1 /*
2 * Copyright 1997-2007 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
24
25 // Portions of code courtesy of Clifford Click
26
27 // Optimization - Graph Style
28
29 #include "incls/_precompiled.incl"
30 #include "incls/_memnode.cpp.incl"
31
32 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
33
34 //=============================================================================
35 uint MemNode::size_of() const { return sizeof(*this); }
36
37 const TypePtr *MemNode::adr_type() const {
38 Node* adr = in(Address);
39 const TypePtr* cross_check = NULL;
40 DEBUG_ONLY(cross_check = _adr_type);
41 return calculate_adr_type(adr->bottom_type(), cross_check);
42 }
43
44 #ifndef PRODUCT
45 void MemNode::dump_spec(outputStream *st) const {
46 if (in(Address) == NULL) return; // node is dead
47 #ifndef ASSERT
48 // fake the missing field
49 const TypePtr* _adr_type = NULL;
50 if (in(Address) != NULL)
51 _adr_type = in(Address)->bottom_type()->isa_ptr();
52 #endif
53 dump_adr_type(this, _adr_type, st);
54
55 Compile* C = Compile::current();
56 if( C->alias_type(_adr_type)->is_volatile() )
57 st->print(" Volatile!");
58 }
59
60 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
61 st->print(" @");
62 if (adr_type == NULL) {
63 st->print("NULL");
64 } else {
65 adr_type->dump_on(st);
66 Compile* C = Compile::current();
67 Compile::AliasType* atp = NULL;
68 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
69 if (atp == NULL)
70 st->print(", idx=?\?;");
71 else if (atp->index() == Compile::AliasIdxBot)
72 st->print(", idx=Bot;");
73 else if (atp->index() == Compile::AliasIdxTop)
74 st->print(", idx=Top;");
75 else if (atp->index() == Compile::AliasIdxRaw)
76 st->print(", idx=Raw;");
77 else {
78 ciField* field = atp->field();
79 if (field) {
80 st->print(", name=");
81 field->print_name_on(st);
82 }
83 st->print(", idx=%d;", atp->index());
84 }
85 }
86 }
87
88 extern void print_alias_types();
89
90 #endif
91
92 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
93 const TypeOopPtr *tinst = t_adr->isa_oopptr();
94 if (tinst == NULL || !tinst->is_instance_field())
95 return mchain; // don't try to optimize non-instance types
96 uint instance_id = tinst->instance_id();
97 Node *prev = NULL;
98 Node *result = mchain;
99 while (prev != result) {
100 prev = result;
101 // skip over a call which does not affect this memory slice
102 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
103 Node *proj_in = result->in(0);
104 if (proj_in->is_Call()) {
105 CallNode *call = proj_in->as_Call();
106 if (!call->may_modify(t_adr, phase)) {
107 result = call->in(TypeFunc::Memory);
108 }
109 } else if (proj_in->is_Initialize()) {
110 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
111 // Stop if this is the initialization for the object instance which
112 // which contains this memory slice, otherwise skip over it.
113 if (alloc != NULL && alloc->_idx != instance_id) {
114 result = proj_in->in(TypeFunc::Memory);
115 }
116 } else if (proj_in->is_MemBar()) {
117 result = proj_in->in(TypeFunc::Memory);
118 }
119 } else if (result->is_MergeMem()) {
120 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
121 }
122 }
123 return result;
124 }
125
126 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
127 const TypeOopPtr *t_oop = t_adr->isa_oopptr();
128 bool is_instance = (t_oop != NULL) && t_oop->is_instance_field();
129 PhaseIterGVN *igvn = phase->is_IterGVN();
130 Node *result = mchain;
131 result = optimize_simple_memory_chain(result, t_adr, phase);
132 if (is_instance && igvn != NULL && result->is_Phi()) {
133 PhiNode *mphi = result->as_Phi();
134 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
135 const TypePtr *t = mphi->adr_type();
136 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM) {
137 // clone the Phi with our address type
138 result = mphi->split_out_instance(t_adr, igvn);
139 } else {
140 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
141 }
142 }
143 return result;
144 }
145
146 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
147 uint alias_idx = phase->C->get_alias_index(tp);
148 Node *mem = mmem;
149 #ifdef ASSERT
150 {
151 // Check that current type is consistent with the alias index used during graph construction
152 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
153 bool consistent = adr_check == NULL || adr_check->empty() ||
154 phase->C->must_alias(adr_check, alias_idx );
155 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
156 if( !consistent && adr_check != NULL && !adr_check->empty() &&
157 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
158 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
159 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
160 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
161 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
162 // don't assert if it is dead code.
163 consistent = true;
164 }
165 if( !consistent ) {
166 st->print("alias_idx==%d, adr_check==", alias_idx);
167 if( adr_check == NULL ) {
168 st->print("NULL");
169 } else {
170 adr_check->dump();
171 }
172 st->cr();
173 print_alias_types();
174 assert(consistent, "adr_check must match alias idx");
175 }
176 }
177 #endif
178 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
179 // means an array I have not precisely typed yet. Do not do any
180 // alias stuff with it any time soon.
181 const TypeOopPtr *tinst = tp->isa_oopptr();
182 if( tp->base() != Type::AnyPtr &&
183 !(tinst &&
184 tinst->klass()->is_java_lang_Object() &&
185 tinst->offset() == Type::OffsetBot) ) {
186 // compress paths and change unreachable cycles to TOP
187 // If not, we can update the input infinitely along a MergeMem cycle
188 // Equivalent code in PhiNode::Ideal
189 Node* m = phase->transform(mmem);
190 // If tranformed to a MergeMem, get the desired slice
191 // Otherwise the returned node represents memory for every slice
192 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
193 // Update input if it is progress over what we have now
194 }
195 return mem;
196 }
197
198 //--------------------------Ideal_common---------------------------------------
199 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
200 // Unhook non-raw memories from complete (macro-expanded) initializations.
201 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
202 // If our control input is a dead region, kill all below the region
203 Node *ctl = in(MemNode::Control);
204 if (ctl && remove_dead_region(phase, can_reshape))
205 return this;
206
207 // Ignore if memory is dead, or self-loop
208 Node *mem = in(MemNode::Memory);
209 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
210 assert( mem != this, "dead loop in MemNode::Ideal" );
211
212 Node *address = in(MemNode::Address);
213 const Type *t_adr = phase->type( address );
214 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
215
216 // Avoid independent memory operations
217 Node* old_mem = mem;
218
219 // The code which unhooks non-raw memories from complete (macro-expanded)
220 // initializations was removed. After macro-expansion all stores catched
221 // by Initialize node became raw stores and there is no information
222 // which memory slices they modify. So it is unsafe to move any memory
223 // operation above these stores. Also in most cases hooked non-raw memories
224 // were already unhooked by using information from detect_ptr_independence()
225 // and find_previous_store().
226
227 if (mem->is_MergeMem()) {
228 MergeMemNode* mmem = mem->as_MergeMem();
229 const TypePtr *tp = t_adr->is_ptr();
230
231 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
232 }
233
234 if (mem != old_mem) {
235 set_req(MemNode::Memory, mem);
236 return this;
237 }
238
239 // let the subclass continue analyzing...
240 return NULL;
241 }
242
243 // Helper function for proving some simple control dominations.
244 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
245 // Already assumes that 'dom' is available at 'sub', and that 'sub'
246 // is not a constant (dominated by the method's StartNode).
247 // Used by MemNode::find_previous_store to prove that the
248 // control input of a memory operation predates (dominates)
249 // an allocation it wants to look past.
250 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
251 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
252 return false; // Conservative answer for dead code
253
254 // Check 'dom'.
255 dom = dom->find_exact_control(dom);
256 if (dom == NULL || dom->is_top())
257 return false; // Conservative answer for dead code
258
259 if (dom->is_Start() || dom->is_Root() || dom == sub)
260 return true;
261
262 // 'dom' dominates 'sub' if its control edge and control edges
263 // of all its inputs dominate or equal to sub's control edge.
264
265 // Currently 'sub' is either Allocate, Initialize or Start nodes.
266 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start(), "expecting only these nodes");
267
268 // Get control edge of 'sub'.
269 sub = sub->find_exact_control(sub->in(0));
270 if (sub == NULL || sub->is_top())
271 return false; // Conservative answer for dead code
272
273 assert(sub->is_CFG(), "expecting control");
274
275 if (sub == dom)
276 return true;
277
278 if (sub->is_Start() || sub->is_Root())
279 return false;
280
281 {
282 // Check all control edges of 'dom'.
283
284 ResourceMark rm;
285 Arena* arena = Thread::current()->resource_area();
286 Node_List nlist(arena);
287 Unique_Node_List dom_list(arena);
288
289 dom_list.push(dom);
290 bool only_dominating_controls = false;
291
292 for (uint next = 0; next < dom_list.size(); next++) {
293 Node* n = dom_list.at(next);
294 if (!n->is_CFG() && n->pinned()) {
295 // Check only own control edge for pinned non-control nodes.
296 n = n->find_exact_control(n->in(0));
297 if (n == NULL || n->is_top())
298 return false; // Conservative answer for dead code
299 assert(n->is_CFG(), "expecting control");
300 }
301 if (n->is_Start() || n->is_Root()) {
302 only_dominating_controls = true;
303 } else if (n->is_CFG()) {
304 if (n->dominates(sub, nlist))
305 only_dominating_controls = true;
306 else
307 return false;
308 } else {
309 // First, own control edge.
310 Node* m = n->find_exact_control(n->in(0));
311 if (m == NULL)
312 continue;
313 if (m->is_top())
314 return false; // Conservative answer for dead code
315 dom_list.push(m);
316
317 // Now, the rest of edges.
318 uint cnt = n->req();
319 for (uint i = 1; i < cnt; i++) {
320 m = n->find_exact_control(n->in(i));
321 if (m == NULL || m->is_top())
322 continue;
323 dom_list.push(m);
324 }
325 }
326 }
327 return only_dominating_controls;
328 }
329 }
330
331 //---------------------detect_ptr_independence---------------------------------
332 // Used by MemNode::find_previous_store to prove that two base
333 // pointers are never equal.
334 // The pointers are accompanied by their associated allocations,
335 // if any, which have been previously discovered by the caller.
336 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
337 Node* p2, AllocateNode* a2,
338 PhaseTransform* phase) {
339 // Attempt to prove that these two pointers cannot be aliased.
340 // They may both manifestly be allocations, and they should differ.
341 // Or, if they are not both allocations, they can be distinct constants.
342 // Otherwise, one is an allocation and the other a pre-existing value.
343 if (a1 == NULL && a2 == NULL) { // neither an allocation
344 return (p1 != p2) && p1->is_Con() && p2->is_Con();
345 } else if (a1 != NULL && a2 != NULL) { // both allocations
346 return (a1 != a2);
347 } else if (a1 != NULL) { // one allocation a1
348 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
349 return all_controls_dominate(p2, a1);
350 } else { //(a2 != NULL) // one allocation a2
351 return all_controls_dominate(p1, a2);
352 }
353 return false;
354 }
355
356
357 // The logic for reordering loads and stores uses four steps:
358 // (a) Walk carefully past stores and initializations which we
359 // can prove are independent of this load.
360 // (b) Observe that the next memory state makes an exact match
361 // with self (load or store), and locate the relevant store.
362 // (c) Ensure that, if we were to wire self directly to the store,
363 // the optimizer would fold it up somehow.
364 // (d) Do the rewiring, and return, depending on some other part of
365 // the optimizer to fold up the load.
366 // This routine handles steps (a) and (b). Steps (c) and (d) are
367 // specific to loads and stores, so they are handled by the callers.
368 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
369 //
370 Node* MemNode::find_previous_store(PhaseTransform* phase) {
371 Node* ctrl = in(MemNode::Control);
372 Node* adr = in(MemNode::Address);
373 intptr_t offset = 0;
374 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
375 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
376
377 if (offset == Type::OffsetBot)
378 return NULL; // cannot unalias unless there are precise offsets
379
380 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
381
382 intptr_t size_in_bytes = memory_size();
383
384 Node* mem = in(MemNode::Memory); // start searching here...
385
386 int cnt = 50; // Cycle limiter
387 for (;;) { // While we can dance past unrelated stores...
388 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
389
390 if (mem->is_Store()) {
391 Node* st_adr = mem->in(MemNode::Address);
392 intptr_t st_offset = 0;
393 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
394 if (st_base == NULL)
395 break; // inscrutable pointer
396 if (st_offset != offset && st_offset != Type::OffsetBot) {
397 const int MAX_STORE = BytesPerLong;
398 if (st_offset >= offset + size_in_bytes ||
399 st_offset <= offset - MAX_STORE ||
400 st_offset <= offset - mem->as_Store()->memory_size()) {
401 // Success: The offsets are provably independent.
402 // (You may ask, why not just test st_offset != offset and be done?
403 // The answer is that stores of different sizes can co-exist
404 // in the same sequence of RawMem effects. We sometimes initialize
405 // a whole 'tile' of array elements with a single jint or jlong.)
406 mem = mem->in(MemNode::Memory);
407 continue; // (a) advance through independent store memory
408 }
409 }
410 if (st_base != base &&
411 detect_ptr_independence(base, alloc,
412 st_base,
413 AllocateNode::Ideal_allocation(st_base, phase),
414 phase)) {
415 // Success: The bases are provably independent.
416 mem = mem->in(MemNode::Memory);
417 continue; // (a) advance through independent store memory
418 }
419
420 // (b) At this point, if the bases or offsets do not agree, we lose,
421 // since we have not managed to prove 'this' and 'mem' independent.
422 if (st_base == base && st_offset == offset) {
423 return mem; // let caller handle steps (c), (d)
424 }
425
426 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
427 InitializeNode* st_init = mem->in(0)->as_Initialize();
428 AllocateNode* st_alloc = st_init->allocation();
429 if (st_alloc == NULL)
430 break; // something degenerated
431 bool known_identical = false;
432 bool known_independent = false;
433 if (alloc == st_alloc)
434 known_identical = true;
435 else if (alloc != NULL)
436 known_independent = true;
437 else if (all_controls_dominate(this, st_alloc))
438 known_independent = true;
439
440 if (known_independent) {
441 // The bases are provably independent: Either they are
442 // manifestly distinct allocations, or else the control
443 // of this load dominates the store's allocation.
444 int alias_idx = phase->C->get_alias_index(adr_type());
445 if (alias_idx == Compile::AliasIdxRaw) {
446 mem = st_alloc->in(TypeFunc::Memory);
447 } else {
448 mem = st_init->memory(alias_idx);
449 }
450 continue; // (a) advance through independent store memory
451 }
452
453 // (b) at this point, if we are not looking at a store initializing
454 // the same allocation we are loading from, we lose.
455 if (known_identical) {
456 // From caller, can_see_stored_value will consult find_captured_store.
457 return mem; // let caller handle steps (c), (d)
458 }
459
460 } else if (addr_t != NULL && addr_t->is_instance_field()) {
461 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
462 if (mem->is_Proj() && mem->in(0)->is_Call()) {
463 CallNode *call = mem->in(0)->as_Call();
464 if (!call->may_modify(addr_t, phase)) {
465 mem = call->in(TypeFunc::Memory);
466 continue; // (a) advance through independent call memory
467 }
468 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
469 mem = mem->in(0)->in(TypeFunc::Memory);
470 continue; // (a) advance through independent MemBar memory
471 } else if (mem->is_MergeMem()) {
472 int alias_idx = phase->C->get_alias_index(adr_type());
473 mem = mem->as_MergeMem()->memory_at(alias_idx);
474 continue; // (a) advance through independent MergeMem memory
475 }
476 }
477
478 // Unless there is an explicit 'continue', we must bail out here,
479 // because 'mem' is an inscrutable memory state (e.g., a call).
480 break;
481 }
482
483 return NULL; // bail out
484 }
485
486 //----------------------calculate_adr_type-------------------------------------
487 // Helper function. Notices when the given type of address hits top or bottom.
488 // Also, asserts a cross-check of the type against the expected address type.
489 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
490 if (t == Type::TOP) return NULL; // does not touch memory any more?
491 #ifdef PRODUCT
492 cross_check = NULL;
493 #else
494 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
495 #endif
496 const TypePtr* tp = t->isa_ptr();
497 if (tp == NULL) {
498 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
499 return TypePtr::BOTTOM; // touches lots of memory
500 } else {
501 #ifdef ASSERT
502 // %%%% [phh] We don't check the alias index if cross_check is
503 // TypeRawPtr::BOTTOM. Needs to be investigated.
