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 t->isa_oopptr() && !t->is_oopptr()->is_instance() &&
138 t->is_oopptr()->cast_to_instance(t_oop->instance_id()) == t_oop) {
139 // clone the Phi with our address type
140 result = mphi->split_out_instance(t_adr, igvn);
141 } else {
142 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
143 }
144 }
145 return result;
146 }
147
148 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
149 uint alias_idx = phase->C->get_alias_index(tp);
150 Node *mem = mmem;
151 #ifdef ASSERT
152 {
153 // Check that current type is consistent with the alias index used during graph construction
154 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
155 bool consistent = adr_check == NULL || adr_check->empty() ||
156 phase->C->must_alias(adr_check, alias_idx );
157 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
158 if( !consistent && adr_check != NULL && !adr_check->empty() &&
159 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
160 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
161 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
162 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
163 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
164 // don't assert if it is dead code.
165 consistent = true;
166 }
167 if( !consistent ) {
168 st->print("alias_idx==%d, adr_check==", alias_idx);
169 if( adr_check == NULL ) {
170 st->print("NULL");
171 } else {
172 adr_check->dump();
173 }
174 st->cr();
175 print_alias_types();
176 assert(consistent, "adr_check must match alias idx");
177 }
178 }
179 #endif
180 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
181 // means an array I have not precisely typed yet. Do not do any
182 // alias stuff with it any time soon.
183 const TypeOopPtr *tinst = tp->isa_oopptr();
184 if( tp->base() != Type::AnyPtr &&
185 !(tinst &&
186 tinst->klass()->is_java_lang_Object() &&
187 tinst->offset() == Type::OffsetBot) ) {
188 // compress paths and change unreachable cycles to TOP
189 // If not, we can update the input infinitely along a MergeMem cycle
190 // Equivalent code in PhiNode::Ideal
191 Node* m = phase->transform(mmem);
192 // If tranformed to a MergeMem, get the desired slice
193 // Otherwise the returned node represents memory for every slice
194 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
195 // Update input if it is progress over what we have now
196 }
197 return mem;
198 }
199
200 //--------------------------Ideal_common---------------------------------------
201 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
202 // Unhook non-raw memories from complete (macro-expanded) initializations.
203 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
204 // If our control input is a dead region, kill all below the region
205 Node *ctl = in(MemNode::Control);
206 if (ctl && remove_dead_region(phase, can_reshape))
207 return this;
208
209 // Ignore if memory is dead, or self-loop
210 Node *mem = in(MemNode::Memory);
211 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
212 assert( mem != this, "dead loop in MemNode::Ideal" );
213
214 Node *address = in(MemNode::Address);
215 const Type *t_adr = phase->type( address );
216 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
217
218 // Avoid independent memory operations
219 Node* old_mem = mem;
220
221 // The code which unhooks non-raw memories from complete (macro-expanded)
222 // initializations was removed. After macro-expansion all stores catched
223 // by Initialize node became raw stores and there is no information
224 // which memory slices they modify. So it is unsafe to move any memory
225 // operation above these stores. Also in most cases hooked non-raw memories
226 // were already unhooked by using information from detect_ptr_independence()
227 // and find_previous_store().
228
229 if (mem->is_MergeMem()) {
230 MergeMemNode* mmem = mem->as_MergeMem();
231 const TypePtr *tp = t_adr->is_ptr();
232
233 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
234 }
235
236 if (mem != old_mem) {
237 set_req(MemNode::Memory, mem);
238 return this;
239 }
240
241 // let the subclass continue analyzing...
242 return NULL;
243 }
244
245 // Helper function for proving some simple control dominations.
246 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
247 // Already assumes that 'dom' is available at 'sub', and that 'sub'
248 // is not a constant (dominated by the method's StartNode).
249 // Used by MemNode::find_previous_store to prove that the
250 // control input of a memory operation predates (dominates)
251 // an allocation it wants to look past.
252 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
253 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
254 return false; // Conservative answer for dead code
255
256 // Check 'dom'. Skip Proj and CatchProj nodes.
257 dom = dom->find_exact_control(dom);
258 if (dom == NULL || dom->is_top())
259 return false; // Conservative answer for dead code
260
261 if (dom == sub) {
262 // For the case when, for example, 'sub' is Initialize and the original
263 // 'dom' is Proj node of the 'sub'.
264 return false;
265 }
266
267 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
268 return true;
269
270 // 'dom' dominates 'sub' if its control edge and control edges
271 // of all its inputs dominate or equal to sub's control edge.
272
273 // Currently 'sub' is either Allocate, Initialize or Start nodes.
274 // Or Region for the check in LoadNode::Ideal();
275 // 'sub' should have sub->in(0) != NULL.
276 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
277 sub->is_Region(), "expecting only these nodes");
278
279 // Get control edge of 'sub'.
280 Node* orig_sub = sub;
281 sub = sub->find_exact_control(sub->in(0));
282 if (sub == NULL || sub->is_top())
283 return false; // Conservative answer for dead code
284
285 assert(sub->is_CFG(), "expecting control");
286
287 if (sub == dom)
288 return true;
289
290 if (sub->is_Start() || sub->is_Root())
291 return false;
292
293 {
294 // Check all control edges of 'dom'.
295
296 ResourceMark rm;
297 Arena* arena = Thread::current()->resource_area();
298 Node_List nlist(arena);
299 Unique_Node_List dom_list(arena);
300
301 dom_list.push(dom);
302 bool only_dominating_controls = false;
303
304 for (uint next = 0; next < dom_list.size(); next++) {
305 Node* n = dom_list.at(next);
306 if (n == orig_sub)
307 return false; // One of dom's inputs dominated by sub.
308 if (!n->is_CFG() && n->pinned()) {
309 // Check only own control edge for pinned non-control nodes.
310 n = n->find_exact_control(n->in(0));
311 if (n == NULL || n->is_top())
312 return false; // Conservative answer for dead code
313 assert(n->is_CFG(), "expecting control");
314 dom_list.push(n);
315 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
316 only_dominating_controls = true;
317 } else if (n->is_CFG()) {
318 if (n->dominates(sub, nlist))
319 only_dominating_controls = true;
320 else
321 return false;
322 } else {
323 // First, own control edge.
324 Node* m = n->find_exact_control(n->in(0));
325 if (m != NULL) {
326 if (m->is_top())
327 return false; // Conservative answer for dead code
328 dom_list.push(m);
329 }
330 // Now, the rest of edges.
331 uint cnt = n->req();
332 for (uint i = 1; i < cnt; i++) {
333 m = n->find_exact_control(n->in(i));
334 if (m == NULL || m->is_top())
335 continue;
336 dom_list.push(m);
337 }
338 }
339 }
340 return only_dominating_controls;
341 }
342 }
343
344 //---------------------detect_ptr_independence---------------------------------
345 // Used by MemNode::find_previous_store to prove that two base
346 // pointers are never equal.
347 // The pointers are accompanied by their associated allocations,
348 // if any, which have been previously discovered by the caller.
349 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
350 Node* p2, AllocateNode* a2,
351 PhaseTransform* phase) {
352 // Attempt to prove that these two pointers cannot be aliased.
353 // They may both manifestly be allocations, and they should differ.
354 // Or, if they are not both allocations, they can be distinct constants.
355 // Otherwise, one is an allocation and the other a pre-existing value.
356 if (a1 == NULL && a2 == NULL) { // neither an allocation
357 return (p1 != p2) && p1->is_Con() && p2->is_Con();
358 } else if (a1 != NULL && a2 != NULL) { // both allocations
359 return (a1 != a2);
360 } else if (a1 != NULL) { // one allocation a1
361 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
362 return all_controls_dominate(p2, a1);
363 } else { //(a2 != NULL) // one allocation a2
364 return all_controls_dominate(p1, a2);
365 }
366 return false;
367 }
368
369
370 // The logic for reordering loads and stores uses four steps:
371 // (a) Walk carefully past stores and initializations which we
372 // can prove are independent of this load.
373 // (b) Observe that the next memory state makes an exact match
374 // with self (load or store), and locate the relevant store.
375 // (c) Ensure that, if we were to wire self directly to the store,
376 // the optimizer would fold it up somehow.
377 // (d) Do the rewiring, and return, depending on some other part of
378 // the optimizer to fold up the load.
379 // This routine handles steps (a) and (b). Steps (c) and (d) are
380 // specific to loads and stores, so they are handled by the callers.
381 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
382 //
383 Node* MemNode::find_previous_store(PhaseTransform* phase) {
384 Node* ctrl = in(MemNode::Control);
385 Node* adr = in(MemNode::Address);
386 intptr_t offset = 0;
387 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
388 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
389
390 if (offset == Type::OffsetBot)
391 return NULL; // cannot unalias unless there are precise offsets
392
393 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
394
395 intptr_t size_in_bytes = memory_size();
396
397 Node* mem = in(MemNode::Memory); // start searching here...
398
399 int cnt = 50; // Cycle limiter
400 for (;;) { // While we can dance past unrelated stores...
401 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
402
403 if (mem->is_Store()) {
404 Node* st_adr = mem->in(MemNode::Address);
405 intptr_t st_offset = 0;
406 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
407 if (st_base == NULL)
408 break; // inscrutable pointer
409 if (st_offset != offset && st_offset != Type::OffsetBot) {
410 const int MAX_STORE = BytesPerLong;
411 if (st_offset >= offset + size_in_bytes ||
412 st_offset <= offset - MAX_STORE ||
413 st_offset <= offset - mem->as_Store()->memory_size()) {
414 // Success: The offsets are provably independent.
415 // (You may ask, why not just test st_offset != offset and be done?
416 // The answer is that stores of different sizes can co-exist
417 // in the same sequence of RawMem effects. We sometimes initialize
418 // a whole 'tile' of array elements with a single jint or jlong.)
419 mem = mem->in(MemNode::Memory);
420 continue; // (a) advance through independent store memory
421 }
422 }
423 if (st_base != base &&
424 detect_ptr_independence(base, alloc,
425 st_base,
426 AllocateNode::Ideal_allocation(st_base, phase),
427 phase)) {
428 // Success: The bases are provably independent.
429 mem = mem->in(MemNode::Memory);
430 continue; // (a) advance through independent store memory
431 }
432
433 // (b) At this point, if the bases or offsets do not agree, we lose,
434 // since we have not managed to prove 'this' and 'mem' independent.
435 if (st_base == base && st_offset == offset) {
436 return mem; // let caller handle steps (c), (d)
437 }
438
439 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
440 InitializeNode* st_init = mem->in(0)->as_Initialize();
441 AllocateNode* st_alloc = st_init->allocation();
442 if (st_alloc == NULL)
443 break; // something degenerated
444 bool known_identical = false;
445 bool known_independent = false;
446 if (alloc == st_alloc)
447 known_identical = true;
448 else if (alloc != NULL)
449 known_independent = true;
450 else if (all_controls_dominate(this, st_alloc))
451 known_independent = true;
452
453 if (known_independent) {
454 // The bases are provably independent: Either they are
455 // manifestly distinct allocations, or else the control
456 // of this load dominates the store's allocation.
457 int alias_idx = phase->C->get_alias_index(adr_type());
458 if (alias_idx == Compile::AliasIdxRaw) {
459 mem = st_alloc->in(TypeFunc::Memory);
460 } else {
461 mem = st_init->memory(alias_idx);
462 }
463 continue; // (a) advance through independent store memory
464 }
465
466 // (b) at this point, if we are not looking at a store initializing
467 // the same allocation we are loading from, we lose.
468 if (known_identical) {
469 // From caller, can_see_stored_value will consult find_captured_store.
470 return mem; // let caller handle steps (c), (d)
471 }
472
473 } else if (addr_t != NULL && addr_t->is_instance_field()) {
474 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
475 if (mem->is_Proj() && mem->in(0)->is_Call()) {
476 CallNode *call = mem->in(0)->as_Call();
477 if (!call->may_modify(addr_t, phase)) {
478 mem = call->in(TypeFunc::Memory);
479 continue; // (a) advance through independent call memory
480 }
481 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
482 mem = mem->in(0)->in(TypeFunc::Memory);
483 continue; // (a) advance through independent MemBar memory
484 } else if (mem->is_MergeMem()) {
485 int alias_idx = phase->C->get_alias_index(adr_type());
486 mem = mem->as_MergeMem()->memory_at(alias_idx);
487 continue; // (a) advance through independent MergeMem memory
488 }
489 }
490
491 // Unless there is an explicit 'continue', we must bail out here,
492 // because 'mem' is an inscrutable memory state (e.g., a call).
493 break;
494 }
495
496 return NULL; // bail out
497 }
498
499 //----------------------calculate_adr_type-------------------------------------
500 // Helper function. Notices when the given type of address hits top or bottom.
501 // Also, asserts a cross-check of the type against the expected address type.
502 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
503 if (t == Type::TOP) return NULL; // does not touch memory any more?
504 #ifdef PRODUCT
505 cross_check = NULL;
506 #else
507 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
508 #endif
509 const TypePtr* tp = t->isa_ptr();
510 if (tp == NULL) {
511 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
512 return TypePtr::BOTTOM; // touches lots of memory
513 } else {
514 #ifdef ASSERT
515 // %%%% [phh] We don't check the alias index if cross_check is
516 // TypeRawPtr::BOTTOM. Needs to be investigated.
517 if (cross_check != NULL &&
518 cross_check != TypePtr::BOTTOM &&
519 cross_check != TypeRawPtr::BOTTOM) {
520 // Recheck the alias index, to see if it has changed (due to a bug).
521 Compile* C = Compile::current();
522 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
523 "must stay in the original alias category");
524 // The type of the address must be contained in the adr_type,
525 // disregarding "null"-ness.
526 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
527 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
528 assert(cross_check->meet(tp_notnull) == cross_check,
529 "real address must not escape from expected memory type");
530 }
531 #endif
532 return tp;
533 }
534 }
535
536 //------------------------adr_phi_is_loop_invariant----------------------------
537 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
538 // loop is loop invariant. Make a quick traversal of Phi and associated
539 // CastPP nodes, looking to see if they are a closed group within the loop.
540 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
541 // The idea is that the phi-nest must boil down to only CastPP nodes
542 // with the same data. This implies that any path into the loop already
543 // includes such a CastPP, and so the original cast, whatever its input,
544 // must be covered by an equivalent cast, with an earlier control input.
545 ResourceMark rm;
546
547 // The loop entry input of the phi should be the unique dominating
548 // node for every Phi/CastPP in the loop.
549 Unique_Node_List closure;
550 closure.push(adr_phi->in(LoopNode::EntryControl));
551
552 // Add the phi node and the cast to the worklist.
553 Unique_Node_List worklist;
554 worklist.push(adr_phi);
555 if( cast != NULL ){
556 if( !cast->is_ConstraintCast() ) return false;
557 worklist.push(cast);
558 }
559
560 // Begin recursive walk of phi nodes.
561 while( worklist.size() ){
562 // Take a node off the worklist
563 Node *n = worklist.pop();
564 if( !closure.member(n) ){
565 // Add it to the closure.
566 closure.push(n);
567 // Make a sanity check to ensure we don't waste too much time here.
568 if( closure.size() > 20) return false;
569 // This node is OK if:
570 // - it is a cast of an identical value
571 // - or it is a phi node (then we add its inputs to the worklist)
572 // Otherwise, the node is not OK, and we presume the cast is not invariant
573 if( n->is_ConstraintCast() ){
574 worklist.push(n->in(1));
575 } else if( n->is_Phi() ) {
576 for( uint i = 1; i < n->req(); i++ ) {
577 worklist.push(n->in(i));
578 }
579 } else {
580 return false;
581 }
582 }
583 }
584
585 // Quit when the worklist is empty, and we've found no offending nodes.
586 return true;
587 }
588
589 //------------------------------Ideal_DU_postCCP-------------------------------
590 // Find any cast-away of null-ness and keep its control. Null cast-aways are
591 // going away in this pass and we need to make this memory op depend on the
592 // gating null check.
593 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
594 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
595 }
596
597 // I tried to leave the CastPP's in. This makes the graph more accurate in
598 // some sense; we get to keep around the knowledge that an oop is not-null
599 // after some test. Alas, the CastPP's interfere with GVN (some values are
600 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
601 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
602 // some of the more trivial cases in the optimizer. Removing more useless
603 // Phi's started allowing Loads to illegally float above null checks. I gave
604 // up on this approach. CNC 10/20/2000
605 // This static method may be called not from MemNode (EncodePNode calls it).
606 // Only the control edge of the node 'n' might be updated.
