1 /*
2 * Copyright (c) 1997, 2011, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "oops/objArrayKlass.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/cfgnode.hpp"
32 #include "opto/compile.hpp"
33 #include "opto/connode.hpp"
34 #include "opto/loopnode.hpp"
35 #include "opto/machnode.hpp"
36 #include "opto/matcher.hpp"
37 #include "opto/memnode.hpp"
38 #include "opto/mulnode.hpp"
39 #include "opto/phaseX.hpp"
40 #include "opto/regmask.hpp"
41
42 // Portions of code courtesy of Clifford Click
43
44 // Optimization - Graph Style
45
46 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
47
48 //=============================================================================
49 uint MemNode::size_of() const { return sizeof(*this); }
50
51 const TypePtr *MemNode::adr_type() const {
52 Node* adr = in(Address);
53 const TypePtr* cross_check = NULL;
54 DEBUG_ONLY(cross_check = _adr_type);
55 return calculate_adr_type(adr->bottom_type(), cross_check);
56 }
57
58 #ifndef PRODUCT
59 void MemNode::dump_spec(outputStream *st) const {
60 if (in(Address) == NULL) return; // node is dead
61 #ifndef ASSERT
62 // fake the missing field
63 const TypePtr* _adr_type = NULL;
64 if (in(Address) != NULL)
65 _adr_type = in(Address)->bottom_type()->isa_ptr();
66 #endif
67 dump_adr_type(this, _adr_type, st);
68
69 Compile* C = Compile::current();
70 if( C->alias_type(_adr_type)->is_volatile() )
71 st->print(" Volatile!");
72 }
73
74 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
75 st->print(" @");
76 if (adr_type == NULL) {
77 st->print("NULL");
78 } else {
79 adr_type->dump_on(st);
80 Compile* C = Compile::current();
81 Compile::AliasType* atp = NULL;
82 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
83 if (atp == NULL)
84 st->print(", idx=?\?;");
85 else if (atp->index() == Compile::AliasIdxBot)
86 st->print(", idx=Bot;");
87 else if (atp->index() == Compile::AliasIdxTop)
88 st->print(", idx=Top;");
89 else if (atp->index() == Compile::AliasIdxRaw)
90 st->print(", idx=Raw;");
91 else {
92 ciField* field = atp->field();
93 if (field) {
94 st->print(", name=");
95 field->print_name_on(st);
96 }
97 st->print(", idx=%d;", atp->index());
98 }
99 }
100 }
101
102 extern void print_alias_types();
103
104 #endif
105
106 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
107 const TypeOopPtr *tinst = t_adr->isa_oopptr();
108 if (tinst == NULL || !tinst->is_known_instance_field())
109 return mchain; // don't try to optimize non-instance types
110 uint instance_id = tinst->instance_id();
111 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
112 Node *prev = NULL;
113 Node *result = mchain;
114 while (prev != result) {
115 prev = result;
116 if (result == start_mem)
117 break; // hit one of our sentinels
118 // skip over a call which does not affect this memory slice
119 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
120 Node *proj_in = result->in(0);
121 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
122 break; // hit one of our sentinels
123 } else if (proj_in->is_Call()) {
124 CallNode *call = proj_in->as_Call();
125 if (!call->may_modify(t_adr, phase)) {
126 result = call->in(TypeFunc::Memory);
127 }
128 } else if (proj_in->is_Initialize()) {
129 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
130 // Stop if this is the initialization for the object instance which
131 // which contains this memory slice, otherwise skip over it.
132 if (alloc != NULL && alloc->_idx != instance_id) {
133 result = proj_in->in(TypeFunc::Memory);
134 }
135 } else if (proj_in->is_MemBar()) {
136 result = proj_in->in(TypeFunc::Memory);
137 } else {
138 assert(false, "unexpected projection");
139 }
140 } else if (result->is_ClearArray()) {
141 if (!ClearArrayNode::step_through(&result, instance_id, phase)) {
142 // Can not bypass initialization of the instance
143 // we are looking for.
144 break;
145 }
146 // Otherwise skip it (the call updated 'result' value).
147 } else if (result->is_MergeMem()) {
148 result = step_through_mergemem(phase, result->as_MergeMem(), t_adr, NULL, tty);
149 }
150 }
151 return result;
152 }
153
154 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, PhaseGVN *phase) {
155 const TypeOopPtr *t_oop = t_adr->isa_oopptr();
156 bool is_instance = (t_oop != NULL) && t_oop->is_known_instance_field();
157 PhaseIterGVN *igvn = phase->is_IterGVN();
158 Node *result = mchain;
159 result = optimize_simple_memory_chain(result, t_adr, phase);
160 if (is_instance && igvn != NULL && result->is_Phi()) {
161 PhiNode *mphi = result->as_Phi();
162 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
163 const TypePtr *t = mphi->adr_type();
164 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
165 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
166 t->is_oopptr()->cast_to_exactness(true)
167 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
168 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
169 // clone the Phi with our address type
170 result = mphi->split_out_instance(t_adr, igvn);
171 } else {
172 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
173 }
174 }
175 return result;
176 }
177
178 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
179 uint alias_idx = phase->C->get_alias_index(tp);
180 Node *mem = mmem;
181 #ifdef ASSERT
182 {
183 // Check that current type is consistent with the alias index used during graph construction
184 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
185 bool consistent = adr_check == NULL || adr_check->empty() ||
186 phase->C->must_alias(adr_check, alias_idx );
187 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
188 if( !consistent && adr_check != NULL && !adr_check->empty() &&
189 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
190 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
191 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
192 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
193 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
194 // don't assert if it is dead code.
195 consistent = true;
196 }
197 if( !consistent ) {
198 st->print("alias_idx==%d, adr_check==", alias_idx);
199 if( adr_check == NULL ) {
200 st->print("NULL");
201 } else {
202 adr_check->dump();
203 }
204 st->cr();
205 print_alias_types();
206 assert(consistent, "adr_check must match alias idx");
207 }
208 }
209 #endif
210 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
211 // means an array I have not precisely typed yet. Do not do any
212 // alias stuff with it any time soon.
213 const TypeOopPtr *toop = tp->isa_oopptr();
214 if( tp->base() != Type::AnyPtr &&
215 !(toop &&
216 toop->klass() != NULL &&
217 toop->klass()->is_java_lang_Object() &&
218 toop->offset() == Type::OffsetBot) ) {
219 // compress paths and change unreachable cycles to TOP
220 // If not, we can update the input infinitely along a MergeMem cycle
221 // Equivalent code in PhiNode::Ideal
222 Node* m = phase->transform(mmem);
223 // If transformed to a MergeMem, get the desired slice
224 // Otherwise the returned node represents memory for every slice
225 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
226 // Update input if it is progress over what we have now
227 }
228 return mem;
229 }
230
231 //--------------------------Ideal_common---------------------------------------
232 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
233 // Unhook non-raw memories from complete (macro-expanded) initializations.
234 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
235 // If our control input is a dead region, kill all below the region
236 Node *ctl = in(MemNode::Control);
237 if (ctl && remove_dead_region(phase, can_reshape))
238 return this;
239 ctl = in(MemNode::Control);
240 // Don't bother trying to transform a dead node
241 if( ctl && ctl->is_top() ) return NodeSentinel;
242
243 PhaseIterGVN *igvn = phase->is_IterGVN();
244 // Wait if control on the worklist.
245 if (ctl && can_reshape && igvn != NULL) {
246 Node* bol = NULL;
247 Node* cmp = NULL;
248 if (ctl->in(0)->is_If()) {
249 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
250 bol = ctl->in(0)->in(1);
251 if (bol->is_Bool())
252 cmp = ctl->in(0)->in(1)->in(1);
253 }
254 if (igvn->_worklist.member(ctl) ||
255 (bol != NULL && igvn->_worklist.member(bol)) ||
256 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
257 // This control path may be dead.
258 // Delay this memory node transformation until the control is processed.
259 phase->is_IterGVN()->_worklist.push(this);
260 return NodeSentinel; // caller will return NULL
261 }
262 }
263 // Ignore if memory is dead, or self-loop
264 Node *mem = in(MemNode::Memory);
265 if( phase->type( mem ) == Type::TOP ) return NodeSentinel; // caller will return NULL
266 assert( mem != this, "dead loop in MemNode::Ideal" );
267
268 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
269 // This memory slice may be dead.
270 // Delay this mem node transformation until the memory is processed.
271 phase->is_IterGVN()->_worklist.push(this);
272 return NodeSentinel; // caller will return NULL
273 }
274
275 Node *address = in(MemNode::Address);
276 const Type *t_adr = phase->type( address );
277 if( t_adr == Type::TOP ) return NodeSentinel; // caller will return NULL
278
279 if( can_reshape && igvn != NULL &&
280 (igvn->_worklist.member(address) ||
281 igvn->_worklist.size() > 0 && (phase->type(address) != adr_type())) ) {
282 // The address's base and type may change when the address is processed.
283 // Delay this mem node transformation until the address is processed.
284 phase->is_IterGVN()->_worklist.push(this);
285 return NodeSentinel; // caller will return NULL
286 }
287
288 // Do NOT remove or optimize the next lines: ensure a new alias index
289 // is allocated for an oop pointer type before Escape Analysis.
290 // Note: C++ will not remove it since the call has side effect.
291 if ( t_adr->isa_oopptr() ) {
292 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
293 }
294
295 #ifdef ASSERT
296 Node* base = NULL;
297 if (address->is_AddP())
298 base = address->in(AddPNode::Base);
299 assert(base == NULL || t_adr->isa_rawptr() ||
300 !phase->type(base)->higher_equal(TypePtr::NULL_PTR), "NULL+offs not RAW address?");
301 #endif
302
303 // Avoid independent memory operations
304 Node* old_mem = mem;
305
306 // The code which unhooks non-raw memories from complete (macro-expanded)
307 // initializations was removed. After macro-expansion all stores catched
308 // by Initialize node became raw stores and there is no information
309 // which memory slices they modify. So it is unsafe to move any memory
310 // operation above these stores. Also in most cases hooked non-raw memories
311 // were already unhooked by using information from detect_ptr_independence()
312 // and find_previous_store().
313
314 if (mem->is_MergeMem()) {
315 MergeMemNode* mmem = mem->as_MergeMem();
316 const TypePtr *tp = t_adr->is_ptr();
317
318 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
319 }
320
321 if (mem != old_mem) {
322 set_req(MemNode::Memory, mem);
323 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
324 return this;
325 }
326
327 // let the subclass continue analyzing...
328 return NULL;
329 }
330
331 // Helper function for proving some simple control dominations.
332 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
333 // Already assumes that 'dom' is available at 'sub', and that 'sub'
334 // is not a constant (dominated by the method's StartNode).
335 // Used by MemNode::find_previous_store to prove that the
336 // control input of a memory operation predates (dominates)
337 // an allocation it wants to look past.
338 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
339 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
340 return false; // Conservative answer for dead code
341
342 // Check 'dom'. Skip Proj and CatchProj nodes.
343 dom = dom->find_exact_control(dom);
344 if (dom == NULL || dom->is_top())
345 return false; // Conservative answer for dead code
346
347 if (dom == sub) {
348 // For the case when, for example, 'sub' is Initialize and the original
349 // 'dom' is Proj node of the 'sub'.
350 return false;
351 }
352
353 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
354 return true;
355
356 // 'dom' dominates 'sub' if its control edge and control edges
357 // of all its inputs dominate or equal to sub's control edge.
358
359 // Currently 'sub' is either Allocate, Initialize or Start nodes.
360 // Or Region for the check in LoadNode::Ideal();
361 // 'sub' should have sub->in(0) != NULL.
362 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
363 sub->is_Region(), "expecting only these nodes");
364
365 // Get control edge of 'sub'.
366 Node* orig_sub = sub;
367 sub = sub->find_exact_control(sub->in(0));
368 if (sub == NULL || sub->is_top())
369 return false; // Conservative answer for dead code
370
371 assert(sub->is_CFG(), "expecting control");
372
373 if (sub == dom)
374 return true;
375
376 if (sub->is_Start() || sub->is_Root())
377 return false;
378
379 {
380 // Check all control edges of 'dom'.
381
382 ResourceMark rm;
383 Arena* arena = Thread::current()->resource_area();
384 Node_List nlist(arena);
385 Unique_Node_List dom_list(arena);
386
387 dom_list.push(dom);
388 bool only_dominating_controls = false;
389
390 for (uint next = 0; next < dom_list.size(); next++) {
391 Node* n = dom_list.at(next);
392 if (n == orig_sub)
393 return false; // One of dom's inputs dominated by sub.
394 if (!n->is_CFG() && n->pinned()) {
395 // Check only own control edge for pinned non-control nodes.
396 n = n->find_exact_control(n->in(0));
397 if (n == NULL || n->is_top())
398 return false; // Conservative answer for dead code
399 assert(n->is_CFG(), "expecting control");
400 dom_list.push(n);
401 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
402 only_dominating_controls = true;
403 } else if (n->is_CFG()) {
404 if (n->dominates(sub, nlist))
405 only_dominating_controls = true;
406 else
407 return false;
408 } else {
409 // First, own control edge.
410 Node* m = n->find_exact_control(n->in(0));
411 if (m != NULL) {
412 if (m->is_top())
413 return false; // Conservative answer for dead code
414 dom_list.push(m);
415 }
416 // Now, the rest of edges.
417 uint cnt = n->req();
418 for (uint i = 1; i < cnt; i++) {
419 m = n->find_exact_control(n->in(i));
420 if (m == NULL || m->is_top())
421 continue;
422 dom_list.push(m);
423 }
424 }
425 }
426 return only_dominating_controls;
427 }
428 }
429
430 //---------------------detect_ptr_independence---------------------------------
431 // Used by MemNode::find_previous_store to prove that two base
432 // pointers are never equal.
433 // The pointers are accompanied by their associated allocations,
434 // if any, which have been previously discovered by the caller.
435 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
436 Node* p2, AllocateNode* a2,
437 PhaseTransform* phase) {
438 // Attempt to prove that these two pointers cannot be aliased.
439 // They may both manifestly be allocations, and they should differ.
440 // Or, if they are not both allocations, they can be distinct constants.
441 // Otherwise, one is an allocation and the other a pre-existing value.
442 if (a1 == NULL && a2 == NULL) { // neither an allocation
443 return (p1 != p2) && p1->is_Con() && p2->is_Con();
444 } else if (a1 != NULL && a2 != NULL) { // both allocations
445 return (a1 != a2);
446 } else if (a1 != NULL) { // one allocation a1
447 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
448 return all_controls_dominate(p2, a1);
449 } else { //(a2 != NULL) // one allocation a2
450 return all_controls_dominate(p1, a2);
451 }
452 return false;
453 }
454
455
456 // The logic for reordering loads and stores uses four steps:
457 // (a) Walk carefully past stores and initializations which we
458 // can prove are independent of this load.
459 // (b) Observe that the next memory state makes an exact match
460 // with self (load or store), and locate the relevant store.
461 // (c) Ensure that, if we were to wire self directly to the store,
462 // the optimizer would fold it up somehow.
463 // (d) Do the rewiring, and return, depending on some other part of
464 // the optimizer to fold up the load.
465 // This routine handles steps (a) and (b). Steps (c) and (d) are
466 // specific to loads and stores, so they are handled by the callers.
467 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
468 //
469 Node* MemNode::find_previous_store(PhaseTransform* phase) {
470 Node* ctrl = in(MemNode::Control);
471 Node* adr = in(MemNode::Address);
472 intptr_t offset = 0;
473 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
474 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
475
476 if (offset == Type::OffsetBot)
477 return NULL; // cannot unalias unless there are precise offsets
478
479 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
480
481 intptr_t size_in_bytes = memory_size();
482
483 Node* mem = in(MemNode::Memory); // start searching here...
484
485 int cnt = 50; // Cycle limiter
486 for (;;) { // While we can dance past unrelated stores...
487 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
488
489 if (mem->is_Store()) {
490 Node* st_adr = mem->in(MemNode::Address);
491 intptr_t st_offset = 0;
492 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
493 if (st_base == NULL)
494 break; // inscrutable pointer
495 if (st_offset != offset && st_offset != Type::OffsetBot) {
496 const int MAX_STORE = BytesPerLong;
497 if (st_offset >= offset + size_in_bytes ||
498 st_offset <= offset - MAX_STORE ||
499 st_offset <= offset - mem->as_Store()->memory_size()) {
500 // Success: The offsets are provably independent.
501 // (You may ask, why not just test st_offset != offset and be done?
502 // The answer is that stores of different sizes can co-exist
503 // in the same sequence of RawMem effects. We sometimes initialize
504 // a whole 'tile' of array elements with a single jint or jlong.)
505 mem = mem->in(MemNode::Memory);
506 continue; // (a) advance through independent store memory
507 }
508 }
509 if (st_base != base &&
510 detect_ptr_independence(base, alloc,
511 st_base,
512 AllocateNode::Ideal_allocation(st_base, phase),
513 phase)) {
514 // Success: The bases are provably independent.
515 mem = mem->in(MemNode::Memory);
516 continue; // (a) advance through independent store memory
517 }
518
519 // (b) At this point, if the bases or offsets do not agree, we lose,
520 // since we have not managed to prove 'this' and 'mem' independent.
521 if (st_base == base && st_offset == offset) {
522 return mem; // let caller handle steps (c), (d)
523 }
524
525 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
526 InitializeNode* st_init = mem->in(0)->as_Initialize();
527 AllocateNode* st_alloc = st_init->allocation();
528 if (st_alloc == NULL)
529 break; // something degenerated
530 bool known_identical = false;
531 bool known_independent = false;
532 if (alloc == st_alloc)
533 known_identical = true;
534 else if (alloc != NULL)
535 known_independent = true;
536 else if (all_controls_dominate(this, st_alloc))
537 known_independent = true;
538
539 if (known_independent) {
540 // The bases are provably independent: Either they are
541 // manifestly distinct allocations, or else the control
542 // of this load dominates the store's allocation.
543 int alias_idx = phase->C->get_alias_index(adr_type());
544 if (alias_idx == Compile::AliasIdxRaw) {
545 mem = st_alloc->in(TypeFunc::Memory);
546 } else {
547 mem = st_init->memory(alias_idx);
548 }
549 continue; // (a) advance through independent store memory
550 }
551
552 // (b) at this point, if we are not looking at a store initializing
553 // the same allocation we are loading from, we lose.
554 if (known_identical) {
555 // From caller, can_see_stored_value will consult find_captured_store.
556 return mem; // let caller handle steps (c), (d)
557 }
558
559 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
560 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
561 if (mem->is_Proj() && mem->in(0)->is_Call()) {
562 CallNode *call = mem->in(0)->as_Call();
563 if (!call->may_modify(addr_t, phase)) {
564 mem = call->in(TypeFunc::Memory);
565 continue; // (a) advance through independent call memory
566 }
567 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
568 mem = mem->in(0)->in(TypeFunc::Memory);
569 continue; // (a) advance through independent MemBar memory
570 } else if (mem->is_ClearArray()) {
571 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
572 // (the call updated 'mem' value)
573 continue; // (a) advance through independent allocation memory
574 } else {
575 // Can not bypass initialization of the instance
576 // we are looking for.
577 return mem;
578 }
579 } else if (mem->is_MergeMem()) {
580 int alias_idx = phase->C->get_alias_index(adr_type());
581 mem = mem->as_MergeMem()->memory_at(alias_idx);
582 continue; // (a) advance through independent MergeMem memory
583 }
584 }
585
586 // Unless there is an explicit 'continue', we must bail out here,
587 // because 'mem' is an inscrutable memory state (e.g., a call).
588 break;
589 }
590
591 return NULL; // bail out
592 }
593
594 //----------------------calculate_adr_type-------------------------------------
595 // Helper function. Notices when the given type of address hits top or bottom.
596 // Also, asserts a cross-check of the type against the expected address type.
597 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
598 if (t == Type::TOP) return NULL; // does not touch memory any more?