504 if (cross_check != NULL &&
505 cross_check != TypePtr::BOTTOM &&
506 cross_check != TypeRawPtr::BOTTOM) {
507 // Recheck the alias index, to see if it has changed (due to a bug).
508 Compile* C = Compile::current();
509 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
510 "must stay in the original alias category");
511 // The type of the address must be contained in the adr_type,
512 // disregarding "null"-ness.
513 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
514 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
515 assert(cross_check->meet(tp_notnull) == cross_check,
516 "real address must not escape from expected memory type");
517 }
518 #endif
519 return tp;
520 }
521 }
522
523 //------------------------adr_phi_is_loop_invariant----------------------------
524 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
525 // loop is loop invariant. Make a quick traversal of Phi and associated
526 // CastPP nodes, looking to see if they are a closed group within the loop.
527 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
528 // The idea is that the phi-nest must boil down to only CastPP nodes
529 // with the same data. This implies that any path into the loop already
530 // includes such a CastPP, and so the original cast, whatever its input,
531 // must be covered by an equivalent cast, with an earlier control input.
532 ResourceMark rm;
533
534 // The loop entry input of the phi should be the unique dominating
535 // node for every Phi/CastPP in the loop.
536 Unique_Node_List closure;
537 closure.push(adr_phi->in(LoopNode::EntryControl));
538
539 // Add the phi node and the cast to the worklist.
540 Unique_Node_List worklist;
541 worklist.push(adr_phi);
542 if( cast != NULL ){
543 if( !cast->is_ConstraintCast() ) return false;
544 worklist.push(cast);
545 }
546
547 // Begin recursive walk of phi nodes.
548 while( worklist.size() ){
549 // Take a node off the worklist
550 Node *n = worklist.pop();
551 if( !closure.member(n) ){
552 // Add it to the closure.
553 closure.push(n);
554 // Make a sanity check to ensure we don't waste too much time here.
555 if( closure.size() > 20) return false;
556 // This node is OK if:
557 // - it is a cast of an identical value
558 // - or it is a phi node (then we add its inputs to the worklist)
559 // Otherwise, the node is not OK, and we presume the cast is not invariant
560 if( n->is_ConstraintCast() ){
561 worklist.push(n->in(1));
562 } else if( n->is_Phi() ) {
563 for( uint i = 1; i < n->req(); i++ ) {
564 worklist.push(n->in(i));
565 }
566 } else {
567 return false;
568 }
569 }
570 }
571
572 // Quit when the worklist is empty, and we've found no offending nodes.
573 return true;
574 }
575
576 //------------------------------Ideal_DU_postCCP-------------------------------
577 // Find any cast-away of null-ness and keep its control. Null cast-aways are
578 // going away in this pass and we need to make this memory op depend on the
579 // gating null check.
580
581 // I tried to leave the CastPP's in. This makes the graph more accurate in
582 // some sense; we get to keep around the knowledge that an oop is not-null
583 // after some test. Alas, the CastPP's interfere with GVN (some values are
584 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
585 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
586 // some of the more trivial cases in the optimizer. Removing more useless
587 // Phi's started allowing Loads to illegally float above null checks. I gave
588 // up on this approach. CNC 10/20/2000
589 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
590 Node *ctr = in(MemNode::Control);
591 Node *mem = in(MemNode::Memory);
592 Node *adr = in(MemNode::Address);
593 Node *skipped_cast = NULL;
594 // Need a null check? Regular static accesses do not because they are
595 // from constant addresses. Array ops are gated by the range check (which
596 // always includes a NULL check). Just check field ops.
597 if( !ctr ) {
598 // Scan upwards for the highest location we can place this memory op.
599 while( true ) {
600 switch( adr->Opcode() ) {
601
602 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
603 adr = adr->in(AddPNode::Base);
604 continue;
605
606 case Op_DecodeN: // No change to NULL-ness, so peek thru
607 adr = adr->in(1);
608 continue;
609
610 case Op_CastPP:
611 // If the CastPP is useless, just peek on through it.
612 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
613 // Remember the cast that we've peeked though. If we peek
614 // through more than one, then we end up remembering the highest
615 // one, that is, if in a loop, the one closest to the top.
616 skipped_cast = adr;
617 adr = adr->in(1);
618 continue;
619 }
620 // CastPP is going away in this pass! We need this memory op to be
621 // control-dependent on the test that is guarding the CastPP.
622 ccp->hash_delete(this);
623 set_req(MemNode::Control, adr->in(0));
624 ccp->hash_insert(this);
625 return this;
626
627 case Op_Phi:
628 // Attempt to float above a Phi to some dominating point.
629 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
630 // If we've already peeked through a Cast (which could have set the
631 // control), we can't float above a Phi, because the skipped Cast
632 // may not be loop invariant.
633 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
634 adr = adr->in(1);
635 continue;
636 }
637 }
638
639 // Intentional fallthrough!
640
641 // No obvious dominating point. The mem op is pinned below the Phi
642 // by the Phi itself. If the Phi goes away (no true value is merged)
643 // then the mem op can float, but not indefinitely. It must be pinned
644 // behind the controls leading to the Phi.
645 case Op_CheckCastPP:
646 // These usually stick around to change address type, however a
647 // useless one can be elided and we still need to pick up a control edge
648 if (adr->in(0) == NULL) {
649 // This CheckCastPP node has NO control and is likely useless. But we
650 // need check further up the ancestor chain for a control input to keep
651 // the node in place. 4959717.
652 skipped_cast = adr;
653 adr = adr->in(1);
654 continue;
655 }
656 ccp->hash_delete(this);
657 set_req(MemNode::Control, adr->in(0));
658 ccp->hash_insert(this);
659 return this;
660
661 // List of "safe" opcodes; those that implicitly block the memory
662 // op below any null check.
663 case Op_CastX2P: // no null checks on native pointers
664 case Op_Parm: // 'this' pointer is not null
665 case Op_LoadP: // Loading from within a klass
666 case Op_LoadN: // Loading from within a klass
667 case Op_LoadKlass: // Loading from within a klass
668 case Op_ConP: // Loading from a klass
669 case Op_CreateEx: // Sucking up the guts of an exception oop
670 case Op_Con: // Reading from TLS
671 case Op_CMoveP: // CMoveP is pinned
672 break; // No progress
673
674 case Op_Proj: // Direct call to an allocation routine
675 case Op_SCMemProj: // Memory state from store conditional ops
676 #ifdef ASSERT
677 {
678 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
679 const Node* call = adr->in(0);
680 if (call->is_CallStaticJava()) {
681 const CallStaticJavaNode* call_java = call->as_CallStaticJava();
682 const TypeTuple *r = call_java->tf()->range();
683 assert(r->cnt() > TypeFunc::Parms, "must return value");
684 const Type* ret_type = r->field_at(TypeFunc::Parms);
685 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
686 // We further presume that this is one of
687 // new_instance_Java, new_array_Java, or
688 // the like, but do not assert for this.
689 } else if (call->is_Allocate()) {
690 // similar case to new_instance_Java, etc.
691 } else if (!call->is_CallLeaf()) {
692 // Projections from fetch_oop (OSR) are allowed as well.
693 ShouldNotReachHere();
694 }
695 }
696 #endif
697 break;
698 default:
699 ShouldNotReachHere();
700 }
701 break;
702 }
703 }
704
705 return NULL; // No progress
706 }
707
708
709 //=============================================================================
710 uint LoadNode::size_of() const { return sizeof(*this); }
711 uint LoadNode::cmp( const Node &n ) const
712 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
713 const Type *LoadNode::bottom_type() const { return _type; }
714 uint LoadNode::ideal_reg() const {
715 return Matcher::base2reg[_type->base()];
716 }
717
718 #ifndef PRODUCT
719 void LoadNode::dump_spec(outputStream *st) const {
720 MemNode::dump_spec(st);
721 if( !Verbose && !WizardMode ) {
722 // standard dump does this in Verbose and WizardMode
723 st->print(" #"); _type->dump_on(st);
724 }
725 }
726 #endif
727
728
729 //----------------------------LoadNode::make-----------------------------------
730 // Polymorphic factory method:
731 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
732 Compile* C = gvn.C;
733
734 // sanity check the alias category against the created node type
735 assert(!(adr_type->isa_oopptr() &&
736 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
737 "use LoadKlassNode instead");
738 assert(!(adr_type->isa_aryptr() &&
739 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
740 "use LoadRangeNode instead");
741 switch (bt) {
742 case T_BOOLEAN:
743 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() );
744 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() );
745 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() );
746 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() );
747 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() );
748 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt );
749 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt );
750 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() );
751 case T_OBJECT:
752 #ifdef _LP64
753 if (adr->bottom_type()->is_narrow()) {
754 const TypeNarrowOop* narrowtype;
755 if (rt->isa_narrowoop()) {
756 narrowtype = rt->is_narrowoop();
757 rt = narrowtype->make_oopptr();
758 } else {
759 narrowtype = rt->is_oopptr()->make_narrowoop();
760 }
761 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, narrowtype));
762
763 return new (C, 2) DecodeNNode(load, rt);
764 } else
765 #endif
766 {
767 assert(!adr->bottom_type()->is_narrow(), "should have got back a narrow oop");
768 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
769 }
770 }
771 ShouldNotReachHere();
772 return (LoadNode*)NULL;
773 }
774
775 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
776 bool require_atomic = true;
777 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
778 }
779
780
781
782
783 //------------------------------hash-------------------------------------------
784 uint LoadNode::hash() const {
785 // unroll addition of interesting fields
786 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
787 }
788
789 //---------------------------can_see_stored_value------------------------------
790 // This routine exists to make sure this set of tests is done the same
791 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
792 // will change the graph shape in a way which makes memory alive twice at the
793 // same time (uses the Oracle model of aliasing), then some
794 // LoadXNode::Identity will fold things back to the equivalence-class model
795 // of aliasing.
796 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
797 Node* ld_adr = in(MemNode::Address);
798
799 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
800 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
801 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
802 atp->field() != NULL && !atp->field()->is_volatile()) {
803 uint alias_idx = atp->index();
804 bool final = atp->field()->is_final();
805 Node* result = NULL;
806 Node* current = st;
807 // Skip through chains of MemBarNodes checking the MergeMems for
808 // new states for the slice of this load. Stop once any other
809 // kind of node is encountered. Loads from final memory can skip
810 // through any kind of MemBar but normal loads shouldn't skip
811 // through MemBarAcquire since the could allow them to move out of
812 // a synchronized region.
813 while (current->is_Proj()) {
814 int opc = current->in(0)->Opcode();
815 if ((final && opc == Op_MemBarAcquire) ||
816 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
817 Node* mem = current->in(0)->in(TypeFunc::Memory);
818 if (mem->is_MergeMem()) {
819 MergeMemNode* merge = mem->as_MergeMem();
820 Node* new_st = merge->memory_at(alias_idx);
821 if (new_st == merge->base_memory()) {
822 // Keep searching
823 current = merge->base_memory();
824 continue;
825 }
826 // Save the new memory state for the slice and fall through
827 // to exit.
828 result = new_st;
829 }
830 }
831 break;
832 }
833 if (result != NULL) {
834 st = result;
835 }
836 }
837
838
839 // Loop around twice in the case Load -> Initialize -> Store.
840 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
841 for (int trip = 0; trip <= 1; trip++) {
842
843 if (st->is_Store()) {
844 Node* st_adr = st->in(MemNode::Address);
845 if (!phase->eqv(st_adr, ld_adr)) {
846 // Try harder before giving up... Match raw and non-raw pointers.
847 intptr_t st_off = 0;
848 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
849 if (alloc == NULL) return NULL;
850 intptr_t ld_off = 0;
851 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
852 if (alloc != allo2) return NULL;
853 if (ld_off != st_off) return NULL;
854 // At this point we have proven something like this setup:
855 // A = Allocate(...)
856 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
857 // S = StoreQ(, AddP(, A.Parm , #Off), V)
858 // (Actually, we haven't yet proven the Q's are the same.)
859 // In other words, we are loading from a casted version of
860 // the same pointer-and-offset that we stored to.
861 // Thus, we are able to replace L by V.
862 }
863 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
864 if (store_Opcode() != st->Opcode())
865 return NULL;
866 return st->in(MemNode::ValueIn);
867 }
868
869 intptr_t offset = 0; // scratch
870
871 // A load from a freshly-created object always returns zero.
872 // (This can happen after LoadNode::Ideal resets the load's memory input
873 // to find_captured_store, which returned InitializeNode::zero_memory.)
874 if (st->is_Proj() && st->in(0)->is_Allocate() &&
875 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
876 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
877 // return a zero value for the load's basic type
878 // (This is one of the few places where a generic PhaseTransform
879 // can create new nodes. Think of it as lazily manifesting
880 // virtually pre-existing constants.)
881 return phase->zerocon(memory_type());
882 }
883
884 // A load from an initialization barrier can match a captured store.
885 if (st->is_Proj() && st->in(0)->is_Initialize()) {
886 InitializeNode* init = st->in(0)->as_Initialize();
887 AllocateNode* alloc = init->allocation();
888 if (alloc != NULL &&
889 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
890 // examine a captured store value
891 st = init->find_captured_store(offset, memory_size(), phase);
892 if (st != NULL)
893 continue; // take one more trip around
894 }
895 }
896
897 break;
898 }
899
900 return NULL;
901 }
902
903 //----------------------is_instance_field_load_with_local_phi------------------
904 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
905 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
906 in(MemNode::Address)->is_AddP() ) {
907 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
908 // Only instances.
909 if( t_oop != NULL && t_oop->is_instance_field() &&
910 t_oop->offset() != Type::OffsetBot &&
911 t_oop->offset() != Type::OffsetTop) {
912 return true;
913 }
914 }
915 return false;
916 }
917
918 //------------------------------Identity---------------------------------------
919 // Loads are identity if previous store is to same address
920 Node *LoadNode::Identity( PhaseTransform *phase ) {
921 // If the previous store-maker is the right kind of Store, and the store is
922 // to the same address, then we are equal to the value stored.
923 Node* mem = in(MemNode::Memory);
924 Node* value = can_see_stored_value(mem, phase);
925 if( value ) {
926 // byte, short & char stores truncate naturally.
927 // A load has to load the truncated value which requires
928 // some sort of masking operation and that requires an
929 // Ideal call instead of an Identity call.
930 if (memory_size() < BytesPerInt) {
931 // If the input to the store does not fit with the load's result type,
932 // it must be truncated via an Ideal call.
933 if (!phase->type(value)->higher_equal(phase->type(this)))
934 return this;
935 }
936 // (This works even when value is a Con, but LoadNode::Value
937 // usually runs first, producing the singleton type of the Con.)
938 return value;
939 }
940
941 // Search for an existing data phi which was generated before for the same
942 // instance's field to avoid infinite genertion of phis in a loop.
943 Node *region = mem->in(0);
944 if (is_instance_field_load_with_local_phi(region)) {
945 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
946 int this_index = phase->C->get_alias_index(addr_t);
947 int this_offset = addr_t->offset();
948 int this_id = addr_t->is_oopptr()->instance_id();
949 const Type* this_type = bottom_type();
950 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
951 Node* phi = region->fast_out(i);
952 if (phi->is_Phi() && phi != mem &&
953 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
954 return phi;
955 }
956 }
957 }
958
959 return this;
960 }
961
962
963 // Returns true if the AliasType refers to the field that holds the
964 // cached box array. Currently only handles the IntegerCache case.
965 static bool is_autobox_cache(Compile::AliasType* atp) {
966 if (atp != NULL && atp->field() != NULL) {
967 ciField* field = atp->field();
968 ciSymbol* klass = field->holder()->name();
969 if (field->name() == ciSymbol::cache_field_name() &&
970 field->holder()->uses_default_loader() &&
971 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
972 return true;
973 }
974 }
975 return false;
976 }
977
978 // Fetch the base value in the autobox array
979 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
980 if (atp != NULL && atp->field() != NULL) {
981 ciField* field = atp->field();
982 ciSymbol* klass = field->holder()->name();
983 if (field->name() == ciSymbol::cache_field_name() &&
984 field->holder()->uses_default_loader() &&
985 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
986 assert(field->is_constant(), "what?");
987 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
988 // Fetch the box object at the base of the array and get its value
989 ciInstance* box = array->obj_at(0)->as_instance();
990 ciInstanceKlass* ik = box->klass()->as_instance_klass();
991 if (ik->nof_nonstatic_fields() == 1) {
992 // This should be true nonstatic_field_at requires calling
993 // nof_nonstatic_fields so check it anyway
994 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
995 cache_offset = c.as_int();
996 }
997 return true;
998 }
999 }
1000 return false;
1001 }
1002
1003 // Returns true if the AliasType refers to the value field of an
1004 // autobox object. Currently only handles Integer.
