Thu, 12 Nov 2009 09:24:21 -0800
6892658: C2 should optimize some stringbuilder patterns
Reviewed-by: kvn, twisti
1 /*
2 * Copyright 1997-2009 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 */
25 // Portions of code courtesy of Clifford Click
27 // Optimization - Graph Style
29 #include "incls/_precompiled.incl"
30 #include "incls/_memnode.cpp.incl"
32 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
34 //=============================================================================
35 uint MemNode::size_of() const { return sizeof(*this); }
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 }
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);
55 Compile* C = Compile::current();
56 if( C->alias_type(_adr_type)->is_volatile() )
57 st->print(" Volatile!");
58 }
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 }
88 extern void print_alias_types();
90 #endif
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_known_instance_field())
95 return mchain; // don't try to optimize non-instance types
96 uint instance_id = tinst->instance_id();
97 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
98 Node *prev = NULL;
99 Node *result = mchain;
100 while (prev != result) {
101 prev = result;
102 if (result == start_mem)
103 break; // hit one of our sentinels
104 // skip over a call which does not affect this memory slice
105 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
106 Node *proj_in = result->in(0);
107 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
108 break; // hit one of our sentinels
109 } else if (proj_in->is_Call()) {
110 CallNode *call = proj_in->as_Call();
111 if (!call->may_modify(t_adr, phase)) {
112 result = call->in(TypeFunc::Memory);
113 }
114 } else if (proj_in->is_Initialize()) {
115 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
116 // Stop if this is the initialization for the object instance which
117 // which contains this memory slice, otherwise skip over it.
118 if (alloc != NULL && alloc->_idx != instance_id) {
119 result = proj_in->in(TypeFunc::Memory);
120 }
121 } else if (proj_in->is_MemBar()) {
122 result = proj_in->in(TypeFunc::Memory);
123 } else {
124 assert(false, "unexpected projection");
125 }
126 } else if (result->is_MergeMem()) {
127 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
128 }
129 }
130 return result;
131 }
133 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
134 const TypeOopPtr *t_oop = t_adr->isa_oopptr();
135 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field();
136 PhaseIterGVN *igvn = phase->is_IterGVN();
137 Node *result = mchain;
138 result = optimize_simple_memory_chain(result, t_adr, phase);
139 if (is_instance && igvn != NULL && result->is_Phi()) {
140 PhiNode *mphi = result->as_Phi();
141 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
142 const TypePtr *t = mphi->adr_type();
143 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
144 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
145 t->is_oopptr()->cast_to_exactness(true)
146 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
147 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
148 // clone the Phi with our address type
149 result = mphi->split_out_instance(t_adr, igvn);
150 } else {
151 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
152 }
153 }
154 return result;
155 }
157 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
158 uint alias_idx = phase->C->get_alias_index(tp);
159 Node *mem = mmem;
160 #ifdef ASSERT
161 {
162 // Check that current type is consistent with the alias index used during graph construction
163 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
164 bool consistent = adr_check == NULL || adr_check->empty() ||
165 phase->C->must_alias(adr_check, alias_idx );
166 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
167 if( !consistent && adr_check != NULL && !adr_check->empty() &&
168 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
169 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
170 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
171 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
172 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
173 // don't assert if it is dead code.
174 consistent = true;
175 }
176 if( !consistent ) {
177 st->print("alias_idx==%d, adr_check==", alias_idx);
178 if( adr_check == NULL ) {
179 st->print("NULL");
180 } else {
181 adr_check->dump();
182 }
183 st->cr();
184 print_alias_types();
185 assert(consistent, "adr_check must match alias idx");
186 }
187 }
188 #endif
189 // TypeInstPtr::NOTNULL+any is an OOP with unknown offset - generally
190 // means an array I have not precisely typed yet. Do not do any
191 // alias stuff with it any time soon.
192 const TypeOopPtr *tinst = tp->isa_oopptr();
193 if( tp->base() != Type::AnyPtr &&
194 !(tinst &&
195 tinst->klass()->is_java_lang_Object() &&
196 tinst->offset() == Type::OffsetBot) ) {
197 // compress paths and change unreachable cycles to TOP
198 // If not, we can update the input infinitely along a MergeMem cycle
199 // Equivalent code in PhiNode::Ideal
200 Node* m = phase->transform(mmem);
201 // If transformed to a MergeMem, get the desired slice
202 // Otherwise the returned node represents memory for every slice
203 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
204 // Update input if it is progress over what we have now
205 }
206 return mem;
207 }
209 //--------------------------Ideal_common---------------------------------------
210 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
211 // Unhook non-raw memories from complete (macro-expanded) initializations.
212 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
213 // If our control input is a dead region, kill all below the region
214 Node *ctl = in(MemNode::Control);
215 if (ctl && remove_dead_region(phase, can_reshape))
216 return this;
217 ctl = in(MemNode::Control);
218 // Don't bother trying to transform a dead node
219 if( ctl && ctl->is_top() ) return NodeSentinel;
221 PhaseIterGVN *igvn = phase->is_IterGVN();
222 // Wait if control on the worklist.
223 if (ctl && can_reshape && igvn != NULL) {
224 Node* bol = NULL;
225 Node* cmp = NULL;
226 if (ctl->in(0)->is_If()) {
227 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
228 bol = ctl->in(0)->in(1);
229 if (bol->is_Bool())
230 cmp = ctl->in(0)->in(1)->in(1);
231 }
232 if (igvn->_worklist.member(ctl) ||
233 (bol != NULL && igvn->_worklist.member(bol)) ||
234 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
235 // This control path may be dead.
236 // Delay this memory node transformation until the control is processed.
237 phase->is_IterGVN()->_worklist.push(this);
238 return NodeSentinel; // caller will return NULL
239 }
240 }
241 // Ignore if memory is dead, or self-loop
242 Node *mem = in(MemNode::Memory);
243 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
244 assert( mem != this, "dead loop in MemNode::Ideal" );
246 Node *address = in(MemNode::Address);
247 const Type *t_adr = phase->type( address );
248 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
250 if( can_reshape && igvn != NULL &&
251 (igvn->_worklist.member(address) || phase->type(address) != adr_type()) ) {
252 // The address's base and type may change when the address is processed.
253 // Delay this mem node transformation until the address is processed.
254 phase->is_IterGVN()->_worklist.push(this);
255 return NodeSentinel; // caller will return NULL
256 }
258 // Do NOT remove or optimize the next lines: ensure a new alias index
259 // is allocated for an oop pointer type before Escape Analysis.
260 // Note: C++ will not remove it since the call has side effect.
261 if ( t_adr->isa_oopptr() ) {
262 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
263 }
265 #ifdef ASSERT
266 Node* base = NULL;
267 if (address->is_AddP())
268 base = address->in(AddPNode::Base);
269 assert(base == NULL || t_adr->isa_rawptr() ||
270 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?");
271 #endif
273 // Avoid independent memory operations
274 Node* old_mem = mem;
276 // The code which unhooks non-raw memories from complete (macro-expanded)
277 // initializations was removed. After macro-expansion all stores catched
278 // by Initialize node became raw stores and there is no information
279 // which memory slices they modify. So it is unsafe to move any memory
280 // operation above these stores. Also in most cases hooked non-raw memories
281 // were already unhooked by using information from detect_ptr_independence()
282 // and find_previous_store().
284 if (mem->is_MergeMem()) {
285 MergeMemNode* mmem = mem->as_MergeMem();
286 const TypePtr *tp = t_adr->is_ptr();
288 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
289 }
291 if (mem != old_mem) {
292 set_req(MemNode::Memory, mem);
293 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
294 return this;
295 }
297 // let the subclass continue analyzing...
298 return NULL;
299 }
301 // Helper function for proving some simple control dominations.
302 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
303 // Already assumes that 'dom' is available at 'sub', and that 'sub'
304 // is not a constant (dominated by the method's StartNode).
305 // Used by MemNode::find_previous_store to prove that the
306 // control input of a memory operation predates (dominates)
307 // an allocation it wants to look past.
308 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
309 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
310 return false; // Conservative answer for dead code
312 // Check 'dom'. Skip Proj and CatchProj nodes.
313 dom = dom->find_exact_control(dom);
314 if (dom == NULL || dom->is_top())
315 return false; // Conservative answer for dead code
317 if (dom == sub) {
318 // For the case when, for example, 'sub' is Initialize and the original
319 // 'dom' is Proj node of the 'sub'.
320 return false;
321 }
323 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
324 return true;
326 // 'dom' dominates 'sub' if its control edge and control edges
327 // of all its inputs dominate or equal to sub's control edge.
329 // Currently 'sub' is either Allocate, Initialize or Start nodes.
330 // Or Region for the check in LoadNode::Ideal();
331 // 'sub' should have sub->in(0) != NULL.
332 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
333 sub->is_Region(), "expecting only these nodes");
335 // Get control edge of 'sub'.
336 Node* orig_sub = sub;
337 sub = sub->find_exact_control(sub->in(0));
338 if (sub == NULL || sub->is_top())
339 return false; // Conservative answer for dead code
341 assert(sub->is_CFG(), "expecting control");
343 if (sub == dom)
344 return true;
346 if (sub->is_Start() || sub->is_Root())
347 return false;
349 {
350 // Check all control edges of 'dom'.
352 ResourceMark rm;
353 Arena* arena = Thread::current()->resource_area();
354 Node_List nlist(arena);
355 Unique_Node_List dom_list(arena);
357 dom_list.push(dom);
358 bool only_dominating_controls = false;
360 for (uint next = 0; next < dom_list.size(); next++) {
361 Node* n = dom_list.at(next);
362 if (n == orig_sub)
363 return false; // One of dom's inputs dominated by sub.
364 if (!n->is_CFG() && n->pinned()) {
365 // Check only own control edge for pinned non-control nodes.
366 n = n->find_exact_control(n->in(0));
367 if (n == NULL || n->is_top())
368 return false; // Conservative answer for dead code
369 assert(n->is_CFG(), "expecting control");
370 dom_list.push(n);
371 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
372 only_dominating_controls = true;
373 } else if (n->is_CFG()) {
374 if (n->dominates(sub, nlist))
375 only_dominating_controls = true;
376 else
377 return false;
378 } else {
379 // First, own control edge.
380 Node* m = n->find_exact_control(n->in(0));
381 if (m != NULL) {
382 if (m->is_top())
383 return false; // Conservative answer for dead code
384 dom_list.push(m);
385 }
386 // Now, the rest of edges.
387 uint cnt = n->req();
388 for (uint i = 1; i < cnt; i++) {
389 m = n->find_exact_control(n->in(i));
390 if (m == NULL || m->is_top())
391 continue;
392 dom_list.push(m);
393 }
394 }
395 }
396 return only_dominating_controls;
397 }
398 }
400 //---------------------detect_ptr_independence---------------------------------
401 // Used by MemNode::find_previous_store to prove that two base
402 // pointers are never equal.
403 // The pointers are accompanied by their associated allocations,
404 // if any, which have been previously discovered by the caller.
405 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
406 Node* p2, AllocateNode* a2,
407 PhaseTransform* phase) {
408 // Attempt to prove that these two pointers cannot be aliased.
409 // They may both manifestly be allocations, and they should differ.
410 // Or, if they are not both allocations, they can be distinct constants.
411 // Otherwise, one is an allocation and the other a pre-existing value.
412 if (a1 == NULL && a2 == NULL) { // neither an allocation
413 return (p1 != p2) && p1->is_Con() && p2->is_Con();
414 } else if (a1 != NULL && a2 != NULL) { // both allocations
415 return (a1 != a2);
416 } else if (a1 != NULL) { // one allocation a1
417 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
418 return all_controls_dominate(p2, a1);
419 } else { //(a2 != NULL) // one allocation a2
420 return all_controls_dominate(p1, a2);
421 }
422 return false;
423 }
426 // The logic for reordering loads and stores uses four steps:
427 // (a) Walk carefully past stores and initializations which we
428 // can prove are independent of this load.
429 // (b) Observe that the next memory state makes an exact match
430 // with self (load or store), and locate the relevant store.
431 // (c) Ensure that, if we were to wire self directly to the store,
432 // the optimizer would fold it up somehow.
433 // (d) Do the rewiring, and return, depending on some other part of
434 // the optimizer to fold up the load.
435 // This routine handles steps (a) and (b). Steps (c) and (d) are
436 // specific to loads and stores, so they are handled by the callers.
437 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
438 //
439 Node* MemNode::find_previous_store(PhaseTransform* phase) {
440 Node* ctrl = in(MemNode::Control);
441 Node* adr = in(MemNode::Address);
442 intptr_t offset = 0;
443 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
444 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
446 if (offset == Type::OffsetBot)
447 return NULL; // cannot unalias unless there are precise offsets
449 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
451 intptr_t size_in_bytes = memory_size();
453 Node* mem = in(MemNode::Memory); // start searching here...
455 int cnt = 50; // Cycle limiter
456 for (;;) { // While we can dance past unrelated stores...
457 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
459 if (mem->is_Store()) {
460 Node* st_adr = mem->in(MemNode::Address);
461 intptr_t st_offset = 0;
462 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
463 if (st_base == NULL)
464 break; // inscrutable pointer
465 if (st_offset != offset && st_offset != Type::OffsetBot) {
466 const int MAX_STORE = BytesPerLong;
467 if (st_offset >= offset + size_in_bytes ||
468 st_offset <= offset - MAX_STORE ||
469 st_offset <= offset - mem->as_Store()->memory_size()) {
470 // Success: The offsets are provably independent.
471 // (You may ask, why not just test st_offset != offset and be done?
472 // The answer is that stores of different sizes can co-exist
473 // in the same sequence of RawMem effects. We sometimes initialize
474 // a whole 'tile' of array elements with a single jint or jlong.)
475 mem = mem->in(MemNode::Memory);
476 continue; // (a) advance through independent store memory
477 }
478 }
479 if (st_base != base &&
480 detect_ptr_independence(base, alloc,
481 st_base,
482 AllocateNode::Ideal_allocation(st_base, phase),
483 phase)) {
484 // Success: The bases are provably independent.
485 mem = mem->in(MemNode::Memory);
486 continue; // (a) advance through independent store memory
487 }
489 // (b) At this point, if the bases or offsets do not agree, we lose,
490 // since we have not managed to prove 'this' and 'mem' independent.
491 if (st_base == base && st_offset == offset) {
492 return mem; // let caller handle steps (c), (d)
493 }
495 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
496 InitializeNode* st_init = mem->in(0)->as_Initialize();
497 AllocateNode* st_alloc = st_init->allocation();
498 if (st_alloc == NULL)
499 break; // something degenerated
500 bool known_identical = false;
501 bool known_independent = false;
502 if (alloc == st_alloc)
503 known_identical = true;
504 else if (alloc != NULL)
505 known_independent = true;
506 else if (all_controls_dominate(this, st_alloc))
507 known_independent = true;
509 if (known_independent) {
510 // The bases are provably independent: Either they are
511 // manifestly distinct allocations, or else the control
512 // of this load dominates the store's allocation.
513 int alias_idx = phase->C->get_alias_index(adr_type());
514 if (alias_idx == Compile::AliasIdxRaw) {
515 mem = st_alloc->in(TypeFunc::Memory);
516 } else {
517 mem = st_init->memory(alias_idx);
518 }
519 continue; // (a) advance through independent store memory
520 }
522 // (b) at this point, if we are not looking at a store initializing
523 // the same allocation we are loading from, we lose.
524 if (known_identical) {
525 // From caller, can_see_stored_value will consult find_captured_store.
526 return mem; // let caller handle steps (c), (d)
527 }
529 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
530 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
531 if (mem->is_Proj() && mem->in(0)->is_Call()) {
532 CallNode *call = mem->in(0)->as_Call();
533 if (!call->may_modify(addr_t, phase)) {
534 mem = call->in(TypeFunc::Memory);
535 continue; // (a) advance through independent call memory
536 }
537 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
538 mem = mem->in(0)->in(TypeFunc::Memory);
539 continue; // (a) advance through independent MemBar memory
540 } else if (mem->is_MergeMem()) {
541 int alias_idx = phase->C->get_alias_index(adr_type());
542 mem = mem->as_MergeMem()->memory_at(alias_idx);
543 continue; // (a) advance through independent MergeMem memory
544 }
545 }
547 // Unless there is an explicit 'continue', we must bail out here,
548 // because 'mem' is an inscrutable memory state (e.g., a call).
549 break;
550 }
552 return NULL; // bail out
553 }
555 //----------------------calculate_adr_type-------------------------------------
556 // Helper function. Notices when the given type of address hits top or bottom.
557 // Also, asserts a cross-check of the type against the expected address type.
558 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
559 if (t == Type::TOP) return NULL; // does not touch memory any more?
560 #ifdef PRODUCT
561 cross_check = NULL;
562 #else
563 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
564 #endif
565 const TypePtr* tp = t->isa_ptr();
566 if (tp == NULL) {
567 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
568 return TypePtr::BOTTOM; // touches lots of memory
569 } else {
570 #ifdef ASSERT
571 // %%%% [phh] We don't check the alias index if cross_check is
572 // TypeRawPtr::BOTTOM. Needs to be investigated.
573 if (cross_check != NULL &&
574 cross_check != TypePtr::BOTTOM &&
575 cross_check != TypeRawPtr::BOTTOM) {
576 // Recheck the alias index, to see if it has changed (due to a bug).