599 #ifdef PRODUCT
600 cross_check = NULL;
601 #else
602 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
603 #endif
604 const TypePtr* tp = t->isa_ptr();
605 if (tp == NULL) {
606 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
607 return TypePtr::BOTTOM; // touches lots of memory
608 } else {
609 #ifdef ASSERT
610 // %%%% [phh] We don't check the alias index if cross_check is
611 // TypeRawPtr::BOTTOM. Needs to be investigated.
612 if (cross_check != NULL &&
613 cross_check != TypePtr::BOTTOM &&
614 cross_check != TypeRawPtr::BOTTOM) {
615 // Recheck the alias index, to see if it has changed (due to a bug).
616 Compile* C = Compile::current();
617 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
618 "must stay in the original alias category");
619 // The type of the address must be contained in the adr_type,
620 // disregarding "null"-ness.
621 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
622 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
623 assert(cross_check->meet(tp_notnull) == cross_check,
624 "real address must not escape from expected memory type");
625 }
626 #endif
627 return tp;
628 }
629 }
630
631 //------------------------adr_phi_is_loop_invariant----------------------------
632 // A helper function for Ideal_DU_postCCP to check if a Phi in a counted
633 // loop is loop invariant. Make a quick traversal of Phi and associated
634 // CastPP nodes, looking to see if they are a closed group within the loop.
635 bool MemNode::adr_phi_is_loop_invariant(Node* adr_phi, Node* cast) {
636 // The idea is that the phi-nest must boil down to only CastPP nodes
637 // with the same data. This implies that any path into the loop already
638 // includes such a CastPP, and so the original cast, whatever its input,
639 // must be covered by an equivalent cast, with an earlier control input.
640 ResourceMark rm;
641
642 // The loop entry input of the phi should be the unique dominating
643 // node for every Phi/CastPP in the loop.
644 Unique_Node_List closure;
645 closure.push(adr_phi->in(LoopNode::EntryControl));
646
647 // Add the phi node and the cast to the worklist.
648 Unique_Node_List worklist;
649 worklist.push(adr_phi);
650 if( cast != NULL ){
651 if( !cast->is_ConstraintCast() ) return false;
652 worklist.push(cast);
653 }
654
655 // Begin recursive walk of phi nodes.
656 while( worklist.size() ){
657 // Take a node off the worklist
658 Node *n = worklist.pop();
659 if( !closure.member(n) ){
660 // Add it to the closure.
661 closure.push(n);
662 // Make a sanity check to ensure we don't waste too much time here.
663 if( closure.size() > 20) return false;
664 // This node is OK if:
665 // - it is a cast of an identical value
666 // - or it is a phi node (then we add its inputs to the worklist)
667 // Otherwise, the node is not OK, and we presume the cast is not invariant
668 if( n->is_ConstraintCast() ){
669 worklist.push(n->in(1));
670 } else if( n->is_Phi() ) {
671 for( uint i = 1; i < n->req(); i++ ) {
672 worklist.push(n->in(i));
673 }
674 } else {
675 return false;
676 }
677 }
678 }
679
680 // Quit when the worklist is empty, and we've found no offending nodes.
681 return true;
682 }
683
684 //------------------------------Ideal_DU_postCCP-------------------------------
685 // Find any cast-away of null-ness and keep its control. Null cast-aways are
686 // going away in this pass and we need to make this memory op depend on the
687 // gating null check.
688 Node *MemNode::Ideal_DU_postCCP( PhaseCCP *ccp ) {
689 return Ideal_common_DU_postCCP(ccp, this, in(MemNode::Address));
690 }
691
692 // I tried to leave the CastPP's in. This makes the graph more accurate in
693 // some sense; we get to keep around the knowledge that an oop is not-null
694 // after some test. Alas, the CastPP's interfere with GVN (some values are
695 // the regular oop, some are the CastPP of the oop, all merge at Phi's which
696 // cannot collapse, etc). This cost us 10% on SpecJVM, even when I removed
697 // some of the more trivial cases in the optimizer. Removing more useless
698 // Phi's started allowing Loads to illegally float above null checks. I gave
699 // up on this approach. CNC 10/20/2000
700 // This static method may be called not from MemNode (EncodePNode calls it).
701 // Only the control edge of the node 'n' might be updated.
702 Node *MemNode::Ideal_common_DU_postCCP( PhaseCCP *ccp, Node* n, Node* adr ) {
703 Node *skipped_cast = NULL;
704 // Need a null check? Regular static accesses do not because they are
705 // from constant addresses. Array ops are gated by the range check (which
706 // always includes a NULL check). Just check field ops.
707 if( n->in(MemNode::Control) == NULL ) {
708 // Scan upwards for the highest location we can place this memory op.
709 while( true ) {
710 switch( adr->Opcode() ) {
711
712 case Op_AddP: // No change to NULL-ness, so peek thru AddP's
713 adr = adr->in(AddPNode::Base);
714 continue;
715
716 case Op_DecodeN: // No change to NULL-ness, so peek thru
717 adr = adr->in(1);
718 continue;
719
720 case Op_CastPP:
721 // If the CastPP is useless, just peek on through it.
722 if( ccp->type(adr) == ccp->type(adr->in(1)) ) {
723 // Remember the cast that we've peeked though. If we peek
724 // through more than one, then we end up remembering the highest
725 // one, that is, if in a loop, the one closest to the top.
726 skipped_cast = adr;
727 adr = adr->in(1);
728 continue;
729 }
730 // CastPP is going away in this pass! We need this memory op to be
731 // control-dependent on the test that is guarding the CastPP.
732 ccp->hash_delete(n);
733 n->set_req(MemNode::Control, adr->in(0));
734 ccp->hash_insert(n);
735 return n;
736
737 case Op_Phi:
738 // Attempt to float above a Phi to some dominating point.
739 if (adr->in(0) != NULL && adr->in(0)->is_CountedLoop()) {
740 // If we've already peeked through a Cast (which could have set the
741 // control), we can't float above a Phi, because the skipped Cast
742 // may not be loop invariant.
743 if (adr_phi_is_loop_invariant(adr, skipped_cast)) {
744 adr = adr->in(1);
745 continue;
746 }
747 }
748
749 // Intentional fallthrough!
750
751 // No obvious dominating point. The mem op is pinned below the Phi
752 // by the Phi itself. If the Phi goes away (no true value is merged)
753 // then the mem op can float, but not indefinitely. It must be pinned
754 // behind the controls leading to the Phi.
755 case Op_CheckCastPP:
756 // These usually stick around to change address type, however a
757 // useless one can be elided and we still need to pick up a control edge
758 if (adr->in(0) == NULL) {
759 // This CheckCastPP node has NO control and is likely useless. But we
760 // need check further up the ancestor chain for a control input to keep
761 // the node in place. 4959717.
762 skipped_cast = adr;
763 adr = adr->in(1);
764 continue;
765 }
766 ccp->hash_delete(n);
767 n->set_req(MemNode::Control, adr->in(0));
768 ccp->hash_insert(n);
769 return n;
770
771 // List of "safe" opcodes; those that implicitly block the memory
772 // op below any null check.
773 case Op_CastX2P: // no null checks on native pointers
774 case Op_Parm: // 'this' pointer is not null
775 case Op_LoadP: // Loading from within a klass
776 case Op_LoadN: // Loading from within a klass
777 case Op_LoadKlass: // Loading from within a klass
778 case Op_LoadNKlass: // Loading from within a klass
779 case Op_ConP: // Loading from a klass
780 case Op_ConN: // Loading from a klass
781 case Op_CreateEx: // Sucking up the guts of an exception oop
782 case Op_Con: // Reading from TLS
783 case Op_CMoveP: // CMoveP is pinned
784 case Op_CMoveN: // CMoveN is pinned
785 break; // No progress
786
787 case Op_Proj: // Direct call to an allocation routine
788 case Op_SCMemProj: // Memory state from store conditional ops
789 #ifdef ASSERT
790 {
791 assert(adr->as_Proj()->_con == TypeFunc::Parms, "must be return value");
792 const Node* call = adr->in(0);
793 if (call->is_CallJava()) {
794 const CallJavaNode* call_java = call->as_CallJava();
795 const TypeTuple *r = call_java->tf()->range();
796 assert(r->cnt() > TypeFunc::Parms, "must return value");
797 const Type* ret_type = r->field_at(TypeFunc::Parms);
798 assert(ret_type && ret_type->isa_ptr(), "must return pointer");
799 // We further presume that this is one of
800 // new_instance_Java, new_array_Java, or
801 // the like, but do not assert for this.
802 } else if (call->is_Allocate()) {
803 // similar case to new_instance_Java, etc.
804 } else if (!call->is_CallLeaf()) {
805 // Projections from fetch_oop (OSR) are allowed as well.
806 ShouldNotReachHere();
807 }
808 }
809 #endif
810 break;
811 default:
812 ShouldNotReachHere();
813 }
814 break;
815 }
816 }
817
818 return NULL; // No progress
819 }
820
821
822 //=============================================================================
823 uint LoadNode::size_of() const { return sizeof(*this); }
824 uint LoadNode::cmp( const Node &n ) const
825 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
826 const Type *LoadNode::bottom_type() const { return _type; }
827 uint LoadNode::ideal_reg() const {
828 return Matcher::base2reg[_type->base()];
829 }
830
831 #ifndef PRODUCT
832 void LoadNode::dump_spec(outputStream *st) const {
833 MemNode::dump_spec(st);
834 if( !Verbose && !WizardMode ) {
835 // standard dump does this in Verbose and WizardMode
836 st->print(" #"); _type->dump_on(st);
837 }
838 }
839 #endif
840
841 #ifdef ASSERT
842 //----------------------------is_immutable_value-------------------------------
843 // Helper function to allow a raw load without control edge for some cases
844 bool LoadNode::is_immutable_value(Node* adr) {
845 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
846 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
847 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
848 in_bytes(JavaThread::osthread_offset())));
849 }
850 #endif
851
852 //----------------------------LoadNode::make-----------------------------------
853 // Polymorphic factory method:
854 Node *LoadNode::make( PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt ) {
855 Compile* C = gvn.C;
856
857 // sanity check the alias category against the created node type
858 assert(!(adr_type->isa_oopptr() &&
859 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
860 "use LoadKlassNode instead");
861 assert(!(adr_type->isa_aryptr() &&
862 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
863 "use LoadRangeNode instead");
864 // Check control edge of raw loads
865 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
866 // oop will be recorded in oop map if load crosses safepoint
867 rt->isa_oopptr() || is_immutable_value(adr),
868 "raw memory operations should have control edge");
869 switch (bt) {
870 case T_BOOLEAN: return new (C, 3) LoadUBNode(ctl, mem, adr, adr_type, rt->is_int() );
871 case T_BYTE: return new (C, 3) LoadBNode (ctl, mem, adr, adr_type, rt->is_int() );
872 case T_INT: return new (C, 3) LoadINode (ctl, mem, adr, adr_type, rt->is_int() );
873 case T_CHAR: return new (C, 3) LoadUSNode(ctl, mem, adr, adr_type, rt->is_int() );
874 case T_SHORT: return new (C, 3) LoadSNode (ctl, mem, adr, adr_type, rt->is_int() );
875 case T_LONG: return new (C, 3) LoadLNode (ctl, mem, adr, adr_type, rt->is_long() );
876 case T_FLOAT: return new (C, 3) LoadFNode (ctl, mem, adr, adr_type, rt );
877 case T_DOUBLE: return new (C, 3) LoadDNode (ctl, mem, adr, adr_type, rt );
878 case T_ADDRESS: return new (C, 3) LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr() );
879 case T_OBJECT:
880 #ifdef _LP64
881 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
882 Node* load = gvn.transform(new (C, 3) LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop()));
883 return new (C, 2) DecodeNNode(load, load->bottom_type()->make_ptr());
884 } else
885 #endif
886 {
887 assert(!adr->bottom_type()->is_ptr_to_narrowoop(), "should have got back a narrow oop");
888 return new (C, 3) LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr());
889 }
890 }
891 ShouldNotReachHere();
892 return (LoadNode*)NULL;
893 }
894
895 LoadLNode* LoadLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt) {
896 bool require_atomic = true;
897 return new (C, 3) LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), require_atomic);
898 }
899
900
901
902
903 //------------------------------hash-------------------------------------------
904 uint LoadNode::hash() const {
905 // unroll addition of interesting fields
906 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
907 }
908
909 //---------------------------can_see_stored_value------------------------------
910 // This routine exists to make sure this set of tests is done the same
911 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
912 // will change the graph shape in a way which makes memory alive twice at the
913 // same time (uses the Oracle model of aliasing), then some
914 // LoadXNode::Identity will fold things back to the equivalence-class model
915 // of aliasing.
916 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
917 Node* ld_adr = in(MemNode::Address);
918
919 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
920 Compile::AliasType* atp = tp != NULL ? phase->C->alias_type(tp) : NULL;
921 if (EliminateAutoBox && atp != NULL && atp->index() >= Compile::AliasIdxRaw &&
922 atp->field() != NULL && !atp->field()->is_volatile()) {
923 uint alias_idx = atp->index();
924 bool final = atp->field()->is_final();
925 Node* result = NULL;
926 Node* current = st;
927 // Skip through chains of MemBarNodes checking the MergeMems for
928 // new states for the slice of this load. Stop once any other
929 // kind of node is encountered. Loads from final memory can skip
930 // through any kind of MemBar but normal loads shouldn't skip
931 // through MemBarAcquire since the could allow them to move out of
932 // a synchronized region.
933 while (current->is_Proj()) {
934 int opc = current->in(0)->Opcode();
935 if ((final && (opc == Op_MemBarAcquire || opc == Op_MemBarAcquireLock)) ||
936 opc == Op_MemBarRelease || opc == Op_MemBarCPUOrder ||
937 opc == Op_MemBarReleaseLock) {
938 Node* mem = current->in(0)->in(TypeFunc::Memory);
939 if (mem->is_MergeMem()) {
940 MergeMemNode* merge = mem->as_MergeMem();
941 Node* new_st = merge->memory_at(alias_idx);
942 if (new_st == merge->base_memory()) {
943 // Keep searching
944 current = merge->base_memory();
945 continue;
946 }
947 // Save the new memory state for the slice and fall through
948 // to exit.
949 result = new_st;
950 }
951 }
952 break;
953 }
954 if (result != NULL) {
955 st = result;
956 }
957 }
958
959
960 // Loop around twice in the case Load -> Initialize -> Store.
961 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
962 for (int trip = 0; trip <= 1; trip++) {
963
964 if (st->is_Store()) {
965 Node* st_adr = st->in(MemNode::Address);
966 if (!phase->eqv(st_adr, ld_adr)) {
967 // Try harder before giving up... Match raw and non-raw pointers.
968 intptr_t st_off = 0;
969 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
970 if (alloc == NULL) return NULL;
971 intptr_t ld_off = 0;
972 AllocateNode* allo2 = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
973 if (alloc != allo2) return NULL;
974 if (ld_off != st_off) return NULL;
975 // At this point we have proven something like this setup:
976 // A = Allocate(...)
977 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
978 // S = StoreQ(, AddP(, A.Parm , #Off), V)
979 // (Actually, we haven't yet proven the Q's are the same.)
980 // In other words, we are loading from a casted version of
981 // the same pointer-and-offset that we stored to.
982 // Thus, we are able to replace L by V.
983 }
984 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
985 if (store_Opcode() != st->Opcode())
986 return NULL;
987 return st->in(MemNode::ValueIn);
988 }
989
990 intptr_t offset = 0; // scratch
991
992 // A load from a freshly-created object always returns zero.
993 // (This can happen after LoadNode::Ideal resets the load's memory input
994 // to find_captured_store, which returned InitializeNode::zero_memory.)
995 if (st->is_Proj() && st->in(0)->is_Allocate() &&
996 st->in(0) == AllocateNode::Ideal_allocation(ld_adr, phase, offset) &&
997 offset >= st->in(0)->as_Allocate()->minimum_header_size()) {
998 // return a zero value for the load's basic type
999 // (This is one of the few places where a generic PhaseTransform
1000 // can create new nodes. Think of it as lazily manifesting
1001 // virtually pre-existing constants.)
1002 return phase->zerocon(memory_type());
1003 }
1004
1005 // A load from an initialization barrier can match a captured store.
1006 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1007 InitializeNode* init = st->in(0)->as_Initialize();
1008 AllocateNode* alloc = init->allocation();
1009 if (alloc != NULL &&
1010 alloc == AllocateNode::Ideal_allocation(ld_adr, phase, offset)) {
1011 // examine a captured store value
1012 st = init->find_captured_store(offset, memory_size(), phase);
1013 if (st != NULL)
1014 continue; // take one more trip around
1015 }
1016 }
1017
1018 break;
1019 }
1020
1021 return NULL;
1022 }
1023
1024 //----------------------is_instance_field_load_with_local_phi------------------
1025 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1026 if( in(MemNode::Memory)->is_Phi() && in(MemNode::Memory)->in(0) == ctrl &&
1027 in(MemNode::Address)->is_AddP() ) {
1028 const TypeOopPtr* t_oop = in(MemNode::Address)->bottom_type()->isa_oopptr();
1029 // Only instances.
1030 if( t_oop != NULL && t_oop->is_known_instance_field() &&
1031 t_oop->offset() != Type::OffsetBot &&
1032 t_oop->offset() != Type::OffsetTop) {
1033 return true;
1034 }
1035 }
1036 return false;
1037 }
1038
1039 //------------------------------Identity---------------------------------------
1040 // Loads are identity if previous store is to same address
1041 Node *LoadNode::Identity( PhaseTransform *phase ) {
1042 // If the previous store-maker is the right kind of Store, and the store is
1043 // to the same address, then we are equal to the value stored.
1044 Node* mem = in(MemNode::Memory);
1045 Node* value = can_see_stored_value(mem, phase);
1046 if( value ) {
1047 // byte, short & char stores truncate naturally.
1048 // A load has to load the truncated value which requires
1049 // some sort of masking operation and that requires an
1050 // Ideal call instead of an Identity call.
1051 if (memory_size() < BytesPerInt) {
1052 // If the input to the store does not fit with the load's result type,
1053 // it must be truncated via an Ideal call.
1054 if (!phase->type(value)->higher_equal(phase->type(this)))
1055 return this;
1056 }
1057 // (This works even when value is a Con, but LoadNode::Value
1058 // usually runs first, producing the singleton type of the Con.)
1059 return value;
1060 }
1061
1062 // Search for an existing data phi which was generated before for the same
1063 // instance's field to avoid infinite generation of phis in a loop.
1064 Node *region = mem->in(0);
1065 if (is_instance_field_load_with_local_phi(region)) {
1066 const TypePtr *addr_t = in(MemNode::Address)->bottom_type()->isa_ptr();
1067 int this_index = phase->C->get_alias_index(addr_t);
1068 int this_offset = addr_t->offset();
1069 int this_id = addr_t->is_oopptr()->instance_id();
1070 const Type* this_type = bottom_type();
1071 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1072 Node* phi = region->fast_out(i);
1073 if (phi->is_Phi() && phi != mem &&
1074 phi->as_Phi()->is_same_inst_field(this_type, this_id, this_index, this_offset)) {
1075 return phi;
1076 }
1077 }
1078 }
1079
1080 return this;
1081 }
1082
1083
1084 // Returns true if the AliasType refers to the field that holds the
1085 // cached box array. Currently only handles the IntegerCache case.
