming control; mem* = incoming memory
3024 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3025 // First allocate uninitialized memory and fill in the header:
3026 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3027 // ctl := alloc.Control; mem* := alloc.Memory*
3028 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3029 // Then initialize to zero the non-header parts of the raw memory block:
3030 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3031 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3032 // After the initialize node executes, the object is ready for service:
3033 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3034 // Suppose its body is immediately initialized as {1,2}:
3035 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3036 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3037 // mem.SLICE(#short[*]) := store2
3038 //
3039 // DETAILS:
3040 // An InitializeNode collects and isolates object initialization after
3041 // an AllocateNode and before the next possible safepoint. As a
3042 // memory barrier (MemBarNode), it keeps critical stores from drifting
3043 // down past any safepoint or any publication of the allocation.
3044 // Before this barrier, a newly-allocated object may have uninitialized bits.
3045 // After this barrier, it may be treated as a real oop, and GC is allowed.
3046 //
3047 // The semantics of the InitializeNode include an implicit zeroing of
3048 // the new object from object header to the end of the object.
3049 // (The object header and end are determined by the AllocateNode.)
3050 //
3051 // Certain stores may be added as direct inputs to the InitializeNode.
3052 // These stores must update raw memory, and they must be to addresses
3053 // derived from the raw address produced by AllocateNode, and with
3054 // a constant offset. They must be ordered by increasing offset.
3055 // The first one is at in(RawStores), the last at in(req()-1).
3056 // Unlike most memory operations, they are not linked in a chain,
3057 // but are displayed in parallel as users of the rawmem output of
3058 // the allocation.
3059 //
3060 // (See comments in InitializeNode::capture_store, which continue
3061 // the example given above.)
3062 //
3063 // When the associated Allocate is macro-expanded, the InitializeNode
3064 // may be rewritten to optimize collected stores. A ClearArrayNode
3065 // may also be created at that point to represent any required zeroing.
3066 // The InitializeNode is then marked 'complete', prohibiting further
3067 // capturing of nearby memory operations.
3068 //
3069 // During macro-expansion, all captured initializations which store
3070 // constant values of 32 bits or smaller are coalesced (if advantageous)
3071 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3072 // initialized in fewer memory operations. Memory words which are
3073 // covered by neither tiles nor non-constant stores are pre-zeroed
3074 // by explicit stores of zero. (The code shape happens to do all
3075 // zeroing first, then all other stores, with both sequences occurring
3076 // in order of ascending offsets.)
3077 //
3078 // Alternatively, code may be inserted between an AllocateNode and its
3079 // InitializeNode, to perform arbitrary initialization of the new object.
3080 // E.g., the object copying intrinsics insert complex data transfers here.
3081 // The initialization must then be marked as 'complete' disable the
3082 // built-in zeroing semantics and the collection of initializing stores.
3083 //
3084 // While an InitializeNode is incomplete, reads from the memory state
3085 // produced by it are optimizable if they match the control edge and
3086 // new oop address associated with the allocation/initialization.
3087 // They return a stored value (if the offset matches) or else zero.
3088 // A write to the memory state, if it matches control and address,
3089 // and if it is to a constant offset, may be 'captured' by the
3090 // InitializeNode. It is cloned as a raw memory operation and rewired
3091 // inside the initialization, to the raw oop produced by the allocation.
3092 // Operations on addresses which are provably distinct (e.g., to
3093 // other AllocateNodes) are allowed to bypass the initialization.
3094 //
3095 // The effect of all this is to consolidate object initialization
3096 // (both arrays and non-arrays, both piecewise and bulk) into a
3097 // single location, where it can be optimized as a unit.
3098 //
3099 // Only stores with an offset less than TrackedInitializationLimit words
3100 // will be considered for capture by an InitializeNode. This puts a
3101 // reasonable limit on the complexity of optimized initializations.
3102
3103 //---------------------------InitializeNode------------------------------------
3104 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3105 : _is_complete(Incomplete), _does_not_escape(false),
3106 MemBarNode(C, adr_type, rawoop)
3107 {
3108 init_class_id(Class_Initialize);
3109
3110 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3111 assert(in(RawAddress) == rawoop, "proper init");
3112 // Note: allocation() can be NULL, for secondary initialization barriers
3113 }
3114
3115 // Since this node is not matched, it will be processed by the
3116 // register allocator. Declare that there are no constraints
3117 // on the allocation of the RawAddress edge.
3118 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3119 // This edge should be set to top, by the set_complete. But be conservative.
3120 if (idx == InitializeNode::RawAddress)
3121 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3122 return RegMask::Empty;
3123 }
3124
3125 Node* InitializeNode::memory(uint alias_idx) {
3126 Node* mem = in(Memory);
3127 if (mem->is_MergeMem()) {
3128 return mem->as_MergeMem()->memory_at(alias_idx);
3129 } else {
3130 // incoming raw memory is not split
3131 return mem;
3132 }
3133 }
3134
3135 bool InitializeNode::is_non_zero() {
3136 if (is_complete()) return false;
3137 remove_extra_zeroes();
3138 return (req() > RawStores);
3139 }
3140
3141 void InitializeNode::set_complete(PhaseGVN* phase) {
3142 assert(!is_complete(), "caller responsibility");
3143 _is_complete = Complete;
3144
3145 // After this node is complete, it contains a bunch of
3146 // raw-memory initializations. There is no need for
3147 // it to have anything to do with non-raw memory effects.
3148 // Therefore, tell all non-raw users to re-optimize themselves,
3149 // after skipping the memory effects of this initialization.
3150 PhaseIterGVN* igvn = phase->is_IterGVN();
3151 if (igvn) igvn->add_users_to_worklist(this);
3152 }
3153
3154 // convenience function
3155 // return false if the init contains any stores already
3156 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3157 InitializeNode* init = initialization();
3158 if (init == NULL || init->is_complete()) return false;
3159 init->remove_extra_zeroes();
3160 // for now, if this allocation has already collected any inits, bail:
3161 if (init->is_non_zero()) return false;
3162 init->set_complete(phase);
3163 return true;
3164 }
3165
3166 void InitializeNode::remove_extra_zeroes() {
3167 if (req() == RawStores) return;
3168 Node* zmem = zero_memory();
3169 uint fill = RawStores;
3170 for (uint i = fill; i < req(); i++) {
3171 Node* n = in(i);
3172 if (n->is_top() || n == zmem) continue; // skip
3173 if (fill < i) set_req(fill, n); // compact
3174 ++fill;
3175 }
3176 // delete any empty spaces created:
3177 while (fill < req()) {
3178 del_req(fill);
3179 }
3180 }
3181
3182 // Helper for remembering which stores go with which offsets.