577 Compile* C = Compile::current();
578 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
579 "must stay in the original alias category");
580 // The type of the address must be contained in the adr_type,
581 // disregarding "null"-ness.
582 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
583 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
584 assert(cross_check->meet(tp_notnull) == cross_check,
585 "real address must not escape from expected memory type");
586 }
587 #endif
588 return tp;
589 }
590 }
592 //------------------------adr_phi_is_loop_invariant----------------------------
593 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
594 // loop is loop invariant. Make a quick traversal of Phi and associated
595 // CastPP nodes, looking to see if they are a closed group within the loop.
596 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
597 // The idea is that the phi-nest must boil down to only CastPP nodes
598 // with the same data. This implies that any path into the loop already
599 // includes such a CastPP, and so the original cast, whatever its input,
600 // must be covered by an equivalent cast, with an earlier control input.
601 ResourceMark rm;
603 // The loop entry input of the phi should be the unique dominating
604 // node for every Phi/CastPP in the loop.
605 Unique_Node_List closure;
606 closure.push(adr_phi->in(LoopNode::EntryControl));
608 // Add the phi node and the cast to the worklist.
609 Unique_Node_List worklist;
610 worklist.push(adr_phi);
611 if( cast != NULL ){
612 if( !cast->is_ConstraintCast() ) return false;
613 worklist.push(cast);
614 }
616 // Begin recursive walk of phi nodes.
617 while( worklist.size() ){
618 // Take a node off the worklist
619 Node *n = worklist.pop();
620 if( !closure.member(n) ){
621 // Add it to the closure.
622 closure.push(n);
623 // Make a sanity check to ensure we don't waste too much time here.
624 if( closure.size() > 20) return false;
625 // This node is OK if:
626 // - it is a cast of an identical value
627 // - or it is a phi node (then we add its inputs to the worklist)
628 // Otherwise, the node is not OK, and we presume the cast is not invariant
629 if( n->is_ConstraintCast() ){
630 worklist.push(n->in(1));
631 } else if( n->is_Phi() ) {
632 for( uint i = 1; i < n->req(); i++ ) {
633 worklist.push(n->in(i));
634 }
635 } else {
636 return false;
637 }
638 }
639 }
641 // Quit when the worklist is empty, and we've found no offending nodes.
642 return true;
643 }
645 //------------------------------Ideal_DU_postCCP-------------------------------
646 // Find any cast-away of null-ness and keep its control. Null cast-aways are
647 // going away in this pass and we need to make this memory op depend on the
648 // gating null check.
649 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
650 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
651 }
653 // I tried to leave the CastPP's in. This makes the graph more accurate in
654 // some sense; we get to keep around the knowledge that an oop is not-null
655 // after some test. Alas, the CastPP's interfere with GVN (some values are
656 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
657 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
658 // some of the more trivial cases in the optimizer. Removing more useless
659 // Phi's started allowing Loads to illegally float above null checks. I gave
660 // up on this approach. CNC 10/20/2000
661 // This static method may be called not from MemNode (EncodePNode calls it).
662 // Only the control edge of the node 'n' might be updated.
663 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
664 Node *skipped_cast = NULL;
665 // Need a null check? Regular static accesses do not because they are
666 // from constant addresses. Array ops are gated by the range check (which
667 // always includes a NULL check). Just check field ops.
668 if( n->in(MemNode::Control) == NULL ) {
669 // Scan upwards for the highest location we can place this memory op.
670 while( true ) {
671 switch( adr->Opcode() ) {
673 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
674 adr = adr->in(AddPNode::Base);
675 continue;
677 case Op_DecodeN: // No change to NULL-ness, so peek thru
678 adr = adr->in(1);
679 continue;
681 case Op_CastPP:
682 // If the CastPP is useless, just peek on through it.
683 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
684 // Remember the cast that we've peeked though. If we peek
685 // through more than one, then we end up remembering the highest
686 // one, that is, if in a loop, the one closest to the top.
687 skipped_cast = adr;
688 adr = adr->in(1);
689 continue;
690 }
691 // CastPP is going away in this pass! We need this memory op to be
692 // control-dependent on the test that is guarding the CastPP.
693 ccp->hash_delete(n);
694 n->set_req(MemNode::Control, adr->in(0));
695 ccp->hash_insert(n);
696 return n;
698 case Op_Phi:
699 // Attempt to float above a Phi to some dominating point.
700 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
701 // If we've already peeked through a Cast (which could have set the
702 // control), we can't float above a Phi, because the skipped Cast
703 // may not be loop invariant.
704 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
705 adr = adr->in(1);
706 continue;
707 }
708 }
710 // Intentional fallthrough!
712 // No obvious dominating point. The mem op is pinned below the Phi
713 // by the Phi itself. If the Phi goes away (no true value is merged)
714 // then the mem op can float, but not indefinitely. It must be pinned
715 // behind the controls leading to the Phi.
716 case Op_CheckCastPP:
717 // These usually stick around to change address type, however a
718 // useless one can be elided and we still need to pick up a control edge
719 if (adr->in(0) == NULL) {
720 // This CheckCastPP node has NO control and is likely useless. But we
721 // need check further up the ancestor chain for a control input to keep
722 // the node in place. 4959717.
723 skipped_cast = adr;
724 adr = adr->in(1);
725 continue;
726 }
727 ccp->hash_delete(n);
728 n->set_req(MemNode::Control, adr->in(0));
729 ccp->hash_insert(n);
730 return n;
732 // List of "safe" opcodes; those that implicitly block the memory
733 // op below any null check.
734 case Op_CastX2P: // no null checks on native pointers
735 case Op_Parm: // 'this' pointer is not null
736 case Op_LoadP: // Loading from within a klass
737 case Op_LoadN: // Loading from within a klass
738 case Op_LoadKlass: // Loading from within a klass
739 case Op_LoadNKlass: // Loading from within a klass
740 case Op_ConP: // Loading from a klass
741 case Op_ConN: // Loading from a klass
742 case Op_CreateEx: // Sucking up the guts of an exception oop
743 case Op_Con: // Reading from TLS
744 case Op_CMoveP: // CMoveP is pinned
745 case Op_CMoveN: // CMoveN is pinned
746 break; // No progress
748 case Op_Proj: // Direct call to an allocation routine
749 case Op_SCMemProj: // Memory state from store conditional ops
750 #ifdef ASSERT
751 {
752 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
753 const Node* call = adr->in(0);
754 if (call->is_CallJava()) {
755 const CallJavaNode* call_java = call->as_CallJava();
756 const TypeTuple *r = call_java->tf()->range();
757 assert(r->cnt() > TypeFunc::Parms, "must return value");
758 const Type* ret_type = r->field_at(TypeFunc::Parms);
759 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
760 // We further presume that this is one of
761 // new_instance_Java, new_array_Java, or
762 // the like, but do not assert for this.
763 } else if (call->is_Allocate()) {
764 // similar case to new_instance_Java, etc.
765 } else if (!call->is_CallLeaf()) {
766 // Projections from fetch_oop (OSR) are allowed as well.
767 ShouldNotReachHere();
768 }
769 }
770 #endif
771 break;
772 default:
773 ShouldNotReachHere();
774 }
775 break;
776 }
777 }
779 return NULL; // No progress
780 }
783 //=============================================================================
784 uint LoadNode::size_of() const { return sizeof(*this); }
785 uint LoadNode::cmp( const Node &n ) const
786 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
787 const Type *LoadNode::bottom_type() const { return _type; }
788 uint LoadNode::ideal_reg() const {
789 return Matcher::base2reg[_type->base()];
790 }
792 #ifndef PRODUCT
793 void LoadNode::dump_spec(outputStream *st) const {
794 MemNode::dump_spec(st);
795 if( !Verbose && !WizardMode ) {
796 // standard dump does this in Verbose and WizardMode
797 st->print(" #"); _type->dump_on(st);
798 }
799 }
800 #endif
803 //----------------------------LoadNode::make-----------------------------------
804 // Polymorphic factory method:
805 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
806 Compile* C = gvn.C;
808 // sanity check the alias category against the created node type
809 assert(!(adr_type->isa_oopptr() &&
810 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
811 "use LoadKlassNode instead");
812 assert(!(adr_type->isa_aryptr() &&
813 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
814 "use LoadRangeNode instead");
815 switch (bt) {
816 case T_BOOLEAN: return new (C, 3) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() );
817 case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() );
818 case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() );
819 case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() );
820 case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() );
821 case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() );
822 case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt );
823 case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt );
824 case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() );
825 case T_OBJECT:
826 #ifdef _LP64
827 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
828 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));
829 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr());
830 } else
831 #endif
832 {
833 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop");
834 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
835 }
836 }
837 ShouldNotReachHere();
838 return (LoadNode*)NULL;
839 }
841 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
842 bool require_atomic = true;
843 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
844 }
849 //------------------------------hash-------------------------------------------
850 uint LoadNode::hash() const {
851 // unroll addition of interesting fields
852 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
853 }
855 //---------------------------can_see_stored_value------------------------------
856 // This routine exists to make sure this set of tests is done the same
857 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
858 // will change the graph shape in a way which makes memory alive twice at the
859 // same time (uses the Oracle model of aliasing), then some
860 // LoadXNode::Identity will fold things back to the equivalence-class model
861 // of aliasing.
862 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
863 Node* ld_adr = in(MemNode::Address);
865 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
866 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
867 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
868 atp->field() != NULL && !atp->field()->is_volatile()) {
869 uint alias_idx = atp->index();
870 bool final = atp->field()->is_final();
871 Node* result = NULL;
872 Node* current = st;
873 // Skip through chains of MemBarNodes checking the MergeMems for
874 // new states for the slice of this load. Stop once any other
875 // kind of node is encountered. Loads from final memory can skip
876 // through any kind of MemBar but normal loads shouldn't skip
877 // through MemBarAcquire since the could allow them to move out of
878 // a synchronized region.
879 while (current->is_Proj()) {
880 int opc = current->in(0)->Opcode();
881 if ((final && opc == Op_MemBarAcquire) ||
882 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder) {
883 Node* mem = current->in(0)->in(TypeFunc::Memory);
884 if (mem->is_MergeMem()) {
885 MergeMemNode* merge = mem->as_MergeMem();
886 Node* new_st = merge->memory_at(alias_idx);
887 if (new_st == merge->base_memory()) {
888 // Keep searching
889 current = merge->base_memory();
890 continue;
891 }
892 // Save the new memory state for the slice and fall through
893 // to exit.
894 result = new_st;
895 }
896 }
897 break;
898 }
899 if (result != NULL) {
900 st = result;
901 }
902 }
905 // Loop around twice in the case Load -> Initialize -> Store.
906 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
907 for (int trip = 0; trip <= 1; trip++) {
909 if (st->is_Store()) {
910 Node* st_adr = st->in(MemNode::Address);
911 if (!phase->eqv(st_adr, ld_adr)) {
912 // Try harder before giving up... Match raw and non-raw pointers.
913 intptr_t st_off = 0;
914 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
915 if (alloc == NULL) return NULL;
916 intptr_t ld_off = 0;
917 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
918 if (alloc != allo2) return NULL;
919 if (ld_off != st_off) return NULL;
920 // At this point we have proven something like this setup:
921 // A = Allocate(...)
922 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
923 // S = StoreQ(, AddP(, A.Parm , #Off), V)
924 // (Actually, we haven't yet proven the Q's are the same.)
925 // In other words, we are loading from a casted version of
926 // the same pointer-and-offset that we stored to.
927 // Thus, we are able to replace L by V.
928 }
929 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
930 if (store_Opcode() != st->Opcode())
931 return NULL;
932 return st->in(MemNode::ValueIn);
933 }
935 intptr_t offset = 0; // scratch
937 // A load from a freshly-created object always returns zero.
938 // (This can happen after LoadNode::Ideal resets the load's memory input
939 // to find_captured_store, which returned InitializeNode::zero_memory.)
940 if (st->is_Proj() && st->in(0)->is_Allocate() &&
941 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
942 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
943 // return a zero value for the load's basic type
944 // (This is one of the few places where a generic PhaseTransform
945 // can create new nodes. Think of it as lazily manifesting
946 // virtually pre-existing constants.)
947 return phase->zerocon(memory_type());
948 }
950 // A load from an initialization barrier can match a captured store.
951 if (st->is_Proj() && st->in(0)->is_Initialize()) {
952 InitializeNode* init = st->in(0)->as_Initialize();
953 AllocateNode* alloc = init->allocation();
954 if (alloc != NULL &&
955 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
956 // examine a captured store value
957 st = init->find_captured_store(offset, memory_size(), phase);
958 if (st != NULL)
959 continue; // take one more trip around
960 }
961 }
963 break;
964 }
966 return NULL;
967 }
969 //----------------------is_instance_field_load_with_local_phi------------------
970 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
971 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
972 in(MemNode::Address)->is_AddP() ) {
973 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
974 // Only instances.
975 if( t_oop != NULL && t_oop->is_known_instance_field() &&
976 t_oop->offset() != Type::OffsetBot &&
977 t_oop->offset() != Type::OffsetTop) {
978 return true;
979 }
980 }
981 return false;
982 }
984 //------------------------------Identity---------------------------------------
985 // Loads are identity if previous store is to same address
986 Node *LoadNode::Identity( PhaseTransform *phase ) {
987 // If the previous store-maker is the right kind of Store, and the store is
988 // to the same address, then we are equal to the value stored.
989 Node* mem = in(MemNode::Memory);
990 Node* value = can_see_stored_value(mem, phase);
991 if( value ) {
992 // byte, short & char stores truncate naturally.
993 // A load has to load the truncated value which requires
994 // some sort of masking operation and that requires an
995 // Ideal call instead of an Identity call.
996 if (memory_size() < BytesPerInt) {
997 // If the input to the store does not fit with the load's result type,
998 // it must be truncated via an Ideal call.
999 if (!phase->type(value)->higher_equal(phase->type(this)))
1000 return this;
1001 }
1002 // (This works even when value is a Con, but LoadNode::Value
1003 // usually runs first, producing the singleton type of the Con.)
1004 return value;
1005 }
1007 // Search for an existing data phi which was generated before for the same
1008 // instance's field to avoid infinite generation of phis in a loop.
1009 Node *region = mem->in(0);
1010 if (is_instance_field_load_with_local_phi(region)) {
1011 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
1012 int this_index = phase->C->get_alias_index(addr_t);
1013 int this_offset = addr_t->offset();
1014 int this_id = addr_t->is_oopptr()->instance_id();
1015 const Type* this_type = bottom_type();
1016 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1017 Node* phi = region->fast_out(i);
1018 if (phi->is_Phi() && phi != mem &&
1019 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
1020 return phi;
1021 }
1022 }
1023 }
1025 return this;
1026 }
1029 // Returns true if the AliasType refers to the field that holds the
1030 // cached box array. Currently only handles the IntegerCache case.
1031 static bool is_autobox_cache(Compile::AliasType* atp) {
1032 if (atp != NULL && atp->field() != NULL) {
1033 ciField* field = atp->field();
1034 ciSymbol* klass = field->holder()->name();
1035 if (field->name() == ciSymbol::cache_field_name() &&
1036 field->holder()->uses_default_loader() &&
1037 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
1038 return true;
1039 }
1040 }
1041 return false;
1042 }
1044 // Fetch the base value in the autobox array
1045 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
1046 if (atp != NULL && atp->field() != NULL) {
1047 ciField* field = atp->field();
1048 ciSymbol* klass = field->holder()->name();
1049 if (field->name() == ciSymbol::cache_field_name() &&
1050 field->holder()->uses_default_loader() &&
1051 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
1052 assert(field->is_constant(), "what?");
1053 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
1054 // Fetch the box object at the base of the array and get its value
1055 ciInstance* box = array->obj_at(0)->as_instance();
1056 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1057 if (ik->nof_nonstatic_fields() == 1) {
1058 // This should be true nonstatic_field_at requires calling
1059 // nof_nonstatic_fields so check it anyway
1060 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1061 cache_offset = c.as_int();
1062 }
1063 return true;
1064 }
1065 }
1066 return false;
1067 }
1069 // Returns true if the AliasType refers to the value field of an
1070 // autobox object. Currently only handles Integer.
1071 static bool is_autobox_object(Compile::AliasType* atp) {
1072 if (atp != NULL && atp->field() != NULL) {
1073 ciField* field = atp->field();
1074 ciSymbol* klass = field->holder()->name();
1075 if (field->name() == ciSymbol::value_name() &&
1076 field->holder()->uses_default_loader() &&
1077 klass == ciSymbol::java_lang_Integer()) {
1078 return true;
1079 }
1080 }
1081 return false;
1082 }
1085 // We're loading from an object which has autobox behaviour.
1086 // If this object is result of a valueOf call we'll have a phi
1087 // merging a newly allocated object and a load from the cache.
1088 // We want to replace this load with the original incoming
1089 // argument to the valueOf call.
1090 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1091 Node* base = in(Address)->in(AddPNode::Base);
1092 if (base->is_Phi() && base->req() == 3) {
1093 AllocateNode* allocation = NULL;
1094 int allocation_index = -1;
1095 int load_index = -1;
1096 for (uint i = 1; i < base->req(); i++) {
1097 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
1098 if (allocation != NULL) {
1099 allocation_index = i;
1100 load_index = 3 - allocation_index;
1101 break;
1102 }
1103 }
1104 bool has_load = ( allocation != NULL &&
1105 (base->in(load_index)->is_Load() ||
1106 base->in(load_index)->is_DecodeN() &&
1107 base->in(load_index)->in(1)->is_Load()) );
1108 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
1109 // Push the loads from the phi that comes from valueOf up
1110 // through it to allow elimination of the loads and the recovery
1111 // of the original value.