1086 static bool is_autobox_cache(Compile::AliasType* atp) {
1087 if (atp != NULL && atp->field() != NULL) {
1088 ciField* field = atp->field();
1089 ciSymbol* klass = field->holder()->name();
1090 if (field->name() == ciSymbol::cache_field_name() &&
1091 field->holder()->uses_default_loader() &&
1092 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
1093 return true;
1094 }
1095 }
1096 return false;
1097 }
1098
1099 // Fetch the base value in the autobox array
1100 static bool fetch_autobox_base(Compile::AliasType* atp, int& cache_offset) {
1101 if (atp != NULL && atp->field() != NULL) {
1102 ciField* field = atp->field();
1103 ciSymbol* klass = field->holder()->name();
1104 if (field->name() == ciSymbol::cache_field_name() &&
1105 field->holder()->uses_default_loader() &&
1106 klass == ciSymbol::java_lang_Integer_IntegerCache()) {
1107 assert(field->is_constant(), "what?");
1108 ciObjArray* array = field->constant_value().as_object()->as_obj_array();
1109 // Fetch the box object at the base of the array and get its value
1110 ciInstance* box = array->obj_at(0)->as_instance();
1111 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1112 if (ik->nof_nonstatic_fields() == 1) {
1113 // This should be true nonstatic_field_at requires calling
1114 // nof_nonstatic_fields so check it anyway
1115 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1116 cache_offset = c.as_int();
1117 }
1118 return true;
1119 }
1120 }
1121 return false;
1122 }
1123
1124 // Returns true if the AliasType refers to the value field of an
1125 // autobox object. Currently only handles Integer.
1126 static bool is_autobox_object(Compile::AliasType* atp) {
1127 if (atp != NULL && atp->field() != NULL) {
1128 ciField* field = atp->field();
1129 ciSymbol* klass = field->holder()->name();
1130 if (field->name() == ciSymbol::value_name() &&
1131 field->holder()->uses_default_loader() &&
1132 klass == ciSymbol::java_lang_Integer()) {
1133 return true;
1134 }
1135 }
1136 return false;
1137 }
1138
1139
1140 // We're loading from an object which has autobox behaviour.
1141 // If this object is result of a valueOf call we'll have a phi
1142 // merging a newly allocated object and a load from the cache.
1143 // We want to replace this load with the original incoming
1144 // argument to the valueOf call.
1145 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1146 Node* base = in(Address)->in(AddPNode::Base);
1147 if (base->is_Phi() && base->req() == 3) {
1148 AllocateNode* allocation = NULL;
1149 int allocation_index = -1;
1150 int load_index = -1;
1151 for (uint i = 1; i < base->req(); i++) {
1152 allocation = AllocateNode::Ideal_allocation(base->in(i), phase);
1153 if (allocation != NULL) {
1154 allocation_index = i;
1155 load_index = 3 - allocation_index;
1156 break;
1157 }
1158 }
1159 bool has_load = ( allocation != NULL &&
1160 (base->in(load_index)->is_Load() ||
1161 base->in(load_index)->is_DecodeN() &&
1162 base->in(load_index)->in(1)->is_Load()) );
1163 if (has_load && in(Memory)->is_Phi() && in(Memory)->in(0) == base->in(0)) {
1164 // Push the loads from the phi that comes from valueOf up
1165 // through it to allow elimination of the loads and the recovery
1166 // of the original value.
1167 Node* mem_phi = in(Memory);
1168 Node* offset = in(Address)->in(AddPNode::Offset);
1169 Node* region = base->in(0);
1170
1171 Node* in1 = clone();
1172 Node* in1_addr = in1->in(Address)->clone();
1173 in1_addr->set_req(AddPNode::Base, base->in(allocation_index));
1174 in1_addr->set_req(AddPNode::Address, base->in(allocation_index));
1175 in1_addr->set_req(AddPNode::Offset, offset);
1176 in1->set_req(0, region->in(allocation_index));
1177 in1->set_req(Address, in1_addr);
1178 in1->set_req(Memory, mem_phi->in(allocation_index));
1179
1180 Node* in2 = clone();
1181 Node* in2_addr = in2->in(Address)->clone();
1182 in2_addr->set_req(AddPNode::Base, base->in(load_index));
1183 in2_addr->set_req(AddPNode::Address, base->in(load_index));
1184 in2_addr->set_req(AddPNode::Offset, offset);
1185 in2->set_req(0, region->in(load_index));
1186 in2->set_req(Address, in2_addr);
1187 in2->set_req(Memory, mem_phi->in(load_index));
1188
1189 in1_addr = phase->transform(in1_addr);
1190 in1 = phase->transform(in1);
1191 in2_addr = phase->transform(in2_addr);
1192 in2 = phase->transform(in2);
1193
1194 PhiNode* result = PhiNode::make_blank(region, this);
1195 result->set_req(allocation_index, in1);
1196 result->set_req(load_index, in2);
1197 return result;
1198 }
1199 } else if (base->is_Load() ||
1200 base->is_DecodeN() && base->in(1)->is_Load()) {
1201 if (base->is_DecodeN()) {
1202 // Get LoadN node which loads cached Integer object
1203 base = base->in(1);
1204 }
1205 // Eliminate the load of Integer.value for integers from the cache
1206 // array by deriving the value from the index into the array.
1207 // Capture the offset of the load and then reverse the computation.
1208 Node* load_base = base->in(Address)->in(AddPNode::Base);
1209 if (load_base->is_DecodeN()) {
1210 // Get LoadN node which loads IntegerCache.cache field
1211 load_base = load_base->in(1);
1212 }
1213 if (load_base != NULL) {
1214 Compile::AliasType* atp = phase->C->alias_type(load_base->adr_type());
1215 intptr_t cache_offset;
1216 int shift = -1;
1217 Node* cache = NULL;
1218 if (is_autobox_cache(atp)) {
1219 shift = exact_log2(type2aelembytes(T_OBJECT));
1220 cache = AddPNode::Ideal_base_and_offset(load_base->in(Address), phase, cache_offset);
1221 }
1222 if (cache != NULL && base->in(Address)->is_AddP()) {
1223 Node* elements[4];
1224 int count = base->in(Address)->as_AddP()->unpack_offsets(elements, ARRAY_SIZE(elements));
1225 int cache_low;
1226 if (count > 0 && fetch_autobox_base(atp, cache_low)) {
1227 int offset = arrayOopDesc::base_offset_in_bytes(memory_type()) - (cache_low << shift);
1228 // Add up all the offsets making of the address of the load
1229 Node* result = elements[0];
1230 for (int i = 1; i < count; i++) {
1231 result = phase->transform(new (phase->C, 3) AddXNode(result, elements[i]));
1232 }
1233 // Remove the constant offset from the address and then
1234 // remove the scaling of the offset to recover the original index.
1235 result = phase->transform(new (phase->C, 3) AddXNode(result, phase->MakeConX(-offset)));
1236 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1237 // Peel the shift off directly but wrap it in a dummy node
1238 // since Ideal can't return existing nodes
1239 result = new (phase->C, 3) RShiftXNode(result->in(1), phase->intcon(0));
1240 } else {
1241 result = new (phase->C, 3) RShiftXNode(result, phase->intcon(shift));
1242 }
1243 #ifdef _LP64
1244 result = new (phase->C, 2) ConvL2INode(phase->transform(result));
1245 #endif
1246 return result;
1247 }
1248 }
1249 }
1250 }
1251 return NULL;
1252 }
1253
1254 //------------------------------split_through_phi------------------------------
1255 // Split instance field load through Phi.
1256 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1257 Node* mem = in(MemNode::Memory);
1258 Node* address = in(MemNode::Address);
1259 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1260 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1261
1262 assert(mem->is_Phi() && (t_oop != NULL) &&
1263 t_oop->is_known_instance_field(), "invalide conditions");
1264
1265 Node *region = mem->in(0);
1266 if (region == NULL) {
1267 return NULL; // Wait stable graph
1268 }
1269 uint cnt = mem->req();
1270 for (uint i = 1; i < cnt; i++) {
1271 Node* rc = region->in(i);
1272 if (rc == NULL || phase->type(rc) == Type::TOP)
1273 return NULL; // Wait stable graph
1274 Node *in = mem->in(i);
1275 if (in == NULL) {
1276 return NULL; // Wait stable graph
1277 }
1278 }
1279 // Check for loop invariant.
1280 if (cnt == 3) {
1281 for (uint i = 1; i < cnt; i++) {
1282 Node *in = mem->in(i);
1283 Node* m = MemNode::optimize_memory_chain(in, addr_t, phase);
1284 if (m == mem) {
1285 set_req(MemNode::Memory, mem->in(cnt - i)); // Skip this phi.
1286 return this;
1287 }
1288 }
1289 }
1290 // Split through Phi (see original code in loopopts.cpp).
1291 assert(phase->C->have_alias_type(addr_t), "instance should have alias type");
1292
1293 // Do nothing here if Identity will find a value
1294 // (to avoid infinite chain of value phis generation).
1295 if (!phase->eqv(this, this->Identity(phase)))
1296 return NULL;
1297
1298 // Skip the split if the region dominates some control edge of the address.
1299 if (!MemNode::all_controls_dominate(address, region))
1300 return NULL;
1301
1302 const Type* this_type = this->bottom_type();
1303 int this_index = phase->C->get_alias_index(addr_t);
1304 int this_offset = addr_t->offset();
1305 int this_iid = addr_t->is_oopptr()->instance_id();
1306 PhaseIterGVN *igvn = phase->is_IterGVN();
1307 Node *phi = new (igvn->C, region->req()) PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1308 for (uint i = 1; i < region->req(); i++) {
1309 Node *x;
1310 Node* the_clone = NULL;
1311 if (region->in(i) == phase->C->top()) {
1312 x = phase->C->top(); // Dead path? Use a dead data op
1313 } else {
1314 x = this->clone(); // Else clone up the data op
1315 the_clone = x; // Remember for possible deletion.
1316 // Alter data node to use pre-phi inputs
1317 if (this->in(0) == region) {
1318 x->set_req(0, region->in(i));
1319 } else {
1320 x->set_req(0, NULL);
1321 }
1322 for (uint j = 1; j < this->req(); j++) {
1323 Node *in = this->in(j);
1324 if (in->is_Phi() && in->in(0) == region)
1325 x->set_req(j, in->in(i)); // Use pre-Phi input for the clone
1326 }
1327 }
1328 // Check for a 'win' on some paths
1329 const Type *t = x->Value(igvn);
1330
1331 bool singleton = t->singleton();
1332
1333 // See comments in PhaseIdealLoop::split_thru_phi().
1334 if (singleton && t == Type::TOP) {
1335 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1336 }
1337
1338 if (singleton) {
1339 x = igvn->makecon(t);
1340 } else {
1341 // We now call Identity to try to simplify the cloned node.
1342 // Note that some Identity methods call phase->type(this).
1343 // Make sure that the type array is big enough for
1344 // our new node, even though we may throw the node away.
1345 // (This tweaking with igvn only works because x is a new node.)
1346 igvn->set_type(x, t);
1347 // If x is a TypeNode, capture any more-precise type permanently into Node
1348 // otherwise it will be not updated during igvn->transform since
1349 // igvn->type(x) is set to x->Value() already.
1350 x->raise_bottom_type(t);
1351 Node *y = x->Identity(igvn);
1352 if (y != x) {
1353 x = y;
1354 } else {
1355 y = igvn->hash_find(x);
1356 if (y) {
1357 x = y;
1358 } else {
1359 // Else x is a new node we are keeping
1360 // We do not need register_new_node_with_optimizer
1361 // because set_type has already been called.
1362 igvn->_worklist.push(x);
1363 }
1364 }
1365 }
1366 if (x != the_clone && the_clone != NULL)
1367 igvn->remove_dead_node(the_clone);
1368 phi->set_req(i, x);
1369 }
1370 // Record Phi
1371 igvn->register_new_node_with_optimizer(phi);
1372 return phi;
1373 }
1374
1375 //------------------------------Ideal------------------------------------------
1376 // If the load is from Field memory and the pointer is non-null, we can
1377 // zero out the control input.
1378 // If the offset is constant and the base is an object allocation,
1379 // try to hook me up to the exact initializing store.
1380 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1381 Node* p = MemNode::Ideal_common(phase, can_reshape);
1382 if (p) return (p == NodeSentinel) ? NULL : p;
1383
1384 Node* ctrl = in(MemNode::Control);
1385 Node* address = in(MemNode::Address);
1386
1387 // Skip up past a SafePoint control. Cannot do this for Stores because
1388 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1389 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1390 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1391 ctrl = ctrl->in(0);
1392 set_req(MemNode::Control,ctrl);
1393 }
1394
1395 intptr_t ignore = 0;
1396 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1397 if (base != NULL
1398 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1399 // Check for useless control edge in some common special cases
1400 if (in(MemNode::Control) != NULL
1401 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1402 && all_controls_dominate(base, phase->C->start())) {
1403 // A method-invariant, non-null address (constant or 'this' argument).
1404 set_req(MemNode::Control, NULL);
1405 }
1406
1407 if (EliminateAutoBox && can_reshape) {
1408 assert(!phase->type(base)->higher_equal(TypePtr::NULL_PTR), "the autobox pointer should be non-null");
1409 Compile::AliasType* atp = phase->C->alias_type(adr_type());
1410 if (is_autobox_object(atp)) {
1411 Node* result = eliminate_autobox(phase);
1412 if (result != NULL) return result;
1413 }
1414 }
1415 }
1416
1417 Node* mem = in(MemNode::Memory);
1418 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1419
1420 if (addr_t != NULL) {
1421 // try to optimize our memory input
1422 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, phase);
1423 if (opt_mem != mem) {
1424 set_req(MemNode::Memory, opt_mem);
1425 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1426 return this;
1427 }
1428 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1429 if (can_reshape && opt_mem->is_Phi() &&
1430 (t_oop != NULL) && t_oop->is_known_instance_field()) {
1431 PhaseIterGVN *igvn = phase->is_IterGVN();
1432 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1433 // Delay this transformation until memory Phi is processed.
1434 phase->is_IterGVN()->_worklist.push(this);
1435 return NULL;
1436 }
1437 // Split instance field load through Phi.
1438 Node* result = split_through_phi(phase);
1439 if (result != NULL) return result;
1440 }
1441 }
1442
1443 // Check for prior store with a different base or offset; make Load
1444 // independent. Skip through any number of them. Bail out if the stores
1445 // are in an endless dead cycle and report no progress. This is a key
1446 // transform for Reflection. However, if after skipping through the Stores
1447 // we can't then fold up against a prior store do NOT do the transform as
1448 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1449 // array memory alive twice: once for the hoisted Load and again after the
1450 // bypassed Store. This situation only works if EVERYBODY who does
1451 // anti-dependence work knows how to bypass. I.e. we need all
1452 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1453 // the alias index stuff. So instead, peek through Stores and IFF we can
1454 // fold up, do so.
1455 Node* prev_mem = find_previous_store(phase);
1456 // Steps (a), (b): Walk past independent stores to find an exact match.
1457 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1458 // (c) See if we can fold up on the spot, but don't fold up here.
1459 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1460 // just return a prior value, which is done by Identity calls.
1461 if (can_see_stored_value(prev_mem, phase)) {
1462 // Make ready for step (d):
1463 set_req(MemNode::Memory, prev_mem);
1464 return this;
1465 }
1466 }
1467
1468 return NULL; // No further progress
1469 }
1470
1471 // Helper to recognize certain Klass fields which are invariant across
1472 // some group of array types (e.g., int[] or all T[] where T < Object).
1473 const Type*
1474 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1475 ciKlass* klass) const {
1476 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1477 // The field is Klass::_modifier_flags. Return its (constant) value.
1478 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1479 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1480 return TypeInt::make(klass->modifier_flags());
1481 }
1482 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1483 // The field is Klass::_access_flags. Return its (constant) value.
1484 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1485 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1486 return TypeInt::make(klass->access_flags());
1487 }
1488 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1489 // The field is Klass::_layout_helper. Return its constant value if known.
1490 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1491 return TypeInt::make(klass->layout_helper());
1492 }
1493
1494 // No match.
1495 return NULL;
1496 }
1497
1498 //------------------------------Value-----------------------------------------
1499 const Type *LoadNode::Value( PhaseTransform *phase ) const {
1500 // Either input is TOP ==> the result is TOP
1501 Node* mem = in(MemNode::Memory);
1502 const Type *t1 = phase->type(mem);
1503 if (t1 == Type::TOP) return Type::TOP;
1504 Node* adr = in(MemNode::Address);
1505 const TypePtr* tp = phase->type(adr)->isa_ptr();
1506 if (tp == NULL || tp->empty()) return Type::TOP;
1507 int off = tp->offset();
1508 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1509 Compile* C = phase->C;
1510
1511 // Try to guess loaded type from pointer type
1512 if (tp->base() == Type::AryPtr) {
1513 const Type *t = tp->is_aryptr()->elem();
1514 // Don't do this for integer types. There is only potential profit if
1515 // the element type t is lower than _type; that is, for int types, if _type is
1516 // more restrictive than t. This only happens here if one is short and the other
1517 // char (both 16 bits), and in those cases we've made an intentional decision
1518 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1519 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1520 //
1521 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1522 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1523 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1524 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1525 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1526 // In fact, that could have been the original type of p1, and p1 could have
1527 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1528 // expression (LShiftL quux 3) independently optimized to the constant 8.
1529 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1530 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1531 // t might actually be lower than _type, if _type is a unique
1532 // concrete subclass of abstract class t.
1533 // Make sure the reference is not into the header, by comparing
1534 // the offset against the offset of the start of the array's data.
1535 // Different array types begin at slightly different offsets (12 vs. 16).
1536 // We choose T_BYTE as an example base type that is least restrictive
1537 // as to alignment, which will therefore produce the smallest
1538 // possible base offset.
1539 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1540 if ((uint)off >= (uint)min_base_off) { // is the offset beyond the header?
1541 const Type* jt = t->join(_type);
1542 // In any case, do not allow the join, per se, to empty out the type.
1543 if (jt->empty() && !t->empty()) {
1544 // This can happen if a interface-typed array narrows to a class type.
1545 jt = _type;
1546 }
1547
1548 if (EliminateAutoBox && adr->is_AddP()) {
1549 // The pointers in the autobox arrays are always non-null
1550 Node* base = adr->in(AddPNode::Base);
1551 if (base != NULL &&
1552 !phase->type(base)->higher_equal(TypePtr::NULL_PTR)) {
1553 Compile::AliasType* atp = C->alias_type(base->adr_type());
1554 if (is_autobox_cache(atp)) {
1555 return jt->join(TypePtr::NOTNULL)->is_ptr();
1556 }
1557 }
1558 }
1559 return jt;
1560 }
1561 }
1562 } else if (tp->base() == Type::InstPtr) {
1563 ciEnv* env = C->env();
1564 const TypeInstPtr* tinst = tp->is_instptr();
1565 ciKlass* klass = tinst->klass();
1566 assert( off != Type::OffsetBot ||
1567 // arrays can be cast to Objects
1568 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1569 // unsafe field access may not have a constant offset
1570 C->has_unsafe_access(),
1571 "Field accesses must be precise" );
1572 // For oop loads, we expect the _type to be precise
1573 if (klass == env->String_klass() &&
1574 adr->is_AddP() && off != Type::OffsetBot) {
1575 // For constant Strings treat the final fields as compile time constants.
1576 Node* base = adr->in(AddPNode::Base);
1577 const TypeOopPtr* t = phase->type(base)->isa_oopptr();
1578 if (t != NULL && t->singleton()) {
1579 ciField* field = env->String_klass()->get_field_by_offset(off, false);
1580 if (field != NULL && field->is_final()) {
1581 ciObject* string = t->const_oop();
1582 ciConstant constant = string->as_instance()->field_value(field);
1583 if (constant.basic_type() == T_INT) {
1584 return TypeInt::make(constant.as_int());
1585 } else if (constant.basic_type() == T_ARRAY) {
1586 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1587 return TypeNarrowOop::make_from_constant(constant.as_object(), true);
1588 } else {
1589 return TypeOopPtr::make_from_constant(constant.as_object(), true);
1590 }
1591 }
1592 }
1593 }
1594 }
1595 // Optimizations for constant objects
1596 ciObject* const_oop = tinst->const_oop();
1597 if (const_oop != NULL) {
1598 // For constant CallSites treat the target field as a compile time constant.