1005 static bool is_autobox_object(Compile::AliasType* atp) {
1006 if (atp != NULL && atp->field() != NULL) {
1007 ciField* field = atp->field();
1008 ciSymbol* klass = field->holder()->name();
1009 if (field->name() == ciSymbol::value_name() &&
1010 field->holder()->uses_default_loader() &&
1011 klass == ciSymbol::java_lang_Integer()) {
1012 return true;
1013 }
1014 }
1015 return false;
1016 }
1017
1018
1019 // We're loading from an object which has autobox behaviour.
1020 // If this object is result of a valueOf call we'll have a phi
1021 // merging a newly allocated object and a load from the cache.
1022 // We want to replace this load with the original incoming
1023 // argument to the valueOf call.
1024 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1025 Node* base = in(Address)->in(AddPNode::Base);
1026 if (base->is_Phi() && base->req() == 3) {
1027 AllocateNode* allocation = NULL;
1028 int allocation_index = -1;
1029 int load_index = -1;
1030 for (uint i = 1; i < base->req(); i++) {
1031 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
1032 if (allocation != NULL) {
1033 allocation_index = i;
1034 load_index = 3 - allocation_index;
1035 break;
1036 }
1037 }
1038 LoadNode* load = NULL;
1039 if (allocation != NULL && base->in(load_index)->is_Load()) {
1040 load = base->in(load_index)->as_Load();
1041 }
1042 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
1043 // Push the loads from the phi that comes from valueOf up
1044 // through it to allow elimination of the loads and the recovery
1045 // of the original value.
1046 Node* mem_phi = in(Memory);
1047 Node* offset = in(Address)->in(AddPNode::Offset);
1048
1049 Node* in1 = clone();
1050 Node* in1_addr = in1->in(Address)->clone();
1051 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
1052 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
1053 in1_addr->set_req(AddPNode::Offset, offset);
1054 in1->set_req(0, base->in(allocation_index));
1055 in1->set_req(Address, in1_addr);
1056 in1->set_req(Memory, mem_phi->in(allocation_index));
1057
1058 Node* in2 = clone();
1059 Node* in2_addr = in2->in(Address)->clone();
1060 in2_addr->set_req(AddPNode::Base, base->in(load_index));
1061 in2_addr->set_req(AddPNode::Address, base->in(load_index));
1062 in2_addr->set_req(AddPNode::Offset, offset);
1063 in2->set_req(0, base->in(load_index));
1064 in2->set_req(Address, in2_addr);
1065 in2->set_req(Memory, mem_phi->in(load_index));
1066
1067 in1_addr = phase->transform(in1_addr);
1068 in1 = phase->transform(in1);
1069 in2_addr = phase->transform(in2_addr);
1070 in2 = phase->transform(in2);
1071
1072 PhiNode* result = PhiNode::make_blank(base->in(0), this);
1073 result->set_req(allocation_index, in1);
1074 result->set_req(load_index, in2);
1075 return result;
1076 }
1077 } else if (base->is_Load()) {
1078 // Eliminate the load of Integer.value for integers from the cache
1079 // array by deriving the value from the index into the array.
1080 // Capture the offset of the load and then reverse the computation.
1081 Node* load_base = base->in(Address)->in(AddPNode::Base);
1082 if (load_base != NULL) {
1083 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
1084 intptr_t cache_offset;
1085 int shift = -1;
1086 Node* cache = NULL;
1087 if (is_autobox_cache(atp)) {
1088 shift = exact_log2(type2aelembytes(T_OBJECT));
1089 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
1090 }
1091 if (cache != NULL && base->in(Address)->is_AddP()) {
1092 Node* elements[4];
1093 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
1094 int cache_low;
1095 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
1096 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
1097 // Add up all the offsets making of the address of the load
1098 Node* result = elements[0];
1099 for (int i = 1; i < count; i++) {
1100 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
1101 }
1102 // Remove the constant offset from the address and then
1103 // remove the scaling of the offset to recover the original index.
1104 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
1105 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1106 // Peel the shift off directly but wrap it in a dummy node
1107 // since Ideal can't return existing nodes
1108 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
1109 } else {
1110 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
1111 }
1112 #ifdef _LP64
1113 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
1114 #endif
1115 return result;
1116 }
1117 }
1118 }
1119 }
1120 return NULL;
1121 }
1122
1123
1124 //------------------------------Ideal------------------------------------------
1125 // If the load is from Field memory and the pointer is non-null, we can
1126 // zero out the control input.
1127 // If the offset is constant and the base is an object allocation,
1128 // try to hook me up to the exact initializing store.
1129 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1130 Node* p = MemNode::Ideal_common(phase, can_reshape);
1131 if (p) return (p == NodeSentinel) ? NULL : p;
1132
1133 Node* ctrl = in(MemNode::Control);
1134 Node* address = in(MemNode::Address);
1135
1136 // Skip up past a SafePoint control. Cannot do this for Stores because
1137 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1138 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1139 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1140 ctrl = ctrl->in(0);
1141 set_req(MemNode::Control,ctrl);
1142 }
1143
1144 // Check for useless control edge in some common special cases
1145 if (in(MemNode::Control) != NULL) {
1146 intptr_t ignore = 0;
1147 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1148 if (base != NULL
1149 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1150 && all_controls_dominate(base, phase->C->start())) {
1151 // A method-invariant, non-null address (constant or 'this' argument).
1152 set_req(MemNode::Control, NULL);
1153 }
1154 }
1155
1156 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
1157 Node* base = in(Address)->in(AddPNode::Base);
1158 if (base != NULL) {
1159 Compile::AliasType* atp = phase->C->alias_type(adr_type());
1160 if (is_autobox_object(atp)) {
1161 Node* result = eliminate_autobox(phase);
1162 if (result != NULL) return result;
1163 }
1164 }
1165 }
1166
1167 Node* mem = in(MemNode::Memory);
1168 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1169
1170 if (addr_t != NULL) {
1171 // try to optimize our memory input
1172 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
1173 if (opt_mem != mem) {
1174 set_req(MemNode::Memory, opt_mem);
1175 return this;
1176 }
1177 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1178 if (can_reshape && opt_mem->is_Phi() &&
1179 (t_oop != NULL) && t_oop->is_instance_field()) {
1180 assert(t_oop->offset() != Type::OffsetBot && t_oop->offset() != Type::OffsetTop, "");
1181 Node *region = opt_mem->in(0);
1182 uint cnt = opt_mem->req();
1183 for( uint i = 1; i < cnt; i++ ) {
1184 Node *in = opt_mem->in(i);
1185 if( in == NULL ) {
1186 region = NULL; // Wait stable graph
1187 break;
1188 }
1189 }
1190 if (region != NULL) {
1191 // Check for loop invariant.
1192 if (cnt == 3) {
1193 for( uint i = 1; i < cnt; i++ ) {
1194 Node *in = opt_mem->in(i);
1195 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
1196 if (m == opt_mem) {
1197 set_req(MemNode::Memory, opt_mem->in(cnt - i)); // Skip this phi.
1198 return this;
1199 }
1200 }
1201 }
1202 // Split through Phi (see original code in loopopts.cpp).
1203 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
1204
1205 // Do nothing here if Identity will find a value
1206 // (to avoid infinite chain of value phis generation).
1207 if ( !phase->eqv(this, this->Identity(phase)) )
1208 return NULL;
1209
1210 const Type* this_type = this->bottom_type();
1211 int this_index = phase->C->get_alias_index(addr_t);
1212 int this_offset = addr_t->offset();
1213 int this_iid = addr_t->is_oopptr()->instance_id();
1214 int wins = 0;
1215 PhaseIterGVN *igvn = phase->is_IterGVN();
1216 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1217 for( uint i = 1; i < region->req(); i++ ) {
1218 Node *x;
1219 Node* the_clone = NULL;
1220 if( region->in(i) == phase->C->top() ) {
1221 x = phase->C->top(); // Dead path? Use a dead data op
1222 } else {
1223 x = this->clone(); // Else clone up the data op
1224 the_clone = x; // Remember for possible deletion.
1225 // Alter data node to use pre-phi inputs
1226 if( this->in(0) == region ) {
1227 x->set_req( 0, region->in(i) );
1228 } else {
1229 x->set_req( 0, NULL );
1230 }
1231 for( uint j = 1; j < this->req(); j++ ) {
1232 Node *in = this->in(j);
1233 if( in->is_Phi() && in->in(0) == region )
1234 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
1235 }
1236 }
1237 // Check for a 'win' on some paths
1238 const Type *t = x->Value(igvn);
1239
1240 bool singleton = t->singleton();
1241
1242 // See comments in PhaseIdealLoop::split_thru_phi().
1243 if( singleton && t == Type::TOP ) {
1244 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1245 }
1246
1247 if( singleton ) {
1248 wins++;
1249 x = igvn->makecon(t);
1250 } else {
1251 // We now call Identity to try to simplify the cloned node.
1252 // Note that some Identity methods call phase->type(this).
1253 // Make sure that the type array is big enough for
1254 // our new node, even though we may throw the node away.
1255 // (This tweaking with igvn only works because x is a new node.)
1256 igvn->set_type(x, t);
1257 Node *y = x->Identity(igvn);
1258 if( y != x ) {
1259 wins++;
1260 x = y;
1261 } else {
1262 y = igvn->hash_find(x);
1263 if( y ) {
1264 wins++;
1265 x = y;
1266 } else {
1267 // Else x is a new node we are keeping
1268 // We do not need register_new_node_with_optimizer
1269 // because set_type has already been called.
1270 igvn->_worklist.push(x);
1271 }
1272 }
1273 }
1274 if (x != the_clone && the_clone != NULL)
1275 igvn->remove_dead_node(the_clone);
1276 phi->set_req(i, x);
1277 }
1278 if( wins > 0 ) {
1279 // Record Phi
1280 igvn->register_new_node_with_optimizer(phi);
1281 return phi;
1282 } else {
1283 igvn->remove_dead_node(phi);
1284 }
1285 }
1286 }
1287 }
1288
1289 // Check for prior store with a different base or offset; make Load
1290 // independent. Skip through any number of them. Bail out if the stores
1291 // are in an endless dead cycle and report no progress. This is a key
1292 // transform for Reflection. However, if after skipping through the Stores
1293 // we can't then fold up against a prior store do NOT do the transform as
1294 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1295 // array memory alive twice: once for the hoisted Load and again after the
1296 // bypassed Store. This situation only works if EVERYBODY who does
1297 // anti-dependence work knows how to bypass. I.e. we need all
1298 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1299 // the alias index stuff. So instead, peek through Stores and IFF we can
1300 // fold up, do so.
1301 Node* prev_mem = find_previous_store(phase);
1302 // Steps (a), (b): Walk past independent stores to find an exact match.
1303 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1304 // (c) See if we can fold up on the spot, but don't fold up here.
1305 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or
1306 // just return a prior value, which is done by Identity calls.
1307 if (can_see_stored_value(prev_mem, phase)) {
1308 // Make ready for step (d):
1309 set_req(MemNode::Memory, prev_mem);
1310 return this;
1311 }
1312 }
1313
1314 return NULL; // No further progress
1315 }
1316
1317 // Helper to recognize certain Klass fields which are invariant across
1318 // some group of array types (e.g., int[] or all T[] where T < Object).
1319 const Type*
1320 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1321 ciKlass* klass) const {
1322 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1323 // The field is Klass::_modifier_flags. Return its (constant) value.
1324 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1325 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1326 return TypeInt::make(klass->modifier_flags());
1327 }
1328 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1329 // The field is Klass::_access_flags. Return its (constant) value.
1330 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1331 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1332 return TypeInt::make(klass->access_flags());
1333 }
1334 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
1335 // The field is Klass::_layout_helper. Return its constant value if known.
1336 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1337 return TypeInt::make(klass->layout_helper());
1338 }
1339
1340 // No match.
1341 return NULL;
1342 }
1343
1344 //------------------------------Value-----------------------------------------
1345 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1346 // Either input is TOP ==> the result is TOP
1347 Node* mem = in(MemNode::Memory);
1348 const Type *t1 = phase->type(mem);
1349 if (t1 == Type::TOP) return Type::TOP;
1350 Node* adr = in(MemNode::Address);
1351 const TypePtr* tp = phase->type(adr)->isa_ptr();
1352 if (tp == NULL || tp->empty()) return Type::TOP;
1353 int off = tp->offset();
1354 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1355
1356 // Try to guess loaded type from pointer type
1357 if (tp->base() == Type::AryPtr) {
1358 const Type *t = tp->is_aryptr()->elem();
1359 // Don't do this for integer types. There is only potential profit if
1360 // the element type t is lower than _type; that is, for int types, if _type is
1361 // more restrictive than t. This only happens here if one is short and the other
1362 // char (both 16 bits), and in those cases we've made an intentional decision
1363 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1364 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1365 //
1366 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1367 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1368 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1369 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1370 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1371 // In fact, that could have been the original type of p1, and p1 could have
1372 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1373 // expression (LShiftL quux 3) independently optimized to the constant 8.
1374 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1375 && Opcode() != Op_LoadKlass) {
1376 // t might actually be lower than _type, if _type is a unique
1377 // concrete subclass of abstract class t.
1378 // Make sure the reference is not into the header, by comparing
1379 // the offset against the offset of the start of the array's data.
1380 // Different array types begin at slightly different offsets (12 vs. 16).
1381 // We choose T_BYTE as an example base type that is least restrictive
1382 // as to alignment, which will therefore produce the smallest
1383 // possible base offset.
1384 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1385 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
1386 const Type* jt = t->join(_type);
1387 // In any case, do not allow the join, per se, to empty out the type.
1388 if (jt->empty() && !t->empty()) {
1389 // This can happen if a interface-typed array narrows to a class type.
1390 jt = _type;
1391 }
1392
1393 if (EliminateAutoBox) {
1394 // The pointers in the autobox arrays are always non-null
1395 Node* base = in(Address)->in(AddPNode::Base);
1396 if (base != NULL) {
1397 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
1398 if (is_autobox_cache(atp)) {
1399 return jt->join(TypePtr::NOTNULL)->is_ptr();
1400 }
1401 }
1402 }
1403 return jt;
1404 }
1405 }
1406 } else if (tp->base() == Type::InstPtr) {
1407 assert( off != Type::OffsetBot ||
1408 // arrays can be cast to Objects
1409 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1410 // unsafe field access may not have a constant offset
1411 phase->C->has_unsafe_access(),
1412 "Field accesses must be precise" );
1413 // For oop loads, we expect the _type to be precise
1414 } else if (tp->base() == Type::KlassPtr) {
1415 assert( off != Type::OffsetBot ||
1416 // arrays can be cast to Objects
1417 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1418 // also allow array-loading from the primary supertype
1419 // array during subtype checks
1420 Opcode() == Op_LoadKlass,
1421 "Field accesses must be precise" );
1422 // For klass/static loads, we expect the _type to be precise
1423 }
1424
1425 const TypeKlassPtr *tkls = tp->isa_klassptr();
1426 if (tkls != NULL && !StressReflectiveCode) {
1427 ciKlass* klass = tkls->klass();
1428 if (klass->is_loaded() && tkls->klass_is_exact()) {
1429 // We are loading a field from a Klass metaobject whose identity
1430 // is known at compile time (the type is "exact" or "precise").
1431 // Check for fields we know are maintained as constants by the VM.
1432 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
1433 // The field is Klass::_super_check_offset. Return its (constant) value.
1434 // (Folds up type checking code.)
1435 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1436 return TypeInt::make(klass->super_check_offset());
1437 }
1438 // Compute index into primary_supers array
1439 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1440 // Check for overflowing; use unsigned compare to handle the negative case.
1441 if( depth < ciKlass::primary_super_limit() ) {
1442 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1443 // (Folds up type checking code.)
1444 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1445 ciKlass *ss = klass->super_of_depth(depth);
1446 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1447 }
1448 const Type* aift = load_array_final_field(tkls, klass);
1449 if (aift != NULL) return aift;
1450 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
1451 && klass->is_array_klass()) {
1452 // The field is arrayKlass::_component_mirror. Return its (constant) value.
1453 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1454 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1455 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1456 }
1457 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
1458 // The field is Klass::_java_mirror. Return its (constant) value.
1459 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1460 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1461 return TypeInstPtr::make(klass->java_mirror());
1462 }
1463 }
1464
1465 // We can still check if we are loading from the primary_supers array at a
1466 // shallow enough depth. Even though the klass is not exact, entries less
1467 // than or equal to its super depth are correct.