607 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
608 Node *skipped_cast = NULL;
609 // Need a null check? Regular static accesses do not because they are
610 // from constant addresses. Array ops are gated by the range check (which
611 // always includes a NULL check). Just check field ops.
612 if( n->in(MemNode::Control) == NULL ) {
613 // Scan upwards for the highest location we can place this memory op.
614 while( true ) {
615 switch( adr->Opcode() ) {
616
617 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
618 adr = adr->in(AddPNode::Base);
619 continue;
620
621 case Op_DecodeN: // No change to NULL-ness, so peek thru
622 adr = adr->in(1);
623 continue;
624
625 case Op_CastPP:
626 // If the CastPP is useless, just peek on through it.
627 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
628 // Remember the cast that we've peeked though. If we peek
629 // through more than one, then we end up remembering the highest
630 // one, that is, if in a loop, the one closest to the top.
631 skipped_cast = adr;
632 adr = adr->in(1);
633 continue;
634 }
635 // CastPP is going away in this pass! We need this memory op to be
636 // control-dependent on the test that is guarding the CastPP.
637 ccp->hash_delete(n);
638 n->set_req(MemNode::Control, adr->in(0));
639 ccp->hash_insert(n);
640 return n;
641
642 case Op_Phi:
643 // Attempt to float above a Phi to some dominating point.
644 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
645 // If we've already peeked through a Cast (which could have set the
646 // control), we can't float above a Phi, because the skipped Cast
647 // may not be loop invariant.
648 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
649 adr = adr->in(1);
650 continue;
651 }
652 }
653
654 // Intentional fallthrough!
655
656 // No obvious dominating point. The mem op is pinned below the Phi
657 // by the Phi itself. If the Phi goes away (no true value is merged)
658 // then the mem op can float, but not indefinitely. It must be pinned
659 // behind the controls leading to the Phi.
660 case Op_CheckCastPP:
661 // These usually stick around to change address type, however a
662 // useless one can be elided and we still need to pick up a control edge
663 if (adr->in(0) == NULL) {
664 // This CheckCastPP node has NO control and is likely useless. But we
665 // need check further up the ancestor chain for a control input to keep
666 // the node in place. 4959717.
667 skipped_cast = adr;
668 adr = adr->in(1);
669 continue;
670 }
671 ccp->hash_delete(n);
672 n->set_req(MemNode::Control, adr->in(0));
673 ccp->hash_insert(n);
674 return n;
675
676 // List of "safe" opcodes; those that implicitly block the memory
677 // op below any null check.
678 case Op_CastX2P: // no null checks on native pointers
679 case Op_Parm: // 'this' pointer is not null
680 case Op_LoadP: // Loading from within a klass
681 case Op_LoadN: // Loading from within a klass
682 case Op_LoadKlass: // Loading from within a klass
683 case Op_LoadNKlass: // Loading from within a klass
684 case Op_ConP: // Loading from a klass
685 case Op_ConN: // Loading from a klass
686 case Op_CreateEx: // Sucking up the guts of an exception oop
687 case Op_Con: // Reading from TLS
688 case Op_CMoveP: // CMoveP is pinned
689 case Op_CMoveN: // CMoveN is pinned
690 break; // No progress
691
692 case Op_Proj: // Direct call to an allocation routine
693 case Op_SCMemProj: // Memory state from store conditional ops
694 #ifdef ASSERT
695 {
696 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
697 const Node* call = adr->in(0);
698 if (call->is_CallJava()) {
699 const CallJavaNode* call_java = call->as_CallJava();
700 const TypeTuple *r = call_java->tf()->range();
701 assert(r->cnt() > TypeFunc::Parms, "must return value");
702 const Type* ret_type = r->field_at(TypeFunc::Parms);
703 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
704 // We further presume that this is one of
705 // new_instance_Java, new_array_Java, or
706 // the like, but do not assert for this.
707 } else if (call->is_Allocate()) {
708 // similar case to new_instance_Java, etc.
709 } else if (!call->is_CallLeaf()) {
710 // Projections from fetch_oop (OSR) are allowed as well.
711 ShouldNotReachHere();
712 }
713 }
714 #endif
715 break;
716 default:
717 ShouldNotReachHere();
718 }
719 break;
720 }
721 }
722
723 return NULL; // No progress
724 }
725
726
727 //=============================================================================
728 uint LoadNode::size_of() const { return sizeof(*this); }
729 uint LoadNode::cmp( const Node &n ) const
730 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
731 const Type *LoadNode::bottom_type() const { return _type; }
732 uint LoadNode::ideal_reg() const {
733 return Matcher::base2reg[_type->base()];
734 }
735
736 #ifndef PRODUCT
737 void LoadNode::dump_spec(outputStream *st) const {
738 MemNode::dump_spec(st);
739 if( !Verbose && !WizardMode ) {
740 // standard dump does this in Verbose and WizardMode
741 st->print(" #"); _type->dump_on(st);
742 }
743 }
744 #endif
745
746
747 //----------------------------LoadNode::make-----------------------------------
748 // Polymorphic factory method:
749 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
750 Compile* C = gvn.C;
751
752 // sanity check the alias category against the created node type
753 assert(!(adr_type->isa_oopptr() &&
754 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
755 "use LoadKlassNode instead");
756 assert(!(adr_type->isa_aryptr() &&
757 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
758 "use LoadRangeNode instead");
759 switch (bt) {
760 case T_BOOLEAN:
761 case T_BYTE: return new (C, 3) LoadBNode(ctl, mem, adr, adr_type, rt->is_int() );
762 case T_INT: return new (C, 3) LoadINode(ctl, mem, adr, adr_type, rt->is_int() );
763 case T_CHAR: return new (C, 3) LoadCNode(ctl, mem, adr, adr_type, rt->is_int() );
764 case T_SHORT: return new (C, 3) LoadSNode(ctl, mem, adr, adr_type, rt->is_int() );
765 case T_LONG: return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long() );
766 case T_FLOAT: return new (C, 3) LoadFNode(ctl, mem, adr, adr_type, rt );
767 case T_DOUBLE: return new (C, 3) LoadDNode(ctl, mem, adr, adr_type, rt );
768 case T_ADDRESS: return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr() );
769 case T_OBJECT:
770 #ifdef _LP64
771 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
772 const TypeNarrowOop* narrowtype;
773 if (rt->isa_narrowoop()) {
774 narrowtype = rt->is_narrowoop();
775 } else {
776 narrowtype = rt->is_oopptr()->make_narrowoop();
777 }
778 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, narrowtype));
779
780 return DecodeNNode::decode(&gvn, load);
781 } else
782 #endif
783 {
784 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop");
785 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
786 }
787 }
788 ShouldNotReachHere();
789 return (LoadNode*)NULL;
790 }
791
792 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
793 bool require_atomic = true;
794 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
795 }
796
797
798
799
800 //------------------------------hash-------------------------------------------
801 uint LoadNode::hash() const {
802 // unroll addition of interesting fields
803 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
804 }
805
806 //---------------------------can_see_stored_value------------------------------
807 // This routine exists to make sure this set of tests is done the same
808 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
809 // will change the graph shape in a way which makes memory alive twice at the
810 // same time (uses the Oracle model of aliasing), then some
811 // LoadXNode::Identity will fold things back to the equivalence-class model
812 // of aliasing.
813 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
814 Node* ld_adr = in(MemNode::Address);
815
816 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
817 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
818 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
819 atp->field() != NULL && !atp->field()->is_volatile()) {
820 uint alias_idx = atp->index();
821 bool final = atp->field()->is_final();
822 Node* result = NULL;
823 Node* current = st;
824 // Skip through chains of MemBarNodes checking the MergeMems for
825 // new states for the slice of this load. Stop once any other
826 // kind of node is encountered. Loads from final memory can skip
827 // through any kind of MemBar but normal loads shouldn't skip
828 // through MemBarAcquire since the could allow them to move out of
829 // a synchronized region.
830 while (current->is_Proj()) {
831 int opc = current->in(0)->Opcode();
832 if ((final && opc == Op_MemBarAcquire) ||
833 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
834 Node* mem = current->in(0)->in(TypeFunc::Memory);
835 if (mem->is_MergeMem()) {
836 MergeMemNode* merge = mem->as_MergeMem();
837 Node* new_st = merge->memory_at(alias_idx);
838 if (new_st == merge->base_memory()) {
839 // Keep searching
840 current = merge->base_memory();
841 continue;
842 }
843 // Save the new memory state for the slice and fall through
844 // to exit.
845 result = new_st;
846 }
847 }
848 break;
849 }
850 if (result != NULL) {
851 st = result;
852 }
853 }
854
855
856 // Loop around twice in the case Load -> Initialize -> Store.
857 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
858 for (int trip = 0; trip <= 1; trip++) {
859
860 if (st->is_Store()) {
861 Node* st_adr = st->in(MemNode::Address);
862 if (!phase->eqv(st_adr, ld_adr)) {
863 // Try harder before giving up... Match raw and non-raw pointers.
864 intptr_t st_off = 0;
865 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
866 if (alloc == NULL) return NULL;
867 intptr_t ld_off = 0;
868 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
869 if (alloc != allo2) return NULL;
870 if (ld_off != st_off) return NULL;
871 // At this point we have proven something like this setup:
872 // A = Allocate(...)
873 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
874 // S = StoreQ(, AddP(, A.Parm , #Off), V)
875 // (Actually, we haven't yet proven the Q's are the same.)
876 // In other words, we are loading from a casted version of
877 // the same pointer-and-offset that we stored to.
878 // Thus, we are able to replace L by V.
879 }
880 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
881 if (store_Opcode() != st->Opcode())
882 return NULL;
883 return st->in(MemNode::ValueIn);
884 }
885
886 intptr_t offset = 0; // scratch
887
888 // A load from a freshly-created object always returns zero.
889 // (This can happen after LoadNode::Ideal resets the load's memory input
890 // to find_captured_store, which returned InitializeNode::zero_memory.)
891 if (st->is_Proj() && st->in(0)->is_Allocate() &&
892 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
893 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
894 // return a zero value for the load's basic type
895 // (This is one of the few places where a generic PhaseTransform
896 // can create new nodes. Think of it as lazily manifesting
897 // virtually pre-existing constants.)
898 return phase->zerocon(memory_type());
899 }
900
901 // A load from an initialization barrier can match a captured store.
902 if (st->is_Proj() && st->in(0)->is_Initialize()) {
903 InitializeNode* init = st->in(0)->as_Initialize();
904 AllocateNode* alloc = init->allocation();
905 if (alloc != NULL &&
906 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
907 // examine a captured store value
908 st = init->find_captured_store(offset, memory_size(), phase);
909 if (st != NULL)
910 continue; // take one more trip around
911 }
912 }
913
914 break;
915 }
916
917 return NULL;
918 }
919
920 //----------------------is_instance_field_load_with_local_phi------------------
921 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
922 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
923 in(MemNode::Address)->is_AddP() ) {
924 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
925 // Only instances.
926 if( t_oop != NULL && t_oop->is_instance_field() &&
927 t_oop->offset() != Type::OffsetBot &&
928 t_oop->offset() != Type::OffsetTop) {
929 return true;
930 }
931 }
932 return false;
933 }
934
935 //------------------------------Identity---------------------------------------
936 // Loads are identity if previous store is to same address
937 Node *LoadNode::Identity( PhaseTransform *phase ) {
938 // If the previous store-maker is the right kind of Store, and the store is
939 // to the same address, then we are equal to the value stored.
940 Node* mem = in(MemNode::Memory);
941 Node* value = can_see_stored_value(mem, phase);
942 if( value ) {
943 // byte, short & char stores truncate naturally.
944 // A load has to load the truncated value which requires
945 // some sort of masking operation and that requires an
946 // Ideal call instead of an Identity call.
947 if (memory_size() < BytesPerInt) {
948 // If the input to the store does not fit with the load's result type,
949 // it must be truncated via an Ideal call.
950 if (!phase->type(value)->higher_equal(phase->type(this)))
951 return this;
952 }
953 // (This works even when value is a Con, but LoadNode::Value
954 // usually runs first, producing the singleton type of the Con.)
955 return value;
956 }
957
958 // Search for an existing data phi which was generated before for the same
959 // instance's field to avoid infinite genertion of phis in a loop.
960 Node *region = mem->in(0);
961 if (is_instance_field_load_with_local_phi(region)) {
962 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
963 int this_index = phase->C->get_alias_index(addr_t);
964 int this_offset = addr_t->offset();
965 int this_id = addr_t->is_oopptr()->instance_id();
966 const Type* this_type = bottom_type();
967 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
968 Node* phi = region->fast_out(i);
969 if (phi->is_Phi() && phi != mem &&
970 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
971 return phi;
972 }
973 }
974 }
975
976 return this;
977 }
978
979
980 // Returns true if the AliasType refers to the field that holds the
981 // cached box array. Currently only handles the IntegerCache case.
982 static bool is_autobox_cache(Compile::AliasType* atp) {
983 if (atp != NULL && atp->field() != NULL) {
984 ciField* field = atp->field();
985 ciSymbol* klass = field->holder()->name();
986 if (field->name() == ciSymbol::cache_field_name() &&
987 field->holder()->uses_default_loader() &&
988 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
989 return true;
990 }
991 }
992 return false;
993 }
994
995 // Fetch the base value in the autobox array
996 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
997 if (atp != NULL && atp->field() != NULL) {
998 ciField* field = atp->field();
999 ciSymbol* klass = field->holder()->name();
1000 if (field->name() == ciSymbol::cache_field_name() &&
1001 field->holder()->uses_default_loader() &&
1002 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
1003 assert(field->is_constant(), "what?");
1004 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
1005 // Fetch the box object at the base of the array and get its value
1006 ciInstance* box = array->obj_at(0)->as_instance();
1007 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1008 if (ik->nof_nonstatic_fields() == 1) {
1009 // This should be true nonstatic_field_at requires calling
1010 // nof_nonstatic_fields so check it anyway
1011 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1012 cache_offset = c.as_int();
1013 }
1014 return true;
1015 }
1016 }
1017 return false;
1018 }
1019
1020 // Returns true if the AliasType refers to the value field of an
1021 // autobox object. Currently only handles Integer.
1022 static bool is_autobox_object(Compile::AliasType* atp) {
1023 if (atp != NULL && atp->field() != NULL) {
1024 ciField* field = atp->field();
1025 ciSymbol* klass = field->holder()->name();
1026 if (field->name() == ciSymbol::value_name() &&
1027 field->holder()->uses_default_loader() &&
1028 klass == ciSymbol::java_lang_Integer()) {
1029 return true;
1030 }
1031 }
1032 return false;
1033 }
1034
1035
1036 // We're loading from an object which has autobox behaviour.
1037 // If this object is result of a valueOf call we'll have a phi
1038 // merging a newly allocated object and a load from the cache.
1039 // We want to replace this load with the original incoming
1040 // argument to the valueOf call.
1041 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1042 Node* base = in(Address)->in(AddPNode::Base);
1043 if (base->is_Phi() && base->req() == 3) {
1044 AllocateNode* allocation = NULL;
1045 int allocation_index = -1;
1046 int load_index = -1;
1047 for (uint i = 1; i < base->req(); i++) {
1048 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
1049 if (allocation != NULL) {
1050 allocation_index = i;
1051 load_index = 3 - allocation_index;
1052 break;
1053 }
1054 }
1055 LoadNode* load = NULL;
1056 if (allocation != NULL && base->in(load_index)->is_Load()) {
1057 load = base->in(load_index)->as_Load();
1058 }
1059 if (load != NULL && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
1060 // Push the loads from the phi that comes from valueOf up
1061 // through it to allow elimination of the loads and the recovery
1062 // of the original value.
1063 Node* mem_phi = in(Memory);
1064 Node* offset = in(Address)->in(AddPNode::Offset);
1065
1066 Node* in1 = clone();
1067 Node* in1_addr = in1->in(Address)->clone();
1068 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
1069 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
1070 in1_addr->set_req(AddPNode::Offset, offset);
1071 in1->set_req(0, base->in(allocation_index));
1072 in1->set_req(Address, in1_addr);
1073 in1->set_req(Memory, mem_phi->in(allocation_index));
1074
1075 Node* in2 = clone();
1076 Node* in2_addr = in2->in(Address)->clone();
1077 in2_addr->set_req(AddPNode::Base, base->in(load_index));
1078 in2_addr->set_req(AddPNode::Address, base->in(load_index));
1079 in2_addr->set_req(AddPNode::Offset, offset);
1080 in2->set_req(0, base->in(load_index));
1081 in2->set_req(Address, in2_addr);
1082 in2->set_req(Memory, mem_phi->in(load_index));
1083
1084 in1_addr = phase->transform(in1_addr);
1085 in1 = phase->transform(in1);
1086 in2_addr = phase->transform(in2_addr);
1087 in2 = phase->transform(in2);
1088
1089 PhiNode* result = PhiNode::make_blank(base->in(0), this);
1090 result->set_req(allocation_index, in1);
1091 result->set_req(load_index, in2);
1092 return result;
1093 }
1094 } else if (base->is_Load()) {
1095 // Eliminate the load of Integer.value for integers from the cache
1096 // array by deriving the value from the index into the array.
