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 LoadDNode* LoadDNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo) {
942 bool require_atomic = true;
943 return new (C) LoadDNode(ctl, mem, adr, adr_type, rt, mo, require_atomic);
944 }
945
946
947
948 //------------------------------hash-------------------------------------------
949 uint LoadNode::hash() const {
950 // unroll addition of interesting fields
951 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
952 }
953
954 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
955 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
956 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
957 bool is_stable_ary = FoldStableValues &&
958 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
959 tp->isa_aryptr()->is_stable();
960
961 return (eliminate_boxing && non_volatile) || is_stable_ary;
962 }
963
964 return false;
965 }
966
967 //---------------------------can_see_stored_value------------------------------
968 // This routine exists to make sure this set of tests is done the same
969 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
970 // will change the graph shape in a way which makes memory alive twice at the
971 // same time (uses the Oracle model of aliasing), then some
972 // LoadXNode::Identity will fold things back to the equivalence-class model
973 // of aliasing.
974 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
975 Node* ld_adr = in(MemNode::Address);
976 intptr_t ld_off = 0;
977 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
978 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
979 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
980 // This is more general than load from boxing objects.
981 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
982 uint alias_idx = atp->index();
983 bool final = !atp->is_rewritable();
984 Node* result = NULL;
985 Node* current = st;
986 // Skip through chains of MemBarNodes checking the MergeMems for
987 // new states for the slice of this load. Stop once any other
988 // kind of node is encountered. Loads from final memory can skip
989 // through any kind of MemBar but normal loads shouldn't skip
990 // through MemBarAcquire since the could allow them to move out of
991 // a synchronized region.
992 while (current->is_Proj()) {
993 int opc = current->in(0)->Opcode();
994 if ((final && (opc == Op_MemBarAcquire ||
995 opc == Op_MemBarAcquireLock ||
996 opc == Op_LoadFence)) ||
997 opc == Op_MemBarRelease ||
998 opc == Op_StoreFence ||
999 opc == Op_MemBarReleaseLock ||
1000 opc == Op_MemBarCPUOrder) {
1001 Node* mem = current->in(0)->in(TypeFunc::Memory);
1002 if (mem->is_MergeMem()) {
1003 MergeMemNode* merge = mem->as_MergeMem();
1004 Node* new_st = merge->memory_at(alias_idx);
1005 if (new_st == merge->base_memory()) {
1006 // Keep searching
1007 current = new_st;
1008 continue;
1009 }
1010 // Save the new memory state for the slice and fall through
1011 // to exit.
1012 result = new_st;
1013 }
1014 }
1015 break;
1016 }
1017 if (result != NULL) {
1018 st = result;
1019 }
1020 }
1021
1022 // Loop around twice in the case Load -> Initialize -> Store.
1023 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1024 for (int trip = 0; trip <= 1; trip++) {
1025
1026 if (st->is_Store()) {
1027 Node* st_adr = st->in(MemNode::Address);
1028 if (!phase->eqv(st_adr, ld_adr)) {
1029 // Try harder before giving up... Match raw and non-raw pointers.
1030 intptr_t st_off = 0;
1031 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
1032 if (alloc == NULL) return NULL;
1033 if (alloc != ld_alloc) return NULL;
1034 if (ld_off != st_off) return NULL;
1035 // At this point we have proven something like this setup:
1036 // A = Allocate(...)
1037 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
1038 // S = StoreQ(, AddP(, A.Parm , #Off), V)
1039 // (Actually, we haven't yet proven the Q's are the same.)
1040 // In other words, we are loading from a casted version of
1041 // the same pointer-and-offset that we stored to.
1042 // Thus, we are able to replace L by V.
1043 }
1044 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1045 if (store_Opcode() != st->Opcode())
1046 return NULL;
1047 return st->in(MemNode::ValueIn);
1048 }
1049
1050 // A load from a freshly-created object always returns zero.
1051 // (This can happen after LoadNode::Ideal resets the load's memory input
1052 // to find_captured_store, which returned InitializeNode::zero_memory.)
1053 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1054 (st->in(0) == ld_alloc) &&
1055 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1056 // return a zero value for the load's basic type
1057 // (This is one of the few places where a generic PhaseTransform
1058 // can create new nodes. Think of it as lazily manifesting
1059 // virtually pre-existing constants.)
1060 return phase->zerocon(memory_type());
1061 }
1062
1063 // A load from an initialization barrier can match a captured store.
1064 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1065 InitializeNode* init = st->in(0)->as_Initialize();
1066 AllocateNode* alloc = init->allocation();
1067 if ((alloc != NULL) && (alloc == ld_alloc)) {
1068 // examine a captured store value
1069 st = init->find_captured_store(ld_off, memory_size(), phase);
1070 if (st != NULL)
1071 continue; // take one more trip around
1072 }
1073 }
1074
1075 // Load boxed value from result of valueOf() call is input parameter.
1076 if (this->is_Load() && ld_adr->is_AddP() &&
1077 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1078 intptr_t ignore = 0;
1079 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1080 if (base != NULL && base->is_Proj() &&
1081 base->as_Proj()->_con == TypeFunc::Parms &&
1082 base->in(0)->is_CallStaticJava() &&
1083 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1084 return base->in(0)->in(TypeFunc::Parms);
1085 }
1086 }
1087
1088 break;
1089 }
1090
1091 return NULL;
1092 }
1093
1094 //----------------------is_instance_field_load_with_local_phi------------------
1095 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1096 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1097 in(Address)->is_AddP() ) {
1098 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1099 // Only instances and boxed values.
1100 if( t_oop != NULL &&
1101 (t_oop->is_ptr_to_boxed_value() ||
1102 t_oop->is_known_instance_field()) &&
1103 t_oop->offset() != Type::OffsetBot &&
1104 t_oop->offset() != Type::OffsetTop) {
1105 return true;
1106 }
1107 }
1108 return false;
1109 }
1110
1111 //------------------------------Identity---------------------------------------
1112 // Loads are identity if previous store is to same address
1113 Node *LoadNode::Identity( PhaseTransform *phase ) {
1114 // If the previous store-maker is the right kind of Store, and the store is
1115 // to the same address, then we are equal to the value stored.
1116 Node* mem = in(Memory);
1117 Node* value = can_see_stored_value(mem, phase);
1118 if( value ) {
1119 // byte, short & char stores truncate naturally.
1120 // A load has to load the truncated value which requires
1121 // some sort of masking operation and that requires an
1122 // Ideal call instead of an Identity call.
1123 if (memory_size() < BytesPerInt) {
1124 // If the input to the store does not fit with the load's result type,
1125 // it must be truncated via an Ideal call.
1126 if (!phase->type(value)->higher_equal(phase->type(this)))
1127 return this;
1128 }
1129 // (This works even when value is a Con, but LoadNode::Value
1130 // usually runs first, producing the singleton type of the Con.)
1131 return value;
1132 }
1133
1134 // Search for an existing data phi which was generated before for the same
1135 // instance's field to avoid infinite generation of phis in a loop.
1136 Node *region = mem->in(0);
1137 if (is_instance_field_load_with_local_phi(region)) {
1138 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1139 int this_index = phase->C->get_alias_index(addr_t);
1140 int this_offset = addr_t->offset();
1141 int this_iid = addr_t->instance_id();
1142 if (!addr_t->is_known_instance() &&
1143 addr_t->is_ptr_to_boxed_value()) {
1144 // Use _idx of address base (could be Phi node) for boxed values.
1145 intptr_t ignore = 0;
1146 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1147 this_iid = base->_idx;
1148 }
1149 const Type* this_type = bottom_type();
1150 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1151 Node* phi = region->fast_out(i);
1152 if (phi->is_Phi() && phi != mem &&
1153 phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) {
1154 return phi;
1155 }
1156 }
1157 }
1158
1159 return this;
1160 }
1161
1162 // We're loading from an object which has autobox behaviour.
1163 // If this object is result of a valueOf call we'll have a phi
1164 // merging a newly allocated object and a load from the cache.
1165 // We want to replace this load with the original incoming
1166 // argument to the valueOf call.
1167 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1168 assert(phase->C->eliminate_boxing(), "sanity");
1169 intptr_t ignore = 0;
1170 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1171 if ((base == NULL) || base->is_Phi()) {
1172 // Push the loads from the phi that comes from valueOf up
1173 // through it to allow elimination of the loads and the recovery
1174 // of the original value. It is done in split_through_phi().
1175 return NULL;
1176 } else if (base->is_Load() ||
1177 base->is_DecodeN() && base->in(1)->is_Load()) {
1178 // Eliminate the load of boxed value for integer types from the cache
1179 // array by deriving the value from the index into the array.
1180 // Capture the offset of the load and then reverse the computation.
1181
1182 // Get LoadN node which loads a boxing object from 'cache' array.
1183 if (base->is_DecodeN()) {
1184 base = base->in(1);
1185 }
1186 if (!base->in(Address)->is_AddP()) {
1187 return NULL; // Complex address
1188 }
1189 AddPNode* address = base->in(Address)->as_AddP();
1190 Node* cache_base = address->in(AddPNode::Base);
1191 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1192 // Get ConP node which is static 'cache' field.
1193 cache_base = cache_base->in(1);
1194 }
1195 if ((cache_base != NULL) && cache_base->is_Con()) {
1196 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1197 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1198 Node* elements[4];
1199 int shift = exact_log2(type2aelembytes(T_OBJECT));
1200 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1201 if ((count > 0) && elements[0]->is_Con() &&
1202 ((count == 1) ||
1203 (count == 2) && elements[1]->Opcode() == Op_LShiftX &&
1204 elements[1]->in(2) == phase->intcon(shift))) {
1205 ciObjArray* array = base_type->const_oop()->as_obj_array();
1206 // Fetch the box object cache[0] at the base of the array and get its value
1207 ciInstance* box = array->obj_at(0)->as_instance();
1208 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1209 assert(ik->is_box_klass(), "sanity");
1210 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1211 if (ik->nof_nonstatic_fields() == 1) {
1212 // This should be true nonstatic_field_at requires calling
1213 // nof_nonstatic_fields so check it anyway
1214 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1215 BasicType bt = c.basic_type();
1216 // Only integer types have boxing cache.
1217 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1218 bt == T_BYTE || bt == T_SHORT ||
1219 bt == T_INT || bt == T_LONG, err_msg_res("wrong type = %s", type2name(bt)));
1220 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1221 if (cache_low != (int)cache_low) {
1222 return NULL; // should not happen since cache is array indexed by value
1223 }
1224 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1225 if (offset != (int)offset) {
1226 return NULL; // should not happen since cache is array indexed by value
1227 }
1228 // Add up all the offsets making of the address of the load
1229 Node* result = elements[0];
1230 for (int i = 1; i < count; i++) {
1231 result = phase->transform(new (phase->C) AddXNode(result, elements[i]));
1232 }
1233 // Remove the constant offset from the address and then
1234 result = phase->transform(new (phase->C) AddXNode(result, phase->MakeConX(-(int)offset)));
1235 // remove the scaling of the offset to recover the original index.
1236 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1237 // Peel the shift off directly but wrap it in a dummy node
1238 // since Ideal can't return existing nodes
1239 result = new (phase->C) RShiftXNode(result->in(1), phase->intcon(0));
1240 } else if (result->is_Add() && result->in(2)->is_Con() &&
1241 result->in(1)->Opcode() == Op_LShiftX &&
1242 result->in(1)->in(2) == phase->intcon(shift)) {
1243 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1244 // but for boxing cache access we know that X<<Z will not overflow
1245 // (there is range check) so we do this optimizatrion by hand here.
1246 Node* add_con = new (phase->C) RShiftXNode(result->in(2), phase->intcon(shift));
1247 result = new (phase->C) AddXNode(result->in(1)->in(1), phase->transform(add_con));
1248 } else {
1249 result = new (phase->C) RShiftXNode(result, phase->intcon(shift));
1250 }
1251 #ifdef _LP64
1252 if (bt != T_LONG) {
1253 result = new (phase->C) ConvL2INode(phase->transform(result));
1254 }
1255 #else
1256 if (bt == T_LONG) {
1257 result = new (phase->C) ConvI2LNode(phase->transform(result));
1258 }
1259 #endif
1260 return result;
1261 }
1262 }
1263 }
1264 }
1265 }
1266 return NULL;
1267 }
1268
1269 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1270 Node* region = phi->in(0);
1271 if (region == NULL) {
1272 return false; // Wait stable graph
1273 }
1274 uint cnt = phi->req();
1275 for (uint i = 1; i < cnt; i++) {
1276 Node* rc = region->in(i);
1277 if (rc == NULL || phase->type(rc) == Type::TOP)
1278 return false; // Wait stable graph
1279 Node* in = phi->in(i);
1280 if (in == NULL || phase->type(in) == Type::TOP)
1281 return false; // Wait stable graph
1282 }
1283 return true;
1284 }
1285 //------------------------------split_through_phi------------------------------
1286 // Split instance or boxed field load through Phi.
1287 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1288 Node* mem = in(Memory);
1289 Node* address = in(Address);
1290 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1291
1292 assert((t_oop != NULL) &&
1293 (t_oop->is_known_instance_field() ||
1294 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1295
1296 Compile* C = phase->C;
1297 intptr_t ignore = 0;
1298 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1299 bool base_is_phi = (base != NULL) && base->is_Phi();
1300 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1301 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1302 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1303
1304 if (!((mem->is_Phi() || base_is_phi) &&
1305 (load_boxed_values || t_oop->is_known_instance_field()))) {
1306 return NULL; // memory is not Phi
1307 }
1308
1309 if (mem->is_Phi()) {
1310 if (!stable_phi(mem->as_Phi(), phase)) {
1311 return NULL; // Wait stable graph
1312 }
1313 uint cnt = mem->req();
1314 // Check for loop invariant memory.
1315 if (cnt == 3) {
1316 for (uint i = 1; i < cnt; i++) {
1317 Node* in = mem->in(i);
1318 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1319 if (m == mem) {
1320 set_req(Memory, mem->in(cnt - i));
1321 return this; // made change
1322 }
1323 }
1324 }
1325 }
1326 if (base_is_phi) {
1327 if (!stable_phi(base->as_Phi(), phase)) {
1328 return NULL; // Wait stable graph
1329 }
1330 uint cnt = base->req();
1331 // Check for loop invariant memory.
1332 if (cnt == 3) {
1333 for (uint i = 1; i < cnt; i++) {
1334 if (base->in(i) == base) {
1335 return NULL; // Wait stable graph
1336 }
1337 }
1338 }
1339 }
1340
1341 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1342
1343 // Split through Phi (see original code in loopopts.cpp).
1344 assert(C->have_alias_type(t_oop), "instance should have alias type");
1345
1346 // Do nothing here if Identity will find a value
1347 // (to avoid infinite chain of value phis generation).
1348 if (!phase->eqv(this, this->Identity(phase)))
1349 return NULL;
1350
1351 // Select Region to split through.
1352 Node* region;
1353 if (!base_is_phi) {
1354 assert(mem->is_Phi(), "sanity");
1355 region = mem->in(0);
1356 // Skip if the region dominates some control edge of the address.
1357 if (!MemNode::all_controls_dominate(address, region))
1358 return NULL;
1359 } else if (!mem->is_Phi()) {
1360 assert(base_is_phi, "sanity");
1361 region = base->in(0);
1362 // Skip if the region dominates some control edge of the memory.
1363 if (!MemNode::all_controls_dominate(mem, region))
1364 return NULL;
1365 } else if (base->in(0) != mem->in(0)) {
1366 assert(base_is_phi && mem->is_Phi(), "sanity");
1367 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1368 region = base->in(0);
1369 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1370 region = mem->in(0);
1371 } else {
1372 return NULL; // complex graph
1373 }
1374 } else {
1375 assert(base->in(0) == mem->in(0), "sanity");
1376 region = mem->in(0);
1377 }
1378
1379 const Type* this_type = this->bottom_type();
1380 int this_index = C->get_alias_index(t_oop);
1381 int this_offset = t_oop->offset();
1382 int this_iid = t_oop->instance_id();
1383 if (!t_oop->is_known_instance() && load_boxed_values) {
1384 // Use _idx of address base for boxed values.