1599 if (const_oop->is_call_site()) {
1600 ciCallSite* call_site = const_oop->as_call_site();
1601 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);
1602 if (field != NULL && field->is_call_site_target()) {
1603 ciMethodHandle* target = call_site->get_target();
1604 if (target != NULL) { // just in case
1605 ciConstant constant(T_OBJECT, target);
1606 const Type* t;
1607 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1608 t = TypeNarrowOop::make_from_constant(constant.as_object(), true);
1609 } else {
1610 t = TypeOopPtr::make_from_constant(constant.as_object(), true);
1611 }
1612 // Add a dependence for invalidation of the optimization.
1613 if (!call_site->is_constant_call_site()) {
1614 C->dependencies()->assert_call_site_target_value(call_site, target);
1615 }
1616 return t;
1617 }
1618 }
1619 }
1620 }
1621 } else if (tp->base() == Type::KlassPtr) {
1622 assert( off != Type::OffsetBot ||
1623 // arrays can be cast to Objects
1624 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1625 // also allow array-loading from the primary supertype
1626 // array during subtype checks
1627 Opcode() == Op_LoadKlass,
1628 "Field accesses must be precise" );
1629 // For klass/static loads, we expect the _type to be precise
1630 }
1631
1632 const TypeKlassPtr *tkls = tp->isa_klassptr();
1633 if (tkls != NULL && !StressReflectiveCode) {
1634 ciKlass* klass = tkls->klass();
1635 if (klass->is_loaded() && tkls->klass_is_exact()) {
1636 // We are loading a field from a Klass metaobject whose identity
1637 // is known at compile time (the type is "exact" or "precise").
1638 // Check for fields we know are maintained as constants by the VM.
1639 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1640 // The field is Klass::_super_check_offset. Return its (constant) value.
1641 // (Folds up type checking code.)
1642 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1643 return TypeInt::make(klass->super_check_offset());
1644 }
1645 // Compute index into primary_supers array
1646 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(klassOop);
1647 // Check for overflowing; use unsigned compare to handle the negative case.
1648 if( depth < ciKlass::primary_super_limit() ) {
1649 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1650 // (Folds up type checking code.)
1651 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1652 ciKlass *ss = klass->super_of_depth(depth);
1653 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1654 }
1655 const Type* aift = load_array_final_field(tkls, klass);
1656 if (aift != NULL) return aift;
1657 if (tkls->offset() == in_bytes(arrayKlass::component_mirror_offset())
1658 && klass->is_array_klass()) {
1659 // The field is arrayKlass::_component_mirror. Return its (constant) value.
1660 // (Folds up aClassConstant.getComponentType, common in Arrays.copyOf.)
1661 assert(Opcode() == Op_LoadP, "must load an oop from _component_mirror");
1662 return TypeInstPtr::make(klass->as_array_klass()->component_mirror());
1663 }
1664 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1665 // The field is Klass::_java_mirror. Return its (constant) value.
1666 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1667 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1668 return TypeInstPtr::make(klass->java_mirror());
1669 }
1670 }
1671
1672 // We can still check if we are loading from the primary_supers array at a
1673 // shallow enough depth. Even though the klass is not exact, entries less
1674 // than or equal to its super depth are correct.
1675 if (klass->is_loaded() ) {
1676 ciType *inner = klass->klass();
1677 while( inner->is_obj_array_klass() )
1678 inner = inner->as_obj_array_klass()->base_element_type();
1679 if( inner->is_instance_klass() &&
1680 !inner->as_instance_klass()->flags().is_interface() ) {
1681 // Compute index into primary_supers array
1682 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(klassOop);
1683 // Check for overflowing; use unsigned compare to handle the negative case.
1684 if( depth < ciKlass::primary_super_limit() &&
1685 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1686 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1687 // (Folds up type checking code.)
1688 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1689 ciKlass *ss = klass->super_of_depth(depth);
1690 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1691 }
1692 }
1693 }
1694
1695 // If the type is enough to determine that the thing is not an array,
1696 // we can give the layout_helper a positive interval type.
1697 // This will help short-circuit some reflective code.
1698 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1699 && !klass->is_array_klass() // not directly typed as an array
1700 && !klass->is_interface() // specifically not Serializable & Cloneable
1701 && !klass->is_java_lang_Object() // not the supertype of all T[]
1702 ) {
1703 // Note: When interfaces are reliable, we can narrow the interface
1704 // test to (klass != Serializable && klass != Cloneable).
1705 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1706 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1707 // The key property of this type is that it folds up tests
1708 // for array-ness, since it proves that the layout_helper is positive.
1709 // Thus, a generic value like the basic object layout helper works fine.
1710 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1711 }
1712 }
1713
1714 // If we are loading from a freshly-allocated object, produce a zero,
1715 // if the load is provably beyond the header of the object.
1716 // (Also allow a variable load from a fresh array to produce zero.)
1717 const TypeOopPtr *tinst = tp->isa_oopptr();
1718 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1719 if (ReduceFieldZeroing || is_instance) {
1720 Node* value = can_see_stored_value(mem,phase);
1721 if (value != NULL && value->is_Con()) {
1722 assert(value->bottom_type()->higher_equal(_type),"sanity");
1723 return value->bottom_type();
1724 }
1725 }
1726
1727 if (is_instance) {
1728 // If we have an instance type and our memory input is the
1729 // programs's initial memory state, there is no matching store,
1730 // so just return a zero of the appropriate type
1731 Node *mem = in(MemNode::Memory);
1732 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1733 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1734 return Type::get_zero_type(_type->basic_type());
1735 }
1736 }
1737 return _type;
1738 }
1739
1740 //------------------------------match_edge-------------------------------------
1741 // Do we Match on this edge index or not? Match only the address.
1742 uint LoadNode::match_edge(uint idx) const {
1743 return idx == MemNode::Address;
1744 }
1745
1746 //--------------------------LoadBNode::Ideal--------------------------------------
1747 //
1748 // If the previous store is to the same address as this load,
1749 // and the value stored was larger than a byte, replace this load
1750 // with the value stored truncated to a byte. If no truncation is
1751 // needed, the replacement is done in LoadNode::Identity().
1752 //
1753 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1754 Node* mem = in(MemNode::Memory);
1755 Node* value = can_see_stored_value(mem,phase);
1756 if( value && !phase->type(value)->higher_equal( _type ) ) {
1757 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(24)) );
1758 return new (phase->C, 3) RShiftINode(result, phase->intcon(24));
1759 }
1760 // Identity call will handle the case where truncation is not needed.
1761 return LoadNode::Ideal(phase, can_reshape);
1762 }
1763
1764 const Type* LoadBNode::Value(PhaseTransform *phase) const {
1765 Node* mem = in(MemNode::Memory);
1766 Node* value = can_see_stored_value(mem,phase);
1767 if (value != NULL && value->is_Con() &&
1768 !value->bottom_type()->higher_equal(_type)) {
1769 // If the input to the store does not fit with the load's result type,
1770 // it must be truncated. We can't delay until Ideal call since
1771 // a singleton Value is needed for split_thru_phi optimization.
1772 int con = value->get_int();
1773 return TypeInt::make((con << 24) >> 24);
1774 }
1775 return LoadNode::Value(phase);
1776 }
1777
1778 //--------------------------LoadUBNode::Ideal-------------------------------------
1779 //
1780 // If the previous store is to the same address as this load,
1781 // and the value stored was larger than a byte, replace this load
1782 // with the value stored truncated to a byte. If no truncation is
1783 // needed, the replacement is done in LoadNode::Identity().
1784 //
1785 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1786 Node* mem = in(MemNode::Memory);
1787 Node* value = can_see_stored_value(mem, phase);
1788 if (value && !phase->type(value)->higher_equal(_type))
1789 return new (phase->C, 3) AndINode(value, phase->intcon(0xFF));
1790 // Identity call will handle the case where truncation is not needed.
1791 return LoadNode::Ideal(phase, can_reshape);
1792 }
1793
1794 const Type* LoadUBNode::Value(PhaseTransform *phase) const {
1795 Node* mem = in(MemNode::Memory);
1796 Node* value = can_see_stored_value(mem,phase);
1797 if (value != NULL && value->is_Con() &&
1798 !value->bottom_type()->higher_equal(_type)) {
1799 // If the input to the store does not fit with the load's result type,
1800 // it must be truncated. We can't delay until Ideal call since
1801 // a singleton Value is needed for split_thru_phi optimization.
1802 int con = value->get_int();
1803 return TypeInt::make(con & 0xFF);
1804 }
1805 return LoadNode::Value(phase);
1806 }
1807
1808 //--------------------------LoadUSNode::Ideal-------------------------------------
1809 //
1810 // If the previous store is to the same address as this load,
1811 // and the value stored was larger than a char, replace this load
1812 // with the value stored truncated to a char. If no truncation is
1813 // needed, the replacement is done in LoadNode::Identity().
1814 //
1815 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1816 Node* mem = in(MemNode::Memory);
1817 Node* value = can_see_stored_value(mem,phase);
1818 if( value && !phase->type(value)->higher_equal( _type ) )
1819 return new (phase->C, 3) AndINode(value,phase->intcon(0xFFFF));
1820 // Identity call will handle the case where truncation is not needed.
1821 return LoadNode::Ideal(phase, can_reshape);
1822 }
1823
1824 const Type* LoadUSNode::Value(PhaseTransform *phase) const {
1825 Node* mem = in(MemNode::Memory);
1826 Node* value = can_see_stored_value(mem,phase);
1827 if (value != NULL && value->is_Con() &&
1828 !value->bottom_type()->higher_equal(_type)) {
1829 // If the input to the store does not fit with the load's result type,
1830 // it must be truncated. We can't delay until Ideal call since
1831 // a singleton Value is needed for split_thru_phi optimization.
1832 int con = value->get_int();
1833 return TypeInt::make(con & 0xFFFF);
1834 }
1835 return LoadNode::Value(phase);
1836 }
1837
1838 //--------------------------LoadSNode::Ideal--------------------------------------
1839 //
1840 // If the previous store is to the same address as this load,
1841 // and the value stored was larger than a short, replace this load
1842 // with the value stored truncated to a short. If no truncation is
1843 // needed, the replacement is done in LoadNode::Identity().
1844 //
1845 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1846 Node* mem = in(MemNode::Memory);
1847 Node* value = can_see_stored_value(mem,phase);
1848 if( value && !phase->type(value)->higher_equal( _type ) ) {
1849 Node *result = phase->transform( new (phase->C, 3) LShiftINode(value, phase->intcon(16)) );
1850 return new (phase->C, 3) RShiftINode(result, phase->intcon(16));
1851 }
1852 // Identity call will handle the case where truncation is not needed.
1853 return LoadNode::Ideal(phase, can_reshape);
1854 }
1855
1856 const Type* LoadSNode::Value(PhaseTransform *phase) const {
1857 Node* mem = in(MemNode::Memory);
1858 Node* value = can_see_stored_value(mem,phase);
1859 if (value != NULL && value->is_Con() &&
1860 !value->bottom_type()->higher_equal(_type)) {
1861 // If the input to the store does not fit with the load's result type,
1862 // it must be truncated. We can't delay until Ideal call since
1863 // a singleton Value is needed for split_thru_phi optimization.
1864 int con = value->get_int();
1865 return TypeInt::make((con << 16) >> 16);
1866 }
1867 return LoadNode::Value(phase);
1868 }
1869
1870 //=============================================================================
1871 //----------------------------LoadKlassNode::make------------------------------
1872 // Polymorphic factory method:
1873 Node *LoadKlassNode::make( PhaseGVN& gvn, Node *mem, Node *adr, const TypePtr* at, const TypeKlassPtr *tk ) {
1874 Compile* C = gvn.C;
1875 Node *ctl = NULL;
1876 // sanity check the alias category against the created node type
1877 const TypeOopPtr *adr_type = adr->bottom_type()->isa_oopptr();
1878 assert(adr_type != NULL, "expecting TypeOopPtr");
1879 #ifdef _LP64
1880 if (adr_type->is_ptr_to_narrowoop()) {
1881 Node* load_klass = gvn.transform(new (C, 3) LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowoop()));
1882 return new (C, 2) DecodeNNode(load_klass, load_klass->bottom_type()->make_ptr());
1883 }
1884 #endif
1885 assert(!adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
1886 return new (C, 3) LoadKlassNode(ctl, mem, adr, at, tk);
1887 }
1888
1889 //------------------------------Value------------------------------------------
1890 const Type *LoadKlassNode::Value( PhaseTransform *phase ) const {
1891 return klass_value_common(phase);
1892 }
1893
1894 const Type *LoadNode::klass_value_common( PhaseTransform *phase ) const {
1895 // Either input is TOP ==> the result is TOP
1896 const Type *t1 = phase->type( in(MemNode::Memory) );
1897 if (t1 == Type::TOP) return Type::TOP;
1898 Node *adr = in(MemNode::Address);
1899 const Type *t2 = phase->type( adr );
1900 if (t2 == Type::TOP) return Type::TOP;
1901 const TypePtr *tp = t2->is_ptr();
1902 if (TypePtr::above_centerline(tp->ptr()) ||
1903 tp->ptr() == TypePtr::Null) return Type::TOP;
1904
1905 // Return a more precise klass, if possible
1906 const TypeInstPtr *tinst = tp->isa_instptr();
1907 if (tinst != NULL) {
1908 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
1909 int offset = tinst->offset();
1910 if (ik == phase->C->env()->Class_klass()
1911 && (offset == java_lang_Class::klass_offset_in_bytes() ||
1912 offset == java_lang_Class::array_klass_offset_in_bytes())) {
1913 // We are loading a special hidden field from a Class mirror object,
1914 // the field which points to the VM's Klass metaobject.
1915 ciType* t = tinst->java_mirror_type();
1916 // java_mirror_type returns non-null for compile-time Class constants.
1917 if (t != NULL) {
1918 // constant oop => constant klass
1919 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
1920 return TypeKlassPtr::make(ciArrayKlass::make(t));
1921 }
1922 if (!t->is_klass()) {
1923 // a primitive Class (e.g., int.class) has NULL for a klass field
1924 return TypePtr::NULL_PTR;
1925 }
1926 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
1927 return TypeKlassPtr::make(t->as_klass());
1928 }
1929 // non-constant mirror, so we can't tell what's going on
1930 }
1931 if( !ik->is_loaded() )
1932 return _type; // Bail out if not loaded
1933 if (offset == oopDesc::klass_offset_in_bytes()) {
1934 if (tinst->klass_is_exact()) {
1935 return TypeKlassPtr::make(ik);
1936 }
1937 // See if we can become precise: no subklasses and no interface
1938 // (Note: We need to support verified interfaces.)
1939 if (!ik->is_interface() && !ik->has_subklass()) {
1940 //assert(!UseExactTypes, "this code should be useless with exact types");
1941 // Add a dependence; if any subclass added we need to recompile
1942 if (!ik->is_final()) {
1943 // %%% should use stronger assert_unique_concrete_subtype instead
1944 phase->C->dependencies()->assert_leaf_type(ik);
1945 }
1946 // Return precise klass
1947 return TypeKlassPtr::make(ik);
1948 }
1949
1950 // Return root of possible klass
1951 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
1952 }
1953 }
1954
1955 // Check for loading klass from an array
1956 const TypeAryPtr *tary = tp->isa_aryptr();
1957 if( tary != NULL ) {
1958 ciKlass *tary_klass = tary->klass();
1959 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
1960 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
1961 if (tary->klass_is_exact()) {
1962 return TypeKlassPtr::make(tary_klass);
1963 }
1964 ciArrayKlass *ak = tary->klass()->as_array_klass();
1965 // If the klass is an object array, we defer the question to the
1966 // array component klass.
1967 if( ak->is_obj_array_klass() ) {
1968 assert( ak->is_loaded(), "" );
1969 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
1970 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
1971 ciInstanceKlass* ik = base_k->as_instance_klass();
1972 // See if we can become precise: no subklasses and no interface
1973 if (!ik->is_interface() && !ik->has_subklass()) {
1974 //assert(!UseExactTypes, "this code should be useless with exact types");
1975 // Add a dependence; if any subclass added we need to recompile
1976 if (!ik->is_final()) {
1977 phase->C->dependencies()->assert_leaf_type(ik);
1978 }
1979 // Return precise array klass
1980 return TypeKlassPtr::make(ak);
1981 }
1982 }
1983 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
1984 } else { // Found a type-array?
1985 //assert(!UseExactTypes, "this code should be useless with exact types");
1986 assert( ak->is_type_array_klass(), "" );
1987 return TypeKlassPtr::make(ak); // These are always precise
1988 }
1989 }
1990 }
1991
1992 // Check for loading klass from an array klass
1993 const TypeKlassPtr *tkls = tp->isa_klassptr();
1994 if (tkls != NULL && !StressReflectiveCode) {
1995 ciKlass* klass = tkls->klass();
1996 if( !klass->is_loaded() )
1997 return _type; // Bail out if not loaded
1998 if( klass->is_obj_array_klass() &&
1999 tkls->offset() == in_bytes(objArrayKlass::element_klass_offset())) {
2000 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2001 // // Always returning precise element type is incorrect,
2002 // // e.g., element type could be object and array may contain strings
2003 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2004
2005 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2006 // according to the element type's subclassing.
2007 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2008 }
2009 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2010 tkls->offset() == in_bytes(Klass::super_offset())) {
2011 ciKlass* sup = klass->as_instance_klass()->super();
2012 // The field is Klass::_super. Return its (constant) value.
2013 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2014 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2015 }
2016 }
2017
2018 // Bailout case
2019 return LoadNode::Value(phase);
2020 }
2021
2022 //------------------------------Identity---------------------------------------
2023 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2024 // Also feed through the klass in Allocate(...klass...)._klass.
2025 Node* LoadKlassNode::Identity( PhaseTransform *phase ) {
2026 return klass_identity_common(phase);
2027 }
2028
2029 Node* LoadNode::klass_identity_common(PhaseTransform *phase ) {
2030 Node* x = LoadNode::Identity(phase);
2031 if (x != this) return x;
2032
2033 // Take apart the address into an oop and and offset.
2034 // Return 'this' if we cannot.
2035 Node* adr = in(MemNode::Address);
2036 intptr_t offset = 0;
2037 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2038 if (base == NULL) return this;
2039 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2040 if (toop == NULL) return this;
2041
2042 // We can fetch the klass directly through an AllocateNode.
2043 // This works even if the klass is not constant (clone or newArray).
2044 if (offset == oopDesc::klass_offset_in_bytes()) {
2045 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2046 if (allocated_klass != NULL) {
2047 return allocated_klass;
2048 }
2049 }
2050
2051 // Simplify k.java_mirror.as_klass to plain k, where k is a klassOop.
2052 // Simplify ak.component_mirror.array_klass to plain ak, ak an arrayKlass.
2053 // See inline_native_Class_query for occurrences of these patterns.
2054 // Java Example: x.getClass().isAssignableFrom(y)
2055 // Java Example: Array.newInstance(x.getClass().getComponentType(), n)
2056 //
2057 // This improves reflective code, often making the Class
2058 // mirror go completely dead. (Current exception: Class
2059 // mirrors may appear in debug info, but we could clean them out by
2060 // introducing a new debug info operator for klassOop.java_mirror).
2061 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2062 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2063 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2064 // We are loading a special hidden field from a Class mirror,
2065 // the field which points to its Klass or arrayKlass metaobject.
2066 if (base->is_Load()) {
2067 Node* adr2 = base->in(MemNode::Address);
2068 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2069 if (tkls != NULL && !tkls->empty()
2070 && (tkls->klass()->is_instance_klass() ||
2071 tkls->klass()->is_array_klass())
2072 && adr2->is_AddP()
2073 ) {
2074 int mirror_field = in_bytes(Klass::java_mirror_offset());
2075 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2076 mirror_field = in_bytes(arrayKlass::component_mirror_offset());
2077 }
2078 if (tkls->offset() == mirror_field) {
2079 return adr2->in(AddPNode::Base);
2080 }
2081 }
2082 }
2083 }
2084
2085 return this;
2086 }
2087
2088
2089 //------------------------------Value------------------------------------------
2090 const Type *LoadNKlassNode::Value( PhaseTransform *phase ) const {
2091 const Type *t = klass_value_common(phase);
2092 if (t == Type::TOP)
2093 return t;
2094
2095 return t->make_narrowoop();
2096 }
2097
2098 //------------------------------Identity---------------------------------------
2099 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2100 // Also feed through the klass in Allocate(...klass...)._klass.