1112 Node* mem_phi = in(Memory);
1113 Node* offset = in(Address)->in(AddPNode::Offset);
1114 Node* region = base->in(0);
1116 Node* in1 = clone();
1117 Node* in1_addr = in1->in(Address)->clone();
1118 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
1119 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
1120 in1_addr->set_req(AddPNode::Offset, offset);
1121 in1->set_req(0, region->in(allocation_index));
1122 in1->set_req(Address, in1_addr);
1123 in1->set_req(Memory, mem_phi->in(allocation_index));
1125 Node* in2 = clone();
1126 Node* in2_addr = in2->in(Address)->clone();
1127 in2_addr->set_req(AddPNode::Base, base->in(load_index));
1128 in2_addr->set_req(AddPNode::Address, base->in(load_index));
1129 in2_addr->set_req(AddPNode::Offset, offset);
1130 in2->set_req(0, region->in(load_index));
1131 in2->set_req(Address, in2_addr);
1132 in2->set_req(Memory, mem_phi->in(load_index));
1134 in1_addr = phase->transform(in1_addr);
1135 in1 = phase->transform(in1);
1136 in2_addr = phase->transform(in2_addr);
1137 in2 = phase->transform(in2);
1139 PhiNode* result = PhiNode::make_blank(region, this);
1140 result->set_req(allocation_index, in1);
1141 result->set_req(load_index, in2);
1142 return result;
1143 }
1144 } else if (base->is_Load() ||
1145 base->is_DecodeN() && base->in(1)->is_Load()) {
1146 if (base->is_DecodeN()) {
1147 // Get LoadN node which loads cached Integer object
1148 base = base->in(1);
1149 }
1150 // Eliminate the load of Integer.value for integers from the cache
1151 // array by deriving the value from the index into the array.
1152 // Capture the offset of the load and then reverse the computation.
1153 Node* load_base = base->in(Address)->in(AddPNode::Base);
1154 if (load_base->is_DecodeN()) {
1155 // Get LoadN node which loads IntegerCache.cache field
1156 load_base = load_base->in(1);
1157 }
1158 if (load_base != NULL) {
1159 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
1160 intptr_t cache_offset;
1161 int shift = -1;
1162 Node* cache = NULL;
1163 if (is_autobox_cache(atp)) {
1164 shift = exact_log2(type2aelembytes(T_OBJECT));
1165 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
1166 }
1167 if (cache != NULL && base->in(Address)->is_AddP()) {
1168 Node* elements[4];
1169 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
1170 int cache_low;
1171 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
1172 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
1173 // Add up all the offsets making of the address of the load
1174 Node* result = elements[0];
1175 for (int i = 1; i < count; i++) {
1176 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
1177 }
1178 // Remove the constant offset from the address and then
1179 // remove the scaling of the offset to recover the original index.
1180 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
1181 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1182 // Peel the shift off directly but wrap it in a dummy node
1183 // since Ideal can't return existing nodes
1184 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
1185 } else {
1186 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
1187 }
1188 #ifdef _LP64
1189 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
1190 #endif
1191 return result;
1192 }
1193 }
1194 }
1195 }
1196 return NULL;
1197 }
1199 //------------------------------split_through_phi------------------------------
1200 // Split instance field load through Phi.
1201 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1202 Node* mem = in(MemNode::Memory);
1203 Node* address = in(MemNode::Address);
1204 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1205 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1207 assert(mem->is_Phi() && (t_oop != NULL) &&
1208 t_oop->is_known_instance_field(), "invalide conditions");
1210 Node *region = mem->in(0);
1211 if (region == NULL) {
1212 return NULL; // Wait stable graph
1213 }
1214 uint cnt = mem->req();
1215 for( uint i = 1; i < cnt; i++ ) {
1216 Node *in = mem->in(i);
1217 if( in == NULL ) {
1218 return NULL; // Wait stable graph
1219 }
1220 }
1221 // Check for loop invariant.
1222 if (cnt == 3) {
1223 for( uint i = 1; i < cnt; i++ ) {
1224 Node *in = mem->in(i);
1225 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
1226 if (m == mem) {
1227 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
1228 return this;
1229 }
1230 }
1231 }
1232 // Split through Phi (see original code in loopopts.cpp).
1233 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
1235 // Do nothing here if Identity will find a value
1236 // (to avoid infinite chain of value phis generation).
1237 if ( !phase->eqv(this, this->Identity(phase)) )
1238 return NULL;
1240 // Skip the split if the region dominates some control edge of the address.
1241 if (cnt == 3 && !MemNode::all_controls_dominate(address, region))
1242 return NULL;
1244 const Type* this_type = this->bottom_type();
1245 int this_index = phase->C->get_alias_index(addr_t);
1246 int this_offset = addr_t->offset();
1247 int this_iid = addr_t->is_oopptr()->instance_id();
1248 int wins = 0;
1249 PhaseIterGVN *igvn = phase->is_IterGVN();
1250 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1251 for( uint i = 1; i < region->req(); i++ ) {
1252 Node *x;
1253 Node* the_clone = NULL;
1254 if( region->in(i) == phase->C->top() ) {
1255 x = phase->C->top(); // Dead path? Use a dead data op
1256 } else {
1257 x = this->clone(); // Else clone up the data op
1258 the_clone = x; // Remember for possible deletion.
1259 // Alter data node to use pre-phi inputs
1260 if( this->in(0) == region ) {
1261 x->set_req( 0, region->in(i) );
1262 } else {
1263 x->set_req( 0, NULL );
1264 }
1265 for( uint j = 1; j < this->req(); j++ ) {
1266 Node *in = this->in(j);
1267 if( in->is_Phi() && in->in(0) == region )
1268 x->set_req( j, in->in(i) ); // Use pre-Phi input for the clone
1269 }
1270 }
1271 // Check for a 'win' on some paths
1272 const Type *t = x->Value(igvn);
1274 bool singleton = t->singleton();
1276 // See comments in PhaseIdealLoop::split_thru_phi().
1277 if( singleton && t == Type::TOP ) {
1278 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1279 }
1281 if( singleton ) {
1282 wins++;
1283 x = igvn->makecon(t);
1284 } else {
1285 // We now call Identity to try to simplify the cloned node.
1286 // Note that some Identity methods call phase->type(this).
1287 // Make sure that the type array is big enough for
1288 // our new node, even though we may throw the node away.
1289 // (This tweaking with igvn only works because x is a new node.)
1290 igvn->set_type(x, t);
1291 // If x is a TypeNode, capture any more-precise type permanently into Node
1292 // otherwise it will be not updated during igvn->transform since
1293 // igvn->type(x) is set to x->Value() already.
1294 x->raise_bottom_type(t);
1295 Node *y = x->Identity(igvn);
1296 if( y != x ) {
1297 wins++;
1298 x = y;
1299 } else {
1300 y = igvn->hash_find(x);
1301 if( y ) {
1302 wins++;
1303 x = y;
1304 } else {
1305 // Else x is a new node we are keeping
1306 // We do not need register_new_node_with_optimizer
1307 // because set_type has already been called.
1308 igvn->_worklist.push(x);
1309 }
1310 }
1311 }
1312 if (x != the_clone && the_clone != NULL)
1313 igvn->remove_dead_node(the_clone);
1314 phi->set_req(i, x);
1315 }
1316 if( wins > 0 ) {
1317 // Record Phi
1318 igvn->register_new_node_with_optimizer(phi);
1319 return phi;
1320 }
1321 igvn->remove_dead_node(phi);
1322 return NULL;
1323 }
1325 //------------------------------Ideal------------------------------------------
1326 // If the load is from Field memory and the pointer is non-null, we can
1327 // zero out the control input.
1328 // If the offset is constant and the base is an object allocation,
1329 // try to hook me up to the exact initializing store.
1330 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1331 Node* p = MemNode::Ideal_common(phase, can_reshape);
1332 if (p) return (p == NodeSentinel) ? NULL : p;
1334 Node* ctrl = in(MemNode::Control);
1335 Node* address = in(MemNode::Address);
1337 // Skip up past a SafePoint control. Cannot do this for Stores because
1338 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1339 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1340 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1341 ctrl = ctrl->in(0);
1342 set_req(MemNode::Control,ctrl);
1343 }
1345 intptr_t ignore = 0;
1346 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1347 if (base != NULL
1348 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1349 // Check for useless control edge in some common special cases
1350 if (in(MemNode::Control) != NULL
1351 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1352 && all_controls_dominate(base, phase->C->start())) {
1353 // A method-invariant, non-null address (constant or 'this' argument).
1354 set_req(MemNode::Control, NULL);
1355 }
1357 if (EliminateAutoBox && can_reshape) {
1358 assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null");
1359 Compile::AliasType* atp = phase->C->alias_type(adr_type());
1360 if (is_autobox_object(atp)) {
1361 Node* result = eliminate_autobox(phase);
1362 if (result != NULL) return result;
1363 }
1364 }
1365 }
1367 Node* mem = in(MemNode::Memory);
1368 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1370 if (addr_t != NULL) {
1371 // try to optimize our memory input
1372 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
1373 if (opt_mem != mem) {
1374 set_req(MemNode::Memory, opt_mem);
1375 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1376 return this;
1377 }
1378 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1379 if (can_reshape && opt_mem->is_Phi() &&
1380 (t_oop != NULL) && t_oop->is_known_instance_field()) {
1381 // Split instance field load through Phi.
1382 Node* result = split_through_phi(phase);
1383 if (result != NULL) return result;
1384 }
1385 }
1387 // Check for prior store with a different base or offset; make Load
1388 // independent. Skip through any number of them. Bail out if the stores
1389 // are in an endless dead cycle and report no progress. This is a key
1390 // transform for Reflection. However, if after skipping through the Stores
1391 // we can't then fold up against a prior store do NOT do the transform as
1392 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1393 // array memory alive twice: once for the hoisted Load and again after the
1394 // bypassed Store. This situation only works if EVERYBODY who does
1395 // anti-dependence work knows how to bypass. I.e. we need all
1396 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1397 // the alias index stuff. So instead, peek through Stores and IFF we can
1398 // fold up, do so.
1399 Node* prev_mem = find_previous_store(phase);
1400 // Steps (a), (b): Walk past independent stores to find an exact match.
1401 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1402 // (c) See if we can fold up on the spot, but don't fold up here.
1403 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1404 // just return a prior value, which is done by Identity calls.
1405 if (can_see_stored_value(prev_mem, phase)) {
1406 // Make ready for step (d):
1407 set_req(MemNode::Memory, prev_mem);
1408 return this;
1409 }
1410 }
1412 return NULL; // No further progress
1413 }
1415 // Helper to recognize certain Klass fields which are invariant across
1416 // some group of array types (e.g., int[] or all T[] where T < Object).
1417 const Type*
1418 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1419 ciKlass* klass) const {
1420 if (tkls->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1421 // The field is Klass::_modifier_flags. Return its (constant) value.
1422 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1423 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1424 return TypeInt::make(klass->modifier_flags());
1425 }
1426 if (tkls->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc)) {
1427 // The field is Klass::_access_flags. Return its (constant) value.
1428 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1429 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1430 return TypeInt::make(klass->access_flags());
1431 }
1432 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)) {
1433 // The field is Klass::_layout_helper. Return its constant value if known.
1434 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1435 return TypeInt::make(klass->layout_helper());
1436 }
1438 // No match.
1439 return NULL;
1440 }
1442 //------------------------------Value-----------------------------------------
1443 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1444 // Either input is TOP ==> the result is TOP
1445 Node* mem = in(MemNode::Memory);
1446 const Type *t1 = phase->type(mem);
1447 if (t1 == Type::TOP) return Type::TOP;
1448 Node* adr = in(MemNode::Address);
1449 const TypePtr* tp = phase->type(adr)->isa_ptr();
1450 if (tp == NULL || tp->empty()) return Type::TOP;
1451 int off = tp->offset();
1452 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1454 // Try to guess loaded type from pointer type
1455 if (tp->base() == Type::AryPtr) {
1456 const Type *t = tp->is_aryptr()->elem();
1457 // Don't do this for integer types. There is only potential profit if
1458 // the element type t is lower than _type; that is, for int types, if _type is
1459 // more restrictive than t. This only happens here if one is short and the other
1460 // char (both 16 bits), and in those cases we've made an intentional decision
1461 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1462 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1463 //
1464 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1465 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1466 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1467 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1468 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1469 // In fact, that could have been the original type of p1, and p1 could have
1470 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1471 // expression (LShiftL quux 3) independently optimized to the constant 8.
1472 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1473 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1474 // t might actually be lower than _type, if _type is a unique
1475 // concrete subclass of abstract class t.
1476 // Make sure the reference is not into the header, by comparing
1477 // the offset against the offset of the start of the array's data.
1478 // Different array types begin at slightly different offsets (12 vs. 16).
1479 // We choose T_BYTE as an example base type that is least restrictive
1480 // as to alignment, which will therefore produce the smallest
1481 // possible base offset.
1482 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1483 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
1484 const Type* jt = t->join(_type);
1485 // In any case, do not allow the join, per se, to empty out the type.
1486 if (jt->empty() && !t->empty()) {
1487 // This can happen if a interface-typed array narrows to a class type.
1488 jt = _type;
1489 }
1491 if (EliminateAutoBox && adr->is_AddP()) {
1492 // The pointers in the autobox arrays are always non-null
1493 Node* base = adr->in(AddPNode::Base);
1494 if (base != NULL &&
1495 !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) {
1496 Compile::AliasType* atp = phase->C->alias_type(base->adr_type());
1497 if (is_autobox_cache(atp)) {
1498 return jt->join(TypePtr::NOTNULL)->is_ptr();
1499 }
1500 }
1501 }
1502 return jt;
1503 }
1504 }
1505 } else if (tp->base() == Type::InstPtr) {
1506 const TypeInstPtr* tinst = tp->is_instptr();
1507 ciKlass* klass = tinst->klass();
1508 assert( off != Type::OffsetBot ||
1509 // arrays can be cast to Objects
1510 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1511 // unsafe field access may not have a constant offset
1512 phase->C->has_unsafe_access(),
1513 "Field accesses must be precise" );
1514 // For oop loads, we expect the _type to be precise
1515 if (OptimizeStringConcat && klass == phase->C->env()->String_klass() &&
1516 adr->is_AddP() && off != Type::OffsetBot) {
1517 // For constant Strings treat the fields as compile time constants.
1518 Node* base = adr->in(AddPNode::Base);
1519 if (base->Opcode() == Op_ConP) {
1520 const TypeOopPtr* t = phase->type(base)->isa_oopptr();
1521 ciObject* string = t->const_oop();
1522 ciConstant constant = string->as_instance()->field_value_by_offset(off);
1523 if (constant.basic_type() == T_INT) {
1524 return TypeInt::make(constant.as_int());
1525 } else if (constant.basic_type() == T_ARRAY) {
1526 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1527 return TypeNarrowOop::make_from_constant(constant.as_object());
1528 } else {
1529 return TypeOopPtr::make_from_constant(constant.as_object());
1530 }
1531 }
1532 }
1533 }
1534 } else if (tp->base() == Type::KlassPtr) {
1535 assert( off != Type::OffsetBot ||
1536 // arrays can be cast to Objects
1537 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1538 // also allow array-loading from the primary supertype
1539 // array during subtype checks
1540 Opcode() == Op_LoadKlass,
1541 "Field accesses must be precise" );
1542 // For klass/static loads, we expect the _type to be precise
1543 }
1545 const TypeKlassPtr *tkls = tp->isa_klassptr();
1546 if (tkls != NULL && !StressReflectiveCode) {
1547 ciKlass* klass = tkls->klass();
1548 if (klass->is_loaded() && tkls->klass_is_exact()) {
1549 // We are loading a field from a Klass metaobject whose identity
1550 // is known at compile time (the type is "exact" or "precise").
1551 // Check for fields we know are maintained as constants by the VM.
1552 if (tkls->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc)) {
1553 // The field is Klass::_super_check_offset. Return its (constant) value.
1554 // (Folds up type checking code.)
1555 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1556 return TypeInt::make(klass->super_check_offset());
1557 }
1558 // Compute index into primary_supers array
1559 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1560 // Check for overflowing; use unsigned compare to handle the negative case.
1561 if( depth < ciKlass::primary_super_limit() ) {
1562 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1563 // (Folds up type checking code.)
1564 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1565 ciKlass *ss = klass->super_of_depth(depth);
1566 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1567 }
1568 const Type* aift = load_array_final_field(tkls, klass);
1569 if (aift != NULL) return aift;
1570 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset()) + (int)sizeof(oopDesc)
1571 && klass->is_array_klass()) {
1572 // The field is arrayKlass::_component_mirror. Return its (constant) value.
1573 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1574 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1575 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1576 }
1577 if (tkls->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc)) {
1578 // The field is Klass::_java_mirror. Return its (constant) value.
1579 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1580 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1581 return TypeInstPtr::make(klass->java_mirror());
1582 }
1583 }
1585 // We can still check if we are loading from the primary_supers array at a
1586 // shallow enough depth. Even though the klass is not exact, entries less
1587 // than or equal to its super depth are correct.