1385 this_iid = base->_idx;
1386 }
1387 PhaseIterGVN* igvn = phase->is_IterGVN();
1388 Node* phi = new (C) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1389 for (uint i = 1; i < region->req(); i++) {
1390 Node* x;
1391 Node* the_clone = NULL;
1392 if (region->in(i) == C->top()) {
1393 x = C->top(); // Dead path? Use a dead data op
1394 } else {
1395 x = this->clone(); // Else clone up the data op
1396 the_clone = x; // Remember for possible deletion.
1397 // Alter data node to use pre-phi inputs
1398 if (this->in(0) == region) {
1399 x->set_req(0, region->in(i));
1400 } else {
1401 x->set_req(0, NULL);
1402 }
1403 if (mem->is_Phi() && (mem->in(0) == region)) {
1404 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1405 }
1406 if (address->is_Phi() && address->in(0) == region) {
1407 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1408 }
1409 if (base_is_phi && (base->in(0) == region)) {
1410 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1411 Node* adr_x = phase->transform(new (C) AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1412 x->set_req(Address, adr_x);
1413 }
1414 }
1415 // Check for a 'win' on some paths
1416 const Type *t = x->Value(igvn);
1417
1418 bool singleton = t->singleton();
1419
1420 // See comments in PhaseIdealLoop::split_thru_phi().
1421 if (singleton && t == Type::TOP) {
1422 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1423 }
1424
1425 if (singleton) {
1426 x = igvn->makecon(t);
1427 } else {
1428 // We now call Identity to try to simplify the cloned node.
1429 // Note that some Identity methods call phase->type(this).
1430 // Make sure that the type array is big enough for
1431 // our new node, even though we may throw the node away.
1432 // (This tweaking with igvn only works because x is a new node.)
1433 igvn->set_type(x, t);
1434 // If x is a TypeNode, capture any more-precise type permanently into Node
1435 // otherwise it will be not updated during igvn->transform since
1436 // igvn->type(x) is set to x->Value() already.
1437 x->raise_bottom_type(t);
1438 Node *y = x->Identity(igvn);
1439 if (y != x) {
1440 x = y;
1441 } else {
1442 y = igvn->hash_find_insert(x);
1443 if (y) {
1444 x = y;
1445 } else {
1446 // Else x is a new node we are keeping
1447 // We do not need register_new_node_with_optimizer
1448 // because set_type has already been called.
1449 igvn->_worklist.push(x);
1450 }
1451 }
1452 }
1453 if (x != the_clone && the_clone != NULL) {
1454 igvn->remove_dead_node(the_clone);
1455 }
1456 phi->set_req(i, x);
1457 }
1458 // Record Phi
1459 igvn->register_new_node_with_optimizer(phi);
1460 return phi;
1461 }
1462
1463 //------------------------------Ideal------------------------------------------
1464 // If the load is from Field memory and the pointer is non-null, we can
1465 // zero out the control input.
1466 // If the offset is constant and the base is an object allocation,
1467 // try to hook me up to the exact initializing store.
1468 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1469 Node* p = MemNode::Ideal_common(phase, can_reshape);
1470 if (p) return (p == NodeSentinel) ? NULL : p;
1471
1472 Node* ctrl = in(MemNode::Control);
1473 Node* address = in(MemNode::Address);
1474
1475 // Skip up past a SafePoint control. Cannot do this for Stores because
1476 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1477 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1478 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1479 ctrl = ctrl->in(0);
1480 set_req(MemNode::Control,ctrl);
1481 }
1482
1483 intptr_t ignore = 0;
1484 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1485 if (base != NULL
1486 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1487 // Check for useless control edge in some common special cases
1488 if (in(MemNode::Control) != NULL
1489 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1490 && all_controls_dominate(base, phase->C->start())) {
1491 // A method-invariant, non-null address (constant or 'this' argument).
1492 set_req(MemNode::Control, NULL);
1493 }
1494 }
1495
1496 Node* mem = in(MemNode::Memory);
1497 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1498
1499 if (can_reshape && (addr_t != NULL)) {
1500 // try to optimize our memory input
1501 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1502 if (opt_mem != mem) {
1503 set_req(MemNode::Memory, opt_mem);
1504 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1505 return this;
1506 }
1507 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1508 if ((t_oop != NULL) &&
1509 (t_oop->is_known_instance_field() ||
1510 t_oop->is_ptr_to_boxed_value())) {
1511 PhaseIterGVN *igvn = phase->is_IterGVN();
1512 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1513 // Delay this transformation until memory Phi is processed.
1514 phase->is_IterGVN()->_worklist.push(this);
1515 return NULL;
1516 }
1517 // Split instance field load through Phi.
1518 Node* result = split_through_phi(phase);
1519 if (result != NULL) return result;
1520
1521 if (t_oop->is_ptr_to_boxed_value()) {
1522 Node* result = eliminate_autobox(phase);
1523 if (result != NULL) return result;
1524 }
1525 }
1526 }
1527
1528 // Check for prior store with a different base or offset; make Load
1529 // independent. Skip through any number of them. Bail out if the stores
1530 // are in an endless dead cycle and report no progress. This is a key
1531 // transform for Reflection. However, if after skipping through the Stores
1532 // we can't then fold up against a prior store do NOT do the transform as
1533 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1534 // array memory alive twice: once for the hoisted Load and again after the
1535 // bypassed Store. This situation only works if EVERYBODY who does
1536 // anti-dependence work knows how to bypass. I.e. we need all
1537 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1538 // the alias index stuff. So instead, peek through Stores and IFF we can
1539 // fold up, do so.
1540 Node* prev_mem = find_previous_store(phase);
1541 // Steps (a), (b): Walk past independent stores to find an exact match.
1542 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1543 // (c) See if we can fold up on the spot, but don't fold up here.
1544 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1545 // just return a prior value, which is done by Identity calls.
1546 if (can_see_stored_value(prev_mem, phase)) {
1547 // Make ready for step (d):
1548 set_req(MemNode::Memory, prev_mem);
1549 return this;
1550 }
1551 }
1552
1553 return NULL; // No further progress
1554 }
1555
1556 // Helper to recognize certain Klass fields which are invariant across
1557 // some group of array types (e.g., int[] or all T[] where T < Object).
1558 const Type*
1559 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1560 ciKlass* klass) const {
1561 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1562 // The field is Klass::_modifier_flags. Return its (constant) value.
1563 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1564 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1565 return TypeInt::make(klass->modifier_flags());
1566 }
1567 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1568 // The field is Klass::_access_flags. Return its (constant) value.
1569 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1570 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1571 return TypeInt::make(klass->access_flags());
1572 }
1573 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1574 // The field is Klass::_layout_helper. Return its constant value if known.
1575 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1576 return TypeInt::make(klass->layout_helper());
1577 }
1578
1579 // No match.
1580 return NULL;
1581 }
1582
1583 // Try to constant-fold a stable array element.
1584 static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) {
1585 assert(ary->const_oop(), "array should be constant");
1586 assert(ary->is_stable(), "array should be stable");
1587
1588 // Decode the results of GraphKit::array_element_address.
1589 ciArray* aobj = ary->const_oop()->as_array();
1590 ciConstant con = aobj->element_value_by_offset(off);
1591
1592 if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) {
1593 const Type* con_type = Type::make_from_constant(con);
1594 if (con_type != NULL) {
1595 if (con_type->isa_aryptr()) {
1596 // Join with the array element type, in case it is also stable.
1597 int dim = ary->stable_dimension();
1598 con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1);
1599 }
1600 if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) {
1601 con_type = con_type->make_narrowoop();
1602 }
1603 #ifndef PRODUCT
1604 if (TraceIterativeGVN) {
1605 tty->print("FoldStableValues: array element [off=%d]: con_type=", off);
1606 con_type->dump(); tty->cr();
1607 }
1608 #endif //PRODUCT
1609 return con_type;
1610 }
1611 }
1612 return NULL;
1613 }
1614
1615 //------------------------------Value-----------------------------------------
1616 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1617 // Either input is TOP ==> the result is TOP
1618 Node* mem = in(MemNode::Memory);
1619 const Type *t1 = phase->type(mem);
1620 if (t1 == Type::TOP) return Type::TOP;
1621 Node* adr = in(MemNode::Address);
1622 const TypePtr* tp = phase->type(adr)->isa_ptr();
1623 if (tp == NULL || tp->empty()) return Type::TOP;
1624 int off = tp->offset();
1625 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1626 Compile* C = phase->C;
1627
1628 // Try to guess loaded type from pointer type
1629 if (tp->isa_aryptr()) {
1630 const TypeAryPtr* ary = tp->is_aryptr();
1631 const Type* t = ary->elem();
1632
1633 // Determine whether the reference is beyond the header or not, by comparing
1634 // the offset against the offset of the start of the array's data.
1635 // Different array types begin at slightly different offsets (12 vs. 16).
1636 // We choose T_BYTE as an example base type that is least restrictive
1637 // as to alignment, which will therefore produce the smallest
1638 // possible base offset.
1639 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1640 const bool off_beyond_header = ((uint)off >= (uint)min_base_off);
1641
1642 // Try to constant-fold a stable array element.
1643 if (FoldStableValues && ary->is_stable() && ary->const_oop() != NULL) {
1644 // Make sure the reference is not into the header and the offset is constant
1645 if (off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1646 const Type* con_type = fold_stable_ary_elem(ary, off, memory_type());
1647 if (con_type != NULL) {
1648 return con_type;
1649 }
1650 }
1651 }
1652
1653 // Don't do this for integer types. There is only potential profit if
1654 // the element type t is lower than _type; that is, for int types, if _type is
1655 // more restrictive than t. This only happens here if one is short and the other
1656 // char (both 16 bits), and in those cases we've made an intentional decision
1657 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1658 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1659 //
1660 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1661 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1662 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1663 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1664 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1665 // In fact, that could have been the original type of p1, and p1 could have
1666 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1667 // expression (LShiftL quux 3) independently optimized to the constant 8.
1668 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1669 && (_type->isa_vect() == NULL)
1670 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1671 // t might actually be lower than _type, if _type is a unique
1672 // concrete subclass of abstract class t.
1673 if (off_beyond_header) { // is the offset beyond the header?
1674 const Type* jt = t->join_speculative(_type);
1675 // In any case, do not allow the join, per se, to empty out the type.
1676 if (jt->empty() && !t->empty()) {
1677 // This can happen if a interface-typed array narrows to a class type.
1678 jt = _type;
1679 }
1680 #ifdef ASSERT
1681 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1682 // The pointers in the autobox arrays are always non-null
1683 Node* base = adr->in(AddPNode::Base);
1684 if ((base != NULL) && base->is_DecodeN()) {
1685 // Get LoadN node which loads IntegerCache.cache field
1686 base = base->in(1);
1687 }
1688 if ((base != NULL) && base->is_Con()) {
1689 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1690 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1691 // It could be narrow oop
1692 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1693 }
1694 }
1695 }
1696 #endif
1697 return jt;
1698 }
1699 }
1700 } else if (tp->base() == Type::InstPtr) {
1701 ciEnv* env = C->env();
1702 const TypeInstPtr* tinst = tp->is_instptr();
1703 ciKlass* klass = tinst->klass();
1704 assert( off != Type::OffsetBot ||
1705 // arrays can be cast to Objects
1706 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1707 // unsafe field access may not have a constant offset
1708 C->has_unsafe_access(),
1709 "Field accesses must be precise" );
1710 // For oop loads, we expect the _type to be precise
1711 if (klass == env->String_klass() &&
1712 adr->is_AddP() && off != Type::OffsetBot) {
1713 // For constant Strings treat the final fields as compile time constants.
1714 Node* base = adr->in(AddPNode::Base);
1715 const TypeOopPtr* t = phase->type(base)->isa_oopptr();
1716 if (t != NULL && t->singleton()) {
1717 ciField* field = env->String_klass()->get_field_by_offset(off, false);
1718 if (field != NULL && field->is_final()) {
1719 ciObject* string = t->const_oop();
1720 ciConstant constant = string->as_instance()->field_value(field);
1721 if (constant.basic_type() == T_INT) {
1722 return TypeInt::make(constant.as_int());
1723 } else if (constant.basic_type() == T_ARRAY) {
1724 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1725 return TypeNarrowOop::make_from_constant(constant.as_object(), true);
1726 } else {
1727 return TypeOopPtr::make_from_constant(constant.as_object(), true);
1728 }
1729 }
1730 }
1731 }
1732 }
1733 // Optimizations for constant objects
1734 ciObject* const_oop = tinst->const_oop();
1735 if (const_oop != NULL) {
1736 // For constant Boxed value treat the target field as a compile time constant.
1737 if (tinst->is_ptr_to_boxed_value()) {
1738 return tinst->get_const_boxed_value();
1739 } else
1740 // For constant CallSites treat the target field as a compile time constant.
1741 if (const_oop->is_call_site()) {
1742 ciCallSite* call_site = const_oop->as_call_site();
1743 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);
1744 if (field != NULL && field->is_call_site_target()) {
1745 ciMethodHandle* target = call_site->get_target();
1746 if (target != NULL) { // just in case
1747 ciConstant constant(T_OBJECT, target);
1748 const Type* t;
1749 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1750 t = TypeNarrowOop::make_from_constant(constant.as_object(), true);
1751 } else {
1752 t = TypeOopPtr::make_from_constant(constant.as_object(), true);
1753 }
1754 // Add a dependence for invalidation of the optimization.
1755 if (!call_site->is_constant_call_site()) {
1756 C->dependencies()->assert_call_site_target_value(call_site, target);
1757 }
1758 return t;
1759 }
1760 }
1761 }
1762 }
1763 } else if (tp->base() == Type::KlassPtr) {
1764 assert( off != Type::OffsetBot ||
1765 // arrays can be cast to Objects
1766 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1767 // also allow array-loading from the primary supertype
1768 // array during subtype checks
1769 Opcode() == Op_LoadKlass,
1770 "Field accesses must be precise" );
1771 // For klass/static loads, we expect the _type to be precise
1772 }
1773
1774 const TypeKlassPtr *tkls = tp->isa_klassptr();
1775 if (tkls != NULL && !StressReflectiveCode) {
1776 ciKlass* klass = tkls->klass();
1777 if (klass->is_loaded() && tkls->klass_is_exact()) {
1778 // We are loading a field from a Klass metaobject whose identity
1779 // is known at compile time (the type is "exact" or "precise").
1780 // Check for fields we know are maintained as constants by the VM.
1781 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1782 // The field is Klass::_super_check_offset. Return its (constant) value.
1783 // (Folds up type checking code.)
1784 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1785 return TypeInt::make(klass->super_check_offset());
1786 }
1787 // Compute index into primary_supers array
1788 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1789 // Check for overflowing; use unsigned compare to handle the negative case.
1790 if( depth < ciKlass::primary_super_limit() ) {
1791 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1792 // (Folds up type checking code.)
1793 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1794 ciKlass *ss = klass->super_of_depth(depth);
1795 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1796 }
1797 const Type* aift = load_array_final_field(tkls, klass);
1798 if (aift != NULL) return aift;
1799 if (tkls->offset() == in_bytes(ArrayKlass::component_mirror_offset())
1800 && klass->is_array_klass()) {
1801 // The field is ArrayKlass::_component_mirror. Return its (constant) value.
1802 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1803 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1804 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1805 }
1806 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1807 // The field is Klass::_java_mirror. Return its (constant) value.
1808 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1809 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1810 return TypeInstPtr::make(klass->java_mirror());
1811 }
1812 }
1813
1814 // We can still check if we are loading from the primary_supers array at a
1815 // shallow enough depth. Even though the klass is not exact, entries less
1816 // than or equal to its super depth are correct.
1817 if (klass->is_loaded() ) {
1818 ciType *inner = klass;
1819 while( inner->is_obj_array_klass() )
1820 inner = inner->as_obj_array_klass()->base_element_type();
1821 if( inner->is_instance_klass() &&
1822 !inner->as_instance_klass()->flags().is_interface() ) {
1823 // Compute index into primary_supers array
1824 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1825 // Check for overflowing; use unsigned compare to handle the negative case.