2101 Node* LoadNKlassNode::Identity( PhaseTransform *phase ) {
2102 Node *x = klass_identity_common(phase);
2103
2104 const Type *t = phase->type( x );
2105 if( t == Type::TOP ) return x;
2106 if( t->isa_narrowoop()) return x;
2107
2108 return phase->transform(new (phase->C, 2) EncodePNode(x, t->make_narrowoop()));
2109 }
2110
2111 //------------------------------Value-----------------------------------------
2112 const Type *LoadRangeNode::Value( PhaseTransform *phase ) const {
2113 // Either input is TOP ==> the result is TOP
2114 const Type *t1 = phase->type( in(MemNode::Memory) );
2115 if( t1 == Type::TOP ) return Type::TOP;
2116 Node *adr = in(MemNode::Address);
2117 const Type *t2 = phase->type( adr );
2118 if( t2 == Type::TOP ) return Type::TOP;
2119 const TypePtr *tp = t2->is_ptr();
2120 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2121 const TypeAryPtr *tap = tp->isa_aryptr();
2122 if( !tap ) return _type;
2123 return tap->size();
2124 }
2125
2126 //-------------------------------Ideal---------------------------------------
2127 // Feed through the length in AllocateArray(...length...)._length.
2128 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2129 Node* p = MemNode::Ideal_common(phase, can_reshape);
2130 if (p) return (p == NodeSentinel) ? NULL : p;
2131
2132 // Take apart the address into an oop and and offset.
2133 // Return 'this' if we cannot.
2134 Node* adr = in(MemNode::Address);
2135 intptr_t offset = 0;
2136 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2137 if (base == NULL) return NULL;
2138 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2139 if (tary == NULL) return NULL;
2140
2141 // We can fetch the length directly through an AllocateArrayNode.
2142 // This works even if the length is not constant (clone or newArray).
2143 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2144 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2145 if (alloc != NULL) {
2146 Node* allocated_length = alloc->Ideal_length();
2147 Node* len = alloc->make_ideal_length(tary, phase);
2148 if (allocated_length != len) {
2149 // New CastII improves on this.
2150 return len;
2151 }
2152 }
2153 }
2154
2155 return NULL;
2156 }
2157
2158 //------------------------------Identity---------------------------------------
2159 // Feed through the length in AllocateArray(...length...)._length.
2160 Node* LoadRangeNode::Identity( PhaseTransform *phase ) {
2161 Node* x = LoadINode::Identity(phase);
2162 if (x != this) return x;
2163
2164 // Take apart the address into an oop and and offset.
2165 // Return 'this' if we cannot.
2166 Node* adr = in(MemNode::Address);
2167 intptr_t offset = 0;
2168 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2169 if (base == NULL) return this;
2170 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2171 if (tary == NULL) return this;
2172
2173 // We can fetch the length directly through an AllocateArrayNode.
2174 // This works even if the length is not constant (clone or newArray).
2175 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2176 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2177 if (alloc != NULL) {
2178 Node* allocated_length = alloc->Ideal_length();
2179 // Do not allow make_ideal_length to allocate a CastII node.
2180 Node* len = alloc->make_ideal_length(tary, phase, false);
2181 if (allocated_length == len) {
2182 // Return allocated_length only if it would not be improved by a CastII.
2183 return allocated_length;
2184 }
2185 }
2186 }
2187
2188 return this;
2189
2190 }
2191
2192 //=============================================================================
2193 //---------------------------StoreNode::make-----------------------------------
2194 // Polymorphic factory method:
2195 StoreNode* StoreNode::make( PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt ) {
2196 Compile* C = gvn.C;
2197 assert( C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2198 ctl != NULL, "raw memory operations should have control edge");
2199
2200 switch (bt) {
2201 case T_BOOLEAN:
2202 case T_BYTE: return new (C, 4) StoreBNode(ctl, mem, adr, adr_type, val);
2203 case T_INT: return new (C, 4) StoreINode(ctl, mem, adr, adr_type, val);
2204 case T_CHAR:
2205 case T_SHORT: return new (C, 4) StoreCNode(ctl, mem, adr, adr_type, val);
2206 case T_LONG: return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val);
2207 case T_FLOAT: return new (C, 4) StoreFNode(ctl, mem, adr, adr_type, val);
2208 case T_DOUBLE: return new (C, 4) StoreDNode(ctl, mem, adr, adr_type, val);
2209 case T_ADDRESS:
2210 case T_OBJECT:
2211 #ifdef _LP64
2212 if (adr->bottom_type()->is_ptr_to_narrowoop() ||
2213 (UseCompressedOops && val->bottom_type()->isa_klassptr() &&
2214 adr->bottom_type()->isa_rawptr())) {
2215 val = gvn.transform(new (C, 2) EncodePNode(val, val->bottom_type()->make_narrowoop()));
2216 return new (C, 4) StoreNNode(ctl, mem, adr, adr_type, val);
2217 } else
2218 #endif
2219 {
2220 return new (C, 4) StorePNode(ctl, mem, adr, adr_type, val);
2221 }
2222 }
2223 ShouldNotReachHere();
2224 return (StoreNode*)NULL;
2225 }
2226
2227 StoreLNode* StoreLNode::make_atomic(Compile *C, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val) {
2228 bool require_atomic = true;
2229 return new (C, 4) StoreLNode(ctl, mem, adr, adr_type, val, require_atomic);
2230 }
2231
2232
2233 //--------------------------bottom_type----------------------------------------
2234 const Type *StoreNode::bottom_type() const {
2235 return Type::MEMORY;
2236 }
2237
2238 //------------------------------hash-------------------------------------------
2239 uint StoreNode::hash() const {
2240 // unroll addition of interesting fields
2241 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2242
2243 // Since they are not commoned, do not hash them:
2244 return NO_HASH;
2245 }
2246
2247 //------------------------------Ideal------------------------------------------
2248 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2249 // When a store immediately follows a relevant allocation/initialization,
2250 // try to capture it into the initialization, or hoist it above.
2251 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2252 Node* p = MemNode::Ideal_common(phase, can_reshape);
2253 if (p) return (p == NodeSentinel) ? NULL : p;
2254
2255 Node* mem = in(MemNode::Memory);
2256 Node* address = in(MemNode::Address);
2257
2258 // Back-to-back stores to same address? Fold em up. Generally
2259 // unsafe if I have intervening uses... Also disallowed for StoreCM
2260 // since they must follow each StoreP operation. Redundant StoreCMs
2261 // are eliminated just before matching in final_graph_reshape.
2262 if (mem->is_Store() && mem->in(MemNode::Address)->eqv_uncast(address) &&
2263 mem->Opcode() != Op_StoreCM) {
2264 // Looking at a dead closed cycle of memory?
2265 assert(mem != mem->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2266
2267 assert(Opcode() == mem->Opcode() ||
2268 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw,
2269 "no mismatched stores, except on raw memory");
2270
2271 if (mem->outcnt() == 1 && // check for intervening uses
2272 mem->as_Store()->memory_size() <= this->memory_size()) {
2273 // If anybody other than 'this' uses 'mem', we cannot fold 'mem' away.
2274 // For example, 'mem' might be the final state at a conditional return.
2275 // Or, 'mem' might be used by some node which is live at the same time
2276 // 'this' is live, which might be unschedulable. So, require exactly
2277 // ONE user, the 'this' store, until such time as we clone 'mem' for
2278 // each of 'mem's uses (thus making the exactly-1-user-rule hold true).
2279 if (can_reshape) { // (%%% is this an anachronism?)
2280 set_req_X(MemNode::Memory, mem->in(MemNode::Memory),
2281 phase->is_IterGVN());
2282 } else {
2283 // It's OK to do this in the parser, since DU info is always accurate,
2284 // and the parser always refers to nodes via SafePointNode maps.
2285 set_req(MemNode::Memory, mem->in(MemNode::Memory));
2286 }
2287 return this;
2288 }
2289 }
2290
2291 // Capture an unaliased, unconditional, simple store into an initializer.
2292 // Or, if it is independent of the allocation, hoist it above the allocation.
2293 if (ReduceFieldZeroing && /*can_reshape &&*/
2294 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2295 InitializeNode* init = mem->in(0)->as_Initialize();
2296 intptr_t offset = init->can_capture_store(this, phase);
2297 if (offset > 0) {
2298 Node* moved = init->capture_store(this, offset, phase);
2299 // If the InitializeNode captured me, it made a raw copy of me,
2300 // and I need to disappear.
2301 if (moved != NULL) {
2302 // %%% hack to ensure that Ideal returns a new node:
2303 mem = MergeMemNode::make(phase->C, mem);
2304 return mem; // fold me away
2305 }
2306 }
2307 }
2308
2309 return NULL; // No further progress
2310 }
2311
2312 //------------------------------Value-----------------------------------------
2313 const Type *StoreNode::Value( PhaseTransform *phase ) const {
2314 // Either input is TOP ==> the result is TOP
2315 const Type *t1 = phase->type( in(MemNode::Memory) );
2316 if( t1 == Type::TOP ) return Type::TOP;
2317 const Type *t2 = phase->type( in(MemNode::Address) );
2318 if( t2 == Type::TOP ) return Type::TOP;
2319 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2320 if( t3 == Type::TOP ) return Type::TOP;
2321 return Type::MEMORY;
2322 }
2323
2324 //------------------------------Identity---------------------------------------
2325 // Remove redundant stores:
2326 // Store(m, p, Load(m, p)) changes to m.
2327 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2328 Node *StoreNode::Identity( PhaseTransform *phase ) {
2329 Node* mem = in(MemNode::Memory);
2330 Node* adr = in(MemNode::Address);
2331 Node* val = in(MemNode::ValueIn);
2332
2333 // Load then Store? Then the Store is useless
2334 if (val->is_Load() &&
2335 val->in(MemNode::Address)->eqv_uncast(adr) &&
2336 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2337 val->as_Load()->store_Opcode() == Opcode()) {
2338 return mem;
2339 }
2340
2341 // Two stores in a row of the same value?
2342 if (mem->is_Store() &&
2343 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2344 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2345 mem->Opcode() == Opcode()) {
2346 return mem;
2347 }
2348
2349 // Store of zero anywhere into a freshly-allocated object?
2350 // Then the store is useless.
2351 // (It must already have been captured by the InitializeNode.)
2352 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2353 // a newly allocated object is already all-zeroes everywhere
2354 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2355 return mem;
2356 }
2357
2358 // the store may also apply to zero-bits in an earlier object
2359 Node* prev_mem = find_previous_store(phase);
2360 // Steps (a), (b): Walk past independent stores to find an exact match.
2361 if (prev_mem != NULL) {
2362 Node* prev_val = can_see_stored_value(prev_mem, phase);
2363 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2364 // prev_val and val might differ by a cast; it would be good
2365 // to keep the more informative of the two.
2366 return mem;
2367 }
2368 }
2369 }
2370
2371 return this;
2372 }
2373
2374 //------------------------------match_edge-------------------------------------
2375 // Do we Match on this edge index or not? Match only memory & value
2376 uint StoreNode::match_edge(uint idx) const {
2377 return idx == MemNode::Address || idx == MemNode::ValueIn;
2378 }
2379
2380 //------------------------------cmp--------------------------------------------
2381 // Do not common stores up together. They generally have to be split
2382 // back up anyways, so do not bother.
2383 uint StoreNode::cmp( const Node &n ) const {
2384 return (&n == this); // Always fail except on self
2385 }
2386
2387 //------------------------------Ideal_masked_input-----------------------------
2388 // Check for a useless mask before a partial-word store
2389 // (StoreB ... (AndI valIn conIa) )
2390 // If (conIa & mask == mask) this simplifies to
2391 // (StoreB ... (valIn) )
2392 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2393 Node *val = in(MemNode::ValueIn);
2394 if( val->Opcode() == Op_AndI ) {
2395 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2396 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2397 set_req(MemNode::ValueIn, val->in(1));
2398 return this;
2399 }
2400 }
2401 return NULL;
2402 }
2403
2404
2405 //------------------------------Ideal_sign_extended_input----------------------
2406 // Check for useless sign-extension before a partial-word store
2407 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2408 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2409 // (StoreB ... (valIn) )
2410 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2411 Node *val = in(MemNode::ValueIn);
2412 if( val->Opcode() == Op_RShiftI ) {
2413 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2414 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2415 Node *shl = val->in(1);
2416 if( shl->Opcode() == Op_LShiftI ) {
2417 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2418 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2419 set_req(MemNode::ValueIn, shl->in(1));
2420 return this;
2421 }
2422 }
2423 }
2424 }
2425 return NULL;
2426 }
2427
2428 //------------------------------value_never_loaded-----------------------------------
2429 // Determine whether there are any possible loads of the value stored.
2430 // For simplicity, we actually check if there are any loads from the
2431 // address stored to, not just for loads of the value stored by this node.
2432 //
2433 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2434 Node *adr = in(Address);
2435 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2436 if (adr_oop == NULL)
2437 return false;
2438 if (!adr_oop->is_known_instance_field())
2439 return false; // if not a distinct instance, there may be aliases of the address
2440 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2441 Node *use = adr->fast_out(i);
2442 int opc = use->Opcode();
2443 if (use->is_Load() || use->is_LoadStore()) {
2444 return false;
2445 }
2446 }
2447 return true;
2448 }
2449
2450 //=============================================================================
2451 //------------------------------Ideal------------------------------------------
2452 // If the store is from an AND mask that leaves the low bits untouched, then
2453 // we can skip the AND operation. If the store is from a sign-extension
2454 // (a left shift, then right shift) we can skip both.
2455 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2456 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2457 if( progress != NULL ) return progress;
2458
2459 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2460 if( progress != NULL ) return progress;
2461
2462 // Finally check the default case
2463 return StoreNode::Ideal(phase, can_reshape);
2464 }
2465
2466 //=============================================================================
2467 //------------------------------Ideal------------------------------------------
2468 // If the store is from an AND mask that leaves the low bits untouched, then
2469 // we can skip the AND operation
2470 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2471 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2472 if( progress != NULL ) return progress;
2473
2474 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2475 if( progress != NULL ) return progress;
2476
2477 // Finally check the default case
2478 return StoreNode::Ideal(phase, can_reshape);
2479 }
2480
2481 //=============================================================================
2482 //------------------------------Identity---------------------------------------
2483 Node *StoreCMNode::Identity( PhaseTransform *phase ) {
2484 // No need to card mark when storing a null ptr
2485 Node* my_store = in(MemNode::OopStore);
2486 if (my_store->is_Store()) {
2487 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2488 if( t1 == TypePtr::NULL_PTR ) {
2489 return in(MemNode::Memory);
2490 }
2491 }
2492 return this;
2493 }
2494
2495 //=============================================================================
2496 //------------------------------Ideal---------------------------------------
2497 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2498 Node* progress = StoreNode::Ideal(phase, can_reshape);
2499 if (progress != NULL) return progress;
2500
2501 Node* my_store = in(MemNode::OopStore);
2502 if (my_store->is_MergeMem()) {
2503 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2504 set_req(MemNode::OopStore, mem);
2505 return this;
2506 }
2507
2508 return NULL;
2509 }
2510
2511 //------------------------------Value-----------------------------------------
2512 const Type *StoreCMNode::Value( PhaseTransform *phase ) const {
2513 // Either input is TOP ==> the result is TOP
2514 const Type *t = phase->type( in(MemNode::Memory) );
2515 if( t == Type::TOP ) return Type::TOP;
2516 t = phase->type( in(MemNode::Address) );
2517 if( t == Type::TOP ) return Type::TOP;
2518 t = phase->type( in(MemNode::ValueIn) );
2519 if( t == Type::TOP ) return Type::TOP;
2520 // If extra input is TOP ==> the result is TOP
2521 t = phase->type( in(MemNode::OopStore) );
2522 if( t == Type::TOP ) return Type::TOP;
2523
2524 return StoreNode::Value( phase );
2525 }
2526
2527
2528 //=============================================================================
2529 //----------------------------------SCMemProjNode------------------------------
2530 const Type * SCMemProjNode::Value( PhaseTransform *phase ) const
2531 {
2532 return bottom_type();
2533 }
2534
2535 //=============================================================================
2536 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : Node(5) {
2537 init_req(MemNode::Control, c );
2538 init_req(MemNode::Memory , mem);
2539 init_req(MemNode::Address, adr);
2540 init_req(MemNode::ValueIn, val);
2541 init_req( ExpectedIn, ex );
2542 init_class_id(Class_LoadStore);
2543
2544 }
2545
2546 //=============================================================================
2547 //-------------------------------adr_type--------------------------------------
2548 // Do we Match on this edge index or not? Do not match memory
2549 const TypePtr* ClearArrayNode::adr_type() const {
2550 Node *adr = in(3);
2551 return MemNode::calculate_adr_type(adr->bottom_type());
2552 }
2553
2554 //------------------------------match_edge-------------------------------------
2555 // Do we Match on this edge index or not? Do not match memory
2556 uint ClearArrayNode::match_edge(uint idx) const {
2557 return idx > 1;
2558 }
2559
2560 //------------------------------Identity---------------------------------------
2561 // Clearing a zero length array does nothing
2562 Node *ClearArrayNode::Identity( PhaseTransform *phase ) {
2563 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2564 }
2565
2566 //------------------------------Idealize---------------------------------------
2567 // Clearing a short array is faster with stores
2568 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2569 const int unit = BytesPerLong;
2570 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2571 if (!t) return NULL;
2572 if (!t->is_con()) return NULL;
2573 intptr_t raw_count = t->get_con();
2574 intptr_t size = raw_count;
2575 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2576 // Clearing nothing uses the Identity call.
2577 // Negative clears are possible on dead ClearArrays
2578 // (see jck test stmt114.stmt11402.val).
2579 if (size <= 0 || size % unit != 0) return NULL;
2580 intptr_t count = size / unit;
2581 // Length too long; use fast hardware clear
2582 if (size > Matcher::init_array_short_size) return NULL;
2583 Node *mem = in(1);
2584 if( phase->type(mem)==Type::TOP ) return NULL;
2585 Node *adr = in(3);
2586 const Type* at = phase->type(adr);
2587 if( at==Type::TOP ) return NULL;
2588 const TypePtr* atp = at->isa_ptr();
2589 // adjust atp to be the correct array element address type
2590 if (atp == NULL) atp = TypePtr::BOTTOM;
2591 else atp = atp->add_offset(Type::OffsetBot);
2592 // Get base for derived pointer purposes
2593 if( adr->Opcode() != Op_AddP ) Unimplemented();
2594 Node *base = adr->in(1);
2595
2596 Node *zero = phase->makecon(TypeLong::ZERO);
2597 Node *off = phase->MakeConX(BytesPerLong);
2598 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2599 count--;
2600 while( count-- ) {
2601 mem = phase->transform(mem);
2602 adr = phase->transform(new (phase->C, 4) AddPNode(base,adr,off));
2603 mem = new (phase->C, 4) StoreLNode(in(0),mem,adr,atp,zero);
2604 }
2605 return mem;
2606 }
2607
2608 //----------------------------step_through----------------------------------
2609 // Return allocation input memory edge if it is different instance
2610 // or itself if it is the one we are looking for.
2611 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2612 Node* n = *np;
2613 assert(n->is_ClearArray(), "sanity");
2614 intptr_t offset;
2615 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2616 // This method is called only before Allocate nodes are expanded during
2617 // macro nodes expansion. Before that ClearArray nodes are only generated
2618 // in LibraryCallKit::generate_arraycopy() which follows allocations.
2619 assert(alloc != NULL, "should have allocation");
2620 if (alloc->_idx == instance_id) {
2621 // Can not bypass initialization of the instance we are looking for.
2622 return false;
2623 }
2624 // Otherwise skip it.