3183 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3184 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3185 intptr_t offset = -1;
3186 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3187 phase, offset);
3188 if (base == NULL) return -1; // something is dead,
3189 if (offset < 0) return -1; // dead, dead
3190 return offset;
3191 }
3192
3193 // Helper for proving that an initialization expression is
3194 // "simple enough" to be folded into an object initialization.
3195 // Attempts to prove that a store's initial value 'n' can be captured
3196 // within the initialization without creating a vicious cycle, such as:
3197 // { Foo p = new Foo(); p.next = p; }
3198 // True for constants and parameters and small combinations thereof.
3199 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3200 if (n == NULL) return true; // (can this really happen?)
3201 if (n->is_Proj()) n = n->in(0);
3202 if (n == this) return false; // found a cycle
3203 if (n->is_Con()) return true;
3204 if (n->is_Start()) return true; // params, etc., are OK
3205 if (n->is_Root()) return true; // even better
3206
3207 Node* ctl = n->in(0);
3208 if (ctl != NULL && !ctl->is_top()) {
3209 if (ctl->is_Proj()) ctl = ctl->in(0);
3210 if (ctl == this) return false;
3211
3212 // If we already know that the enclosing memory op is pinned right after
3213 // the init, then any control flow that the store has picked up
3214 // must have preceded the init, or else be equal to the init.
3215 // Even after loop optimizations (which might change control edges)
3216 // a store is never pinned *before* the availability of its inputs.
3217 if (!MemNode::all_controls_dominate(n, this))
3218 return false; // failed to prove a good control
3219 }
3220
3221 // Check data edges for possible dependencies on 'this'.
3222 if ((count += 1) > 20) return false; // complexity limit
3223 for (uint i = 1; i < n->req(); i++) {
3224 Node* m = n->in(i);
3225 if (m == NULL || m == n || m->is_top()) continue;
3226 uint first_i = n->find_edge(m);
3227 if (i != first_i) continue; // process duplicate edge just once
3228 if (!detect_init_independence(m, count)) {
3229 return false;
3230 }
3231 }
3232
3233 return true;
3234 }
3235
3236 // Here are all the checks a Store must pass before it can be moved into
3237 // an initialization. Returns zero if a check fails.
3238 // On success, returns the (constant) offset to which the store applies,
3239 // within the initialized memory.
3240 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3241 const int FAIL = 0;
3242 if (st->req() != MemNode::ValueIn + 1)
3243 return FAIL; // an inscrutable StoreNode (card mark?)
3244 Node* ctl = st->in(MemNode::Control);
3245 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3246 return FAIL; // must be unconditional after the initialization
3247 Node* mem = st->in(MemNode::Memory);
3248 if (!(mem->is_Proj() && mem->in(0) == this))
3249 return FAIL; // must not be preceded by other stores
3250 Node* adr = st->in(MemNode::Address);
3251 intptr_t offset;
3252 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3253 if (alloc == NULL)
3254 return FAIL; // inscrutable address
3255 if (alloc != allocation())
3256 return FAIL; // wrong allocation! (store needs to float up)
3257 Node* val = st->in(MemNode::ValueIn);
3258 int complexity_count = 0;
3259 if (!detect_init_independence(val, complexity_count))
3260 return FAIL; // stored value must be 'simple enough'
3261
3262 // The Store can be captured only if nothing after the allocation
3263 // and before the Store is using the memory location that the store
3264 // overwrites.
3265 bool failed = false;
3266 // If is_complete_with_arraycopy() is true the shape of the graph is
3267 // well defined and is safe so no need for extra checks.
3268 if (!is_complete_with_arraycopy()) {
3269 // We are going to look at each use of the memory state following
3270 // the allocation to make sure nothing reads the memory that the
3271 // Store writes.
3272 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3273 int alias_idx = phase->C->get_alias_index(t_adr);
3274 ResourceMark rm;
3275 Unique_Node_List mems;
3276 mems.push(mem);
3277 Node* unique_merge = NULL;
3278 for (uint next = 0; next < mems.size(); ++next) {
3279 Node *m = mems.at(next);
3280 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3281 Node *n = m->fast_out(j);
3282 if (n->outcnt() == 0) {
3283 continue;
3284 }
3285 if (n == st) {
3286 continue;
3287 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3288 // If the control of this use is different from the control
3289 // of the Store which is right after the InitializeNode then
3290 // this node cannot be between the InitializeNode and the
3291 // Store.
3292 continue;
3293 } else if (n->is_MergeMem()) {
3294 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3295 // We can hit a MergeMemNode (that will likely go away
3296 // later) that is a direct use of the memory state
3297 // following the InitializeNode on the same slice as the
3298 // store node that we'd like to capture. We need to check
3299 // the uses of the MergeMemNode.
3300 mems.push(n);
3301 }
3302 } else if (n->is_Mem()) {
3303 Node* other_adr = n->in(MemNode::Address);
3304 if (other_adr == adr) {
3305 failed = true;
3306 break;
3307 } else {
3308 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3309 if (other_t_adr != NULL) {
3310 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3311 if (other_alias_idx == alias_idx) {
3312 // A load from the same memory slice as the store right
3313 // after the InitializeNode. We check the control of the
3314 // object/array that is loaded from. If it's the same as
3315 // the store control then we cannot capture the store.
3316 assert(!n->is_Store(), "2 stores to same slice on same control?");
3317 Node* base = other_adr;
3318 assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name()));
3319 base = base->in(AddPNode::Base);
3320 if (base != NULL) {
3321 base = base->uncast();
3322 if (base->is_Proj() && base->in(0) == alloc) {
3323 failed = true;
3324 break;
3325 }
3326 }
3327 }
3328 }
3329 }
3330 } else {
3331 failed = true;
3332 break;
3333 }
3334 }
3335 }
3336 }
3337 if (failed) {
3338 if (!can_reshape) {
3339 // We decided we couldn't capture the store during parsing. We
3340 // should try again during the next IGVN once the graph is
3341 // cleaner.
3342 phase->C->record_for_igvn(st);
3343 }
3344 return FAIL;
3345 }
3346
3347 return offset; // success
3348 }
3349
3350 // Find the captured store in(i) which corresponds to the range
3351 // [start..start+size) in the initialized object.
3352 // If there is one, return its index i. If there isn't, return the
3353 // negative of the index where it should be inserted.
3354 // Return 0 if the queried range overlaps an initialization boundary
3355 // or if dead code is encountered.
3356 // If size_in_bytes is zero, do not bother with overlap checks.