1826 if( depth < ciKlass::primary_super_limit() &&
1827 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1828 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1829 // (Folds up type checking code.)
1830 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1831 ciKlass *ss = klass->super_of_depth(depth);
1832 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1833 }
1834 }
1835 }
1836
1837 // If the type is enough to determine that the thing is not an array,
1838 // we can give the layout_helper a positive interval type.
1839 // This will help short-circuit some reflective code.
1840 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1841 && !klass->is_array_klass() // not directly typed as an array
1842 && !klass->is_interface() // specifically not Serializable & Cloneable
1843 && !klass->is_java_lang_Object() // not the supertype of all T[]
1844 ) {
1845 // Note: When interfaces are reliable, we can narrow the interface
1846 // test to (klass != Serializable && klass != Cloneable).
1847 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1848 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1849 // The key property of this type is that it folds up tests
1850 // for array-ness, since it proves that the layout_helper is positive.
1851 // Thus, a generic value like the basic object layout helper works fine.
1852 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1853 }
1854 }
1855
1856 // If we are loading from a freshly-allocated object, produce a zero,
1857 // if the load is provably beyond the header of the object.
1858 // (Also allow a variable load from a fresh array to produce zero.)
1859 const TypeOopPtr *tinst = tp->isa_oopptr();
1860 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1861 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1862 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1863 Node* value = can_see_stored_value(mem,phase);
1864 if (value != NULL && value->is_Con()) {
1865 assert(value->bottom_type()->higher_equal(_type),"sanity");
1866 return value->bottom_type();
1867 }
1868 }
1869
1870 if (is_instance) {
1871 // If we have an instance type and our memory input is the
1872 // programs's initial memory state, there is no matching store,
1873 // so just return a zero of the appropriate type
1874 Node *mem = in(MemNode::Memory);
1875 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1876 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1877 return Type::get_zero_type(_type->basic_type());
1878 }
1879 }
1880 return _type;
1881 }
1882
1883 //------------------------------match_edge-------------------------------------
1884 // Do we Match on this edge index or not? Match only the address.
1885 uint LoadNode::match_edge(uint idx) const {
1886 return idx == MemNode::Address;
1887 }
1888
1889 //--------------------------LoadBNode::Ideal--------------------------------------
1890 //
1891 // If the previous store is to the same address as this load,
1892 // and the value stored was larger than a byte, replace this load
1893 // with the value stored truncated to a byte. If no truncation is
1894 // needed, the replacement is done in LoadNode::Identity().
1895 //
1896 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1897 Node* mem = in(MemNode::Memory);
1898 Node* value = can_see_stored_value(mem,phase);
1899 if( value && !phase->type(value)->higher_equal( _type ) ) {
1900 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(24)) );
1901 return new (phase->C) RShiftINode(result, phase->intcon(24));
1902 }
1903 // Identity call will handle the case where truncation is not needed.
1904 return LoadNode::Ideal(phase, can_reshape);
1905 }
1906
1907 const Type* LoadBNode::Value(PhaseTransform *phase) const {
1908 Node* mem = in(MemNode::Memory);
1909 Node* value = can_see_stored_value(mem,phase);
1910 if (value != NULL && value->is_Con() &&
1911 !value->bottom_type()->higher_equal(_type)) {
1912 // If the input to the store does not fit with the load's result type,
1913 // it must be truncated. We can't delay until Ideal call since
1914 // a singleton Value is needed for split_thru_phi optimization.
1915 int con = value->get_int();
1916 return TypeInt::make((con << 24) >> 24);
1917 }
1918 return LoadNode::Value(phase);
1919 }
1920
1921 //--------------------------LoadUBNode::Ideal-------------------------------------
1922 //
1923 // If the previous store is to the same address as this load,
1924 // and the value stored was larger than a byte, replace this load
1925 // with the value stored truncated to a byte. If no truncation is
1926 // needed, the replacement is done in LoadNode::Identity().
1927 //
1928 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1929 Node* mem = in(MemNode::Memory);
1930 Node* value = can_see_stored_value(mem, phase);
1931 if (value && !phase->type(value)->higher_equal(_type))
1932 return new (phase->C) AndINode(value, phase->intcon(0xFF));
1933 // Identity call will handle the case where truncation is not needed.
1934 return LoadNode::Ideal(phase, can_reshape);
1935 }
1936
1937 const Type* LoadUBNode::Value(PhaseTransform *phase) const {
1938 Node* mem = in(MemNode::Memory);
1939 Node* value = can_see_stored_value(mem,phase);
1940 if (value != NULL && value->is_Con() &&
1941 !value->bottom_type()->higher_equal(_type)) {
1942 // If the input to the store does not fit with the load's result type,
1943 // it must be truncated. We can't delay until Ideal call since
1944 // a singleton Value is needed for split_thru_phi optimization.
1945 int con = value->get_int();
1946 return TypeInt::make(con & 0xFF);
1947 }
1948 return LoadNode::Value(phase);
1949 }
1950
1951 //--------------------------LoadUSNode::Ideal-------------------------------------
1952 //
1953 // If the previous store is to the same address as this load,
1954 // and the value stored was larger than a char, replace this load
1955 // with the value stored truncated to a char. If no truncation is
1956 // needed, the replacement is done in LoadNode::Identity().
1957 //
1958 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1959 Node* mem = in(MemNode::Memory);
1960 Node* value = can_see_stored_value(mem,phase);
1961 if( value && !phase->type(value)->higher_equal( _type ) )
1962 return new (phase->C) AndINode(value,phase->intcon(0xFFFF));
1963 // Identity call will handle the case where truncation is not needed.
1964 return LoadNode::Ideal(phase, can_reshape);
1965 }
1966
1967 const Type* LoadUSNode::Value(PhaseTransform *phase) const {
1968 Node* mem = in(MemNode::Memory);
1969 Node* value = can_see_stored_value(mem,phase);
1970 if (value != NULL && value->is_Con() &&
1971 !value->bottom_type()->higher_equal(_type)) {
1972 // If the input to the store does not fit with the load's result type,
1973 // it must be truncated. We can't delay until Ideal call since
1974 // a singleton Value is needed for split_thru_phi optimization.
1975 int con = value->get_int();
1976 return TypeInt::make(con & 0xFFFF);
1977 }
1978 return LoadNode::Value(phase);
1979 }
1980
1981 //--------------------------LoadSNode::Ideal--------------------------------------
1982 //
1983 // If the previous store is to the same address as this load,
1984 // and the value stored was larger than a short, replace this load
1985 // with the value stored truncated to a short. If no truncation is
1986 // needed, the replacement is done in LoadNode::Identity().
1987 //
1988 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1989 Node* mem = in(MemNode::Memory);
1990 Node* value = can_see_stored_value(mem,phase);
1991 if( value && !phase->type(value)->higher_equal( _type ) ) {
1992 Node *result = phase->transform( new (phase->C) LShiftINode(value, phase->intcon(16)) );
1993 return new (phase->C) RShiftINode(result, phase->intcon(16));
1994 }
1995 // Identity call will handle the case where truncation is not needed.
1996 return LoadNode::Ideal(phase, can_reshape);
1997 }
1998
1999 const Type* LoadSNode::Value(PhaseTransform *phase) const {
2000 Node* mem = in(MemNode::Memory);
2001 Node* value = can_see_stored_value(mem,phase);
2002 if (value != NULL && value->is_Con() &&
2003 !value->bottom_type()->higher_equal(_type)) {
2004 // If the input to the store does not fit with the load's result type,
2005 // it must be truncated. We can't delay until Ideal call since
2006 // a singleton Value is needed for split_thru_phi optimization.
2007 int con = value->get_int();
2008 return TypeInt::make((con << 16) >> 16);
2009 }
2010 return LoadNode::Value(phase);
2011 }
2012
2013 //=============================================================================
2014 //----------------------------LoadKlassNode::make------------------------------
2015 // Polymorphic factory method:
2016 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
2017 Compile* C = gvn.C;
2018 Node *ctl = NULL;
2019 // sanity check the alias category against the created node type
2020 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2021 assert(adr_type != NULL, "expecting TypeKlassPtr");
2022 #ifdef _LP64
2023 if (adr_type->is_ptr_to_narrowklass()) {
2024 assert(UseCompressedClassPointers, "no compressed klasses");
2025 Node* load_klass = gvn.transform(new (C) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2026 return new (C) DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2027 }
2028 #endif
2029 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2030 return new (C) LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2031 }
2032
2033 //------------------------------Value------------------------------------------
2034 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
2035 return klass_value_common(phase);
2036 }
2037
2038 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
2039 // Either input is TOP ==> the result is TOP
2040 const Type *t1 = phase->type( in(MemNode::Memory) );
2041 if (t1 == Type::TOP) return Type::TOP;
2042 Node *adr = in(MemNode::Address);
2043 const Type *t2 = phase->type( adr );
2044 if (t2 == Type::TOP) return Type::TOP;
2045 const TypePtr *tp = t2->is_ptr();
2046 if (TypePtr::above_centerline(tp->ptr()) ||
2047 tp->ptr() == TypePtr::Null) return Type::TOP;
2048
2049 // Return a more precise klass, if possible
2050 const TypeInstPtr *tinst = tp->isa_instptr();
2051 if (tinst != NULL) {
2052 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2053 int offset = tinst->offset();
2054 if (ik == phase->C->env()->Class_klass()
2055 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2056 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2057 // We are loading a special hidden field from a Class mirror object,
2058 // the field which points to the VM's Klass metaobject.
2059 ciType* t = tinst->java_mirror_type();
2060 // java_mirror_type returns non-null for compile-time Class constants.
2061 if (t != NULL) {
2062 // constant oop => constant klass
2063 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2064 if (t->is_void()) {
2065 // We cannot create a void array. Since void is a primitive type return null
2066 // klass. Users of this result need to do a null check on the returned klass.
2067 return TypePtr::NULL_PTR;
2068 }
2069 return TypeKlassPtr::make(ciArrayKlass::make(t));
2070 }
2071 if (!t->is_klass()) {
2072 // a primitive Class (e.g., int.class) has NULL for a klass field
2073 return TypePtr::NULL_PTR;
2074 }
2075 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2076 return TypeKlassPtr::make(t->as_klass());
2077 }
2078 // non-constant mirror, so we can't tell what's going on
2079 }
2080 if( !ik->is_loaded() )
2081 return _type; // Bail out if not loaded
2082 if (offset == oopDesc::klass_offset_in_bytes()) {
2083 if (tinst->klass_is_exact()) {
2084 return TypeKlassPtr::make(ik);
2085 }
2086 // See if we can become precise: no subklasses and no interface
2087 // (Note: We need to support verified interfaces.)
2088 if (!ik->is_interface() && !ik->has_subklass()) {
2089 //assert(!UseExactTypes, "this code should be useless with exact types");
2090 // Add a dependence; if any subclass added we need to recompile
2091 if (!ik->is_final()) {
2092 // %%% should use stronger assert_unique_concrete_subtype instead
2093 phase->C->dependencies()->assert_leaf_type(ik);
2094 }
2095 // Return precise klass
2096 return TypeKlassPtr::make(ik);
2097 }
2098
2099 // Return root of possible klass
2100 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2101 }
2102 }
2103
2104 // Check for loading klass from an array
2105 const TypeAryPtr *tary = tp->isa_aryptr();
2106 if( tary != NULL ) {
2107 ciKlass *tary_klass = tary->klass();
2108 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2109 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2110 if (tary->klass_is_exact()) {
2111 return TypeKlassPtr::make(tary_klass);
2112 }
2113 ciArrayKlass *ak = tary->klass()->as_array_klass();
2114 // If the klass is an object array, we defer the question to the
2115 // array component klass.
2116 if( ak->is_obj_array_klass() ) {
2117 assert( ak->is_loaded(), "" );
2118 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2119 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2120 ciInstanceKlass* ik = base_k->as_instance_klass();
2121 // See if we can become precise: no subklasses and no interface
2122 if (!ik->is_interface() && !ik->has_subklass()) {
2123 //assert(!UseExactTypes, "this code should be useless with exact types");
2124 // Add a dependence; if any subclass added we need to recompile
2125 if (!ik->is_final()) {
2126 phase->C->dependencies()->assert_leaf_type(ik);
2127 }
2128 // Return precise array klass
2129 return TypeKlassPtr::make(ak);
2130 }
2131 }
2132 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2133 } else { // Found a type-array?
2134 //assert(!UseExactTypes, "this code should be useless with exact types");
2135 assert( ak->is_type_array_klass(), "" );
2136 return TypeKlassPtr::make(ak); // These are always precise
2137 }
2138 }
2139 }
2140
2141 // Check for loading klass from an array klass
2142 const TypeKlassPtr *tkls = tp->isa_klassptr();
2143 if (tkls != NULL && !StressReflectiveCode) {
2144 ciKlass* klass = tkls->klass();
2145 if( !klass->is_loaded() )
2146 return _type; // Bail out if not loaded
2147 if( klass->is_obj_array_klass() &&
2148 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2149 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2150 // // Always returning precise element type is incorrect,
2151 // // e.g., element type could be object and array may contain strings
2152 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2153
2154 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2155 // according to the element type's subclassing.
2156 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2157 }
2158 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2159 tkls->offset() == in_bytes(Klass::super_offset())) {
2160 ciKlass* sup = klass->as_instance_klass()->super();
2161 // The field is Klass::_super. Return its (constant) value.
2162 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2163 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2164 }
2165 }
2166
2167 // Bailout case
2168 return LoadNode::Value(phase);
2169 }
2170
2171 //------------------------------Identity---------------------------------------
2172 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2173 // Also feed through the klass in Allocate(...klass...)._klass.
2174 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
2175 return klass_identity_common(phase);
2176 }
2177
2178 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
2179 Node* x = LoadNode::Identity(phase);
2180 if (x != this) return x;
2181
2182 // Take apart the address into an oop and and offset.
2183 // Return 'this' if we cannot.
2184 Node* adr = in(MemNode::Address);
2185 intptr_t offset = 0;
2186 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2187 if (base == NULL) return this;
2188 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2189 if (toop == NULL) return this;
2190
2191 // We can fetch the klass directly through an AllocateNode.
2192 // This works even if the klass is not constant (clone or newArray).
2193 if (offset == oopDesc::klass_offset_in_bytes()) {
2194 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2195 if (allocated_klass != NULL) {
2196 return allocated_klass;
2197 }
2198 }
2199
2200 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2201 // Simplify ak.component_mirror.array_klass to plain ak, ak an ArrayKlass.
2202 // See inline_native_Class_query for occurrences of these patterns.
2203 // Java Example: x.getClass().isAssignableFrom(y)
2204 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
2205 //
2206 // This improves reflective code, often making the Class
2207 // mirror go completely dead. (Current exception: Class
2208 // mirrors may appear in debug info, but we could clean them out by
2209 // introducing a new debug info operator for Klass*.java_mirror).
2210 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2211 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2212 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2213 // We are loading a special hidden field from a Class mirror,
2214 // the field which points to its Klass or ArrayKlass metaobject.
2215 if (base->is_Load()) {
2216 Node* adr2 = base->in(MemNode::Address);
2217 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2218 if (tkls != NULL && !tkls->empty()
2219 && (tkls->klass()->is_instance_klass() ||
2220 tkls->klass()->is_array_klass())
2221 && adr2->is_AddP()
2222 ) {
2223 int mirror_field = in_bytes(Klass::java_mirror_offset());
2224 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2225 mirror_field = in_bytes(ArrayKlass::component_mirror_offset());
2226 }
2227 if (tkls->offset() == mirror_field) {
2228 return adr2->in(AddPNode::Base);
2229 }
2230 }
2231 }
2232 }
2233
2234 return this;
2235 }
2236
2237
2238 //------------------------------Value------------------------------------------
2239 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
2240 const Type *t = klass_value_common(phase);
2241 if (t == Type::TOP)
2242 return t;
2243
2244 return t->make_narrowklass();
2245 }
2246
2247 //------------------------------Identity---------------------------------------
2248 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2249 // Also feed through the klass in Allocate(...klass...)._klass.
