1 /*
2 * Copyright (c) 1997, 2021, 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/javaClasses.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "gc/shared/tlab_globals.hpp"
31 #include "memory/allocation.inline.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "oops/objArrayKlass.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/arraycopynode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/regalloc.hpp"
38 #include "opto/compile.hpp"
39 #include "opto/connode.hpp"
40 #include "opto/convertnode.hpp"
41 #include "opto/loopnode.hpp"
42 #include "opto/machnode.hpp"
43 #include "opto/matcher.hpp"
44 #include "opto/memnode.hpp"
45 #include "opto/mulnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/phaseX.hpp"
48 #include "opto/regmask.hpp"
49 #include "opto/rootnode.hpp"
50 #include "opto/vectornode.hpp"
51 #include "utilities/align.hpp"
52 #include "utilities/copy.hpp"
53 #include "utilities/macros.hpp"
54 #include "utilities/powerOfTwo.hpp"
55 #include "utilities/vmError.hpp"
56
57 // Portions of code courtesy of Clifford Click
58
59 // Optimization - Graph Style
60
61 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
62
63 //=============================================================================
64 uint MemNode::size_of() const { return sizeof(*this); }
65
66 const TypePtr *MemNode::adr_type() const {
67 Node* adr = in(Address);
68 if (adr == NULL) return NULL; // node is dead
69 const TypePtr* cross_check = NULL;
70 DEBUG_ONLY(cross_check = _adr_type);
71 return calculate_adr_type(adr->bottom_type(), cross_check);
72 }
73
74 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
75 if (adr != NULL) {
76 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
77 return true;
78 }
79 }
80 return false;
81 }
82
83 #ifndef PRODUCT
84 void MemNode::dump_spec(outputStream *st) const {
85 if (in(Address) == NULL) return; // node is dead
86 #ifndef ASSERT
87 // fake the missing field
88 const TypePtr* _adr_type = NULL;
89 if (in(Address) != NULL)
90 _adr_type = in(Address)->bottom_type()->isa_ptr();
91 #endif
92 dump_adr_type(this, _adr_type, st);
93
94 Compile* C = Compile::current();
95 if (C->alias_type(_adr_type)->is_volatile()) {
96 st->print(" Volatile!");
97 }
98 if (_unaligned_access) {
99 st->print(" unaligned");
100 }
101 if (_mismatched_access) {
102 st->print(" mismatched");
103 }
104 if (_unsafe_access) {
105 st->print(" unsafe");
106 }
107 }
108
109 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
110 st->print(" @");
111 if (adr_type == NULL) {
112 st->print("NULL");
113 } else {
114 adr_type->dump_on(st);
115 Compile* C = Compile::current();
116 Compile::AliasType* atp = NULL;
117 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
118 if (atp == NULL)
119 st->print(", idx=?\?;");
120 else if (atp->index() == Compile::AliasIdxBot)
121 st->print(", idx=Bot;");
122 else if (atp->index() == Compile::AliasIdxTop)
123 st->print(", idx=Top;");
124 else if (atp->index() == Compile::AliasIdxRaw)
125 st->print(", idx=Raw;");
126 else {
127 ciField* field = atp->field();
128 if (field) {
129 st->print(", name=");
130 field->print_name_on(st);
131 }
132 st->print(", idx=%d;", atp->index());
133 }
134 }
135 }
136
137 extern void print_alias_types();
138
139 #endif
140
141 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
142 assert((t_oop != NULL), "sanity");
143 bool is_instance = t_oop->is_known_instance_field();
144 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
145 (load != NULL) && load->is_Load() &&
146 (phase->is_IterGVN() != NULL);
147 if (!(is_instance || is_boxed_value_load))
148 return mchain; // don't try to optimize non-instance types
149 uint instance_id = t_oop->instance_id();
150 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
151 Node *prev = NULL;
152 Node *result = mchain;
153 while (prev != result) {
154 prev = result;
155 if (result == start_mem)
156 break; // hit one of our sentinels
157 // skip over a call which does not affect this memory slice
158 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
159 Node *proj_in = result->in(0);
160 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
161 break; // hit one of our sentinels
162 } else if (proj_in->is_Call()) {
163 // ArrayCopyNodes processed here as well
164 CallNode *call = proj_in->as_Call();
165 if (!call->may_modify(t_oop, phase)) { // returns false for instances
166 result = call->in(TypeFunc::Memory);
167 }
168 } else if (proj_in->is_Initialize()) {
169 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
170 // Stop if this is the initialization for the object instance which
171 // contains this memory slice, otherwise skip over it.
172 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
173 break;
174 }
175 if (is_instance) {
176 result = proj_in->in(TypeFunc::Memory);
177 } else if (is_boxed_value_load) {
178 Node* klass = alloc->in(AllocateNode::KlassNode);
179 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
180 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
181 result = proj_in->in(TypeFunc::Memory); // not related allocation
182 }
183 }
184 } else if (proj_in->is_MemBar()) {
185 ArrayCopyNode* ac = NULL;
186 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
187 break;
188 }
189 result = proj_in->in(TypeFunc::Memory);
190 } else {
191 assert(false, "unexpected projection");
192 }
193 } else if (result->is_ClearArray()) {
194 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
195 // Can not bypass initialization of the instance
196 // we are looking for.
197 break;
198 }
199 // Otherwise skip it (the call updated 'result' value).
200 } else if (result->is_MergeMem()) {
201 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
202 }
203 }
204 return result;
205 }
206
207 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
208 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
209 if (t_oop == NULL)
210 return mchain; // don't try to optimize non-oop types
211 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
212 bool is_instance = t_oop->is_known_instance_field();
213 PhaseIterGVN *igvn = phase->is_IterGVN();
214 if (is_instance && igvn != NULL && result->is_Phi()) {
215 PhiNode *mphi = result->as_Phi();
216 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
217 const TypePtr *t = mphi->adr_type();
218 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
219 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
220 t->is_oopptr()->cast_to_exactness(true)
221 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
222 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) {
223 // clone the Phi with our address type
224 result = mphi->split_out_instance(t_adr, igvn);
225 } else {
226 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
227 }
228 }
229 return result;
230 }
231
232 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
233 uint alias_idx = phase->C->get_alias_index(tp);
234 Node *mem = mmem;
235 #ifdef ASSERT
236 {
237 // Check that current type is consistent with the alias index used during graph construction
238 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
239 bool consistent = adr_check == NULL || adr_check->empty() ||
240 phase->C->must_alias(adr_check, alias_idx );
241 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
242 if( !consistent && adr_check != NULL && !adr_check->empty() &&
243 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
244 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
245 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
246 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
247 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
248 // don't assert if it is dead code.
249 consistent = true;
250 }
251 if( !consistent ) {
252 st->print("alias_idx==%d, adr_check==", alias_idx);
253 if( adr_check == NULL ) {
254 st->print("NULL");
255 } else {
256 adr_check->dump();
257 }
258 st->cr();
259 print_alias_types();
260 assert(consistent, "adr_check must match alias idx");
261 }
262 }
263 #endif
264 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
265 // means an array I have not precisely typed yet. Do not do any
266 // alias stuff with it any time soon.
267 const TypeOopPtr *toop = tp->isa_oopptr();
268 if( tp->base() != Type::AnyPtr &&
269 !(toop &&
270 toop->klass() != NULL &&
271 toop->klass()->is_java_lang_Object() &&
272 toop->offset() == Type::OffsetBot) ) {
273 // compress paths and change unreachable cycles to TOP
274 // If not, we can update the input infinitely along a MergeMem cycle
275 // Equivalent code in PhiNode::Ideal
276 Node* m = phase->transform(mmem);
277 // If transformed to a MergeMem, get the desired slice
278 // Otherwise the returned node represents memory for every slice
279 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
280 // Update input if it is progress over what we have now
281 }
282 return mem;
283 }
284
285 //--------------------------Ideal_common---------------------------------------
286 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
287 // Unhook non-raw memories from complete (macro-expanded) initializations.
288 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
289 // If our control input is a dead region, kill all below the region
290 Node *ctl = in(MemNode::Control);
291 if (ctl && remove_dead_region(phase, can_reshape))
292 return this;
293 ctl = in(MemNode::Control);
294 // Don't bother trying to transform a dead node
295 if (ctl && ctl->is_top()) return NodeSentinel;
296
297 PhaseIterGVN *igvn = phase->is_IterGVN();
298 // Wait if control on the worklist.
299 if (ctl && can_reshape && igvn != NULL) {
300 Node* bol = NULL;
301 Node* cmp = NULL;
302 if (ctl->in(0)->is_If()) {
303 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
304 bol = ctl->in(0)->in(1);
305 if (bol->is_Bool())
306 cmp = ctl->in(0)->in(1)->in(1);
307 }
308 if (igvn->_worklist.member(ctl) ||
309 (bol != NULL && igvn->_worklist.member(bol)) ||
310 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
311 // This control path may be dead.
312 // Delay this memory node transformation until the control is processed.
313 igvn->_worklist.push(this);
314 return NodeSentinel; // caller will return NULL
315 }
316 }
317 // Ignore if memory is dead, or self-loop
318 Node *mem = in(MemNode::Memory);
319 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
320 assert(mem != this, "dead loop in MemNode::Ideal");
321
322 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
323 // This memory slice may be dead.
324 // Delay this mem node transformation until the memory is processed.
325 igvn->_worklist.push(this);
326 return NodeSentinel; // caller will return NULL
327 }
328
329 Node *address = in(MemNode::Address);
330 const Type *t_adr = phase->type(address);
331 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
332
333 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
334 // Unsafe off-heap access with zero address. Remove access and other control users
335 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
336 assert(ctl != NULL, "unsafe accesses should be control dependent");
337 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
338 Node* u = ctl->fast_out(i);
339 if (u != ctl) {
340 igvn->rehash_node_delayed(u);
341 int nb = u->replace_edge(ctl, phase->C->top(), igvn);
342 --i, imax -= nb;
343 }
344 }
345 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
346 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
347 phase->C->root()->add_req(halt);
348 return this;
349 }
350
351 if (can_reshape && igvn != NULL &&
352 (igvn->_worklist.member(address) ||
353 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
354 // The address's base and type may change when the address is processed.