3357 int InitializeNode::captured_store_insertion_point(intptr_t start,
3358 int size_in_bytes,
3359 PhaseTransform* phase) {
3360 const int FAIL = 0, MAX_STORE = BytesPerLong;
3361
3362 if (is_complete())
3363 return FAIL; // arraycopy got here first; punt
3364
3365 assert(allocation() != NULL, "must be present");
3366
3367 // no negatives, no header fields:
3368 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3369
3370 // after a certain size, we bail out on tracking all the stores:
3371 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3372 if (start >= ti_limit) return FAIL;
3373
3374 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3375 if (i >= limit) return -(int)i; // not found; here is where to put it
3376
3377 Node* st = in(i);
3378 intptr_t st_off = get_store_offset(st, phase);
3379 if (st_off < 0) {
3380 if (st != zero_memory()) {
3381 return FAIL; // bail out if there is dead garbage
3382 }
3383 } else if (st_off > start) {
3384 // ...we are done, since stores are ordered
3385 if (st_off < start + size_in_bytes) {
3386 return FAIL; // the next store overlaps
3387 }
3388 return -(int)i; // not found; here is where to put it
3389 } else if (st_off < start) {
3390 if (size_in_bytes != 0 &&
3391 start < st_off + MAX_STORE &&
3392 start < st_off + st->as_Store()->memory_size()) {
3393 return FAIL; // the previous store overlaps
3394 }
3395 } else {
3396 if (size_in_bytes != 0 &&
3397 st->as_Store()->memory_size() != size_in_bytes) {
3398 return FAIL; // mismatched store size
3399 }
3400 return i;
3401 }
3402
3403 ++i;
3404 }
3405 }
3406
3407 // Look for a captured store which initializes at the offset 'start'
3408 // with the given size. If there is no such store, and no other
3409 // initialization interferes, then return zero_memory (the memory
3410 // projection of the AllocateNode).
3411 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3412 PhaseTransform* phase) {
3413 assert(stores_are_sane(phase), "");
3414 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3415 if (i == 0) {
3416 return NULL; // something is dead
3417 } else if (i < 0) {
3418 return zero_memory(); // just primordial zero bits here
3419 } else {
3420 Node* st = in(i); // here is the store at this position
3421 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3422 return st;
3423 }
3424 }
3425
3426 // Create, as a raw pointer, an address within my new object at 'offset'.
3427 Node* InitializeNode::make_raw_address(intptr_t offset,
3428 PhaseTransform* phase) {
3429 Node* addr = in(RawAddress);
3430 if (offset != 0) {
3431 Compile* C = phase->C;
3432 addr = phase->transform( new (C) AddPNode(C->top(), addr,
3433 phase->MakeConX(offset)) );
3434 }
3435 return addr;
3436 }
3437
3438 // Clone the given store, converting it into a raw store
3439 // initializing a field or element of my new object.
3440 // Caller is responsible for retiring the original store,
3441 // with subsume_node or the like.
3442 //
3443 // From the example above InitializeNode::InitializeNode,
3444 // here are the old stores to be captured:
3445 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3446 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3447 //
3448 // Here is the changed code; note the extra edges on init:
3449 // alloc = (Allocate ...)
3450 // rawoop = alloc.RawAddress
3451 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3452 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3453 // init = (Initialize alloc.Control alloc.Memory rawoop
3454 // rawstore1 rawstore2)
3455 //
3456 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3457 PhaseTransform* phase, bool can_reshape) {
3458 assert(stores_are_sane(phase), "");
3459
3460 if (start < 0) return NULL;
3461 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3462
3463 Compile* C = phase->C;
3464 int size_in_bytes = st->memory_size();
3465 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3466 if (i == 0) return NULL; // bail out
3467 Node* prev_mem = NULL; // raw memory for the captured store
3468 if (i > 0) {
3469 prev_mem = in(i); // there is a pre-existing store under this one
3470 set_req(i, C->top()); // temporarily disconnect it
3471 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3472 } else {
3473 i = -i; // no pre-existing store
3474 prev_mem = zero_memory(); // a slice of the newly allocated object
3475 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3476 set_req(--i, C->top()); // reuse this edge; it has been folded away
3477 else
3478 ins_req(i, C->top()); // build a new edge
3479 }
3480 Node* new_st = st->clone();
3481 new_st->set_req(MemNode::Control, in(Control));
3482 new_st->set_req(MemNode::Memory, prev_mem);
3483 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3484 new_st = phase->transform(new_st);
3485
3486 // At this point, new_st might have swallowed a pre-existing store
3487 // at the same offset, or perhaps new_st might have disappeared,
3488 // if it redundantly stored the same value (or zero to fresh memory).
3489
3490 // In any case, wire it in:
3491 set_req(i, new_st);
3492
3493 // The caller may now kill the old guy.
3494 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3495 assert(check_st == new_st || check_st == NULL, "must be findable");
3496 assert(!is_complete(), "");
3497 return new_st;
3498 }
3499
3500 static bool store_constant(jlong* tiles, int num_tiles,
3501 intptr_t st_off, int st_size,
3502 jlong con) {
3503 if ((st_off & (st_size-1)) != 0)
3504 return false; // strange store offset (assume size==2**N)
3505 address addr = (address)tiles + st_off;
3506 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3507 switch (st_size) {
3508 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3509 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3510 case sizeof(jint): *(jint*) addr = (jint) con; break;
3511 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3512 default: return false; // strange store size (detect size!=2**N here)
3513 }
3514 return true; // return success to caller
3515 }
3516
3517 // Coalesce subword constants into int constants and possibly
3518 // into long constants. The goal, if the CPU permits,
3519 // is to initialize the object with a small number of 64-bit tiles.
3520 // Also, convert floating-point constants to bit patterns.
3521 // Non-constants are not relevant to this pass.
3522 //
3523 // In terms of the running example on InitializeNode::InitializeNode
3524 // and InitializeNode::capture_store, here is the transformation
3525 // of rawstore1 and rawstore2 into rawstore12:
3526 // alloc = (Allocate ...)
3527 // rawoop = alloc.RawAddress
3528 // tile12 = 0x00010002
3529 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3530 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3531 //
3532 void
3533 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3534 Node* size_in_bytes,
3535 PhaseGVN* phase) {
3536 Compile* C = phase->C;
3537
3538 assert(stores_are_sane(phase), "");
3539 // Note: After this pass, they are not completely sane,
3540 // since there may be some overlaps.
3541
3542 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3543
3544 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3545 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3546 size_limit = MIN2(size_limit, ti_limit);
3547 size_limit = align_size_up(size_limit, BytesPerLong);
3548 int num_tiles = size_limit / BytesPerLong;
3549
3550 // allocate space for the tile map:
3551 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3552 jlong tiles_buf[small_len];
3553 Node* nodes_buf[small_len];
3554 jlong inits_buf[small_len];
3555 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3556 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3557 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3558 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3559 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3560 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3561 // tiles: exact bitwise model of all primitive constants
3562 // nodes: last constant-storing node subsumed into the tiles model
3563 // inits: which bytes (in each tile) are touched by any initializations
3564
3565 //// Pass A: Fill in the tile model with any relevant stores.