1468 if (klass->is_loaded() ) {
1469 ciType *inner = klass->klass();
1470 while( inner->is_obj_array_klass() )
1471 inner = inner->as_obj_array_klass()->base_element_type();
1472 if( inner->is_instance_klass() &&
1473 !inner->as_instance_klass()->flags().is_interface() ) {
1474 // Compute index into primary_supers array
1475 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1476 // Check for overflowing; use unsigned compare to handle the negative case.
1477 if( depth < ciKlass::primary_super_limit() &&
1478 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1479 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1480 // (Folds up type checking code.)
1481 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1482 ciKlass *ss = klass->super_of_depth(depth);
1483 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1484 }
1485 }
1486 }
1487
1488 // If the type is enough to determine that the thing is not an array,
1489 // we can give the layout_helper a positive interval type.
1490 // This will help short-circuit some reflective code.
1491 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
1492 && !klass->is_array_klass() // not directly typed as an array
1493 && !klass->is_interface() // specifically not Serializable & Cloneable
1494 && !klass->is_java_lang_Object() // not the supertype of all T[]
1495 ) {
1496 // Note: When interfaces are reliable, we can narrow the interface
1497 // test to (klass != Serializable && klass != Cloneable).
1498 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1499 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1500 // The key property of this type is that it folds up tests
1501 // for array-ness, since it proves that the layout_helper is positive.
1502 // Thus, a generic value like the basic object layout helper works fine.
1503 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1504 }
1505 }
1506
1507 // If we are loading from a freshly-allocated object, produce a zero,
1508 // if the load is provably beyond the header of the object.
1509 // (Also allow a variable load from a fresh array to produce zero.)
1510 if (ReduceFieldZeroing) {
1511 Node* value = can_see_stored_value(mem,phase);
1512 if (value != NULL && value->is_Con())
1513 return value->bottom_type();
1514 }
1515
1516 const TypeOopPtr *tinst = tp->isa_oopptr();
1517 if (tinst != NULL && tinst->is_instance_field()) {
1518 // If we have an instance type and our memory input is the
1519 // programs's initial memory state, there is no matching store,
1520 // so just return a zero of the appropriate type
1521 Node *mem = in(MemNode::Memory);
1522 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1523 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1524 return Type::get_zero_type(_type->basic_type());
1525 }
1526 }
1527 return _type;
1528 }
1529
1530 //------------------------------match_edge-------------------------------------
1531 // Do we Match on this edge index or not? Match only the address.
1532 uint LoadNode::match_edge(uint idx) const {
1533 return idx == MemNode::Address;
1534 }
1535
1536 //--------------------------LoadBNode::Ideal--------------------------------------
1537 //
1538 // If the previous store is to the same address as this load,
1539 // and the value stored was larger than a byte, replace this load
1540 // with the value stored truncated to a byte. If no truncation is
1541 // needed, the replacement is done in LoadNode::Identity().
1542 //
1543 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1544 Node* mem = in(MemNode::Memory);
1545 Node* value = can_see_stored_value(mem,phase);
1546 if( value && !phase->type(value)->higher_equal( _type ) ) {
1547 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
1548 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
1549 }
1550 // Identity call will handle the case where truncation is not needed.
1551 return LoadNode::Ideal(phase, can_reshape);
1552 }
1553
1554 //--------------------------LoadCNode::Ideal--------------------------------------
1555 //
1556 // If the previous store is to the same address as this load,
1557 // and the value stored was larger than a char, replace this load
1558 // with the value stored truncated to a char. If no truncation is
1559 // needed, the replacement is done in LoadNode::Identity().
1560 //
1561 Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1562 Node* mem = in(MemNode::Memory);
1563 Node* value = can_see_stored_value(mem,phase);
1564 if( value && !phase->type(value)->higher_equal( _type ) )
1565 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
1566 // Identity call will handle the case where truncation is not needed.
1567 return LoadNode::Ideal(phase, can_reshape);
1568 }
1569
1570 //--------------------------LoadSNode::Ideal--------------------------------------
1571 //
1572 // If the previous store is to the same address as this load,
1573 // and the value stored was larger than a short, replace this load
1574 // with the value stored truncated to a short. If no truncation is
1575 // needed, the replacement is done in LoadNode::Identity().
1576 //
1577 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1578 Node* mem = in(MemNode::Memory);
1579 Node* value = can_see_stored_value(mem,phase);
1580 if( value && !phase->type(value)->higher_equal( _type ) ) {
1581 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
1582 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
1583 }
1584 // Identity call will handle the case where truncation is not needed.
1585 return LoadNode::Ideal(phase, can_reshape);
1586 }
1587
1588 //=============================================================================
1589 //------------------------------Value------------------------------------------
1590 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1591 // Either input is TOP ==> the result is TOP
1592 const Type *t1 = phase->type( in(MemNode::Memory) );
1593 if (t1 == Type::TOP) return Type::TOP;
1594 Node *adr = in(MemNode::Address);
1595 const Type *t2 = phase->type( adr );
1596 if (t2 == Type::TOP) return Type::TOP;
1597 const TypePtr *tp = t2->is_ptr();
1598 if (TypePtr::above_centerline(tp->ptr()) ||
1599 tp->ptr() == TypePtr::Null) return Type::TOP;
1600
1601 // Return a more precise klass, if possible
1602 const TypeInstPtr *tinst = tp->isa_instptr();
1603 if (tinst != NULL) {
1604 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1605 int offset = tinst->offset();
1606 if (ik == phase->C->env()->Class_klass()
1607 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1608 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1609 // We are loading a special hidden field from a Class mirror object,
1610 // the field which points to the VM's Klass metaobject.
1611 ciType* t = tinst->java_mirror_type();
1612 // java_mirror_type returns non-null for compile-time Class constants.
1613 if (t != NULL) {
1614 // constant oop => constant klass
1615 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1616 return TypeKlassPtr::make(ciArrayKlass::make(t));
1617 }
1618 if (!t->is_klass()) {
1619 // a primitive Class (e.g., int.class) has NULL for a klass field
1620 return TypePtr::NULL_PTR;
1621 }
1622 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1623 return TypeKlassPtr::make(t->as_klass());
1624 }
1625 // non-constant mirror, so we can't tell what's going on
1626 }
1627 if( !ik->is_loaded() )
1628 return _type; // Bail out if not loaded
1629 if (offset == oopDesc::klass_offset_in_bytes()) {
1630 if (tinst->klass_is_exact()) {
1631 return TypeKlassPtr::make(ik);
1632 }
1633 // See if we can become precise: no subklasses and no interface
1634 // (Note: We need to support verified interfaces.)
1635 if (!ik->is_interface() && !ik->has_subklass()) {
1636 //assert(!UseExactTypes, "this code should be useless with exact types");
1637 // Add a dependence; if any subclass added we need to recompile
1638 if (!ik->is_final()) {
1639 // %%% should use stronger assert_unique_concrete_subtype instead
1640 phase->C->dependencies()->assert_leaf_type(ik);
1641 }
1642 // Return precise klass
1643 return TypeKlassPtr::make(ik);
1644 }
1645
1646 // Return root of possible klass
1647 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1648 }
1649 }
1650
1651 // Check for loading klass from an array
1652 const TypeAryPtr *tary = tp->isa_aryptr();
1653 if( tary != NULL ) {
1654 ciKlass *tary_klass = tary->klass();
1655 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1656 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1657 if (tary->klass_is_exact()) {
1658 return TypeKlassPtr::make(tary_klass);
1659 }
1660 ciArrayKlass *ak = tary->klass()->as_array_klass();
1661 // If the klass is an object array, we defer the question to the
1662 // array component klass.
1663 if( ak->is_obj_array_klass() ) {
1664 assert( ak->is_loaded(), "" );
1665 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1666 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1667 ciInstanceKlass* ik = base_k->as_instance_klass();
1668 // See if we can become precise: no subklasses and no interface
1669 if (!ik->is_interface() && !ik->has_subklass()) {
1670 //assert(!UseExactTypes, "this code should be useless with exact types");
1671 // Add a dependence; if any subclass added we need to recompile
1672 if (!ik->is_final()) {
1673 phase->C->dependencies()->assert_leaf_type(ik);
1674 }
1675 // Return precise array klass
1676 return TypeKlassPtr::make(ak);
1677 }
1678 }
1679 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1680 } else { // Found a type-array?
1681 //assert(!UseExactTypes, "this code should be useless with exact types");
1682 assert( ak->is_type_array_klass(), "" );
1683 return TypeKlassPtr::make(ak); // These are always precise
1684 }
1685 }
1686 }
1687
1688 // Check for loading klass from an array klass
1689 const TypeKlassPtr *tkls = tp->isa_klassptr();
1690 if (tkls != NULL && !StressReflectiveCode) {
1691 ciKlass* klass = tkls->klass();
1692 if( !klass->is_loaded() )
1693 return _type; // Bail out if not loaded
1694 if( klass->is_obj_array_klass() &&
1695 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
1696 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
1697 // // Always returning precise element type is incorrect,
1698 // // e.g., element type could be object and array may contain strings
1699 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
1700
1701 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
1702 // according to the element type's subclassing.
1703 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
1704 }
1705 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
1706 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
1707 ciKlass* sup = klass->as_instance_klass()->super();
1708 // The field is Klass::_super. Return its (constant) value.
1709 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
1710 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
1711 }
1712 }
1713
1714 // Bailout case
1715 return LoadNode::Value(phase);
1716 }
1717
1718 //------------------------------Identity---------------------------------------
1719 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
1720 // Also feed through the klass in Allocate(...klass...)._klass.
1721 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
1722 Node* x = LoadNode::Identity(phase);
1723 if (x != this) return x;
1724
1725 // Take apart the address into an oop and and offset.
1726 // Return 'this' if we cannot.
1727 Node* adr = in(MemNode::Address);
1728 intptr_t offset = 0;
1729 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1730 if (base == NULL) return this;
1731 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
1732 if (toop == NULL) return this;
1733
1734 // We can fetch the klass directly through an AllocateNode.
1735 // This works even if the klass is not constant (clone or newArray).
1736 if (offset == oopDesc::klass_offset_in_bytes()) {
1737 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
1738 if (allocated_klass != NULL) {
1739 return allocated_klass;
1740 }
1741 }
1742
1743 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
1744 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
1745 // See inline_native_Class_query for occurrences of these patterns.
1746 // Java Example: x.getClass().isAssignableFrom(y)
1747 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
1748 //
1749 // This improves reflective code, often making the Class
1750 // mirror go completely dead. (Current exception: Class
1751 // mirrors may appear in debug info, but we could clean them out by
1752 // introducing a new debug info operator for klassOop.java_mirror).
1753 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
1754 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1755 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1756 // We are loading a special hidden field from a Class mirror,
1757 // the field which points to its Klass or arrayKlass metaobject.
1758 if (base->is_Load()) {
1759 Node* adr2 = base->in(MemNode::Address);
1760 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1761 if (tkls != NULL && !tkls->empty()
1762 && (tkls->klass()->is_instance_klass() ||
1763 tkls->klass()->is_array_klass())
1764 && adr2->is_AddP()
1765 ) {
1766 int mirror_field = Klass::java_mirror_offset_in_bytes();
1767 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1768 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
1769 }
1770 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
1771 return adr2->in(AddPNode::Base);
1772 }
1773 }
1774 }
1775 }
1776
1777 return this;
1778 }
1779
1780 //------------------------------Value-----------------------------------------
1781 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
1782 // Either input is TOP ==> the result is TOP
1783 const Type *t1 = phase->type( in(MemNode::Memory) );
1784 if( t1 == Type::TOP ) return Type::TOP;
1785 Node *adr = in(MemNode::Address);
1786 const Type *t2 = phase->type( adr );
1787 if( t2 == Type::TOP ) return Type::TOP;
1788 const TypePtr *tp = t2->is_ptr();
1789 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
1790 const TypeAryPtr *tap = tp->isa_aryptr();
1791 if( !tap ) return _type;
1792 return tap->size();
1793 }
1794
1795 //------------------------------Identity---------------------------------------
1796 // Feed through the length in AllocateArray(...length...)._length.
1797 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
1798 Node* x = LoadINode::Identity(phase);
1799 if (x != this) return x;
1800
1801 // Take apart the address into an oop and and offset.
1802 // Return 'this' if we cannot.
1803 Node* adr = in(MemNode::Address);
1804 intptr_t offset = 0;
1805 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1806 if (base == NULL) return this;
1807 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
1808 if (tary == NULL) return this;
1809
1810 // We can fetch the length directly through an AllocateArrayNode.
1811 // This works even if the length is not constant (clone or newArray).
1812 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1813 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase);
1814 if (allocated_length != NULL) {
1815 return allocated_length;
1816 }
1817 }
1818
1819 return this;
1820
1821 }
1822 //=============================================================================
1823 //---------------------------StoreNode::make-----------------------------------
1824 // Polymorphic factory method:
1825 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
1826 Compile* C = gvn.C;
1827
1828 switch (bt) {
1829 case T_BOOLEAN:
1830 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
1831 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
1832 case T_CHAR:
1833 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
1834 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
1835 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
1836 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
1837 case T_ADDRESS:
1838 case T_OBJECT:
1839 #ifdef _LP64
1840 if (adr->bottom_type()->is_narrow() ||
1841 (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
1842 adr->bottom_type()->isa_rawptr())) {
1843 const TypePtr* type = val->bottom_type()->is_ptr();
1844 Node* cp;
1845 if (type->isa_oopptr()) {
1846 const TypeNarrowOop* etype = type->is_oopptr()->make_narrowoop();
1847 cp = gvn.transform(new (C, 2) EncodePNode(val, etype));
1848 } else if (type == TypePtr::NULL_PTR) {
1849 cp = gvn.transform(new (C, 1) ConNNode(TypeNarrowOop::NULL_PTR));
1850 } else {
1851 ShouldNotReachHere();
1852 }
1853 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, cp);
1854 } else
1855 #endif
1856 {
1857 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
1858 }
1859 }
1860 ShouldNotReachHere();
1861 return (StoreNode*)NULL;
1862 }
1863
1864 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
1865 bool require_atomic = true;
1866 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
1867 }
1868
1869
1870 //--------------------------bottom_type----------------------------------------
1871 const Type *StoreNode::bottom_type() const {
1872 return Type::MEMORY;
1873 }
1874
1875 //------------------------------hash-------------------------------------------
1876 uint StoreNode::hash() const {
1877 // unroll addition of interesting fields
1878 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
1879
1880 // Since they are not commoned, do not hash them:
1881 return NO_HASH;
1882 }
1883
1884 //------------------------------Ideal------------------------------------------
1885 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
1886 // When a store immediately follows a relevant allocation/initialization,
1887 // try to capture it into the initialization, or hoist it above.
1888 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1889 Node* p = MemNode::Ideal_common(phase, can_reshape);
1890 if (p) return (p == NodeSentinel) ? NULL : p;
1891
1892 Node* mem = in(MemNode::Memory);
1893 Node* address = in(MemNode::Address);
1894
1895 // Back-to-back stores to same address? Fold em up.
1896 // Generally unsafe if I have intervening uses...
1897 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
1898 // Looking at a dead closed cycle of memory?
1899 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
1900
1901 assert(Opcode() == mem->Opcode() ||
1902 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
1903 "no mismatched stores, except on raw memory");
1904
1905 if (mem->outcnt() == 1 && // check for intervening uses
1906 mem->as_Store()->memory_size() <= this->memory_size()) {
1907 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
1908 // For example, 'mem' might be the final state at a conditional return.
1909 // Or, 'mem' might be used by some node which is live at the same time
1910 // 'this' is live, which might be unschedulable. So, require exactly
1911 // ONE user, the 'this' store, until such time as we clone 'mem' for
1912 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
1913 if (can_reshape) { // (%%% is this an anachronism?)
1914 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
1915 phase->is_IterGVN());
1916 } else {
1917 // It's OK to do this in the parser, since DU info is always accurate,
1918 // and the parser always refers to nodes via SafePointNode maps.
1919 set_req(MemNode::Memory, mem->in(MemNode::Memory));
1920 }
1921 return this;
1922 }
1923 }
1924
1925 // Capture an unaliased, unconditional, simple store into an initializer.
1926 // Or, if it is independent of the allocation, hoist it above the allocation.
1927 if (ReduceFieldZeroing && /*can_reshape &&*/
1928 mem->is_Proj() && mem->in(0)->is_Initialize()) {
1929 InitializeNode* init = mem->in(0)->as_Initialize();
1930 intptr_t offset = init->can_capture_store(this, phase);
1931 if (offset > 0) {
1932 Node* moved = init->capture_store(this, offset, phase);
1933 // If the InitializeNode captured me, it made a raw copy of me,
1934 // and I need to disappear.