355 // Delay this mem node transformation until the address is processed.
356 igvn->_worklist.push(this);
357 return NodeSentinel; // caller will return NULL
358 }
359
360 // Do NOT remove or optimize the next lines: ensure a new alias index
361 // is allocated for an oop pointer type before Escape Analysis.
362 // Note: C++ will not remove it since the call has side effect.
363 if (t_adr->isa_oopptr()) {
364 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
365 }
366
367 Node* base = NULL;
368 if (address->is_AddP()) {
369 base = address->in(AddPNode::Base);
370 }
371 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
372 !t_adr->isa_rawptr()) {
373 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
374 // Skip this node optimization if its address has TOP base.
375 return NodeSentinel; // caller will return NULL
376 }
377
378 // Avoid independent memory operations
379 Node* old_mem = mem;
380
381 // The code which unhooks non-raw memories from complete (macro-expanded)
382 // initializations was removed. After macro-expansion all stores catched
383 // by Initialize node became raw stores and there is no information
384 // which memory slices they modify. So it is unsafe to move any memory
385 // operation above these stores. Also in most cases hooked non-raw memories
386 // were already unhooked by using information from detect_ptr_independence()
387 // and find_previous_store().
388
389 if (mem->is_MergeMem()) {
390 MergeMemNode* mmem = mem->as_MergeMem();
391 const TypePtr *tp = t_adr->is_ptr();
392
393 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
394 }
395
396 if (mem != old_mem) {
397 set_req(MemNode::Memory, mem);
398 if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) {
399 igvn->_worklist.push(old_mem);
400 }
401 if (phase->type(mem) == Type::TOP) return NodeSentinel;
402 return this;
403 }
404
405 // let the subclass continue analyzing...
406 return NULL;
407 }
408
409 // Helper function for proving some simple control dominations.
410 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
411 // Already assumes that 'dom' is available at 'sub', and that 'sub'
412 // is not a constant (dominated by the method's StartNode).
413 // Used by MemNode::find_previous_store to prove that the
414 // control input of a memory operation predates (dominates)
415 // an allocation it wants to look past.
416 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
417 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
418 return false; // Conservative answer for dead code
419
420 // Check 'dom'. Skip Proj and CatchProj nodes.
421 dom = dom->find_exact_control(dom);
422 if (dom == NULL || dom->is_top())
423 return false; // Conservative answer for dead code
424
425 if (dom == sub) {
426 // For the case when, for example, 'sub' is Initialize and the original
427 // 'dom' is Proj node of the 'sub'.
428 return false;
429 }
430
431 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
432 return true;
433
434 // 'dom' dominates 'sub' if its control edge and control edges
435 // of all its inputs dominate or equal to sub's control edge.
436
437 // Currently 'sub' is either Allocate, Initialize or Start nodes.
438 // Or Region for the check in LoadNode::Ideal();
439 // 'sub' should have sub->in(0) != NULL.
440 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
441 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
442
443 // Get control edge of 'sub'.
444 Node* orig_sub = sub;
445 sub = sub->find_exact_control(sub->in(0));
446 if (sub == NULL || sub->is_top())
447 return false; // Conservative answer for dead code
448
449 assert(sub->is_CFG(), "expecting control");
450
451 if (sub == dom)
452 return true;
453
454 if (sub->is_Start() || sub->is_Root())
455 return false;
456
457 {
458 // Check all control edges of 'dom'.
459
460 ResourceMark rm;
461 Node_List nlist;
462 Unique_Node_List dom_list;
463
464 dom_list.push(dom);
465 bool only_dominating_controls = false;
466
467 for (uint next = 0; next < dom_list.size(); next++) {
468 Node* n = dom_list.at(next);
469 if (n == orig_sub)
470 return false; // One of dom's inputs dominated by sub.
471 if (!n->is_CFG() && n->pinned()) {
472 // Check only own control edge for pinned non-control nodes.
473 n = n->find_exact_control(n->in(0));
474 if (n == NULL || n->is_top())
475 return false; // Conservative answer for dead code
476 assert(n->is_CFG(), "expecting control");
477 dom_list.push(n);
478 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
479 only_dominating_controls = true;
480 } else if (n->is_CFG()) {
481 if (n->dominates(sub, nlist))
482 only_dominating_controls = true;
483 else
484 return false;
485 } else {
486 // First, own control edge.
487 Node* m = n->find_exact_control(n->in(0));
488 if (m != NULL) {
489 if (m->is_top())
490 return false; // Conservative answer for dead code
491 dom_list.push(m);
492 }
493 // Now, the rest of edges.
494 uint cnt = n->req();
495 for (uint i = 1; i < cnt; i++) {
496 m = n->find_exact_control(n->in(i));
497 if (m == NULL || m->is_top())
498 continue;
499 dom_list.push(m);
500 }
501 }
502 }
503 return only_dominating_controls;
504 }
505 }
506
507 //---------------------detect_ptr_independence---------------------------------
508 // Used by MemNode::find_previous_store to prove that two base
509 // pointers are never equal.
510 // The pointers are accompanied by their associated allocations,
511 // if any, which have been previously discovered by the caller.
512 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
513 Node* p2, AllocateNode* a2,
514 PhaseTransform* phase) {
515 // Attempt to prove that these two pointers cannot be aliased.
516 // They may both manifestly be allocations, and they should differ.
517 // Or, if they are not both allocations, they can be distinct constants.
518 // Otherwise, one is an allocation and the other a pre-existing value.
519 if (a1 == NULL && a2 == NULL) { // neither an allocation
520 return (p1 != p2) && p1->is_Con() && p2->is_Con();
521 } else if (a1 != NULL && a2 != NULL) { // both allocations
522 return (a1 != a2);
523 } else if (a1 != NULL) { // one allocation a1
524 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
525 return all_controls_dominate(p2, a1);
526 } else { //(a2 != NULL) // one allocation a2
527 return all_controls_dominate(p1, a2);
528 }
529 return false;
530 }
531
532
533 // Find an arraycopy ac that produces the memory state represented by parameter mem.
534 // Return ac if
535 // (a) can_see_stored_value=true and ac must have set the value for this load or if
536 // (b) can_see_stored_value=false and ac could have set the value for this load or if
537 // (c) can_see_stored_value=false and ac cannot have set the value for this load.
538 // In case (c) change the parameter mem to the memory input of ac to skip it
539 // when searching stored value.
540 // Otherwise return NULL.
541 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
542 ArrayCopyNode* ac = find_array_copy_clone(phase, ld_alloc, mem);
543 if (ac != NULL) {
544 Node* ld_addp = in(MemNode::Address);
545 Node* src = ac->in(ArrayCopyNode::Src);
546 const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
547
548 // This is a load from a cloned array. The corresponding arraycopy ac must
549 // have set the value for the load and we can return ac but only if the load
550 // is known to be within bounds. This is checked below.
551 if (ary_t != NULL && ld_addp->is_AddP()) {
552 Node* ld_offs = ld_addp->in(AddPNode::Offset);
553 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type();
554 jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
555 jlong elemsize = type2aelembytes(ary_elem);
556
557 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
558 const TypeInt* sizetype = ary_t->size();
559
560 if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
561 // The load is known to be within bounds. It receives its value from ac.
562 return ac;
563 }
564 // The load is known to be out-of-bounds.
565 }
566 // The load could be out-of-bounds. It must not be hoisted but must remain
567 // dependent on the runtime range check. This is achieved by returning NULL.
568 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
569 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
570
571 if (ac->is_arraycopy_validated() ||
572 ac->is_copyof_validated() ||
573 ac->is_copyofrange_validated()) {
574 Node* ld_addp = in(MemNode::Address);
575 if (ld_addp->is_AddP()) {
576 Node* ld_base = ld_addp->in(AddPNode::Address);
577 Node* ld_offs = ld_addp->in(AddPNode::Offset);
578
579 Node* dest = ac->in(ArrayCopyNode::Dest);
580
581 if (dest == ld_base) {
582 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
583 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
584 return ac;
585 }
586 if (!can_see_stored_value) {
587 mem = ac->in(TypeFunc::Memory);
588 return ac;
589 }
590 }
591 }
592 }
593 }
594 return NULL;
595 }
596
597 ArrayCopyNode* MemNode::find_array_copy_clone(PhaseTransform* phase, Node* ld_alloc, Node* mem) const {
598 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
599 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
600 if (ld_alloc != NULL) {
601 // Check if there is an array copy for a clone
602 Node* mb = mem->in(0);
603 ArrayCopyNode* ac = NULL;
604 if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
605 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
606 ac = mb->in(0)->in(0)->as_ArrayCopy();
607 } else {
608 // Step over GC barrier when ReduceInitialCardMarks is disabled
609 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
610 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
611
612 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
613 ac = control_proj_ac->in(0)->as_ArrayCopy();
614 }
615 }
616
617 if (ac != NULL && ac->is_clonebasic()) {
618 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase);
619 if (alloc != NULL && alloc == ld_alloc) {
620 return ac;
621 }
622 }
623 }
624 }
625 return NULL;
626 }
627
628 // The logic for reordering loads and stores uses four steps:
629 // (a) Walk carefully past stores and initializations which we
630 // can prove are independent of this load.
631 // (b) Observe that the next memory state makes an exact match
632 // with self (load or store), and locate the relevant store.
633 // (c) Ensure that, if we were to wire self directly to the store,
634 // the optimizer would fold it up somehow.
635 // (d) Do the rewiring, and return, depending on some other part of
636 // the optimizer to fold up the load.
637 // This routine handles steps (a) and (b). Steps (c) and (d) are
638 // specific to loads and stores, so they are handled by the callers.
