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