1935 if (moved != NULL) {
1936 // %%% hack to ensure that Ideal returns a new node:
1937 mem = MergeMemNode::make(phase->C, mem);
1938 return mem; // fold me away
1939 }
1940 }
1941 }
1942
1943 return NULL; // No further progress
1944 }
1945
1946 //------------------------------Value-----------------------------------------
1947 const Type *StoreNode::Value( PhaseTransform *phase ) const {
1948 // Either input is TOP ==> the result is TOP
1949 const Type *t1 = phase->type( in(MemNode::Memory) );
1950 if( t1 == Type::TOP ) return Type::TOP;
1951 const Type *t2 = phase->type( in(MemNode::Address) );
1952 if( t2 == Type::TOP ) return Type::TOP;
1953 const Type *t3 = phase->type( in(MemNode::ValueIn) );
1954 if( t3 == Type::TOP ) return Type::TOP;
1955 return Type::MEMORY;
1956 }
1957
1958 //------------------------------Identity---------------------------------------
1959 // Remove redundant stores:
1960 // Store(m, p, Load(m, p)) changes to m.
1961 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
1962 Node *StoreNode::Identity( PhaseTransform *phase ) {
1963 Node* mem = in(MemNode::Memory);
1964 Node* adr = in(MemNode::Address);
1965 Node* val = in(MemNode::ValueIn);
1966
1967 // Load then Store? Then the Store is useless
1968 if (val->is_Load() &&
1969 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
1970 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
1971 val->as_Load()->store_Opcode() == Opcode()) {
1972 return mem;
1973 }
1974
1975 // Two stores in a row of the same value?
1976 if (mem->is_Store() &&
1977 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
1978 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
1979 mem->Opcode() == Opcode()) {
1980 return mem;
1981 }
1982
1983 // Store of zero anywhere into a freshly-allocated object?
1984 // Then the store is useless.
1985 // (It must already have been captured by the InitializeNode.)
1986 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
1987 // a newly allocated object is already all-zeroes everywhere
1988 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
1989 return mem;
1990 }
1991
1992 // the store may also apply to zero-bits in an earlier object
1993 Node* prev_mem = find_previous_store(phase);
1994 // Steps (a), (b): Walk past independent stores to find an exact match.
1995 if (prev_mem != NULL) {
1996 Node* prev_val = can_see_stored_value(prev_mem, phase);
1997 if (prev_val != NULL && phase->eqv(prev_val, val)) {
1998 // prev_val and val might differ by a cast; it would be good
1999 // to keep the more informative of the two.
2000 return mem;
2001 }
2002 }
2003 }
2004
2005 return this;
2006 }
2007
2008 //------------------------------match_edge-------------------------------------
2009 // Do we Match on this edge index or not? Match only memory & value
2010 uint StoreNode::match_edge(uint idx) const {
2011 return idx == MemNode::Address || idx == MemNode::ValueIn;
2012 }
2013
2014 //------------------------------cmp--------------------------------------------
2015 // Do not common stores up together. They generally have to be split
2016 // back up anyways, so do not bother.
2017 uint StoreNode::cmp( const Node &n ) const {
2018 return (&n == this); // Always fail except on self
2019 }
2020
2021 //------------------------------Ideal_masked_input-----------------------------
2022 // Check for a useless mask before a partial-word store
2023 // (StoreB ... (AndI valIn conIa) )
2024 // If (conIa & mask == mask) this simplifies to
2025 // (StoreB ... (valIn) )
2026 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2027 Node *val = in(MemNode::ValueIn);
2028 if( val->Opcode() == Op_AndI ) {
2029 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2030 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2031 set_req(MemNode::ValueIn, val->in(1));
2032 return this;
2033 }
2034 }
2035 return NULL;
2036 }
2037
2038
2039 //------------------------------Ideal_sign_extended_input----------------------
2040 // Check for useless sign-extension before a partial-word store
2041 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2042 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2043 // (StoreB ... (valIn) )
2044 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2045 Node *val = in(MemNode::ValueIn);
2046 if( val->Opcode() == Op_RShiftI ) {
2047 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2048 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2049 Node *shl = val->in(1);
2050 if( shl->Opcode() == Op_LShiftI ) {
2051 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2052 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2053 set_req(MemNode::ValueIn, shl->in(1));
2054 return this;
2055 }
2056 }
2057 }
2058 }
2059 return NULL;
2060 }
2061
2062 //------------------------------value_never_loaded-----------------------------------
2063 // Determine whether there are any possible loads of the value stored.
2064 // For simplicity, we actually check if there are any loads from the
2065 // address stored to, not just for loads of the value stored by this node.
2066 //
2067 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2068 Node *adr = in(Address);
2069 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2070 if (adr_oop == NULL)
2071 return false;
2072 if (!adr_oop->is_instance_field())
2073 return false; // if not a distinct instance, there may be aliases of the address
2074 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2075 Node *use = adr->fast_out(i);
2076 int opc = use->Opcode();
2077 if (use->is_Load() || use->is_LoadStore()) {
2078 return false;
2079 }
2080 }
2081 return true;
2082 }
2083
2084 //=============================================================================
2085 //------------------------------Ideal------------------------------------------
2086 // If the store is from an AND mask that leaves the low bits untouched, then
2087 // we can skip the AND operation. If the store is from a sign-extension
2088 // (a left shift, then right shift) we can skip both.
2089 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2090 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2091 if( progress != NULL ) return progress;
2092
2093 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2094 if( progress != NULL ) return progress;
2095
2096 // Finally check the default case
2097 return StoreNode::Ideal(phase, can_reshape);
2098 }
2099
2100 //=============================================================================
2101 //------------------------------Ideal------------------------------------------
2102 // If the store is from an AND mask that leaves the low bits untouched, then
2103 // we can skip the AND operation
2104 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2105 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2106 if( progress != NULL ) return progress;
2107
2108 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2109 if( progress != NULL ) return progress;
2110
2111 // Finally check the default case
2112 return StoreNode::Ideal(phase, can_reshape);
2113 }
2114
2115 //=============================================================================
2116 //------------------------------Identity---------------------------------------
2117 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2118 // No need to card mark when storing a null ptr
2119 Node* my_store = in(MemNode::OopStore);
2120 if (my_store->is_Store()) {
2121 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2122 if( t1 == TypePtr::NULL_PTR ) {
2123 return in(MemNode::Memory);
2124 }
2125 }
2126 return this;
2127 }
2128
2129 //------------------------------Value-----------------------------------------
2130 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2131 // Either input is TOP ==> the result is TOP
2132 const Type *t = phase->type( in(MemNode::Memory) );
2133 if( t == Type::TOP ) return Type::TOP;
2134 t = phase->type( in(MemNode::Address) );
2135 if( t == Type::TOP ) return Type::TOP;
2136 t = phase->type( in(MemNode::ValueIn) );
2137 if( t == Type::TOP ) return Type::TOP;
2138 // If extra input is TOP ==> the result is TOP
2139 t = phase->type( in(MemNode::OopStore) );
2140 if( t == Type::TOP ) return Type::TOP;
2141
2142 return StoreNode::Value( phase );
2143 }
2144
2145
2146 //=============================================================================
2147 //----------------------------------SCMemProjNode------------------------------
2148 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2149 {
2150 return bottom_type();
2151 }
2152
2153 //=============================================================================
2154 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
2155 init_req(MemNode::Control, c );
2156 init_req(MemNode::Memory , mem);
2157 init_req(MemNode::Address, adr);
2158 init_req(MemNode::ValueIn, val);
2159 init_req( ExpectedIn, ex );
2160 init_class_id(Class_LoadStore);
2161
2162 }
2163
2164 //=============================================================================
2165 //-------------------------------adr_type--------------------------------------
2166 // Do we Match on this edge index or not? Do not match memory
2167 const TypePtr* ClearArrayNode::adr_type() const {
2168 Node *adr = in(3);
2169 return MemNode::calculate_adr_type(adr->bottom_type());
2170 }
2171
2172 //------------------------------match_edge-------------------------------------
2173 // Do we Match on this edge index or not? Do not match memory
2174 uint ClearArrayNode::match_edge(uint idx) const {
2175 return idx > 1;
2176 }
2177
2178 //------------------------------Identity---------------------------------------
2179 // Clearing a zero length array does nothing
2180 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2181 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2182 }
2183
2184 //------------------------------Idealize---------------------------------------
2185 // Clearing a short array is faster with stores
2186 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2187 const int unit = BytesPerLong;
2188 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2189 if (!t) return NULL;
2190 if (!t->is_con()) return NULL;
2191 intptr_t raw_count = t->get_con();
2192 intptr_t size = raw_count;
2193 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2194 // Clearing nothing uses the Identity call.
2195 // Negative clears are possible on dead ClearArrays
2196 // (see jck test stmt114.stmt11402.val).
2197 if (size <= 0 || size % unit != 0) return NULL;
2198 intptr_t count = size / unit;
2199 // Length too long; use fast hardware clear
2200 if (size > Matcher::init_array_short_size) return NULL;
2201 Node *mem = in(1);
2202 if( phase->type(mem)==Type::TOP ) return NULL;
2203 Node *adr = in(3);
2204 const Type* at = phase->type(adr);
2205 if( at==Type::TOP ) return NULL;
2206 const TypePtr* atp = at->isa_ptr();
2207 // adjust atp to be the correct array element address type
2208 if (atp == NULL) atp = TypePtr::BOTTOM;
2209 else atp = atp->add_offset(Type::OffsetBot);
2210 // Get base for derived pointer purposes
2211 if( adr->Opcode() != Op_AddP ) Unimplemented();
2212 Node *base = adr->in(1);
2213
2214 Node *zero = phase->makecon(TypeLong::ZERO);
2215 Node *off = phase->MakeConX(BytesPerLong);
2216 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2217 count--;
2218 while( count-- ) {
2219 mem = phase->transform(mem);
2220 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
2221 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2222 }
2223 return mem;
2224 }
2225
2226 //----------------------------clear_memory-------------------------------------
2227 // Generate code to initialize object storage to zero.
2228 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2229 intptr_t start_offset,
2230 Node* end_offset,
2231 PhaseGVN* phase) {
2232 Compile* C = phase->C;
2233 intptr_t offset = start_offset;
2234
2235 int unit = BytesPerLong;
2236 if ((offset % unit) != 0) {
2237 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
2238 adr = phase->transform(adr);
2239 const TypePtr* atp = TypeRawPtr::BOTTOM;
2240 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2241 mem = phase->transform(mem);
2242 offset += BytesPerInt;
2243 }
2244 assert((offset % unit) == 0, "");
2245
2246 // Initialize the remaining stuff, if any, with a ClearArray.
2247 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2248 }
2249
2250 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2251 Node* start_offset,
2252 Node* end_offset,
2253 PhaseGVN* phase) {
2254 if (start_offset == end_offset) {
2255 // nothing to do
2256 return mem;
2257 }
2258
2259 Compile* C = phase->C;
2260 int unit = BytesPerLong;
2261 Node* zbase = start_offset;
2262 Node* zend = end_offset;
2263
2264 // Scale to the unit required by the CPU:
2265 if (!Matcher::init_array_count_is_in_bytes) {
2266 Node* shift = phase->intcon(exact_log2(unit));
2267 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
2268 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
2269 }
2270
2271 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
2272 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
2273
2274 // Bulk clear double-words
2275 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
2276 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
2277 return phase->transform(mem);
2278 }
2279
2280 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2281 intptr_t start_offset,
2282 intptr_t end_offset,
2283 PhaseGVN* phase) {
2284 if (start_offset == end_offset) {
2285 // nothing to do
2286 return mem;
2287 }
2288
2289 Compile* C = phase->C;
2290 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2291 intptr_t done_offset = end_offset;
2292 if ((done_offset % BytesPerLong) != 0) {
2293 done_offset -= BytesPerInt;
2294 }
2295 if (done_offset > start_offset) {
2296 mem = clear_memory(ctl, mem, dest,
2297 start_offset, phase->MakeConX(done_offset), phase);
2298 }
2299 if (done_offset < end_offset) { // emit the final 32-bit store
2300 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
2301 adr = phase->transform(adr);
2302 const TypePtr* atp = TypeRawPtr::BOTTOM;
2303 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2304 mem = phase->transform(mem);
2305 done_offset += BytesPerInt;
2306 }
2307 assert(done_offset == end_offset, "");
2308 return mem;
2309 }
2310
2311 //=============================================================================
2312 // Do we match on this edge? No memory edges
2313 uint StrCompNode::match_edge(uint idx) const {
2314 return idx == 5 || idx == 6;
2315 }
2316
2317 //------------------------------Ideal------------------------------------------
2318 // Return a node which is more "ideal" than the current node. Strip out
2319 // control copies
2320 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
2321 return remove_dead_region(phase, can_reshape) ? this : NULL;
2322 }
2323
2324
2325 //=============================================================================
2326 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2327 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2328 _adr_type(C->get_adr_type(alias_idx))
2329 {
2330 init_class_id(Class_MemBar);
2331 Node* top = C->top();
2332 init_req(TypeFunc::I_O,top);
2333 init_req(TypeFunc::FramePtr,top);
2334 init_req(TypeFunc::ReturnAdr,top);
2335 if (precedent != NULL)
2336 init_req(TypeFunc::Parms, precedent);
2337 }
2338
2339 //------------------------------cmp--------------------------------------------
2340 uint MemBarNode::hash() const { return NO_HASH; }
2341 uint MemBarNode::cmp( const Node &n ) const {
2342 return (&n == this); // Always fail except on self
2343 }
2344
2345 //------------------------------make-------------------------------------------
2346 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2347 int len = Precedent + (pn == NULL? 0: 1);
2348 switch (opcode) {
2349 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
2350 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
2351 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
2352 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
2353 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
2354 default: ShouldNotReachHere(); return NULL;
2355 }
2356 }
2357
2358 //------------------------------Ideal------------------------------------------
2359 // Return a node which is more "ideal" than the current node. Strip out
2360 // control copies
2361 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2362 if (remove_dead_region(phase, can_reshape)) return this;
2363 return NULL;
2364 }
2365
2366 //------------------------------Value------------------------------------------
2367 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2368 if( !in(0) ) return Type::TOP;
2369 if( phase->type(in(0)) == Type::TOP )
2370 return Type::TOP;
2371 return TypeTuple::MEMBAR;
2372 }
2373
2374 //------------------------------match------------------------------------------
2375 // Construct projections for memory.
2376 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2377 switch (proj->_con) {
2378 case TypeFunc::Control:
2379 case TypeFunc::Memory:
2380 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2381 }
2382 ShouldNotReachHere();
2383 return NULL;
2384 }
2385
2386 //===========================InitializeNode====================================
2387 // SUMMARY:
2388 // This node acts as a memory barrier on raw memory, after some raw stores.
2389 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2390 // The Initialize can 'capture' suitably constrained stores as raw inits.
2391 // It can coalesce related raw stores into larger units (called 'tiles').
2392 // It can avoid zeroing new storage for memory units which have raw inits.
2393 // At macro-expansion, it is marked 'complete', and does not optimize further.
2394 //
2395 // EXAMPLE:
2396 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2397 // ctl = incoming control; mem* = incoming memory
2398 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
2399 // First allocate uninitialized memory and fill in the header:
2400 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
2401 // ctl := alloc.Control; mem* := alloc.Memory*
2402 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
2403 // Then initialize to zero the non-header parts of the raw memory block:
2404 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2405 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2406 // After the initialize node executes, the object is ready for service:
2407 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2408 // Suppose its body is immediately initialized as {1,2}:
2409 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2410 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2411 // mem.SLICE(#short[*]) := store2
2412 //
2413 // DETAILS:
2414 // An InitializeNode collects and isolates object initialization after
2415 // an AllocateNode and before the next possible safepoint. As a
2416 // memory barrier (MemBarNode), it keeps critical stores from drifting
2417 // down past any safepoint or any publication of the allocation.
2418 // Before this barrier, a newly-allocated object may have uninitialized bits.
2419 // After this barrier, it may be treated as a real oop, and GC is allowed.
2420 //
2421 // The semantics of the InitializeNode include an implicit zeroing of
2422 // the new object from object header to the end of the object.
2423 // (The object header and end are determined by the AllocateNode.)
2424 //
2425 // Certain stores may be added as direct inputs to the InitializeNode.
2426 // These stores must update raw memory, and they must be to addresses
2427 // derived from the raw address produced by AllocateNode, and with
2428 // a constant offset. They must be ordered by increasing offset.
2429 // The first one is at in(RawStores), the last at in(req()-1).