1097 // Capture the offset of the load and then reverse the computation.
1098 Node* load_base = base->in(Address)->in(AddPNode::Base);
1099 if (load_base != NULL) {
1100 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
1101 intptr_t cache_offset;
1102 int shift = -1;
1103 Node* cache = NULL;
1104 if (is_autobox_cache(atp)) {
1105 shift = exact_log2(type2aelembytes(T_OBJECT));
1106 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
1107 }
1108 if (cache != NULL && base->in(Address)->is_AddP()) {
1109 Node* elements[4];
1110 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
1111 int cache_low;
1112 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
1113 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
1114 // Add up all the offsets making of the address of the load
1115 Node* result = elements[0];
1116 for (int i = 1; i < count; i++) {
1117 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
1118 }
1119 // Remove the constant offset from the address and then
1120 // remove the scaling of the offset to recover the original index.
1121 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
1122 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1123 // Peel the shift off directly but wrap it in a dummy node
1124 // since Ideal can't return existing nodes
1125 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
1126 } else {
1127 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
1128 }
1129 #ifdef _LP64
1130 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
1131 #endif
1132 return result;
1133 }
1134 }
1135 }
1136 }
1137 return NULL;
1138 }
1139
1140 //------------------------------split_through_phi------------------------------
1141 // Split instance field load through Phi.
1142 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1143 Node* mem = in(MemNode::Memory);
1144 Node* address = in(MemNode::Address);
1145 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1146 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1147
1148 assert(mem->is_Phi() && (t_oop != NULL) &&
1149 t_oop->is_instance_field(), "invalide conditions");
1150
1151 Node *region = mem->in(0);
1152 if (region == NULL) {
1153 return NULL; // Wait stable graph
1154 }
1155 uint cnt = mem->req();
1156 for( uint i = 1; i < cnt; i++ ) {
1157 Node *in = mem->in(i);
1158 if( in == NULL ) {
1159 return NULL; // Wait stable graph
1160 }
1161 }
1162 // Check for loop invariant.
1163 if (cnt == 3) {
1164 for( uint i = 1; i < cnt; i++ ) {
1165 Node *in = mem->in(i);
1166 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
1167 if (m == mem) {
1168 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
1169 return this;
1170 }
1171 }
1172 }
1173 // Split through Phi (see original code in loopopts.cpp).
1174 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
1175
1176 // Do nothing here if Identity will find a value
1177 // (to avoid infinite chain of value phis generation).
1178 if ( !phase->eqv(this, this->Identity(phase)) )
1179 return NULL;
1180
1181 // Skip the split if the region dominates some control edge of the address.
1182 if (cnt == 3 && !MemNode::all_controls_dominate(address, region))
1183 return NULL;
1184
1185 const Type* this_type = this->bottom_type();
1186 int this_index = phase->C->get_alias_index(addr_t);
1187 int this_offset = addr_t->offset();
1188 int this_iid = addr_t->is_oopptr()->instance_id();
1189 int wins = 0;
1190 PhaseIterGVN *igvn = phase->is_IterGVN();
1191 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1192 for( uint i = 1; i < region->req(); i++ ) {
1193 Node *x;
1194 Node* the_clone = NULL;
1195 if( region->in(i) == phase->C->top() ) {
1196 x = phase->C->top(); // Dead path? Use a dead data op
1197 } else {
1198 x = this->clone(); // Else clone up the data op
1199 the_clone = x; // Remember for possible deletion.
1200 // Alter data node to use pre-phi inputs
1201 if( this->in(0) == region ) {
1202 x->set_req( 0, region->in(i) );
1203 } else {
1204 x->set_req( 0, NULL );
1205 }
1206 for( uint j = 1; j < this->req(); j++ ) {
1207 Node *in = this->in(j);
1208 if( in->is_Phi() && in->in(0) == region )
1209 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
1210 }
1211 }
1212 // Check for a 'win' on some paths
1213 const Type *t = x->Value(igvn);
1214
1215 bool singleton = t->singleton();
1216
1217 // See comments in PhaseIdealLoop::split_thru_phi().
1218 if( singleton && t == Type::TOP ) {
1219 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1220 }
1221
1222 if( singleton ) {
1223 wins++;
1224 x = igvn->makecon(t);
1225 } else {
1226 // We now call Identity to try to simplify the cloned node.
1227 // Note that some Identity methods call phase->type(this).
1228 // Make sure that the type array is big enough for
1229 // our new node, even though we may throw the node away.
1230 // (This tweaking with igvn only works because x is a new node.)
1231 igvn->set_type(x, t);
1232 Node *y = x->Identity(igvn);
1233 if( y != x ) {
1234 wins++;
1235 x = y;
1236 } else {
1237 y = igvn->hash_find(x);
1238 if( y ) {
1239 wins++;
1240 x = y;
1241 } else {
1242 // Else x is a new node we are keeping
1243 // We do not need register_new_node_with_optimizer
1244 // because set_type has already been called.
1245 igvn->_worklist.push(x);
1246 }
1247 }
1248 }
1249 if (x != the_clone && the_clone != NULL)
1250 igvn->remove_dead_node(the_clone);
1251 phi->set_req(i, x);
1252 }
1253 if( wins > 0 ) {
1254 // Record Phi
1255 igvn->register_new_node_with_optimizer(phi);
1256 return phi;
1257 }
1258 igvn->remove_dead_node(phi);
1259 return NULL;
1260 }
1261
1262 //------------------------------Ideal------------------------------------------
1263 // If the load is from Field memory and the pointer is non-null, we can
1264 // zero out the control input.
1265 // If the offset is constant and the base is an object allocation,
1266 // try to hook me up to the exact initializing store.
1267 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1268 Node* p = MemNode::Ideal_common(phase, can_reshape);
1269 if (p) return (p == NodeSentinel) ? NULL : p;
1270
1271 Node* ctrl = in(MemNode::Control);
1272 Node* address = in(MemNode::Address);
1273
1274 // Skip up past a SafePoint control. Cannot do this for Stores because
1275 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1276 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1277 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1278 ctrl = ctrl->in(0);
1279 set_req(MemNode::Control,ctrl);
1280 }
1281
1282 // Check for useless control edge in some common special cases
1283 if (in(MemNode::Control) != NULL) {
1284 intptr_t ignore = 0;
1285 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1286 if (base != NULL
1287 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1288 && all_controls_dominate(base, phase->C->start())) {
1289 // A method-invariant, non-null address (constant or 'this' argument).
1290 set_req(MemNode::Control, NULL);
1291 }
1292 }
1293
1294 if (EliminateAutoBox && can_reshape && in(Address)->is_AddP()) {
1295 Node* base = in(Address)->in(AddPNode::Base);
1296 if (base != NULL) {
1297 Compile::AliasType* atp = phase->C->alias_type(adr_type());
1298 if (is_autobox_object(atp)) {
1299 Node* result = eliminate_autobox(phase);
1300 if (result != NULL) return result;
1301 }
1302 }
1303 }
1304
1305 Node* mem = in(MemNode::Memory);
1306 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1307
1308 if (addr_t != NULL) {
1309 // try to optimize our memory input
1310 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
1311 if (opt_mem != mem) {
1312 set_req(MemNode::Memory, opt_mem);
1313 return this;
1314 }
1315 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1316 if (can_reshape && opt_mem->is_Phi() &&
1317 (t_oop != NULL) && t_oop->is_instance_field()) {
1318 // Split instance field load through Phi.
1319 Node* result = split_through_phi(phase);
1320 if (result != NULL) return result;
1321 }
1322 }
1323
1324 // Check for prior store with a different base or offset; make Load
1325 // independent. Skip through any number of them. Bail out if the stores
1326 // are in an endless dead cycle and report no progress. This is a key
1327 // transform for Reflection. However, if after skipping through the Stores
1328 // we can't then fold up against a prior store do NOT do the transform as
1329 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1330 // array memory alive twice: once for the hoisted Load and again after the
1331 // bypassed Store. This situation only works if EVERYBODY who does
1332 // anti-dependence work knows how to bypass. I.e. we need all
1333 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1334 // the alias index stuff. So instead, peek through Stores and IFF we can
1335 // fold up, do so.
1336 Node* prev_mem = find_previous_store(phase);
1337 // Steps (a), (b): Walk past independent stores to find an exact match.
1338 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1339 // (c) See if we can fold up on the spot, but don't fold up here.
1340 // Fold-up might require truncation (for LoadB/LoadS/LoadC) or
1341 // just return a prior value, which is done by Identity calls.
1342 if (can_see_stored_value(prev_mem, phase)) {
1343 // Make ready for step (d):
1344 set_req(MemNode::Memory, prev_mem);
1345 return this;
1346 }
1347 }
1348
1349 return NULL; // No further progress
1350 }
1351
1352 // Helper to recognize certain Klass fields which are invariant across
1353 // some group of array types (e.g., int[] or all T[] where T < Object).
1354 const Type*
1355 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1356 ciKlass* klass) const {
1357 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1358 // The field is Klass::_modifier_flags. Return its (constant) value.
1359 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1360 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1361 return TypeInt::make(klass->modifier_flags());
1362 }
1363 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1364 // The field is Klass::_access_flags. Return its (constant) value.
1365 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1366 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1367 return TypeInt::make(klass->access_flags());
1368 }
1369 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
1370 // The field is Klass::_layout_helper. Return its constant value if known.
1371 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1372 return TypeInt::make(klass->layout_helper());
1373 }
1374
1375 // No match.
1376 return NULL;
1377 }
1378
1379 //------------------------------Value-----------------------------------------
1380 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1381 // Either input is TOP ==> the result is TOP
1382 Node* mem = in(MemNode::Memory);
1383 const Type *t1 = phase->type(mem);
1384 if (t1 == Type::TOP) return Type::TOP;
1385 Node* adr = in(MemNode::Address);
1386 const TypePtr* tp = phase->type(adr)->isa_ptr();
1387 if (tp == NULL || tp->empty()) return Type::TOP;
1388 int off = tp->offset();
1389 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1390
1391 // Try to guess loaded type from pointer type
1392 if (tp->base() == Type::AryPtr) {
1393 const Type *t = tp->is_aryptr()->elem();
1394 // Don't do this for integer types. There is only potential profit if
1395 // the element type t is lower than _type; that is, for int types, if _type is
1396 // more restrictive than t. This only happens here if one is short and the other
1397 // char (both 16 bits), and in those cases we've made an intentional decision
1398 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1399 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1400 //
1401 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1402 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1403 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1404 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1405 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1406 // In fact, that could have been the original type of p1, and p1 could have
1407 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1408 // expression (LShiftL quux 3) independently optimized to the constant 8.
1409 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1410 && Opcode() != Op_LoadKlass) {
1411 // t might actually be lower than _type, if _type is a unique
1412 // concrete subclass of abstract class t.
1413 // Make sure the reference is not into the header, by comparing
1414 // the offset against the offset of the start of the array's data.
1415 // Different array types begin at slightly different offsets (12 vs. 16).
1416 // We choose T_BYTE as an example base type that is least restrictive
1417 // as to alignment, which will therefore produce the smallest
1418 // possible base offset.
1419 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1420 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
1421 const Type* jt = t->join(_type);
1422 // In any case, do not allow the join, per se, to empty out the type.
1423 if (jt->empty() && !t->empty()) {
1424 // This can happen if a interface-typed array narrows to a class type.
1425 jt = _type;
1426 }
1427
1428 if (EliminateAutoBox) {
1429 // The pointers in the autobox arrays are always non-null
1430 Node* base = in(Address)->in(AddPNode::Base);
1431 if (base != NULL) {
1432 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
1433 if (is_autobox_cache(atp)) {
1434 return jt->join(TypePtr::NOTNULL)->is_ptr();
1435 }
1436 }
1437 }
1438 return jt;
1439 }
1440 }
1441 } else if (tp->base() == Type::InstPtr) {
1442 assert( off != Type::OffsetBot ||
1443 // arrays can be cast to Objects
1444 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1445 // unsafe field access may not have a constant offset
1446 phase->C->has_unsafe_access(),
1447 "Field accesses must be precise" );
1448 // For oop loads, we expect the _type to be precise
1449 } else if (tp->base() == Type::KlassPtr) {
1450 assert( off != Type::OffsetBot ||
1451 // arrays can be cast to Objects
1452 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1453 // also allow array-loading from the primary supertype
1454 // array during subtype checks
1455 Opcode() == Op_LoadKlass,
1456 "Field accesses must be precise" );
1457 // For klass/static loads, we expect the _type to be precise
1458 }
1459
1460 const TypeKlassPtr *tkls = tp->isa_klassptr();
1461 if (tkls != NULL && !StressReflectiveCode) {
1462 ciKlass* klass = tkls->klass();
1463 if (klass->is_loaded() && tkls->klass_is_exact()) {
1464 // We are loading a field from a Klass metaobject whose identity
1465 // is known at compile time (the type is "exact" or "precise").
1466 // Check for fields we know are maintained as constants by the VM.
1467 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
1468 // The field is Klass::_super_check_offset. Return its (constant) value.
1469 // (Folds up type checking code.)
1470 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1471 return TypeInt::make(klass->super_check_offset());
1472 }
1473 // Compute index into primary_supers array
1474 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1475 // Check for overflowing; use unsigned compare to handle the negative case.
1476 if( depth < ciKlass::primary_super_limit() ) {
1477 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1478 // (Folds up type checking code.)
1479 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1480 ciKlass *ss = klass->super_of_depth(depth);
1481 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1482 }
1483 const Type* aift = load_array_final_field(tkls, klass);
1484 if (aift != NULL) return aift;
1485 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
1486 && klass->is_array_klass()) {
1487 // The field is arrayKlass::_component_mirror. Return its (constant) value.
1488 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1489 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1490 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1491 }
1492 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
1493 // The field is Klass::_java_mirror. Return its (constant) value.
1494 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1495 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1496 return TypeInstPtr::make(klass->java_mirror());
1497 }
1498 }
1499
1500 // We can still check if we are loading from the primary_supers array at a
1501 // shallow enough depth. Even though the klass is not exact, entries less
1502 // than or equal to its super depth are correct.
1503 if (klass->is_loaded() ) {
1504 ciType *inner = klass->klass();
1505 while( inner->is_obj_array_klass() )
1506 inner = inner->as_obj_array_klass()->base_element_type();
1507 if( inner->is_instance_klass() &&
1508 !inner->as_instance_klass()->flags().is_interface() ) {
1509 // Compute index into primary_supers array
1510 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1511 // Check for overflowing; use unsigned compare to handle the negative case.
1512 if( depth < ciKlass::primary_super_limit() &&
1513 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1514 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1515 // (Folds up type checking code.)
1516 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1517 ciKlass *ss = klass->super_of_depth(depth);
1518 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1519 }
1520 }
1521 }
1522
1523 // If the type is enough to determine that the thing is not an array,
1524 // we can give the layout_helper a positive interval type.
1525 // This will help short-circuit some reflective code.
1526 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
1527 && !klass->is_array_klass() // not directly typed as an array
1528 && !klass->is_interface() // specifically not Serializable & Cloneable
1529 && !klass->is_java_lang_Object() // not the supertype of all T[]
1530 ) {
1531 // Note: When interfaces are reliable, we can narrow the interface
1532 // test to (klass != Serializable && klass != Cloneable).
1533 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1534 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1535 // The key property of this type is that it folds up tests
1536 // for array-ness, since it proves that the layout_helper is positive.
1537 // Thus, a generic value like the basic object layout helper works fine.
1538 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1539 }
1540 }
1541
1542 // If we are loading from a freshly-allocated object, produce a zero,
1543 // if the load is provably beyond the header of the object.
1544 // (Also allow a variable load from a fresh array to produce zero.)