2250 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
2251 Node *x = klass_identity_common(phase);
2252
2253 const Type *t = phase->type( x );
2254 if( t == Type::TOP ) return x;
2255 if( t->isa_narrowklass()) return x;
2256 assert (!t->isa_narrowoop(), "no narrow oop here");
2257
2258 return phase->transform(new (phase->C) EncodePKlassNode(x, t->make_narrowklass()));
2259 }
2260
2261 //------------------------------Value-----------------------------------------
2262 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
2263 // Either input is TOP ==> the result is TOP
2264 const Type *t1 = phase->type( in(MemNode::Memory) );
2265 if( t1 == Type::TOP ) return Type::TOP;
2266 Node *adr = in(MemNode::Address);
2267 const Type *t2 = phase->type( adr );
2268 if( t2 == Type::TOP ) return Type::TOP;
2269 const TypePtr *tp = t2->is_ptr();
2270 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2271 const TypeAryPtr *tap = tp->isa_aryptr();
2272 if( !tap ) return _type;
2273 return tap->size();
2274 }
2275
2276 //-------------------------------Ideal---------------------------------------
2277 // Feed through the length in AllocateArray(...length...)._length.
2278 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2279 Node* p = MemNode::Ideal_common(phase, can_reshape);
2280 if (p) return (p == NodeSentinel) ? NULL : p;
2281
2282 // Take apart the address into an oop and and offset.
2283 // Return 'this' if we cannot.
2284 Node* adr = in(MemNode::Address);
2285 intptr_t offset = 0;
2286 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2287 if (base == NULL) return NULL;
2288 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2289 if (tary == NULL) return NULL;
2290
2291 // We can fetch the length directly through an AllocateArrayNode.
2292 // This works even if the length is not constant (clone or newArray).
2293 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2294 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2295 if (alloc != NULL) {
2296 Node* allocated_length = alloc->Ideal_length();
2297 Node* len = alloc->make_ideal_length(tary, phase);
2298 if (allocated_length != len) {
2299 // New CastII improves on this.
2300 return len;
2301 }
2302 }
2303 }
2304
2305 return NULL;
2306 }
2307
2308 //------------------------------Identity---------------------------------------
2309 // Feed through the length in AllocateArray(...length...)._length.
2310 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
2311 Node* x = LoadINode::Identity(phase);
2312 if (x != this) return x;
2313
2314 // Take apart the address into an oop and and offset.
2315 // Return 'this' if we cannot.
2316 Node* adr = in(MemNode::Address);
2317 intptr_t offset = 0;
2318 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2319 if (base == NULL) return this;
2320 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2321 if (tary == NULL) return this;
2322
2323 // We can fetch the length directly through an AllocateArrayNode.
2324 // This works even if the length is not constant (clone or newArray).
2325 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2326 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2327 if (alloc != NULL) {
2328 Node* allocated_length = alloc->Ideal_length();
2329 // Do not allow make_ideal_length to allocate a CastII node.
2330 Node* len = alloc->make_ideal_length(tary, phase, false);
2331 if (allocated_length == len) {
2332 // Return allocated_length only if it would not be improved by a CastII.
2333 return allocated_length;
2334 }
2335 }
2336 }
2337
2338 return this;
2339
2340 }
2341
2342 //=============================================================================
2343 //---------------------------StoreNode::make-----------------------------------
2344 // Polymorphic factory method:
2345 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2346 assert((mo == unordered || mo == release), "unexpected");
2347 Compile* C = gvn.C;
2348 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2349 ctl != NULL, "raw memory operations should have control edge");
2350
2351 switch (bt) {
2352 case T_BOOLEAN:
2353 case T_BYTE: return new (C) StoreBNode(ctl, mem, adr, adr_type, val, mo);
2354 case T_INT: return new (C) StoreINode(ctl, mem, adr, adr_type, val, mo);
2355 case T_CHAR:
2356 case T_SHORT: return new (C) StoreCNode(ctl, mem, adr, adr_type, val, mo);
2357 case T_LONG: return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo);
2358 case T_FLOAT: return new (C) StoreFNode(ctl, mem, adr, adr_type, val, mo);
2359 case T_DOUBLE: return new (C) StoreDNode(ctl, mem, adr, adr_type, val, mo);
2360 case T_METADATA:
2361 case T_ADDRESS:
2362 case T_OBJECT:
2363 #ifdef _LP64
2364 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2365 val = gvn.transform(new (C) EncodePNode(val, val->bottom_type()->make_narrowoop()));
2366 return new (C) StoreNNode(ctl, mem, adr, adr_type, val, mo);
2367 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2368 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2369 adr->bottom_type()->isa_rawptr())) {
2370 val = gvn.transform(new (C) EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2371 return new (C) StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2372 }
2373 #endif
2374 {
2375 return new (C) StorePNode(ctl, mem, adr, adr_type, val, mo);
2376 }
2377 }
2378 ShouldNotReachHere();
2379 return (StoreNode*)NULL;
2380 }
2381
2382 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2383 bool require_atomic = true;
2384 return new (C) StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2385 }
2386
2387 StoreDNode* StoreDNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2388 bool require_atomic = true;
2389 return new (C) StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2390 }
2391
2392
2393 //--------------------------bottom_type----------------------------------------
2394 const Type *StoreNode::bottom_type() const {
2395 return Type::MEMORY;
2396 }
2397
2398 //------------------------------hash-------------------------------------------
2399 uint StoreNode::hash() const {
2400 // unroll addition of interesting fields
2401 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2402
2403 // Since they are not commoned, do not hash them:
2404 return NO_HASH;
2405 }
2406
2407 //------------------------------Ideal------------------------------------------
2408 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2409 // When a store immediately follows a relevant allocation/initialization,
2410 // try to capture it into the initialization, or hoist it above.
2411 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2412 Node* p = MemNode::Ideal_common(phase, can_reshape);
2413 if (p) return (p == NodeSentinel) ? NULL : p;
2414
2415 Node* mem = in(MemNode::Memory);
2416 Node* address = in(MemNode::Address);
2417
2418 // Back-to-back stores to same address? Fold em up. Generally
2419 // unsafe if I have intervening uses... Also disallowed for StoreCM
2420 // since they must follow each StoreP operation. Redundant StoreCMs
2421 // are eliminated just before matching in final_graph_reshape.
2422 if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) &&
2423 mem->Opcode() != Op_StoreCM) {
2424 // Looking at a dead closed cycle of memory?
2425 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2426
2427 assert(Opcode() == mem->Opcode() ||
2428 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
2429 "no mismatched stores, except on raw memory");
2430
2431 if (mem->outcnt() == 1 && // check for intervening uses
2432 mem->as_Store()->memory_size() <= this->memory_size()) {
2433 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
2434 // For example, 'mem' might be the final state at a conditional return.
2435 // Or, 'mem' might be used by some node which is live at the same time
2436 // 'this' is live, which might be unschedulable. So, require exactly
2437 // ONE user, the 'this' store, until such time as we clone 'mem' for
2438 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
2439 if (can_reshape) { // (%%% is this an anachronism?)
2440 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
2441 phase->is_IterGVN());
2442 } else {
2443 // It's OK to do this in the parser, since DU info is always accurate,
2444 // and the parser always refers to nodes via SafePointNode maps.
2445 set_req(MemNode::Memory, mem->in(MemNode::Memory));
2446 }
2447 return this;
2448 }
2449 }
2450
2451 // Capture an unaliased, unconditional, simple store into an initializer.
2452 // Or, if it is independent of the allocation, hoist it above the allocation.
2453 if (ReduceFieldZeroing && /*can_reshape &&*/
2454 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2455 InitializeNode* init = mem->in(0)->as_Initialize();
2456 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2457 if (offset > 0) {
2458 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2459 // If the InitializeNode captured me, it made a raw copy of me,
2460 // and I need to disappear.
2461 if (moved != NULL) {
2462 // %%% hack to ensure that Ideal returns a new node:
2463 mem = MergeMemNode::make(phase->C, mem);
2464 return mem; // fold me away
2465 }
2466 }
2467 }
2468
2469 return NULL; // No further progress
2470 }
2471
2472 //------------------------------Value-----------------------------------------
2473 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2474 // Either input is TOP ==> the result is TOP
2475 const Type *t1 = phase->type( in(MemNode::Memory) );
2476 if( t1 == Type::TOP ) return Type::TOP;
2477 const Type *t2 = phase->type( in(MemNode::Address) );
2478 if( t2 == Type::TOP ) return Type::TOP;
2479 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2480 if( t3 == Type::TOP ) return Type::TOP;
2481 return Type::MEMORY;
2482 }
2483
2484 //------------------------------Identity---------------------------------------
2485 // Remove redundant stores:
2486 // Store(m, p, Load(m, p)) changes to m.
2487 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2488 Node *StoreNode::Identity( PhaseTransform *phase ) {
2489 Node* mem = in(MemNode::Memory);
2490 Node* adr = in(MemNode::Address);
2491 Node* val = in(MemNode::ValueIn);
2492
2493 // Load then Store? Then the Store is useless
2494 if (val->is_Load() &&
2495 val->in(MemNode::Address)->eqv_uncast(adr) &&
2496 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2497 val->as_Load()->store_Opcode() == Opcode()) {
2498 return mem;
2499 }
2500
2501 // Two stores in a row of the same value?
2502 if (mem->is_Store() &&
2503 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2504 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2505 mem->Opcode() == Opcode()) {
2506 return mem;
2507 }
2508
2509 // Store of zero anywhere into a freshly-allocated object?
2510 // Then the store is useless.
2511 // (It must already have been captured by the InitializeNode.)
2512 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2513 // a newly allocated object is already all-zeroes everywhere
2514 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2515 return mem;
2516 }
2517
2518 // the store may also apply to zero-bits in an earlier object
2519 Node* prev_mem = find_previous_store(phase);
2520 // Steps (a), (b): Walk past independent stores to find an exact match.
2521 if (prev_mem != NULL) {
2522 Node* prev_val = can_see_stored_value(prev_mem, phase);
2523 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2524 // prev_val and val might differ by a cast; it would be good
2525 // to keep the more informative of the two.
2526 return mem;
2527 }
2528 }
2529 }
2530
2531 return this;
2532 }
2533
2534 //------------------------------match_edge-------------------------------------
2535 // Do we Match on this edge index or not? Match only memory & value
2536 uint StoreNode::match_edge(uint idx) const {
2537 return idx == MemNode::Address || idx == MemNode::ValueIn;
2538 }
2539
2540 //------------------------------cmp--------------------------------------------
2541 // Do not common stores up together. They generally have to be split
2542 // back up anyways, so do not bother.
2543 uint StoreNode::cmp( const Node &n ) const {
2544 return (&n == this); // Always fail except on self
2545 }
2546
2547 //------------------------------Ideal_masked_input-----------------------------
2548 // Check for a useless mask before a partial-word store
2549 // (StoreB ... (AndI valIn conIa) )
2550 // If (conIa & mask == mask) this simplifies to
2551 // (StoreB ... (valIn) )
2552 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2553 Node *val = in(MemNode::ValueIn);
2554 if( val->Opcode() == Op_AndI ) {
2555 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2556 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2557 set_req(MemNode::ValueIn, val->in(1));
2558 return this;
2559 }
2560 }
2561 return NULL;
2562 }
2563
2564
2565 //------------------------------Ideal_sign_extended_input----------------------
2566 // Check for useless sign-extension before a partial-word store
2567 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2568 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2569 // (StoreB ... (valIn) )
2570 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2571 Node *val = in(MemNode::ValueIn);
2572 if( val->Opcode() == Op_RShiftI ) {
2573 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2574 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2575 Node *shl = val->in(1);
2576 if( shl->Opcode() == Op_LShiftI ) {
2577 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2578 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2579 set_req(MemNode::ValueIn, shl->in(1));
2580 return this;
2581 }
2582 }
2583 }
2584 }
2585 return NULL;
2586 }
2587
2588 //------------------------------value_never_loaded-----------------------------------
2589 // Determine whether there are any possible loads of the value stored.
2590 // For simplicity, we actually check if there are any loads from the
2591 // address stored to, not just for loads of the value stored by this node.
2592 //
2593 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2594 Node *adr = in(Address);
2595 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2596 if (adr_oop == NULL)
2597 return false;
2598 if (!adr_oop->is_known_instance_field())
2599 return false; // if not a distinct instance, there may be aliases of the address
2600 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2601 Node *use = adr->fast_out(i);
2602 int opc = use->Opcode();
2603 if (use->is_Load() || use->is_LoadStore()) {
2604 return false;
2605 }
2606 }
2607 return true;
2608 }
2609
2610 //=============================================================================
2611 //------------------------------Ideal------------------------------------------
2612 // If the store is from an AND mask that leaves the low bits untouched, then
2613 // we can skip the AND operation. If the store is from a sign-extension
2614 // (a left shift, then right shift) we can skip both.
2615 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2616 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2617 if( progress != NULL ) return progress;
2618
2619 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2620 if( progress != NULL ) return progress;
2621
2622 // Finally check the default case
2623 return StoreNode::Ideal(phase, can_reshape);
2624 }
2625
2626 //=============================================================================
2627 //------------------------------Ideal------------------------------------------
2628 // If the store is from an AND mask that leaves the low bits untouched, then
2629 // we can skip the AND operation
2630 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2631 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2632 if( progress != NULL ) return progress;
2633
2634 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2635 if( progress != NULL ) return progress;
2636
2637 // Finally check the default case
2638 return StoreNode::Ideal(phase, can_reshape);
2639 }
2640
2641 //=============================================================================
2642 //------------------------------Identity---------------------------------------
2643 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2644 // No need to card mark when storing a null ptr
2645 Node* my_store = in(MemNode::OopStore);
2646 if (my_store->is_Store()) {
2647 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2648 if( t1 == TypePtr::NULL_PTR ) {
2649 return in(MemNode::Memory);
2650 }
2651 }
2652 return this;
2653 }
2654
2655 //=============================================================================
2656 //------------------------------Ideal---------------------------------------
2657 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2658 Node* progress = StoreNode::Ideal(phase, can_reshape);
2659 if (progress != NULL) return progress;
2660
2661 Node* my_store = in(MemNode::OopStore);
2662 if (my_store->is_MergeMem()) {
2663 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2664 set_req(MemNode::OopStore, mem);
2665 return this;
2666 }
2667
2668 return NULL;
2669 }
2670
2671 //------------------------------Value-----------------------------------------
2672 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2673 // Either input is TOP ==> the result is TOP
2674 const Type *t = phase->type( in(MemNode::Memory) );
2675 if( t == Type::TOP ) return Type::TOP;
2676 t = phase->type( in(MemNode::Address) );
2677 if( t == Type::TOP ) return Type::TOP;
2678 t = phase->type( in(MemNode::ValueIn) );
2679 if( t == Type::TOP ) return Type::TOP;
2680 // If extra input is TOP ==> the result is TOP
2681 t = phase->type( in(MemNode::OopStore) );
2682 if( t == Type::TOP ) return Type::TOP;
2683
2684 return StoreNode::Value( phase );
2685 }
2686
2687
2688 //=============================================================================
2689 //----------------------------------SCMemProjNode------------------------------
2690 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2691 {
2692 return bottom_type();
2693 }
2694
2695 //=============================================================================
2696 //----------------------------------LoadStoreNode------------------------------
2697 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2698 : Node(required),
2699 _type(rt),
2700 _adr_type(at)
2701 {
2702 init_req(MemNode::Control, c );
2703 init_req(MemNode::Memory , mem);
2704 init_req(MemNode::Address, adr);
2705 init_req(MemNode::ValueIn, val);
2706 init_class_id(Class_LoadStore);
2707 }
2708
2709 uint LoadStoreNode::ideal_reg() const {
2710 return _type->ideal_reg();
2711 }
2712
2713 bool LoadStoreNode::result_not_used() const {
2714 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2715 Node *x = fast_out(i);
2716 if (x->Opcode() == Op_SCMemProj) continue;
2717 return false;
2718 }
2719 return true;
2720 }
2721
2722 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2723
2724 //=============================================================================
2725 //----------------------------------LoadStoreConditionalNode--------------------
2726 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2727 init_req(ExpectedIn, ex );
2728 }
2729
2730 //=============================================================================
2731 //-------------------------------adr_type--------------------------------------
2732 // Do we Match on this edge index or not? Do not match memory
2733 const TypePtr* ClearArrayNode::adr_type() const {
2734 Node *adr = in(3);
2735 return MemNode::calculate_adr_type(adr->bottom_type());
2736 }
2737
2738 //------------------------------match_edge-------------------------------------
2739 // Do we Match on this edge index or not? Do not match memory
2740 uint ClearArrayNode::match_edge(uint idx) const {
2741 return idx > 1;
2742 }
2743
2744 //------------------------------Identity---------------------------------------
2745 // Clearing a zero length array does nothing
2746 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2747 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2748 }
2749
2750 //------------------------------Idealize---------------------------------------
2751 // Clearing a short array is faster with stores
2752 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2753 const int unit = BytesPerLong;
2754 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2755 if (!t) return NULL;
2756 if (!t->is_con()) return NULL;
2757 intptr_t raw_count = t->get_con();
2758 intptr_t size = raw_count;
2759 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2760 // Clearing nothing uses the Identity call.