1588 if (klass->is_loaded() ) {
1589 ciType *inner = klass->klass();
1590 while( inner->is_obj_array_klass() )
1591 inner = inner->as_obj_array_klass()->base_element_type();
1592 if( inner->is_instance_klass() &&
1593 !inner->as_instance_klass()->flags().is_interface() ) {
1594 // Compute index into primary_supers array
1595 juint depth = (tkls->offset() - (Klass::primary_supers_offset_in_bytes() + (int)sizeof(oopDesc))) / sizeof(klassOop);
1596 // Check for overflowing; use unsigned compare to handle the negative case.
1597 if( depth < ciKlass::primary_super_limit() &&
1598 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1599 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1600 // (Folds up type checking code.)
1601 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1602 ciKlass *ss = klass->super_of_depth(depth);
1603 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1604 }
1605 }
1606 }
1608 // If the type is enough to determine that the thing is not an array,
1609 // we can give the layout_helper a positive interval type.
1610 // This will help short-circuit some reflective code.
1611 if (tkls->offset() == Klass::layout_helper_offset_in_bytes() + (int)sizeof(oopDesc)
1612 && !klass->is_array_klass() // not directly typed as an array
1613 && !klass->is_interface() // specifically not Serializable & Cloneable
1614 && !klass->is_java_lang_Object() // not the supertype of all T[]
1615 ) {
1616 // Note: When interfaces are reliable, we can narrow the interface
1617 // test to (klass != Serializable && klass != Cloneable).
1618 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1619 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1620 // The key property of this type is that it folds up tests
1621 // for array-ness, since it proves that the layout_helper is positive.
1622 // Thus, a generic value like the basic object layout helper works fine.
1623 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1624 }
1625 }
1627 // If we are loading from a freshly-allocated object, produce a zero,
1628 // if the load is provably beyond the header of the object.
1629 // (Also allow a variable load from a fresh array to produce zero.)
1630 if (ReduceFieldZeroing) {
1631 Node* value = can_see_stored_value(mem,phase);
1632 if (value != NULL && value->is_Con())
1633 return value->bottom_type();
1634 }
1636 const TypeOopPtr *tinst = tp->isa_oopptr();
1637 if (tinst != NULL && tinst->is_known_instance_field()) {
1638 // If we have an instance type and our memory input is the
1639 // programs's initial memory state, there is no matching store,
1640 // so just return a zero of the appropriate type
1641 Node *mem = in(MemNode::Memory);
1642 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1643 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1644 return Type::get_zero_type(_type->basic_type());
1645 }
1646 }
1647 return _type;
1648 }
1650 //------------------------------match_edge-------------------------------------
1651 // Do we Match on this edge index or not? Match only the address.
1652 uint LoadNode::match_edge(uint idx) const {
1653 return idx == MemNode::Address;
1654 }
1656 //--------------------------LoadBNode::Ideal--------------------------------------
1657 //
1658 // If the previous store is to the same address as this load,
1659 // and the value stored was larger than a byte, replace this load
1660 // with the value stored truncated to a byte. If no truncation is
1661 // needed, the replacement is done in LoadNode::Identity().
1662 //
1663 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1664 Node* mem = in(MemNode::Memory);
1665 Node* value = can_see_stored_value(mem,phase);
1666 if( value && !phase->type(value)->higher_equal( _type ) ) {
1667 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
1668 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
1669 }
1670 // Identity call will handle the case where truncation is not needed.
1671 return LoadNode::Ideal(phase, can_reshape);
1672 }
1674 //--------------------------LoadUBNode::Ideal-------------------------------------
1675 //
1676 // If the previous store is to the same address as this load,
1677 // and the value stored was larger than a byte, replace this load
1678 // with the value stored truncated to a byte. If no truncation is
1679 // needed, the replacement is done in LoadNode::Identity().
1680 //
1681 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1682 Node* mem = in(MemNode::Memory);
1683 Node* value = can_see_stored_value(mem, phase);
1684 if (value && !phase->type(value)->higher_equal(_type))
1685 return new (phase->C, 3) AndINode(value, phase->intcon(0xFF));
1686 // Identity call will handle the case where truncation is not needed.
1687 return LoadNode::Ideal(phase, can_reshape);
1688 }
1690 //--------------------------LoadUSNode::Ideal-------------------------------------
1691 //
1692 // If the previous store is to the same address as this load,
1693 // and the value stored was larger than a char, replace this load
1694 // with the value stored truncated to a char. If no truncation is
1695 // needed, the replacement is done in LoadNode::Identity().
1696 //
1697 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1698 Node* mem = in(MemNode::Memory);
1699 Node* value = can_see_stored_value(mem,phase);
1700 if( value && !phase->type(value)->higher_equal( _type ) )
1701 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
1702 // Identity call will handle the case where truncation is not needed.
1703 return LoadNode::Ideal(phase, can_reshape);
1704 }
1706 //--------------------------LoadSNode::Ideal--------------------------------------
1707 //
1708 // If the previous store is to the same address as this load,
1709 // and the value stored was larger than a short, replace this load
1710 // with the value stored truncated to a short. If no truncation is
1711 // needed, the replacement is done in LoadNode::Identity().
1712 //
1713 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1714 Node* mem = in(MemNode::Memory);
1715 Node* value = can_see_stored_value(mem,phase);
1716 if( value && !phase->type(value)->higher_equal( _type ) ) {
1717 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
1718 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
1719 }
1720 // Identity call will handle the case where truncation is not needed.
1721 return LoadNode::Ideal(phase, can_reshape);
1722 }
1724 //=============================================================================
1725 //----------------------------LoadKlassNode::make------------------------------
1726 // Polymorphic factory method:
1727 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
1728 Compile* C = gvn.C;
1729 Node *ctl = NULL;
1730 // sanity check the alias category against the created node type
1731 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr();
1732 assert(adr_type != NULL, "expecting TypeOopPtr");
1733 #ifdef _LP64
1734 if (adr_type->is_ptr_to_narrowoop()) {
1735 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop()));
1736 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr());
1737 }
1738 #endif
1739 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
1740 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk);
1741 }
1743 //------------------------------Value------------------------------------------
1744 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1745 return klass_value_common(phase);
1746 }
1748 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
1749 // Either input is TOP ==> the result is TOP
1750 const Type *t1 = phase->type( in(MemNode::Memory) );
1751 if (t1 == Type::TOP) return Type::TOP;
1752 Node *adr = in(MemNode::Address);
1753 const Type *t2 = phase->type( adr );
1754 if (t2 == Type::TOP) return Type::TOP;
1755 const TypePtr *tp = t2->is_ptr();
1756 if (TypePtr::above_centerline(tp->ptr()) ||
1757 tp->ptr() == TypePtr::Null) return Type::TOP;
1759 // Return a more precise klass, if possible
1760 const TypeInstPtr *tinst = tp->isa_instptr();
1761 if (tinst != NULL) {
1762 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1763 int offset = tinst->offset();
1764 if (ik == phase->C->env()->Class_klass()
1765 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1766 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1767 // We are loading a special hidden field from a Class mirror object,
1768 // the field which points to the VM's Klass metaobject.
1769 ciType* t = tinst->java_mirror_type();
1770 // java_mirror_type returns non-null for compile-time Class constants.
1771 if (t != NULL) {
1772 // constant oop => constant klass
1773 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1774 return TypeKlassPtr::make(ciArrayKlass::make(t));
1775 }
1776 if (!t->is_klass()) {
1777 // a primitive Class (e.g., int.class) has NULL for a klass field
1778 return TypePtr::NULL_PTR;
1779 }
1780 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1781 return TypeKlassPtr::make(t->as_klass());
1782 }
1783 // non-constant mirror, so we can't tell what's going on
1784 }
1785 if( !ik->is_loaded() )
1786 return _type; // Bail out if not loaded
1787 if (offset == oopDesc::klass_offset_in_bytes()) {
1788 if (tinst->klass_is_exact()) {
1789 return TypeKlassPtr::make(ik);
1790 }
1791 // See if we can become precise: no subklasses and no interface
1792 // (Note: We need to support verified interfaces.)
1793 if (!ik->is_interface() && !ik->has_subklass()) {
1794 //assert(!UseExactTypes, "this code should be useless with exact types");
1795 // Add a dependence; if any subclass added we need to recompile
1796 if (!ik->is_final()) {
1797 // %%% should use stronger assert_unique_concrete_subtype instead
1798 phase->C->dependencies()->assert_leaf_type(ik);
1799 }
1800 // Return precise klass
1801 return TypeKlassPtr::make(ik);
1802 }
1804 // Return root of possible klass
1805 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1806 }
1807 }
1809 // Check for loading klass from an array
1810 const TypeAryPtr *tary = tp->isa_aryptr();
1811 if( tary != NULL ) {
1812 ciKlass *tary_klass = tary->klass();
1813 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1814 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1815 if (tary->klass_is_exact()) {
1816 return TypeKlassPtr::make(tary_klass);
1817 }
1818 ciArrayKlass *ak = tary->klass()->as_array_klass();
1819 // If the klass is an object array, we defer the question to the
1820 // array component klass.
1821 if( ak->is_obj_array_klass() ) {
1822 assert( ak->is_loaded(), "" );
1823 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1824 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1825 ciInstanceKlass* ik = base_k->as_instance_klass();
1826 // See if we can become precise: no subklasses and no interface
1827 if (!ik->is_interface() && !ik->has_subklass()) {
1828 //assert(!UseExactTypes, "this code should be useless with exact types");
1829 // Add a dependence; if any subclass added we need to recompile
1830 if (!ik->is_final()) {
1831 phase->C->dependencies()->assert_leaf_type(ik);
1832 }
1833 // Return precise array klass
1834 return TypeKlassPtr::make(ak);
1835 }
1836 }
1837 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1838 } else { // Found a type-array?
1839 //assert(!UseExactTypes, "this code should be useless with exact types");
1840 assert( ak->is_type_array_klass(), "" );
1841 return TypeKlassPtr::make(ak); // These are always precise
1842 }
1843 }
1844 }
1846 // Check for loading klass from an array klass
1847 const TypeKlassPtr *tkls = tp->isa_klassptr();
1848 if (tkls != NULL && !StressReflectiveCode) {
1849 ciKlass* klass = tkls->klass();
1850 if( !klass->is_loaded() )
1851 return _type; // Bail out if not loaded
1852 if( klass->is_obj_array_klass() &&
1853 (uint)tkls->offset() == objArrayKlass::element_klass_offset_in_bytes() + sizeof(oopDesc)) {
1854 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
1855 // // Always returning precise element type is incorrect,
1856 // // e.g., element type could be object and array may contain strings
1857 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
1859 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
1860 // according to the element type's subclassing.
1861 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
1862 }
1863 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
1864 (uint)tkls->offset() == Klass::super_offset_in_bytes() + sizeof(oopDesc)) {
1865 ciKlass* sup = klass->as_instance_klass()->super();
1866 // The field is Klass::_super. Return its (constant) value.
1867 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
1868 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
1869 }
1870 }
1872 // Bailout case
1873 return LoadNode::Value(phase);
1874 }
1876 //------------------------------Identity---------------------------------------
1877 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
1878 // Also feed through the klass in Allocate(...klass...)._klass.
1879 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
1880 return klass_identity_common(phase);
1881 }
1883 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
1884 Node* x = LoadNode::Identity(phase);
1885 if (x != this) return x;
1887 // Take apart the address into an oop and and offset.
1888 // Return 'this' if we cannot.
1889 Node* adr = in(MemNode::Address);
1890 intptr_t offset = 0;
1891 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1892 if (base == NULL) return this;
1893 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
1894 if (toop == NULL) return this;
1896 // We can fetch the klass directly through an AllocateNode.
1897 // This works even if the klass is not constant (clone or newArray).
1898 if (offset == oopDesc::klass_offset_in_bytes()) {
1899 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
1900 if (allocated_klass != NULL) {
1901 return allocated_klass;
1902 }
1903 }
1905 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
1906 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
1907 // See inline_native_Class_query for occurrences of these patterns.
1908 // Java Example: x.getClass().isAssignableFrom(y)
1909 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
1910 //
1911 // This improves reflective code, often making the Class
1912 // mirror go completely dead. (Current exception: Class
1913 // mirrors may appear in debug info, but we could clean them out by
1914 // introducing a new debug info operator for klassOop.java_mirror).
1915 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
1916 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1917 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1918 // We are loading a special hidden field from a Class mirror,
1919 // the field which points to its Klass or arrayKlass metaobject.
1920 if (base->is_Load()) {
1921 Node* adr2 = base->in(MemNode::Address);
1922 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1923 if (tkls != NULL && !tkls->empty()
1924 && (tkls->klass()->is_instance_klass() ||
1925 tkls->klass()->is_array_klass())
1926 && adr2->is_AddP()
1927 ) {
1928 int mirror_field = Klass::java_mirror_offset_in_bytes();
1929 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1930 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
1931 }
1932 if (tkls->offset() == mirror_field + (int)sizeof(oopDesc)) {
1933 return adr2->in(AddPNode::Base);
1934 }
1935 }
1936 }
1937 }
1939 return this;
1940 }
1943 //------------------------------Value------------------------------------------
1944 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
1945 const Type *t = klass_value_common(phase);
1946 if (t == Type::TOP)
1947 return t;
1949 return t->make_narrowoop();
1950 }
1952 //------------------------------Identity---------------------------------------
1953 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
1954 // Also feed through the klass in Allocate(...klass...)._klass.
1955 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
1956 Node *x = klass_identity_common(phase);
1958 const Type *t = phase->type( x );
1959 if( t == Type::TOP ) return x;
1960 if( t->isa_narrowoop()) return x;
1962 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop()));
1963 }
1965 //------------------------------Value-----------------------------------------
1966 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
1967 // Either input is TOP ==> the result is TOP
1968 const Type *t1 = phase->type( in(MemNode::Memory) );
1969 if( t1 == Type::TOP ) return Type::TOP;
1970 Node *adr = in(MemNode::Address);
1971 const Type *t2 = phase->type( adr );
1972 if( t2 == Type::TOP ) return Type::TOP;
1973 const TypePtr *tp = t2->is_ptr();
1974 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
1975 const TypeAryPtr *tap = tp->isa_aryptr();
1976 if( !tap ) return _type;
1977 return tap->size();
1978 }
1980 //-------------------------------Ideal---------------------------------------
1981 // Feed through the length in AllocateArray(...length...)._length.
1982 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1983 Node* p = MemNode::Ideal_common(phase, can_reshape);
1984 if (p) return (p == NodeSentinel) ? NULL : p;
1986 // Take apart the address into an oop and and offset.
1987 // Return 'this' if we cannot.
1988 Node* adr = in(MemNode::Address);
1989 intptr_t offset = 0;
1990 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1991 if (base == NULL) return NULL;
1992 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
1993 if (tary == NULL) return NULL;
1995 // We can fetch the length directly through an AllocateArrayNode.
1996 // This works even if the length is not constant (clone or newArray).
1997 if (offset == arrayOopDesc::length_offset_in_bytes()) {
1998 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
1999 if (alloc != NULL) {
2000 Node* allocated_length = alloc->Ideal_length();
2001 Node* len = alloc->make_ideal_length(tary, phase);
2002 if (allocated_length != len) {
2003 // New CastII improves on this.
2004 return len;
2005 }
2006 }
2007 }
2009 return NULL;
2010 }
2012 //------------------------------Identity---------------------------------------
2013 // Feed through the length in AllocateArray(...length...)._length.
2014 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
2015 Node* x = LoadINode::Identity(phase);
2016 if (x != this) return x;
2018 // Take apart the address into an oop and and offset.
2019 // Return 'this' if we cannot.
2020 Node* adr = in(MemNode::Address);
2021 intptr_t offset = 0;
2022 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2023 if (base == NULL) return this;
2024 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2025 if (tary == NULL) return this;
2027 // We can fetch the length directly through an AllocateArrayNode.
2028 // This works even if the length is not constant (clone or newArray).
2029 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2030 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2031 if (alloc != NULL) {
2032 Node* allocated_length = alloc->Ideal_length();
2033 // Do not allow make_ideal_length to allocate a CastII node.
2034 Node* len = alloc->make_ideal_length(tary, phase, false);
2035 if (allocated_length == len) {
2036 // Return allocated_length only if it would not be improved by a CastII.
2037 return allocated_length;
2038 }
2039 }
2040 }
2042 return this;
2044 }
2046 //=============================================================================
2047 //---------------------------StoreNode::make-----------------------------------
2048 // Polymorphic factory method:
2049 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
2050 Compile* C = gvn.C;
2052 switch (bt) {
2053 case T_BOOLEAN:
2054 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
2055 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
2056 case T_CHAR:
2057 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
2058 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
2059 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
2060 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
2061 case T_ADDRESS:
2062 case T_OBJECT:
2063 #ifdef _LP64
2064 if (adr->bottom_type()->is_ptr_to_narrowoop() ||
2065 (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
2066 adr->bottom_type()->isa_rawptr())) {
2067 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop()));
2068 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val);
2069 } else
2070 #endif
2071 {
2072 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
2073 }
2074 }
2075 ShouldNotReachHere();
2076 return (StoreNode*)NULL;
2077 }
2079 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
2080 bool require_atomic = true;
2081 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
2082 }
2085 //--------------------------bottom_type----------------------------------------
2086 const Type *StoreNode::bottom_type() const {
2087 return Type::MEMORY;
2088 }
2090 //------------------------------hash-------------------------------------------
2091 uint StoreNode::hash() const {
2092 // unroll addition of interesting fields
2093 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2095 // Since they are not commoned, do not hash them:
2096 return NO_HASH;
2097 }
2099 //------------------------------Ideal------------------------------------------
2100 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2101 // When a store immediately follows a relevant allocation/initialization,
2102 // try to capture it into the initialization, or hoist it above.