1545 if (ReduceFieldZeroing) {
1546 Node* value = can_see_stored_value(mem,phase);
1547 if (value != NULL && value->is_Con())
1548 return value->bottom_type();
1549 }
1550
1551 const TypeOopPtr *tinst = tp->isa_oopptr();
1552 if (tinst != NULL && tinst->is_instance_field()) {
1553 // If we have an instance type and our memory input is the
1554 // programs's initial memory state, there is no matching store,
1555 // so just return a zero of the appropriate type
1556 Node *mem = in(MemNode::Memory);
1557 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1558 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1559 return Type::get_zero_type(_type->basic_type());
1560 }
1561 }
1562 return _type;
1563 }
1564
1565 //------------------------------match_edge-------------------------------------
1566 // Do we Match on this edge index or not? Match only the address.
1567 uint LoadNode::match_edge(uint idx) const {
1568 return idx == MemNode::Address;
1569 }
1570
1571 //--------------------------LoadBNode::Ideal--------------------------------------
1572 //
1573 // If the previous store is to the same address as this load,
1574 // and the value stored was larger than a byte, replace this load
1575 // with the value stored truncated to a byte. If no truncation is
1576 // needed, the replacement is done in LoadNode::Identity().
1577 //
1578 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1579 Node* mem = in(MemNode::Memory);
1580 Node* value = can_see_stored_value(mem,phase);
1581 if( value && !phase->type(value)->higher_equal( _type ) ) {
1582 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
1583 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
1584 }
1585 // Identity call will handle the case where truncation is not needed.
1586 return LoadNode::Ideal(phase, can_reshape);
1587 }
1588
1589 //--------------------------LoadCNode::Ideal--------------------------------------
1590 //
1591 // If the previous store is to the same address as this load,
1592 // and the value stored was larger than a char, replace this load
1593 // with the value stored truncated to a char. If no truncation is
1594 // needed, the replacement is done in LoadNode::Identity().
1595 //
1596 Node *LoadCNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1597 Node* mem = in(MemNode::Memory);
1598 Node* value = can_see_stored_value(mem,phase);
1599 if( value && !phase->type(value)->higher_equal( _type ) )
1600 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
1601 // Identity call will handle the case where truncation is not needed.
1602 return LoadNode::Ideal(phase, can_reshape);
1603 }
1604
1605 //--------------------------LoadSNode::Ideal--------------------------------------
1606 //
1607 // If the previous store is to the same address as this load,
1608 // and the value stored was larger than a short, replace this load
1609 // with the value stored truncated to a short. If no truncation is
1610 // needed, the replacement is done in LoadNode::Identity().
1611 //
1612 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1613 Node* mem = in(MemNode::Memory);
1614 Node* value = can_see_stored_value(mem,phase);
1615 if( value && !phase->type(value)->higher_equal( _type ) ) {
1616 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
1617 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
1618 }
1619 // Identity call will handle the case where truncation is not needed.
1620 return LoadNode::Ideal(phase, can_reshape);
1621 }
1622
1623 //=============================================================================
1624 //----------------------------LoadKlassNode::make------------------------------
1625 // Polymorphic factory method:
1626 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
1627 Compile* C = gvn.C;
1628 Node *ctl = NULL;
1629 // sanity check the alias category against the created node type
1630 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr();
1631 assert(adr_type != NULL, "expecting TypeOopPtr");
1632 #ifdef _LP64
1633 if (adr_type->is_ptr_to_narrowoop()) {
1634 const TypeNarrowOop* narrowtype = tk->is_oopptr()->make_narrowoop();
1635 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, narrowtype));
1636 return DecodeNNode::decode(&gvn, load_klass);
1637 }
1638 #endif
1639 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
1640 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk);
1641 }
1642
1643 //------------------------------Value------------------------------------------
1644 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1645 return klass_value_common(phase);
1646 }
1647
1648 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
1649 // Either input is TOP ==> the result is TOP
1650 const Type *t1 = phase->type( in(MemNode::Memory) );
1651 if (t1 == Type::TOP) return Type::TOP;
1652 Node *adr = in(MemNode::Address);
1653 const Type *t2 = phase->type( adr );
1654 if (t2 == Type::TOP) return Type::TOP;
1655 const TypePtr *tp = t2->is_ptr();
1656 if (TypePtr::above_centerline(tp->ptr()) ||
1657 tp->ptr() == TypePtr::Null) return Type::TOP;
1658
1659 // Return a more precise klass, if possible
1660 const TypeInstPtr *tinst = tp->isa_instptr();
1661 if (tinst != NULL) {
1662 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1663 int offset = tinst->offset();
1664 if (ik == phase->C->env()->Class_klass()
1665 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1666 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1667 // We are loading a special hidden field from a Class mirror object,
1668 // the field which points to the VM's Klass metaobject.
1669 ciType* t = tinst->java_mirror_type();
1670 // java_mirror_type returns non-null for compile-time Class constants.
1671 if (t != NULL) {
1672 // constant oop => constant klass
1673 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1674 return TypeKlassPtr::make(ciArrayKlass::make(t));
1675 }
1676 if (!t->is_klass()) {
1677 // a primitive Class (e.g., int.class) has NULL for a klass field
1678 return TypePtr::NULL_PTR;
1679 }
1680 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1681 return TypeKlassPtr::make(t->as_klass());
1682 }
1683 // non-constant mirror, so we can't tell what's going on
1684 }
1685 if( !ik->is_loaded() )
1686 return _type; // Bail out if not loaded
1687 if (offset == oopDesc::klass_offset_in_bytes()) {
1688 if (tinst->klass_is_exact()) {
1689 return TypeKlassPtr::make(ik);
1690 }
1691 // See if we can become precise: no subklasses and no interface
1692 // (Note: We need to support verified interfaces.)
1693 if (!ik->is_interface() && !ik->has_subklass()) {
1694 //assert(!UseExactTypes, "this code should be useless with exact types");
1695 // Add a dependence; if any subclass added we need to recompile
1696 if (!ik->is_final()) {
1697 // %%% should use stronger assert_unique_concrete_subtype instead
1698 phase->C->dependencies()->assert_leaf_type(ik);
1699 }
1700 // Return precise klass
1701 return TypeKlassPtr::make(ik);
1702 }
1703
1704 // Return root of possible klass
1705 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1706 }
1707 }
1708
1709 // Check for loading klass from an array
1710 const TypeAryPtr *tary = tp->isa_aryptr();
1711 if( tary != NULL ) {
1712 ciKlass *tary_klass = tary->klass();
1713 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1714 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1715 if (tary->klass_is_exact()) {
1716 return TypeKlassPtr::make(tary_klass);
1717 }
1718 ciArrayKlass *ak = tary->klass()->as_array_klass();
1719 // If the klass is an object array, we defer the question to the
1720 // array component klass.
1721 if( ak->is_obj_array_klass() ) {
1722 assert( ak->is_loaded(), "" );
1723 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1724 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1725 ciInstanceKlass* ik = base_k->as_instance_klass();
1726 // See if we can become precise: no subklasses and no interface
1727 if (!ik->is_interface() && !ik->has_subklass()) {
1728 //assert(!UseExactTypes, "this code should be useless with exact types");
1729 // Add a dependence; if any subclass added we need to recompile
1730 if (!ik->is_final()) {
1731 phase->C->dependencies()->assert_leaf_type(ik);
1732 }
1733 // Return precise array klass
1734 return TypeKlassPtr::make(ak);
1735 }
1736 }
1737 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1738 } else { // Found a type-array?
1739 //assert(!UseExactTypes, "this code should be useless with exact types");
1740 assert( ak->is_type_array_klass(), "" );
1741 return TypeKlassPtr::make(ak); // These are always precise
1742 }
1743 }
1744 }
1745
1746 // Check for loading klass from an array klass
1747 const TypeKlassPtr *tkls = tp->isa_klassptr();
1748 if (tkls != NULL && !StressReflectiveCode) {
1749 ciKlass* klass = tkls->klass();
1750 if( !klass->is_loaded() )
1751 return _type; // Bail out if not loaded
1752 if( klass->is_obj_array_klass() &&
1753 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
1754 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
1755 // // Always returning precise element type is incorrect,
1756 // // e.g., element type could be object and array may contain strings
1757 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
1758
1759 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
1760 // according to the element type's subclassing.
1761 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
1762 }
1763 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
1764 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
1765 ciKlass* sup = klass->as_instance_klass()->super();
1766 // The field is Klass::_super. Return its (constant) value.
1767 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
1768 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
1769 }
1770 }
1771
1772 // Bailout case
1773 return LoadNode::Value(phase);
1774 }
1775
1776 //------------------------------Identity---------------------------------------
1777 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
1778 // Also feed through the klass in Allocate(...klass...)._klass.
1779 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
1780 return klass_identity_common(phase);
1781 }
1782
1783 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
1784 Node* x = LoadNode::Identity(phase);
1785 if (x != this) return x;
1786
1787 // Take apart the address into an oop and and offset.
1788 // Return 'this' if we cannot.
1789 Node* adr = in(MemNode::Address);
1790 intptr_t offset = 0;
1791 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1792 if (base == NULL) return this;
1793 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
1794 if (toop == NULL) return this;
1795
1796 // We can fetch the klass directly through an AllocateNode.
1797 // This works even if the klass is not constant (clone or newArray).
1798 if (offset == oopDesc::klass_offset_in_bytes()) {
1799 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
1800 if (allocated_klass != NULL) {
1801 return allocated_klass;
1802 }
1803 }
1804
1805 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
1806 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
1807 // See inline_native_Class_query for occurrences of these patterns.
1808 // Java Example: x.getClass().isAssignableFrom(y)
1809 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
1810 //
1811 // This improves reflective code, often making the Class
1812 // mirror go completely dead. (Current exception: Class
1813 // mirrors may appear in debug info, but we could clean them out by
1814 // introducing a new debug info operator for klassOop.java_mirror).
1815 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
1816 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1817 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1818 // We are loading a special hidden field from a Class mirror,
1819 // the field which points to its Klass or arrayKlass metaobject.
1820 if (base->is_Load()) {
1821 Node* adr2 = base->in(MemNode::Address);
1822 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1823 if (tkls != NULL && !tkls->empty()
1824 && (tkls->klass()->is_instance_klass() ||
1825 tkls->klass()->is_array_klass())
1826 && adr2->is_AddP()
1827 ) {
1828 int mirror_field = Klass::java_mirror_offset_in_bytes();
1829 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1830 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
1831 }
1832 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
1833 return adr2->in(AddPNode::Base);
1834 }
1835 }
1836 }
1837 }
1838
1839 return this;
1840 }
1841
1842
1843 //------------------------------Value------------------------------------------
1844 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
1845 const Type *t = klass_value_common(phase);
1846
1847 if (t == TypePtr::NULL_PTR) {
1848 return TypeNarrowOop::NULL_PTR;
1849 }
1850 if (t != Type::TOP && !t->isa_narrowoop()) {
1851 assert(t->is_oopptr(), "sanity");
1852 t = t->is_oopptr()->make_narrowoop();
1853 }
1854 return t;
1855 }
1856
1857 //------------------------------Identity---------------------------------------
1858 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
1859 // Also feed through the klass in Allocate(...klass...)._klass.
1860 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
1861 Node *x = klass_identity_common(phase);
1862
1863 const Type *t = phase->type( x );
1864 if( t == Type::TOP ) return x;
1865 if( t->isa_narrowoop()) return x;
1866
1867 return EncodePNode::encode(phase, x);
1868 }
1869
1870 //------------------------------Value-----------------------------------------
1871 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
1872 // Either input is TOP ==> the result is TOP
1873 const Type *t1 = phase->type( in(MemNode::Memory) );
1874 if( t1 == Type::TOP ) return Type::TOP;
1875 Node *adr = in(MemNode::Address);
1876 const Type *t2 = phase->type( adr );
1877 if( t2 == Type::TOP ) return Type::TOP;
1878 const TypePtr *tp = t2->is_ptr();
1879 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
1880 const TypeAryPtr *tap = tp->isa_aryptr();
1881 if( !tap ) return _type;
1882 return tap->size();
1883 }
1884
1885 //------------------------------Identity---------------------------------------
1886 // Feed through the length in AllocateArray(...length...)._length.
1887 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
1888 Node* x = LoadINode::Identity(phase);
1889 if (x != this) return x;
1890
1891 // Take apart the address into an oop and and offset.
1892 // Return 'this' if we cannot.
1893 Node* adr = in(MemNode::Address);
1894 intptr_t offset = 0;
1895 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1896 if (base == NULL) return this;
1897 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
1898 if (tary == NULL) return this;
1899
1900 // We can fetch the length directly through an AllocateArrayNode.
1901 // This works even if the length is not constant (clone or newArray).
1902 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1903 Node* allocated_length = AllocateArrayNode::Ideal_length(base, phase);
1904 if (allocated_length != NULL) {
1905 return allocated_length;
1906 }
1907 }
1908
1909 return this;
1910
1911 }
1912 //=============================================================================
1913 //---------------------------StoreNode::make-----------------------------------
1914 // Polymorphic factory method:
1915 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
1916 Compile* C = gvn.C;
1917
1918 switch (bt) {
1919 case T_BOOLEAN:
1920 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
1921 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
1922 case T_CHAR:
1923 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
1924 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
1925 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
1926 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
1927 case T_ADDRESS:
1928 case T_OBJECT:
1929 #ifdef _LP64
1930 if (adr->bottom_type()->is_ptr_to_narrowoop() ||
1931 (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
1932 adr->bottom_type()->isa_rawptr())) {
1933 const TypePtr* type = val->bottom_type()->is_ptr();
1934 Node* cp = EncodePNode::encode(&gvn, val);
1935 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, cp);
1936 } else
1937 #endif
1938 {
1939 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
1940 }
1941 }
1942 ShouldNotReachHere();
1943 return (StoreNode*)NULL;
1944 }
1945
1946 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
1947 bool require_atomic = true;
1948 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
1949 }
1950
1951
1952 //--------------------------bottom_type----------------------------------------
1953 const Type *StoreNode::bottom_type() const {
1954 return Type::MEMORY;
1955 }
1956
1957 //------------------------------hash-------------------------------------------
1958 uint StoreNode::hash() const {
1959 // unroll addition of interesting fields
1960 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
1961
1962 // Since they are not commoned, do not hash them:
1963 return NO_HASH;
1964 }
1965
1966 //------------------------------Ideal------------------------------------------
1967 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
1968 // When a store immediately follows a relevant allocation/initialization,
1969 // try to capture it into the initialization, or hoist it above.
1970 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1971 Node* p = MemNode::Ideal_common(phase, can_reshape);
1972 if (p) return (p == NodeSentinel) ? NULL : p;
1973
1974 Node* mem = in(MemNode::Memory);
1975 Node* address = in(MemNode::Address);
1976
1977 // Back-to-back stores to same address? Fold em up.
1978 // Generally unsafe if I have intervening uses...
1979 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
1980 // Looking at a dead closed cycle of memory?
1981 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
1982
1983 assert(Opcode() == mem->Opcode() ||
1984 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
1985 "no mismatched stores, except on raw memory");
1986
1987 if (mem->outcnt() == 1 && // check for intervening uses
1988 mem->as_Store()->memory_size() <= this->memory_size()) {
1989 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
1990 // For example, 'mem' might be the final state at a conditional return.
1991 // Or, 'mem' might be used by some node which is live at the same time
1992 // 'this' is live, which might be unschedulable. So, require exactly
1993 // ONE user, the 'this' store, until such time as we clone 'mem' for
1994 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
1995 if (can_reshape) { // (%%% is this an anachronism?)
1996 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
1997 phase->is_IterGVN());
1998 } else {
1999 // It's OK to do this in the parser, since DU info is always accurate,
2000 // and the parser always refers to nodes via SafePointNode maps.
2001 set_req(MemNode::Memory, mem->in(MemNode::Memory));
2002 }
2003 return this;
2004 }
2005 }
2006
2007 // Capture an unaliased, unconditional, simple store into an initializer.
2008 // Or, if it is independent of the allocation, hoist it above the allocation.
2009 if (ReduceFieldZeroing && /*can_reshape &&*/
2010 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2011 InitializeNode* init = mem->in(0)->as_Initialize();
2012 intptr_t offset = init->can_capture_store(this, phase);
2013 if (offset > 0) {
2014 Node* moved = init->capture_store(this, offset, phase);
2015 // If the InitializeNode captured me, it made a raw copy of me,
2016 // and I need to disappear.