2430 // Unlike most memory operations, they are not linked in a chain,
2431 // but are displayed in parallel as users of the rawmem output of
2432 // the allocation.
2433 //
2434 // (See comments in InitializeNode::capture_store, which continue
2435 // the example given above.)
2436 //
2437 // When the associated Allocate is macro-expanded, the InitializeNode
2438 // may be rewritten to optimize collected stores. A ClearArrayNode
2439 // may also be created at that point to represent any required zeroing.
2440 // The InitializeNode is then marked 'complete', prohibiting further
2441 // capturing of nearby memory operations.
2442 //
2443 // During macro-expansion, all captured initializations which store
2444 // constant values of 32 bits or smaller are coalesced (if advantagous)
2445 // into larger 'tiles' 32 or 64 bits. This allows an object to be
2446 // initialized in fewer memory operations. Memory words which are
2447 // covered by neither tiles nor non-constant stores are pre-zeroed
2448 // by explicit stores of zero. (The code shape happens to do all
2449 // zeroing first, then all other stores, with both sequences occurring
2450 // in order of ascending offsets.)
2451 //
2452 // Alternatively, code may be inserted between an AllocateNode and its
2453 // InitializeNode, to perform arbitrary initialization of the new object.
2454 // E.g., the object copying intrinsics insert complex data transfers here.
2455 // The initialization must then be marked as 'complete' disable the
2456 // built-in zeroing semantics and the collection of initializing stores.
2457 //
2458 // While an InitializeNode is incomplete, reads from the memory state
2459 // produced by it are optimizable if they match the control edge and
2460 // new oop address associated with the allocation/initialization.
2461 // They return a stored value (if the offset matches) or else zero.
2462 // A write to the memory state, if it matches control and address,
2463 // and if it is to a constant offset, may be 'captured' by the
2464 // InitializeNode. It is cloned as a raw memory operation and rewired
2465 // inside the initialization, to the raw oop produced by the allocation.
2466 // Operations on addresses which are provably distinct (e.g., to
2467 // other AllocateNodes) are allowed to bypass the initialization.
2468 //
2469 // The effect of all this is to consolidate object initialization
2470 // (both arrays and non-arrays, both piecewise and bulk) into a
2471 // single location, where it can be optimized as a unit.
2472 //
2473 // Only stores with an offset less than TrackedInitializationLimit words
2474 // will be considered for capture by an InitializeNode. This puts a
2475 // reasonable limit on the complexity of optimized initializations.
2476
2477 //---------------------------InitializeNode------------------------------------
2478 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
2479 : _is_complete(false),
2480 MemBarNode(C, adr_type, rawoop)
2481 {
2482 init_class_id(Class_Initialize);
2483
2484 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
2485 assert(in(RawAddress) == rawoop, "proper init");
2486 // Note: allocation() can be NULL, for secondary initialization barriers
2487 }
2488
2489 // Since this node is not matched, it will be processed by the
2490 // register allocator. Declare that there are no constraints
2491 // on the allocation of the RawAddress edge.
2492 const RegMask &InitializeNode::in_RegMask(uint idx) const {
2493 // This edge should be set to top, by the set_complete. But be conservative.
2494 if (idx == InitializeNode::RawAddress)
2495 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
2496 return RegMask::Empty;
2497 }
2498
2499 Node* InitializeNode::memory(uint alias_idx) {
2500 Node* mem = in(Memory);
2501 if (mem->is_MergeMem()) {
2502 return mem->as_MergeMem()->memory_at(alias_idx);
2503 } else {
2504 // incoming raw memory is not split
2505 return mem;
2506 }
2507 }
2508
2509 bool InitializeNode::is_non_zero() {
2510 if (is_complete()) return false;
2511 remove_extra_zeroes();
2512 return (req() > RawStores);
2513 }
2514
2515 void InitializeNode::set_complete(PhaseGVN* phase) {
2516 assert(!is_complete(), "caller responsibility");
2517 _is_complete = true;
2518
2519 // After this node is complete, it contains a bunch of
2520 // raw-memory initializations. There is no need for
2521 // it to have anything to do with non-raw memory effects.
2522 // Therefore, tell all non-raw users to re-optimize themselves,
2523 // after skipping the memory effects of this initialization.
2524 PhaseIterGVN* igvn = phase->is_IterGVN();
2525 if (igvn) igvn->add_users_to_worklist(this);
2526 }
2527
2528 // convenience function
2529 // return false if the init contains any stores already
2530 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
2531 InitializeNode* init = initialization();
2532 if (init == NULL || init->is_complete()) return false;
2533 init->remove_extra_zeroes();
2534 // for now, if this allocation has already collected any inits, bail:
2535 if (init->is_non_zero()) return false;
2536 init->set_complete(phase);
2537 return true;
2538 }
2539
2540 void InitializeNode::remove_extra_zeroes() {
2541 if (req() == RawStores) return;
2542 Node* zmem = zero_memory();
2543 uint fill = RawStores;
2544 for (uint i = fill; i < req(); i++) {
2545 Node* n = in(i);
2546 if (n->is_top() || n == zmem) continue; // skip
2547 if (fill < i) set_req(fill, n); // compact
2548 ++fill;
2549 }
2550 // delete any empty spaces created:
2551 while (fill < req()) {
2552 del_req(fill);
2553 }
2554 }
2555
2556 // Helper for remembering which stores go with which offsets.
2557 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
2558 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
2559 intptr_t offset = -1;
2560 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
2561 phase, offset);
2562 if (base == NULL) return -1; // something is dead,
2563 if (offset < 0) return -1; // dead, dead
2564 return offset;
2565 }
2566
2567 // Helper for proving that an initialization expression is
2568 // "simple enough" to be folded into an object initialization.
2569 // Attempts to prove that a store's initial value 'n' can be captured
2570 // within the initialization without creating a vicious cycle, such as:
2571 // { Foo p = new Foo(); p.next = p; }
2572 // True for constants and parameters and small combinations thereof.
2573 bool InitializeNode::detect_init_independence(Node* n,
2574 bool st_is_pinned,
2575 int& count) {
2576 if (n == NULL) return true; // (can this really happen?)
2577 if (n->is_Proj()) n = n->in(0);
2578 if (n == this) return false; // found a cycle
2579 if (n->is_Con()) return true;
2580 if (n->is_Start()) return true; // params, etc., are OK
2581 if (n->is_Root()) return true; // even better
2582
2583 Node* ctl = n->in(0);
2584 if (ctl != NULL && !ctl->is_top()) {
2585 if (ctl->is_Proj()) ctl = ctl->in(0);
2586 if (ctl == this) return false;
2587
2588 // If we already know that the enclosing memory op is pinned right after
2589 // the init, then any control flow that the store has picked up
2590 // must have preceded the init, or else be equal to the init.
2591 // Even after loop optimizations (which might change control edges)
2592 // a store is never pinned *before* the availability of its inputs.
2593 if (!MemNode::all_controls_dominate(n, this))
2594 return false; // failed to prove a good control
2595
2596 }
2597
2598 // Check data edges for possible dependencies on 'this'.
2599 if ((count += 1) > 20) return false; // complexity limit
2600 for (uint i = 1; i < n->req(); i++) {
2601 Node* m = n->in(i);
2602 if (m == NULL || m == n || m->is_top()) continue;
2603 uint first_i = n->find_edge(m);
2604 if (i != first_i) continue; // process duplicate edge just once
2605 if (!detect_init_independence(m, st_is_pinned, count)) {
2606 return false;
2607 }
2608 }
2609
2610 return true;
2611 }
2612
2613 // Here are all the checks a Store must pass before it can be moved into
2614 // an initialization. Returns zero if a check fails.
2615 // On success, returns the (constant) offset to which the store applies,
2616 // within the initialized memory.
2617 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
2618 const int FAIL = 0;
2619 if (st->req() != MemNode::ValueIn + 1)
2620 return FAIL; // an inscrutable StoreNode (card mark?)
2621 Node* ctl = st->in(MemNode::Control);
2622 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
2623 return FAIL; // must be unconditional after the initialization
2624 Node* mem = st->in(MemNode::Memory);
2625 if (!(mem->is_Proj() && mem->in(0) == this))
2626 return FAIL; // must not be preceded by other stores
2627 Node* adr = st->in(MemNode::Address);
2628 intptr_t offset;
2629 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
2630 if (alloc == NULL)
2631 return FAIL; // inscrutable address
2632 if (alloc != allocation())
2633 return FAIL; // wrong allocation! (store needs to float up)
2634 Node* val = st->in(MemNode::ValueIn);
2635 int complexity_count = 0;
2636 if (!detect_init_independence(val, true, complexity_count))
2637 return FAIL; // stored value must be 'simple enough'
2638
2639 return offset; // success
2640 }
2641
2642 // Find the captured store in(i) which corresponds to the range
2643 // [start..start+size) in the initialized object.
2644 // If there is one, return its index i. If there isn't, return the
2645 // negative of the index where it should be inserted.
2646 // Return 0 if the queried range overlaps an initialization boundary
2647 // or if dead code is encountered.
2648 // If size_in_bytes is zero, do not bother with overlap checks.
2649 int InitializeNode::captured_store_insertion_point(intptr_t start,
2650 int size_in_bytes,
2651 PhaseTransform* phase) {
2652 const int FAIL = 0, MAX_STORE = BytesPerLong;
2653
2654 if (is_complete())
2655 return FAIL; // arraycopy got here first; punt
2656
2657 assert(allocation() != NULL, "must be present");
2658
2659 // no negatives, no header fields:
2660 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
2661
2662 // after a certain size, we bail out on tracking all the stores:
2663 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2664 if (start >= ti_limit) return FAIL;
2665
2666 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
2667 if (i >= limit) return -(int)i; // not found; here is where to put it
2668
2669 Node* st = in(i);
2670 intptr_t st_off = get_store_offset(st, phase);
2671 if (st_off < 0) {
2672 if (st != zero_memory()) {
2673 return FAIL; // bail out if there is dead garbage
2674 }
2675 } else if (st_off > start) {
2676 // ...we are done, since stores are ordered
2677 if (st_off < start + size_in_bytes) {
2678 return FAIL; // the next store overlaps
2679 }
2680 return -(int)i; // not found; here is where to put it
2681 } else if (st_off < start) {
2682 if (size_in_bytes != 0 &&
2683 start < st_off + MAX_STORE &&
2684 start < st_off + st->as_Store()->memory_size()) {
2685 return FAIL; // the previous store overlaps
2686 }
2687 } else {
2688 if (size_in_bytes != 0 &&
2689 st->as_Store()->memory_size() != size_in_bytes) {
2690 return FAIL; // mismatched store size
2691 }
2692 return i;
2693 }
2694
2695 ++i;
2696 }
2697 }
2698
2699 // Look for a captured store which initializes at the offset 'start'
2700 // with the given size. If there is no such store, and no other
2701 // initialization interferes, then return zero_memory (the memory
2702 // projection of the AllocateNode).
2703 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
2704 PhaseTransform* phase) {
2705 assert(stores_are_sane(phase), "");
2706 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2707 if (i == 0) {
2708 return NULL; // something is dead
2709 } else if (i < 0) {
2710 return zero_memory(); // just primordial zero bits here
2711 } else {
2712 Node* st = in(i); // here is the store at this position
2713 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
2714 return st;
2715 }
2716 }
2717
2718 // Create, as a raw pointer, an address within my new object at 'offset'.
2719 Node* InitializeNode::make_raw_address(intptr_t offset,
2720 PhaseTransform* phase) {
2721 Node* addr = in(RawAddress);
2722 if (offset != 0) {
2723 Compile* C = phase->C;
2724 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
2725 phase->MakeConX(offset)) );
2726 }
2727 return addr;
2728 }
2729
2730 // Clone the given store, converting it into a raw store
2731 // initializing a field or element of my new object.
2732 // Caller is responsible for retiring the original store,
2733 // with subsume_node or the like.
2734 //
2735 // From the example above InitializeNode::InitializeNode,
2736 // here are the old stores to be captured:
2737 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2738 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2739 //
2740 // Here is the changed code; note the extra edges on init:
2741 // alloc = (Allocate ...)
2742 // rawoop = alloc.RawAddress
2743 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
2744 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
2745 // init = (Initialize alloc.Control alloc.Memory rawoop
2746 // rawstore1 rawstore2)
2747 //
2748 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
2749 PhaseTransform* phase) {
2750 assert(stores_are_sane(phase), "");
2751
2752 if (start < 0) return NULL;
2753 assert(can_capture_store(st, phase) == start, "sanity");
2754
2755 Compile* C = phase->C;
2756 int size_in_bytes = st->memory_size();
2757 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2758 if (i == 0) return NULL; // bail out
2759 Node* prev_mem = NULL; // raw memory for the captured store
2760 if (i > 0) {
2761 prev_mem = in(i); // there is a pre-existing store under this one
2762 set_req(i, C->top()); // temporarily disconnect it
2763 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
2764 } else {
2765 i = -i; // no pre-existing store
2766 prev_mem = zero_memory(); // a slice of the newly allocated object
2767 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
2768 set_req(--i, C->top()); // reuse this edge; it has been folded away
2769 else
2770 ins_req(i, C->top()); // build a new edge
2771 }
2772 Node* new_st = st->clone();
2773 new_st->set_req(MemNode::Control, in(Control));
2774 new_st->set_req(MemNode::Memory, prev_mem);
2775 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
2776 new_st = phase->transform(new_st);
2777
2778 // At this point, new_st might have swallowed a pre-existing store
2779 // at the same offset, or perhaps new_st might have disappeared,
2780 // if it redundantly stored the same value (or zero to fresh memory).
2781
2782 // In any case, wire it in:
2783 set_req(i, new_st);
2784
2785 // The caller may now kill the old guy.
2786 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
2787 assert(check_st == new_st || check_st == NULL, "must be findable");
2788 assert(!is_complete(), "");
2789 return new_st;
2790 }
2791
2792 static bool store_constant(jlong* tiles, int num_tiles,
2793 intptr_t st_off, int st_size,
2794 jlong con) {
2795 if ((st_off & (st_size-1)) != 0)
2796 return false; // strange store offset (assume size==2**N)
2797 address addr = (address)tiles + st_off;
2798 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
2799 switch (st_size) {
2800 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
2801 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
2802 case sizeof(jint): *(jint*) addr = (jint) con; break;
2803 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
2804 default: return false; // strange store size (detect size!=2**N here)
2805 }
2806 return true; // return success to caller
2807 }
2808
2809 // Coalesce subword constants into int constants and possibly
2810 // into long constants. The goal, if the CPU permits,
2811 // is to initialize the object with a small number of 64-bit tiles.
2812 // Also, convert floating-point constants to bit patterns.
2813 // Non-constants are not relevant to this pass.
2814 //
2815 // In terms of the running example on InitializeNode::InitializeNode
2816 // and InitializeNode::capture_store, here is the transformation
2817 // of rawstore1 and rawstore2 into rawstore12:
2818 // alloc = (Allocate ...)
2819 // rawoop = alloc.RawAddress
2820 // tile12 = 0x00010002
2821 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
2822 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
2823 //
2824 void
2825 InitializeNode::coalesce_subword_stores(intptr_t header_size,
2826 Node* size_in_bytes,
2827 PhaseGVN* phase) {
2828 Compile* C = phase->C;
2829
2830 assert(stores_are_sane(phase), "");
2831 // Note: After this pass, they are not completely sane,
2832 // since there may be some overlaps.
2833
2834 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
2835
2836 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2837 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
2838 size_limit = MIN2(size_limit, ti_limit);
2839 size_limit = align_size_up(size_limit, BytesPerLong);
2840 int num_tiles = size_limit / BytesPerLong;
2841
2842 // allocate space for the tile map:
2843 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
2844 jlong tiles_buf[small_len];
2845 Node* nodes_buf[small_len];
2846 jlong inits_buf[small_len];
2847 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
2848 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2849 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
2850 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
2851 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
2852 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2853 // tiles: exact bitwise model of all primitive constants
2854 // nodes: last constant-storing node subsumed into the tiles model
2855 // inits: which bytes (in each tile) are touched by any initializations
2856
2857 //// Pass A: Fill in the tile model with any relevant stores.
2858
2859 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
2860 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
2861 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
2862 Node* zmem = zero_memory(); // initially zero memory state
2863 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
2864 Node* st = in(i);
2865 intptr_t st_off = get_store_offset(st, phase);
2866
2867 // Figure out the store's offset and constant value:
2868 if (st_off < header_size) continue; //skip (ignore header)
2869 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
2870 int st_size = st->as_Store()->memory_size();
2871 if (st_off + st_size > size_limit) break;
2872
2873 // Record which bytes are touched, whether by constant or not.