2625 InitializeNode* init = alloc->initialization();
2626 if (init != NULL)
2627 *np = init->in(TypeFunc::Memory);
2628 else
2629 *np = alloc->in(TypeFunc::Memory);
2630 return true;
2631 }
2632
2633 //----------------------------clear_memory-------------------------------------
2634 // Generate code to initialize object storage to zero.
2635 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2636 intptr_t start_offset,
2637 Node* end_offset,
2638 PhaseGVN* phase) {
2639 Compile* C = phase->C;
2640 intptr_t offset = start_offset;
2641
2642 int unit = BytesPerLong;
2643 if ((offset % unit) != 0) {
2644 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(offset));
2645 adr = phase->transform(adr);
2646 const TypePtr* atp = TypeRawPtr::BOTTOM;
2647 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2648 mem = phase->transform(mem);
2649 offset += BytesPerInt;
2650 }
2651 assert((offset % unit) == 0, "");
2652
2653 // Initialize the remaining stuff, if any, with a ClearArray.
2654 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2655 }
2656
2657 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2658 Node* start_offset,
2659 Node* end_offset,
2660 PhaseGVN* phase) {
2661 if (start_offset == end_offset) {
2662 // nothing to do
2663 return mem;
2664 }
2665
2666 Compile* C = phase->C;
2667 int unit = BytesPerLong;
2668 Node* zbase = start_offset;
2669 Node* zend = end_offset;
2670
2671 // Scale to the unit required by the CPU:
2672 if (!Matcher::init_array_count_is_in_bytes) {
2673 Node* shift = phase->intcon(exact_log2(unit));
2674 zbase = phase->transform( new(C,3) URShiftXNode(zbase, shift) );
2675 zend = phase->transform( new(C,3) URShiftXNode(zend, shift) );
2676 }
2677
2678 Node* zsize = phase->transform( new(C,3) SubXNode(zend, zbase) );
2679 Node* zinit = phase->zerocon((unit == BytesPerLong) ? T_LONG : T_INT);
2680
2681 // Bulk clear double-words
2682 Node* adr = phase->transform( new(C,4) AddPNode(dest, dest, start_offset) );
2683 mem = new (C, 4) ClearArrayNode(ctl, mem, zsize, adr);
2684 return phase->transform(mem);
2685 }
2686
2687 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2688 intptr_t start_offset,
2689 intptr_t end_offset,
2690 PhaseGVN* phase) {
2691 if (start_offset == end_offset) {
2692 // nothing to do
2693 return mem;
2694 }
2695
2696 Compile* C = phase->C;
2697 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2698 intptr_t done_offset = end_offset;
2699 if ((done_offset % BytesPerLong) != 0) {
2700 done_offset -= BytesPerInt;
2701 }
2702 if (done_offset > start_offset) {
2703 mem = clear_memory(ctl, mem, dest,
2704 start_offset, phase->MakeConX(done_offset), phase);
2705 }
2706 if (done_offset < end_offset) { // emit the final 32-bit store
2707 Node* adr = new (C, 4) AddPNode(dest, dest, phase->MakeConX(done_offset));
2708 adr = phase->transform(adr);
2709 const TypePtr* atp = TypeRawPtr::BOTTOM;
2710 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT);
2711 mem = phase->transform(mem);
2712 done_offset += BytesPerInt;
2713 }
2714 assert(done_offset == end_offset, "");
2715 return mem;
2716 }
2717
2718 //=============================================================================
2719 // Do not match memory edge.
2720 uint StrIntrinsicNode::match_edge(uint idx) const {
2721 return idx == 2 || idx == 3;
2722 }
2723
2724 //------------------------------Ideal------------------------------------------
2725 // Return a node which is more "ideal" than the current node. Strip out
2726 // control copies
2727 Node *StrIntrinsicNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2728 if (remove_dead_region(phase, can_reshape)) return this;
2729 // Don't bother trying to transform a dead node
2730 if (in(0) && in(0)->is_top()) return NULL;
2731
2732 if (can_reshape) {
2733 Node* mem = phase->transform(in(MemNode::Memory));
2734 // If transformed to a MergeMem, get the desired slice
2735 uint alias_idx = phase->C->get_alias_index(adr_type());
2736 mem = mem->is_MergeMem() ? mem->as_MergeMem()->memory_at(alias_idx) : mem;
2737 if (mem != in(MemNode::Memory)) {
2738 set_req(MemNode::Memory, mem);
2739 return this;
2740 }
2741 }
2742 return NULL;
2743 }
2744
2745 //------------------------------Value------------------------------------------
2746 const Type *StrIntrinsicNode::Value( PhaseTransform *phase ) const {
2747 if (in(0) && phase->type(in(0)) == Type::TOP) return Type::TOP;
2748 return bottom_type();
2749 }
2750
2751 //=============================================================================
2752 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2753 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2754 _adr_type(C->get_adr_type(alias_idx))
2755 {
2756 init_class_id(Class_MemBar);
2757 Node* top = C->top();
2758 init_req(TypeFunc::I_O,top);
2759 init_req(TypeFunc::FramePtr,top);
2760 init_req(TypeFunc::ReturnAdr,top);
2761 if (precedent != NULL)
2762 init_req(TypeFunc::Parms, precedent);
2763 }
2764
2765 //------------------------------cmp--------------------------------------------
2766 uint MemBarNode::hash() const { return NO_HASH; }
2767 uint MemBarNode::cmp( const Node &n ) const {
2768 return (&n == this); // Always fail except on self
2769 }
2770
2771 //------------------------------make-------------------------------------------
2772 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2773 int len = Precedent + (pn == NULL? 0: 1);
2774 switch (opcode) {
2775 case Op_MemBarAcquire: return new(C, len) MemBarAcquireNode(C, atp, pn);
2776 case Op_MemBarRelease: return new(C, len) MemBarReleaseNode(C, atp, pn);
2777 case Op_MemBarAcquireLock: return new(C, len) MemBarAcquireLockNode(C, atp, pn);
2778 case Op_MemBarReleaseLock: return new(C, len) MemBarReleaseLockNode(C, atp, pn);
2779 case Op_MemBarVolatile: return new(C, len) MemBarVolatileNode(C, atp, pn);
2780 case Op_MemBarCPUOrder: return new(C, len) MemBarCPUOrderNode(C, atp, pn);
2781 case Op_Initialize: return new(C, len) InitializeNode(C, atp, pn);
2782 case Op_MemBarStoreStore: return new(C, len) MemBarStoreStoreNode(C, atp, pn);
2783 default: ShouldNotReachHere(); return NULL;
2784 }
2785 }
2786
2787 //------------------------------Ideal------------------------------------------
2788 // Return a node which is more "ideal" than the current node. Strip out
2789 // control copies
2790 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2791 if (remove_dead_region(phase, can_reshape)) return this;
2792 // Don't bother trying to transform a dead node
2793 if (in(0) && in(0)->is_top()) return NULL;
2794
2795 // Eliminate volatile MemBars for scalar replaced objects.
2796 if (can_reshape && req() == (Precedent+1) &&
2797 (Opcode() == Op_MemBarAcquire || Opcode() == Op_MemBarVolatile)) {
2798 // Volatile field loads and stores.
2799 Node* my_mem = in(MemBarNode::Precedent);
2800 if (my_mem != NULL && my_mem->is_Mem()) {
2801 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
2802 // Check for scalar replaced object reference.
2803 if( t_oop != NULL && t_oop->is_known_instance_field() &&
2804 t_oop->offset() != Type::OffsetBot &&
2805 t_oop->offset() != Type::OffsetTop) {
2806 // Replace MemBar projections by its inputs.
2807 PhaseIterGVN* igvn = phase->is_IterGVN();
2808 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
2809 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
2810 // Must return either the original node (now dead) or a new node
2811 // (Do not return a top here, since that would break the uniqueness of top.)
2812 return new (phase->C, 1) ConINode(TypeInt::ZERO);
2813 }
2814 }
2815 }
2816 return NULL;
2817 }
2818
2819 //------------------------------Value------------------------------------------
2820 const Type *MemBarNode::Value( PhaseTransform *phase ) const {
2821 if( !in(0) ) return Type::TOP;
2822 if( phase->type(in(0)) == Type::TOP )
2823 return Type::TOP;
2824 return TypeTuple::MEMBAR;
2825 }
2826
2827 //------------------------------match------------------------------------------
2828 // Construct projections for memory.
2829 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
2830 switch (proj->_con) {
2831 case TypeFunc::Control:
2832 case TypeFunc::Memory:
2833 return new (m->C, 1) MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
2834 }
2835 ShouldNotReachHere();
2836 return NULL;
2837 }
2838
2839 //===========================InitializeNode====================================
2840 // SUMMARY:
2841 // This node acts as a memory barrier on raw memory, after some raw stores.
2842 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
2843 // The Initialize can 'capture' suitably constrained stores as raw inits.
2844 // It can coalesce related raw stores into larger units (called 'tiles').
2845 // It can avoid zeroing new storage for memory units which have raw inits.
2846 // At macro-expansion, it is marked 'complete', and does not optimize further.
2847 //
2848 // EXAMPLE:
2849 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
2850 // ctl = incoming control; mem* = incoming memory
2851 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
2852 // First allocate uninitialized memory and fill in the header:
2853 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
2854 // ctl := alloc.Control; mem* := alloc.Memory*
2855 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
2856 // Then initialize to zero the non-header parts of the raw memory block:
2857 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
2858 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
2859 // After the initialize node executes, the object is ready for service:
2860 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
2861 // Suppose its body is immediately initialized as {1,2}:
2862 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
2863 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
2864 // mem.SLICE(#short[*]) := store2
2865 //
2866 // DETAILS:
2867 // An InitializeNode collects and isolates object initialization after
2868 // an AllocateNode and before the next possible safepoint. As a
2869 // memory barrier (MemBarNode), it keeps critical stores from drifting
2870 // down past any safepoint or any publication of the allocation.
2871 // Before this barrier, a newly-allocated object may have uninitialized bits.
2872 // After this barrier, it may be treated as a real oop, and GC is allowed.
2873 //
2874 // The semantics of the InitializeNode include an implicit zeroing of
2875 // the new object from object header to the end of the object.
2876 // (The object header and end are determined by the AllocateNode.)
2877 //
2878 // Certain stores may be added as direct inputs to the InitializeNode.
2879 // These stores must update raw memory, and they must be to addresses
2880 // derived from the raw address produced by AllocateNode, and with
2881 // a constant offset. They must be ordered by increasing offset.
2882 // The first one is at in(RawStores), the last at in(req()-1).
2883 // Unlike most memory operations, they are not linked in a chain,
2884 // but are displayed in parallel as users of the rawmem output of
2885 // the allocation.
2886 //
2887 // (See comments in InitializeNode::capture_store, which continue
2888 // the example given above.)
2889 //
2890 // When the associated Allocate is macro-expanded, the InitializeNode
2891 // may be rewritten to optimize collected stores. A ClearArrayNode
2892 // may also be created at that point to represent any required zeroing.
2893 // The InitializeNode is then marked 'complete', prohibiting further
2894 // capturing of nearby memory operations.
2895 //
2896 // During macro-expansion, all captured initializations which store
2897 // constant values of 32 bits or smaller are coalesced (if advantageous)
2898 // into larger 'tiles' 32 or 64 bits. This allows an object to be
2899 // initialized in fewer memory operations. Memory words which are
2900 // covered by neither tiles nor non-constant stores are pre-zeroed
2901 // by explicit stores of zero. (The code shape happens to do all
2902 // zeroing first, then all other stores, with both sequences occurring
2903 // in order of ascending offsets.)
2904 //
2905 // Alternatively, code may be inserted between an AllocateNode and its
2906 // InitializeNode, to perform arbitrary initialization of the new object.
2907 // E.g., the object copying intrinsics insert complex data transfers here.
2908 // The initialization must then be marked as 'complete' disable the
2909 // built-in zeroing semantics and the collection of initializing stores.
2910 //
2911 // While an InitializeNode is incomplete, reads from the memory state
2912 // produced by it are optimizable if they match the control edge and
2913 // new oop address associated with the allocation/initialization.
2914 // They return a stored value (if the offset matches) or else zero.
2915 // A write to the memory state, if it matches control and address,
2916 // and if it is to a constant offset, may be 'captured' by the
2917 // InitializeNode. It is cloned as a raw memory operation and rewired
2918 // inside the initialization, to the raw oop produced by the allocation.
2919 // Operations on addresses which are provably distinct (e.g., to
2920 // other AllocateNodes) are allowed to bypass the initialization.
2921 //
2922 // The effect of all this is to consolidate object initialization
2923 // (both arrays and non-arrays, both piecewise and bulk) into a
2924 // single location, where it can be optimized as a unit.
2925 //
2926 // Only stores with an offset less than TrackedInitializationLimit words
2927 // will be considered for capture by an InitializeNode. This puts a
2928 // reasonable limit on the complexity of optimized initializations.
2929
2930 //---------------------------InitializeNode------------------------------------
2931 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
2932 : _is_complete(Incomplete), _does_not_escape(false),
2933 MemBarNode(C, adr_type, rawoop)
2934 {
2935 init_class_id(Class_Initialize);
2936
2937 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
2938 assert(in(RawAddress) == rawoop, "proper init");
2939 // Note: allocation() can be NULL, for secondary initialization barriers
2940 }
2941
2942 // Since this node is not matched, it will be processed by the
2943 // register allocator. Declare that there are no constraints
2944 // on the allocation of the RawAddress edge.
2945 const RegMask &InitializeNode::in_RegMask(uint idx) const {
2946 // This edge should be set to top, by the set_complete. But be conservative.
2947 if (idx == InitializeNode::RawAddress)
2948 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
2949 return RegMask::Empty;
2950 }
2951
2952 Node* InitializeNode::memory(uint alias_idx) {
2953 Node* mem = in(Memory);
2954 if (mem->is_MergeMem()) {
2955 return mem->as_MergeMem()->memory_at(alias_idx);
2956 } else {
2957 // incoming raw memory is not split
2958 return mem;
2959 }
2960 }
2961
2962 bool InitializeNode::is_non_zero() {
2963 if (is_complete()) return false;
2964 remove_extra_zeroes();
2965 return (req() > RawStores);
2966 }
2967
2968 void InitializeNode::set_complete(PhaseGVN* phase) {
2969 assert(!is_complete(), "caller responsibility");
2970 _is_complete = Complete;
2971
2972 // After this node is complete, it contains a bunch of
2973 // raw-memory initializations. There is no need for
2974 // it to have anything to do with non-raw memory effects.
2975 // Therefore, tell all non-raw users to re-optimize themselves,
2976 // after skipping the memory effects of this initialization.
2977 PhaseIterGVN* igvn = phase->is_IterGVN();
2978 if (igvn) igvn->add_users_to_worklist(this);
2979 }
2980
2981 // convenience function
2982 // return false if the init contains any stores already
2983 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
2984 InitializeNode* init = initialization();
2985 if (init == NULL || init->is_complete()) return false;
2986 init->remove_extra_zeroes();
2987 // for now, if this allocation has already collected any inits, bail:
2988 if (init->is_non_zero()) return false;
2989 init->set_complete(phase);
2990 return true;
2991 }
2992
2993 void InitializeNode::remove_extra_zeroes() {
2994 if (req() == RawStores) return;
2995 Node* zmem = zero_memory();
2996 uint fill = RawStores;
2997 for (uint i = fill; i < req(); i++) {
2998 Node* n = in(i);
2999 if (n->is_top() || n == zmem) continue; // skip
3000 if (fill < i) set_req(fill, n); // compact
3001 ++fill;
3002 }
3003 // delete any empty spaces created:
3004 while (fill < req()) {
3005 del_req(fill);
3006 }
3007 }
3008
3009 // Helper for remembering which stores go with which offsets.
3010 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3011 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3012 intptr_t offset = -1;
3013 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3014 phase, offset);
3015 if (base == NULL) return -1; // something is dead,
3016 if (offset < 0) return -1; // dead, dead
3017 return offset;
3018 }
3019
3020 // Helper for proving that an initialization expression is
3021 // "simple enough" to be folded into an object initialization.
3022 // Attempts to prove that a store's initial value 'n' can be captured
3023 // within the initialization without creating a vicious cycle, such as:
3024 // { Foo p = new Foo(); p.next = p; }
3025 // True for constants and parameters and small combinations thereof.
3026 bool InitializeNode::detect_init_independence(Node* n,
3027 bool st_is_pinned,
3028 int& count) {
3029 if (n == NULL) return true; // (can this really happen?)
3030 if (n->is_Proj()) n = n->in(0);
3031 if (n == this) return false; // found a cycle
3032 if (n->is_Con()) return true;
3033 if (n->is_Start()) return true; // params, etc., are OK
3034 if (n->is_Root()) return true; // even better
3035
3036 Node* ctl = n->in(0);
3037 if (ctl != NULL && !ctl->is_top()) {
3038 if (ctl->is_Proj()) ctl = ctl->in(0);
3039 if (ctl == this) return false;
3040
3041 // If we already know that the enclosing memory op is pinned right after
3042 // the init, then any control flow that the store has picked up
3043 // must have preceded the init, or else be equal to the init.
3044 // Even after loop optimizations (which might change control edges)
3045 // a store is never pinned *before* the availability of its inputs.
3046 if (!MemNode::all_controls_dominate(n, this))
3047 return false; // failed to prove a good control
3048
3049 }
3050
3051 // Check data edges for possible dependencies on 'this'.
3052 if ((count += 1) > 20) return false; // complexity limit
3053 for (uint i = 1; i < n->req(); i++) {
3054 Node* m = n->in(i);
3055 if (m == NULL || m == n || m->is_top()) continue;
3056 uint first_i = n->find_edge(m);
3057 if (i != first_i) continue; // process duplicate edge just once
3058 if (!detect_init_independence(m, st_is_pinned, count)) {
3059 return false;
3060 }
3061 }
3062
3063 return true;
3064 }
3065
3066 // Here are all the checks a Store must pass before it can be moved into
3067 // an initialization. Returns zero if a check fails.
3068 // On success, returns the (constant) offset to which the store applies,
3069 // within the initialized memory.
3070 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase) {
3071 const int FAIL = 0;
3072 if (st->req() != MemNode::ValueIn + 1)
3073 return FAIL; // an inscrutable StoreNode (card mark?)
3074 Node* ctl = st->in(MemNode::Control);
3075 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3076 return FAIL; // must be unconditional after the initialization
3077 Node* mem = st->in(MemNode::Memory);
3078 if (!(mem->is_Proj() && mem->in(0) == this))
3079 return FAIL; // must not be preceded by other stores
3080 Node* adr = st->in(MemNode::Address);
3081 intptr_t offset;
3082 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3083 if (alloc == NULL)
3084 return FAIL; // inscrutable address
3085 if (alloc != allocation())
3086 return FAIL; // wrong allocation! (store needs to float up)
3087 Node* val = st->in(MemNode::ValueIn);
3088 int complexity_count = 0;
3089 if (!detect_init_independence(val, true, complexity_count))
3090 return FAIL; // stored value must be 'simple enough'
3091
3092 return offset; // success
3093 }
3094
3095 // Find the captured store in(i) which corresponds to the range
3096 // [start..start+size) in the initialized object.
3097 // If there is one, return its index i. If there isn't, return the
3098 // negative of the index where it should be inserted.
3099 // Return 0 if the queried range overlaps an initialization boundary
3100 // or if dead code is encountered.
3101 // If size_in_bytes is zero, do not bother with overlap checks.