2017 if (moved != NULL) {
2018 // %%% hack to ensure that Ideal returns a new node:
2019 mem = MergeMemNode::make(phase->C, mem);
2020 return mem; // fold me away
2021 }
2022 }
2023 }
2024
2025 return NULL; // No further progress
2026 }
2027
2028 //------------------------------Value-----------------------------------------
2029 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2030 // Either input is TOP ==> the result is TOP
2031 const Type *t1 = phase->type( in(MemNode::Memory) );
2032 if( t1 == Type::TOP ) return Type::TOP;
2033 const Type *t2 = phase->type( in(MemNode::Address) );
2034 if( t2 == Type::TOP ) return Type::TOP;
2035 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2036 if( t3 == Type::TOP ) return Type::TOP;
2037 return Type::MEMORY;
2038 }
2039
2040 //------------------------------Identity---------------------------------------
2041 // Remove redundant stores:
2042 // Store(m, p, Load(m, p)) changes to m.
2043 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2044 Node *StoreNode::Identity( PhaseTransform *phase ) {
2045 Node* mem = in(MemNode::Memory);
2046 Node* adr = in(MemNode::Address);
2047 Node* val = in(MemNode::ValueIn);
2048
2049 // Load then Store? Then the Store is useless
2050 if (val->is_Load() &&
2051 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
2052 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
2053 val->as_Load()->store_Opcode() == Opcode()) {
2054 return mem;
2055 }
2056
2057 // Two stores in a row of the same value?
2058 if (mem->is_Store() &&
2059 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
2060 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
2061 mem->Opcode() == Opcode()) {
2062 return mem;
2063 }
2064
2065 // Store of zero anywhere into a freshly-allocated object?
2066 // Then the store is useless.
2067 // (It must already have been captured by the InitializeNode.)
2068 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2069 // a newly allocated object is already all-zeroes everywhere
2070 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2071 return mem;
2072 }
2073
2074 // the store may also apply to zero-bits in an earlier object
2075 Node* prev_mem = find_previous_store(phase);
2076 // Steps (a), (b): Walk past independent stores to find an exact match.
2077 if (prev_mem != NULL) {
2078 Node* prev_val = can_see_stored_value(prev_mem, phase);
2079 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2080 // prev_val and val might differ by a cast; it would be good
2081 // to keep the more informative of the two.
2082 return mem;
2083 }
2084 }
2085 }
2086
2087 return this;
2088 }
2089
2090 //------------------------------match_edge-------------------------------------
2091 // Do we Match on this edge index or not? Match only memory & value
2092 uint StoreNode::match_edge(uint idx) const {
2093 return idx == MemNode::Address || idx == MemNode::ValueIn;
2094 }
2095
2096 //------------------------------cmp--------------------------------------------
2097 // Do not common stores up together. They generally have to be split
2098 // back up anyways, so do not bother.
2099 uint StoreNode::cmp( const Node &n ) const {
2100 return (&n == this); // Always fail except on self
2101 }
2102
2103 //------------------------------Ideal_masked_input-----------------------------
2104 // Check for a useless mask before a partial-word store
2105 // (StoreB ... (AndI valIn conIa) )
2106 // If (conIa & mask == mask) this simplifies to
2107 // (StoreB ... (valIn) )
2108 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2109 Node *val = in(MemNode::ValueIn);
2110 if( val->Opcode() == Op_AndI ) {
2111 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2112 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2113 set_req(MemNode::ValueIn, val->in(1));
2114 return this;
2115 }
2116 }
2117 return NULL;
2118 }
2119
2120
2121 //------------------------------Ideal_sign_extended_input----------------------
2122 // Check for useless sign-extension before a partial-word store
2123 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2124 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2125 // (StoreB ... (valIn) )
2126 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2127 Node *val = in(MemNode::ValueIn);
2128 if( val->Opcode() == Op_RShiftI ) {
2129 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2130 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2131 Node *shl = val->in(1);
2132 if( shl->Opcode() == Op_LShiftI ) {
2133 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2134 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2135 set_req(MemNode::ValueIn, shl->in(1));
2136 return this;
2137 }
2138 }
2139 }
2140 }
2141 return NULL;
2142 }
2143
2144 //------------------------------value_never_loaded-----------------------------------
2145 // Determine whether there are any possible loads of the value stored.
2146 // For simplicity, we actually check if there are any loads from the
2147 // address stored to, not just for loads of the value stored by this node.
2148 //
2149 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2150 Node *adr = in(Address);
2151 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2152 if (adr_oop == NULL)
2153 return false;
2154 if (!adr_oop->is_instance_field())
2155 return false; // if not a distinct instance, there may be aliases of the address
2156 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2157 Node *use = adr->fast_out(i);
2158 int opc = use->Opcode();
2159 if (use->is_Load() || use->is_LoadStore()) {
2160 return false;
2161 }
2162 }
2163 return true;
2164 }
2165
2166 //=============================================================================
2167 //------------------------------Ideal------------------------------------------
2168 // If the store is from an AND mask that leaves the low bits untouched, then
2169 // we can skip the AND operation. If the store is from a sign-extension
2170 // (a left shift, then right shift) we can skip both.
2171 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2172 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2173 if( progress != NULL ) return progress;
2174
2175 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2176 if( progress != NULL ) return progress;
2177
2178 // Finally check the default case
2179 return StoreNode::Ideal(phase, can_reshape);
2180 }
2181
2182 //=============================================================================
2183 //------------------------------Ideal------------------------------------------
2184 // If the store is from an AND mask that leaves the low bits untouched, then
2185 // we can skip the AND operation
2186 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2187 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2188 if( progress != NULL ) return progress;
2189
2190 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2191 if( progress != NULL ) return progress;
2192
2193 // Finally check the default case
2194 return StoreNode::Ideal(phase, can_reshape);
2195 }
2196
2197 //=============================================================================
2198 //------------------------------Identity---------------------------------------
2199 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2200 // No need to card mark when storing a null ptr
2201 Node* my_store = in(MemNode::OopStore);
2202 if (my_store->is_Store()) {
2203 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2204 if( t1 == TypePtr::NULL_PTR ) {
2205 return in(MemNode::Memory);
2206 }
2207 }
2208 return this;
2209 }
2210
2211 //------------------------------Value-----------------------------------------
2212 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2213 // Either input is TOP ==> the result is TOP
2214 const Type *t = phase->type( in(MemNode::Memory) );
2215 if( t == Type::TOP ) return Type::TOP;
2216 t = phase->type( in(MemNode::Address) );
2217 if( t == Type::TOP ) return Type::TOP;
2218 t = phase->type( in(MemNode::ValueIn) );
2219 if( t == Type::TOP ) return Type::TOP;
2220 // If extra input is TOP ==> the result is TOP
2221 t = phase->type( in(MemNode::OopStore) );
2222 if( t == Type::TOP ) return Type::TOP;
2223
2224 return StoreNode::Value( phase );
2225 }
2226
2227
2228 //=============================================================================
2229 //----------------------------------SCMemProjNode------------------------------
2230 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2231 {
2232 return bottom_type();
2233 }
2234
2235 //=============================================================================
2236 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
2237 init_req(MemNode::Control, c );
2238 init_req(MemNode::Memory , mem);
2239 init_req(MemNode::Address, adr);
2240 init_req(MemNode::ValueIn, val);
2241 init_req( ExpectedIn, ex );
2242 init_class_id(Class_LoadStore);
2243
2244 }
2245
2246 //=============================================================================
2247 //-------------------------------adr_type--------------------------------------
2248 // Do we Match on this edge index or not? Do not match memory
2249 const TypePtr* ClearArrayNode::adr_type() const {
2250 Node *adr = in(3);
2251 return MemNode::calculate_adr_type(adr->bottom_type());
2252 }
2253
2254 //------------------------------match_edge-------------------------------------
2255 // Do we Match on this edge index or not? Do not match memory
2256 uint ClearArrayNode::match_edge(uint idx) const {
2257 return idx > 1;
2258 }
2259
2260 //------------------------------Identity---------------------------------------
2261 // Clearing a zero length array does nothing
2262 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2263 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2264 }
2265
2266 //------------------------------Idealize---------------------------------------
2267 // Clearing a short array is faster with stores
2268 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2269 const int unit = BytesPerLong;
2270 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2271 if (!t) return NULL;
2272 if (!t->is_con()) return NULL;
2273 intptr_t raw_count = t->get_con();
2274 intptr_t size = raw_count;
2275 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2276 // Clearing nothing uses the Identity call.
2277 // Negative clears are possible on dead ClearArrays
2278 // (see jck test stmt114.stmt11402.val).
2279 if (size <= 0 || size % unit != 0) return NULL;
2280 intptr_t count = size / unit;
2281 // Length too long; use fast hardware clear
2282 if (size > Matcher::init_array_short_size) return NULL;
2283 Node *mem = in(1);
2284 if( phase->type(mem)==Type::TOP ) return NULL;
2285 Node *adr = in(3);
2286 const Type* at = phase->type(adr);
2287 if( at==Type::TOP ) return NULL;
2288 const TypePtr* atp = at->isa_ptr();
2289 // adjust atp to be the correct array element address type
2290 if (atp == NULL) atp = TypePtr::BOTTOM;
2291 else atp = atp->add_offset(Type::OffsetBot);
2292 // Get base for derived pointer purposes
2293 if( adr->Opcode() != Op_AddP ) Unimplemented();
2294 Node *base = adr->in(1);
2295
2296 Node *zero = phase->makecon(TypeLong::ZERO);
2297 Node *off = phase->MakeConX(BytesPerLong);
2298 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2299 count--;
2300 while( count-- ) {
2301 mem = phase->transform(mem);
2302 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
2303 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2304 }
2305 return mem;
2306 }
2307
2308 //----------------------------clear_memory-------------------------------------
2309 // Generate code to initialize object storage to zero.
2310 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2311 intptr_t start_offset,
2312 Node* end_offset,
2313 PhaseGVN* phase) {
2314 Compile* C = phase->C;
2315 intptr_t offset = start_offset;
2316
2317 int unit = BytesPerLong;
2318 if ((offset % unit) != 0) {
2319 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
2320 adr = phase->transform(adr);
2321 const TypePtr* atp = TypeRawPtr::BOTTOM;
2322 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2323 mem = phase->transform(mem);
2324 offset += BytesPerInt;
2325 }
2326 assert((offset % unit) == 0, "");
2327
2328 // Initialize the remaining stuff, if any, with a ClearArray.
2329 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2330 }
2331
2332 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2333 Node* start_offset,
2334 Node* end_offset,
2335 PhaseGVN* phase) {
2336 if (start_offset == end_offset) {
2337 // nothing to do
2338 return mem;
2339 }
2340
2341 Compile* C = phase->C;
2342 int unit = BytesPerLong;
2343 Node* zbase = start_offset;
2344 Node* zend = end_offset;
2345
2346 // Scale to the unit required by the CPU:
2347 if (!Matcher::init_array_count_is_in_bytes) {
2348 Node* shift = phase->intcon(exact_log2(unit));
2349 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
2350 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
2351 }
2352
2353 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
2354 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
2355
2356 // Bulk clear double-words
2357 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
2358 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
2359 return phase->transform(mem);
2360 }
2361
2362 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2363 intptr_t start_offset,
2364 intptr_t end_offset,
2365 PhaseGVN* phase) {
2366 if (start_offset == end_offset) {
2367 // nothing to do
2368 return mem;
2369 }
2370
2371 Compile* C = phase->C;
2372 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2373 intptr_t done_offset = end_offset;
2374 if ((done_offset % BytesPerLong) != 0) {
2375 done_offset -= BytesPerInt;
2376 }
2377 if (done_offset > start_offset) {
2378 mem = clear_memory(ctl, mem, dest,
2379 start_offset, phase->MakeConX(done_offset), phase);
2380 }
2381 if (done_offset < end_offset) { // emit the final 32-bit store
2382 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
2383 adr = phase->transform(adr);
2384 const TypePtr* atp = TypeRawPtr::BOTTOM;
2385 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2386 mem = phase->transform(mem);
2387 done_offset += BytesPerInt;
2388 }
2389 assert(done_offset == end_offset, "");
2390 return mem;
2391 }
2392
2393 //=============================================================================
2394 // Do we match on this edge? No memory edges
2395 uint StrCompNode::match_edge(uint idx) const {
2396 return idx == 5 || idx == 6;
2397 }
2398
2399 //------------------------------Ideal------------------------------------------
2400 // Return a node which is more "ideal" than the current node. Strip out
2401 // control copies
2402 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
2403 return remove_dead_region(phase, can_reshape) ? this : NULL;
2404 }
2405
2406 //------------------------------Ideal------------------------------------------
2407 // Return a node which is more "ideal" than the current node. Strip out
2408 // control copies
2409 Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){
2410 return remove_dead_region(phase, can_reshape) ? this : NULL;
2411 }
2412
2413
2414 //=============================================================================
2415 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2416 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2417 _adr_type(C->get_adr_type(alias_idx))
2418 {
2419 init_class_id(Class_MemBar);
2420 Node* top = C->top();
2421 init_req(TypeFunc::I_O,top);
2422 init_req(TypeFunc::FramePtr,top);
2423 init_req(TypeFunc::ReturnAdr,top);
2424 if (precedent != NULL)
2425 init_req(TypeFunc::Parms, precedent);
2426 }
2427
2428 //------------------------------cmp--------------------------------------------
2429 uint MemBarNode::hash() const { return NO_HASH; }
2430 uint MemBarNode::cmp( const Node &n ) const {
2431 return (&n == this); // Always fail except on self
2432 }
2433
2434 //------------------------------make-------------------------------------------
2435 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2436 int len = Precedent + (pn == NULL? 0: 1);
2437 switch (opcode) {
2438 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
2439 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
2440 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
2441 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
2442 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
2443 default: ShouldNotReachHere(); return NULL;
2444 }
2445 }
2446
2447 //------------------------------Ideal------------------------------------------
2448 // Return a node which is more "ideal" than the current node. Strip out
2449 // control copies
2450 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2451 if (remove_dead_region(phase, can_reshape)) return this;
2452 return NULL;
2453 }
2454
2455 //------------------------------Value------------------------------------------
2456 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2457 if( !in(0) ) return Type::TOP;
2458 if( phase->type(in(0)) == Type::TOP )
2459 return Type::TOP;
2460 return TypeTuple::MEMBAR;
2461 }
2462
2463 //------------------------------match------------------------------------------
2464 // Construct projections for memory.
2465 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2466 switch (proj->_con) {
2467 case TypeFunc::Control:
2468 case TypeFunc::Memory:
2469 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2470 }
2471 ShouldNotReachHere();
2472 return NULL;
2473 }
2474
2475 //===========================InitializeNode====================================
2476 // SUMMARY:
2477 // This node acts as a memory barrier on raw memory, after some raw stores.
2478 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2479 // The Initialize can 'capture' suitably constrained stores as raw inits.
2480 // It can coalesce related raw stores into larger units (called 'tiles').
2481 // It can avoid zeroing new storage for memory units which have raw inits.
2482 // At macro-expansion, it is marked 'complete', and does not optimize further.
2483 //
2484 // EXAMPLE:
2485 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2486 // ctl = incoming control; mem* = incoming memory
2487 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
2488 // First allocate uninitialized memory and fill in the header:
2489 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
2490 // ctl := alloc.Control; mem* := alloc.Memory*
2491 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
2492 // Then initialize to zero the non-header parts of the raw memory block:
2493 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2494 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2495 // After the initialize node executes, the object is ready for service:
2496 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2497 // Suppose its body is immediately initialized as {1,2}:
2498 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2499 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2500 // mem.SLICE(#short[*]) := store2
2501 //
2502 // DETAILS:
2503 // An InitializeNode collects and isolates object initialization after
2504 // an AllocateNode and before the next possible safepoint. As a
2505 // memory barrier (MemBarNode), it keeps critical stores from drifting
2506 // down past any safepoint or any publication of the allocation.
2507 // Before this barrier, a newly-allocated object may have uninitialized bits.
2508 // After this barrier, it may be treated as a real oop, and GC is allowed.
2509 //
2510 // The semantics of the InitializeNode include an implicit zeroing of
2511 // the new object from object header to the end of the object.
2512 // (The object header and end are determined by the AllocateNode.)
2513 //
2514 // Certain stores may be added as direct inputs to the InitializeNode.
2515 // These stores must update raw memory, and they must be to addresses
2516 // derived from the raw address produced by AllocateNode, and with
2517 // a constant offset. They must be ordered by increasing offset.
2518 // The first one is at in(RawStores), the last at in(req()-1).
2519 // Unlike most memory operations, they are not linked in a chain,
2520 // but are displayed in parallel as users of the rawmem output of
2521 // the allocation.