2874 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
2875 continue; // skip (strange store size)
2876
2877 const Type* val = phase->type(st->in(MemNode::ValueIn));
2878 if (!val->singleton()) continue; //skip (non-con store)
2879 BasicType type = val->basic_type();
2880
2881 jlong con = 0;
2882 switch (type) {
2883 case T_INT: con = val->is_int()->get_con(); break;
2884 case T_LONG: con = val->is_long()->get_con(); break;
2885 case T_FLOAT: con = jint_cast(val->getf()); break;
2886 case T_DOUBLE: con = jlong_cast(val->getd()); break;
2887 default: continue; //skip (odd store type)
2888 }
2889
2890 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
2891 st->Opcode() == Op_StoreL) {
2892 continue; // This StoreL is already optimal.
2893 }
2894
2895 // Store down the constant.
2896 store_constant(tiles, num_tiles, st_off, st_size, con);
2897
2898 intptr_t j = st_off >> LogBytesPerLong;
2899
2900 if (type == T_INT && st_size == BytesPerInt
2901 && (st_off & BytesPerInt) == BytesPerInt) {
2902 jlong lcon = tiles[j];
2903 if (!Matcher::isSimpleConstant64(lcon) &&
2904 st->Opcode() == Op_StoreI) {
2905 // This StoreI is already optimal by itself.
2906 jint* intcon = (jint*) &tiles[j];
2907 intcon[1] = 0; // undo the store_constant()
2908
2909 // If the previous store is also optimal by itself, back up and
2910 // undo the action of the previous loop iteration... if we can.
2911 // But if we can't, just let the previous half take care of itself.
2912 st = nodes[j];
2913 st_off -= BytesPerInt;
2914 con = intcon[0];
2915 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
2916 assert(st_off >= header_size, "still ignoring header");
2917 assert(get_store_offset(st, phase) == st_off, "must be");
2918 assert(in(i-1) == zmem, "must be");
2919 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
2920 assert(con == tcon->is_int()->get_con(), "must be");
2921 // Undo the effects of the previous loop trip, which swallowed st:
2922 intcon[0] = 0; // undo store_constant()
2923 set_req(i-1, st); // undo set_req(i, zmem)
2924 nodes[j] = NULL; // undo nodes[j] = st
2925 --old_subword; // undo ++old_subword
2926 }
2927 continue; // This StoreI is already optimal.
2928 }
2929 }
2930
2931 // This store is not needed.
2932 set_req(i, zmem);
2933 nodes[j] = st; // record for the moment
2934 if (st_size < BytesPerLong) // something has changed
2935 ++old_subword; // includes int/float, but who's counting...
2936 else ++old_long;
2937 }
2938
2939 if ((old_subword + old_long) == 0)
2940 return; // nothing more to do
2941
2942 //// Pass B: Convert any non-zero tiles into optimal constant stores.
2943 // Be sure to insert them before overlapping non-constant stores.
2944 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
2945 for (int j = 0; j < num_tiles; j++) {
2946 jlong con = tiles[j];
2947 jlong init = inits[j];
2948 if (con == 0) continue;
2949 jint con0, con1; // split the constant, address-wise
2950 jint init0, init1; // split the init map, address-wise
2951 { union { jlong con; jint intcon[2]; } u;
2952 u.con = con;
2953 con0 = u.intcon[0];
2954 con1 = u.intcon[1];
2955 u.con = init;
2956 init0 = u.intcon[0];
2957 init1 = u.intcon[1];
2958 }
2959
2960 Node* old = nodes[j];
2961 assert(old != NULL, "need the prior store");
2962 intptr_t offset = (j * BytesPerLong);
2963
2964 bool split = !Matcher::isSimpleConstant64(con);
2965
2966 if (offset < header_size) {
2967 assert(offset + BytesPerInt >= header_size, "second int counts");
2968 assert(*(jint*)&tiles[j] == 0, "junk in header");
2969 split = true; // only the second word counts
2970 // Example: int a[] = { 42 ... }
2971 } else if (con0 == 0 && init0 == -1) {
2972 split = true; // first word is covered by full inits
2973 // Example: int a[] = { ... foo(), 42 ... }
2974 } else if (con1 == 0 && init1 == -1) {
2975 split = true; // second word is covered by full inits
2976 // Example: int a[] = { ... 42, foo() ... }
2977 }
2978
2979 // Here's a case where init0 is neither 0 nor -1:
2980 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
2981 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
2982 // In this case the tile is not split; it is (jlong)42.
2983 // The big tile is stored down, and then the foo() value is inserted.
2984 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
2985
2986 Node* ctl = old->in(MemNode::Control);
2987 Node* adr = make_raw_address(offset, phase);
2988 const TypePtr* atp = TypeRawPtr::BOTTOM;
2989
2990 // One or two coalesced stores to plop down.
2991 Node* st[2];
2992 intptr_t off[2];
2993 int nst = 0;
2994 if (!split) {
2995 ++new_long;
2996 off[nst] = offset;
2997 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
2998 phase->longcon(con), T_LONG);
2999 } else {
3000 // Omit either if it is a zero.
3001 if (con0 != 0) {
3002 ++new_int;
3003 off[nst] = offset;
3004 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3005 phase->intcon(con0), T_INT);
3006 }
3007 if (con1 != 0) {
3008 ++new_int;
3009 offset += BytesPerInt;
3010 adr = make_raw_address(offset, phase);
3011 off[nst] = offset;
3012 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3013 phase->intcon(con1), T_INT);
3014 }
3015 }
3016
3017 // Insert second store first, then the first before the second.
3018 // Insert each one just before any overlapping non-constant stores.
3019 while (nst > 0) {
3020 Node* st1 = st[--nst];
3021 C->copy_node_notes_to(st1, old);
3022 st1 = phase->transform(st1);
3023 offset = off[nst];
3024 assert(offset >= header_size, "do not smash header");
3025 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3026 guarantee(ins_idx != 0, "must re-insert constant store");
3027 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3028 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3029 set_req(--ins_idx, st1);
3030 else
3031 ins_req(ins_idx, st1);
3032 }
3033 }
3034
3035 if (PrintCompilation && WizardMode)
3036 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3037 old_subword, old_long, new_int, new_long);
3038 if (C->log() != NULL)
3039 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3040 old_subword, old_long, new_int, new_long);
3041
3042 // Clean up any remaining occurrences of zmem:
3043 remove_extra_zeroes();
3044 }
3045
3046 // Explore forward from in(start) to find the first fully initialized
3047 // word, and return its offset. Skip groups of subword stores which
3048 // together initialize full words. If in(start) is itself part of a
3049 // fully initialized word, return the offset of in(start). If there
3050 // are no following full-word stores, or if something is fishy, return
3051 // a negative value.
3052 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3053 int int_map = 0;
3054 intptr_t int_map_off = 0;
3055 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3056
3057 for (uint i = start, limit = req(); i < limit; i++) {
3058 Node* st = in(i);
3059
3060 intptr_t st_off = get_store_offset(st, phase);
3061 if (st_off < 0) break; // return conservative answer
3062
3063 int st_size = st->as_Store()->memory_size();
3064 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3065 return st_off; // we found a complete word init
3066 }
3067
3068 // update the map:
3069
3070 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3071 if (this_int_off != int_map_off) {
3072 // reset the map:
3073 int_map = 0;
3074 int_map_off = this_int_off;
3075 }
3076
3077 int subword_off = st_off - this_int_off;
3078 int_map |= right_n_bits(st_size) << subword_off;
3079 if ((int_map & FULL_MAP) == FULL_MAP) {
3080 return this_int_off; // we found a complete word init
3081 }
3082
3083 // Did this store hit or cross the word boundary?
3084 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3085 if (next_int_off == this_int_off + BytesPerInt) {
3086 // We passed the current int, without fully initializing it.
3087 int_map_off = next_int_off;
3088 int_map >>= BytesPerInt;
3089 } else if (next_int_off > this_int_off + BytesPerInt) {
3090 // We passed the current and next int.
3091 return this_int_off + BytesPerInt;
3092 }
3093 }
3094
3095 return -1;
3096 }
3097
3098
3099 // Called when the associated AllocateNode is expanded into CFG.
3100 // At this point, we may perform additional optimizations.
3101 // Linearize the stores by ascending offset, to make memory
3102 // activity as coherent as possible.
3103 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3104 intptr_t header_size,
3105 Node* size_in_bytes,
3106 PhaseGVN* phase) {
3107 assert(!is_complete(), "not already complete");
3108 assert(stores_are_sane(phase), "");
3109 assert(allocation() != NULL, "must be present");
3110
3111 remove_extra_zeroes();
3112
3113 if (ReduceFieldZeroing || ReduceBulkZeroing)
3114 // reduce instruction count for common initialization patterns
3115 coalesce_subword_stores(header_size, size_in_bytes, phase);
3116
3117 Node* zmem = zero_memory(); // initially zero memory state
3118 Node* inits = zmem; // accumulating a linearized chain of inits
3119 #ifdef ASSERT
3120 intptr_t first_offset = allocation()->minimum_header_size();
3121 intptr_t last_init_off = first_offset; // previous init offset
3122 intptr_t last_init_end = first_offset; // previous init offset+size
3123 intptr_t last_tile_end = first_offset; // previous tile offset+size
3124 #endif
3125 intptr_t zeroes_done = header_size;
3126
3127 bool do_zeroing = true; // we might give up if inits are very sparse
3128 int big_init_gaps = 0; // how many large gaps have we seen?
3129
3130 if (ZeroTLAB) do_zeroing = false;
3131 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3132
3133 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3134 Node* st = in(i);
3135 intptr_t st_off = get_store_offset(st, phase);
3136 if (st_off < 0)
3137 break; // unknown junk in the inits
3138 if (st->in(MemNode::Memory) != zmem)
3139 break; // complicated store chains somehow in list
3140
3141 int st_size = st->as_Store()->memory_size();
3142 intptr_t next_init_off = st_off + st_size;
3143
3144 if (do_zeroing && zeroes_done < next_init_off) {
3145 // See if this store needs a zero before it or under it.
3146 intptr_t zeroes_needed = st_off;
3147
3148 if (st_size < BytesPerInt) {
3149 // Look for subword stores which only partially initialize words.
3150 // If we find some, we must lay down some word-level zeroes first,
3151 // underneath the subword stores.
3152 //
3153 // Examples:
3154 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3155 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3156 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3157 //
3158 // Note: coalesce_subword_stores may have already done this,
3159 // if it was prompted by constant non-zero subword initializers.
3160 // But this case can still arise with non-constant stores.
3161
3162 intptr_t next_full_store = find_next_fullword_store(i, phase);
3163
3164 // In the examples above:
3165 // in(i) p q r s x y z
3166 // st_off 12 13 14 15 12 13 14
3167 // st_size 1 1 1 1 1 1 1
3168 // next_full_s. 12 16 16 16 16 16 16
3169 // z's_done 12 16 16 16 12 16 12
3170 // z's_needed 12 16 16 16 16 16 16
3171 // zsize 0 0 0 0 4 0 4
3172 if (next_full_store < 0) {
3173 // Conservative tack: Zero to end of current word.
3174 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3175 } else {
3176 // Zero to beginning of next fully initialized word.
3177 // Or, don't zero at all, if we are already in that word.
3178 assert(next_full_store >= zeroes_needed, "must go forward");
3179 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3180 zeroes_needed = next_full_store;
3181 }
3182 }
3183
3184 if (zeroes_needed > zeroes_done) {
3185 intptr_t zsize = zeroes_needed - zeroes_done;
3186 // Do some incremental zeroing on rawmem, in parallel with inits.
3187 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3188 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3189 zeroes_done, zeroes_needed,
3190 phase);
3191 zeroes_done = zeroes_needed;
3192 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3193 do_zeroing = false; // leave the hole, next time
3194 }
3195 }
3196
3197 // Collect the store and move on:
3198 st->set_req(MemNode::Memory, inits);
3199 inits = st; // put it on the linearized chain
3200 set_req(i, zmem); // unhook from previous position
3201
3202 if (zeroes_done == st_off)
3203 zeroes_done = next_init_off;
3204
3205 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3206
3207 #ifdef ASSERT
3208 // Various order invariants. Weaker than stores_are_sane because
3209 // a large constant tile can be filled in by smaller non-constant stores.
3210 assert(st_off >= last_init_off, "inits do not reverse");
3211 last_init_off = st_off;
3212 const Type* val = NULL;
3213 if (st_size >= BytesPerInt &&
3214 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3215 (int)val->basic_type() < (int)T_OBJECT) {
3216 assert(st_off >= last_tile_end, "tiles do not overlap");
3217 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3218 last_tile_end = MAX2(last_tile_end, next_init_off);
3219 } else {
3220 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3221 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3222 assert(st_off >= last_init_end, "inits do not overlap");
3223 last_init_end = next_init_off; // it's a non-tile
3224 }
3225 #endif //ASSERT
3226 }
3227
3228 remove_extra_zeroes(); // clear out all the zmems left over
3229 add_req(inits);
3230
3231 if (!ZeroTLAB) {
3232 // If anything remains to be zeroed, zero it all now.
3233 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3234 // if it is the last unused 4 bytes of an instance, forget about it
3235 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3236 if (zeroes_done + BytesPerLong >= size_limit) {
3237 assert(allocation() != NULL, "");
3238 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3239 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3240 if (zeroes_done == k->layout_helper())
3241 zeroes_done = size_limit;
3242 }
3243 if (zeroes_done < size_limit) {
3244 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3245 zeroes_done, size_in_bytes, phase);
3246 }
3247 }
3248
3249 set_complete(phase);
3250 return rawmem;
3251 }
3252
3253
3254 #ifdef ASSERT
3255 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3256 if (is_complete())
3257 return true; // stores could be anything at this point
3258 assert(allocation() != NULL, "must be present");
3259 intptr_t last_off = allocation()->minimum_header_size();
3260 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3261 Node* st = in(i);
3262 intptr_t st_off = get_store_offset(st, phase);
3263 if (st_off < 0) continue; // ignore dead garbage
3264 if (last_off > st_off) {
3265 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
3266 this->dump(2);
3267 assert(false, "ascending store offsets");
3268 return false;
3269 }
3270 last_off = st_off + st->as_Store()->memory_size();
3271 }
3272 return true;
3273 }
3274 #endif //ASSERT
3275
3276
3277
3278
3279 //============================MergeMemNode=====================================
3280 //
3281 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3282 // contributing store or call operations. Each contributor provides the memory
3283 // state for a particular "alias type" (see Compile::alias_type). For example,
3284 // if a MergeMem has an input X for alias category #6, then any memory reference
3285 // to alias category #6 may use X as its memory state input, as an exact equivalent
3286 // to using the MergeMem as a whole.
3287 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3288 //
3289 // (Here, the <N> notation gives the index of the relevant adr_type.)
3290 //
3291 // In one special case (and more cases in the future), alias categories overlap.
3292 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
3293 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
3294 // it is exactly equivalent to that state W:
3295 // MergeMem(<Bot>: W) <==> W
3296 //
3297 // Usually, the merge has more than one input. In that case, where inputs
3298 // overlap (i.e., one is Bot), the narrower alias type determines the memory
3299 // state for that type, and the wider alias type (Bot) fills in everywhere else:
3300 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
3301 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
3302 //
3303 // A merge can take a "wide" memory state as one of its narrow inputs.
3304 // This simply means that the merge observes out only the relevant parts of
3305 // the wide input. That is, wide memory states arriving at narrow merge inputs
3306 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
3307 //
3308 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
3309 // and that memory slices "leak through":
3310 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
3311 //
3312 // But, in such a cascade, repeated memory slices can "block the leak":
3313 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
3314 //
3315 // In the last example, Y is not part of the combined memory state of the
3316 // outermost MergeMem. The system must, of course, prevent unschedulable
3317 // memory states from arising, so you can be sure that the state Y is somehow
3318 // a precursor to state Y'.
3319 //
3320 //
3321 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
3322 // of each MergeMemNode array are exactly the numerical alias indexes, including
3323 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
3324 // Compile::alias_type (and kin) produce and manage these indexes.
3325 //
3326 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
3327 // (Note that this provides quick access to the top node inside MergeMem methods,
3328 // without the need to reach out via TLS to Compile::current.)
3329 //
3330 // As a consequence of what was just described, a MergeMem that represents a full
3331 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
3332 // containing all alias categories.