639 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
640 //
641 Node* MemNode::find_previous_store(PhaseTransform* phase) {
642 Node* ctrl = in(MemNode::Control);
643 Node* adr = in(MemNode::Address);
644 intptr_t offset = 0;
645 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
646 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
647
648 if (offset == Type::OffsetBot)
649 return NULL; // cannot unalias unless there are precise offsets
650
651 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
652 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
653
654 intptr_t size_in_bytes = memory_size();
655
656 Node* mem = in(MemNode::Memory); // start searching here...
657
658 int cnt = 50; // Cycle limiter
659 for (;;) { // While we can dance past unrelated stores...
660 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
661
662 Node* prev = mem;
663 if (mem->is_Store()) {
664 Node* st_adr = mem->in(MemNode::Address);
665 intptr_t st_offset = 0;
666 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
667 if (st_base == NULL)
668 break; // inscrutable pointer
669
670 // For raw accesses it's not enough to prove that constant offsets don't intersect.
671 // We need the bases to be the equal in order for the offset check to make sense.
672 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
673 break;
674 }
675
676 if (st_offset != offset && st_offset != Type::OffsetBot) {
677 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
678 assert(mem->as_Store()->memory_size() <= MAX_STORE, "");
679 if (st_offset >= offset + size_in_bytes ||
680 st_offset <= offset - MAX_STORE ||
681 st_offset <= offset - mem->as_Store()->memory_size()) {
682 // Success: The offsets are provably independent.
683 // (You may ask, why not just test st_offset != offset and be done?
684 // The answer is that stores of different sizes can co-exist
685 // in the same sequence of RawMem effects. We sometimes initialize
686 // a whole 'tile' of array elements with a single jint or jlong.)
687 mem = mem->in(MemNode::Memory);
688 continue; // (a) advance through independent store memory
689 }
690 }
691 if (st_base != base &&
692 detect_ptr_independence(base, alloc,
693 st_base,
694 AllocateNode::Ideal_allocation(st_base, phase),
695 phase)) {
696 // Success: The bases are provably independent.
697 mem = mem->in(MemNode::Memory);
698 continue; // (a) advance through independent store memory
699 }
700
701 // (b) At this point, if the bases or offsets do not agree, we lose,
702 // since we have not managed to prove 'this' and 'mem' independent.
703 if (st_base == base && st_offset == offset) {
704 return mem; // let caller handle steps (c), (d)
705 }
706
707 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
708 InitializeNode* st_init = mem->in(0)->as_Initialize();
709 AllocateNode* st_alloc = st_init->allocation();
710 if (st_alloc == NULL)
711 break; // something degenerated
712 bool known_identical = false;
713 bool known_independent = false;
714 if (alloc == st_alloc)
715 known_identical = true;
716 else if (alloc != NULL)
717 known_independent = true;
718 else if (all_controls_dominate(this, st_alloc))
719 known_independent = true;
720
721 if (known_independent) {
722 // The bases are provably independent: Either they are
723 // manifestly distinct allocations, or else the control
724 // of this load dominates the store's allocation.
725 int alias_idx = phase->C->get_alias_index(adr_type());
726 if (alias_idx == Compile::AliasIdxRaw) {
727 mem = st_alloc->in(TypeFunc::Memory);
728 } else {
729 mem = st_init->memory(alias_idx);
730 }
731 continue; // (a) advance through independent store memory
732 }
733
734 // (b) at this point, if we are not looking at a store initializing
735 // the same allocation we are loading from, we lose.
736 if (known_identical) {
737 // From caller, can_see_stored_value will consult find_captured_store.
738 return mem; // let caller handle steps (c), (d)
739 }
740
741 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
742 if (prev != mem) {
743 // Found an arraycopy but it doesn't affect that load
744 continue;
745 }
746 // Found an arraycopy that may affect that load
747 return mem;
748 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
749 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
750 if (mem->is_Proj() && mem->in(0)->is_Call()) {
751 // ArrayCopyNodes processed here as well.
752 CallNode *call = mem->in(0)->as_Call();
753 if (!call->may_modify(addr_t, phase)) {
754 mem = call->in(TypeFunc::Memory);
755 continue; // (a) advance through independent call memory
756 }
757 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
758 ArrayCopyNode* ac = NULL;
759 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
760 break;
761 }
762 mem = mem->in(0)->in(TypeFunc::Memory);
763 continue; // (a) advance through independent MemBar memory
764 } else if (mem->is_ClearArray()) {
765 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
766 // (the call updated 'mem' value)
767 continue; // (a) advance through independent allocation memory
768 } else {
769 // Can not bypass initialization of the instance
770 // we are looking for.
771 return mem;
772 }
773 } else if (mem->is_MergeMem()) {
774 int alias_idx = phase->C->get_alias_index(adr_type());
775 mem = mem->as_MergeMem()->memory_at(alias_idx);
776 continue; // (a) advance through independent MergeMem memory
777 }
778 }
779
780 // Unless there is an explicit 'continue', we must bail out here,
781 // because 'mem' is an inscrutable memory state (e.g., a call).
782 break;
783 }
784
785 return NULL; // bail out
786 }
787
788 //----------------------calculate_adr_type-------------------------------------
789 // Helper function. Notices when the given type of address hits top or bottom.
790 // Also, asserts a cross-check of the type against the expected address type.
791 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
792 if (t == Type::TOP) return NULL; // does not touch memory any more?
793 #ifdef ASSERT
794 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL;
795 #endif
796 const TypePtr* tp = t->isa_ptr();
797 if (tp == NULL) {
798 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
799 return TypePtr::BOTTOM; // touches lots of memory
800 } else {
801 #ifdef ASSERT
802 // %%%% [phh] We don't check the alias index if cross_check is
803 // TypeRawPtr::BOTTOM. Needs to be investigated.
804 if (cross_check != NULL &&
805 cross_check != TypePtr::BOTTOM &&
806 cross_check != TypeRawPtr::BOTTOM) {
807 // Recheck the alias index, to see if it has changed (due to a bug).
808 Compile* C = Compile::current();
809 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
810 "must stay in the original alias category");
811 // The type of the address must be contained in the adr_type,
812 // disregarding "null"-ness.
813 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
814 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
815 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
816 "real address must not escape from expected memory type");
817 }
818 #endif
819 return tp;
820 }
821 }
822
823 //=============================================================================
824 // Should LoadNode::Ideal() attempt to remove control edges?
825 bool LoadNode::can_remove_control() const {
826 return !has_pinned_control_dependency();
827 }
828 uint LoadNode::size_of() const { return sizeof(*this); }
829 bool LoadNode::cmp( const Node &n ) const
830 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
831 const Type *LoadNode::bottom_type() const { return _type; }
832 uint LoadNode::ideal_reg() const {
833 return _type->ideal_reg();
834 }
835
836 #ifndef PRODUCT
837 void LoadNode::dump_spec(outputStream *st) const {
838 MemNode::dump_spec(st);
839 if( !Verbose && !WizardMode ) {
840 // standard dump does this in Verbose and WizardMode
841 st->print(" #"); _type->dump_on(st);
842 }
843 if (!depends_only_on_test()) {
844 st->print(" (does not depend only on test, ");
845 if (control_dependency() == UnknownControl) {
846 st->print("unknown control");
847 } else if (control_dependency() == Pinned) {
848 st->print("pinned");
849 } else if (adr_type() == TypeRawPtr::BOTTOM) {
850 st->print("raw access");
851 } else {
852 st->print("unknown reason");
853 }
854 st->print(")");
855 }
856 }
857 #endif
858
859 #ifdef ASSERT
860 //----------------------------is_immutable_value-------------------------------
861 // Helper function to allow a raw load without control edge for some cases
862 bool LoadNode::is_immutable_value(Node* adr) {
863 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
864 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
865 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
866 in_bytes(JavaThread::osthread_offset()) ||
867 adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
868 in_bytes(JavaThread::threadObj_offset())));
869 }
870 #endif
871
872 //----------------------------LoadNode::make-----------------------------------
873 // Polymorphic factory method:
874 Node* LoadNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, BasicType bt, MemOrd mo,
875 ControlDependency control_dependency, bool require_atomic_access, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
876 Compile* C = gvn.C;
877
878 // sanity check the alias category against the created node type
879 assert(!(adr_type->isa_oopptr() &&
880 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
881 "use LoadKlassNode instead");
882 assert(!(adr_type->isa_aryptr() &&
883 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
884 "use LoadRangeNode instead");
885 // Check control edge of raw loads
886 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
887 // oop will be recorded in oop map if load crosses safepoint
888 rt->isa_oopptr() || is_immutable_value(adr),
889 "raw memory operations should have control edge");
890 LoadNode* load = NULL;
891 switch (bt) {
892 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
893 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
894 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
895 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
896 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
897 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic_access); break;
898 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
899 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic_access); break;
900 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
901 case T_OBJECT:
902 #ifdef _LP64
903 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
904 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
905 } else
906 #endif
907 {
908 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
909 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
910 }
911 break;
912 default:
913 ShouldNotReachHere();
914 break;
915 }
916 assert(load != NULL, "LoadNode should have been created");
917 if (unaligned) {
918 load->set_unaligned_access();
919 }
920 if (mismatched) {
921 load->set_mismatched_access();
922 }
923 if (unsafe) {
924 load->set_unsafe_access();
925 }
926 load->set_barrier_data(barrier_data);
927 if (load->Opcode() == Op_LoadN) {
928 Node* ld = gvn.transform(load);
929 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
930 }
931
932 return load;
933 }
934
935 //------------------------------hash-------------------------------------------
936 uint LoadNode::hash() const {
937 // unroll addition of interesting fields
938 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
939 }
940
941 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
942 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
943 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
944 bool is_stable_ary = FoldStableValues &&
945 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
946 tp->isa_aryptr()->is_stable();
947
948 return (eliminate_boxing && non_volatile) || is_stable_ary;
949 }
950
951 return false;
952 }
953
954 // Is the value loaded previously stored by an arraycopy? If so return
955 // a load node that reads from the source array so we may be able to
956 // optimize out the ArrayCopy node later.