3566
3567 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3568 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3569 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3570 Node* zmem = zero_memory(); // initially zero memory state
3571 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3572 Node* st = in(i);
3573 intptr_t st_off = get_store_offset(st, phase);
3574
3575 // Figure out the store's offset and constant value:
3576 if (st_off < header_size) continue; //skip (ignore header)
3577 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3578 int st_size = st->as_Store()->memory_size();
3579 if (st_off + st_size > size_limit) break;
3580
3581 // Record which bytes are touched, whether by constant or not.
3582 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3583 continue; // skip (strange store size)
3584
3585 const Type* val = phase->type(st->in(MemNode::ValueIn));
3586 if (!val->singleton()) continue; //skip (non-con store)
3587 BasicType type = val->basic_type();
3588
3589 jlong con = 0;
3590 switch (type) {
3591 case T_INT: con = val->is_int()->get_con(); break;
3592 case T_LONG: con = val->is_long()->get_con(); break;
3593 case T_FLOAT: con = jint_cast(val->getf()); break;
3594 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3595 default: continue; //skip (odd store type)
3596 }
3597
3598 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3599 st->Opcode() == Op_StoreL) {
3600 continue; // This StoreL is already optimal.
3601 }
3602
3603 // Store down the constant.
3604 store_constant(tiles, num_tiles, st_off, st_size, con);
3605
3606 intptr_t j = st_off >> LogBytesPerLong;
3607
3608 if (type == T_INT && st_size == BytesPerInt
3609 && (st_off & BytesPerInt) == BytesPerInt) {
3610 jlong lcon = tiles[j];
3611 if (!Matcher::isSimpleConstant64(lcon) &&
3612 st->Opcode() == Op_StoreI) {
3613 // This StoreI is already optimal by itself.
3614 jint* intcon = (jint*) &tiles[j];
3615 intcon[1] = 0; // undo the store_constant()
3616
3617 // If the previous store is also optimal by itself, back up and
3618 // undo the action of the previous loop iteration... if we can.
3619 // But if we can't, just let the previous half take care of itself.
3620 st = nodes[j];
3621 st_off -= BytesPerInt;
3622 con = intcon[0];
3623 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3624 assert(st_off >= header_size, "still ignoring header");
3625 assert(get_store_offset(st, phase) == st_off, "must be");
3626 assert(in(i-1) == zmem, "must be");
3627 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3628 assert(con == tcon->is_int()->get_con(), "must be");
3629 // Undo the effects of the previous loop trip, which swallowed st:
3630 intcon[0] = 0; // undo store_constant()
3631 set_req(i-1, st); // undo set_req(i, zmem)
3632 nodes[j] = NULL; // undo nodes[j] = st
3633 --old_subword; // undo ++old_subword
3634 }
3635 continue; // This StoreI is already optimal.
3636 }
3637 }
3638
3639 // This store is not needed.
3640 set_req(i, zmem);
3641 nodes[j] = st; // record for the moment
3642 if (st_size < BytesPerLong) // something has changed
3643 ++old_subword; // includes int/float, but who's counting...
3644 else ++old_long;
3645 }
3646
3647 if ((old_subword + old_long) == 0)
3648 return; // nothing more to do
3649
3650 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3651 // Be sure to insert them before overlapping non-constant stores.
3652 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3653 for (int j = 0; j < num_tiles; j++) {
3654 jlong con = tiles[j];
3655 jlong init = inits[j];
3656 if (con == 0) continue;
3657 jint con0, con1; // split the constant, address-wise
3658 jint init0, init1; // split the init map, address-wise
3659 { union { jlong con; jint intcon[2]; } u;
3660 u.con = con;
3661 con0 = u.intcon[0];
3662 con1 = u.intcon[1];
3663 u.con = init;
3664 init0 = u.intcon[0];
3665 init1 = u.intcon[1];
3666 }
3667
3668 Node* old = nodes[j];
3669 assert(old != NULL, "need the prior store");
3670 intptr_t offset = (j * BytesPerLong);
3671
3672 bool split = !Matcher::isSimpleConstant64(con);
3673
3674 if (offset < header_size) {
3675 assert(offset + BytesPerInt >= header_size, "second int counts");
3676 assert(*(jint*)&tiles[j] == 0, "junk in header");
3677 split = true; // only the second word counts
3678 // Example: int a[] = { 42 ... }
3679 } else if (con0 == 0 && init0 == -1) {
3680 split = true; // first word is covered by full inits
3681 // Example: int a[] = { ... foo(), 42 ... }
3682 } else if (con1 == 0 && init1 == -1) {
3683 split = true; // second word is covered by full inits
3684 // Example: int a[] = { ... 42, foo() ... }
3685 }
3686
3687 // Here's a case where init0 is neither 0 nor -1:
3688 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3689 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3690 // In this case the tile is not split; it is (jlong)42.
3691 // The big tile is stored down, and then the foo() value is inserted.
3692 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3693
3694 Node* ctl = old->in(MemNode::Control);
3695 Node* adr = make_raw_address(offset, phase);
3696 const TypePtr* atp = TypeRawPtr::BOTTOM;
3697
3698 // One or two coalesced stores to plop down.
3699 Node* st[2];
3700 intptr_t off[2];
3701 int nst = 0;
3702 if (!split) {
3703 ++new_long;
3704 off[nst] = offset;
3705 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3706 phase->longcon(con), T_LONG, MemNode::unordered);
3707 } else {
3708 // Omit either if it is a zero.
3709 if (con0 != 0) {
3710 ++new_int;
3711 off[nst] = offset;
3712 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3713 phase->intcon(con0), T_INT, MemNode::unordered);
3714 }
3715 if (con1 != 0) {
3716 ++new_int;
3717 offset += BytesPerInt;
3718 adr = make_raw_address(offset, phase);
3719 off[nst] = offset;
3720 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3721 phase->intcon(con1), T_INT, MemNode::unordered);
3722 }
3723 }
3724
3725 // Insert second store first, then the first before the second.
3726 // Insert each one just before any overlapping non-constant stores.
3727 while (nst > 0) {
3728 Node* st1 = st[--nst];
3729 C->copy_node_notes_to(st1, old);
3730 st1 = phase->transform(st1);
3731 offset = off[nst];
3732 assert(offset >= header_size, "do not smash header");
3733 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3734 guarantee(ins_idx != 0, "must re-insert constant store");
3735 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3736 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3737 set_req(--ins_idx, st1);
3738 else
3739 ins_req(ins_idx, st1);
3740 }
3741 }
3742
3743 if (PrintCompilation && WizardMode)
3744 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3745 old_subword, old_long, new_int, new_long);
3746 if (C->log() != NULL)
3747 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3748 old_subword, old_long, new_int, new_long);
3749
3750 // Clean up any remaining occurrences of zmem:
3751 remove_extra_zeroes();
3752 }
3753
3754 // Explore forward from in(start) to find the first fully initialized
3755 // word, and return its offset. Skip groups of subword stores which
3756 // together initialize full words. If in(start) is itself part of a
3757 // fully initialized word, return the offset of in(start). If there
3758 // are no following full-word stores, or if something is fishy, return
3759 // a negative value.