3102 int InitializeNode::captured_store_insertion_point(intptr_t start,
3103 int size_in_bytes,
3104 PhaseTransform* phase) {
3105 const int FAIL = 0, MAX_STORE = BytesPerLong;
3106
3107 if (is_complete())
3108 return FAIL; // arraycopy got here first; punt
3109
3110 assert(allocation() != NULL, "must be present");
3111
3112 // no negatives, no header fields:
3113 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3114
3115 // after a certain size, we bail out on tracking all the stores:
3116 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3117 if (start >= ti_limit) return FAIL;
3118
3119 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3120 if (i >= limit) return -(int)i; // not found; here is where to put it
3121
3122 Node* st = in(i);
3123 intptr_t st_off = get_store_offset(st, phase);
3124 if (st_off < 0) {
3125 if (st != zero_memory()) {
3126 return FAIL; // bail out if there is dead garbage
3127 }
3128 } else if (st_off > start) {
3129 // ...we are done, since stores are ordered
3130 if (st_off < start + size_in_bytes) {
3131 return FAIL; // the next store overlaps
3132 }
3133 return -(int)i; // not found; here is where to put it
3134 } else if (st_off < start) {
3135 if (size_in_bytes != 0 &&
3136 start < st_off + MAX_STORE &&
3137 start < st_off + st->as_Store()->memory_size()) {
3138 return FAIL; // the previous store overlaps
3139 }
3140 } else {
3141 if (size_in_bytes != 0 &&
3142 st->as_Store()->memory_size() != size_in_bytes) {
3143 return FAIL; // mismatched store size
3144 }
3145 return i;
3146 }
3147
3148 ++i;
3149 }
3150 }
3151
3152 // Look for a captured store which initializes at the offset 'start'
3153 // with the given size. If there is no such store, and no other
3154 // initialization interferes, then return zero_memory (the memory
3155 // projection of the AllocateNode).
3156 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3157 PhaseTransform* phase) {
3158 assert(stores_are_sane(phase), "");
3159 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3160 if (i == 0) {
3161 return NULL; // something is dead
3162 } else if (i < 0) {
3163 return zero_memory(); // just primordial zero bits here
3164 } else {
3165 Node* st = in(i); // here is the store at this position
3166 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3167 return st;
3168 }
3169 }
3170
3171 // Create, as a raw pointer, an address within my new object at 'offset'.
3172 Node* InitializeNode::make_raw_address(intptr_t offset,
3173 PhaseTransform* phase) {
3174 Node* addr = in(RawAddress);
3175 if (offset != 0) {
3176 Compile* C = phase->C;
3177 addr = phase->transform( new (C, 4) AddPNode(C->top(), addr,
3178 phase->MakeConX(offset)) );
3179 }
3180 return addr;
3181 }
3182
3183 // Clone the given store, converting it into a raw store
3184 // initializing a field or element of my new object.
3185 // Caller is responsible for retiring the original store,
3186 // with subsume_node or the like.
3187 //
3188 // From the example above InitializeNode::InitializeNode,
3189 // here are the old stores to be captured:
3190 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3191 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3192 //
3193 // Here is the changed code; note the extra edges on init:
3194 // alloc = (Allocate ...)
3195 // rawoop = alloc.RawAddress
3196 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3197 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3198 // init = (Initialize alloc.Control alloc.Memory rawoop
3199 // rawstore1 rawstore2)
3200 //
3201 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3202 PhaseTransform* phase) {
3203 assert(stores_are_sane(phase), "");
3204
3205 if (start < 0) return NULL;
3206 assert(can_capture_store(st, phase) == start, "sanity");
3207
3208 Compile* C = phase->C;
3209 int size_in_bytes = st->memory_size();
3210 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3211 if (i == 0) return NULL; // bail out
3212 Node* prev_mem = NULL; // raw memory for the captured store
3213 if (i > 0) {
3214 prev_mem = in(i); // there is a pre-existing store under this one
3215 set_req(i, C->top()); // temporarily disconnect it
3216 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3217 } else {
3218 i = -i; // no pre-existing store
3219 prev_mem = zero_memory(); // a slice of the newly allocated object
3220 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3221 set_req(--i, C->top()); // reuse this edge; it has been folded away
3222 else
3223 ins_req(i, C->top()); // build a new edge
3224 }
3225 Node* new_st = st->clone();
3226 new_st->set_req(MemNode::Control, in(Control));
3227 new_st->set_req(MemNode::Memory, prev_mem);
3228 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3229 new_st = phase->transform(new_st);
3230
3231 // At this point, new_st might have swallowed a pre-existing store
3232 // at the same offset, or perhaps new_st might have disappeared,
3233 // if it redundantly stored the same value (or zero to fresh memory).
3234
3235 // In any case, wire it in:
3236 set_req(i, new_st);
3237
3238 // The caller may now kill the old guy.
3239 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3240 assert(check_st == new_st || check_st == NULL, "must be findable");
3241 assert(!is_complete(), "");
3242 return new_st;
3243 }
3244
3245 static bool store_constant(jlong* tiles, int num_tiles,
3246 intptr_t st_off, int st_size,
3247 jlong con) {
3248 if ((st_off & (st_size-1)) != 0)
3249 return false; // strange store offset (assume size==2**N)
3250 address addr = (address)tiles + st_off;
3251 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3252 switch (st_size) {
3253 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3254 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3255 case sizeof(jint): *(jint*) addr = (jint) con; break;
3256 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3257 default: return false; // strange store size (detect size!=2**N here)
3258 }
3259 return true; // return success to caller
3260 }
3261
3262 // Coalesce subword constants into int constants and possibly
3263 // into long constants. The goal, if the CPU permits,
3264 // is to initialize the object with a small number of 64-bit tiles.
3265 // Also, convert floating-point constants to bit patterns.
3266 // Non-constants are not relevant to this pass.
3267 //
3268 // In terms of the running example on InitializeNode::InitializeNode
3269 // and InitializeNode::capture_store, here is the transformation
3270 // of rawstore1 and rawstore2 into rawstore12:
3271 // alloc = (Allocate ...)
3272 // rawoop = alloc.RawAddress
3273 // tile12 = 0x00010002
3274 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3275 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3276 //
3277 void
3278 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3279 Node* size_in_bytes,
3280 PhaseGVN* phase) {
3281 Compile* C = phase->C;
3282
3283 assert(stores_are_sane(phase), "");
3284 // Note: After this pass, they are not completely sane,
3285 // since there may be some overlaps.
3286
3287 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3288
3289 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3290 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3291 size_limit = MIN2(size_limit, ti_limit);
3292 size_limit = align_size_up(size_limit, BytesPerLong);
3293 int num_tiles = size_limit / BytesPerLong;
3294
3295 // allocate space for the tile map:
3296 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3297 jlong tiles_buf[small_len];
3298 Node* nodes_buf[small_len];
3299 jlong inits_buf[small_len];
3300 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3301 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3302 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3303 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3304 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3305 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3306 // tiles: exact bitwise model of all primitive constants
3307 // nodes: last constant-storing node subsumed into the tiles model
3308 // inits: which bytes (in each tile) are touched by any initializations
3309
3310 //// Pass A: Fill in the tile model with any relevant stores.
3311
3312 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3313 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3314 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3315 Node* zmem = zero_memory(); // initially zero memory state
3316 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3317 Node* st = in(i);
3318 intptr_t st_off = get_store_offset(st, phase);
3319
3320 // Figure out the store's offset and constant value:
3321 if (st_off < header_size) continue; //skip (ignore header)
3322 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3323 int st_size = st->as_Store()->memory_size();
3324 if (st_off + st_size > size_limit) break;
3325
3326 // Record which bytes are touched, whether by constant or not.
3327 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3328 continue; // skip (strange store size)
3329
3330 const Type* val = phase->type(st->in(MemNode::ValueIn));
3331 if (!val->singleton()) continue; //skip (non-con store)
3332 BasicType type = val->basic_type();
3333
3334 jlong con = 0;
3335 switch (type) {
3336 case T_INT: con = val->is_int()->get_con(); break;
3337 case T_LONG: con = val->is_long()->get_con(); break;
3338 case T_FLOAT: con = jint_cast(val->getf()); break;
3339 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3340 default: continue; //skip (odd store type)
3341 }
3342
3343 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3344 st->Opcode() == Op_StoreL) {
3345 continue; // This StoreL is already optimal.
3346 }
3347
3348 // Store down the constant.
3349 store_constant(tiles, num_tiles, st_off, st_size, con);
3350
3351 intptr_t j = st_off >> LogBytesPerLong;
3352
3353 if (type == T_INT && st_size == BytesPerInt
3354 && (st_off & BytesPerInt) == BytesPerInt) {
3355 jlong lcon = tiles[j];
3356 if (!Matcher::isSimpleConstant64(lcon) &&
3357 st->Opcode() == Op_StoreI) {
3358 // This StoreI is already optimal by itself.
3359 jint* intcon = (jint*) &tiles[j];
3360 intcon[1] = 0; // undo the store_constant()
3361
3362 // If the previous store is also optimal by itself, back up and
3363 // undo the action of the previous loop iteration... if we can.
3364 // But if we can't, just let the previous half take care of itself.
3365 st = nodes[j];
3366 st_off -= BytesPerInt;
3367 con = intcon[0];
3368 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3369 assert(st_off >= header_size, "still ignoring header");
3370 assert(get_store_offset(st, phase) == st_off, "must be");
3371 assert(in(i-1) == zmem, "must be");
3372 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3373 assert(con == tcon->is_int()->get_con(), "must be");
3374 // Undo the effects of the previous loop trip, which swallowed st:
3375 intcon[0] = 0; // undo store_constant()
3376 set_req(i-1, st); // undo set_req(i, zmem)
3377 nodes[j] = NULL; // undo nodes[j] = st
3378 --old_subword; // undo ++old_subword
3379 }
3380 continue; // This StoreI is already optimal.
3381 }
3382 }
3383
3384 // This store is not needed.
3385 set_req(i, zmem);
3386 nodes[j] = st; // record for the moment
3387 if (st_size < BytesPerLong) // something has changed
3388 ++old_subword; // includes int/float, but who's counting...
3389 else ++old_long;
3390 }
3391
3392 if ((old_subword + old_long) == 0)
3393 return; // nothing more to do
3394
3395 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3396 // Be sure to insert them before overlapping non-constant stores.
3397 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3398 for (int j = 0; j < num_tiles; j++) {
3399 jlong con = tiles[j];
3400 jlong init = inits[j];
3401 if (con == 0) continue;
3402 jint con0, con1; // split the constant, address-wise
3403 jint init0, init1; // split the init map, address-wise
3404 { union { jlong con; jint intcon[2]; } u;
3405 u.con = con;
3406 con0 = u.intcon[0];
3407 con1 = u.intcon[1];
3408 u.con = init;
3409 init0 = u.intcon[0];
3410 init1 = u.intcon[1];
3411 }
3412
3413 Node* old = nodes[j];
3414 assert(old != NULL, "need the prior store");
3415 intptr_t offset = (j * BytesPerLong);
3416
3417 bool split = !Matcher::isSimpleConstant64(con);
3418
3419 if (offset < header_size) {
3420 assert(offset + BytesPerInt >= header_size, "second int counts");
3421 assert(*(jint*)&tiles[j] == 0, "junk in header");
3422 split = true; // only the second word counts
3423 // Example: int a[] = { 42 ... }
3424 } else if (con0 == 0 && init0 == -1) {
3425 split = true; // first word is covered by full inits
3426 // Example: int a[] = { ... foo(), 42 ... }
3427 } else if (con1 == 0 && init1 == -1) {
3428 split = true; // second word is covered by full inits
3429 // Example: int a[] = { ... 42, foo() ... }
3430 }
3431
3432 // Here's a case where init0 is neither 0 nor -1:
3433 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3434 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3435 // In this case the tile is not split; it is (jlong)42.
3436 // The big tile is stored down, and then the foo() value is inserted.
3437 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3438
3439 Node* ctl = old->in(MemNode::Control);
3440 Node* adr = make_raw_address(offset, phase);
3441 const TypePtr* atp = TypeRawPtr::BOTTOM;
3442
3443 // One or two coalesced stores to plop down.
3444 Node* st[2];
3445 intptr_t off[2];
3446 int nst = 0;
3447 if (!split) {
3448 ++new_long;
3449 off[nst] = offset;
3450 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3451 phase->longcon(con), T_LONG);
3452 } else {
3453 // Omit either if it is a zero.
3454 if (con0 != 0) {
3455 ++new_int;
3456 off[nst] = offset;
3457 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3458 phase->intcon(con0), T_INT);
3459 }
3460 if (con1 != 0) {
3461 ++new_int;
3462 offset += BytesPerInt;
3463 adr = make_raw_address(offset, phase);
3464 off[nst] = offset;
3465 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3466 phase->intcon(con1), T_INT);
3467 }
3468 }
3469
3470 // Insert second store first, then the first before the second.
3471 // Insert each one just before any overlapping non-constant stores.
3472 while (nst > 0) {
3473 Node* st1 = st[--nst];
3474 C->copy_node_notes_to(st1, old);
3475 st1 = phase->transform(st1);
3476 offset = off[nst];
3477 assert(offset >= header_size, "do not smash header");
3478 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3479 guarantee(ins_idx != 0, "must re-insert constant store");
3480 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3481 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3482 set_req(--ins_idx, st1);
3483 else
3484 ins_req(ins_idx, st1);
3485 }
3486 }
3487
3488 if (PrintCompilation && WizardMode)
3489 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3490 old_subword, old_long, new_int, new_long);
3491 if (C->log() != NULL)
3492 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3493 old_subword, old_long, new_int, new_long);
3494
3495 // Clean up any remaining occurrences of zmem:
3496 remove_extra_zeroes();
3497 }
3498
3499 // Explore forward from in(start) to find the first fully initialized
3500 // word, and return its offset. Skip groups of subword stores which
3501 // together initialize full words. If in(start) is itself part of a
3502 // fully initialized word, return the offset of in(start). If there
3503 // are no following full-word stores, or if something is fishy, return
3504 // a negative value.
3505 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3506 int int_map = 0;
3507 intptr_t int_map_off = 0;
3508 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3509
3510 for (uint i = start, limit = req(); i < limit; i++) {
3511 Node* st = in(i);
3512
3513 intptr_t st_off = get_store_offset(st, phase);
3514 if (st_off < 0) break; // return conservative answer
3515
3516 int st_size = st->as_Store()->memory_size();
3517 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3518 return st_off; // we found a complete word init
3519 }
3520
3521 // update the map:
3522
3523 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3524 if (this_int_off != int_map_off) {
3525 // reset the map:
3526 int_map = 0;
3527 int_map_off = this_int_off;
3528 }
3529
3530 int subword_off = st_off - this_int_off;
3531 int_map |= right_n_bits(st_size) << subword_off;
3532 if ((int_map & FULL_MAP) == FULL_MAP) {
3533 return this_int_off; // we found a complete word init
3534 }
3535
3536 // Did this store hit or cross the word boundary?
3537 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3538 if (next_int_off == this_int_off + BytesPerInt) {
3539 // We passed the current int, without fully initializing it.
3540 int_map_off = next_int_off;
3541 int_map >>= BytesPerInt;
3542 } else if (next_int_off > this_int_off + BytesPerInt) {
3543 // We passed the current and next int.
3544 return this_int_off + BytesPerInt;
3545 }
3546 }
3547
3548 return -1;
3549 }
3550
3551
3552 // Called when the associated AllocateNode is expanded into CFG.
3553 // At this point, we may perform additional optimizations.
3554 // Linearize the stores by ascending offset, to make memory
3555 // activity as coherent as possible.
3556 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3557 intptr_t header_size,
3558 Node* size_in_bytes,
3559 PhaseGVN* phase) {
3560 assert(!is_complete(), "not already complete");
3561 assert(stores_are_sane(phase), "");
3562 assert(allocation() != NULL, "must be present");
3563
3564 remove_extra_zeroes();
3565
3566 if (ReduceFieldZeroing || ReduceBulkZeroing)
3567 // reduce instruction count for common initialization patterns
3568 coalesce_subword_stores(header_size, size_in_bytes, phase);
3569
3570 Node* zmem = zero_memory(); // initially zero memory state
3571 Node* inits = zmem; // accumulating a linearized chain of inits
3572 #ifdef ASSERT
3573 intptr_t first_offset = allocation()->minimum_header_size();
3574 intptr_t last_init_off = first_offset; // previous init offset
3575 intptr_t last_init_end = first_offset; // previous init offset+size
3576 intptr_t last_tile_end = first_offset; // previous tile offset+size
3577 #endif
3578 intptr_t zeroes_done = header_size;
3579
3580 bool do_zeroing = true; // we might give up if inits are very sparse
3581 int big_init_gaps = 0; // how many large gaps have we seen?
3582
3583 if (ZeroTLAB) do_zeroing = false;
3584 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3585
3586 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3587 Node* st = in(i);
3588 intptr_t st_off = get_store_offset(st, phase);
3589 if (st_off < 0)
3590 break; // unknown junk in the inits
3591 if (st->in(MemNode::Memory) != zmem)
3592 break; // complicated store chains somehow in list
3593
3594 int st_size = st->as_Store()->memory_size();
3595 intptr_t next_init_off = st_off + st_size;
3596
3597 if (do_zeroing && zeroes_done < next_init_off) {
3598 // See if this store needs a zero before it or under it.
3599 intptr_t zeroes_needed = st_off;
3600
3601 if (st_size < BytesPerInt) {
3602 // Look for subword stores which only partially initialize words.
3603 // If we find some, we must lay down some word-level zeroes first,
3604 // underneath the subword stores.
3605 //
3606 // Examples:
3607 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3608 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3609 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3610 //
3611 // Note: coalesce_subword_stores may have already done this,
3612 // if it was prompted by constant non-zero subword initializers.
3613 // But this case can still arise with non-constant stores.
3614
3615 intptr_t next_full_store = find_next_fullword_store(i, phase);
3616
3617 // In the examples above:
3618 // in(i) p q r s x y z
3619 // st_off 12 13 14 15 12 13 14
3620 // st_size 1 1 1 1 1 1 1
3621 // next_full_s. 12 16 16 16 16 16 16
3622 // z's_done 12 16 16 16 12 16 12
3623 // z's_needed 12 16 16 16 16 16 16
3624 // zsize 0 0 0 0 4 0 4
3625 if (next_full_store < 0) {
3626 // Conservative tack: Zero to end of current word.
3627 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3628 } else {
3629 // Zero to beginning of next fully initialized word.
3630 // Or, don't zero at all, if we are already in that word.
3631 assert(next_full_store >= zeroes_needed, "must go forward");
3632 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3633 zeroes_needed = next_full_store;
3634 }
3635 }
3636
3637 if (zeroes_needed > zeroes_done) {
3638 intptr_t zsize = zeroes_needed - zeroes_done;
3639 // Do some incremental zeroing on rawmem, in parallel with inits.
3640 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3641 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3642 zeroes_done, zeroes_needed,
3643 phase);
3644 zeroes_done = zeroes_needed;
3645 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3646 do_zeroing = false; // leave the hole, next time
3647 }
3648 }
3649
3650 // Collect the store and move on:
3651 st->set_req(MemNode::Memory, inits);
3652 inits = st; // put it on the linearized chain
3653 set_req(i, zmem); // unhook from previous position
3654
3655 if (zeroes_done == st_off)
3656 zeroes_done = next_init_off;
3657
3658 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3659
3660 #ifdef ASSERT
3661 // Various order invariants. Weaker than stores_are_sane because
3662 // a large constant tile can be filled in by smaller non-constant stores.
3663 assert(st_off >= last_init_off, "inits do not reverse");
3664 last_init_off = st_off;
3665 const Type* val = NULL;
3666 if (st_size >= BytesPerInt &&
3667 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3668 (int)val->basic_type() < (int)T_OBJECT) {
3669 assert(st_off >= last_tile_end, "tiles do not overlap");
3670 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3671 last_tile_end = MAX2(last_tile_end, next_init_off);
3672 } else {
3673 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3674 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3675 assert(st_off >= last_init_end, "inits do not overlap");
3676 last_init_end = next_init_off; // it's a non-tile
3677 }
3678 #endif //ASSERT
3679 }
3680
3681 remove_extra_zeroes(); // clear out all the zmems left over
3682 add_req(inits);
3683
3684 if (!ZeroTLAB) {
3685 // If anything remains to be zeroed, zero it all now.