2761 // Negative clears are possible on dead ClearArrays
2762 // (see jck test stmt114.stmt11402.val).
2763 if (size <= 0 || size % unit != 0) return NULL;
2764 intptr_t count = size / unit;
2765 // Length too long; use fast hardware clear
2766 if (size > Matcher::init_array_short_size) return NULL;
2767 Node *mem = in(1);
2768 if( phase->type(mem)==Type::TOP ) return NULL;
2769 Node *adr = in(3);
2770 const Type* at = phase->type(adr);
2771 if( at==Type::TOP ) return NULL;
2772 const TypePtr* atp = at->isa_ptr();
2773 // adjust atp to be the correct array element address type
2774 if (atp == NULL) atp = TypePtr::BOTTOM;
2775 else atp = atp->add_offset(Type::OffsetBot);
2776 // Get base for derived pointer purposes
2777 if( adr->Opcode() != Op_AddP ) Unimplemented();
2778 Node *base = adr->in(1);
2779
2780 Node *zero = phase->makecon(TypeLong::ZERO);
2781 Node *off = phase->MakeConX(BytesPerLong);
2782 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2783 count--;
2784 while( count-- ) {
2785 mem = phase->transform(mem);
2786 adr = phase->transform(new (phase->C) AddPNode(base,adr,off));
2787 mem = new (phase->C) StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2788 }
2789 return mem;
2790 }
2791
2792 //----------------------------step_through----------------------------------
2793 // Return allocation input memory edge if it is different instance
2794 // or itself if it is the one we are looking for.
2795 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2796 Node* n = *np;
2797 assert(n->is_ClearArray(), "sanity");
2798 intptr_t offset;
2799 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2800 // This method is called only before Allocate nodes are expanded during
2801 // macro nodes expansion. Before that ClearArray nodes are only generated
2802 // in LibraryCallKit::generate_arraycopy() which follows allocations.
2803 assert(alloc != NULL, "should have allocation");
2804 if (alloc->_idx == instance_id) {
2805 // Can not bypass initialization of the instance we are looking for.
2806 return false;
2807 }
2808 // Otherwise skip it.
2809 InitializeNode* init = alloc->initialization();
2810 if (init != NULL)
2811 *np = init->in(TypeFunc::Memory);
2812 else
2813 *np = alloc->in(TypeFunc::Memory);
2814 return true;
2815 }
2816
2817 //----------------------------clear_memory-------------------------------------
2818 // Generate code to initialize object storage to zero.
2819 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2820 intptr_t start_offset,
2821 Node* end_offset,
2822 PhaseGVN* phase) {
2823 Compile* C = phase->C;
2824 intptr_t offset = start_offset;
2825
2826 int unit = BytesPerLong;
2827 if ((offset % unit) != 0) {
2828 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(offset));
2829 adr = phase->transform(adr);
2830 const TypePtr* atp = TypeRawPtr::BOTTOM;
2831 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2832 mem = phase->transform(mem);
2833 offset += BytesPerInt;
2834 }
2835 assert((offset % unit) == 0, "");
2836
2837 // Initialize the remaining stuff, if any, with a ClearArray.
2838 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2839 }
2840
2841 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2842 Node* start_offset,
2843 Node* end_offset,
2844 PhaseGVN* phase) {
2845 if (start_offset == end_offset) {
2846 // nothing to do
2847 return mem;
2848 }
2849
2850 Compile* C = phase->C;
2851 int unit = BytesPerLong;
2852 Node* zbase = start_offset;
2853 Node* zend = end_offset;
2854
2855 // Scale to the unit required by the CPU:
2856 if (!Matcher::init_array_count_is_in_bytes) {
2857 Node* shift = phase->intcon(exact_log2(unit));
2858 zbase = phase->transform( new(C) URShiftXNode(zbase, shift) );
2859 zend = phase->transform( new(C) URShiftXNode(zend, shift) );
2860 }
2861
2862 // Bulk clear double-words
2863 Node* zsize = phase->transform( new(C) SubXNode(zend, zbase) );
2864 Node* adr = phase->transform( new(C) AddPNode(dest, dest, start_offset) );
2865 mem = new (C) ClearArrayNode(ctl, mem, zsize, adr);
2866 return phase->transform(mem);
2867 }
2868
2869 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2870 intptr_t start_offset,
2871 intptr_t end_offset,
2872 PhaseGVN* phase) {
2873 if (start_offset == end_offset) {
2874 // nothing to do
2875 return mem;
2876 }
2877
2878 Compile* C = phase->C;
2879 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2880 intptr_t done_offset = end_offset;
2881 if ((done_offset % BytesPerLong) != 0) {
2882 done_offset -= BytesPerInt;
2883 }
2884 if (done_offset > start_offset) {
2885 mem = clear_memory(ctl, mem, dest,
2886 start_offset, phase->MakeConX(done_offset), phase);
2887 }
2888 if (done_offset < end_offset) { // emit the final 32-bit store
2889 Node* adr = new (C) AddPNode(dest, dest, phase->MakeConX(done_offset));
2890 adr = phase->transform(adr);
2891 const TypePtr* atp = TypeRawPtr::BOTTOM;
2892 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2893 mem = phase->transform(mem);
2894 done_offset += BytesPerInt;
2895 }
2896 assert(done_offset == end_offset, "");
2897 return mem;
2898 }
2899
2900 //=============================================================================
2901 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2902 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2903 _adr_type(C->get_adr_type(alias_idx))
2904 {
2905 init_class_id(Class_MemBar);
2906 Node* top = C->top();
2907 init_req(TypeFunc::I_O,top);
2908 init_req(TypeFunc::FramePtr,top);
2909 init_req(TypeFunc::ReturnAdr,top);
2910 if (precedent != NULL)
2911 init_req(TypeFunc::Parms, precedent);
2912 }
2913
2914 //------------------------------cmp--------------------------------------------
2915 uint MemBarNode::hash() const { return NO_HASH; }
2916 uint MemBarNode::cmp( const Node &n ) const {
2917 return (&n == this); // Always fail except on self
2918 }
2919
2920 //------------------------------make-------------------------------------------
2921 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2922 switch (opcode) {
2923 case Op_MemBarAcquire: return new(C) MemBarAcquireNode(C, atp, pn);
2924 case Op_LoadFence: return new(C) LoadFenceNode(C, atp, pn);
2925 case Op_MemBarRelease: return new(C) MemBarReleaseNode(C, atp, pn);
2926 case Op_StoreFence: return new(C) StoreFenceNode(C, atp, pn);
2927 case Op_MemBarAcquireLock: return new(C) MemBarAcquireLockNode(C, atp, pn);
2928 case Op_MemBarReleaseLock: return new(C) MemBarReleaseLockNode(C, atp, pn);
2929 case Op_MemBarVolatile: return new(C) MemBarVolatileNode(C, atp, pn);
2930 case Op_MemBarCPUOrder: return new(C) MemBarCPUOrderNode(C, atp, pn);
2931 case Op_Initialize: return new(C) InitializeNode(C, atp, pn);
2932 case Op_MemBarStoreStore: return new(C) MemBarStoreStoreNode(C, atp, pn);
2933 default: ShouldNotReachHere(); return NULL;
2934 }
2935 }
2936
2937 //------------------------------Ideal------------------------------------------
2938 // Return a node which is more "ideal" than the current node. Strip out
2939 // control copies
2940 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2941 if (remove_dead_region(phase, can_reshape)) return this;
2942 // Don't bother trying to transform a dead node
2943 if (in(0) && in(0)->is_top()) {
2944 return NULL;
2945 }
2946
2947 // Eliminate volatile MemBars for scalar replaced objects.
2948 if (can_reshape && req() == (Precedent+1)) {
2949 bool eliminate = false;
2950 int opc = Opcode();
2951 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
2952 // Volatile field loads and stores.
2953 Node* my_mem = in(MemBarNode::Precedent);
2954 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
2955 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
2956 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
2957 // replace this Precedent (decodeN) with the Load instead.
2958 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
2959 Node* load_node = my_mem->in(1);
2960 set_req(MemBarNode::Precedent, load_node);
2961 phase->is_IterGVN()->_worklist.push(my_mem);
2962 my_mem = load_node;
2963 } else {
2964 assert(my_mem->unique_out() == this, "sanity");
2965 del_req(Precedent);
2966 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
2967 my_mem = NULL;
2968 }
2969 }
2970 if (my_mem != NULL && my_mem->is_Mem()) {
2971 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
2972 // Check for scalar replaced object reference.
2973 if( t_oop != NULL && t_oop->is_known_instance_field() &&
2974 t_oop->offset() != Type::OffsetBot &&
2975 t_oop->offset() != Type::OffsetTop) {
2976 eliminate = true;
2977 }
2978 }
2979 } else if (opc == Op_MemBarRelease) {
2980 // Final field stores.
2981 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
2982 if ((alloc != NULL) && alloc->is_Allocate() &&
2983 alloc->as_Allocate()->_is_non_escaping) {
2984 // The allocated object does not escape.
2985 eliminate = true;
2986 }
2987 }
2988 if (eliminate) {
2989 // Replace MemBar projections by its inputs.
2990 PhaseIterGVN* igvn = phase->is_IterGVN();
2991 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
2992 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
2993 // Must return either the original node (now dead) or a new node
2994 // (Do not return a top here, since that would break the uniqueness of top.)
2995 return new (phase->C) ConINode(TypeInt::ZERO);
2996 }
2997 }
2998 return NULL;
2999 }
3000
3001 //------------------------------Value------------------------------------------
3002 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
3003 if( !in(0) ) return Type::TOP;
3004 if( phase->type(in(0)) == Type::TOP )
3005 return Type::TOP;
3006 return TypeTuple::MEMBAR;
3007 }
3008
3009 //------------------------------match------------------------------------------
3010 // Construct projections for memory.
3011 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3012 switch (proj->_con) {
3013 case TypeFunc::Control:
3014 case TypeFunc::Memory:
3015 return new (m->C) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3016 }
3017 ShouldNotReachHere();
3018 return NULL;
3019 }
3020
3021 //===========================InitializeNode====================================
3022 // SUMMARY:
3023 // This node acts as a memory barrier on raw memory, after some raw stores.
3024 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3025 // The Initialize can 'capture' suitably constrained stores as raw inits.
3026 // It can coalesce related raw stores into larger units (called 'tiles').
3027 // It can avoid zeroing new storage for memory units which have raw inits.
3028 // At macro-expansion, it is marked 'complete', and does not optimize further.
3029 //
3030 // EXAMPLE:
3031 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3032 // ctl = incoming control; mem* = incoming memory
3033 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3034 // First allocate uninitialized memory and fill in the header:
3035 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3036 // ctl := alloc.Control; mem* := alloc.Memory*
3037 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3038 // Then initialize to zero the non-header parts of the raw memory block:
3039 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3040 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3041 // After the initialize node executes, the object is ready for service:
3042 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3043 // Suppose its body is immediately initialized as {1,2}:
3044 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3045 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3046 // mem.SLICE(#short[*]) := store2
3047 //
3048 // DETAILS:
3049 // An InitializeNode collects and isolates object initialization after
3050 // an AllocateNode and before the next possible safepoint. As a
3051 // memory barrier (MemBarNode), it keeps critical stores from drifting
3052 // down past any safepoint or any publication of the allocation.
3053 // Before this barrier, a newly-allocated object may have uninitialized bits.
3054 // After this barrier, it may be treated as a real oop, and GC is allowed.
3055 //
3056 // The semantics of the InitializeNode include an implicit zeroing of
3057 // the new object from object header to the end of the object.
3058 // (The object header and end are determined by the AllocateNode.)
3059 //
3060 // Certain stores may be added as direct inputs to the InitializeNode.
3061 // These stores must update raw memory, and they must be to addresses
3062 // derived from the raw address produced by AllocateNode, and with
3063 // a constant offset. They must be ordered by increasing offset.
3064 // The first one is at in(RawStores), the last at in(req()-1).
3065 // Unlike most memory operations, they are not linked in a chain,
3066 // but are displayed in parallel as users of the rawmem output of
3067 // the allocation.
3068 //
3069 // (See comments in InitializeNode::capture_store, which continue
3070 // the example given above.)
3071 //
3072 // When the associated Allocate is macro-expanded, the InitializeNode
3073 // may be rewritten to optimize collected stores. A ClearArrayNode
3074 // may also be created at that point to represent any required zeroing.
3075 // The InitializeNode is then marked 'complete', prohibiting further
3076 // capturing of nearby memory operations.
3077 //
3078 // During macro-expansion, all captured initializations which store
3079 // constant values of 32 bits or smaller are coalesced (if advantageous)
3080 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3081 // initialized in fewer memory operations. Memory words which are
3082 // covered by neither tiles nor non-constant stores are pre-zeroed
3083 // by explicit stores of zero. (The code shape happens to do all
3084 // zeroing first, then all other stores, with both sequences occurring
3085 // in order of ascending offsets.)
3086 //
3087 // Alternatively, code may be inserted between an AllocateNode and its
3088 // InitializeNode, to perform arbitrary initialization of the new object.
3089 // E.g., the object copying intrinsics insert complex data transfers here.
3090 // The initialization must then be marked as 'complete' disable the
3091 // built-in zeroing semantics and the collection of initializing stores.
3092 //
3093 // While an InitializeNode is incomplete, reads from the memory state
3094 // produced by it are optimizable if they match the control edge and
3095 // new oop address associated with the allocation/initialization.
3096 // They return a stored value (if the offset matches) or else zero.
3097 // A write to the memory state, if it matches control and address,
3098 // and if it is to a constant offset, may be 'captured' by the
3099 // InitializeNode. It is cloned as a raw memory operation and rewired
3100 // inside the initialization, to the raw oop produced by the allocation.
3101 // Operations on addresses which are provably distinct (e.g., to
3102 // other AllocateNodes) are allowed to bypass the initialization.
3103 //
3104 // The effect of all this is to consolidate object initialization
3105 // (both arrays and non-arrays, both piecewise and bulk) into a
3106 // single location, where it can be optimized as a unit.
3107 //
3108 // Only stores with an offset less than TrackedInitializationLimit words
3109 // will be considered for capture by an InitializeNode. This puts a
3110 // reasonable limit on the complexity of optimized initializations.
3111
3112 //---------------------------InitializeNode------------------------------------
3113 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3114 : _is_complete(Incomplete), _does_not_escape(false),
3115 MemBarNode(C, adr_type, rawoop)
3116 {
3117 init_class_id(Class_Initialize);
3118
3119 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3120 assert(in(RawAddress) == rawoop, "proper init");
3121 // Note: allocation() can be NULL, for secondary initialization barriers
3122 }
3123
3124 // Since this node is not matched, it will be processed by the
3125 // register allocator. Declare that there are no constraints
3126 // on the allocation of the RawAddress edge.
3127 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3128 // This edge should be set to top, by the set_complete. But be conservative.