3760 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3761 int int_map = 0;
3762 intptr_t int_map_off = 0;
3763 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3764
3765 for (uint i = start, limit = req(); i < limit; i++) {
3766 Node* st = in(i);
3767
3768 intptr_t st_off = get_store_offset(st, phase);
3769 if (st_off < 0) break; // return conservative answer
3770
3771 int st_size = st->as_Store()->memory_size();
3772 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3773 return st_off; // we found a complete word init
3774 }
3775
3776 // update the map:
3777
3778 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3779 if (this_int_off != int_map_off) {
3780 // reset the map:
3781 int_map = 0;
3782 int_map_off = this_int_off;
3783 }
3784
3785 int subword_off = st_off - this_int_off;
3786 int_map |= right_n_bits(st_size) << subword_off;
3787 if ((int_map & FULL_MAP) == FULL_MAP) {
3788 return this_int_off; // we found a complete word init
3789 }
3790
3791 // Did this store hit or cross the word boundary?
3792 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3793 if (next_int_off == this_int_off + BytesPerInt) {
3794 // We passed the current int, without fully initializing it.
3795 int_map_off = next_int_off;
3796 int_map >>= BytesPerInt;
3797 } else if (next_int_off > this_int_off + BytesPerInt) {
3798 // We passed the current and next int.
3799 return this_int_off + BytesPerInt;
3800 }
3801 }
3802
3803 return -1;
3804 }
3805
3806
3807 // Called when the associated AllocateNode is expanded into CFG.
3808 // At this point, we may perform additional optimizations.
3809 // Linearize the stores by ascending offset, to make memory
3810 // activity as coherent as possible.
3811 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3812 intptr_t header_size,
3813 Node* size_in_bytes,
3814 PhaseGVN* phase) {
3815 assert(!is_complete(), "not already complete");
3816 assert(stores_are_sane(phase), "");
3817 assert(allocation() != NULL, "must be present");
3818
3819 remove_extra_zeroes();
3820
3821 if (ReduceFieldZeroing || ReduceBulkZeroing)
3822 // reduce instruction count for common initialization patterns
3823 coalesce_subword_stores(header_size, size_in_bytes, phase);
3824
3825 Node* zmem = zero_memory(); // initially zero memory state
3826 Node* inits = zmem; // accumulating a linearized chain of inits
3827 #ifdef ASSERT
3828 intptr_t first_offset = allocation()->minimum_header_size();
3829 intptr_t last_init_off = first_offset; // previous init offset
3830 intptr_t last_init_end = first_offset; // previous init offset+size
3831 intptr_t last_tile_end = first_offset; // previous tile offset+size
3832 #endif
3833 intptr_t zeroes_done = header_size;
3834
3835 bool do_zeroing = true; // we might give up if inits are very sparse
3836 int big_init_gaps = 0; // how many large gaps have we seen?
3837
3838 if (ZeroTLAB) do_zeroing = false;
3839 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3840
3841 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3842 Node* st = in(i);
3843 intptr_t st_off = get_store_offset(st, phase);
3844 if (st_off < 0)
3845 break; // unknown junk in the inits
3846 if (st->in(MemNode::Memory) != zmem)
3847 break; // complicated store chains somehow in list
3848
3849 int st_size = st->as_Store()->memory_size();
3850 intptr_t next_init_off = st_off + st_size;
3851
3852 if (do_zeroing && zeroes_done < next_init_off) {
3853 // See if this store needs a zero before it or under it.
3854 intptr_t zeroes_needed = st_off;
3855
3856 if (st_size < BytesPerInt) {
3857 // Look for subword stores which only partially initialize words.
3858 // If we find some, we must lay down some word-level zeroes first,
3859 // underneath the subword stores.
3860 //
3861 // Examples:
3862 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3863 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3864 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3865 //
3866 // Note: coalesce_subword_stores may have already done this,
3867 // if it was prompted by constant non-zero subword initializers.
3868 // But this case can still arise with non-constant stores.
3869
3870 intptr_t next_full_store = find_next_fullword_store(i, phase);
3871
3872 // In the examples above:
3873 // in(i) p q r s x y z
3874 // st_off 12 13 14 15 12 13 14
3875 // st_size 1 1 1 1 1 1 1
3876 // next_full_s. 12 16 16 16 16 16 16
3877 // z's_done 12 16 16 16 12 16 12
3878 // z's_needed 12 16 16 16 16 16 16
3879 // zsize 0 0 0 0 4 0 4
3880 if (next_full_store < 0) {
3881 // Conservative tack: Zero to end of current word.
3882 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3883 } else {
3884 // Zero to beginning of next fully initialized word.
3885 // Or, don't zero at all, if we are already in that word.
3886 assert(next_full_store >= zeroes_needed, "must go forward");
3887 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3888 zeroes_needed = next_full_store;
3889 }
3890 }
3891
3892 if (zeroes_needed > zeroes_done) {
3893 intptr_t zsize = zeroes_needed - zeroes_done;
3894 // Do some incremental zeroing on rawmem, in parallel with inits.
3895 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3896 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3897 zeroes_done, zeroes_needed,
3898 phase);
3899 zeroes_done = zeroes_needed;
3900 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3901 do_zeroing = false; // leave the hole, next time
3902 }
3903 }
3904
3905 // Collect the store and move on:
3906 st->set_req(MemNode::Memory, inits);
3907 inits = st; // put it on the linearized chain
3908 set_req(i, zmem); // unhook from previous position
3909
3910 if (zeroes_done == st_off)
3911 zeroes_done = next_init_off;
3912
3913 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3914
3915 #ifdef ASSERT
3916 // Various order invariants. Weaker than stores_are_sane because
3917 // a large constant tile can be filled in by smaller non-constant stores.
3918 assert(st_off >= last_init_off, "inits do not reverse");
3919 last_init_off = st_off;
3920 const Type* val = NULL;
3921 if (st_size >= BytesPerInt &&
3922 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3923 (int)val->basic_type() < (int)T_OBJECT) {
3924 assert(st_off >= last_tile_end, "tiles do not overlap");
3925 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3926 last_tile_end = MAX2(last_tile_end, next_init_off);
3927 } else {
3928 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3929 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3930 assert(st_off >= last_init_end, "inits do not overlap");
3931 last_init_end = next_init_off; // it's a non-tile
3932 }
3933 #endif //ASSERT
3934 }
3935
3936 remove_extra_zeroes(); // clear out all the zmems left over
3937 add_req(inits);
3938
3939 if (!ZeroTLAB) {
3940 // If anything remains to be zeroed, zero it all now.