2522 //
2523 // (See comments in InitializeNode::capture_store, which continue
2524 // the example given above.)
2525 //
2526 // When the associated Allocate is macro-expanded, the InitializeNode
2527 // may be rewritten to optimize collected stores. A ClearArrayNode
2528 // may also be created at that point to represent any required zeroing.
2529 // The InitializeNode is then marked 'complete', prohibiting further
2530 // capturing of nearby memory operations.
2531 //
2532 // During macro-expansion, all captured initializations which store
2533 // constant values of 32 bits or smaller are coalesced (if advantagous)
2534 // into larger 'tiles' 32 or 64 bits. This allows an object to be
2535 // initialized in fewer memory operations. Memory words which are
2536 // covered by neither tiles nor non-constant stores are pre-zeroed
2537 // by explicit stores of zero. (The code shape happens to do all
2538 // zeroing first, then all other stores, with both sequences occurring
2539 // in order of ascending offsets.)
2540 //
2541 // Alternatively, code may be inserted between an AllocateNode and its
2542 // InitializeNode, to perform arbitrary initialization of the new object.
2543 // E.g., the object copying intrinsics insert complex data transfers here.
2544 // The initialization must then be marked as 'complete' disable the
2545 // built-in zeroing semantics and the collection of initializing stores.
2546 //
2547 // While an InitializeNode is incomplete, reads from the memory state
2548 // produced by it are optimizable if they match the control edge and
2549 // new oop address associated with the allocation/initialization.
2550 // They return a stored value (if the offset matches) or else zero.
2551 // A write to the memory state, if it matches control and address,
2552 // and if it is to a constant offset, may be 'captured' by the
2553 // InitializeNode. It is cloned as a raw memory operation and rewired
2554 // inside the initialization, to the raw oop produced by the allocation.
2555 // Operations on addresses which are provably distinct (e.g., to
2556 // other AllocateNodes) are allowed to bypass the initialization.
2557 //
2558 // The effect of all this is to consolidate object initialization
2559 // (both arrays and non-arrays, both piecewise and bulk) into a
2560 // single location, where it can be optimized as a unit.
2561 //
2562 // Only stores with an offset less than TrackedInitializationLimit words
2563 // will be considered for capture by an InitializeNode. This puts a
2564 // reasonable limit on the complexity of optimized initializations.
2565
2566 //---------------------------InitializeNode------------------------------------
2567 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
2568 : _is_complete(false),
2569 MemBarNode(C, adr_type, rawoop)
2570 {
2571 init_class_id(Class_Initialize);
2572
2573 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
2574 assert(in(RawAddress) == rawoop, "proper init");
2575 // Note: allocation() can be NULL, for secondary initialization barriers
2576 }
2577
2578 // Since this node is not matched, it will be processed by the
2579 // register allocator. Declare that there are no constraints
2580 // on the allocation of the RawAddress edge.
2581 const RegMask &InitializeNode::in_RegMask(uint idx) const {
2582 // This edge should be set to top, by the set_complete. But be conservative.
2583 if (idx == InitializeNode::RawAddress)
2584 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
2585 return RegMask::Empty;
2586 }
2587
2588 Node* InitializeNode::memory(uint alias_idx) {
2589 Node* mem = in(Memory);
2590 if (mem->is_MergeMem()) {
2591 return mem->as_MergeMem()->memory_at(alias_idx);
2592 } else {
2593 // incoming raw memory is not split
2594 return mem;
2595 }
2596 }
2597
2598 bool InitializeNode::is_non_zero() {
2599 if (is_complete()) return false;
2600 remove_extra_zeroes();
2601 return (req() > RawStores);
2602 }
2603
2604 void InitializeNode::set_complete(PhaseGVN* phase) {
2605 assert(!is_complete(), "caller responsibility");
2606 _is_complete = true;
2607
2608 // After this node is complete, it contains a bunch of
2609 // raw-memory initializations. There is no need for
2610 // it to have anything to do with non-raw memory effects.
2611 // Therefore, tell all non-raw users to re-optimize themselves,
2612 // after skipping the memory effects of this initialization.
2613 PhaseIterGVN* igvn = phase->is_IterGVN();
2614 if (igvn) igvn->add_users_to_worklist(this);
2615 }
2616
2617 // convenience function
2618 // return false if the init contains any stores already
2619 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
2620 InitializeNode* init = initialization();
2621 if (init == NULL || init->is_complete()) return false;
2622 init->remove_extra_zeroes();
2623 // for now, if this allocation has already collected any inits, bail:
2624 if (init->is_non_zero()) return false;
2625 init->set_complete(phase);
2626 return true;
2627 }
2628
2629 void InitializeNode::remove_extra_zeroes() {
2630 if (req() == RawStores) return;
2631 Node* zmem = zero_memory();
2632 uint fill = RawStores;
2633 for (uint i = fill; i < req(); i++) {
2634 Node* n = in(i);
2635 if (n->is_top() || n == zmem) continue; // skip
2636 if (fill < i) set_req(fill, n); // compact
2637 ++fill;
2638 }
2639 // delete any empty spaces created:
2640 while (fill < req()) {
2641 del_req(fill);
2642 }
2643 }
2644
2645 // Helper for remembering which stores go with which offsets.
2646 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
2647 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
2648 intptr_t offset = -1;
2649 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
2650 phase, offset);
2651 if (base == NULL) return -1; // something is dead,
2652 if (offset < 0) return -1; // dead, dead
2653 return offset;
2654 }
2655
2656 // Helper for proving that an initialization expression is
2657 // "simple enough" to be folded into an object initialization.
2658 // Attempts to prove that a store's initial value 'n' can be captured
2659 // within the initialization without creating a vicious cycle, such as:
2660 // { Foo p = new Foo(); p.next = p; }
2661 // True for constants and parameters and small combinations thereof.
2662 bool InitializeNode::detect_init_independence(Node* n,
2663 bool st_is_pinned,
2664 int& count) {
2665 if (n == NULL) return true; // (can this really happen?)
2666 if (n->is_Proj()) n = n->in(0);
2667 if (n == this) return false; // found a cycle
2668 if (n->is_Con()) return true;
2669 if (n->is_Start()) return true; // params, etc., are OK
2670 if (n->is_Root()) return true; // even better
2671
2672 Node* ctl = n->in(0);
2673 if (ctl != NULL && !ctl->is_top()) {
2674 if (ctl->is_Proj()) ctl = ctl->in(0);
2675 if (ctl == this) return false;
2676
2677 // If we already know that the enclosing memory op is pinned right after
2678 // the init, then any control flow that the store has picked up
2679 // must have preceded the init, or else be equal to the init.
2680 // Even after loop optimizations (which might change control edges)
2681 // a store is never pinned *before* the availability of its inputs.
2682 if (!MemNode::all_controls_dominate(n, this))
2683 return false; // failed to prove a good control
2684
2685 }
2686
2687 // Check data edges for possible dependencies on 'this'.
2688 if ((count += 1) > 20) return false; // complexity limit
2689 for (uint i = 1; i < n->req(); i++) {
2690 Node* m = n->in(i);
2691 if (m == NULL || m == n || m->is_top()) continue;
2692 uint first_i = n->find_edge(m);
2693 if (i != first_i) continue; // process duplicate edge just once
2694 if (!detect_init_independence(m, st_is_pinned, count)) {
2695 return false;
2696 }
2697 }
2698
2699 return true;
2700 }
2701
2702 // Here are all the checks a Store must pass before it can be moved into
2703 // an initialization. Returns zero if a check fails.
2704 // On success, returns the (constant) offset to which the store applies,
2705 // within the initialized memory.
2706 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
2707 const int FAIL = 0;
2708 if (st->req() != MemNode::ValueIn + 1)
2709 return FAIL; // an inscrutable StoreNode (card mark?)
2710 Node* ctl = st->in(MemNode::Control);
2711 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
2712 return FAIL; // must be unconditional after the initialization
2713 Node* mem = st->in(MemNode::Memory);
2714 if (!(mem->is_Proj() && mem->in(0) == this))
2715 return FAIL; // must not be preceded by other stores
2716 Node* adr = st->in(MemNode::Address);
2717 intptr_t offset;
2718 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
2719 if (alloc == NULL)
2720 return FAIL; // inscrutable address
2721 if (alloc != allocation())
2722 return FAIL; // wrong allocation! (store needs to float up)
2723 Node* val = st->in(MemNode::ValueIn);
2724 int complexity_count = 0;
2725 if (!detect_init_independence(val, true, complexity_count))
2726 return FAIL; // stored value must be 'simple enough'
2727
2728 return offset; // success
2729 }
2730
2731 // Find the captured store in(i) which corresponds to the range
2732 // [start..start+size) in the initialized object.
2733 // If there is one, return its index i. If there isn't, return the
2734 // negative of the index where it should be inserted.
2735 // Return 0 if the queried range overlaps an initialization boundary
2736 // or if dead code is encountered.
2737 // If size_in_bytes is zero, do not bother with overlap checks.
2738 int InitializeNode::captured_store_insertion_point(intptr_t start,
2739 int size_in_bytes,
2740 PhaseTransform* phase) {
2741 const int FAIL = 0, MAX_STORE = BytesPerLong;
2742
2743 if (is_complete())
2744 return FAIL; // arraycopy got here first; punt
2745
2746 assert(allocation() != NULL, "must be present");
2747
2748 // no negatives, no header fields:
2749 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
2750
2751 // after a certain size, we bail out on tracking all the stores:
2752 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2753 if (start >= ti_limit) return FAIL;
2754
2755 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
2756 if (i >= limit) return -(int)i; // not found; here is where to put it
2757
2758 Node* st = in(i);
2759 intptr_t st_off = get_store_offset(st, phase);
2760 if (st_off < 0) {
2761 if (st != zero_memory()) {
2762 return FAIL; // bail out if there is dead garbage
2763 }
2764 } else if (st_off > start) {
2765 // ...we are done, since stores are ordered
2766 if (st_off < start + size_in_bytes) {
2767 return FAIL; // the next store overlaps
2768 }
2769 return -(int)i; // not found; here is where to put it
2770 } else if (st_off < start) {
2771 if (size_in_bytes != 0 &&
2772 start < st_off + MAX_STORE &&
2773 start < st_off + st->as_Store()->memory_size()) {
2774 return FAIL; // the previous store overlaps
2775 }
2776 } else {
2777 if (size_in_bytes != 0 &&
2778 st->as_Store()->memory_size() != size_in_bytes) {
2779 return FAIL; // mismatched store size
2780 }
2781 return i;
2782 }
2783
2784 ++i;
2785 }
2786 }
2787
2788 // Look for a captured store which initializes at the offset 'start'
2789 // with the given size. If there is no such store, and no other
2790 // initialization interferes, then return zero_memory (the memory
2791 // projection of the AllocateNode).
2792 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
2793 PhaseTransform* phase) {
2794 assert(stores_are_sane(phase), "");
2795 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2796 if (i == 0) {
2797 return NULL; // something is dead
2798 } else if (i < 0) {
2799 return zero_memory(); // just primordial zero bits here
2800 } else {
2801 Node* st = in(i); // here is the store at this position
2802 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
2803 return st;
2804 }
2805 }
2806
2807 // Create, as a raw pointer, an address within my new object at 'offset'.
2808 Node* InitializeNode::make_raw_address(intptr_t offset,
2809 PhaseTransform* phase) {
2810 Node* addr = in(RawAddress);
2811 if (offset != 0) {
2812 Compile* C = phase->C;
2813 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
2814 phase->MakeConX(offset)) );
2815 }
2816 return addr;
2817 }
2818
2819 // Clone the given store, converting it into a raw store
2820 // initializing a field or element of my new object.
2821 // Caller is responsible for retiring the original store,
2822 // with subsume_node or the like.
2823 //
2824 // From the example above InitializeNode::InitializeNode,
2825 // here are the old stores to be captured:
2826 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2827 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2828 //
2829 // Here is the changed code; note the extra edges on init:
2830 // alloc = (Allocate ...)
2831 // rawoop = alloc.RawAddress
2832 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
2833 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
2834 // init = (Initialize alloc.Control alloc.Memory rawoop
2835 // rawstore1 rawstore2)
2836 //
2837 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
2838 PhaseTransform* phase) {
2839 assert(stores_are_sane(phase), "");
2840
2841 if (start < 0) return NULL;
2842 assert(can_capture_store(st, phase) == start, "sanity");
2843
2844 Compile* C = phase->C;
2845 int size_in_bytes = st->memory_size();
2846 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2847 if (i == 0) return NULL; // bail out
2848 Node* prev_mem = NULL; // raw memory for the captured store
2849 if (i > 0) {
2850 prev_mem = in(i); // there is a pre-existing store under this one
2851 set_req(i, C->top()); // temporarily disconnect it
2852 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
2853 } else {
2854 i = -i; // no pre-existing store
2855 prev_mem = zero_memory(); // a slice of the newly allocated object
2856 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
2857 set_req(--i, C->top()); // reuse this edge; it has been folded away
2858 else
2859 ins_req(i, C->top()); // build a new edge
2860 }
2861 Node* new_st = st->clone();
2862 new_st->set_req(MemNode::Control, in(Control));
2863 new_st->set_req(MemNode::Memory, prev_mem);
2864 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
2865 new_st = phase->transform(new_st);
2866
2867 // At this point, new_st might have swallowed a pre-existing store
2868 // at the same offset, or perhaps new_st might have disappeared,
2869 // if it redundantly stored the same value (or zero to fresh memory).
2870
2871 // In any case, wire it in:
2872 set_req(i, new_st);
2873
2874 // The caller may now kill the old guy.
2875 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
2876 assert(check_st == new_st || check_st == NULL, "must be findable");
2877 assert(!is_complete(), "");
2878 return new_st;
2879 }
2880
2881 static bool store_constant(jlong* tiles, int num_tiles,
2882 intptr_t st_off, int st_size,
2883 jlong con) {
2884 if ((st_off & (st_size-1)) != 0)
2885 return false; // strange store offset (assume size==2**N)
2886 address addr = (address)tiles + st_off;
2887 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
2888 switch (st_size) {
2889 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
2890 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
2891 case sizeof(jint): *(jint*) addr = (jint) con; break;
2892 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
2893 default: return false; // strange store size (detect size!=2**N here)
2894 }
2895 return true; // return success to caller
2896 }
2897
2898 // Coalesce subword constants into int constants and possibly
2899 // into long constants. The goal, if the CPU permits,
2900 // is to initialize the object with a small number of 64-bit tiles.
2901 // Also, convert floating-point constants to bit patterns.
2902 // Non-constants are not relevant to this pass.
2903 //
2904 // In terms of the running example on InitializeNode::InitializeNode
2905 // and InitializeNode::capture_store, here is the transformation
2906 // of rawstore1 and rawstore2 into rawstore12:
2907 // alloc = (Allocate ...)
2908 // rawoop = alloc.RawAddress
2909 // tile12 = 0x00010002
2910 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
2911 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
2912 //
2913 void
2914 InitializeNode::coalesce_subword_stores(intptr_t header_size,
2915 Node* size_in_bytes,
2916 PhaseGVN* phase) {
2917 Compile* C = phase->C;
2918
2919 assert(stores_are_sane(phase), "");
2920 // Note: After this pass, they are not completely sane,
2921 // since there may be some overlaps.
2922
2923 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
2924
2925 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2926 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
2927 size_limit = MIN2(size_limit, ti_limit);
2928 size_limit = align_size_up(size_limit, BytesPerLong);
2929 int num_tiles = size_limit / BytesPerLong;
2930
2931 // allocate space for the tile map:
2932 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
2933 jlong tiles_buf[small_len];
2934 Node* nodes_buf[small_len];
2935 jlong inits_buf[small_len];
2936 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
2937 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2938 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
2939 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
2940 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
2941 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
2942 // tiles: exact bitwise model of all primitive constants
2943 // nodes: last constant-storing node subsumed into the tiles model
2944 // inits: which bytes (in each tile) are touched by any initializations
2945
2946 //// Pass A: Fill in the tile model with any relevant stores.
2947
2948 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
2949 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
2950 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
2951 Node* zmem = zero_memory(); // initially zero memory state
2952 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
2953 Node* st = in(i);
2954 intptr_t st_off = get_store_offset(st, phase);
2955
2956 // Figure out the store's offset and constant value:
2957 if (st_off < header_size) continue; //skip (ignore header)
2958 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
2959 int st_size = st->as_Store()->memory_size();
2960 if (st_off + st_size > size_limit) break;
2961
2962 // Record which bytes are touched, whether by constant or not.