957 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
958 Node* ld_adr = in(MemNode::Address);
959 intptr_t ld_off = 0;
960 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
961 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
962 if (ac != NULL) {
963 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
964
965 Node* mem = ac->in(TypeFunc::Memory);
966 Node* ctl = ac->in(0);
967 Node* src = ac->in(ArrayCopyNode::Src);
968
969 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
970 return NULL;
971 }
972
973 LoadNode* ld = clone()->as_Load();
974 Node* addp = in(MemNode::Address)->clone();
975 if (ac->as_ArrayCopy()->is_clonebasic()) {
976 assert(ld_alloc != NULL, "need an alloc");
977 assert(addp->is_AddP(), "address must be addp");
978 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
979 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
980 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
981 addp->set_req(AddPNode::Base, src);
982 addp->set_req(AddPNode::Address, src);
983 } else {
984 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
985 ac->as_ArrayCopy()->is_copyof_validated() ||
986 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
987 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
988 addp->set_req(AddPNode::Base, src);
989 addp->set_req(AddPNode::Address, src);
990
991 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
992 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type();
993 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
994 uint shift = exact_log2(type2aelembytes(ary_elem));
995
996 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
997 #ifdef _LP64
998 diff = phase->transform(new ConvI2LNode(diff));
999 #endif
1000 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1001
1002 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1003 addp->set_req(AddPNode::Offset, offset);
1004 }
1005 addp = phase->transform(addp);
1006 #ifdef ASSERT
1007 const TypePtr* adr_type = phase->type(addp)->is_ptr();
1008 ld->_adr_type = adr_type;
1009 #endif
1010 ld->set_req(MemNode::Address, addp);
1011 ld->set_req(0, ctl);
1012 ld->set_req(MemNode::Memory, mem);
1013 // load depends on the tests that validate the arraycopy
1014 ld->_control_dependency = UnknownControl;
1015 return ld;
1016 }
1017 return NULL;
1018 }
1019
1020
1021 //---------------------------can_see_stored_value------------------------------
1022 // This routine exists to make sure this set of tests is done the same
1023 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
1024 // will change the graph shape in a way which makes memory alive twice at the
1025 // same time (uses the Oracle model of aliasing), then some
1026 // LoadXNode::Identity will fold things back to the equivalence-class model
1027 // of aliasing.
1028 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
1029 Node* ld_adr = in(MemNode::Address);
1030 intptr_t ld_off = 0;
1031 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1032 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base, phase);
1033 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1034 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
1035 // This is more general than load from boxing objects.
1036 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1037 uint alias_idx = atp->index();
1038 Node* result = NULL;
1039 Node* current = st;
1040 // Skip through chains of MemBarNodes checking the MergeMems for
1041 // new states for the slice of this load. Stop once any other
1042 // kind of node is encountered. Loads from final memory can skip
1043 // through any kind of MemBar but normal loads shouldn't skip
1044 // through MemBarAcquire since the could allow them to move out of
1045 // a synchronized region. It is not safe to step over MemBarCPUOrder,
1046 // because alias info above them may be inaccurate (e.g., due to
1047 // mixed/mismatched unsafe accesses).
1048 bool is_final_mem = !atp->is_rewritable();
1049 while (current->is_Proj()) {
1050 int opc = current->in(0)->Opcode();
1051 if ((is_final_mem && (opc == Op_MemBarAcquire ||
1052 opc == Op_MemBarAcquireLock ||
1053 opc == Op_LoadFence)) ||
1054 opc == Op_MemBarRelease ||
1055 opc == Op_StoreFence ||
1056 opc == Op_MemBarReleaseLock ||
1057 opc == Op_MemBarStoreStore ||
1058 opc == Op_StoreStoreFence) {
1059 Node* mem = current->in(0)->in(TypeFunc::Memory);
1060 if (mem->is_MergeMem()) {
1061 MergeMemNode* merge = mem->as_MergeMem();
1062 Node* new_st = merge->memory_at(alias_idx);
1063 if (new_st == merge->base_memory()) {
1064 // Keep searching
1065 current = new_st;
1066 continue;
1067 }
1068 // Save the new memory state for the slice and fall through
1069 // to exit.
1070 result = new_st;
1071 }
1072 }
1073 break;
1074 }
1075 if (result != NULL) {
1076 st = result;
1077 }
1078 }
1079
1080 // Loop around twice in the case Load -> Initialize -> Store.
1081 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1082 for (int trip = 0; trip <= 1; trip++) {
1083
1084 if (st->is_Store()) {
1085 Node* st_adr = st->in(MemNode::Address);
1086 if (st_adr != ld_adr) {
1087 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1088 intptr_t st_off = 0;
1089 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1090 if (ld_base == NULL) return NULL;
1091 if (st_base == NULL) return NULL;
1092 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return NULL;
1093 if (ld_off != st_off) return NULL;
1094 if (ld_off == Type::OffsetBot) return NULL;
1095 // Same base, same offset.
1096 // Possible improvement for arrays: check index value instead of absolute offset.
1097
1098 // At this point we have proven something like this setup:
1099 // B = << base >>
1100 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1101 // S = StoreQ(AddP( B , #Off), V)
1102 // (Actually, we haven't yet proven the Q's are the same.)
1103 // In other words, we are loading from a casted version of
1104 // the same pointer-and-offset that we stored to.
1105 // Casted version may carry a dependency and it is respected.
1106 // Thus, we are able to replace L by V.
1107 }
1108 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1109 if (store_Opcode() != st->Opcode()) {
1110 return NULL;
1111 }
1112 // LoadVector/StoreVector needs additional check to ensure the types match.
1113 if (store_Opcode() == Op_StoreVector) {
1114 const TypeVect* in_vt = st->as_StoreVector()->vect_type();
1115 const TypeVect* out_vt = as_LoadVector()->vect_type();
1116 if (in_vt != out_vt) {
1117 return NULL;
1118 }
1119 }
1120 return st->in(MemNode::ValueIn);
1121 }
1122
1123 // A load from a freshly-created object always returns zero.
1124 // (This can happen after LoadNode::Ideal resets the load's memory input
1125 // to find_captured_store, which returned InitializeNode::zero_memory.)
1126 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1127 (st->in(0) == ld_alloc) &&
1128 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1129 // return a zero value for the load's basic type
1130 // (This is one of the few places where a generic PhaseTransform
1131 // can create new nodes. Think of it as lazily manifesting
1132 // virtually pre-existing constants.)
1133 if (memory_type() != T_VOID) {
1134 if (ReduceBulkZeroing || find_array_copy_clone(phase, ld_alloc, in(MemNode::Memory)) == NULL) {
1135 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1136 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1137 // by the ArrayCopyNode.
1138 return phase->zerocon(memory_type());
1139 }
1140 } else {
1141 // TODO: materialize all-zero vector constant
1142 assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1143 }
1144 }
1145
1146 // A load from an initialization barrier can match a captured store.
1147 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1148 InitializeNode* init = st->in(0)->as_Initialize();
1149 AllocateNode* alloc = init->allocation();
1150 if ((alloc != NULL) && (alloc == ld_alloc)) {
1151 // examine a captured store value
1152 st = init->find_captured_store(ld_off, memory_size(), phase);
1153 if (st != NULL) {
1154 continue; // take one more trip around
1155 }
1156 }
1157 }
1158
1159 // Load boxed value from result of valueOf() call is input parameter.
1160 if (this->is_Load() && ld_adr->is_AddP() &&
1161 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1162 intptr_t ignore = 0;
1163 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1164 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1165 base = bs->step_over_gc_barrier(base);
1166 if (base != NULL && base->is_Proj() &&
1167 base->as_Proj()->_con == TypeFunc::Parms &&
1168 base->in(0)->is_CallStaticJava() &&
1169 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1170 return base->in(0)->in(TypeFunc::Parms);
1171 }
1172 }
1173
1174 break;
1175 }
1176
1177 return NULL;
1178 }
1179
1180 //----------------------is_instance_field_load_with_local_phi------------------
1181 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1182 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1183 in(Address)->is_AddP() ) {
1184 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1185 // Only instances and boxed values.
1186 if( t_oop != NULL &&
1187 (t_oop->is_ptr_to_boxed_value() ||
1188 t_oop->is_known_instance_field()) &&
1189 t_oop->offset() != Type::OffsetBot &&
1190 t_oop->offset() != Type::OffsetTop) {
1191 return true;
1192 }
1193 }
1194 return false;
1195 }
1196
1197 //------------------------------Identity---------------------------------------
1198 // Loads are identity if previous store is to same address
1199 Node* LoadNode::Identity(PhaseGVN* phase) {
1200 // If the previous store-maker is the right kind of Store, and the store is
1201 // to the same address, then we are equal to the value stored.
1202 Node* mem = in(Memory);
1203 Node* value = can_see_stored_value(mem, phase);
1204 if( value ) {
1205 // byte, short & char stores truncate naturally.
1206 // A load has to load the truncated value which requires
1207 // some sort of masking operation and that requires an
1208 // Ideal call instead of an Identity call.
1209 if (memory_size() < BytesPerInt) {
1210 // If the input to the store does not fit with the load's result type,
1211 // it must be truncated via an Ideal call.
1212 if (!phase->type(value)->higher_equal(phase->type(this)))
1213 return this;
1214 }
1215 // (This works even when value is a Con, but LoadNode::Value
1216 // usually runs first, producing the singleton type of the Con.)
1217 if (!has_pinned_control_dependency() || value->is_Con()) {
1218 return value;
1219 } else {
1220 return this;
1221 }
1222 }
1223
1224 if (has_pinned_control_dependency()) {
1225 return this;
1226 }
1227 // Search for an existing data phi which was generated before for the same
1228 // instance's field to avoid infinite generation of phis in a loop.