3941 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3942 // if it is the last unused 4 bytes of an instance, forget about it
3943 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3944 if (zeroes_done + BytesPerLong >= size_limit) {
3945 assert(allocation() != NULL, "");
3946 if (allocation()->Opcode() == Op_Allocate) {
3947 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3948 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3949 if (zeroes_done == k->layout_helper())
3950 zeroes_done = size_limit;
3951 }
3952 }
3953 if (zeroes_done < size_limit) {
3954 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3955 zeroes_done, size_in_bytes, phase);
3956 }
3957 }
3958
3959 set_complete(phase);
3960 return rawmem;
3961 }
3962
3963
3964 #ifdef ASSERT
3965 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3966 if (is_complete())
3967 return true; // stores could be anything at this point
3968 assert(allocation() != NULL, "must be present");
3969 intptr_t last_off = allocation()->minimum_header_size();
3970 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3971 Node* st = in(i);
3972 intptr_t st_off = get_store_offset(st, phase);
3973 if (st_off < 0) continue; // ignore dead garbage
3974 if (last_off > st_off) {
3975 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
3976 this->dump(2);
3977 assert(false, "ascending store offsets");
3978 return false;
3979 }
3980 last_off = st_off + st->as_Store()->memory_size();
3981 }
3982 return true;
3983 }
3984 #endif //ASSERT
3985
3986
3987
3988
3989 //============================MergeMemNode=====================================
3990 //
3991 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3992 // contributing store or call operations. Each contributor provides the memory
3993 // state for a particular "alias type" (see Compile::alias_type). For example,
3994 // if a MergeMem has an input X for alias category #6, then any memory reference
3995 // to alias category #6 may use X as its memory state input, as an exact equivalent
3996 // to using the MergeMem as a whole.
3997 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3998 //
3999 // (Here, the <N> notation gives the index of the relevant adr_type.)
4000 //
4001 // In one special case (and more cases in the future), alias categories overlap.
4002 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4003 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4004 // it is exactly equivalent to that state W:
4005 // MergeMem(<Bot>: W) <==> W
4006 //
4007 // Usually, the merge has more than one input. In that case, where inputs
4008 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4009 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4010 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4011 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4012 //
4013 // A merge can take a "wide" memory state as one of its narrow inputs.
4014 // This simply means that the merge observes out only the relevant parts of
4015 // the wide input. That is, wide memory states arriving at narrow merge inputs
4016 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4017 //
4018 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4019 // and that memory slices "leak through":
4020 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4021 //
4022 // But, in such a cascade, repeated memory slices can "block the leak":
4023 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4024 //
4025 // In the last example, Y is not part of the combined memory state of the
4026 // outermost MergeMem. The system must, of course, prevent unschedulable
4027 // memory states from arising, so you can be sure that the state Y is somehow
4028 // a precursor to state Y'.
4029 //
4030 //
4031 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4032 // of each MergeMemNode array are exactly the numerical alias indexes, including
4033 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4034 // Compile::alias_type (and kin) produce and manage these indexes.
4035 //
4036 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4037 // (Note that this provides quick access to the top node inside MergeMem methods,
4038 // without the need to reach out via TLS to Compile::current.)
4039 //
4040 // As a consequence of what was just described, a MergeMem that represents a full
4041 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4042 // containing all alias categories.
4043 //
4044 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4045 //
4046 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4047 // a memory state for the alias type <N>, or else the top node, meaning that
4048 // there is no particular input for that alias type. Note that the length of
4049 // a MergeMem is variable, and may be extended at any time to accommodate new
4050 // memory states at larger alias indexes. When merges grow, they are of course
4051 // filled with "top" in the unused in() positions.
4052 //
4053 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4054 // (Top was chosen because it works smoothly with passes like GCM.)
4055 //
4056 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4057 // the type of random VM bits like TLS references.) Since it is always the
4058 // first non-Bot memory slice, some low-level loops use it to initialize an
4059 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4060 //
4061 //
4062 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4063 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4064 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4065 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4066 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4067 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4068 //
4069 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4070 // really that different from the other memory inputs. An abbreviation called
4071 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4072 //
4073 //
4074 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4075 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4076 // that "emerges though" the base memory will be marked as excluding the alias types
4077 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4078 //
4079 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4080 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4081 //
4082 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4083 // (It is currently unimplemented.) As you can see, the resulting merge is
4084 // actually a disjoint union of memory states, rather than an overlay.
4085 //
4086
4087 //------------------------------MergeMemNode-----------------------------------
4088 Node* MergeMemNode::make_empty_memory() {
4089 Node* empty_memory = (Node*) Compile::current()->top();
4090 assert(empty_memory->is_top(), "correct sentinel identity");
4091 return empty_memory;
4092 }
4093
4094 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4095 init_class_id(Class_MergeMem);
4096 // all inputs are nullified in Node::Node(int)
4097 // set_input(0, NULL); // no control input
4098
4099 // Initialize the edges uniformly to top, for starters.
4100 Node* empty_mem = make_empty_memory();
4101 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4102 init_req(i,empty_mem);
4103 }
4104 assert(empty_memory() == empty_mem, "");
4105
4106 if( new_base != NULL && new_base->is_MergeMem() ) {
4107 MergeMemNode* mdef = new_base->as_MergeMem();
4108 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4109 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4110 mms.set_memory(mms.memory2());
4111 }
4112 assert(base_memory() == mdef->base_memory(), "");
4113 } else {
4114 set_base_memory(new_base);
4115 }
4116 }
4117
4118 // Make a new, untransformed MergeMem with the same base as 'mem'.
4119 // If mem is itself a MergeMem, populate the result with the same edges.
4120 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
4121 return new(C) MergeMemNode(mem);
4122 }
4123
4124 //------------------------------cmp--------------------------------------------
4125 uint MergeMemNode::hash() const { return NO_HASH; }
4126 uint MergeMemNode::cmp( const Node &n ) const {
4127 return (&n == this); // Always fail except on self
4128 }
4129
4130 //------------------------------Identity---------------------------------------
4131 Node* MergeMemNode::Identity(PhaseTransform *phase) {
4132 // Identity if this merge point does not record any interesting memory
4133 // disambiguations.
4134 Node* base_mem = base_memory();
4135 Node* empty_mem = empty_memory();
4136 if (base_mem != empty_mem) { // Memory path is not dead?