2963 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
2964 continue; // skip (strange store size)
2965
2966 const Type* val = phase->type(st->in(MemNode::ValueIn));
2967 if (!val->singleton()) continue; //skip (non-con store)
2968 BasicType type = val->basic_type();
2969
2970 jlong con = 0;
2971 switch (type) {
2972 case T_INT: con = val->is_int()->get_con(); break;
2973 case T_LONG: con = val->is_long()->get_con(); break;
2974 case T_FLOAT: con = jint_cast(val->getf()); break;
2975 case T_DOUBLE: con = jlong_cast(val->getd()); break;
2976 default: continue; //skip (odd store type)
2977 }
2978
2979 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
2980 st->Opcode() == Op_StoreL) {
2981 continue; // This StoreL is already optimal.
2982 }
2983
2984 // Store down the constant.
2985 store_constant(tiles, num_tiles, st_off, st_size, con);
2986
2987 intptr_t j = st_off >> LogBytesPerLong;
2988
2989 if (type == T_INT && st_size == BytesPerInt
2990 && (st_off & BytesPerInt) == BytesPerInt) {
2991 jlong lcon = tiles[j];
2992 if (!Matcher::isSimpleConstant64(lcon) &&
2993 st->Opcode() == Op_StoreI) {
2994 // This StoreI is already optimal by itself.
2995 jint* intcon = (jint*) &tiles[j];
2996 intcon[1] = 0; // undo the store_constant()
2997
2998 // If the previous store is also optimal by itself, back up and
2999 // undo the action of the previous loop iteration... if we can.
3000 // But if we can't, just let the previous half take care of itself.
3001 st = nodes[j];
3002 st_off -= BytesPerInt;
3003 con = intcon[0];
3004 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3005 assert(st_off >= header_size, "still ignoring header");
3006 assert(get_store_offset(st, phase) == st_off, "must be");
3007 assert(in(i-1) == zmem, "must be");
3008 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3009 assert(con == tcon->is_int()->get_con(), "must be");
3010 // Undo the effects of the previous loop trip, which swallowed st:
3011 intcon[0] = 0; // undo store_constant()
3012 set_req(i-1, st); // undo set_req(i, zmem)
3013 nodes[j] = NULL; // undo nodes[j] = st
3014 --old_subword; // undo ++old_subword
3015 }
3016 continue; // This StoreI is already optimal.
3017 }
3018 }
3019
3020 // This store is not needed.
3021 set_req(i, zmem);
3022 nodes[j] = st; // record for the moment
3023 if (st_size < BytesPerLong) // something has changed
3024 ++old_subword; // includes int/float, but who's counting...
3025 else ++old_long;
3026 }
3027
3028 if ((old_subword + old_long) == 0)
3029 return; // nothing more to do
3030
3031 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3032 // Be sure to insert them before overlapping non-constant stores.
3033 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3034 for (int j = 0; j < num_tiles; j++) {
3035 jlong con = tiles[j];
3036 jlong init = inits[j];
3037 if (con == 0) continue;
3038 jint con0, con1; // split the constant, address-wise
3039 jint init0, init1; // split the init map, address-wise
3040 { union { jlong con; jint intcon[2]; } u;
3041 u.con = con;
3042 con0 = u.intcon[0];
3043 con1 = u.intcon[1];
3044 u.con = init;
3045 init0 = u.intcon[0];
3046 init1 = u.intcon[1];
3047 }
3048
3049 Node* old = nodes[j];
3050 assert(old != NULL, "need the prior store");
3051 intptr_t offset = (j * BytesPerLong);
3052
3053 bool split = !Matcher::isSimpleConstant64(con);
3054
3055 if (offset < header_size) {
3056 assert(offset + BytesPerInt >= header_size, "second int counts");
3057 assert(*(jint*)&tiles[j] == 0, "junk in header");
3058 split = true; // only the second word counts
3059 // Example: int a[] = { 42 ... }
3060 } else if (con0 == 0 && init0 == -1) {
3061 split = true; // first word is covered by full inits
3062 // Example: int a[] = { ... foo(), 42 ... }
3063 } else if (con1 == 0 && init1 == -1) {
3064 split = true; // second word is covered by full inits
3065 // Example: int a[] = { ... 42, foo() ... }
3066 }
3067
3068 // Here's a case where init0 is neither 0 nor -1:
3069 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3070 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3071 // In this case the tile is not split; it is (jlong)42.
3072 // The big tile is stored down, and then the foo() value is inserted.
3073 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3074
3075 Node* ctl = old->in(MemNode::Control);
3076 Node* adr = make_raw_address(offset, phase);
3077 const TypePtr* atp = TypeRawPtr::BOTTOM;
3078
3079 // One or two coalesced stores to plop down.
3080 Node* st[2];
3081 intptr_t off[2];
3082 int nst = 0;
3083 if (!split) {
3084 ++new_long;
3085 off[nst] = offset;
3086 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3087 phase->longcon(con), T_LONG);
3088 } else {
3089 // Omit either if it is a zero.
3090 if (con0 != 0) {
3091 ++new_int;
3092 off[nst] = offset;
3093 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3094 phase->intcon(con0), T_INT);
3095 }
3096 if (con1 != 0) {
3097 ++new_int;
3098 offset += BytesPerInt;
3099 adr = make_raw_address(offset, phase);
3100 off[nst] = offset;
3101 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3102 phase->intcon(con1), T_INT);
3103 }
3104 }
3105
3106 // Insert second store first, then the first before the second.
3107 // Insert each one just before any overlapping non-constant stores.
3108 while (nst > 0) {
3109 Node* st1 = st[--nst];
3110 C->copy_node_notes_to(st1, old);
3111 st1 = phase->transform(st1);
3112 offset = off[nst];
3113 assert(offset >= header_size, "do not smash header");
3114 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3115 guarantee(ins_idx != 0, "must re-insert constant store");
3116 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3117 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3118 set_req(--ins_idx, st1);
3119 else
3120 ins_req(ins_idx, st1);
3121 }
3122 }
3123
3124 if (PrintCompilation && WizardMode)
3125 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3126 old_subword, old_long, new_int, new_long);
3127 if (C->log() != NULL)
3128 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3129 old_subword, old_long, new_int, new_long);
3130
3131 // Clean up any remaining occurrences of zmem:
3132 remove_extra_zeroes();
3133 }
3134
3135 // Explore forward from in(start) to find the first fully initialized
3136 // word, and return its offset. Skip groups of subword stores which
3137 // together initialize full words. If in(start) is itself part of a
3138 // fully initialized word, return the offset of in(start). If there
3139 // are no following full-word stores, or if something is fishy, return
3140 // a negative value.
3141 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3142 int int_map = 0;
3143 intptr_t int_map_off = 0;
3144 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3145
3146 for (uint i = start, limit = req(); i < limit; i++) {
3147 Node* st = in(i);
3148
3149 intptr_t st_off = get_store_offset(st, phase);
3150 if (st_off < 0) break; // return conservative answer
3151
3152 int st_size = st->as_Store()->memory_size();
3153 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3154 return st_off; // we found a complete word init
3155 }
3156
3157 // update the map:
3158
3159 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3160 if (this_int_off != int_map_off) {
3161 // reset the map:
3162 int_map = 0;
3163 int_map_off = this_int_off;
3164 }
3165
3166 int subword_off = st_off - this_int_off;
3167 int_map |= right_n_bits(st_size) << subword_off;
3168 if ((int_map & FULL_MAP) == FULL_MAP) {
3169 return this_int_off; // we found a complete word init
3170 }
3171
3172 // Did this store hit or cross the word boundary?
3173 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3174 if (next_int_off == this_int_off + BytesPerInt) {
3175 // We passed the current int, without fully initializing it.
3176 int_map_off = next_int_off;
3177 int_map >>= BytesPerInt;
3178 } else if (next_int_off > this_int_off + BytesPerInt) {
3179 // We passed the current and next int.
3180 return this_int_off + BytesPerInt;
3181 }
3182 }
3183
3184 return -1;
3185 }
3186
3187
3188 // Called when the associated AllocateNode is expanded into CFG.
3189 // At this point, we may perform additional optimizations.
3190 // Linearize the stores by ascending offset, to make memory
3191 // activity as coherent as possible.
3192 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3193 intptr_t header_size,
3194 Node* size_in_bytes,
3195 PhaseGVN* phase) {
3196 assert(!is_complete(), "not already complete");
3197 assert(stores_are_sane(phase), "");
3198 assert(allocation() != NULL, "must be present");
3199
3200 remove_extra_zeroes();
3201
3202 if (ReduceFieldZeroing || ReduceBulkZeroing)
3203 // reduce instruction count for common initialization patterns
3204 coalesce_subword_stores(header_size, size_in_bytes, phase);
3205
3206 Node* zmem = zero_memory(); // initially zero memory state
3207 Node* inits = zmem; // accumulating a linearized chain of inits
3208 #ifdef ASSERT
3209 intptr_t first_offset = allocation()->minimum_header_size();
3210 intptr_t last_init_off = first_offset; // previous init offset
3211 intptr_t last_init_end = first_offset; // previous init offset+size
3212 intptr_t last_tile_end = first_offset; // previous tile offset+size
3213 #endif
3214 intptr_t zeroes_done = header_size;
3215
3216 bool do_zeroing = true; // we might give up if inits are very sparse
3217 int big_init_gaps = 0; // how many large gaps have we seen?
3218
3219 if (ZeroTLAB) do_zeroing = false;
3220 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3221
3222 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3223 Node* st = in(i);
3224 intptr_t st_off = get_store_offset(st, phase);
3225 if (st_off < 0)
3226 break; // unknown junk in the inits
3227 if (st->in(MemNode::Memory) != zmem)
3228 break; // complicated store chains somehow in list
3229
3230 int st_size = st->as_Store()->memory_size();
3231 intptr_t next_init_off = st_off + st_size;
3232
3233 if (do_zeroing && zeroes_done < next_init_off) {
3234 // See if this store needs a zero before it or under it.
3235 intptr_t zeroes_needed = st_off;
3236
3237 if (st_size < BytesPerInt) {
3238 // Look for subword stores which only partially initialize words.
3239 // If we find some, we must lay down some word-level zeroes first,
3240 // underneath the subword stores.
3241 //
3242 // Examples:
3243 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3244 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3245 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3246 //
3247 // Note: coalesce_subword_stores may have already done this,
3248 // if it was prompted by constant non-zero subword initializers.
3249 // But this case can still arise with non-constant stores.
3250
3251 intptr_t next_full_store = find_next_fullword_store(i, phase);
3252
3253 // In the examples above:
3254 // in(i) p q r s x y z
3255 // st_off 12 13 14 15 12 13 14
3256 // st_size 1 1 1 1 1 1 1
3257 // next_full_s. 12 16 16 16 16 16 16
3258 // z's_done 12 16 16 16 12 16 12
3259 // z's_needed 12 16 16 16 16 16 16
3260 // zsize 0 0 0 0 4 0 4
3261 if (next_full_store < 0) {
3262 // Conservative tack: Zero to end of current word.
3263 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3264 } else {
3265 // Zero to beginning of next fully initialized word.
3266 // Or, don't zero at all, if we are already in that word.
3267 assert(next_full_store >= zeroes_needed, "must go forward");
3268 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3269 zeroes_needed = next_full_store;
3270 }
3271 }
3272
3273 if (zeroes_needed > zeroes_done) {
3274 intptr_t zsize = zeroes_needed - zeroes_done;
3275 // Do some incremental zeroing on rawmem, in parallel with inits.
3276 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3277 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3278 zeroes_done, zeroes_needed,
3279 phase);
3280 zeroes_done = zeroes_needed;
3281 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3282 do_zeroing = false; // leave the hole, next time
3283 }
3284 }
3285
3286 // Collect the store and move on:
3287 st->set_req(MemNode::Memory, inits);
3288 inits = st; // put it on the linearized chain
3289 set_req(i, zmem); // unhook from previous position
3290
3291 if (zeroes_done == st_off)
3292 zeroes_done = next_init_off;
3293
3294 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3295
3296 #ifdef ASSERT
3297 // Various order invariants. Weaker than stores_are_sane because
3298 // a large constant tile can be filled in by smaller non-constant stores.
3299 assert(st_off >= last_init_off, "inits do not reverse");
3300 last_init_off = st_off;
3301 const Type* val = NULL;
3302 if (st_size >= BytesPerInt &&
3303 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3304 (int)val->basic_type() < (int)T_OBJECT) {
3305 assert(st_off >= last_tile_end, "tiles do not overlap");
3306 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3307 last_tile_end = MAX2(last_tile_end, next_init_off);
3308 } else {
3309 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3310 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3311 assert(st_off >= last_init_end, "inits do not overlap");
3312 last_init_end = next_init_off; // it's a non-tile
3313 }
3314 #endif //ASSERT
3315 }
3316
3317 remove_extra_zeroes(); // clear out all the zmems left over
3318 add_req(inits);
3319
3320 if (!ZeroTLAB) {
3321 // If anything remains to be zeroed, zero it all now.
3322 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3323 // if it is the last unused 4 bytes of an instance, forget about it
3324 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3325 if (zeroes_done + BytesPerLong >= size_limit) {
3326 assert(allocation() != NULL, "");
3327 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3328 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3329 if (zeroes_done == k->layout_helper())
3330 zeroes_done = size_limit;
3331 }
3332 if (zeroes_done < size_limit) {
3333 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3334 zeroes_done, size_in_bytes, phase);
3335 }
3336 }
3337
3338 set_complete(phase);
3339 return rawmem;
3340 }
3341
3342
3343 #ifdef ASSERT
3344 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3345 if (is_complete())
3346 return true; // stores could be anything at this point
3347 assert(allocation() != NULL, "must be present");
3348 intptr_t last_off = allocation()->minimum_header_size();
3349 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3350 Node* st = in(i);
3351 intptr_t st_off = get_store_offset(st, phase);
3352 if (st_off < 0) continue; // ignore dead garbage
3353 if (last_off > st_off) {
3354 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
3355 this->dump(2);
3356 assert(false, "ascending store offsets");
3357 return false;
3358 }
3359 last_off = st_off + st->as_Store()->memory_size();
3360 }
3361 return true;
3362 }
3363 #endif //ASSERT
3364
3365
3366
3367
3368 //============================MergeMemNode=====================================
3369 //
3370 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3371 // contributing store or call operations. Each contributor provides the memory
3372 // state for a particular "alias type" (see Compile::alias_type). For example,
3373 // if a MergeMem has an input X for alias category #6, then any memory reference
3374 // to alias category #6 may use X as its memory state input, as an exact equivalent
3375 // to using the MergeMem as a whole.
3376 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3377 //
3378 // (Here, the <N> notation gives the index of the relevant adr_type.)
3379 //
3380 // In one special case (and more cases in the future), alias categories overlap.
3381 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
3382 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
3383 // it is exactly equivalent to that state W:
3384 // MergeMem(<Bot>: W) <==> W
3385 //
3386 // Usually, the merge has more than one input. In that case, where inputs
3387 // overlap (i.e., one is Bot), the narrower alias type determines the memory
3388 // state for that type, and the wider alias type (Bot) fills in everywhere else:
3389 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
3390 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
3391 //
3392 // A merge can take a "wide" memory state as one of its narrow inputs.
3393 // This simply means that the merge observes out only the relevant parts of
3394 // the wide input. That is, wide memory states arriving at narrow merge inputs
3395 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
3396 //
3397 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
3398 // and that memory slices "leak through":
3399 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
3400 //
3401 // But, in such a cascade, repeated memory slices can "block the leak":
3402 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
3403 //
3404 // In the last example, Y is not part of the combined memory state of the
3405 // outermost MergeMem. The system must, of course, prevent unschedulable
3406 // memory states from arising, so you can be sure that the state Y is somehow
3407 // a precursor to state Y'.
3408 //
3409 //
3410 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
3411 // of each MergeMemNode array are exactly the numerical alias indexes, including
3412 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
3413 // Compile::alias_type (and kin) produce and manage these indexes.
3414 //
3415 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
3416 // (Note that this provides quick access to the top node inside MergeMem methods,
3417 // without the need to reach out via TLS to Compile::current.)
3418 //
3419 // As a consequence of what was just described, a MergeMem that represents a full
3420 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
3421 // containing all alias categories.
3422 //
3423 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
3424 //
3425 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
3426 // a memory state for the alias type <N>, or else the top node, meaning that
3427 // there is no particular input for that alias type. Note that the length of
3428 // a MergeMem is variable, and may be extended at any time to accommodate new
3429 // memory states at larger alias indexes. When merges grow, they are of course
3430 // filled with "top" in the unused in() positions.