1229 Node *region = mem->in(0);
1230 if (is_instance_field_load_with_local_phi(region)) {
1231 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1232 int this_index = phase->C->get_alias_index(addr_t);
1233 int this_offset = addr_t->offset();
1234 int this_iid = addr_t->instance_id();
1235 if (!addr_t->is_known_instance() &&
1236 addr_t->is_ptr_to_boxed_value()) {
1237 // Use _idx of address base (could be Phi node) for boxed values.
1238 intptr_t ignore = 0;
1239 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1240 if (base == NULL) {
1241 return this;
1242 }
1243 this_iid = base->_idx;
1244 }
1245 const Type* this_type = bottom_type();
1246 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1247 Node* phi = region->fast_out(i);
1248 if (phi->is_Phi() && phi != mem &&
1249 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1250 return phi;
1251 }
1252 }
1253 }
1254
1255 return this;
1256 }
1257
1258 // Construct an equivalent unsigned load.
1259 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1260 BasicType bt = T_ILLEGAL;
1261 const Type* rt = NULL;
1262 switch (Opcode()) {
1263 case Op_LoadUB: return this;
1264 case Op_LoadUS: return this;
1265 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1266 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1267 default:
1268 assert(false, "no unsigned variant: %s", Name());
1269 return NULL;
1270 }
1271 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1272 raw_adr_type(), rt, bt, _mo, _control_dependency,
1273 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1274 }
1275
1276 // Construct an equivalent signed load.
1277 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1278 BasicType bt = T_ILLEGAL;
1279 const Type* rt = NULL;
1280 switch (Opcode()) {
1281 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1282 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1283 case Op_LoadB: // fall through
1284 case Op_LoadS: // fall through
1285 case Op_LoadI: // fall through
1286 case Op_LoadL: return this;
1287 default:
1288 assert(false, "no signed variant: %s", Name());
1289 return NULL;
1290 }
1291 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1292 raw_adr_type(), rt, bt, _mo, _control_dependency,
1293 false /*require_atomic_access*/, is_unaligned_access(), is_mismatched_access());
1294 }
1295
1296 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1297 BasicType bt = rt->basic_type();
1298 switch (Opcode()) {
1299 case Op_LoadI: return (bt == T_FLOAT);
1300 case Op_LoadL: return (bt == T_DOUBLE);
1301 case Op_LoadF: return (bt == T_INT);
1302 case Op_LoadD: return (bt == T_LONG);
1303
1304 default: return false;
1305 }
1306 }
1307
1308 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1309 BasicType bt = rt->basic_type();
1310 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1311 bool is_mismatched = is_mismatched_access();
1312 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1313 if (raw_type == NULL) {
1314 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1315 }
1316 const int op = Opcode();
1317 bool require_atomic_access = (op == Op_LoadL && ((LoadLNode*)this)->require_atomic_access()) ||
1318 (op == Op_LoadD && ((LoadDNode*)this)->require_atomic_access());
1319 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1320 raw_adr_type(), rt, bt, _mo, _control_dependency,
1321 require_atomic_access, is_unaligned_access(), is_mismatched);
1322 }
1323
1324 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1325 BasicType bt = vt->basic_type();
1326 switch (Opcode()) {
1327 case Op_StoreI: return (bt == T_FLOAT);
1328 case Op_StoreL: return (bt == T_DOUBLE);
1329 case Op_StoreF: return (bt == T_INT);
1330 case Op_StoreD: return (bt == T_LONG);
1331
1332 default: return false;
1333 }
1334 }
1335
1336 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1337 BasicType bt = vt->basic_type();
1338 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1339 const int op = Opcode();
1340 bool require_atomic_access = (op == Op_StoreL && ((StoreLNode*)this)->require_atomic_access()) ||
1341 (op == Op_StoreD && ((StoreDNode*)this)->require_atomic_access());
1342 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1343 raw_adr_type(), val, bt, _mo, require_atomic_access);
1344
1345 bool is_mismatched = is_mismatched_access();
1346 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1347 if (raw_type == NULL) {
1348 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1349 }
1350 if (is_mismatched) {
1351 st->set_mismatched_access();
1352 }
1353 return st;
1354 }
1355
1356 // We're loading from an object which has autobox behaviour.
1357 // If this object is result of a valueOf call we'll have a phi
1358 // merging a newly allocated object and a load from the cache.
1359 // We want to replace this load with the original incoming
1360 // argument to the valueOf call.
1361 Node* LoadNode::eliminate_autobox(PhaseIterGVN* igvn) {
1362 assert(igvn->C->eliminate_boxing(), "sanity");
1363 intptr_t ignore = 0;
1364 Node* base = AddPNode::Ideal_base_and_offset(in(Address), igvn, ignore);
1365 if ((base == NULL) || base->is_Phi()) {
1366 // Push the loads from the phi that comes from valueOf up
1367 // through it to allow elimination of the loads and the recovery
1368 // of the original value. It is done in split_through_phi().
1369 return NULL;
1370 } else if (base->is_Load() ||
1371 (base->is_DecodeN() && base->in(1)->is_Load())) {
1372 // Eliminate the load of boxed value for integer types from the cache
1373 // array by deriving the value from the index into the array.
1374 // Capture the offset of the load and then reverse the computation.
1375
1376 // Get LoadN node which loads a boxing object from 'cache' array.
1377 if (base->is_DecodeN()) {
1378 base = base->in(1);
1379 }
1380 if (!base->in(Address)->is_AddP()) {
1381 return NULL; // Complex address
1382 }
1383 AddPNode* address = base->in(Address)->as_AddP();
1384 Node* cache_base = address->in(AddPNode::Base);
1385 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1386 // Get ConP node which is static 'cache' field.
1387 cache_base = cache_base->in(1);
1388 }
1389 if ((cache_base != NULL) && cache_base->is_Con()) {
1390 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1391 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1392 Node* elements[4];
1393 int shift = exact_log2(type2aelembytes(T_OBJECT));
1394 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1395 if (count > 0 && elements[0]->is_Con() &&
1396 (count == 1 ||
1397 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1398 elements[1]->in(2) == igvn->intcon(shift)))) {
1399 ciObjArray* array = base_type->const_oop()->as_obj_array();
1400 // Fetch the box object cache[0] at the base of the array and get its value
1401 ciInstance* box = array->obj_at(0)->as_instance();
1402 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1403 assert(ik->is_box_klass(), "sanity");
1404 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1405 if (ik->nof_nonstatic_fields() == 1) {
1406 // This should be true nonstatic_field_at requires calling
1407 // nof_nonstatic_fields so check it anyway
1408 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1409 BasicType bt = c.basic_type();
1410 // Only integer types have boxing cache.
1411 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1412 bt == T_BYTE || bt == T_SHORT ||
1413 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1414 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1415 if (cache_low != (int)cache_low) {
1416 return NULL; // should not happen since cache is array indexed by value
1417 }
1418 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1419 if (offset != (int)offset) {
1420 return NULL; // should not happen since cache is array indexed by value
1421 }
1422 // Add up all the offsets making of the address of the load
1423 Node* result = elements[0];
1424 for (int i = 1; i < count; i++) {
1425 result = igvn->transform(new AddXNode(result, elements[i]));
1426 }
1427 // Remove the constant offset from the address and then
1428 result = igvn->transform(new AddXNode(result, igvn->MakeConX(-(int)offset)));
1429 // remove the scaling of the offset to recover the original index.
1430 if (result->Opcode() == Op_LShiftX && result->in(2) == igvn->intcon(shift)) {
1431 // Peel the shift off directly but wrap it in a dummy node
1432 // since Ideal can't return existing nodes
1433 igvn->_worklist.push(result); // remove dead node later
1434 result = new RShiftXNode(result->in(1), igvn->intcon(0));
1435 } else if (result->is_Add() && result->in(2)->is_Con() &&
1436 result->in(1)->Opcode() == Op_LShiftX &&
1437 result->in(1)->in(2) == igvn->intcon(shift)) {
1438 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1439 // but for boxing cache access we know that X<<Z will not overflow
1440 // (there is range check) so we do this optimizatrion by hand here.
1441 igvn->_worklist.push(result); // remove dead node later
1442 Node* add_con = new RShiftXNode(result->in(2), igvn->intcon(shift));
1443 result = new AddXNode(result->in(1)->in(1), igvn->transform(add_con));
1444 } else {
1445 result = new RShiftXNode(result, igvn->intcon(shift));
1446 }
1447 #ifdef _LP64
1448 if (bt != T_LONG) {
1449 result = new ConvL2INode(igvn->transform(result));
1450 }
1451 #else
1452 if (bt == T_LONG) {
1453 result = new ConvI2LNode(igvn->transform(result));
1454 }
1455 #endif
1456 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1457 // Need to preserve unboxing load type if it is unsigned.
1458 switch(this->Opcode()) {
1459 case Op_LoadUB:
1460 result = new AndINode(igvn->transform(result), igvn->intcon(0xFF));
1461 break;
1462 case Op_LoadUS:
1463 result = new AndINode(igvn->transform(result), igvn->intcon(0xFFFF));
1464 break;
1465 }
1466 return result;
1467 }
1468 }
1469 }
1470 }
1471 }
1472 return NULL;
1473 }
1474
1475 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1476 Node* region = phi->in(0);
1477 if (region == NULL) {
1478 return false; // Wait stable graph
1479 }
1480 uint cnt = phi->req();
1481 for (uint i = 1; i < cnt; i++) {
1482 Node* rc = region->in(i);
1483 if (rc == NULL || phase->type(rc) == Type::TOP)
1484 return false; // Wait stable graph
1485 Node* in = phi->in(i);
1486 if (in == NULL || phase->type(in) == Type::TOP)
1487 return false; // Wait stable graph
1488 }
1489 return true;
1490 }
1491 //------------------------------split_through_phi------------------------------
1492 // Split instance or boxed field load through Phi.