4137 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4138 Node* mem = in(i);
4139 if (mem != empty_mem && mem != base_mem) {
4140 return this; // Many memory splits; no change
4141 }
4142 }
4143 }
4144 return base_mem; // No memory splits; ID on the one true input
4145 }
4146
4147 //------------------------------Ideal------------------------------------------
4148 // This method is invoked recursively on chains of MergeMem nodes
4149 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4150 // Remove chain'd MergeMems
4151 //
4152 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4153 // relative to the "in(Bot)". Since we are patching both at the same time,
4154 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4155 // but rewrite each "in(i)" relative to the new "in(Bot)".
4156 Node *progress = NULL;
4157
4158
4159 Node* old_base = base_memory();
4160 Node* empty_mem = empty_memory();
4161 if (old_base == empty_mem)
4162 return NULL; // Dead memory path.
4163
4164 MergeMemNode* old_mbase;
4165 if (old_base != NULL && old_base->is_MergeMem())
4166 old_mbase = old_base->as_MergeMem();
4167 else
4168 old_mbase = NULL;
4169 Node* new_base = old_base;
4170
4171 // simplify stacked MergeMems in base memory
4172 if (old_mbase) new_base = old_mbase->base_memory();
4173
4174 // the base memory might contribute new slices beyond my req()
4175 if (old_mbase) grow_to_match(old_mbase);
4176
4177 // Look carefully at the base node if it is a phi.
4178 PhiNode* phi_base;
4179 if (new_base != NULL && new_base->is_Phi())
4180 phi_base = new_base->as_Phi();
4181 else
4182 phi_base = NULL;
4183
4184 Node* phi_reg = NULL;
4185 uint phi_len = (uint)-1;
4186 if (phi_base != NULL && !phi_base->is_copy()) {
4187 // do not examine phi if degraded to a copy
4188 phi_reg = phi_base->region();
4189 phi_len = phi_base->req();
4190 // see if the phi is unfinished
4191 for (uint i = 1; i < phi_len; i++) {
4192 if (phi_base->in(i) == NULL) {
4193 // incomplete phi; do not look at it yet!
4194 phi_reg = NULL;
4195 phi_len = (uint)-1;
4196 break;
4197 }
4198 }
4199 }
4200
4201 // Note: We do not call verify_sparse on entry, because inputs
4202 // can normalize to the base_memory via subsume_node or similar
4203 // mechanisms. This method repairs that damage.
4204
4205 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4206
4207 // Look at each slice.
4208 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4209 Node* old_in = in(i);
4210 // calculate the old memory value
4211 Node* old_mem = old_in;
4212 if (old_mem == empty_mem) old_mem = old_base;
4213 assert(old_mem == memory_at(i), "");
4214
4215 // maybe update (reslice) the old memory value
4216
4217 // simplify stacked MergeMems
4218 Node* new_mem = old_mem;
4219 MergeMemNode* old_mmem;
4220 if (old_mem != NULL && old_mem->is_MergeMem())
4221 old_mmem = old_mem->as_MergeMem();
4222 else
4223 old_mmem = NULL;
4224 if (old_mmem == this) {
4225 // This can happen if loops break up and safepoints disappear.
4226 // A merge of BotPtr (default) with a RawPtr memory derived from a
4227 // safepoint can be rewritten to a merge of the same BotPtr with
4228 // the BotPtr phi coming into the loop. If that phi disappears
4229 // also, we can end up with a self-loop of the mergemem.
4230 // In general, if loops degenerate and memory effects disappear,
4231 // a mergemem can be left looking at itself. This simply means
4232 // that the mergemem's default should be used, since there is
4233 // no longer any apparent effect on this slice.
4234 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4235 // from start. Update the input to TOP.
4236 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4237 }
4238 else if (old_mmem != NULL) {
4239 new_mem = old_mmem->memory_at(i);
4240 }
4241 // else preceding memory was not a MergeMem
4242
4243 // replace equivalent phis (unfortunately, they do not GVN together)
4244 if (new_mem != NULL && new_mem != new_base &&
4245 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4246 if (new_mem->is_Phi()) {
4247 PhiNode* phi_mem = new_mem->as_Phi();
4248 for (uint i = 1; i < phi_len; i++) {
4249 if (phi_base->in(i) != phi_mem->in(i)) {
4250 phi_mem = NULL;
4251 break;
4252 }
4253 }
4254 if (phi_mem != NULL) {
4255 // equivalent phi nodes; revert to the def
4256 new_mem = new_base;
4257 }
4258 }
4259 }
4260
4261 // maybe store down a new value
4262 Node* new_in = new_mem;
4263 if (new_in == new_base) new_in = empty_mem;
4264
4265 if (new_in != old_in) {
4266 // Warning: Do not combine this "if" with the previous "if"
4267 // A memory slice might have be be rewritten even if it is semantically
4268 // unchanged, if the base_memory value has changed.
4269 set_req(i, new_in);
4270 progress = this; // Report progress
4271 }
4272 }
4273
4274 if (new_base != old_base) {
4275 set_req(Compile::AliasIdxBot, new_base);
4276 // Don't use set_base_memory(new_base), because we need to update du.
4277 assert(base_memory() == new_base, "");
4278 progress = this;
4279 }
4280
4281 if( base_memory() == this ) {
4282 // a self cycle indicates this memory path is dead
4283 set_req(Compile::AliasIdxBot, empty_mem);
4284 }
4285
4286 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4287 // Recursion must occur after the self cycle check above
4288 if( base_memory()->is_MergeMem() ) {
4289 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4290 Node *m = phase->transform(new_mbase); // Rollup any cycles
4291 if( m != NULL && (m->is_top() ||
4292 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
4293 // propagate rollup of dead cycle to self
4294 set_req(Compile::AliasIdxBot, empty_mem);
4295 }
4296 }
4297
4298 if( base_memory() == empty_mem ) {
4299 progress = this;
4300 // Cut inputs during Parse phase only.
4301 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4302 if( !can_reshape ) {
4303 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4304 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4305 }
4306 }
4307 }
4308
4309 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4310 // Check if PhiNode::Ideal's "Split phis through memory merges"
4311 // transform should be attempted. Look for this->phi->this cycle.