3431 //
3432 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
3433 // (Top was chosen because it works smoothly with passes like GCM.)
3434 //
3435 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
3436 // the type of random VM bits like TLS references.) Since it is always the
3437 // first non-Bot memory slice, some low-level loops use it to initialize an
3438 // index variable: for (i = AliasIdxRaw; i < req(); i++).
3439 //
3440 //
3441 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
3442 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
3443 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
3444 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
3445 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
3446 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
3447 //
3448 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
3449 // really that different from the other memory inputs. An abbreviation called
3450 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
3451 //
3452 //
3453 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
3454 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
3455 // that "emerges though" the base memory will be marked as excluding the alias types
3456 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
3457 //
3458 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
3459 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
3460 //
3461 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
3462 // (It is currently unimplemented.) As you can see, the resulting merge is
3463 // actually a disjoint union of memory states, rather than an overlay.
3464 //
3465
3466 //------------------------------MergeMemNode-----------------------------------
3467 Node* MergeMemNode::make_empty_memory() {
3468 Node* empty_memory = (Node*) Compile::current()->top();
3469 assert(empty_memory->is_top(), "correct sentinel identity");
3470 return empty_memory;
3471 }
3472
3473 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3474 init_class_id(Class_MergeMem);
3475 // all inputs are nullified in Node::Node(int)
3476 // set_input(0, NULL); // no control input
3477
3478 // Initialize the edges uniformly to top, for starters.
3479 Node* empty_mem = make_empty_memory();
3480 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
3481 init_req(i,empty_mem);
3482 }
3483 assert(empty_memory() == empty_mem, "");
3484
3485 if( new_base != NULL && new_base->is_MergeMem() ) {
3486 MergeMemNode* mdef = new_base->as_MergeMem();
3487 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
3488 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
3489 mms.set_memory(mms.memory2());
3490 }
3491 assert(base_memory() == mdef->base_memory(), "");
3492 } else {
3493 set_base_memory(new_base);
3494 }
3495 }
3496
3497 // Make a new, untransformed MergeMem with the same base as 'mem'.
3498 // If mem is itself a MergeMem, populate the result with the same edges.
3499 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
3500 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
3501 }
3502
3503 //------------------------------cmp--------------------------------------------
3504 uint MergeMemNode::hash() const { return NO_HASH; }
3505 uint MergeMemNode::cmp( const Node &n ) const {
3506 return (&n == this); // Always fail except on self
3507 }
3508
3509 //------------------------------Identity---------------------------------------
3510 Node* MergeMemNode::Identity(PhaseTransform *phase) {
3511 // Identity if this merge point does not record any interesting memory
3512 // disambiguations.
3513 Node* base_mem = base_memory();
3514 Node* empty_mem = empty_memory();
3515 if (base_mem != empty_mem) { // Memory path is not dead?
3516 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3517 Node* mem = in(i);
3518 if (mem != empty_mem && mem != base_mem) {
3519 return this; // Many memory splits; no change
3520 }
3521 }
3522 }
3523 return base_mem; // No memory splits; ID on the one true input
3524 }
3525
3526 //------------------------------Ideal------------------------------------------
3527 // This method is invoked recursively on chains of MergeMem nodes
3528 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3529 // Remove chain'd MergeMems
3530 //
3531 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
3532 // relative to the "in(Bot)". Since we are patching both at the same time,
3533 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
3534 // but rewrite each "in(i)" relative to the new "in(Bot)".
3535 Node *progress = NULL;
3536
3537
3538 Node* old_base = base_memory();
3539 Node* empty_mem = empty_memory();
3540 if (old_base == empty_mem)
3541 return NULL; // Dead memory path.
3542
3543 MergeMemNode* old_mbase;
3544 if (old_base != NULL && old_base->is_MergeMem())
3545 old_mbase = old_base->as_MergeMem();
3546 else
3547 old_mbase = NULL;
3548 Node* new_base = old_base;
3549
3550 // simplify stacked MergeMems in base memory
3551 if (old_mbase) new_base = old_mbase->base_memory();
3552
3553 // the base memory might contribute new slices beyond my req()
3554 if (old_mbase) grow_to_match(old_mbase);
3555
3556 // Look carefully at the base node if it is a phi.
3557 PhiNode* phi_base;
3558 if (new_base != NULL && new_base->is_Phi())
3559 phi_base = new_base->as_Phi();
3560 else
3561 phi_base = NULL;
3562
3563 Node* phi_reg = NULL;
3564 uint phi_len = (uint)-1;
3565 if (phi_base != NULL && !phi_base->is_copy()) {
3566 // do not examine phi if degraded to a copy
3567 phi_reg = phi_base->region();
3568 phi_len = phi_base->req();
3569 // see if the phi is unfinished
3570 for (uint i = 1; i < phi_len; i++) {
3571 if (phi_base->in(i) == NULL) {
3572 // incomplete phi; do not look at it yet!
3573 phi_reg = NULL;
3574 phi_len = (uint)-1;
3575 break;
3576 }
3577 }
3578 }
3579
3580 // Note: We do not call verify_sparse on entry, because inputs
3581 // can normalize to the base_memory via subsume_node or similar
3582 // mechanisms. This method repairs that damage.
3583
3584 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
3585
3586 // Look at each slice.
3587 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3588 Node* old_in = in(i);
3589 // calculate the old memory value
3590 Node* old_mem = old_in;
3591 if (old_mem == empty_mem) old_mem = old_base;
3592 assert(old_mem == memory_at(i), "");
3593
3594 // maybe update (reslice) the old memory value
3595
3596 // simplify stacked MergeMems
3597 Node* new_mem = old_mem;
3598 MergeMemNode* old_mmem;
3599 if (old_mem != NULL && old_mem->is_MergeMem())
3600 old_mmem = old_mem->as_MergeMem();
3601 else
3602 old_mmem = NULL;
3603 if (old_mmem == this) {
3604 // This can happen if loops break up and safepoints disappear.
3605 // A merge of BotPtr (default) with a RawPtr memory derived from a
3606 // safepoint can be rewritten to a merge of the same BotPtr with
3607 // the BotPtr phi coming into the loop. If that phi disappears
3608 // also, we can end up with a self-loop of the mergemem.
3609 // In general, if loops degenerate and memory effects disappear,
3610 // a mergemem can be left looking at itself. This simply means
3611 // that the mergemem's default should be used, since there is
3612 // no longer any apparent effect on this slice.
3613 // Note: If a memory slice is a MergeMem cycle, it is unreachable
3614 // from start. Update the input to TOP.
3615 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
3616 }
3617 else if (old_mmem != NULL) {
3618 new_mem = old_mmem->memory_at(i);
3619 }
3620 // else preceeding memory was not a MergeMem
3621
3622 // replace equivalent phis (unfortunately, they do not GVN together)
3623 if (new_mem != NULL && new_mem != new_base &&
3624 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
3625 if (new_mem->is_Phi()) {
3626 PhiNode* phi_mem = new_mem->as_Phi();
3627 for (uint i = 1; i < phi_len; i++) {
3628 if (phi_base->in(i) != phi_mem->in(i)) {
3629 phi_mem = NULL;
3630 break;
3631 }
3632 }
3633 if (phi_mem != NULL) {
3634 // equivalent phi nodes; revert to the def
3635 new_mem = new_base;
3636 }
3637 }
3638 }
3639
3640 // maybe store down a new value
3641 Node* new_in = new_mem;
3642 if (new_in == new_base) new_in = empty_mem;
3643
3644 if (new_in != old_in) {
3645 // Warning: Do not combine this "if" with the previous "if"
3646 // A memory slice might have be be rewritten even if it is semantically
3647 // unchanged, if the base_memory value has changed.
3648 set_req(i, new_in);
3649 progress = this; // Report progress
3650 }
3651 }
3652
3653 if (new_base != old_base) {
3654 set_req(Compile::AliasIdxBot, new_base);
3655 // Don't use set_base_memory(new_base), because we need to update du.
3656 assert(base_memory() == new_base, "");
3657 progress = this;
3658 }
3659
3660 if( base_memory() == this ) {
3661 // a self cycle indicates this memory path is dead
3662 set_req(Compile::AliasIdxBot, empty_mem);
3663 }
3664
3665 // Resolve external cycles by calling Ideal on a MergeMem base_memory
3666 // Recursion must occur after the self cycle check above
3667 if( base_memory()->is_MergeMem() ) {
3668 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
3669 Node *m = phase->transform(new_mbase); // Rollup any cycles
3670 if( m != NULL && (m->is_top() ||
3671 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
3672 // propagate rollup of dead cycle to self
3673 set_req(Compile::AliasIdxBot, empty_mem);
3674 }
3675 }
3676
3677 if( base_memory() == empty_mem ) {
3678 progress = this;
3679 // Cut inputs during Parse phase only.
3680 // During Optimize phase a dead MergeMem node will be subsumed by Top.
3681 if( !can_reshape ) {
3682 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3683 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
3684 }
3685 }
3686 }
3687
3688 if( !progress && base_memory()->is_Phi() && can_reshape ) {
3689 // Check if PhiNode::Ideal's "Split phis through memory merges"
3690 // transform should be attempted. Look for this->phi->this cycle.
3691 uint merge_width = req();
3692 if (merge_width > Compile::AliasIdxRaw) {
3693 PhiNode* phi = base_memory()->as_Phi();
3694 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
3695 if (phi->in(i) == this) {
3696 phase->is_IterGVN()->_worklist.push(phi);
3697 break;
3698 }
3699 }
3700 }
3701 }
3702
3703 assert(progress || verify_sparse(), "please, no dups of base");
3704 return progress;
3705 }
3706
3707 //-------------------------set_base_memory-------------------------------------
3708 void MergeMemNode::set_base_memory(Node *new_base) {
3709 Node* empty_mem = empty_memory();
3710 set_req(Compile::AliasIdxBot, new_base);
3711 assert(memory_at(req()) == new_base, "must set default memory");
3712 // Clear out other occurrences of new_base:
3713 if (new_base != empty_mem) {
3714 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3715 if (in(i) == new_base) set_req(i, empty_mem);
3716 }
3717 }
3718 }
3719
3720 //------------------------------out_RegMask------------------------------------
3721 const RegMask &MergeMemNode::out_RegMask() const {
3722 return RegMask::Empty;
3723 }
3724
3725 //------------------------------dump_spec--------------------------------------
3726 #ifndef PRODUCT
3727 void MergeMemNode::dump_spec(outputStream *st) const {
3728 st->print(" {");
3729 Node* base_mem = base_memory();
3730 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
3731 Node* mem = memory_at(i);
3732 if (mem == base_mem) { st->print(" -"); continue; }
3733 st->print( " N%d:", mem->_idx );
3734 Compile::current()->get_adr_type(i)->dump_on(st);
3735 }
3736 st->print(" }");
3737 }
3738 #endif // !PRODUCT
3739
3740
3741 #ifdef ASSERT
3742 static bool might_be_same(Node* a, Node* b) {
3743 if (a == b) return true;
3744 if (!(a->is_Phi() || b->is_Phi())) return false;
3745 // phis shift around during optimization
3746 return true; // pretty stupid...
3747 }
3748
3749 // verify a narrow slice (either incoming or outgoing)
3750 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
3751 if (!VerifyAliases) return; // don't bother to verify unless requested
3752 if (is_error_reported()) return; // muzzle asserts when debugging an error
3753 if (Node::in_dump()) return; // muzzle asserts when printing
3754 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
3755 assert(n != NULL, "");
3756 // Elide intervening MergeMem's
3757 while (n->is_MergeMem()) {
3758 n = n->as_MergeMem()->memory_at(alias_idx);
3759 }
3760 Compile* C = Compile::current();
3761 const TypePtr* n_adr_type = n->adr_type();
3762 if (n == m->empty_memory()) {
3763 // Implicit copy of base_memory()
3764 } else if (n_adr_type != TypePtr::BOTTOM) {
3765 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
3766 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
3767 } else {
3768 // A few places like make_runtime_call "know" that VM calls are narrow,
3769 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
3770 bool expected_wide_mem = false;
3771 if (n == m->base_memory()) {
3772 expected_wide_mem = true;
3773 } else if (alias_idx == Compile::AliasIdxRaw ||
3774 n == m->memory_at(Compile::AliasIdxRaw)) {
3775 expected_wide_mem = true;
3776 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
3777 // memory can "leak through" calls on channels that
3778 // are write-once. Allow this also.
3779 expected_wide_mem = true;
3780 }
3781 assert(expected_wide_mem, "expected narrow slice replacement");
3782 }
3783 }
3784 #else // !ASSERT
3785 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
3786 #endif
3787
3788
3789 //-----------------------------memory_at---------------------------------------
3790 Node* MergeMemNode::memory_at(uint alias_idx) const {
3791 assert(alias_idx >= Compile::AliasIdxRaw ||
3792 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
3793 "must avoid base_memory and AliasIdxTop");
3794
3795 // Otherwise, it is a narrow slice.
3796 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
3797 Compile *C = Compile::current();
3798 if (is_empty_memory(n)) {
3799 // the array is sparse; empty slots are the "top" node
3800 n = base_memory();
3801 assert(Node::in_dump()
3802 || n == NULL || n->bottom_type() == Type::TOP
3803 || n->adr_type() == TypePtr::BOTTOM
3804 || n->adr_type() == TypeRawPtr::BOTTOM
3805 || Compile::current()->AliasLevel() == 0,
3806 "must be a wide memory");
3807 // AliasLevel == 0 if we are organizing the memory states manually.
3808 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
3809 } else {
3810 // make sure the stored slice is sane
3811 #ifdef ASSERT
3812 if (is_error_reported() || Node::in_dump()) {
3813 } else if (might_be_same(n, base_memory())) {
3814 // Give it a pass: It is a mostly harmless repetition of the base.
3815 // This can arise normally from node subsumption during optimization.
3816 } else {
3817 verify_memory_slice(this, alias_idx, n);
3818 }
3819 #endif
3820 }
3821 return n;
3822 }
3823
3824 //---------------------------set_memory_at-------------------------------------
3825 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
3826 verify_memory_slice(this, alias_idx, n);
3827 Node* empty_mem = empty_memory();
3828 if (n == base_memory()) n = empty_mem; // collapse default
3829 uint need_req = alias_idx+1;
3830 if (req() < need_req) {
3831 if (n == empty_mem) return; // already the default, so do not grow me
3832 // grow the sparse array
3833 do {
3834 add_req(empty_mem);
3835 } while (req() < need_req);
3836 }
3837 set_req( alias_idx, n );
3838 }
3839
3840
3841
3842 //--------------------------iteration_setup------------------------------------
3843 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
3844 if (other != NULL) {
3845 grow_to_match(other);
3846 // invariant: the finite support of mm2 is within mm->req()
3847 #ifdef ASSERT
3848 for (uint i = req(); i < other->req(); i++) {
3849 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
3850 }
3851 #endif
3852 }
3853 // Replace spurious copies of base_memory by top.
3854 Node* base_mem = base_memory();
3855 if (base_mem != NULL && !base_mem->is_top()) {
3856 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
3857 if (in(i) == base_mem)
3858 set_req(i, empty_memory());
3859 }
3860 }
3861 }
3862
3863 //---------------------------grow_to_match-------------------------------------
3864 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
3865 Node* empty_mem = empty_memory();
3866 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
3867 // look for the finite support of the other memory
3868 for (uint i = other->req(); --i >= req(); ) {
3869 if (other->in(i) != empty_mem) {
3870 uint new_len = i+1;
3871 while (req() < new_len) add_req(empty_mem);
3872 break;
3873 }
3874 }
3875 }
3876
3877 //---------------------------verify_sparse-------------------------------------
3878 #ifndef PRODUCT
3879 bool MergeMemNode::verify_sparse() const {
3880 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
3881 Node* base_mem = base_memory();
3882 // The following can happen in degenerate cases, since empty==top.
3883 if (is_empty_memory(base_mem)) return true;
3884 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3885 assert(in(i) != NULL, "sane slice");
3886 if (in(i) == base_mem) return false; // should have been the sentinel value!
3887 }
3888 return true;
3889 }
3890
3891 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
3892 Node* n;
3893 n = mm->in(idx);
3894 if (mem == n) return true; // might be empty_memory()
3895 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
3896 if (mem == n) return true;
3897 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
3898 if (mem == n) return true;
3899 if (n == NULL) break;
3900 }
3901 return false;
3902 }
3903 #endif // !PRODUCT