2103 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2104 Node* p = MemNode::Ideal_common(phase, can_reshape);
2105 if (p) return (p == NodeSentinel) ? NULL : p;
2107 Node* mem = in(MemNode::Memory);
2108 Node* address = in(MemNode::Address);
2110 // Back-to-back stores to same address? Fold em up.
2111 // Generally unsafe if I have intervening uses...
2112 if (mem->is_Store() && phase->eqv_uncast(mem->in(MemNode::Address), address)) {
2113 // Looking at a dead closed cycle of memory?
2114 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2116 assert(Opcode() == mem->Opcode() ||
2117 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
2118 "no mismatched stores, except on raw memory");
2120 if (mem->outcnt() == 1 && // check for intervening uses
2121 mem->as_Store()->memory_size() <= this->memory_size()) {
2122 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
2123 // For example, 'mem' might be the final state at a conditional return.
2124 // Or, 'mem' might be used by some node which is live at the same time
2125 // 'this' is live, which might be unschedulable. So, require exactly
2126 // ONE user, the 'this' store, until such time as we clone 'mem' for
2127 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
2128 if (can_reshape) { // (%%% is this an anachronism?)
2129 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
2130 phase->is_IterGVN());
2131 } else {
2132 // It's OK to do this in the parser, since DU info is always accurate,
2133 // and the parser always refers to nodes via SafePointNode maps.
2134 set_req(MemNode::Memory, mem->in(MemNode::Memory));
2135 }
2136 return this;
2137 }
2138 }
2140 // Capture an unaliased, unconditional, simple store into an initializer.
2141 // Or, if it is independent of the allocation, hoist it above the allocation.
2142 if (ReduceFieldZeroing && /*can_reshape &&*/
2143 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2144 InitializeNode* init = mem->in(0)->as_Initialize();
2145 intptr_t offset = init->can_capture_store(this, phase);
2146 if (offset > 0) {
2147 Node* moved = init->capture_store(this, offset, phase);
2148 // If the InitializeNode captured me, it made a raw copy of me,
2149 // and I need to disappear.
2150 if (moved != NULL) {
2151 // %%% hack to ensure that Ideal returns a new node:
2152 mem = MergeMemNode::make(phase->C, mem);
2153 return mem; // fold me away
2154 }
2155 }
2156 }
2158 return NULL; // No further progress
2159 }
2161 //------------------------------Value-----------------------------------------
2162 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2163 // Either input is TOP ==> the result is TOP
2164 const Type *t1 = phase->type( in(MemNode::Memory) );
2165 if( t1 == Type::TOP ) return Type::TOP;
2166 const Type *t2 = phase->type( in(MemNode::Address) );
2167 if( t2 == Type::TOP ) return Type::TOP;
2168 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2169 if( t3 == Type::TOP ) return Type::TOP;
2170 return Type::MEMORY;
2171 }
2173 //------------------------------Identity---------------------------------------
2174 // Remove redundant stores:
2175 // Store(m, p, Load(m, p)) changes to m.
2176 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2177 Node *StoreNode::Identity( PhaseTransform *phase ) {
2178 Node* mem = in(MemNode::Memory);
2179 Node* adr = in(MemNode::Address);
2180 Node* val = in(MemNode::ValueIn);
2182 // Load then Store? Then the Store is useless
2183 if (val->is_Load() &&
2184 phase->eqv_uncast( val->in(MemNode::Address), adr ) &&
2185 phase->eqv_uncast( val->in(MemNode::Memory ), mem ) &&
2186 val->as_Load()->store_Opcode() == Opcode()) {
2187 return mem;
2188 }
2190 // Two stores in a row of the same value?
2191 if (mem->is_Store() &&
2192 phase->eqv_uncast( mem->in(MemNode::Address), adr ) &&
2193 phase->eqv_uncast( mem->in(MemNode::ValueIn), val ) &&
2194 mem->Opcode() == Opcode()) {
2195 return mem;
2196 }
2198 // Store of zero anywhere into a freshly-allocated object?
2199 // Then the store is useless.
2200 // (It must already have been captured by the InitializeNode.)
2201 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2202 // a newly allocated object is already all-zeroes everywhere
2203 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2204 return mem;
2205 }
2207 // the store may also apply to zero-bits in an earlier object
2208 Node* prev_mem = find_previous_store(phase);
2209 // Steps (a), (b): Walk past independent stores to find an exact match.
2210 if (prev_mem != NULL) {
2211 Node* prev_val = can_see_stored_value(prev_mem, phase);
2212 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2213 // prev_val and val might differ by a cast; it would be good
2214 // to keep the more informative of the two.
2215 return mem;
2216 }
2217 }
2218 }
2220 return this;
2221 }
2223 //------------------------------match_edge-------------------------------------
2224 // Do we Match on this edge index or not? Match only memory & value
2225 uint StoreNode::match_edge(uint idx) const {
2226 return idx == MemNode::Address || idx == MemNode::ValueIn;
2227 }
2229 //------------------------------cmp--------------------------------------------
2230 // Do not common stores up together. They generally have to be split
2231 // back up anyways, so do not bother.
2232 uint StoreNode::cmp( const Node &n ) const {
2233 return (&n == this); // Always fail except on self
2234 }
2236 //------------------------------Ideal_masked_input-----------------------------
2237 // Check for a useless mask before a partial-word store
2238 // (StoreB ... (AndI valIn conIa) )
2239 // If (conIa & mask == mask) this simplifies to
2240 // (StoreB ... (valIn) )
2241 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2242 Node *val = in(MemNode::ValueIn);
2243 if( val->Opcode() == Op_AndI ) {
2244 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2245 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2246 set_req(MemNode::ValueIn, val->in(1));
2247 return this;
2248 }
2249 }
2250 return NULL;
2251 }
2254 //------------------------------Ideal_sign_extended_input----------------------
2255 // Check for useless sign-extension before a partial-word store
2256 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2257 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2258 // (StoreB ... (valIn) )
2259 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2260 Node *val = in(MemNode::ValueIn);
2261 if( val->Opcode() == Op_RShiftI ) {
2262 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2263 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2264 Node *shl = val->in(1);
2265 if( shl->Opcode() == Op_LShiftI ) {
2266 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2267 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2268 set_req(MemNode::ValueIn, shl->in(1));
2269 return this;
2270 }
2271 }
2272 }
2273 }
2274 return NULL;
2275 }
2277 //------------------------------value_never_loaded-----------------------------------
2278 // Determine whether there are any possible loads of the value stored.
2279 // For simplicity, we actually check if there are any loads from the
2280 // address stored to, not just for loads of the value stored by this node.
2281 //
2282 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2283 Node *adr = in(Address);
2284 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2285 if (adr_oop == NULL)
2286 return false;
2287 if (!adr_oop->is_known_instance_field())
2288 return false; // if not a distinct instance, there may be aliases of the address
2289 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2290 Node *use = adr->fast_out(i);
2291 int opc = use->Opcode();
2292 if (use->is_Load() || use->is_LoadStore()) {
2293 return false;
2294 }
2295 }
2296 return true;
2297 }
2299 //=============================================================================
2300 //------------------------------Ideal------------------------------------------
2301 // If the store is from an AND mask that leaves the low bits untouched, then
2302 // we can skip the AND operation. If the store is from a sign-extension
2303 // (a left shift, then right shift) we can skip both.
2304 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2305 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2306 if( progress != NULL ) return progress;
2308 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2309 if( progress != NULL ) return progress;
2311 // Finally check the default case
2312 return StoreNode::Ideal(phase, can_reshape);
2313 }
2315 //=============================================================================
2316 //------------------------------Ideal------------------------------------------
2317 // If the store is from an AND mask that leaves the low bits untouched, then
2318 // we can skip the AND operation
2319 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2320 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2321 if( progress != NULL ) return progress;
2323 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2324 if( progress != NULL ) return progress;
2326 // Finally check the default case
2327 return StoreNode::Ideal(phase, can_reshape);
2328 }
2330 //=============================================================================
2331 //------------------------------Identity---------------------------------------
2332 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2333 // No need to card mark when storing a null ptr
2334 Node* my_store = in(MemNode::OopStore);
2335 if (my_store->is_Store()) {
2336 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2337 if( t1 == TypePtr::NULL_PTR ) {
2338 return in(MemNode::Memory);
2339 }
2340 }
2341 return this;
2342 }
2344 //=============================================================================
2345 //------------------------------Ideal---------------------------------------
2346 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2347 Node* progress = StoreNode::Ideal(phase, can_reshape);
2348 if (progress != NULL) return progress;
2350 Node* my_store = in(MemNode::OopStore);
2351 if (my_store->is_MergeMem()) {
2352 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2353 set_req(MemNode::OopStore, mem);
2354 return this;
2355 }
2357 return NULL;
2358 }
2360 //------------------------------Value-----------------------------------------
2361 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2362 // Either input is TOP ==> the result is TOP
2363 const Type *t = phase->type( in(MemNode::Memory) );
2364 if( t == Type::TOP ) return Type::TOP;
2365 t = phase->type( in(MemNode::Address) );
2366 if( t == Type::TOP ) return Type::TOP;
2367 t = phase->type( in(MemNode::ValueIn) );
2368 if( t == Type::TOP ) return Type::TOP;
2369 // If extra input is TOP ==> the result is TOP
2370 t = phase->type( in(MemNode::OopStore) );
2371 if( t == Type::TOP ) return Type::TOP;
2373 return StoreNode::Value( phase );
2374 }
2377 //=============================================================================
2378 //----------------------------------SCMemProjNode------------------------------
2379 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2380 {
2381 return bottom_type();
2382 }
2384 //=============================================================================
2385 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
2386 init_req(MemNode::Control, c );
2387 init_req(MemNode::Memory , mem);
2388 init_req(MemNode::Address, adr);
2389 init_req(MemNode::ValueIn, val);
2390 init_req( ExpectedIn, ex );
2391 init_class_id(Class_LoadStore);
2393 }
2395 //=============================================================================
2396 //-------------------------------adr_type--------------------------------------
2397 // Do we Match on this edge index or not? Do not match memory
2398 const TypePtr* ClearArrayNode::adr_type() const {
2399 Node *adr = in(3);
2400 return MemNode::calculate_adr_type(adr->bottom_type());
2401 }
2403 //------------------------------match_edge-------------------------------------
2404 // Do we Match on this edge index or not? Do not match memory
2405 uint ClearArrayNode::match_edge(uint idx) const {
2406 return idx > 1;
2407 }
2409 //------------------------------Identity---------------------------------------
2410 // Clearing a zero length array does nothing
2411 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2412 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2413 }
2415 //------------------------------Idealize---------------------------------------
2416 // Clearing a short array is faster with stores
2417 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2418 const int unit = BytesPerLong;
2419 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2420 if (!t) return NULL;
2421 if (!t->is_con()) return NULL;
2422 intptr_t raw_count = t->get_con();
2423 intptr_t size = raw_count;
2424 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2425 // Clearing nothing uses the Identity call.
2426 // Negative clears are possible on dead ClearArrays
2427 // (see jck test stmt114.stmt11402.val).
2428 if (size <= 0 || size % unit != 0) return NULL;
2429 intptr_t count = size / unit;
2430 // Length too long; use fast hardware clear
2431 if (size > Matcher::init_array_short_size) return NULL;
2432 Node *mem = in(1);
2433 if( phase->type(mem)==Type::TOP ) return NULL;
2434 Node *adr = in(3);
2435 const Type* at = phase->type(adr);
2436 if( at==Type::TOP ) return NULL;
2437 const TypePtr* atp = at->isa_ptr();
2438 // adjust atp to be the correct array element address type
2439 if (atp == NULL) atp = TypePtr::BOTTOM;
2440 else atp = atp->add_offset(Type::OffsetBot);
2441 // Get base for derived pointer purposes
2442 if( adr->Opcode() != Op_AddP ) Unimplemented();
2443 Node *base = adr->in(1);
2445 Node *zero = phase->makecon(TypeLong::ZERO);
2446 Node *off = phase->MakeConX(BytesPerLong);
2447 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2448 count--;
2449 while( count-- ) {
2450 mem = phase->transform(mem);
2451 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
2452 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2453 }
2454 return mem;
2455 }
2457 //----------------------------clear_memory-------------------------------------
2458 // Generate code to initialize object storage to zero.
2459 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2460 intptr_t start_offset,
2461 Node* end_offset,
2462 PhaseGVN* phase) {
2463 Compile* C = phase->C;
2464 intptr_t offset = start_offset;
2466 int unit = BytesPerLong;
2467 if ((offset % unit) != 0) {
2468 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
2469 adr = phase->transform(adr);
2470 const TypePtr* atp = TypeRawPtr::BOTTOM;
2471 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2472 mem = phase->transform(mem);
2473 offset += BytesPerInt;
2474 }
2475 assert((offset % unit) == 0, "");
2477 // Initialize the remaining stuff, if any, with a ClearArray.
2478 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2479 }
2481 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2482 Node* start_offset,
2483 Node* end_offset,
2484 PhaseGVN* phase) {
2485 if (start_offset == end_offset) {
2486 // nothing to do
2487 return mem;
2488 }
2490 Compile* C = phase->C;
2491 int unit = BytesPerLong;
2492 Node* zbase = start_offset;
2493 Node* zend = end_offset;
2495 // Scale to the unit required by the CPU:
2496 if (!Matcher::init_array_count_is_in_bytes) {
2497 Node* shift = phase->intcon(exact_log2(unit));
2498 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
2499 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
2500 }
2502 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
2503 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
2505 // Bulk clear double-words
2506 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
2507 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
2508 return phase->transform(mem);
2509 }
2511 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2512 intptr_t start_offset,
2513 intptr_t end_offset,
2514 PhaseGVN* phase) {
2515 if (start_offset == end_offset) {
2516 // nothing to do
2517 return mem;
2518 }
2520 Compile* C = phase->C;
2521 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2522 intptr_t done_offset = end_offset;
2523 if ((done_offset % BytesPerLong) != 0) {
2524 done_offset -= BytesPerInt;
2525 }
2526 if (done_offset > start_offset) {
2527 mem = clear_memory(ctl, mem, dest,
2528 start_offset, phase->MakeConX(done_offset), phase);
2529 }
2530 if (done_offset < end_offset) { // emit the final 32-bit store
2531 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
2532 adr = phase->transform(adr);
2533 const TypePtr* atp = TypeRawPtr::BOTTOM;
2534 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2535 mem = phase->transform(mem);
2536 done_offset += BytesPerInt;
2537 }
2538 assert(done_offset == end_offset, "");
2539 return mem;
2540 }
2542 //=============================================================================
2543 // Do we match on this edge? No memory edges
2544 uint StrCompNode::match_edge(uint idx) const {
2545 return idx == 2 || idx == 3; // StrComp (Binary str1 cnt1) (Binary str2 cnt2)
2546 }
2548 //------------------------------Ideal------------------------------------------
2549 // Return a node which is more "ideal" than the current node. Strip out
2550 // control copies
2551 Node *StrCompNode::Ideal(PhaseGVN *phase, bool can_reshape){
2552 return remove_dead_region(phase, can_reshape) ? this : NULL;
2553 }
2555 //=============================================================================
2556 // Do we match on this edge? No memory edges
2557 uint StrEqualsNode::match_edge(uint idx) const {
2558 return idx == 2 || idx == 3; // StrEquals (Binary str1 str2) cnt
2559 }
2561 //------------------------------Ideal------------------------------------------
2562 // Return a node which is more "ideal" than the current node. Strip out
2563 // control copies
2564 Node *StrEqualsNode::Ideal(PhaseGVN *phase, bool can_reshape){
2565 return remove_dead_region(phase, can_reshape) ? this : NULL;
2566 }
2568 //=============================================================================
2569 // Do we match on this edge? No memory edges
2570 uint StrIndexOfNode::match_edge(uint idx) const {
2571 return idx == 2 || idx == 3; // StrIndexOf (Binary str1 cnt1) (Binary str2 cnt2)
2572 }
2574 //------------------------------Ideal------------------------------------------
2575 // Return a node which is more "ideal" than the current node. Strip out
2576 // control copies
2577 Node *StrIndexOfNode::Ideal(PhaseGVN *phase, bool can_reshape){
2578 return remove_dead_region(phase, can_reshape) ? this : NULL;
2579 }
2581 //=============================================================================
2582 // Do we match on this edge? No memory edges
2583 uint AryEqNode::match_edge(uint idx) const {
2584 return idx == 2 || idx == 3; // StrEquals ary1 ary2
2585 }
2586 //------------------------------Ideal------------------------------------------
2587 // Return a node which is more "ideal" than the current node. Strip out
2588 // control copies
2589 Node *AryEqNode::Ideal(PhaseGVN *phase, bool can_reshape){
2590 return remove_dead_region(phase, can_reshape) ? this : NULL;
2591 }
2593 //=============================================================================
2594 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2595 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2596 _adr_type(C->get_adr_type(alias_idx))
2597 {
2598 init_class_id(Class_MemBar);
2599 Node* top = C->top();
2600 init_req(TypeFunc::I_O,top);
2601 init_req(TypeFunc::FramePtr,top);
2602 init_req(TypeFunc::ReturnAdr,top);
2603 if (precedent != NULL)
2604 init_req(TypeFunc::Parms, precedent);
2605 }
2607 //------------------------------cmp--------------------------------------------
2608 uint MemBarNode::hash() const { return NO_HASH; }
2609 uint MemBarNode::cmp( const Node &n ) const {
2610 return (&n == this); // Always fail except on self
2611 }
2613 //------------------------------make-------------------------------------------
2614 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2615 int len = Precedent + (pn == NULL? 0: 1);
2616 switch (opcode) {
2617 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
2618 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
2619 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
2620 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
2621 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
2622 default: ShouldNotReachHere(); return NULL;
2623 }
2624 }
2626 //------------------------------Ideal------------------------------------------
2627 // Return a node which is more "ideal" than the current node. Strip out
2628 // control copies
2629 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2630 return remove_dead_region(phase, can_reshape) ? this : NULL;
2631 }
2633 //------------------------------Value------------------------------------------
2634 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2635 if( !in(0) ) return Type::TOP;
2636 if( phase->type(in(0)) == Type::TOP )
2637 return Type::TOP;
2638 return TypeTuple::MEMBAR;
2639 }
2641 //------------------------------match------------------------------------------
2642 // Construct projections for memory.