1493 Node* LoadNode::split_through_phi(PhaseGVN* phase) {
1494 if (req() > 3) {
1495 assert(is_LoadVector() && Opcode() != Op_LoadVector, "load has too many inputs");
1496 // LoadVector subclasses such as LoadVectorMasked have extra inputs that the logic below doesn't take into account
1497 return NULL;
1498 }
1499 Node* mem = in(Memory);
1500 Node* address = in(Address);
1501 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1502
1503 assert((t_oop != NULL) &&
1504 (t_oop->is_known_instance_field() ||
1505 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1506
1507 Compile* C = phase->C;
1508 intptr_t ignore = 0;
1509 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1510 bool base_is_phi = (base != NULL) && base->is_Phi();
1511 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1512 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1513 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1514
1515 if (!((mem->is_Phi() || base_is_phi) &&
1516 (load_boxed_values || t_oop->is_known_instance_field()))) {
1517 return NULL; // memory is not Phi
1518 }
1519
1520 if (mem->is_Phi()) {
1521 if (!stable_phi(mem->as_Phi(), phase)) {
1522 return NULL; // Wait stable graph
1523 }
1524 uint cnt = mem->req();
1525 // Check for loop invariant memory.
1526 if (cnt == 3) {
1527 for (uint i = 1; i < cnt; i++) {
1528 Node* in = mem->in(i);
1529 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1530 if (m == mem) {
1531 if (i == 1) {
1532 // if the first edge was a loop, check second edge too.
1533 // If both are replaceable - we are in an infinite loop
1534 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1535 if (n == mem) {
1536 break;
1537 }
1538 }
1539 set_req(Memory, mem->in(cnt - i));
1540 return this; // made change
1541 }
1542 }
1543 }
1544 }
1545 if (base_is_phi) {
1546 if (!stable_phi(base->as_Phi(), phase)) {
1547 return NULL; // Wait stable graph
1548 }
1549 uint cnt = base->req();
1550 // Check for loop invariant memory.
1551 if (cnt == 3) {
1552 for (uint i = 1; i < cnt; i++) {
1553 if (base->in(i) == base) {
1554 return NULL; // Wait stable graph
1555 }
1556 }
1557 }
1558 }
1559
1560 // Split through Phi (see original code in loopopts.cpp).
1561 assert(C->have_alias_type(t_oop), "instance should have alias type");
1562
1563 // Do nothing here if Identity will find a value
1564 // (to avoid infinite chain of value phis generation).
1565 if (this != Identity(phase)) {
1566 return NULL;
1567 }
1568
1569 // Select Region to split through.
1570 Node* region;
1571 if (!base_is_phi) {
1572 assert(mem->is_Phi(), "sanity");
1573 region = mem->in(0);
1574 // Skip if the region dominates some control edge of the address.
1575 if (!MemNode::all_controls_dominate(address, region))
1576 return NULL;
1577 } else if (!mem->is_Phi()) {
1578 assert(base_is_phi, "sanity");
1579 region = base->in(0);
1580 // Skip if the region dominates some control edge of the memory.
1581 if (!MemNode::all_controls_dominate(mem, region))
1582 return NULL;
1583 } else if (base->in(0) != mem->in(0)) {
1584 assert(base_is_phi && mem->is_Phi(), "sanity");
1585 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1586 region = base->in(0);
1587 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1588 region = mem->in(0);
1589 } else {
1590 return NULL; // complex graph
1591 }
1592 } else {
1593 assert(base->in(0) == mem->in(0), "sanity");
1594 region = mem->in(0);
1595 }
1596
1597 const Type* this_type = this->bottom_type();
1598 int this_index = C->get_alias_index(t_oop);
1599 int this_offset = t_oop->offset();
1600 int this_iid = t_oop->instance_id();
1601 if (!t_oop->is_known_instance() && load_boxed_values) {
1602 // Use _idx of address base for boxed values.
1603 this_iid = base->_idx;
1604 }
1605 PhaseIterGVN* igvn = phase->is_IterGVN();
1606 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1607 for (uint i = 1; i < region->req(); i++) {
1608 Node* x;
1609 Node* the_clone = NULL;
1610 Node* in = region->in(i);
1611 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1612 in != NULL && in->is_OuterStripMinedLoop()) {
1613 // No node should go in the outer strip mined loop
1614 in = in->in(LoopNode::EntryControl);
1615 }
1616 if (in == NULL || in == C->top()) {
1617 x = C->top(); // Dead path? Use a dead data op
1618 } else {
1619 x = this->clone(); // Else clone up the data op
1620 the_clone = x; // Remember for possible deletion.
1621 // Alter data node to use pre-phi inputs
1622 if (this->in(0) == region) {
1623 x->set_req(0, in);
1624 } else {
1625 x->set_req(0, NULL);
1626 }
1627 if (mem->is_Phi() && (mem->in(0) == region)) {
1628 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1629 }
1630 if (address->is_Phi() && address->in(0) == region) {
1631 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1632 }
1633 if (base_is_phi && (base->in(0) == region)) {
1634 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1635 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1636 x->set_req(Address, adr_x);
1637 }
1638 }
1639 // Check for a 'win' on some paths
1640 const Type *t = x->Value(igvn);
1641
1642 bool singleton = t->singleton();
1643
1644 // See comments in PhaseIdealLoop::split_thru_phi().
1645 if (singleton && t == Type::TOP) {
1646 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1647 }
1648
1649 if (singleton) {
1650 x = igvn->makecon(t);
1651 } else {
1652 // We now call Identity to try to simplify the cloned node.
1653 // Note that some Identity methods call phase->type(this).
1654 // Make sure that the type array is big enough for
1655 // our new node, even though we may throw the node away.
1656 // (This tweaking with igvn only works because x is a new node.)
1657 igvn->set_type(x, t);
1658 // If x is a TypeNode, capture any more-precise type permanently into Node
1659 // otherwise it will be not updated during igvn->transform since
1660 // igvn->type(x) is set to x->Value() already.
1661 x->raise_bottom_type(t);
1662 Node* y = x->Identity(igvn);
1663 if (y != x) {
1664 x = y;
1665 } else {
1666 y = igvn->hash_find_insert(x);
1667 if (y) {
1668 x = y;
1669 } else {
1670 // Else x is a new node we are keeping
1671 // We do not need register_new_node_with_optimizer
1672 // because set_type has already been called.
1673 igvn->_worklist.push(x);
1674 }
1675 }
1676 }
1677 if (x != the_clone && the_clone != NULL) {
1678 igvn->remove_dead_node(the_clone);
1679 }
1680 phi->set_req(i, x);
1681 }
1682 // Record Phi
1683 igvn->register_new_node_with_optimizer(phi);
1684 return phi;
1685 }
1686
1687 AllocateNode* LoadNode::is_new_object_mark_load(PhaseGVN *phase) const {
1688 if (Opcode() == Op_LoadX) {
1689 Node* address = in(MemNode::Address);
1690 AllocateNode* alloc = AllocateNode::Ideal_allocation(address, phase);
1691 Node* mem = in(MemNode::Memory);
1692 if (alloc != NULL && mem->is_Proj() &&
1693 mem->in(0) != NULL &&
1694 mem->in(0) == alloc->initialization() &&
1695 alloc->initialization()->proj_out_or_null(0) != NULL) {
1696 return alloc;
1697 }
1698 }
1699 return NULL;
1700 }
1701
1702
1703 //------------------------------Ideal------------------------------------------
1704 // If the load is from Field memory and the pointer is non-null, it might be possible to
1705 // zero out the control input.
1706 // If the offset is constant and the base is an object allocation,
1707 // try to hook me up to the exact initializing store.
1708 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1709 if (has_pinned_control_dependency()) {
1710 return NULL;
1711 }
1712 Node* p = MemNode::Ideal_common(phase, can_reshape);
1713 if (p) return (p == NodeSentinel) ? NULL : p;
1714
1715 Node* ctrl = in(MemNode::Control);
1716 Node* address = in(MemNode::Address);
1717 bool progress = false;
1718
1719 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1720 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1721
1722 // Skip up past a SafePoint control. Cannot do this for Stores because
1723 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1724 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1725 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1726 !addr_mark &&
1727 (depends_only_on_test() || has_unknown_control_dependency())) {
1728 ctrl = ctrl->in(0);
1729 set_req(MemNode::Control,ctrl);
1730 progress = true;
1731 }
1732
1733 intptr_t ignore = 0;
1734 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1735 if (base != NULL
1736 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1737 // Check for useless control edge in some common special cases
1738 if (in(MemNode::Control) != NULL
1739 && can_remove_control()
1740 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1741 && all_controls_dominate(base, phase->C->start())) {
1742 // A method-invariant, non-null address (constant or 'this' argument).
1743 set_req(MemNode::Control, NULL);
1744 progress = true;
1745 }
1746 }
1747
1748 Node* mem = in(MemNode::Memory);
1749 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1750
1751 if (can_reshape && (addr_t != NULL)) {
1752 // try to optimize our memory input
1753 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1754 if (opt_mem != mem) {
1755 set_req_X(MemNode::Memory, opt_mem, phase);
1756 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1757 return this;
1758 }
1759 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1760 if ((t_oop != NULL) &&
1761 (t_oop->is_known_instance_field() ||
1762 t_oop->is_ptr_to_boxed_value())) {
1763 PhaseIterGVN *igvn = phase->is_IterGVN();
1764 assert(igvn != NULL, "must be PhaseIterGVN when can_reshape is true");
1765 if (igvn->_worklist.member(opt_mem)) {
1766 // Delay this transformation until memory Phi is processed.
1767 igvn->_worklist.push(this);
1768 return NULL;
1769 }
1770 // Split instance field load through Phi.