3686 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3687 // if it is the last unused 4 bytes of an instance, forget about it
3688 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3689 if (zeroes_done + BytesPerLong >= size_limit) {
3690 assert(allocation() != NULL, "");
3691 if (allocation()->Opcode() == Op_Allocate) {
3692 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3693 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3694 if (zeroes_done == k->layout_helper())
3695 zeroes_done = size_limit;
3696 }
3697 }
3698 if (zeroes_done < size_limit) {
3699 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3700 zeroes_done, size_in_bytes, phase);
3701 }
3702 }
3703
3704 set_complete(phase);
3705 return rawmem;
3706 }
3707
3708
3709 #ifdef ASSERT
3710 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3711 if (is_complete())
3712 return true; // stores could be anything at this point
3713 assert(allocation() != NULL, "must be present");
3714 intptr_t last_off = allocation()->minimum_header_size();
3715 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3716 Node* st = in(i);
3717 intptr_t st_off = get_store_offset(st, phase);
3718 if (st_off < 0) continue; // ignore dead garbage
3719 if (last_off > st_off) {
3720 tty->print_cr("*** bad store offset at %d: %d > %d", i, last_off, st_off);
3721 this->dump(2);
3722 assert(false, "ascending store offsets");
3723 return false;
3724 }
3725 last_off = st_off + st->as_Store()->memory_size();
3726 }
3727 return true;
3728 }
3729 #endif //ASSERT
3730
3731
3732
3733
3734 //============================MergeMemNode=====================================
3735 //
3736 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3737 // contributing store or call operations. Each contributor provides the memory
3738 // state for a particular "alias type" (see Compile::alias_type). For example,
3739 // if a MergeMem has an input X for alias category #6, then any memory reference
3740 // to alias category #6 may use X as its memory state input, as an exact equivalent
3741 // to using the MergeMem as a whole.
3742 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
3743 //
3744 // (Here, the <N> notation gives the index of the relevant adr_type.)
3745 //
3746 // In one special case (and more cases in the future), alias categories overlap.
3747 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
3748 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
3749 // it is exactly equivalent to that state W:
3750 // MergeMem(<Bot>: W) <==> W
3751 //
3752 // Usually, the merge has more than one input. In that case, where inputs
3753 // overlap (i.e., one is Bot), the narrower alias type determines the memory
3754 // state for that type, and the wider alias type (Bot) fills in everywhere else:
3755 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
3756 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
3757 //
3758 // A merge can take a "wide" memory state as one of its narrow inputs.
3759 // This simply means that the merge observes out only the relevant parts of
3760 // the wide input. That is, wide memory states arriving at narrow merge inputs
3761 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
3762 //
3763 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
3764 // and that memory slices "leak through":
3765 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
3766 //
3767 // But, in such a cascade, repeated memory slices can "block the leak":
3768 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
3769 //
3770 // In the last example, Y is not part of the combined memory state of the
3771 // outermost MergeMem. The system must, of course, prevent unschedulable
3772 // memory states from arising, so you can be sure that the state Y is somehow
3773 // a precursor to state Y'.
3774 //
3775 //
3776 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
3777 // of each MergeMemNode array are exactly the numerical alias indexes, including
3778 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
3779 // Compile::alias_type (and kin) produce and manage these indexes.
3780 //
3781 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
3782 // (Note that this provides quick access to the top node inside MergeMem methods,
3783 // without the need to reach out via TLS to Compile::current.)
3784 //
3785 // As a consequence of what was just described, a MergeMem that represents a full
3786 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
3787 // containing all alias categories.
3788 //
3789 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
3790 //
3791 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
3792 // a memory state for the alias type <N>, or else the top node, meaning that
3793 // there is no particular input for that alias type. Note that the length of
3794 // a MergeMem is variable, and may be extended at any time to accommodate new
3795 // memory states at larger alias indexes. When merges grow, they are of course
3796 // filled with "top" in the unused in() positions.
3797 //
3798 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
3799 // (Top was chosen because it works smoothly with passes like GCM.)
3800 //
3801 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
3802 // the type of random VM bits like TLS references.) Since it is always the
3803 // first non-Bot memory slice, some low-level loops use it to initialize an
3804 // index variable: for (i = AliasIdxRaw; i < req(); i++).
3805 //
3806 //
3807 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
3808 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
3809 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
3810 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
3811 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
3812 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
3813 //
3814 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
3815 // really that different from the other memory inputs. An abbreviation called
3816 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
3817 //
3818 //
3819 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
3820 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
3821 // that "emerges though" the base memory will be marked as excluding the alias types
3822 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
3823 //
3824 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
3825 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
3826 //
3827 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
3828 // (It is currently unimplemented.) As you can see, the resulting merge is
3829 // actually a disjoint union of memory states, rather than an overlay.
3830 //
3831
3832 //------------------------------MergeMemNode-----------------------------------
3833 Node* MergeMemNode::make_empty_memory() {
3834 Node* empty_memory = (Node*) Compile::current()->top();
3835 assert(empty_memory->is_top(), "correct sentinel identity");
3836 return empty_memory;
3837 }
3838
3839 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
3840 init_class_id(Class_MergeMem);
3841 // all inputs are nullified in Node::Node(int)
3842 // set_input(0, NULL); // no control input
3843
3844 // Initialize the edges uniformly to top, for starters.
3845 Node* empty_mem = make_empty_memory();
3846 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
3847 init_req(i,empty_mem);
3848 }
3849 assert(empty_memory() == empty_mem, "");
3850
3851 if( new_base != NULL && new_base->is_MergeMem() ) {
3852 MergeMemNode* mdef = new_base->as_MergeMem();
3853 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
3854 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
3855 mms.set_memory(mms.memory2());
3856 }
3857 assert(base_memory() == mdef->base_memory(), "");
3858 } else {
3859 set_base_memory(new_base);
3860 }
3861 }
3862
3863 // Make a new, untransformed MergeMem with the same base as 'mem'.
3864 // If mem is itself a MergeMem, populate the result with the same edges.
3865 MergeMemNode* MergeMemNode::make(Compile* C, Node* mem) {
3866 return new(C, 1+Compile::AliasIdxRaw) MergeMemNode(mem);
3867 }
3868
3869 //------------------------------cmp--------------------------------------------
3870 uint MergeMemNode::hash() const { return NO_HASH; }
3871 uint MergeMemNode::cmp( const Node &n ) const {
3872 return (&n == this); // Always fail except on self
3873 }
3874
3875 //------------------------------Identity---------------------------------------
3876 Node* MergeMemNode::Identity(PhaseTransform *phase) {
3877 // Identity if this merge point does not record any interesting memory
3878 // disambiguations.
3879 Node* base_mem = base_memory();
3880 Node* empty_mem = empty_memory();
3881 if (base_mem != empty_mem) { // Memory path is not dead?
3882 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3883 Node* mem = in(i);
3884 if (mem != empty_mem && mem != base_mem) {
3885 return this; // Many memory splits; no change
3886 }
3887 }
3888 }
3889 return base_mem; // No memory splits; ID on the one true input
3890 }
3891
3892 //------------------------------Ideal------------------------------------------
3893 // This method is invoked recursively on chains of MergeMem nodes
3894 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3895 // Remove chain'd MergeMems
3896 //
3897 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
3898 // relative to the "in(Bot)". Since we are patching both at the same time,
3899 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
3900 // but rewrite each "in(i)" relative to the new "in(Bot)".
3901 Node *progress = NULL;
3902
3903
3904 Node* old_base = base_memory();
3905 Node* empty_mem = empty_memory();
3906 if (old_base == empty_mem)
3907 return NULL; // Dead memory path.
3908
3909 MergeMemNode* old_mbase;
3910 if (old_base != NULL && old_base->is_MergeMem())
3911 old_mbase = old_base->as_MergeMem();
3912 else
3913 old_mbase = NULL;
3914 Node* new_base = old_base;
3915
3916 // simplify stacked MergeMems in base memory
3917 if (old_mbase) new_base = old_mbase->base_memory();
3918
3919 // the base memory might contribute new slices beyond my req()
3920 if (old_mbase) grow_to_match(old_mbase);
3921
3922 // Look carefully at the base node if it is a phi.
3923 PhiNode* phi_base;
3924 if (new_base != NULL && new_base->is_Phi())
3925 phi_base = new_base->as_Phi();
3926 else
3927 phi_base = NULL;
3928
3929 Node* phi_reg = NULL;
3930 uint phi_len = (uint)-1;
3931 if (phi_base != NULL && !phi_base->is_copy()) {
3932 // do not examine phi if degraded to a copy
3933 phi_reg = phi_base->region();
3934 phi_len = phi_base->req();
3935 // see if the phi is unfinished
3936 for (uint i = 1; i < phi_len; i++) {
3937 if (phi_base->in(i) == NULL) {
3938 // incomplete phi; do not look at it yet!
3939 phi_reg = NULL;
3940 phi_len = (uint)-1;
3941 break;
3942 }
3943 }
3944 }
3945
3946 // Note: We do not call verify_sparse on entry, because inputs
3947 // can normalize to the base_memory via subsume_node or similar
3948 // mechanisms. This method repairs that damage.
3949
3950 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
3951
3952 // Look at each slice.
3953 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
3954 Node* old_in = in(i);
3955 // calculate the old memory value
3956 Node* old_mem = old_in;
3957 if (old_mem == empty_mem) old_mem = old_base;
3958 assert(old_mem == memory_at(i), "");
3959
3960 // maybe update (reslice) the old memory value
3961
3962 // simplify stacked MergeMems
3963 Node* new_mem = old_mem;
3964 MergeMemNode* old_mmem;
3965 if (old_mem != NULL && old_mem->is_MergeMem())
3966 old_mmem = old_mem->as_MergeMem();
3967 else
3968 old_mmem = NULL;
3969 if (old_mmem == this) {
3970 // This can happen if loops break up and safepoints disappear.
3971 // A merge of BotPtr (default) with a RawPtr memory derived from a
3972 // safepoint can be rewritten to a merge of the same BotPtr with
3973 // the BotPtr phi coming into the loop. If that phi disappears
3974 // also, we can end up with a self-loop of the mergemem.
3975 // In general, if loops degenerate and memory effects disappear,
3976 // a mergemem can be left looking at itself. This simply means
3977 // that the mergemem's default should be used, since there is
3978 // no longer any apparent effect on this slice.
3979 // Note: If a memory slice is a MergeMem cycle, it is unreachable
3980 // from start. Update the input to TOP.
3981 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
3982 }
3983 else if (old_mmem != NULL) {
3984 new_mem = old_mmem->memory_at(i);
3985 }
3986 // else preceding memory was not a MergeMem
3987
3988 // replace equivalent phis (unfortunately, they do not GVN together)
3989 if (new_mem != NULL && new_mem != new_base &&
3990 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
3991 if (new_mem->is_Phi()) {
3992 PhiNode* phi_mem = new_mem->as_Phi();
3993 for (uint i = 1; i < phi_len; i++) {
3994 if (phi_base->in(i) != phi_mem->in(i)) {
3995 phi_mem = NULL;
3996 break;
3997 }
3998 }
3999 if (phi_mem != NULL) {
4000 // equivalent phi nodes; revert to the def
4001 new_mem = new_base;
4002 }
4003 }
4004 }
4005
4006 // maybe store down a new value
4007 Node* new_in = new_mem;
4008 if (new_in == new_base) new_in = empty_mem;
4009
4010 if (new_in != old_in) {
4011 // Warning: Do not combine this "if" with the previous "if"
4012 // A memory slice might have be be rewritten even if it is semantically
4013 // unchanged, if the base_memory value has changed.
4014 set_req(i, new_in);
4015 progress = this; // Report progress
4016 }
4017 }
4018
4019 if (new_base != old_base) {
4020 set_req(Compile::AliasIdxBot, new_base);
4021 // Don't use set_base_memory(new_base), because we need to update du.
4022 assert(base_memory() == new_base, "");
4023 progress = this;
4024 }
4025
4026 if( base_memory() == this ) {
4027 // a self cycle indicates this memory path is dead
4028 set_req(Compile::AliasIdxBot, empty_mem);
4029 }
4030
4031 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4032 // Recursion must occur after the self cycle check above
4033 if( base_memory()->is_MergeMem() ) {
4034 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4035 Node *m = phase->transform(new_mbase); // Rollup any cycles
4036 if( m != NULL && (m->is_top() ||
4037 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
4038 // propagate rollup of dead cycle to self
4039 set_req(Compile::AliasIdxBot, empty_mem);
4040 }
4041 }
4042
4043 if( base_memory() == empty_mem ) {
4044 progress = this;
4045 // Cut inputs during Parse phase only.
4046 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4047 if( !can_reshape ) {
4048 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4049 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4050 }
4051 }
4052 }
4053
4054 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4055 // Check if PhiNode::Ideal's "Split phis through memory merges"
4056 // transform should be attempted. Look for this->phi->this cycle.
4057 uint merge_width = req();
4058 if (merge_width > Compile::AliasIdxRaw) {
4059 PhiNode* phi = base_memory()->as_Phi();
4060 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4061 if (phi->in(i) == this) {
4062 phase->is_IterGVN()->_worklist.push(phi);
4063 break;
4064 }
4065 }
4066 }
4067 }
4068
4069 assert(progress || verify_sparse(), "please, no dups of base");
4070 return progress;
4071 }
4072
4073 //-------------------------set_base_memory-------------------------------------
4074 void MergeMemNode::set_base_memory(Node *new_base) {
4075 Node* empty_mem = empty_memory();
4076 set_req(Compile::AliasIdxBot, new_base);
4077 assert(memory_at(req()) == new_base, "must set default memory");
4078 // Clear out other occurrences of new_base:
4079 if (new_base != empty_mem) {
4080 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4081 if (in(i) == new_base) set_req(i, empty_mem);
4082 }
4083 }
4084 }
4085
4086 //------------------------------out_RegMask------------------------------------
4087 const RegMask &MergeMemNode::out_RegMask() const {
4088 return RegMask::Empty;
4089 }
4090
4091 //------------------------------dump_spec--------------------------------------
4092 #ifndef PRODUCT
4093 void MergeMemNode::dump_spec(outputStream *st) const {
4094 st->print(" {");
4095 Node* base_mem = base_memory();
4096 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4097 Node* mem = memory_at(i);
4098 if (mem == base_mem) { st->print(" -"); continue; }
4099 st->print( " N%d:", mem->_idx );
4100 Compile::current()->get_adr_type(i)->dump_on(st);
4101 }
4102 st->print(" }");
4103 }
4104 #endif // !PRODUCT
4105
4106
4107 #ifdef ASSERT
4108 static bool might_be_same(Node* a, Node* b) {
4109 if (a == b) return true;
4110 if (!(a->is_Phi() || b->is_Phi())) return false;
4111 // phis shift around during optimization
4112 return true; // pretty stupid...
4113 }
4114
4115 // verify a narrow slice (either incoming or outgoing)
4116 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4117 if (!VerifyAliases) return; // don't bother to verify unless requested
4118 if (is_error_reported()) return; // muzzle asserts when debugging an error
4119 if (Node::in_dump()) return; // muzzle asserts when printing
4120 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4121 assert(n != NULL, "");
4122 // Elide intervening MergeMem's
4123 while (n->is_MergeMem()) {
4124 n = n->as_MergeMem()->memory_at(alias_idx);
4125 }
4126 Compile* C = Compile::current();
4127 const TypePtr* n_adr_type = n->adr_type();
4128 if (n == m->empty_memory()) {
4129 // Implicit copy of base_memory()
4130 } else if (n_adr_type != TypePtr::BOTTOM) {
4131 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4132 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4133 } else {
4134 // A few places like make_runtime_call "know" that VM calls are narrow,
4135 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4136 bool expected_wide_mem = false;
4137 if (n == m->base_memory()) {
4138 expected_wide_mem = true;
4139 } else if (alias_idx == Compile::AliasIdxRaw ||
4140 n == m->memory_at(Compile::AliasIdxRaw)) {
4141 expected_wide_mem = true;
4142 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4143 // memory can "leak through" calls on channels that
4144 // are write-once. Allow this also.
4145 expected_wide_mem = true;
4146 }
4147 assert(expected_wide_mem, "expected narrow slice replacement");
4148 }
4149 }
4150 #else // !ASSERT
4151 #define verify_memory_slice(m,i,n) (0) // PRODUCT version is no-op
4152 #endif
4153
4154
4155 //-----------------------------memory_at---------------------------------------
4156 Node* MergeMemNode::memory_at(uint alias_idx) const {
4157 assert(alias_idx >= Compile::AliasIdxRaw ||
4158 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4159 "must avoid base_memory and AliasIdxTop");
4160
4161 // Otherwise, it is a narrow slice.
4162 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4163 Compile *C = Compile::current();
4164 if (is_empty_memory(n)) {
4165 // the array is sparse; empty slots are the "top" node
4166 n = base_memory();
4167 assert(Node::in_dump()
4168 || n == NULL || n->bottom_type() == Type::TOP
4169 || n->adr_type() == NULL // address is TOP
4170 || n->adr_type() == TypePtr::BOTTOM
4171 || n->adr_type() == TypeRawPtr::BOTTOM
4172 || Compile::current()->AliasLevel() == 0,
4173 "must be a wide memory");
4174 // AliasLevel == 0 if we are organizing the memory states manually.
4175 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4176 } else {
4177 // make sure the stored slice is sane
4178 #ifdef ASSERT
4179 if (is_error_reported() || Node::in_dump()) {
4180 } else if (might_be_same(n, base_memory())) {
4181 // Give it a pass: It is a mostly harmless repetition of the base.
4182 // This can arise normally from node subsumption during optimization.
4183 } else {
4184 verify_memory_slice(this, alias_idx, n);
4185 }
4186 #endif
4187 }
4188 return n;
4189 }
4190
4191 //---------------------------set_memory_at-------------------------------------
4192 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4193 verify_memory_slice(this, alias_idx, n);
4194 Node* empty_mem = empty_memory();
4195 if (n == base_memory()) n = empty_mem; // collapse default
4196 uint need_req = alias_idx+1;
4197 if (req() < need_req) {
4198 if (n == empty_mem) return; // already the default, so do not grow me
4199 // grow the sparse array
4200 do {
4201 add_req(empty_mem);
4202 } while (req() < need_req);
4203 }
4204 set_req( alias_idx, n );
4205 }
4206
4207
4208
4209 //--------------------------iteration_setup------------------------------------
4210 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4211 if (other != NULL) {
4212 grow_to_match(other);
4213 // invariant: the finite support of mm2 is within mm->req()
4214 #ifdef ASSERT
4215 for (uint i = req(); i < other->req(); i++) {
4216 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4217 }
4218 #endif
4219 }
4220 // Replace spurious copies of base_memory by top.
4221 Node* base_mem = base_memory();
4222 if (base_mem != NULL && !base_mem->is_top()) {
4223 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4224 if (in(i) == base_mem)
4225 set_req(i, empty_memory());
4226 }
4227 }
4228 }
4229
4230 //---------------------------grow_to_match-------------------------------------
4231 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4232 Node* empty_mem = empty_memory();
4233 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4234 // look for the finite support of the other memory
4235 for (uint i = other->req(); --i >= req(); ) {
4236 if (other->in(i) != empty_mem) {
4237 uint new_len = i+1;
4238 while (req() < new_len) add_req(empty_mem);
4239 break;
4240 }
4241 }
4242 }
4243
4244 //---------------------------verify_sparse-------------------------------------
4245 #ifndef PRODUCT
4246 bool MergeMemNode::verify_sparse() const {
4247 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4248 Node* base_mem = base_memory();
4249 // The following can happen in degenerate cases, since empty==top.
4250 if (is_empty_memory(base_mem)) return true;
4251 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4252 assert(in(i) != NULL, "sane slice");
4253 if (in(i) == base_mem) return false; // should have been the sentinel value!
4254 }
4255 return true;
4256 }
4257
4258 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4259 Node* n;
4260 n = mm->in(idx);
4261 if (mem == n) return true; // might be empty_memory()
4262 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4263 if (mem == n) return true;
4264 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4265 if (mem == n) return true;
4266 if (n == NULL) break;
4267 }
4268 return false;
4269 }
4270 #endif // !PRODUCT