2643 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2644 switch (proj->_con) {
2645 case TypeFunc::Control:
2646 case TypeFunc::Memory:
2647 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2648 }
2649 ShouldNotReachHere();
2650 return NULL;
2651 }
2653 //===========================InitializeNode====================================
2654 // SUMMARY:
2655 // This node acts as a memory barrier on raw memory, after some raw stores.
2656 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2657 // The Initialize can 'capture' suitably constrained stores as raw inits.
2658 // It can coalesce related raw stores into larger units (called 'tiles').
2659 // It can avoid zeroing new storage for memory units which have raw inits.
2660 // At macro-expansion, it is marked 'complete', and does not optimize further.
2661 //
2662 // EXAMPLE:
2663 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2664 // ctl = incoming control; mem* = incoming memory
2665 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
2666 // First allocate uninitialized memory and fill in the header:
2667 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
2668 // ctl := alloc.Control; mem* := alloc.Memory*
2669 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
2670 // Then initialize to zero the non-header parts of the raw memory block:
2671 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2672 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2673 // After the initialize node executes, the object is ready for service:
2674 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2675 // Suppose its body is immediately initialized as {1,2}:
2676 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2677 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2678 // mem.SLICE(#short[*]) := store2
2679 //
2680 // DETAILS:
2681 // An InitializeNode collects and isolates object initialization after
2682 // an AllocateNode and before the next possible safepoint. As a
2683 // memory barrier (MemBarNode), it keeps critical stores from drifting
2684 // down past any safepoint or any publication of the allocation.
2685 // Before this barrier, a newly-allocated object may have uninitialized bits.
2686 // After this barrier, it may be treated as a real oop, and GC is allowed.
2687 //
2688 // The semantics of the InitializeNode include an implicit zeroing of
2689 // the new object from object header to the end of the object.
2690 // (The object header and end are determined by the AllocateNode.)
2691 //
2692 // Certain stores may be added as direct inputs to the InitializeNode.
2693 // These stores must update raw memory, and they must be to addresses
2694 // derived from the raw address produced by AllocateNode, and with
2695 // a constant offset. They must be ordered by increasing offset.
2696 // The first one is at in(RawStores), the last at in(req()-1).
2697 // Unlike most memory operations, they are not linked in a chain,
2698 // but are displayed in parallel as users of the rawmem output of
2699 // the allocation.
2700 //
2701 // (See comments in InitializeNode::capture_store, which continue
2702 // the example given above.)
2703 //
2704 // When the associated Allocate is macro-expanded, the InitializeNode
2705 // may be rewritten to optimize collected stores. A ClearArrayNode
2706 // may also be created at that point to represent any required zeroing.
2707 // The InitializeNode is then marked 'complete', prohibiting further
2708 // capturing of nearby memory operations.
2709 //
2710 // During macro-expansion, all captured initializations which store
2711 // constant values of 32 bits or smaller are coalesced (if advantageous)
2712 // into larger 'tiles' 32 or 64 bits. This allows an object to be
2713 // initialized in fewer memory operations. Memory words which are
2714 // covered by neither tiles nor non-constant stores are pre-zeroed
2715 // by explicit stores of zero. (The code shape happens to do all
2716 // zeroing first, then all other stores, with both sequences occurring
2717 // in order of ascending offsets.)
2718 //
2719 // Alternatively, code may be inserted between an AllocateNode and its
2720 // InitializeNode, to perform arbitrary initialization of the new object.
2721 // E.g., the object copying intrinsics insert complex data transfers here.
2722 // The initialization must then be marked as 'complete' disable the
2723 // built-in zeroing semantics and the collection of initializing stores.
2724 //
2725 // While an InitializeNode is incomplete, reads from the memory state
2726 // produced by it are optimizable if they match the control edge and
2727 // new oop address associated with the allocation/initialization.
2728 // They return a stored value (if the offset matches) or else zero.
2729 // A write to the memory state, if it matches control and address,
2730 // and if it is to a constant offset, may be 'captured' by the
2731 // InitializeNode. It is cloned as a raw memory operation and rewired
2732 // inside the initialization, to the raw oop produced by the allocation.
2733 // Operations on addresses which are provably distinct (e.g., to
2734 // other AllocateNodes) are allowed to bypass the initialization.
2735 //
2736 // The effect of all this is to consolidate object initialization
2737 // (both arrays and non-arrays, both piecewise and bulk) into a
2738 // single location, where it can be optimized as a unit.
2739 //
2740 // Only stores with an offset less than TrackedInitializationLimit words
2741 // will be considered for capture by an InitializeNode. This puts a
2742 // reasonable limit on the complexity of optimized initializations.
2744 //---------------------------InitializeNode------------------------------------
2745 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
2746 : _is_complete(false),
2747 MemBarNode(C, adr_type, rawoop)
2748 {
2749 init_class_id(Class_Initialize);
2751 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
2752 assert(in(RawAddress) == rawoop, "proper init");
2753 // Note: allocation() can be NULL, for secondary initialization barriers
2754 }
2756 // Since this node is not matched, it will be processed by the
2757 // register allocator. Declare that there are no constraints
2758 // on the allocation of the RawAddress edge.
2759 const RegMask &InitializeNode::in_RegMask(uint idx) const {
2760 // This edge should be set to top, by the set_complete. But be conservative.
2761 if (idx == InitializeNode::RawAddress)
2762 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
2763 return RegMask::Empty;
2764 }
2766 Node* InitializeNode::memory(uint alias_idx) {
2767 Node* mem = in(Memory);
2768 if (mem->is_MergeMem()) {
2769 return mem->as_MergeMem()->memory_at(alias_idx);
2770 } else {
2771 // incoming raw memory is not split
2772 return mem;
2773 }
2774 }
2776 bool InitializeNode::is_non_zero() {
2777 if (is_complete()) return false;
2778 remove_extra_zeroes();
2779 return (req() > RawStores);
2780 }
2782 void InitializeNode::set_complete(PhaseGVN* phase) {
2783 assert(!is_complete(), "caller responsibility");
2784 _is_complete = true;
2786 // After this node is complete, it contains a bunch of
2787 // raw-memory initializations. There is no need for
2788 // it to have anything to do with non-raw memory effects.
2789 // Therefore, tell all non-raw users to re-optimize themselves,
2790 // after skipping the memory effects of this initialization.
2791 PhaseIterGVN* igvn = phase->is_IterGVN();
2792 if (igvn) igvn->add_users_to_worklist(this);
2793 }
2795 // convenience function
2796 // return false if the init contains any stores already
2797 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
2798 InitializeNode* init = initialization();
2799 if (init == NULL || init->is_complete()) return false;
2800 init->remove_extra_zeroes();
2801 // for now, if this allocation has already collected any inits, bail:
2802 if (init->is_non_zero()) return false;
2803 init->set_complete(phase);
2804 return true;
2805 }
2807 void InitializeNode::remove_extra_zeroes() {
2808 if (req() == RawStores) return;
2809 Node* zmem = zero_memory();
2810 uint fill = RawStores;
2811 for (uint i = fill; i < req(); i++) {
2812 Node* n = in(i);
2813 if (n->is_top() || n == zmem) continue; // skip
2814 if (fill < i) set_req(fill, n); // compact
2815 ++fill;
2816 }
2817 // delete any empty spaces created:
2818 while (fill < req()) {
2819 del_req(fill);
2820 }
2821 }
2823 // Helper for remembering which stores go with which offsets.
2824 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
2825 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
2826 intptr_t offset = -1;
2827 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
2828 phase, offset);
2829 if (base == NULL) return -1; // something is dead,
2830 if (offset < 0) return -1; // dead, dead
2831 return offset;
2832 }
2834 // Helper for proving that an initialization expression is
2835 // "simple enough" to be folded into an object initialization.
2836 // Attempts to prove that a store's initial value 'n' can be captured
2837 // within the initialization without creating a vicious cycle, such as:
2838 // { Foo p = new Foo(); p.next = p; }
2839 // True for constants and parameters and small combinations thereof.
2840 bool InitializeNode::detect_init_independence(Node* n,
2841 bool st_is_pinned,
2842 int& count) {
2843 if (n == NULL) return true; // (can this really happen?)
2844 if (n->is_Proj()) n = n->in(0);
2845 if (n == this) return false; // found a cycle
2846 if (n->is_Con()) return true;
2847 if (n->is_Start()) return true; // params, etc., are OK
2848 if (n->is_Root()) return true; // even better
2850 Node* ctl = n->in(0);
2851 if (ctl != NULL && !ctl->is_top()) {
2852 if (ctl->is_Proj()) ctl = ctl->in(0);
2853 if (ctl == this) return false;
2855 // If we already know that the enclosing memory op is pinned right after
2856 // the init, then any control flow that the store has picked up
2857 // must have preceded the init, or else be equal to the init.
2858 // Even after loop optimizations (which might change control edges)
2859 // a store is never pinned *before* the availability of its inputs.
2860 if (!MemNode::all_controls_dominate(n, this))
2861 return false; // failed to prove a good control
2863 }
2865 // Check data edges for possible dependencies on 'this'.
2866 if ((count += 1) > 20) return false; // complexity limit
2867 for (uint i = 1; i < n->req(); i++) {
2868 Node* m = n->in(i);
2869 if (m == NULL || m == n || m->is_top()) continue;
2870 uint first_i = n->find_edge(m);
2871 if (i != first_i) continue; // process duplicate edge just once
2872 if (!detect_init_independence(m, st_is_pinned, count)) {
2873 return false;
2874 }
2875 }
2877 return true;
2878 }
2880 // Here are all the checks a Store must pass before it can be moved into
2881 // an initialization. Returns zero if a check fails.
2882 // On success, returns the (constant) offset to which the store applies,
2883 // within the initialized memory.
2884 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
2885 const int FAIL = 0;
2886 if (st->req() != MemNode::ValueIn + 1)
2887 return FAIL; // an inscrutable StoreNode (card mark?)
2888 Node* ctl = st->in(MemNode::Control);
2889 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
2890 return FAIL; // must be unconditional after the initialization
2891 Node* mem = st->in(MemNode::Memory);
2892 if (!(mem->is_Proj() && mem->in(0) == this))
2893 return FAIL; // must not be preceded by other stores
2894 Node* adr = st->in(MemNode::Address);
2895 intptr_t offset;
2896 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
2897 if (alloc == NULL)
2898 return FAIL; // inscrutable address
2899 if (alloc != allocation())
2900 return FAIL; // wrong allocation! (store needs to float up)
2901 Node* val = st->in(MemNode::ValueIn);
2902 int complexity_count = 0;
2903 if (!detect_init_independence(val, true, complexity_count))
2904 return FAIL; // stored value must be 'simple enough'
2906 return offset; // success
2907 }
2909 // Find the captured store in(i) which corresponds to the range
2910 // [start..start+size) in the initialized object.
2911 // If there is one, return its index i. If there isn't, return the
2912 // negative of the index where it should be inserted.
2913 // Return 0 if the queried range overlaps an initialization boundary
2914 // or if dead code is encountered.
2915 // If size_in_bytes is zero, do not bother with overlap checks.
2916 int InitializeNode::captured_store_insertion_point(intptr_t start,
2917 int size_in_bytes,
2918 PhaseTransform* phase) {
2919 const int FAIL = 0, MAX_STORE = BytesPerLong;
2921 if (is_complete())
2922 return FAIL; // arraycopy got here first; punt
2924 assert(allocation() != NULL, "must be present");
2926 // no negatives, no header fields:
2927 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
2929 // after a certain size, we bail out on tracking all the stores:
2930 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
2931 if (start >= ti_limit) return FAIL;
2933 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
2934 if (i >= limit) return -(int)i; // not found; here is where to put it
2936 Node* st = in(i);
2937 intptr_t st_off = get_store_offset(st, phase);
2938 if (st_off < 0) {
2939 if (st != zero_memory()) {
2940 return FAIL; // bail out if there is dead garbage
2941 }
2942 } else if (st_off > start) {
2943 // ...we are done, since stores are ordered
2944 if (st_off < start + size_in_bytes) {
2945 return FAIL; // the next store overlaps
2946 }
2947 return -(int)i; // not found; here is where to put it
2948 } else if (st_off < start) {
2949 if (size_in_bytes != 0 &&
2950 start < st_off + MAX_STORE &&
2951 start < st_off + st->as_Store()->memory_size()) {
2952 return FAIL; // the previous store overlaps
2953 }
2954 } else {
2955 if (size_in_bytes != 0 &&
2956 st->as_Store()->memory_size() != size_in_bytes) {
2957 return FAIL; // mismatched store size
2958 }
2959 return i;
2960 }
2962 ++i;
2963 }
2964 }
2966 // Look for a captured store which initializes at the offset 'start'
2967 // with the given size. If there is no such store, and no other
2968 // initialization interferes, then return zero_memory (the memory
2969 // projection of the AllocateNode).
2970 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
2971 PhaseTransform* phase) {
2972 assert(stores_are_sane(phase), "");
2973 int i = captured_store_insertion_point(start, size_in_bytes, phase);
2974 if (i == 0) {
2975 return NULL; // something is dead
2976 } else if (i < 0) {
2977 return zero_memory(); // just primordial zero bits here
2978 } else {
2979 Node* st = in(i); // here is the store at this position
2980 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
2981 return st;
2982 }
2983 }
2985 // Create, as a raw pointer, an address within my new object at 'offset'.
2986 Node* InitializeNode::make_raw_address(intptr_t offset,
2987 PhaseTransform* phase) {
2988 Node* addr = in(RawAddress);
2989 if (offset != 0) {
2990 Compile* C = phase->C;
2991 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
2992 phase->MakeConX(offset)) );
2993 }
2994 return addr;
2995 }
2997 // Clone the given store, converting it into a raw store
2998 // initializing a field or element of my new object.
2999 // Caller is responsible for retiring the original store,
3000 // with subsume_node or the like.
3001 //
3002 // From the example above InitializeNode::InitializeNode,
3003 // here are the old stores to be captured:
3004 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3005 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3006 //
3007 // Here is the changed code; note the extra edges on init:
3008 // alloc = (Allocate ...)
3009 // rawoop = alloc.RawAddress
3010 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3011 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3012 // init = (Initialize alloc.Control alloc.Memory rawoop
3013 // rawstore1 rawstore2)
3014 //
3015 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3016 PhaseTransform* phase) {
3017 assert(stores_are_sane(phase), "");
3019 if (start < 0) return NULL;
3020 assert(can_capture_store(st, phase) == start, "sanity");
3022 Compile* C = phase->C;
3023 int size_in_bytes = st->memory_size();
3024 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3025 if (i == 0) return NULL; // bail out
3026 Node* prev_mem = NULL; // raw memory for the captured store
3027 if (i > 0) {
3028 prev_mem = in(i); // there is a pre-existing store under this one
3029 set_req(i, C->top()); // temporarily disconnect it
3030 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3031 } else {
3032 i = -i; // no pre-existing store
3033 prev_mem = zero_memory(); // a slice of the newly allocated object
3034 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3035 set_req(--i, C->top()); // reuse this edge; it has been folded away
3036 else
3037 ins_req(i, C->top()); // build a new edge
3038 }
3039 Node* new_st = st->clone();
3040 new_st->set_req(MemNode::Control, in(Control));
3041 new_st->set_req(MemNode::Memory, prev_mem);
3042 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3043 new_st = phase->transform(new_st);
3045 // At this point, new_st might have swallowed a pre-existing store
3046 // at the same offset, or perhaps new_st might have disappeared,
3047 // if it redundantly stored the same value (or zero to fresh memory).
3049 // In any case, wire it in:
3050 set_req(i, new_st);
3052 // The caller may now kill the old guy.