1771 Node* result = split_through_phi(phase);
1772 if (result != NULL) return result;
1773
1774 if (t_oop->is_ptr_to_boxed_value()) {
1775 Node* result = eliminate_autobox(igvn);
1776 if (result != NULL) return result;
1777 }
1778 }
1779 }
1780
1781 // Is there a dominating load that loads the same value? Leave
1782 // anything that is not a load of a field/array element (like
1783 // barriers etc.) alone
1784 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) {
1785 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1786 Node *use = mem->fast_out(i);
1787 if (use != this &&
1788 use->Opcode() == Opcode() &&
1789 use->in(0) != NULL &&
1790 use->in(0) != in(0) &&
1791 use->in(Address) == in(Address)) {
1792 Node* ctl = in(0);
1793 for (int i = 0; i < 10 && ctl != NULL; i++) {
1794 ctl = IfNode::up_one_dom(ctl);
1795 if (ctl == use->in(0)) {
1796 set_req(0, use->in(0));
1797 return this;
1798 }
1799 }
1800 }
1801 }
1802 }
1803
1804 // Check for prior store with a different base or offset; make Load
1805 // independent. Skip through any number of them. Bail out if the stores
1806 // are in an endless dead cycle and report no progress. This is a key
1807 // transform for Reflection. However, if after skipping through the Stores
1808 // we can't then fold up against a prior store do NOT do the transform as
1809 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1810 // array memory alive twice: once for the hoisted Load and again after the
1811 // bypassed Store. This situation only works if EVERYBODY who does
1812 // anti-dependence work knows how to bypass. I.e. we need all
1813 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1814 // the alias index stuff. So instead, peek through Stores and IFF we can
1815 // fold up, do so.
1816 Node* prev_mem = find_previous_store(phase);
1817 if (prev_mem != NULL) {
1818 Node* value = can_see_arraycopy_value(prev_mem, phase);
1819 if (value != NULL) {
1820 return value;
1821 }
1822 }
1823 // Steps (a), (b): Walk past independent stores to find an exact match.
1824 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1825 // (c) See if we can fold up on the spot, but don't fold up here.
1826 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1827 // just return a prior value, which is done by Identity calls.
1828 if (can_see_stored_value(prev_mem, phase)) {
1829 // Make ready for step (d):
1830 set_req_X(MemNode::Memory, prev_mem, phase);
1831 return this;
1832 }
1833 }
1834
1835 AllocateNode* alloc = is_new_object_mark_load(phase);
1836 if (alloc != NULL && alloc->Opcode() == Op_Allocate && UseBiasedLocking) {
1837 InitializeNode* init = alloc->initialization();
1838 Node* control = init->proj_out(0);
1839 return alloc->make_ideal_mark(phase, address, control, mem);
1840 }
1841
1842 return progress ? this : NULL;
1843 }
1844
1845 // Helper to recognize certain Klass fields which are invariant across
1846 // some group of array types (e.g., int[] or all T[] where T < Object).
1847 const Type*
1848 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1849 ciKlass* klass) const {
1850 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1851 // The field is Klass::_modifier_flags. Return its (constant) value.
1852 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1853 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1854 return TypeInt::make(klass->modifier_flags());
1855 }
1856 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1857 // The field is Klass::_access_flags. Return its (constant) value.
1858 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1859 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1860 return TypeInt::make(klass->access_flags());
1861 }
1862 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1863 // The field is Klass::_layout_helper. Return its constant value if known.
1864 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1865 return TypeInt::make(klass->layout_helper());
1866 }
1867
1868 // No match.
1869 return NULL;
1870 }
1871
1872 //------------------------------Value-----------------------------------------
1873 const Type* LoadNode::Value(PhaseGVN* phase) const {
1874 // Either input is TOP ==> the result is TOP
1875 Node* mem = in(MemNode::Memory);
1876 const Type *t1 = phase->type(mem);
1877 if (t1 == Type::TOP) return Type::TOP;
1878 Node* adr = in(MemNode::Address);
1879 const TypePtr* tp = phase->type(adr)->isa_ptr();
1880 if (tp == NULL || tp->empty()) return Type::TOP;
1881 int off = tp->offset();
1882 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1883 Compile* C = phase->C;
1884
1885 // Try to guess loaded type from pointer type
1886 if (tp->isa_aryptr()) {
1887 const TypeAryPtr* ary = tp->is_aryptr();
1888 const Type* t = ary->elem();
1889
1890 // Determine whether the reference is beyond the header or not, by comparing
1891 // the offset against the offset of the start of the array's data.
1892 // Different array types begin at slightly different offsets (12 vs. 16).
1893 // We choose T_BYTE as an example base type that is least restrictive
1894 // as to alignment, which will therefore produce the smallest
1895 // possible base offset.
1896 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1897 const bool off_beyond_header = (off >= min_base_off);
1898
1899 // Try to constant-fold a stable array element.
1900 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1901 // Make sure the reference is not into the header and the offset is constant
1902 ciObject* aobj = ary->const_oop();
1903 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1904 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1905 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1906 stable_dimension,
1907 memory_type(), is_unsigned());
1908 if (con_type != NULL) {
1909 return con_type;
1910 }
1911 }
1912 }
1913
1914 // Don't do this for integer types. There is only potential profit if
1915 // the element type t is lower than _type; that is, for int types, if _type is
1916 // more restrictive than t. This only happens here if one is short and the other
1917 // char (both 16 bits), and in those cases we've made an intentional decision
1918 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1919 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1920 //
1921 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1922 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1923 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1924 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1925 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1926 // In fact, that could have been the original type of p1, and p1 could have
1927 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1928 // expression (LShiftL quux 3) independently optimized to the constant 8.
1929 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1930 && (_type->isa_vect() == NULL)
1931 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1932 // t might actually be lower than _type, if _type is a unique
1933 // concrete subclass of abstract class t.
1934 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
1935 const Type* jt = t->join_speculative(_type);
1936 // In any case, do not allow the join, per se, to empty out the type.
1937 if (jt->empty() && !t->empty()) {
1938 // This can happen if a interface-typed array narrows to a class type.
1939 jt = _type;
1940 }
1941 #ifdef ASSERT
1942 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1943 // The pointers in the autobox arrays are always non-null
1944 Node* base = adr->in(AddPNode::Base);
1945 if ((base != NULL) && base->is_DecodeN()) {
1946 // Get LoadN node which loads IntegerCache.cache field
1947 base = base->in(1);
1948 }
1949 if ((base != NULL) && base->is_Con()) {
1950 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1951 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1952 // It could be narrow oop
1953 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1954 }
1955 }
1956 }
1957 #endif
1958 return jt;
1959 }
1960 }
1961 } else if (tp->base() == Type::InstPtr) {
1962 assert( off != Type::OffsetBot ||
1963 // arrays can be cast to Objects
1964 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1965 // unsafe field access may not have a constant offset
1966 C->has_unsafe_access(),
1967 "Field accesses must be precise" );
1968 // For oop loads, we expect the _type to be precise.
1969
1970 // Optimize loads from constant fields.
1971 const TypeInstPtr* tinst = tp->is_instptr();
1972 ciObject* const_oop = tinst->const_oop();
1973 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1974 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
1975 if (con_type != NULL) {
1976 return con_type;
1977 }
1978 }
1979 } else if (tp->base() == Type::KlassPtr) {
1980 assert( off != Type::OffsetBot ||
1981 // arrays can be cast to Objects
1982 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1983 // also allow array-loading from the primary supertype
1984 // array during subtype checks
1985 Opcode() == Op_LoadKlass,
1986 "Field accesses must be precise" );
1987 // For klass/static loads, we expect the _type to be precise
1988 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
1989 /* With mirrors being an indirect in the Klass*
1990 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
1991 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
1992 *
1993 * So check the type and klass of the node before the LoadP.
1994 */
1995 Node* adr2 = adr->in(MemNode::Address);
1996 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1997 if (tkls != NULL && !StressReflectiveCode) {
1998 ciKlass* klass = tkls->klass();
1999 if (klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
2000 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2001 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
2002 return TypeInstPtr::make(klass->java_mirror());
2003 }
2004 }
2005 }
2006
2007 const TypeKlassPtr *tkls = tp->isa_klassptr();
2008 if (tkls != NULL && !StressReflectiveCode) {
2009 ciKlass* klass = tkls->klass();
2010 if (klass->is_loaded() && tkls->klass_is_exact()) {
2011 // We are loading a field from a Klass metaobject whose identity
2012 // is known at compile time (the type is "exact" or "precise").
2013 // Check for fields we know are maintained as constants by the VM.
2014 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
2015 // The field is Klass::_super_check_offset. Return its (constant) value.
2016 // (Folds up type checking code.)
2017 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
2018 return TypeInt::make(klass->super_check_offset());
2019 }
2020 // Compute index into primary_supers array
2021 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2022 // Check for overflowing; use unsigned compare to handle the negative case.
2023 if( depth < ciKlass::primary_super_limit() ) {
2024 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2025 // (Folds up type checking code.)
2026 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2027 ciKlass *ss = klass->super_of_depth(depth);
2028 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
2029 }
2030 const Type* aift = load_array_final_field(tkls, klass);
2031 if (aift != NULL) return aift;
2032 }
2033
2034 // We can still check if we are loading from the primary_supers array at a
2035 // shallow enough depth. Even though the klass is not exact, entries less
2036 // than or equal to its super depth are correct.
2037 if (klass->is_loaded() ) {
2038 ciType *inner = klass;
2039 while( inner->is_obj_array_klass() )
2040 inner = inner->as_obj_array_klass()->base_element_type();
2041 if( inner->is_instance_klass() &&
2042 !inner->as_instance_klass()->flags().is_interface() ) {
2043 // Compute index into primary_supers array
2044 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2045 // Check for overflowing; use unsigned compare to handle the negative case.
2046 if( depth < ciKlass::primary_super_limit() &&
2047 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
2048 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2049 // (Folds up type checking code.)
2050 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2051 ciKlass *ss = klass->super_of_depth(depth);
2052 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
2053 }
2054 }
2055 }
2056
2057 // If the type is enough to determine that the thing is not an array,
2058 // we can give the layout_helper a positive interval type.