3129 if (idx == InitializeNode::RawAddress)
3130 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3131 return RegMask::Empty;
3132 }
3133
3134 Node* InitializeNode::memory(uint alias_idx) {
3135 Node* mem = in(Memory);
3136 if (mem->is_MergeMem()) {
3137 return mem->as_MergeMem()->memory_at(alias_idx);
3138 } else {
3139 // incoming raw memory is not split
3140 return mem;
3141 }
3142 }
3143
3144 bool InitializeNode::is_non_zero() {
3145 if (is_complete()) return false;
3146 remove_extra_zeroes();
3147 return (req() > RawStores);
3148 }
3149
3150 void InitializeNode::set_complete(PhaseGVN* phase) {
3151 assert(!is_complete(), "caller responsibility");
3152 _is_complete = Complete;
3153
3154 // After this node is complete, it contains a bunch of
3155 // raw-memory initializations. There is no need for
3156 // it to have anything to do with non-raw memory effects.
3157 // Therefore, tell all non-raw users to re-optimize themselves,
3158 // after skipping the memory effects of this initialization.
3159 PhaseIterGVN* igvn = phase->is_IterGVN();
3160 if (igvn) igvn->add_users_to_worklist(this);
3161 }
3162
3163 // convenience function
3164 // return false if the init contains any stores already
3165 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3166 InitializeNode* init = initialization();
3167 if (init == NULL || init->is_complete()) return false;
3168 init->remove_extra_zeroes();
3169 // for now, if this allocation has already collected any inits, bail:
3170 if (init->is_non_zero()) return false;
3171 init->set_complete(phase);
3172 return true;
3173 }
3174
3175 void InitializeNode::remove_extra_zeroes() {
3176 if (req() == RawStores) return;
3177 Node* zmem = zero_memory();
3178 uint fill = RawStores;
3179 for (uint i = fill; i < req(); i++) {
3180 Node* n = in(i);
3181 if (n->is_top() || n == zmem) continue; // skip
3182 if (fill < i) set_req(fill, n); // compact
3183 ++fill;
3184 }
3185 // delete any empty spaces created:
3186 while (fill < req()) {
3187 del_req(fill);
3188 }
3189 }
3190
3191 // Helper for remembering which stores go with which offsets.
3192 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3193 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3194 intptr_t offset = -1;
3195 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3196 phase, offset);
3197 if (base == NULL) return -1; // something is dead,
3198 if (offset < 0) return -1; // dead, dead
3199 return offset;
3200 }
3201
3202 // Helper for proving that an initialization expression is
3203 // "simple enough" to be folded into an object initialization.
3204 // Attempts to prove that a store's initial value 'n' can be captured
3205 // within the initialization without creating a vicious cycle, such as:
3206 // { Foo p = new Foo(); p.next = p; }
3207 // True for constants and parameters and small combinations thereof.
3208 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3209 if (n == NULL) return true; // (can this really happen?)
3210 if (n->is_Proj()) n = n->in(0);
3211 if (n == this) return false; // found a cycle
3212 if (n->is_Con()) return true;
3213 if (n->is_Start()) return true; // params, etc., are OK
3214 if (n->is_Root()) return true; // even better
3215
3216 Node* ctl = n->in(0);
3217 if (ctl != NULL && !ctl->is_top()) {
3218 if (ctl->is_Proj()) ctl = ctl->in(0);
3219 if (ctl == this) return false;
3220
3221 // If we already know that the enclosing memory op is pinned right after
3222 // the init, then any control flow that the store has picked up
3223 // must have preceded the init, or else be equal to the init.
3224 // Even after loop optimizations (which might change control edges)
3225 // a store is never pinned *before* the availability of its inputs.
3226 if (!MemNode::all_controls_dominate(n, this))
3227 return false; // failed to prove a good control
3228 }
3229
3230 // Check data edges for possible dependencies on 'this'.
3231 if ((count += 1) > 20) return false; // complexity limit
3232 for (uint i = 1; i < n->req(); i++) {
3233 Node* m = n->in(i);
3234 if (m == NULL || m == n || m->is_top()) continue;
3235 uint first_i = n->find_edge(m);
3236 if (i != first_i) continue; // process duplicate edge just once
3237 if (!detect_init_independence(m, count)) {
3238 return false;
3239 }
3240 }
3241
3242 return true;
3243 }
3244
3245 // Here are all the checks a Store must pass before it can be moved into
3246 // an initialization. Returns zero if a check fails.
3247 // On success, returns the (constant) offset to which the store applies,
3248 // within the initialized memory.
3249 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3250 const int FAIL = 0;
3251 if (st->req() != MemNode::ValueIn + 1)
3252 return FAIL; // an inscrutable StoreNode (card mark?)
3253 Node* ctl = st->in(MemNode::Control);
3254 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3255 return FAIL; // must be unconditional after the initialization
3256 Node* mem = st->in(MemNode::Memory);
3257 if (!(mem->is_Proj() && mem->in(0) == this))
3258 return FAIL; // must not be preceded by other stores
3259 Node* adr = st->in(MemNode::Address);
3260 intptr_t offset;
3261 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3262 if (alloc == NULL)
3263 return FAIL; // inscrutable address
3264 if (alloc != allocation())
3265 return FAIL; // wrong allocation! (store needs to float up)
3266 Node* val = st->in(MemNode::ValueIn);
3267 int complexity_count = 0;
3268 if (!detect_init_independence(val, complexity_count))
3269 return FAIL; // stored value must be 'simple enough'
3270
3271 // The Store can be captured only if nothing after the allocation
3272 // and before the Store is using the memory location that the store
3273 // overwrites.
3274 bool failed = false;
3275 // If is_complete_with_arraycopy() is true the shape of the graph is
3276 // well defined and is safe so no need for extra checks.
3277 if (!is_complete_with_arraycopy()) {
3278 // We are going to look at each use of the memory state following
3279 // the allocation to make sure nothing reads the memory that the
3280 // Store writes.
3281 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3282 int alias_idx = phase->C->get_alias_index(t_adr);
3283 ResourceMark rm;
3284 Unique_Node_List mems;
3285 mems.push(mem);
3286 Node* unique_merge = NULL;
3287 for (uint next = 0; next < mems.size(); ++next) {
3288 Node *m = mems.at(next);
3289 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3290 Node *n = m->fast_out(j);
3291 if (n->outcnt() == 0) {
3292 continue;
3293 }
3294 if (n == st) {
3295 continue;
3296 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3297 // If the control of this use is different from the control
3298 // of the Store which is right after the InitializeNode then
3299 // this node cannot be between the InitializeNode and the
3300 // Store.
3301 continue;
3302 } else if (n->is_MergeMem()) {
3303 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3304 // We can hit a MergeMemNode (that will likely go away
3305 // later) that is a direct use of the memory state
3306 // following the InitializeNode on the same slice as the
3307 // store node that we'd like to capture. We need to check
3308 // the uses of the MergeMemNode.
3309 mems.push(n);
3310 }
3311 } else if (n->is_Mem()) {
3312 Node* other_adr = n->in(MemNode::Address);
3313 if (other_adr == adr) {
3314 failed = true;
3315 break;
3316 } else {
3317 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3318 if (other_t_adr != NULL) {
3319 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3320 if (other_alias_idx == alias_idx) {
3321 // A load from the same memory slice as the store right
3322 // after the InitializeNode. We check the control of the
3323 // object/array that is loaded from. If it's the same as
3324 // the store control then we cannot capture the store.
3325 assert(!n->is_Store(), "2 stores to same slice on same control?");
3326 Node* base = other_adr;
3327 assert(base->is_AddP(), err_msg_res("should be addp but is %s", base->Name()));
3328 base = base->in(AddPNode::Base);
3329 if (base != NULL) {
3330 base = base->uncast();
3331 if (base->is_Proj() && base->in(0) == alloc) {
3332 failed = true;
3333 break;
3334 }
3335 }
3336 }
3337 }
3338 }
3339 } else {
3340 failed = true;
3341 break;
3342 }
3343 }
3344 }
3345 }
3346 if (failed) {
3347 if (!can_reshape) {
3348 // We decided we couldn't capture the store during parsing. We
3349 // should try again during the next IGVN once the graph is
3350 // cleaner.
3351 phase->C->record_for_igvn(st);
3352 }
3353 return FAIL;
3354 }
3355
3356 return offset; // success
3357 }
3358
3359 // Find the captured store in(i) which corresponds to the range
3360 // [start..start+size) in the initialized object.
3361 // If there is one, return its index i. If there isn't, return the
3362 // negative of the index where it should be inserted.
3363 // Return 0 if the queried range overlaps an initialization boundary
3364 // or if dead code is encountered.
3365 // If size_in_bytes is zero, do not bother with overlap checks.
3366 int InitializeNode::captured_store_insertion_point(intptr_t start,
3367 int size_in_bytes,
3368 PhaseTransform* phase) {
3369 const int FAIL = 0, MAX_STORE = BytesPerLong;
3370
3371 if (is_complete())
3372 return FAIL; // arraycopy got here first; punt
3373
3374 assert(allocation() != NULL, "must be present");
3375
3376 // no negatives, no header fields:
3377 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3378
3379 // after a certain size, we bail out on tracking all the stores:
3380 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3381 if (start >= ti_limit) return FAIL;
3382
3383 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3384 if (i >= limit) return -(int)i; // not found; here is where to put it
3385
3386 Node* st = in(i);
3387 intptr_t st_off = get_store_offset(st, phase);
3388 if (st_off < 0) {
3389 if (st != zero_memory()) {
3390 return FAIL; // bail out if there is dead garbage
3391 }
3392 } else if (st_off > start) {
3393 // ...we are done, since stores are ordered
3394 if (st_off < start + size_in_bytes) {
3395 return FAIL; // the next store overlaps
3396 }
3397 return -(int)i; // not found; here is where to put it
3398 } else if (st_off < start) {
3399 if (size_in_bytes != 0 &&
3400 start < st_off + MAX_STORE &&
3401 start < st_off + st->as_Store()->memory_size()) {
3402 return FAIL; // the previous store overlaps
3403 }
3404 } else {
3405 if (size_in_bytes != 0 &&
3406 st->as_Store()->memory_size() != size_in_bytes) {
3407 return FAIL; // mismatched store size
3408 }
3409 return i;
3410 }
3411
3412 ++i;
3413 }
3414 }
3415
3416 // Look for a captured store which initializes at the offset 'start'
3417 // with the given size. If there is no such store, and no other
3418 // initialization interferes, then return zero_memory (the memory
3419 // projection of the AllocateNode).
3420 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3421 PhaseTransform* phase) {
3422 assert(stores_are_sane(phase), "");
3423 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3424 if (i == 0) {
3425 return NULL; // something is dead
3426 } else if (i < 0) {
3427 return zero_memory(); // just primordial zero bits here
3428 } else {
3429 Node* st = in(i); // here is the store at this position
3430 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3431 return st;
3432 }
3433 }
3434
3435 // Create, as a raw pointer, an address within my new object at 'offset'.
3436 Node* InitializeNode::make_raw_address(intptr_t offset,
3437 PhaseTransform* phase) {
3438 Node* addr = in(RawAddress);
3439 if (offset != 0) {
3440 Compile* C = phase->C;
3441 addr = phase->transform( new (C) AddPNode(C->top(), addr,
3442 phase->MakeConX(offset)) );
3443 }
3444 return addr;
3445 }
3446
3447 // Clone the given store, converting it into a raw store
3448 // initializing a field or element of my new object.
3449 // Caller is responsible for retiring the original store,
3450 // with subsume_node or the like.
3451 //
3452 // From the example above InitializeNode::InitializeNode,
3453 // here are the old stores to be captured:
3454 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3455 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3456 //
3457 // Here is the changed code; note the extra edges on init:
3458 // alloc = (Allocate ...)
3459 // rawoop = alloc.RawAddress
3460 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3461 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3462 // init = (Initialize alloc.Control alloc.Memory rawoop
3463 // rawstore1 rawstore2)
3464 //
3465 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3466 PhaseTransform* phase, bool can_reshape) {
3467 assert(stores_are_sane(phase), "");
3468
3469 if (start < 0) return NULL;
3470 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3471
3472 Compile* C = phase->C;
3473 int size_in_bytes = st->memory_size();
3474 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3475 if (i == 0) return NULL; // bail out
3476 Node* prev_mem = NULL; // raw memory for the captured store
3477 if (i > 0) {
3478 prev_mem = in(i); // there is a pre-existing store under this one
3479 set_req(i, C->top()); // temporarily disconnect it
3480 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3481 } else {
3482 i = -i; // no pre-existing store
3483 prev_mem = zero_memory(); // a slice of the newly allocated object
3484 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3485 set_req(--i, C->top()); // reuse this edge; it has been folded away
3486 else
3487 ins_req(i, C->top()); // build a new edge
3488 }
3489 Node* new_st = st->clone();
3490 new_st->set_req(MemNode::Control, in(Control));
3491 new_st->set_req(MemNode::Memory, prev_mem);
3492 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3493 new_st = phase->transform(new_st);
3494
3495 // At this point, new_st might have swallowed a pre-existing store
3496 // at the same offset, or perhaps new_st might have disappeared,
3497 // if it redundantly stored the same value (or zero to fresh memory).
3498
3499 // In any case, wire it in:
3500 set_req(i, new_st);
3501
3502 // The caller may now kill the old guy.
3503 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3504 assert(check_st == new_st || check_st == NULL, "must be findable");
3505 assert(!is_complete(), "");
3506 return new_st;
3507 }
3508
3509 static bool store_constant(jlong* tiles, int num_tiles,
3510 intptr_t st_off, int st_size,
3511 jlong con) {
3512 if ((st_off & (st_size-1)) != 0)
3513 return false; // strange store offset (assume size==2**N)
3514 address addr = (address)tiles + st_off;
3515 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3516 switch (st_size) {
3517 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3518 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3519 case sizeof(jint): *(jint*) addr = (jint) con; break;
3520 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3521 default: return false; // strange store size (detect size!=2**N here)
3522 }
3523 return true; // return success to caller
3524 }
3525
3526 // Coalesce subword constants into int constants and possibly
3527 // into long constants. The goal, if the CPU permits,
3528 // is to initialize the object with a small number of 64-bit tiles.
3529 // Also, convert floating-point constants to bit patterns.
3530 // Non-constants are not relevant to this pass.
3531 //
3532 // In terms of the running example on InitializeNode::InitializeNode
3533 // and InitializeNode::capture_store, here is the transformation
3534 // of rawstore1 and rawstore2 into rawstore12:
3535 // alloc = (Allocate ...)
3536 // rawoop = alloc.RawAddress
3537 // tile12 = 0x00010002
3538 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3539 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3540 //
3541 void
3542 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3543 Node* size_in_bytes,
3544 PhaseGVN* phase) {
3545 Compile* C = phase->C;
3546
3547 assert(stores_are_sane(phase), "");
3548 // Note: After this pass, they are not completely sane,
3549 // since there may be some overlaps.
3550
3551 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3552
3553 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3554 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3555 size_limit = MIN2(size_limit, ti_limit);
3556 size_limit = align_size_up(size_limit, BytesPerLong);
3557 int num_tiles = size_limit / BytesPerLong;
3558
3559 // allocate space for the tile map:
3560 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3561 jlong tiles_buf[small_len];
3562 Node* nodes_buf[small_len];
3563 jlong inits_buf[small_len];
3564 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3565 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3566 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3567 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3568 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3569 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3570 // tiles: exact bitwise model of all primitive constants
3571 // nodes: last constant-storing node subsumed into the tiles model
3572 // inits: which bytes (in each tile) are touched by any initializations
3573
3574 //// Pass A: Fill in the tile model with any relevant stores.
3575
3576 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3577 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3578 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3579 Node* zmem = zero_memory(); // initially zero memory state
3580 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3581 Node* st = in(i);
3582 intptr_t st_off = get_store_offset(st, phase);
3583
3584 // Figure out the store's offset and constant value:
3585 if (st_off < header_size) continue; //skip (ignore header)
3586 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3587 int st_size = st->as_Store()->memory_size();
3588 if (st_off + st_size > size_limit) break;
3589
3590 // Record which bytes are touched, whether by constant or not.