3053 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3054 assert(check_st == new_st || check_st == NULL, "must be findable");
3055 assert(!is_complete(), "");
3056 return new_st;
3057 }
3059 static bool store_constant(jlong* tiles, int num_tiles,
3060 intptr_t st_off, int st_size,
3061 jlong con) {
3062 if ((st_off & (st_size-1)) != 0)
3063 return false; // strange store offset (assume size==2**N)
3064 address addr = (address)tiles + st_off;
3065 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3066 switch (st_size) {
3067 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3068 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3069 case sizeof(jint): *(jint*) addr = (jint) con; break;
3070 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3071 default: return false; // strange store size (detect size!=2**N here)
3072 }
3073 return true; // return success to caller
3074 }
3076 // Coalesce subword constants into int constants and possibly
3077 // into long constants. The goal, if the CPU permits,
3078 // is to initialize the object with a small number of 64-bit tiles.
3079 // Also, convert floating-point constants to bit patterns.
3080 // Non-constants are not relevant to this pass.
3081 //
3082 // In terms of the running example on InitializeNode::InitializeNode
3083 // and InitializeNode::capture_store, here is the transformation
3084 // of rawstore1 and rawstore2 into rawstore12:
3085 // alloc = (Allocate ...)
3086 // rawoop = alloc.RawAddress
3087 // tile12 = 0x00010002
3088 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3089 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3090 //
3091 void
3092 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3093 Node* size_in_bytes,
3094 PhaseGVN* phase) {
3095 Compile* C = phase->C;
3097 assert(stores_are_sane(phase), "");
3098 // Note: After this pass, they are not completely sane,
3099 // since there may be some overlaps.
3101 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3103 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3104 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3105 size_limit = MIN2(size_limit, ti_limit);
3106 size_limit = align_size_up(size_limit, BytesPerLong);
3107 int num_tiles = size_limit / BytesPerLong;
3109 // allocate space for the tile map:
3110 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3111 jlong tiles_buf[small_len];
3112 Node* nodes_buf[small_len];
3113 jlong inits_buf[small_len];
3114 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3115 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3116 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3117 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3118 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3119 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3120 // tiles: exact bitwise model of all primitive constants
3121 // nodes: last constant-storing node subsumed into the tiles model
3122 // inits: which bytes (in each tile) are touched by any initializations
3124 //// Pass A: Fill in the tile model with any relevant stores.
3126 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3127 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3128 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3129 Node* zmem = zero_memory(); // initially zero memory state
3130 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3131 Node* st = in(i);
3132 intptr_t st_off = get_store_offset(st, phase);
3134 // Figure out the store's offset and constant value:
3135 if (st_off < header_size) continue; //skip (ignore header)
3136 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3137 int st_size = st->as_Store()->memory_size();
3138 if (st_off + st_size > size_limit) break;
3140 // Record which bytes are touched, whether by constant or not.
3141 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3142 continue; // skip (strange store size)
3144 const Type* val = phase->type(st->in(MemNode::ValueIn));
3145 if (!val->singleton()) continue; //skip (non-con store)
3146 BasicType type = val->basic_type();
3148 jlong con = 0;
3149 switch (type) {
3150 case T_INT: con = val->is_int()->get_con(); break;
3151 case T_LONG: con = val->is_long()->get_con(); break;
3152 case T_FLOAT: con = jint_cast(val->getf()); break;
3153 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3154 default: continue; //skip (odd store type)
3155 }
3157 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3158 st->Opcode() == Op_StoreL) {
3159 continue; // This StoreL is already optimal.
3160 }
3162 // Store down the constant.
3163 store_constant(tiles, num_tiles, st_off, st_size, con);
3165 intptr_t j = st_off >> LogBytesPerLong;
3167 if (type == T_INT && st_size == BytesPerInt
3168 && (st_off & BytesPerInt) == BytesPerInt) {
3169 jlong lcon = tiles[j];
3170 if (!Matcher::isSimpleConstant64(lcon) &&
3171 st->Opcode() == Op_StoreI) {
3172 // This StoreI is already optimal by itself.
3173 jint* intcon = (jint*) &tiles[j];
3174 intcon[1] = 0; // undo the store_constant()
3176 // If the previous store is also optimal by itself, back up and
3177 // undo the action of the previous loop iteration... if we can.
3178 // But if we can't, just let the previous half take care of itself.
3179 st = nodes[j];
3180 st_off -= BytesPerInt;
3181 con = intcon[0];
3182 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3183 assert(st_off >= header_size, "still ignoring header");
3184 assert(get_store_offset(st, phase) == st_off, "must be");
3185 assert(in(i-1) == zmem, "must be");
3186 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3187 assert(con == tcon->is_int()->get_con(), "must be");
3188 // Undo the effects of the previous loop trip, which swallowed st:
3189 intcon[0] = 0; // undo store_constant()
3190 set_req(i-1, st); // undo set_req(i, zmem)
3191 nodes[j] = NULL; // undo nodes[j] = st
3192 --old_subword; // undo ++old_subword
3193 }
3194 continue; // This StoreI is already optimal.
3195 }
3196 }
3198 // This store is not needed.
3199 set_req(i, zmem);
3200 nodes[j] = st; // record for the moment
3201 if (st_size < BytesPerLong) // something has changed
3202 ++old_subword; // includes int/float, but who's counting...
3203 else ++old_long;
3204 }
3206 if ((old_subword + old_long) == 0)
3207 return; // nothing more to do
3209 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3210 // Be sure to insert them before overlapping non-constant stores.
3211 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3212 for (int j = 0; j < num_tiles; j++) {
3213 jlong con = tiles[j];
3214 jlong init = inits[j];
3215 if (con == 0) continue;
3216 jint con0, con1; // split the constant, address-wise
3217 jint init0, init1; // split the init map, address-wise
3218 { union { jlong con; jint intcon[2]; } u;
3219 u.con = con;
3220 con0 = u.intcon[0];
3221 con1 = u.intcon[1];
3222 u.con = init;
3223 init0 = u.intcon[0];
3224 init1 = u.intcon[1];
3225 }
3227 Node* old = nodes[j];
3228 assert(old != NULL, "need the prior store");
3229 intptr_t offset = (j * BytesPerLong);
3231 bool split = !Matcher::isSimpleConstant64(con);
3233 if (offset < header_size) {
3234 assert(offset + BytesPerInt >= header_size, "second int counts");
3235 assert(*(jint*)&tiles[j] == 0, "junk in header");
3236 split = true; // only the second word counts
3237 // Example: int a[] = { 42 ... }
3238 } else if (con0 == 0 && init0 == -1) {
3239 split = true; // first word is covered by full inits
3240 // Example: int a[] = { ... foo(), 42 ... }
3241 } else if (con1 == 0 && init1 == -1) {
3242 split = true; // second word is covered by full inits
3243 // Example: int a[] = { ... 42, foo() ... }
3244 }
3246 // Here's a case where init0 is neither 0 nor -1:
3247 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3248 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3249 // In this case the tile is not split; it is (jlong)42.
3250 // The big tile is stored down, and then the foo() value is inserted.
3251 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3253 Node* ctl = old->in(MemNode::Control);
3254 Node* adr = make_raw_address(offset, phase);
3255 const TypePtr* atp = TypeRawPtr::BOTTOM;
3257 // One or two coalesced stores to plop down.
3258 Node* st[2];
3259 intptr_t off[2];
3260 int nst = 0;
3261 if (!split) {
3262 ++new_long;
3263 off[nst] = offset;
3264 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3265 phase->longcon(con), T_LONG);
3266 } else {
3267 // Omit either if it is a zero.
3268 if (con0 != 0) {
3269 ++new_int;
3270 off[nst] = offset;
3271 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3272 phase->intcon(con0), T_INT);
3273 }
3274 if (con1 != 0) {
3275 ++new_int;
3276 offset += BytesPerInt;
3277 adr = make_raw_address(offset, phase);
3278 off[nst] = offset;
3279 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3280 phase->intcon(con1), T_INT);
3281 }
3282 }
3284 // Insert second store first, then the first before the second.
3285 // Insert each one just before any overlapping non-constant stores.
3286 while (nst > 0) {
3287 Node* st1 = st[--nst];
3288 C->copy_node_notes_to(st1, old);
3289 st1 = phase->transform(st1);
3290 offset = off[nst];
3291 assert(offset >= header_size, "do not smash header");
3292 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3293 guarantee(ins_idx != 0, "must re-insert constant store");
3294 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3295 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3296 set_req(--ins_idx, st1);
3297 else
3298 ins_req(ins_idx, st1);
3299 }
3300 }
3302 if (PrintCompilation && WizardMode)
3303 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3304 old_subword, old_long, new_int, new_long);
3305 if (C->log() != NULL)
3306 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3307 old_subword, old_long, new_int, new_long);
3309 // Clean up any remaining occurrences of zmem:
3310 remove_extra_zeroes();
3311 }
3313 // Explore forward from in(start) to find the first fully initialized
3314 // word, and return its offset. Skip groups of subword stores which
3315 // together initialize full words. If in(start) is itself part of a
3316 // fully initialized word, return the offset of in(start). If there
3317 // are no following full-word stores, or if something is fishy, return
3318 // a negative value.
3319 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3320 int int_map = 0;
3321 intptr_t int_map_off = 0;
3322 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3324 for (uint i = start, limit = req(); i < limit; i++) {
3325 Node* st = in(i);
3327 intptr_t st_off = get_store_offset(st, phase);
3328 if (st_off < 0) break; // return conservative answer
3330 int st_size = st->as_Store()->memory_size();
3331 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3332 return st_off; // we found a complete word init
3333 }
3335 // update the map:
3337 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3338 if (this_int_off != int_map_off) {
3339 // reset the map:
3340 int_map = 0;
3341 int_map_off = this_int_off;
3342 }
3344 int subword_off = st_off - this_int_off;
3345 int_map |= right_n_bits(st_size) << subword_off;
3346 if ((int_map & FULL_MAP) == FULL_MAP) {
3347 return this_int_off; // we found a complete word init
3348 }
3350 // Did this store hit or cross the word boundary?
3351 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3352 if (next_int_off == this_int_off + BytesPerInt) {
3353 // We passed the current int, without fully initializing it.
3354 int_map_off = next_int_off;
3355 int_map >>= BytesPerInt;
3356 } else if (next_int_off > this_int_off + BytesPerInt) {
3357 // We passed the current and next int.
3358 return this_int_off + BytesPerInt;
3359 }
3360 }
3362 return -1;
3363 }
3366 // Called when the associated AllocateNode is expanded into CFG.
3367 // At this point, we may perform additional optimizations.
3368 // Linearize the stores by ascending offset, to make memory
3369 // activity as coherent as possible.
3370 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3371 intptr_t header_size,
3372 Node* size_in_bytes,
3373 PhaseGVN* phase) {
3374 assert(!is_complete(), "not already complete");
3375 assert(stores_are_sane(phase), "");
3376 assert(allocation() != NULL, "must be present");
3378 remove_extra_zeroes();
3380 if (ReduceFieldZeroing || ReduceBulkZeroing)
3381 // reduce instruction count for common initialization patterns
3382 coalesce_subword_stores(header_size, size_in_bytes, phase);
3384 Node* zmem = zero_memory(); // initially zero memory state
3385 Node* inits = zmem; // accumulating a linearized chain of inits
3386 #ifdef ASSERT
3387 intptr_t first_offset = allocation()->minimum_header_size();
3388 intptr_t last_init_off = first_offset; // previous init offset
3389 intptr_t last_init_end = first_offset; // previous init offset+size
3390 intptr_t last_tile_end = first_offset; // previous tile offset+size
3391 #endif
3392 intptr_t zeroes_done = header_size;
3394 bool do_zeroing = true; // we might give up if inits are very sparse
3395 int big_init_gaps = 0; // how many large gaps have we seen?
3397 if (ZeroTLAB) do_zeroing = false;
3398 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3400 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3401 Node* st = in(i);
3402 intptr_t st_off = get_store_offset(st, phase);
3403 if (st_off < 0)
3404 break; // unknown junk in the inits
3405 if (st->in(MemNode::Memory) != zmem)
3406 break; // complicated store chains somehow in list
3408 int st_size = st->as_Store()->memory_size();
3409 intptr_t next_init_off = st_off + st_size;
3411 if (do_zeroing && zeroes_done < next_init_off) {
3412 // See if this store needs a zero before it or under it.
3413 intptr_t zeroes_needed = st_off;
3415 if (st_size < BytesPerInt) {
3416 // Look for subword stores which only partially initialize words.
3417 // If we find some, we must lay down some word-level zeroes first,
3418 // underneath the subword stores.
3419 //
3420 // Examples:
3421 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3422 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3423 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3424 //
3425 // Note: coalesce_subword_stores may have already done this,
3426 // if it was prompted by constant non-zero subword initializers.
3427 // But this case can still arise with non-constant stores.
3429 intptr_t next_full_store = find_next_fullword_store(i, phase);
3431 // In the examples above:
3432 // in(i) p q r s x y z
3433 // st_off 12 13 14 15 12 13 14
3434 // st_size 1 1 1 1 1 1 1
3435 // next_full_s. 12 16 16 16 16 16 16
3436 // z's_done 12 16 16 16 12 16 12
3437 // z's_needed 12 16 16 16 16 16 16
3438 // zsize 0 0 0 0 4 0 4
3439 if (next_full_store < 0) {
3440 // Conservative tack: Zero to end of current word.
3441 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3442 } else {
3443 // Zero to beginning of next fully initialized word.
3444 // Or, don't zero at all, if we are already in that word.
3445 assert(next_full_store >= zeroes_needed, "must go forward");
3446 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3447 zeroes_needed = next_full_store;
3448 }
3449 }
3451 if (zeroes_needed > zeroes_done) {
3452 intptr_t zsize = zeroes_needed - zeroes_done;
3453 // Do some incremental zeroing on rawmem, in parallel with inits.
3454 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3455 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3456 zeroes_done, zeroes_needed,
3457 phase);
3458 zeroes_done = zeroes_needed;
3459 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3460 do_zeroing = false; // leave the hole, next time
3461 }
3462 }
3464 // Collect the store and move on:
3465 st->set_req(MemNode::Memory, inits);
3466 inits = st; // put it on the linearized chain
3467 set_req(i, zmem); // unhook from previous position
3469 if (zeroes_done == st_off)
3470 zeroes_done = next_init_off;
3472 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3474 #ifdef ASSERT
3475 // Various order invariants. Weaker than stores_are_sane because
3476 // a large constant tile can be filled in by smaller non-constant stores.
3477 assert(st_off >= last_init_off, "inits do not reverse");
3478 last_init_off = st_off;
3479 const Type* val = NULL;
3480 if (st_size >= BytesPerInt &&
3481 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3482 (int)val->basic_type() < (int)T_OBJECT) {
3483 assert(st_off >= last_tile_end, "tiles do not overlap");
3484 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3485 last_tile_end = MAX2(last_tile_end, next_init_off);
3486 } else {
3487 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3488 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3489 assert(st_off >= last_init_end, "inits do not overlap");
3490 last_init_end = next_init_off; // it's a non-tile
3491 }
3492 #endif //ASSERT
3493 }
3495 remove_extra_zeroes(); // clear out all the zmems left over
3496 add_req(inits);
3498 if (!ZeroTLAB) {
3499 // If anything remains to be zeroed, zero it all now.
3500 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3501 // if it is the last unused 4 bytes of an instance, forget about it
3502 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3503 if (zeroes_done + BytesPerLong >= size_limit) {
3504 assert(allocation() != NULL, "");
3505 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3506 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3507 if (zeroes_done == k->layout_helper())
3508 zeroes_done = size_limit;
3509 }
3510 if (zeroes_done < size_limit) {
3511 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3512 zeroes_done, size_in_bytes, phase);
3513 }
3514 }
3516 set_complete(phase);
3517 return rawmem;
3518 }
3521 #ifdef ASSERT
3522 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3523 if (is_complete())
3524 return true; // stores could be anything at this point
3525 assert(allocation() != NULL, "must be present");
3526 intptr_t last_off = allocation()->minimum_header_size();
3527 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3528 Node* st = in(i);
3529 intptr_t st_off = get_store_offset(st, phase);
3530 if (st_off < 0) continue; // ignore dead garbage
3531 if (last_off > st_off) {
3532 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
3533 this->dump(2);
3534 assert(false, "ascending store offsets");
3535 return false;
3536 }
3537 last_off = st_off + st->as_Store()->memory_size();
3538 }
3539 return true;
3540 }
3541 #endif //ASSERT
3546 //============================MergeMemNode=====================================
3547 //
3548 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3549 // contributing store or call operations. Each contributor provides the memory
3550 // state for a particular "alias type" (see Compile::alias_type). For example,
3551 // if a MergeMem has an input X for alias category #6, then any memory reference
3552 // to alias category #6 may use X as its memory state input, as an exact equivalent
3553 // to using the MergeMem as a whole.
3554 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3555 //
3556 // (Here, the <N> notation gives the index of the relevant adr_type.)
3557 //
3558 // In one special case (and more cases in the future), alias categories overlap.
3559 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
3560 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
3561 // it is exactly equivalent to that state W:
3562 // MergeMem(<Bot>: W) <==> W
3563 //
3564 // Usually, the merge has more than one input. In that case, where inputs
3565 // overlap (i.e., one is Bot), the narrower alias type determines the memory
3566 // state for that type, and the wider alias type (Bot) fills in everywhere else:
3567 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
3568 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
3569 //
3570 // A merge can take a "wide" memory state as one of its narrow inputs.