3333 //
3334 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
3335 //
3336 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
3337 // a memory state for the alias type <N>, or else the top node, meaning that
3338 // there is no particular input for that alias type. Note that the length of
3339 // a MergeMem is variable, and may be extended at any time to accommodate new
3340 // memory states at larger alias indexes. When merges grow, they are of course
3341 // filled with "top" in the unused in() positions.
3342 //
3343 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
3344 // (Top was chosen because it works smoothly with passes like GCM.)
3345 //
3346 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
3347 // the type of random VM bits like TLS references.) Since it is always the
3348 // first non-Bot memory slice, some low-level loops use it to initialize an
3349 // index variable: for (i = AliasIdxRaw; i < req(); i++).
3350 //
3351 //
3352 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
3353 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
3354 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
3355 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
3356 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
3357 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
3358 //
3359 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
3360 // really that different from the other memory inputs. An abbreviation called
3361 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
3362 //
3363 //
3364 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
3365 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
3366 // that "emerges though" the base memory will be marked as excluding the alias types
3367 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
3368 //
3369 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
3370 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
3371 //
3372 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
3373 // (It is currently unimplemented.) As you can see, the resulting merge is
3374 // actually a disjoint union of memory states, rather than an overlay.
3375 //
3376
3377 //------------------------------MergeMemNode-----------------------------------
3378 Node* MergeMemNode::make_empty_memory() {
3379 Node* empty_memory = (Node*) Compile::current()->top();
3380 assert(empty_memory->is_top(), "correct sentinel identity");
3381 return empty_memory;
3382 }
3383
3384 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3385 init_class_id(Class_MergeMem);
3386 // all inputs are nullified in Node::Node(int)
3387 // set_input(0, NULL); // no control input
3388
3389 // Initialize the edges uniformly to top, for starters.
3390 Node* empty_mem = make_empty_memory();
3391 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
3392 init_req(i,empty_mem);
3393 }
3394 assert(empty_memory() == empty_mem, "");
3395
3396 if( new_base != NULL && new_base->is_MergeMem() ) {
3397 MergeMemNode* mdef = new_base->as_MergeMem();
3398 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
3399 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
3400 mms.set_memory(mms.memory2());
3401 }
3402 assert(base_memory() == mdef->base_memory(), "");
3403 } else {
3404 set_base_memory(new_base);
3405 }
3406 }
3407
3408 // Make a new, untransformed MergeMem with the same base as 'mem'.
3409 // If mem is itself a MergeMem, populate the result with the same edges.
3410 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
3411 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
3412 }
3413
3414 //------------------------------cmp--------------------------------------------
3415 uint MergeMemNode::hash() const { return NO_HASH; }
3416 uint MergeMemNode::cmp( const Node &n ) const {
3417 return (&n == this); // Always fail except on self
3418 }
3419
3420 //------------------------------Identity---------------------------------------
3421 Node* MergeMemNode::Identity(PhaseTransform *phase) {
3422 // Identity if this merge point does not record any interesting memory
3423 // disambiguations.
3424 Node* base_mem = base_memory();
3425 Node* empty_mem = empty_memory();
3426 if (base_mem != empty_mem) { // Memory path is not dead?
3427 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3428 Node* mem = in(i);
3429 if (mem != empty_mem && mem != base_mem) {
3430 return this; // Many memory splits; no change
3431 }
3432 }
3433 }
3434 return base_mem; // No memory splits; ID on the one true input
3435 }
3436
3437 //------------------------------Ideal------------------------------------------
3438 // This method is invoked recursively on chains of MergeMem nodes
3439 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3440 // Remove chain'd MergeMems
3441 //
3442 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
3443 // relative to the "in(Bot)". Since we are patching both at the same time,
3444 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
3445 // but rewrite each "in(i)" relative to the new "in(Bot)".
3446 Node *progress = NULL;
3447
3448
3449 Node* old_base = base_memory();
3450 Node* empty_mem = empty_memory();
3451 if (old_base == empty_mem)
3452 return NULL; // Dead memory path.
3453
3454 MergeMemNode* old_mbase;
3455 if (old_base != NULL && old_base->is_MergeMem())
3456 old_mbase = old_base->as_MergeMem();
3457 else
3458 old_mbase = NULL;
3459 Node* new_base = old_base;
3460
3461 // simplify stacked MergeMems in base memory
3462 if (old_mbase) new_base = old_mbase->base_memory();
3463
3464 // the base memory might contribute new slices beyond my req()
3465 if (old_mbase) grow_to_match(old_mbase);
3466
3467 // Look carefully at the base node if it is a phi.
3468 PhiNode* phi_base;
3469 if (new_base != NULL && new_base->is_Phi())
3470 phi_base = new_base->as_Phi();
3471 else
3472 phi_base = NULL;
3473
3474 Node* phi_reg = NULL;
3475 uint phi_len = (uint)-1;
3476 if (phi_base != NULL && !phi_base->is_copy()) {
3477 // do not examine phi if degraded to a copy
3478 phi_reg = phi_base->region();
3479 phi_len = phi_base->req();
3480 // see if the phi is unfinished
3481 for (uint i = 1; i < phi_len; i++) {
3482 if (phi_base->in(i) == NULL) {
3483 // incomplete phi; do not look at it yet!
3484 phi_reg = NULL;
3485 phi_len = (uint)-1;
3486 break;
3487 }
3488 }
3489 }
3490
3491 // Note: We do not call verify_sparse on entry, because inputs
3492 // can normalize to the base_memory via subsume_node or similar
3493 // mechanisms. This method repairs that damage.
3494
3495 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
3496
3497 // Look at each slice.
3498 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3499 Node* old_in = in(i);
3500 // calculate the old memory value
3501 Node* old_mem = old_in;
3502 if (old_mem == empty_mem) old_mem = old_base;
3503 assert(old_mem == memory_at(i), "");
3504
3505 // maybe update (reslice) the old memory value
3506
3507 // simplify stacked MergeMems
3508 Node* new_mem = old_mem;
3509 MergeMemNode* old_mmem;
3510 if (old_mem != NULL && old_mem->is_MergeMem())
3511 old_mmem = old_mem->as_MergeMem();
3512 else
3513 old_mmem = NULL;
3514 if (old_mmem == this) {
3515 // This can happen if loops break up and safepoints disappear.
3516 // A merge of BotPtr (default) with a RawPtr memory derived from a
3517 // safepoint can be rewritten to a merge of the same BotPtr with
3518 // the BotPtr phi coming into the loop. If that phi disappears
3519 // also, we can end up with a self-loop of the mergemem.
3520 // In general, if loops degenerate and memory effects disappear,
3521 // a mergemem can be left looking at itself. This simply means
3522 // that the mergemem's default should be used, since there is
3523 // no longer any apparent effect on this slice.
3524 // Note: If a memory slice is a MergeMem cycle, it is unreachable
3525 // from start. Update the input to TOP.
3526 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
3527 }
3528 else if (old_mmem != NULL) {
3529 new_mem = old_mmem->memory_at(i);
3530 }
3531 // else preceeding memory was not a MergeMem
3532
3533 // replace equivalent phis (unfortunately, they do not GVN together)
3534 if (new_mem != NULL && new_mem != new_base &&
3535 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
3536 if (new_mem->is_Phi()) {
3537 PhiNode* phi_mem = new_mem->as_Phi();
3538 for (uint i = 1; i < phi_len; i++) {
3539 if (phi_base->in(i) != phi_mem->in(i)) {
3540 phi_mem = NULL;
3541 break;
3542 }
3543 }
3544 if (phi_mem != NULL) {
3545 // equivalent phi nodes; revert to the def
3546 new_mem = new_base;
3547 }
3548 }
3549 }
3550
3551 // maybe store down a new value
3552 Node* new_in = new_mem;
3553 if (new_in == new_base) new_in = empty_mem;
3554
3555 if (new_in != old_in) {
3556 // Warning: Do not combine this "if" with the previous "if"
3557 // A memory slice might have be be rewritten even if it is semantically
3558 // unchanged, if the base_memory value has changed.
3559 set_req(i, new_in);
3560 progress = this; // Report progress
3561 }
3562 }
3563
3564 if (new_base != old_base) {
3565 set_req(Compile::AliasIdxBot, new_base);
3566 // Don't use set_base_memory(new_base), because we need to update du.
3567 assert(base_memory() == new_base, "");
3568 progress = this;
3569 }
3570
3571 if( base_memory() == this ) {
3572 // a self cycle indicates this memory path is dead
3573 set_req(Compile::AliasIdxBot, empty_mem);
3574 }
3575
3576 // Resolve external cycles by calling Ideal on a MergeMem base_memory
3577 // Recursion must occur after the self cycle check above
3578 if( base_memory()->is_MergeMem() ) {
3579 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
3580 Node *m = phase->transform(new_mbase); // Rollup any cycles
3581 if( m != NULL && (m->is_top() ||
3582 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
3583 // propagate rollup of dead cycle to self
3584 set_req(Compile::AliasIdxBot, empty_mem);
3585 }
3586 }
3587
3588 if( base_memory() == empty_mem ) {
3589 progress = this;
3590 // Cut inputs during Parse phase only.
3591 // During Optimize phase a dead MergeMem node will be subsumed by Top.
3592 if( !can_reshape ) {
3593 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3594 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
3595 }
3596 }
3597 }
3598
3599 if( !progress && base_memory()->is_Phi() && can_reshape ) {
3600 // Check if PhiNode::Ideal's "Split phis through memory merges"
3601 // transform should be attempted. Look for this->phi->this cycle.
3602 uint merge_width = req();
3603 if (merge_width > Compile::AliasIdxRaw) {
3604 PhiNode* phi = base_memory()->as_Phi();
3605 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
3606 if (phi->in(i) == this) {
3607 phase->is_IterGVN()->_worklist.push(phi);
3608 break;
3609 }
3610 }
3611 }
3612 }
3613
3614 assert(progress || verify_sparse(), "please, no dups of base");
3615 return progress;
3616 }
3617
3618 //-------------------------set_base_memory-------------------------------------
3619 void MergeMemNode::set_base_memory(Node *new_base) {
3620 Node* empty_mem = empty_memory();
3621 set_req(Compile::AliasIdxBot, new_base);
3622 assert(memory_at(req()) == new_base, "must set default memory");
3623 // Clear out other occurrences of new_base:
3624 if (new_base != empty_mem) {
3625 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3626 if (in(i) == new_base) set_req(i, empty_mem);
3627 }
3628 }
3629 }
3630
3631 //------------------------------out_RegMask------------------------------------
3632 const RegMask &MergeMemNode::out_RegMask() const {
3633 return RegMask::Empty;
3634 }
3635
3636 //------------------------------dump_spec--------------------------------------
3637 #ifndef PRODUCT
3638 void MergeMemNode::dump_spec(outputStream *st) const {
3639 st->print(" {");
3640 Node* base_mem = base_memory();
3641 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
3642 Node* mem = memory_at(i);
3643 if (mem == base_mem) { st->print(" -"); continue; }
3644 st->print( " N%d:", mem->_idx );
3645 Compile::current()->get_adr_type(i)->dump_on(st);
3646 }
3647 st->print(" }");
3648 }
3649 #endif // !PRODUCT
3650
3651
3652 #ifdef ASSERT
3653 static bool might_be_same(Node* a, Node* b) {
3654 if (a == b) return true;
3655 if (!(a->is_Phi() || b->is_Phi())) return false;
3656 // phis shift around during optimization
3657 return true; // pretty stupid...
3658 }
3659
3660 // verify a narrow slice (either incoming or outgoing)
3661 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
3662 if (!VerifyAliases) return; // don't bother to verify unless requested
3663 if (is_error_reported()) return; // muzzle asserts when debugging an error
3664 if (Node::in_dump()) return; // muzzle asserts when printing
3665 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
3666 assert(n != NULL, "");
3667 // Elide intervening MergeMem's
3668 while (n->is_MergeMem()) {
3669 n = n->as_MergeMem()->memory_at(alias_idx);
3670 }
3671 Compile* C = Compile::current();
3672 const TypePtr* n_adr_type = n->adr_type();
3673 if (n == m->empty_memory()) {
3674 // Implicit copy of base_memory()
3675 } else if (n_adr_type != TypePtr::BOTTOM) {
3676 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
3677 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
3678 } else {
3679 // A few places like make_runtime_call "know" that VM calls are narrow,
3680 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
3681 bool expected_wide_mem = false;
3682 if (n == m->base_memory()) {
3683 expected_wide_mem = true;
3684 } else if (alias_idx == Compile::AliasIdxRaw ||
3685 n == m->memory_at(Compile::AliasIdxRaw)) {
3686 expected_wide_mem = true;
3687 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
3688 // memory can "leak through" calls on channels that
3689 // are write-once. Allow this also.
3690 expected_wide_mem = true;
3691 }
3692 assert(expected_wide_mem, "expected narrow slice replacement");
3693 }
3694 }
3695 #else // !ASSERT
3696 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
3697 #endif
3698
3699
3700 //-----------------------------memory_at---------------------------------------
3701 Node* MergeMemNode::memory_at(uint alias_idx) const {
3702 assert(alias_idx >= Compile::AliasIdxRaw ||
3703 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
3704 "must avoid base_memory and AliasIdxTop");
3705
3706 // Otherwise, it is a narrow slice.
3707 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
3708 Compile *C = Compile::current();
3709 if (is_empty_memory(n)) {
3710 // the array is sparse; empty slots are the "top" node
3711 n = base_memory();
3712 assert(Node::in_dump()
3713 || n == NULL || n->bottom_type() == Type::TOP
3714 || n->adr_type() == TypePtr::BOTTOM
3715 || n->adr_type() == TypeRawPtr::BOTTOM
3716 || Compile::current()->AliasLevel() == 0,
3717 "must be a wide memory");
3718 // AliasLevel == 0 if we are organizing the memory states manually.
3719 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
3720 } else {
3721 // make sure the stored slice is sane
3722 #ifdef ASSERT
3723 if (is_error_reported() || Node::in_dump()) {
3724 } else if (might_be_same(n, base_memory())) {
3725 // Give it a pass: It is a mostly harmless repetition of the base.
3726 // This can arise normally from node subsumption during optimization.
3727 } else {
3728 verify_memory_slice(this, alias_idx, n);
3729 }
3730 #endif
3731 }
3732 return n;
3733 }
3734
3735 //---------------------------set_memory_at-------------------------------------
3736 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
3737 verify_memory_slice(this, alias_idx, n);
3738 Node* empty_mem = empty_memory();
3739 if (n == base_memory()) n = empty_mem; // collapse default
3740 uint need_req = alias_idx+1;
3741 if (req() < need_req) {
3742 if (n == empty_mem) return; // already the default, so do not grow me
3743 // grow the sparse array
3744 do {
3745 add_req(empty_mem);
3746 } while (req() < need_req);
3747 }
3748 set_req( alias_idx, n );
3749 }
3750
3751
3752
3753 //--------------------------iteration_setup------------------------------------
3754 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
3755 if (other != NULL) {
3756 grow_to_match(other);
3757 // invariant: the finite support of mm2 is within mm->req()
3758 #ifdef ASSERT
3759 for (uint i = req(); i < other->req(); i++) {
3760 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
3761 }
3762 #endif
3763 }
3764 // Replace spurious copies of base_memory by top.
3765 Node* base_mem = base_memory();
3766 if (base_mem != NULL && !base_mem->is_top()) {
3767 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
3768 if (in(i) == base_mem)
3769 set_req(i, empty_memory());
3770 }
3771 }
3772 }
3773
3774 //---------------------------grow_to_match-------------------------------------
3775 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
3776 Node* empty_mem = empty_memory();
3777 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
3778 // look for the finite support of the other memory
3779 for (uint i = other->req(); --i >= req(); ) {
3780 if (other->in(i) != empty_mem) {
3781 uint new_len = i+1;
3782 while (req() < new_len) add_req(empty_mem);
3783 break;
3784 }
3785 }
3786 }
3787
3788 //---------------------------verify_sparse-------------------------------------
3789 #ifndef PRODUCT
3790 bool MergeMemNode::verify_sparse() const {
3791 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
3792 Node* base_mem = base_memory();
3793 // The following can happen in degenerate cases, since empty==top.
3794 if (is_empty_memory(base_mem)) return true;
3795 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3796 assert(in(i) != NULL, "sane slice");
3797 if (in(i) == base_mem) return false; // should have been the sentinel value!
3798 }
3799 return true;
3800 }
3801
3802 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
3803 Node* n;
3804 n = mm->in(idx);
3805 if (mem == n) return true; // might be empty_memory()
3806 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
3807 if (mem == n) return true;
3808 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
3809 if (mem == n) return true;
3810 if (n == NULL) break;
3811 }
3812 return false;
3813 }
3814 #endif // !PRODUCT