3591 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3592 continue; // skip (strange store size)
3593
3594 const Type* val = phase->type(st->in(MemNode::ValueIn));
3595 if (!val->singleton()) continue; //skip (non-con store)
3596 BasicType type = val->basic_type();
3597
3598 jlong con = 0;
3599 switch (type) {
3600 case T_INT: con = val->is_int()->get_con(); break;
3601 case T_LONG: con = val->is_long()->get_con(); break;
3602 case T_FLOAT: con = jint_cast(val->getf()); break;
3603 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3604 default: continue; //skip (odd store type)
3605 }
3606
3607 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3608 st->Opcode() == Op_StoreL) {
3609 continue; // This StoreL is already optimal.
3610 }
3611
3612 // Store down the constant.
3613 store_constant(tiles, num_tiles, st_off, st_size, con);
3614
3615 intptr_t j = st_off >> LogBytesPerLong;
3616
3617 if (type == T_INT && st_size == BytesPerInt
3618 && (st_off & BytesPerInt) == BytesPerInt) {
3619 jlong lcon = tiles[j];
3620 if (!Matcher::isSimpleConstant64(lcon) &&
3621 st->Opcode() == Op_StoreI) {
3622 // This StoreI is already optimal by itself.
3623 jint* intcon = (jint*) &tiles[j];
3624 intcon[1] = 0; // undo the store_constant()
3625
3626 // If the previous store is also optimal by itself, back up and
3627 // undo the action of the previous loop iteration... if we can.
3628 // But if we can't, just let the previous half take care of itself.
3629 st = nodes[j];
3630 st_off -= BytesPerInt;
3631 con = intcon[0];
3632 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3633 assert(st_off >= header_size, "still ignoring header");
3634 assert(get_store_offset(st, phase) == st_off, "must be");
3635 assert(in(i-1) == zmem, "must be");
3636 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3637 assert(con == tcon->is_int()->get_con(), "must be");
3638 // Undo the effects of the previous loop trip, which swallowed st:
3639 intcon[0] = 0; // undo store_constant()
3640 set_req(i-1, st); // undo set_req(i, zmem)
3641 nodes[j] = NULL; // undo nodes[j] = st
3642 --old_subword; // undo ++old_subword
3643 }
3644 continue; // This StoreI is already optimal.
3645 }
3646 }
3647
3648 // This store is not needed.
3649 set_req(i, zmem);
3650 nodes[j] = st; // record for the moment
3651 if (st_size < BytesPerLong) // something has changed
3652 ++old_subword; // includes int/float, but who's counting...
3653 else ++old_long;
3654 }
3655
3656 if ((old_subword + old_long) == 0)
3657 return; // nothing more to do
3658
3659 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3660 // Be sure to insert them before overlapping non-constant stores.
3661 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3662 for (int j = 0; j < num_tiles; j++) {
3663 jlong con = tiles[j];
3664 jlong init = inits[j];
3665 if (con == 0) continue;
3666 jint con0, con1; // split the constant, address-wise
3667 jint init0, init1; // split the init map, address-wise
3668 { union { jlong con; jint intcon[2]; } u;
3669 u.con = con;
3670 con0 = u.intcon[0];
3671 con1 = u.intcon[1];
3672 u.con = init;
3673 init0 = u.intcon[0];
3674 init1 = u.intcon[1];
3675 }
3676
3677 Node* old = nodes[j];
3678 assert(old != NULL, "need the prior store");
3679 intptr_t offset = (j * BytesPerLong);
3680
3681 bool split = !Matcher::isSimpleConstant64(con);
3682
3683 if (offset < header_size) {
3684 assert(offset + BytesPerInt >= header_size, "second int counts");
3685 assert(*(jint*)&tiles[j] == 0, "junk in header");
3686 split = true; // only the second word counts
3687 // Example: int a[] = { 42 ... }
3688 } else if (con0 == 0 && init0 == -1) {
3689 split = true; // first word is covered by full inits
3690 // Example: int a[] = { ... foo(), 42 ... }
3691 } else if (con1 == 0 && init1 == -1) {
3692 split = true; // second word is covered by full inits
3693 // Example: int a[] = { ... 42, foo() ... }
3694 }
3695
3696 // Here's a case where init0 is neither 0 nor -1:
3697 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3698 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3699 // In this case the tile is not split; it is (jlong)42.
3700 // The big tile is stored down, and then the foo() value is inserted.
3701 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3702
3703 Node* ctl = old->in(MemNode::Control);
3704 Node* adr = make_raw_address(offset, phase);
3705 const TypePtr* atp = TypeRawPtr::BOTTOM;
3706
3707 // One or two coalesced stores to plop down.
3708 Node* st[2];
3709 intptr_t off[2];
3710 int nst = 0;
3711 if (!split) {
3712 ++new_long;
3713 off[nst] = offset;
3714 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3715 phase->longcon(con), T_LONG, MemNode::unordered);
3716 } else {
3717 // Omit either if it is a zero.
3718 if (con0 != 0) {
3719 ++new_int;
3720 off[nst] = offset;
3721 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3722 phase->intcon(con0), T_INT, MemNode::unordered);
3723 }
3724 if (con1 != 0) {
3725 ++new_int;
3726 offset += BytesPerInt;
3727 adr = make_raw_address(offset, phase);
3728 off[nst] = offset;
3729 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3730 phase->intcon(con1), T_INT, MemNode::unordered);
3731 }
3732 }
3733
3734 // Insert second store first, then the first before the second.
3735 // Insert each one just before any overlapping non-constant stores.
3736 while (nst > 0) {
3737 Node* st1 = st[--nst];
3738 C->copy_node_notes_to(st1, old);
3739 st1 = phase->transform(st1);
3740 offset = off[nst];
3741 assert(offset >= header_size, "do not smash header");
3742 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3743 guarantee(ins_idx != 0, "must re-insert constant store");
3744 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3745 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3746 set_req(--ins_idx, st1);
3747 else
3748 ins_req(ins_idx, st1);
3749 }
3750 }
3751
3752 if (PrintCompilation && WizardMode)
3753 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3754 old_subword, old_long, new_int, new_long);
3755 if (C->log() != NULL)
3756 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3757 old_subword, old_long, new_int, new_long);
3758
3759 // Clean up any remaining occurrences of zmem:
3760 remove_extra_zeroes();
3761 }
3762
3763 // Explore forward from in(start) to find the first fully initialized
3764 // word, and return its offset. Skip groups of subword stores which
3765 // together initialize full words. If in(start) is itself part of a
3766 // fully initialized word, return the offset of in(start). If there
3767 // are no following full-word stores, or if something is fishy, return
3768 // a negative value.
3769 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3770 int int_map = 0;
3771 intptr_t int_map_off = 0;
3772 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3773
3774 for (uint i = start, limit = req(); i < limit; i++) {
3775 Node* st = in(i);
3776
3777 intptr_t st_off = get_store_offset(st, phase);
3778 if (st_off < 0) break; // return conservative answer
3779
3780 int st_size = st->as_Store()->memory_size();
3781 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3782 return st_off; // we found a complete word init
3783 }
3784
3785 // update the map:
3786
3787 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3788 if (this_int_off != int_map_off) {
3789 // reset the map:
3790 int_map = 0;
3791 int_map_off = this_int_off;
3792 }
3793
3794 int subword_off = st_off - this_int_off;
3795 int_map |= right_n_bits(st_size) << subword_off;
3796 if ((int_map & FULL_MAP) == FULL_MAP) {
3797 return this_int_off; // we found a complete word init
3798 }
3799
3800 // Did this store hit or cross the word boundary?
3801 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3802 if (next_int_off == this_int_off + BytesPerInt) {
3803 // We passed the current int, without fully initializing it.
3804 int_map_off = next_int_off;
3805 int_map >>= BytesPerInt;
3806 } else if (next_int_off > this_int_off + BytesPerInt) {
3807 // We passed the current and next int.
3808 return this_int_off + BytesPerInt;
3809 }
3810 }
3811
3812 return -1;
3813 }
3814
3815
3816 // Called when the associated AllocateNode is expanded into CFG.
3817 // At this point, we may perform additional optimizations.
3818 // Linearize the stores by ascending offset, to make memory
3819 // activity as coherent as possible.
3820 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3821 intptr_t header_size,
3822 Node* size_in_bytes,
3823 PhaseGVN* phase) {
3824 assert(!is_complete(), "not already complete");
3825 assert(stores_are_sane(phase), "");
3826 assert(allocation() != NULL, "must be present");
3827
3828 remove_extra_zeroes();
3829
3830 if (ReduceFieldZeroing || ReduceBulkZeroing)
3831 // reduce instruction count for common initialization patterns
3832 coalesce_subword_stores(header_size, size_in_bytes, phase);
3833
3834 Node* zmem = zero_memory(); // initially zero memory state
3835 Node* inits = zmem; // accumulating a linearized chain of inits
3836 #ifdef ASSERT
3837 intptr_t first_offset = allocation()->minimum_header_size();
3838 intptr_t last_init_off = first_offset; // previous init offset
3839 intptr_t last_init_end = first_offset; // previous init offset+size
3840 intptr_t last_tile_end = first_offset; // previous tile offset+size
3841 #endif
3842 intptr_t zeroes_done = header_size;
3843
3844 bool do_zeroing = true; // we might give up if inits are very sparse
3845 int big_init_gaps = 0; // how many large gaps have we seen?
3846
3847 if (ZeroTLAB) do_zeroing = false;
3848 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3849
3850 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3851 Node* st = in(i);
3852 intptr_t st_off = get_store_offset(st, phase);
3853 if (st_off < 0)
3854 break; // unknown junk in the inits
3855 if (st->in(MemNode::Memory) != zmem)
3856 break; // complicated store chains somehow in list
3857
3858 int st_size = st->as_Store()->memory_size();
3859 intptr_t next_init_off = st_off + st_size;
3860
3861 if (do_zeroing && zeroes_done < next_init_off) {
3862 // See if this store needs a zero before it or under it.
3863 intptr_t zeroes_needed = st_off;
3864
3865 if (st_size < BytesPerInt) {
3866 // Look for subword stores which only partially initialize words.
3867 // If we find some, we must lay down some word-level zeroes first,
3868 // underneath the subword stores.
3869 //
3870 // Examples:
3871 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3872 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3873 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3874 //
3875 // Note: coalesce_subword_stores may have already done this,
3876 // if it was prompted by constant non-zero subword initializers.
3877 // But this case can still arise with non-constant stores.
3878
3879 intptr_t next_full_store = find_next_fullword_store(i, phase);
3880
3881 // In the examples above:
3882 // in(i) p q r s x y z
3883 // st_off 12 13 14 15 12 13 14
3884 // st_size 1 1 1 1 1 1 1
3885 // next_full_s. 12 16 16 16 16 16 16
3886 // z's_done 12 16 16 16 12 16 12
3887 // z's_needed 12 16 16 16 16 16 16
3888 // zsize 0 0 0 0 4 0 4
3889 if (next_full_store < 0) {
3890 // Conservative tack: Zero to end of current word.
3891 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3892 } else {
3893 // Zero to beginning of next fully initialized word.
3894 // Or, don't zero at all, if we are already in that word.
3895 assert(next_full_store >= zeroes_needed, "must go forward");
3896 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3897 zeroes_needed = next_full_store;
3898 }
3899 }
3900
3901 if (zeroes_needed > zeroes_done) {
3902 intptr_t zsize = zeroes_needed - zeroes_done;
3903 // Do some incremental zeroing on rawmem, in parallel with inits.
3904 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3905 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3906 zeroes_done, zeroes_needed,
3907 phase);
3908 zeroes_done = zeroes_needed;
3909 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3910 do_zeroing = false; // leave the hole, next time
3911 }
3912 }
3913
3914 // Collect the store and move on:
3915 st->set_req(MemNode::Memory, inits);
3916 inits = st; // put it on the linearized chain
3917 set_req(i, zmem); // unhook from previous position
3918
3919 if (zeroes_done == st_off)
3920 zeroes_done = next_init_off;
3921
3922 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3923
3924 #ifdef ASSERT
3925 // Various order invariants. Weaker than stores_are_sane because
3926 // a large constant tile can be filled in by smaller non-constant stores.
3927 assert(st_off >= last_init_off, "inits do not reverse");
3928 last_init_off = st_off;
3929 const Type* val = NULL;
3930 if (st_size >= BytesPerInt &&
3931 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3932 (int)val->basic_type() < (int)T_OBJECT) {
3933 assert(st_off >= last_tile_end, "tiles do not overlap");
3934 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3935 last_tile_end = MAX2(last_tile_end, next_init_off);
3936 } else {
3937 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3938 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3939 assert(st_off >= last_init_end, "inits do not overlap");
3940 last_init_end = next_init_off; // it's a non-tile
3941 }
3942 #endif //ASSERT
3943 }
3944
3945 remove_extra_zeroes(); // clear out all the zmems left over
3946 add_req(inits);
3947
3948 if (!ZeroTLAB) {
3949 // If anything remains to be zeroed, zero it all now.
3950 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3951 // if it is the last unused 4 bytes of an instance, forget about it
3952 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3953 if (zeroes_done + BytesPerLong >= size_limit) {
3954 assert(allocation() != NULL, "");
3955 if (allocation()->Opcode() == Op_Allocate) {
3956 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3957 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3958 if (zeroes_done == k->layout_helper())
3959 zeroes_done = size_limit;
3960 }
3961 }
3962 if (zeroes_done < size_limit) {
3963 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3964 zeroes_done, size_in_bytes, phase);
3965 }
3966 }
3967
3968 set_complete(phase);
3969 return rawmem;
3970 }
3971
3972
3973 #ifdef ASSERT
3974 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3975 if (is_complete())
3976 return true; // stores could be anything at this point
3977 assert(allocation() != NULL, "must be present");
3978 intptr_t last_off = allocation()->minimum_header_size();
3979 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3980 Node* st = in(i);
3981 intptr_t st_off = get_store_offset(st, phase);
3982 if (st_off < 0) continue; // ignore dead garbage
3983 if (last_off > st_off) {
3984 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
3985 this->dump(2);
3986 assert(false, "ascending store offsets");
3987 return false;
3988 }
3989 last_off = st_off + st->as_Store()->memory_size();
3990 }
3991 return true;
3992 }
3993 #endif //ASSERT
3994
3995
3996
3997
3998 //============================MergeMemNode=====================================
3999 //
4000 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4001 // contributing store or call operations. Each contributor provides the memory
4002 // state for a particular "alias type" (see Compile::alias_type). For example,
4003 // if a MergeMem has an input X for alias category #6, then any memory reference
4004 // to alias category #6 may use X as its memory state input, as an exact equivalent
4005 // to using the MergeMem as a whole.
4006 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4007 //
4008 // (Here, the <N> notation gives the index of the relevant adr_type.)
4009 //
4010 // In one special case (and more cases in the future), alias categories overlap.
4011 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4012 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4013 // it is exactly equivalent to that state W:
4014 // MergeMem(<Bot>: W) <==> W
4015 //
4016 // Usually, the merge has more than one input. In that case, where inputs
4017 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4018 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4019 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4020 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4021 //
4022 // A merge can take a "wide" memory state as one of its narrow inputs.
4023 // This simply means that the merge observes out only the relevant parts of
4024 // the wide input. That is, wide memory states arriving at narrow merge inputs
4025 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4026 //
4027 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4028 // and that memory slices "leak through":
4029 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4030 //
4031 // But, in such a cascade, repeated memory slices can "block the leak":
4032 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4033 //
4034 // In the last example, Y is not part of the combined memory state of the
4035 // outermost MergeMem. The system must, of course, prevent unschedulable
4036 // memory states from arising, so you can be sure that the state Y is somehow
4037 // a precursor to state Y'.
4038 //
4039 //
4040 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4041 // of each MergeMemNode array are exactly the numerical alias indexes, including
4042 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4043 // Compile::alias_type (and kin) produce and manage these indexes.
4044 //
4045 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4046 // (Note that this provides quick access to the top node inside MergeMem methods,
4047 // without the need to reach out via TLS to Compile::current.)
4048 //
4049 // As a consequence of what was just described, a MergeMem that represents a full
4050 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4051 // containing all alias categories.
4052 //
4053 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4054 //
4055 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4056 // a memory state for the alias type <N>, or else the top node, meaning that
4057 // there is no particular input for that alias type. Note that the length of
4058 // a MergeMem is variable, and may be extended at any time to accommodate new
4059 // memory states at larger alias indexes. When merges grow, they are of course
4060 // filled with "top" in the unused in() positions.