3571 // This simply means that the merge observes out only the relevant parts of
3572 // the wide input. That is, wide memory states arriving at narrow merge inputs
3573 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
3574 //
3575 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
3576 // and that memory slices "leak through":
3577 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
3578 //
3579 // But, in such a cascade, repeated memory slices can "block the leak":
3580 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
3581 //
3582 // In the last example, Y is not part of the combined memory state of the
3583 // outermost MergeMem. The system must, of course, prevent unschedulable
3584 // memory states from arising, so you can be sure that the state Y is somehow
3585 // a precursor to state Y'.
3586 //
3587 //
3588 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
3589 // of each MergeMemNode array are exactly the numerical alias indexes, including
3590 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
3591 // Compile::alias_type (and kin) produce and manage these indexes.
3592 //
3593 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
3594 // (Note that this provides quick access to the top node inside MergeMem methods,
3595 // without the need to reach out via TLS to Compile::current.)
3596 //
3597 // As a consequence of what was just described, a MergeMem that represents a full
3598 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
3599 // containing all alias categories.
3600 //
3601 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
3602 //
3603 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
3604 // a memory state for the alias type <N>, or else the top node, meaning that
3605 // there is no particular input for that alias type. Note that the length of
3606 // a MergeMem is variable, and may be extended at any time to accommodate new
3607 // memory states at larger alias indexes. When merges grow, they are of course
3608 // filled with "top" in the unused in() positions.
3609 //
3610 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
3611 // (Top was chosen because it works smoothly with passes like GCM.)
3612 //
3613 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
3614 // the type of random VM bits like TLS references.) Since it is always the
3615 // first non-Bot memory slice, some low-level loops use it to initialize an
3616 // index variable: for (i = AliasIdxRaw; i < req(); i++).
3617 //
3618 //
3619 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
3620 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
3621 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
3622 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
3623 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
3624 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
3625 //
3626 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
3627 // really that different from the other memory inputs. An abbreviation called
3628 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
3629 //
3630 //
3631 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
3632 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
3633 // that "emerges though" the base memory will be marked as excluding the alias types
3634 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
3635 //
3636 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
3637 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
3638 //
3639 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
3640 // (It is currently unimplemented.) As you can see, the resulting merge is
3641 // actually a disjoint union of memory states, rather than an overlay.
3642 //
3644 //------------------------------MergeMemNode-----------------------------------
3645 Node* MergeMemNode::make_empty_memory() {
3646 Node* empty_memory = (Node*) Compile::current()->top();
3647 assert(empty_memory->is_top(), "correct sentinel identity");
3648 return empty_memory;
3649 }
3651 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3652 init_class_id(Class_MergeMem);
3653 // all inputs are nullified in Node::Node(int)
3654 // set_input(0, NULL); // no control input
3656 // Initialize the edges uniformly to top, for starters.
3657 Node* empty_mem = make_empty_memory();
3658 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
3659 init_req(i,empty_mem);
3660 }
3661 assert(empty_memory() == empty_mem, "");
3663 if( new_base != NULL && new_base->is_MergeMem() ) {
3664 MergeMemNode* mdef = new_base->as_MergeMem();
3665 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
3666 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
3667 mms.set_memory(mms.memory2());
3668 }
3669 assert(base_memory() == mdef->base_memory(), "");
3670 } else {
3671 set_base_memory(new_base);
3672 }
3673 }
3675 // Make a new, untransformed MergeMem with the same base as 'mem'.
3676 // If mem is itself a MergeMem, populate the result with the same edges.
3677 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
3678 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
3679 }
3681 //------------------------------cmp--------------------------------------------
3682 uint MergeMemNode::hash() const { return NO_HASH; }
3683 uint MergeMemNode::cmp( const Node &n ) const {
3684 return (&n == this); // Always fail except on self
3685 }
3687 //------------------------------Identity---------------------------------------
3688 Node* MergeMemNode::Identity(PhaseTransform *phase) {
3689 // Identity if this merge point does not record any interesting memory
3690 // disambiguations.
3691 Node* base_mem = base_memory();
3692 Node* empty_mem = empty_memory();
3693 if (base_mem != empty_mem) { // Memory path is not dead?
3694 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3695 Node* mem = in(i);
3696 if (mem != empty_mem && mem != base_mem) {
3697 return this; // Many memory splits; no change
3698 }
3699 }
3700 }
3701 return base_mem; // No memory splits; ID on the one true input
3702 }
3704 //------------------------------Ideal------------------------------------------
3705 // This method is invoked recursively on chains of MergeMem nodes
3706 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3707 // Remove chain'd MergeMems
3708 //
3709 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
3710 // relative to the "in(Bot)". Since we are patching both at the same time,
3711 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
3712 // but rewrite each "in(i)" relative to the new "in(Bot)".
3713 Node *progress = NULL;
3716 Node* old_base = base_memory();
3717 Node* empty_mem = empty_memory();
3718 if (old_base == empty_mem)
3719 return NULL; // Dead memory path.
3721 MergeMemNode* old_mbase;
3722 if (old_base != NULL && old_base->is_MergeMem())
3723 old_mbase = old_base->as_MergeMem();
3724 else
3725 old_mbase = NULL;
3726 Node* new_base = old_base;
3728 // simplify stacked MergeMems in base memory
3729 if (old_mbase) new_base = old_mbase->base_memory();
3731 // the base memory might contribute new slices beyond my req()
3732 if (old_mbase) grow_to_match(old_mbase);
3734 // Look carefully at the base node if it is a phi.
3735 PhiNode* phi_base;
3736 if (new_base != NULL && new_base->is_Phi())
3737 phi_base = new_base->as_Phi();
3738 else
3739 phi_base = NULL;
3741 Node* phi_reg = NULL;
3742 uint phi_len = (uint)-1;
3743 if (phi_base != NULL && !phi_base->is_copy()) {
3744 // do not examine phi if degraded to a copy
3745 phi_reg = phi_base->region();
3746 phi_len = phi_base->req();
3747 // see if the phi is unfinished
3748 for (uint i = 1; i < phi_len; i++) {
3749 if (phi_base->in(i) == NULL) {
3750 // incomplete phi; do not look at it yet!
3751 phi_reg = NULL;
3752 phi_len = (uint)-1;
3753 break;
3754 }
3755 }
3756 }
3758 // Note: We do not call verify_sparse on entry, because inputs
3759 // can normalize to the base_memory via subsume_node or similar
3760 // mechanisms. This method repairs that damage.
3762 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
3764 // Look at each slice.
3765 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3766 Node* old_in = in(i);
3767 // calculate the old memory value
3768 Node* old_mem = old_in;
3769 if (old_mem == empty_mem) old_mem = old_base;
3770 assert(old_mem == memory_at(i), "");
3772 // maybe update (reslice) the old memory value
3774 // simplify stacked MergeMems
3775 Node* new_mem = old_mem;
3776 MergeMemNode* old_mmem;
3777 if (old_mem != NULL && old_mem->is_MergeMem())
3778 old_mmem = old_mem->as_MergeMem();
3779 else
3780 old_mmem = NULL;
3781 if (old_mmem == this) {
3782 // This can happen if loops break up and safepoints disappear.
3783 // A merge of BotPtr (default) with a RawPtr memory derived from a
3784 // safepoint can be rewritten to a merge of the same BotPtr with
3785 // the BotPtr phi coming into the loop. If that phi disappears
3786 // also, we can end up with a self-loop of the mergemem.
3787 // In general, if loops degenerate and memory effects disappear,
3788 // a mergemem can be left looking at itself. This simply means
3789 // that the mergemem's default should be used, since there is
3790 // no longer any apparent effect on this slice.
3791 // Note: If a memory slice is a MergeMem cycle, it is unreachable
3792 // from start. Update the input to TOP.
3793 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
3794 }
3795 else if (old_mmem != NULL) {
3796 new_mem = old_mmem->memory_at(i);
3797 }
3798 // else preceding memory was not a MergeMem
3800 // replace equivalent phis (unfortunately, they do not GVN together)
3801 if (new_mem != NULL && new_mem != new_base &&
3802 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
3803 if (new_mem->is_Phi()) {
3804 PhiNode* phi_mem = new_mem->as_Phi();
3805 for (uint i = 1; i < phi_len; i++) {
3806 if (phi_base->in(i) != phi_mem->in(i)) {
3807 phi_mem = NULL;
3808 break;
3809 }
3810 }
3811 if (phi_mem != NULL) {
3812 // equivalent phi nodes; revert to the def
3813 new_mem = new_base;
3814 }
3815 }
3816 }
3818 // maybe store down a new value
3819 Node* new_in = new_mem;
3820 if (new_in == new_base) new_in = empty_mem;
3822 if (new_in != old_in) {
3823 // Warning: Do not combine this "if" with the previous "if"
3824 // A memory slice might have be be rewritten even if it is semantically
3825 // unchanged, if the base_memory value has changed.
3826 set_req(i, new_in);
3827 progress = this; // Report progress
3828 }
3829 }
3831 if (new_base != old_base) {
3832 set_req(Compile::AliasIdxBot, new_base);
3833 // Don't use set_base_memory(new_base), because we need to update du.
3834 assert(base_memory() == new_base, "");
3835 progress = this;
3836 }
3838 if( base_memory() == this ) {
3839 // a self cycle indicates this memory path is dead
3840 set_req(Compile::AliasIdxBot, empty_mem);
3841 }
3843 // Resolve external cycles by calling Ideal on a MergeMem base_memory
3844 // Recursion must occur after the self cycle check above
3845 if( base_memory()->is_MergeMem() ) {
3846 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
3847 Node *m = phase->transform(new_mbase); // Rollup any cycles
3848 if( m != NULL && (m->is_top() ||
3849 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
3850 // propagate rollup of dead cycle to self
3851 set_req(Compile::AliasIdxBot, empty_mem);
3852 }
3853 }
3855 if( base_memory() == empty_mem ) {
3856 progress = this;
3857 // Cut inputs during Parse phase only.
3858 // During Optimize phase a dead MergeMem node will be subsumed by Top.
3859 if( !can_reshape ) {
3860 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3861 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
3862 }
3863 }
3864 }
3866 if( !progress && base_memory()->is_Phi() && can_reshape ) {
3867 // Check if PhiNode::Ideal's "Split phis through memory merges"
3868 // transform should be attempted. Look for this->phi->this cycle.
3869 uint merge_width = req();
3870 if (merge_width > Compile::AliasIdxRaw) {
3871 PhiNode* phi = base_memory()->as_Phi();
3872 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
3873 if (phi->in(i) == this) {
3874 phase->is_IterGVN()->_worklist.push(phi);
3875 break;
3876 }
3877 }
3878 }
3879 }
3881 assert(progress || verify_sparse(), "please, no dups of base");
3882 return progress;
3883 }
3885 //-------------------------set_base_memory-------------------------------------
3886 void MergeMemNode::set_base_memory(Node *new_base) {
3887 Node* empty_mem = empty_memory();
3888 set_req(Compile::AliasIdxBot, new_base);
3889 assert(memory_at(req()) == new_base, "must set default memory");
3890 // Clear out other occurrences of new_base:
3891 if (new_base != empty_mem) {
3892 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3893 if (in(i) == new_base) set_req(i, empty_mem);
3894 }
3895 }
3896 }
3898 //------------------------------out_RegMask------------------------------------
3899 const RegMask &MergeMemNode::out_RegMask() const {
3900 return RegMask::Empty;
3901 }
3903 //------------------------------dump_spec--------------------------------------
3904 #ifndef PRODUCT
3905 void MergeMemNode::dump_spec(outputStream *st) const {
3906 st->print(" {");
3907 Node* base_mem = base_memory();
3908 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
3909 Node* mem = memory_at(i);
3910 if (mem == base_mem) { st->print(" -"); continue; }
3911 st->print( " N%d:", mem->_idx );
3912 Compile::current()->get_adr_type(i)->dump_on(st);
3913 }
3914 st->print(" }");
3915 }
3916 #endif // !PRODUCT
3919 #ifdef ASSERT
3920 static bool might_be_same(Node* a, Node* b) {
3921 if (a == b) return true;
3922 if (!(a->is_Phi() || b->is_Phi())) return false;
3923 // phis shift around during optimization
3924 return true; // pretty stupid...
3925 }
3927 // verify a narrow slice (either incoming or outgoing)
3928 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
3929 if (!VerifyAliases) return; // don't bother to verify unless requested
3930 if (is_error_reported()) return; // muzzle asserts when debugging an error
3931 if (Node::in_dump()) return; // muzzle asserts when printing
3932 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
3933 assert(n != NULL, "");
3934 // Elide intervening MergeMem's
3935 while (n->is_MergeMem()) {
3936 n = n->as_MergeMem()->memory_at(alias_idx);
3937 }
3938 Compile* C = Compile::current();
3939 const TypePtr* n_adr_type = n->adr_type();
3940 if (n == m->empty_memory()) {
3941 // Implicit copy of base_memory()
3942 } else if (n_adr_type != TypePtr::BOTTOM) {
3943 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
3944 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
3945 } else {
3946 // A few places like make_runtime_call "know" that VM calls are narrow,
3947 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
3948 bool expected_wide_mem = false;
3949 if (n == m->base_memory()) {
3950 expected_wide_mem = true;
3951 } else if (alias_idx == Compile::AliasIdxRaw ||
3952 n == m->memory_at(Compile::AliasIdxRaw)) {
3953 expected_wide_mem = true;
3954 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
3955 // memory can "leak through" calls on channels that
3956 // are write-once. Allow this also.
3957 expected_wide_mem = true;
3958 }
3959 assert(expected_wide_mem, "expected narrow slice replacement");
3960 }
3961 }
3962 #else // !ASSERT
3963 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
3964 #endif
3967 //-----------------------------memory_at---------------------------------------
3968 Node* MergeMemNode::memory_at(uint alias_idx) const {
3969 assert(alias_idx >= Compile::AliasIdxRaw ||
3970 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
3971 "must avoid base_memory and AliasIdxTop");
3973 // Otherwise, it is a narrow slice.
3974 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
3975 Compile *C = Compile::current();
3976 if (is_empty_memory(n)) {
3977 // the array is sparse; empty slots are the "top" node
3978 n = base_memory();
3979 assert(Node::in_dump()
3980 || n == NULL || n->bottom_type() == Type::TOP
3981 || n->adr_type() == TypePtr::BOTTOM
3982 || n->adr_type() == TypeRawPtr::BOTTOM
3983 || Compile::current()->AliasLevel() == 0,
3984 "must be a wide memory");
3985 // AliasLevel == 0 if we are organizing the memory states manually.
3986 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
3987 } else {
3988 // make sure the stored slice is sane
3989 #ifdef ASSERT
3990 if (is_error_reported() || Node::in_dump()) {
3991 } else if (might_be_same(n, base_memory())) {
3992 // Give it a pass: It is a mostly harmless repetition of the base.
3993 // This can arise normally from node subsumption during optimization.
3994 } else {
3995 verify_memory_slice(this, alias_idx, n);
3996 }
3997 #endif
3998 }
3999 return n;
4000 }
4002 //---------------------------set_memory_at-------------------------------------
4003 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4004 verify_memory_slice(this, alias_idx, n);
4005 Node* empty_mem = empty_memory();
4006 if (n == base_memory()) n = empty_mem; // collapse default
4007 uint need_req = alias_idx+1;
4008 if (req() < need_req) {
4009 if (n == empty_mem) return; // already the default, so do not grow me
4010 // grow the sparse array
4011 do {
4012 add_req(empty_mem);
4013 } while (req() < need_req);
4014 }
4015 set_req( alias_idx, n );
4016 }
4020 //--------------------------iteration_setup------------------------------------
4021 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4022 if (other != NULL) {
4023 grow_to_match(other);
4024 // invariant: the finite support of mm2 is within mm->req()
4025 #ifdef ASSERT
4026 for (uint i = req(); i < other->req(); i++) {
4027 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4028 }
4029 #endif
4030 }
4031 // Replace spurious copies of base_memory by top.
4032 Node* base_mem = base_memory();
4033 if (base_mem != NULL && !base_mem->is_top()) {
4034 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4035 if (in(i) == base_mem)
4036 set_req(i, empty_memory());
4037 }
4038 }
4039 }
4041 //---------------------------grow_to_match-------------------------------------
4042 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4043 Node* empty_mem = empty_memory();
4044 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4045 // look for the finite support of the other memory
4046 for (uint i = other->req(); --i >= req(); ) {
4047 if (other->in(i) != empty_mem) {
4048 uint new_len = i+1;
4049 while (req() < new_len) add_req(empty_mem);
4050 break;
4051 }
4052 }
4053 }
4055 //---------------------------verify_sparse-------------------------------------
4056 #ifndef PRODUCT
4057 bool MergeMemNode::verify_sparse() const {
4058 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4059 Node* base_mem = base_memory();
4060 // The following can happen in degenerate cases, since empty==top.
4061 if (is_empty_memory(base_mem)) return true;
4062 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4063 assert(in(i) != NULL, "sane slice");
4064 if (in(i) == base_mem) return false; // should have been the sentinel value!
4065 }
4066 return true;
4067 }
4069 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4070 Node* n;
4071 n = mm->in(idx);
4072 if (mem == n) return true; // might be empty_memory()
4073 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4074 if (mem == n) return true;
4075 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4076 if (mem == n) return true;
4077 if (n == NULL) break;
4078 }
4079 return false;
4080 }
4081 #endif // !PRODUCT