4312 uint merge_width = req();
4313 if (merge_width > Compile::AliasIdxRaw) {
4314 PhiNode* phi = base_memory()->as_Phi();
4315 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4316 if (phi->in(i) == this) {
4317 phase->is_IterGVN()->_worklist.push(phi);
4318 break;
4319 }
4320 }
4321 }
4322 }
4323
4324 assert(progress || verify_sparse(), "please, no dups of base");
4325 return progress;
4326 }
4327
4328 //-------------------------set_base_memory-------------------------------------
4329 void MergeMemNode::set_base_memory(Node *new_base) {
4330 Node* empty_mem = empty_memory();
4331 set_req(Compile::AliasIdxBot, new_base);
4332 assert(memory_at(req()) == new_base, "must set default memory");
4333 // Clear out other occurrences of new_base:
4334 if (new_base != empty_mem) {
4335 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4336 if (in(i) == new_base) set_req(i, empty_mem);
4337 }
4338 }
4339 }
4340
4341 //------------------------------out_RegMask------------------------------------
4342 const RegMask &MergeMemNode::out_RegMask() const {
4343 return RegMask::Empty;
4344 }
4345
4346 //------------------------------dump_spec--------------------------------------
4347 #ifndef PRODUCT
4348 void MergeMemNode::dump_spec(outputStream *st) const {
4349 st->print(" {");
4350 Node* base_mem = base_memory();
4351 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4352 Node* mem = memory_at(i);
4353 if (mem == base_mem) { st->print(" -"); continue; }
4354 st->print( " N%d:", mem->_idx );
4355 Compile::current()->get_adr_type(i)->dump_on(st);
4356 }
4357 st->print(" }");
4358 }
4359 #endif // !PRODUCT
4360
4361
4362 #ifdef ASSERT
4363 static bool might_be_same(Node* a, Node* b) {
4364 if (a == b) return true;
4365 if (!(a->is_Phi() || b->is_Phi())) return false;
4366 // phis shift around during optimization
4367 return true; // pretty stupid...
4368 }
4369
4370 // verify a narrow slice (either incoming or outgoing)
4371 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4372 if (!VerifyAliases) return; // don't bother to verify unless requested
4373 if (is_error_reported()) return; // muzzle asserts when debugging an error
4374 if (Node::in_dump()) return; // muzzle asserts when printing
4375 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4376 assert(n != NULL, "");
4377 // Elide intervening MergeMem's
4378 while (n->is_MergeMem()) {
4379 n = n->as_MergeMem()->memory_at(alias_idx);
4380 }
4381 Compile* C = Compile::current();
4382 const TypePtr* n_adr_type = n->adr_type();
4383 if (n == m->empty_memory()) {
4384 // Implicit copy of base_memory()
4385 } else if (n_adr_type != TypePtr::BOTTOM) {
4386 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4387 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4388 } else {
4389 // A few places like make_runtime_call "know" that VM calls are narrow,
4390 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4391 bool expected_wide_mem = false;
4392 if (n == m->base_memory()) {
4393 expected_wide_mem = true;
4394 } else if (alias_idx == Compile::AliasIdxRaw ||
4395 n == m->memory_at(Compile::AliasIdxRaw)) {
4396 expected_wide_mem = true;
4397 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4398 // memory can "leak through" calls on channels that
4399 // are write-once. Allow this also.
4400 expected_wide_mem = true;
4401 }
4402 assert(expected_wide_mem, "expected narrow slice replacement");
4403 }
4404 }
4405 #else // !ASSERT
4406 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4407 #endif
4408
4409
4410 //-----------------------------memory_at---------------------------------------
4411 Node* MergeMemNode::memory_at(uint alias_idx) const {
4412 assert(alias_idx >= Compile::AliasIdxRaw ||
4413 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4414 "must avoid base_memory and AliasIdxTop");
4415
4416 // Otherwise, it is a narrow slice.
4417 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4418 Compile *C = Compile::current();
4419 if (is_empty_memory(n)) {
4420 // the array is sparse; empty slots are the "top" node
4421 n = base_memory();
4422 assert(Node::in_dump()
4423 || n == NULL || n->bottom_type() == Type::TOP
4424 || n->adr_type() == NULL // address is TOP
4425 || n->adr_type() == TypePtr::BOTTOM
4426 || n->adr_type() == TypeRawPtr::BOTTOM
4427 || Compile::current()->AliasLevel() == 0,
4428 "must be a wide memory");
4429 // AliasLevel == 0 if we are organizing the memory states manually.
4430 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4431 } else {
4432 // make sure the stored slice is sane
4433 #ifdef ASSERT
4434 if (is_error_reported() || Node::in_dump()) {
4435 } else if (might_be_same(n, base_memory())) {
4436 // Give it a pass: It is a mostly harmless repetition of the base.
4437 // This can arise normally from node subsumption during optimization.
4438 } else {
4439 verify_memory_slice(this, alias_idx, n);
4440 }
4441 #endif
4442 }
4443 return n;
4444 }
4445
4446 //---------------------------set_memory_at-------------------------------------
4447 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4448 verify_memory_slice(this, alias_idx, n);
4449 Node* empty_mem = empty_memory();
4450 if (n == base_memory()) n = empty_mem; // collapse default
4451 uint need_req = alias_idx+1;
4452 if (req() < need_req) {
4453 if (n == empty_mem) return; // already the default, so do not grow me
4454 // grow the sparse array
4455 do {
4456 add_req(empty_mem);
4457 } while (req() < need_req);
4458 }
4459 set_req( alias_idx, n );
4460 }
4461
4462
4463
4464 //--------------------------iteration_setup------------------------------------
4465 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4466 if (other != NULL) {
4467 grow_to_match(other);
4468 // invariant: the finite support of mm2 is within mm->req()
4469 #ifdef ASSERT
4470 for (uint i = req(); i < other->req(); i++) {
4471 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4472 }
4473 #endif
4474 }
4475 // Replace spurious copies of base_memory by top.
4476 Node* base_mem = base_memory();
4477 if (base_mem != NULL && !base_mem->is_top()) {
4478 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4479 if (in(i) == base_mem)
4480 set_req(i, empty_memory());
4481 }
4482 }
4483 }
4484
4485 //---------------------------grow_to_match-------------------------------------
4486 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4487 Node* empty_mem = empty_memory();
4488 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4489 // look for the finite support of the other memory
4490 for (uint i = other->req(); --i >= req(); ) {
4491 if (other->in(i) != empty_mem) {
4492 uint new_len = i+1;
4493 while (req() < new_len) add_req(empty_mem);
4494 break;
4495 }
4496 }
4497 }
4498
4499 //---------------------------verify_sparse-------------------------------------
4500 #ifndef PRODUCT
4501 bool MergeMemNode::verify_sparse() const {
4502 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4503 Node* base_mem = base_memory();
4504 // The following can happen in degenerate cases, since empty==top.
4505 if (is_empty_memory(base_mem)) return true;
4506 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4507 assert(in(i) != NULL, "sane slice");
4508 if (in(i) == base_mem) return false; // should have been the sentinel value!
4509 }
4510 return true;
4511 }
4512
4513 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4514 Node* n;
4515 n = mm->in(idx);
4516 if (mem == n) return true; // might be empty_memory()
4517 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4518 if (mem == n) return true;
4519 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4520 if (mem == n) return true;
4521 if (n == NULL) break;
4522 }
4523 return false;
4524 }
4525 #endif // !PRODUCT