4061 //
4062 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4063 // (Top was chosen because it works smoothly with passes like GCM.)
4064 //
4065 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4066 // the type of random VM bits like TLS references.) Since it is always the
4067 // first non-Bot memory slice, some low-level loops use it to initialize an
4068 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4069 //
4070 //
4071 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4072 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4073 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4074 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4075 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4076 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4077 //
4078 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4079 // really that different from the other memory inputs. An abbreviation called
4080 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4081 //
4082 //
4083 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4084 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4085 // that "emerges though" the base memory will be marked as excluding the alias types
4086 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4087 //
4088 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4089 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4090 //
4091 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4092 // (It is currently unimplemented.) As you can see, the resulting merge is
4093 // actually a disjoint union of memory states, rather than an overlay.
4094 //
4095
4096 //------------------------------MergeMemNode-----------------------------------
4097 Node* MergeMemNode::make_empty_memory() {
4098 Node* empty_memory = (Node*) Compile::current()->top();
4099 assert(empty_memory->is_top(), "correct sentinel identity");
4100 return empty_memory;
4101 }
4102
4103 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4104 init_class_id(Class_MergeMem);
4105 // all inputs are nullified in Node::Node(int)
4106 // set_input(0, NULL); // no control input
4107
4108 // Initialize the edges uniformly to top, for starters.
4109 Node* empty_mem = make_empty_memory();
4110 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4111 init_req(i,empty_mem);
4112 }
4113 assert(empty_memory() == empty_mem, "");
4114
4115 if( new_base != NULL && new_base->is_MergeMem() ) {
4116 MergeMemNode* mdef = new_base->as_MergeMem();
4117 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4118 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4119 mms.set_memory(mms.memory2());
4120 }
4121 assert(base_memory() == mdef->base_memory(), "");
4122 } else {
4123 set_base_memory(new_base);
4124 }
4125 }
4126
4127 // Make a new, untransformed MergeMem with the same base as 'mem'.
4128 // If mem is itself a MergeMem, populate the result with the same edges.
4129 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
4130 return new(C) MergeMemNode(mem);
4131 }
4132
4133 //------------------------------cmp--------------------------------------------
4134 uint MergeMemNode::hash() const { return NO_HASH; }
4135 uint MergeMemNode::cmp( const Node &n ) const {
4136 return (&n == this); // Always fail except on self
4137 }
4138
4139 //------------------------------Identity---------------------------------------
4140 Node* MergeMemNode::Identity(PhaseTransform *phase) {
4141 // Identity if this merge point does not record any interesting memory
4142 // disambiguations.
4143 Node* base_mem = base_memory();
4144 Node* empty_mem = empty_memory();
4145 if (base_mem != empty_mem) { // Memory path is not dead?
4146 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4147 Node* mem = in(i);
4148 if (mem != empty_mem && mem != base_mem) {
4149 return this; // Many memory splits; no change
4150 }
4151 }
4152 }
4153 return base_mem; // No memory splits; ID on the one true input
4154 }
4155
4156 //------------------------------Ideal------------------------------------------
4157 // This method is invoked recursively on chains of MergeMem nodes
4158 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4159 // Remove chain'd MergeMems
4160 //
4161 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4162 // relative to the "in(Bot)". Since we are patching both at the same time,
4163 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4164 // but rewrite each "in(i)" relative to the new "in(Bot)".
4165 Node *progress = NULL;
4166
4167
4168 Node* old_base = base_memory();
4169 Node* empty_mem = empty_memory();
4170 if (old_base == empty_mem)
4171 return NULL; // Dead memory path.
4172
4173 MergeMemNode* old_mbase;
4174 if (old_base != NULL && old_base->is_MergeMem())
4175 old_mbase = old_base->as_MergeMem();
4176 else
4177 old_mbase = NULL;
4178 Node* new_base = old_base;
4179
4180 // simplify stacked MergeMems in base memory
4181 if (old_mbase) new_base = old_mbase->base_memory();
4182
4183 // the base memory might contribute new slices beyond my req()
4184 if (old_mbase) grow_to_match(old_mbase);
4185
4186 // Look carefully at the base node if it is a phi.
4187 PhiNode* phi_base;
4188 if (new_base != NULL && new_base->is_Phi())
4189 phi_base = new_base->as_Phi();
4190 else
4191 phi_base = NULL;
4192
4193 Node* phi_reg = NULL;
4194 uint phi_len = (uint)-1;
4195 if (phi_base != NULL && !phi_base->is_copy()) {
4196 // do not examine phi if degraded to a copy
4197 phi_reg = phi_base->region();
4198 phi_len = phi_base->req();
4199 // see if the phi is unfinished
4200 for (uint i = 1; i < phi_len; i++) {
4201 if (phi_base->in(i) == NULL) {
4202 // incomplete phi; do not look at it yet!
4203 phi_reg = NULL;
4204 phi_len = (uint)-1;
4205 break;
4206 }
4207 }
4208 }
4209
4210 // Note: We do not call verify_sparse on entry, because inputs
4211 // can normalize to the base_memory via subsume_node or similar
4212 // mechanisms. This method repairs that damage.
4213
4214 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4215
4216 // Look at each slice.
4217 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4218 Node* old_in = in(i);
4219 // calculate the old memory value
4220 Node* old_mem = old_in;
4221 if (old_mem == empty_mem) old_mem = old_base;
4222 assert(old_mem == memory_at(i), "");
4223
4224 // maybe update (reslice) the old memory value
4225
4226 // simplify stacked MergeMems
4227 Node* new_mem = old_mem;
4228 MergeMemNode* old_mmem;
4229 if (old_mem != NULL && old_mem->is_MergeMem())
4230 old_mmem = old_mem->as_MergeMem();
4231 else
4232 old_mmem = NULL;
4233 if (old_mmem == this) {
4234 // This can happen if loops break up and safepoints disappear.
4235 // A merge of BotPtr (default) with a RawPtr memory derived from a
4236 // safepoint can be rewritten to a merge of the same BotPtr with
4237 // the BotPtr phi coming into the loop. If that phi disappears
4238 // also, we can end up with a self-loop of the mergemem.
4239 // In general, if loops degenerate and memory effects disappear,
4240 // a mergemem can be left looking at itself. This simply means
4241 // that the mergemem's default should be used, since there is
4242 // no longer any apparent effect on this slice.
4243 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4244 // from start. Update the input to TOP.
4245 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4246 }
4247 else if (old_mmem != NULL) {
4248 new_mem = old_mmem->memory_at(i);
4249 }
4250 // else preceding memory was not a MergeMem
4251
4252 // replace equivalent phis (unfortunately, they do not GVN together)
4253 if (new_mem != NULL && new_mem != new_base &&
4254 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4255 if (new_mem->is_Phi()) {
4256 PhiNode* phi_mem = new_mem->as_Phi();
4257 for (uint i = 1; i < phi_len; i++) {
4258 if (phi_base->in(i) != phi_mem->in(i)) {
4259 phi_mem = NULL;
4260 break;
4261 }
4262 }
4263 if (phi_mem != NULL) {
4264 // equivalent phi nodes; revert to the def
4265 new_mem = new_base;
4266 }
4267 }
4268 }
4269
4270 // maybe store down a new value
4271 Node* new_in = new_mem;
4272 if (new_in == new_base) new_in = empty_mem;
4273
4274 if (new_in != old_in) {
4275 // Warning: Do not combine this "if" with the previous "if"
4276 // A memory slice might have be be rewritten even if it is semantically
4277 // unchanged, if the base_memory value has changed.
4278 set_req(i, new_in);
4279 progress = this; // Report progress
4280 }
4281 }
4282
4283 if (new_base != old_base) {
4284 set_req(Compile::AliasIdxBot, new_base);
4285 // Don't use set_base_memory(new_base), because we need to update du.
4286 assert(base_memory() == new_base, "");
4287 progress = this;
4288 }
4289
4290 if( base_memory() == this ) {
4291 // a self cycle indicates this memory path is dead
4292 set_req(Compile::AliasIdxBot, empty_mem);
4293 }
4294
4295 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4296 // Recursion must occur after the self cycle check above
4297 if( base_memory()->is_MergeMem() ) {
4298 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4299 Node *m = phase->transform(new_mbase); // Rollup any cycles
4300 if( m != NULL && (m->is_top() ||
4301 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
4302 // propagate rollup of dead cycle to self
4303 set_req(Compile::AliasIdxBot, empty_mem);
4304 }
4305 }
4306
4307 if( base_memory() == empty_mem ) {
4308 progress = this;
4309 // Cut inputs during Parse phase only.
4310 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4311 if( !can_reshape ) {
4312 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4313 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4314 }
4315 }
4316 }
4317
4318 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4319 // Check if PhiNode::Ideal's "Split phis through memory merges"
4320 // transform should be attempted. Look for this->phi->this cycle.
4321 uint merge_width = req();
4322 if (merge_width > Compile::AliasIdxRaw) {
4323 PhiNode* phi = base_memory()->as_Phi();
4324 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4325 if (phi->in(i) == this) {
4326 phase->is_IterGVN()->_worklist.push(phi);
4327 break;
4328 }
4329 }
4330 }
4331 }
4332
4333 assert(progress || verify_sparse(), "please, no dups of base");
4334 return progress;
4335 }
4336
4337 //-------------------------set_base_memory-------------------------------------
4338 void MergeMemNode::set_base_memory(Node *new_base) {
4339 Node* empty_mem = empty_memory();
4340 set_req(Compile::AliasIdxBot, new_base);
4341 assert(memory_at(req()) == new_base, "must set default memory");
4342 // Clear out other occurrences of new_base:
4343 if (new_base != empty_mem) {
4344 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4345 if (in(i) == new_base) set_req(i, empty_mem);
4346 }
4347 }
4348 }
4349
4350 //------------------------------out_RegMask------------------------------------
4351 const RegMask &MergeMemNode::out_RegMask() const {
4352 return RegMask::Empty;
4353 }
4354
4355 //------------------------------dump_spec--------------------------------------
4356 #ifndef PRODUCT
4357 void MergeMemNode::dump_spec(outputStream *st) const {
4358 st->print(" {");
4359 Node* base_mem = base_memory();
4360 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4361 Node* mem = memory_at(i);
4362 if (mem == base_mem) { st->print(" -"); continue; }
4363 st->print( " N%d:", mem->_idx );
4364 Compile::current()->get_adr_type(i)->dump_on(st);
4365 }
4366 st->print(" }");
4367 }
4368 #endif // !PRODUCT
4369
4370
4371 #ifdef ASSERT
4372 static bool might_be_same(Node* a, Node* b) {
4373 if (a == b) return true;
4374 if (!(a->is_Phi() || b->is_Phi())) return false;
4375 // phis shift around during optimization
4376 return true; // pretty stupid...
4377 }
4378
4379 // verify a narrow slice (either incoming or outgoing)
4380 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4381 if (!VerifyAliases) return; // don't bother to verify unless requested
4382 if (is_error_reported()) return; // muzzle asserts when debugging an error
4383 if (Node::in_dump()) return; // muzzle asserts when printing
4384 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4385 assert(n != NULL, "");
4386 // Elide intervening MergeMem's
4387 while (n->is_MergeMem()) {
4388 n = n->as_MergeMem()->memory_at(alias_idx);
4389 }
4390 Compile* C = Compile::current();
4391 const TypePtr* n_adr_type = n->adr_type();
4392 if (n == m->empty_memory()) {
4393 // Implicit copy of base_memory()
4394 } else if (n_adr_type != TypePtr::BOTTOM) {
4395 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4396 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4397 } else {
4398 // A few places like make_runtime_call "know" that VM calls are narrow,
4399 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4400 bool expected_wide_mem = false;
4401 if (n == m->base_memory()) {
4402 expected_wide_mem = true;
4403 } else if (alias_idx == Compile::AliasIdxRaw ||
4404 n == m->memory_at(Compile::AliasIdxRaw)) {
4405 expected_wide_mem = true;
4406 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4407 // memory can "leak through" calls on channels that
4408 // are write-once. Allow this also.
4409 expected_wide_mem = true;
4410 }
4411 assert(expected_wide_mem, "expected narrow slice replacement");
4412 }
4413 }
4414 #else // !ASSERT
4415 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4416 #endif
4417
4418
4419 //-----------------------------memory_at---------------------------------------
4420 Node* MergeMemNode::memory_at(uint alias_idx) const {
4421 assert(alias_idx >= Compile::AliasIdxRaw ||
4422 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4423 "must avoid base_memory and AliasIdxTop");
4424
4425 // Otherwise, it is a narrow slice.
4426 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4427 Compile *C = Compile::current();
4428 if (is_empty_memory(n)) {
4429 // the array is sparse; empty slots are the "top" node
4430 n = base_memory();
4431 assert(Node::in_dump()
4432 || n == NULL || n->bottom_type() == Type::TOP
4433 || n->adr_type() == NULL // address is TOP
4434 || n->adr_type() == TypePtr::BOTTOM
4435 || n->adr_type() == TypeRawPtr::BOTTOM
4436 || Compile::current()->AliasLevel() == 0,
4437 "must be a wide memory");
4438 // AliasLevel == 0 if we are organizing the memory states manually.
4439 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4440 } else {
4441 // make sure the stored slice is sane
4442 #ifdef ASSERT
4443 if (is_error_reported() || Node::in_dump()) {
4444 } else if (might_be_same(n, base_memory())) {
4445 // Give it a pass: It is a mostly harmless repetition of the base.
4446 // This can arise normally from node subsumption during optimization.
4447 } else {
4448 verify_memory_slice(this, alias_idx, n);
4449 }
4450 #endif
4451 }
4452 return n;
4453 }
4454
4455 //---------------------------set_memory_at-------------------------------------
4456 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4457 verify_memory_slice(this, alias_idx, n);
4458 Node* empty_mem = empty_memory();
4459 if (n == base_memory()) n = empty_mem; // collapse default
4460 uint need_req = alias_idx+1;
4461 if (req() < need_req) {
4462 if (n == empty_mem) return; // already the default, so do not grow me
4463 // grow the sparse array
4464 do {
4465 add_req(empty_mem);
4466 } while (req() < need_req);
4467 }
4468 set_req( alias_idx, n );
4469 }
4470
4471
4472
4473 //--------------------------iteration_setup------------------------------------
4474 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4475 if (other != NULL) {
4476 grow_to_match(other);
4477 // invariant: the finite support of mm2 is within mm->req()
4478 #ifdef ASSERT
4479 for (uint i = req(); i < other->req(); i++) {
4480 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4481 }
4482 #endif
4483 }
4484 // Replace spurious copies of base_memory by top.
4485 Node* base_mem = base_memory();
4486 if (base_mem != NULL && !base_mem->is_top()) {
4487 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4488 if (in(i) == base_mem)
4489 set_req(i, empty_memory());
4490 }
4491 }
4492 }
4493
4494 //---------------------------grow_to_match-------------------------------------
4495 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4496 Node* empty_mem = empty_memory();
4497 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4498 // look for the finite support of the other memory
4499 for (uint i = other->req(); --i >= req(); ) {
4500 if (other->in(i) != empty_mem) {
4501 uint new_len = i+1;
4502 while (req() < new_len) add_req(empty_mem);
4503 break;
4504 }
4505 }
4506 }
4507
4508 //---------------------------verify_sparse-------------------------------------
4509 #ifndef PRODUCT
4510 bool MergeMemNode::verify_sparse() const {
4511 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4512 Node* base_mem = base_memory();
4513 // The following can happen in degenerate cases, since empty==top.
4514 if (is_empty_memory(base_mem)) return true;
4515 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4516 assert(in(i) != NULL, "sane slice");
4517 if (in(i) == base_mem) return false; // should have been the sentinel value!
4518 }
4519 return true;
4520 }
4521
4522 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4523 Node* n;
4524 n = mm->in(idx);
4525 if (mem == n) return true; // might be empty_memory()
4526 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4527 if (mem == n) return true;
4528 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4529 if (mem == n) return true;
4530 if (n == NULL) break;
4531 }
4532 return false;
4533 }
4534 #endif // !PRODUCT