2059 // This will help short-circuit some reflective code.
2060 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
2061 && !klass->is_array_klass() // not directly typed as an array
2062 && !klass->is_interface() // specifically not Serializable & Cloneable
2063 && !klass->is_java_lang_Object() // not the supertype of all T[]
2064 ) {
2065 // Note: When interfaces are reliable, we can narrow the interface
2066 // test to (klass != Serializable && klass != Cloneable).
2067 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2068 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2069 // The key property of this type is that it folds up tests
2070 // for array-ness, since it proves that the layout_helper is positive.
2071 // Thus, a generic value like the basic object layout helper works fine.
2072 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2073 }
2074 }
2075
2076 // If we are loading from a freshly-allocated object, produce a zero,
2077 // if the load is provably beyond the header of the object.
2078 // (Also allow a variable load from a fresh array to produce zero.)
2079 const TypeOopPtr *tinst = tp->isa_oopptr();
2080 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
2081 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
2082 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
2083 Node* value = can_see_stored_value(mem,phase);
2084 if (value != NULL && value->is_Con()) {
2085 assert(value->bottom_type()->higher_equal(_type),"sanity");
2086 return value->bottom_type();
2087 }
2088 }
2089
2090 bool is_vect = (_type->isa_vect() != NULL);
2091 if (is_instance && !is_vect) {
2092 // If we have an instance type and our memory input is the
2093 // programs's initial memory state, there is no matching store,
2094 // so just return a zero of the appropriate type -
2095 // except if it is vectorized - then we have no zero constant.
2096 Node *mem = in(MemNode::Memory);
2097 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2098 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2099 return Type::get_zero_type(_type->basic_type());
2100 }
2101 }
2102
2103 Node* alloc = is_new_object_mark_load(phase);
2104 if (alloc != NULL && !(alloc->Opcode() == Op_Allocate && UseBiasedLocking)) {
2105 return TypeX::make(markWord::prototype().value());
2106 }
2107
2108 return _type;
2109 }
2110
2111 //------------------------------match_edge-------------------------------------
2112 // Do we Match on this edge index or not? Match only the address.
2113 uint LoadNode::match_edge(uint idx) const {
2114 return idx == MemNode::Address;
2115 }
2116
2117 //--------------------------LoadBNode::Ideal--------------------------------------
2118 //
2119 // If the previous store is to the same address as this load,
2120 // and the value stored was larger than a byte, replace this load
2121 // with the value stored truncated to a byte. If no truncation is
2122 // needed, the replacement is done in LoadNode::Identity().
2123 //
2124 Node* LoadBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2125 Node* mem = in(MemNode::Memory);
2126 Node* value = can_see_stored_value(mem,phase);
2127 if (value != NULL) {
2128 Node* narrow = Compile::narrow_value(T_BYTE, value, _type, phase, false);
2129 if (narrow != value) {
2130 return narrow;
2131 }
2132 }
2133 // Identity call will handle the case where truncation is not needed.
2134 return LoadNode::Ideal(phase, can_reshape);
2135 }
2136
2137 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2138 Node* mem = in(MemNode::Memory);
2139 Node* value = can_see_stored_value(mem,phase);
2140 if (value != NULL && value->is_Con() &&
2141 !value->bottom_type()->higher_equal(_type)) {
2142 // If the input to the store does not fit with the load's result type,
2143 // it must be truncated. We can't delay until Ideal call since
2144 // a singleton Value is needed for split_thru_phi optimization.
2145 int con = value->get_int();
2146 return TypeInt::make((con << 24) >> 24);
2147 }
2148 return LoadNode::Value(phase);
2149 }
2150
2151 //--------------------------LoadUBNode::Ideal-------------------------------------
2152 //
2153 // If the previous store is to the same address as this load,
2154 // and the value stored was larger than a byte, replace this load
2155 // with the value stored truncated to a byte. If no truncation is
2156 // needed, the replacement is done in LoadNode::Identity().
2157 //
2158 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2159 Node* mem = in(MemNode::Memory);
2160 Node* value = can_see_stored_value(mem, phase);
2161 if (value != NULL) {
2162 Node* narrow = Compile::narrow_value(T_BOOLEAN, value, _type, phase, false);
2163 if (narrow != value) {
2164 return narrow;
2165 }
2166 }
2167 // Identity call will handle the case where truncation is not needed.
2168 return LoadNode::Ideal(phase, can_reshape);
2169 }
2170
2171 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2172 Node* mem = in(MemNode::Memory);
2173 Node* value = can_see_stored_value(mem,phase);
2174 if (value != NULL && value->is_Con() &&
2175 !value->bottom_type()->higher_equal(_type)) {
2176 // If the input to the store does not fit with the load's result type,
2177 // it must be truncated. We can't delay until Ideal call since
2178 // a singleton Value is needed for split_thru_phi optimization.
2179 int con = value->get_int();
2180 return TypeInt::make(con & 0xFF);
2181 }
2182 return LoadNode::Value(phase);
2183 }
2184
2185 //--------------------------LoadUSNode::Ideal-------------------------------------
2186 //
2187 // If the previous store is to the same address as this load,
2188 // and the value stored was larger than a char, replace this load
2189 // with the value stored truncated to a char. If no truncation is
2190 // needed, the replacement is done in LoadNode::Identity().
2191 //
2192 Node* LoadUSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2193 Node* mem = in(MemNode::Memory);
2194 Node* value = can_see_stored_value(mem,phase);
2195 if (value != NULL) {
2196 Node* narrow = Compile::narrow_value(T_CHAR, value, _type, phase, false);
2197 if (narrow != value) {
2198 return narrow;
2199 }
2200 }
2201 // Identity call will handle the case where truncation is not needed.
2202 return LoadNode::Ideal(phase, can_reshape);
2203 }
2204
2205 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2206 Node* mem = in(MemNode::Memory);
2207 Node* value = can_see_stored_value(mem,phase);
2208 if (value != NULL && value->is_Con() &&
2209 !value->bottom_type()->higher_equal(_type)) {
2210 // If the input to the store does not fit with the load's result type,
2211 // it must be truncated. We can't delay until Ideal call since
2212 // a singleton Value is needed for split_thru_phi optimization.
2213 int con = value->get_int();
2214 return TypeInt::make(con & 0xFFFF);
2215 }
2216 return LoadNode::Value(phase);
2217 }
2218
2219 //--------------------------LoadSNode::Ideal--------------------------------------
2220 //
2221 // If the previous store is to the same address as this load,
2222 // and the value stored was larger than a short, replace this load
2223 // with the value stored truncated to a short. If no truncation is
2224 // needed, the replacement is done in LoadNode::Identity().
2225 //
2226 Node* LoadSNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2227 Node* mem = in(MemNode::Memory);
2228 Node* value = can_see_stored_value(mem,phase);
2229 if (value != NULL) {
2230 Node* narrow = Compile::narrow_value(T_SHORT, value, _type, phase, false);
2231 if (narrow != value) {
2232 return narrow;
2233 }
2234 }
2235 // Identity call will handle the case where truncation is not needed.
2236 return LoadNode::Ideal(phase, can_reshape);
2237 }
2238
2239 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2240 Node* mem = in(MemNode::Memory);
2241 Node* value = can_see_stored_value(mem,phase);
2242 if (value != NULL && value->is_Con() &&
2243 !value->bottom_type()->higher_equal(_type)) {
2244 // If the input to the store does not fit with the load's result type,
2245 // it must be truncated. We can't delay until Ideal call since
2246 // a singleton Value is needed for split_thru_phi optimization.
2247 int con = value->get_int();
2248 return TypeInt::make((con << 16) >> 16);
2249 }
2250 return LoadNode::Value(phase);
2251 }
2252
2253 //=============================================================================
2254 //----------------------------LoadKlassNode::make------------------------------
2255 // Polymorphic factory method:
2256 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2257 // sanity check the alias category against the created node type
2258 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2259 assert(adr_type != NULL, "expecting TypeKlassPtr");
2260 #ifdef _LP64
2261 if (adr_type->is_ptr_to_narrowklass()) {
2262 assert(UseCompressedClassPointers, "no compressed klasses");
2263 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2264 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2265 }
2266 #endif
2267 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2268 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2269 }
2270
2271 //------------------------------Value------------------------------------------
2272 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2273 return klass_value_common(phase);
2274 }
2275
2276 // In most cases, LoadKlassNode does not have the control input set. If the control
2277 // input is set, it must not be removed (by LoadNode::Ideal()).
2278 bool LoadKlassNode::can_remove_control() const {
2279 return false;
2280 }
2281
2282 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2283 // Either input is TOP ==> the result is TOP
2284 const Type *t1 = phase->type( in(MemNode::Memory) );
2285 if (t1 == Type::TOP) return Type::TOP;
2286 Node *adr = in(MemNode::Address);
2287 const Type *t2 = phase->type( adr );
2288 if (t2 == Type::TOP) return Type::TOP;
2289 const TypePtr *tp = t2->is_ptr();
2290 if (TypePtr::above_centerline(tp->ptr()) ||
2291 tp->ptr() == TypePtr::Null) return Type::TOP;
2292
2293 // Return a more precise klass, if possible
2294 const TypeInstPtr *tinst = tp->isa_instptr();
2295 if (tinst != NULL) {
2296 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2297 int offset = tinst->offset();
2298 if (ik == phase->C->env()->Class_klass()
2299 && (offset == java_lang_Class::klass_offset() ||
2300 offset == java_lang_Class::array_klass_offset())) {
2301 // We are loading a special hidden field from a Class mirror object,
2302 // the field which points to the VM's Klass metaobject.
2303 ciType* t = tinst->java_mirror_type();
2304 // java_mirror_type returns non-null for compile-time Class constants.
2305 if (t != NULL) {
2306 // constant oop => constant klass
2307 if (offset == java_lang_Class::array_klass_offset()) {
2308 if (t->is_void()) {
2309 // We cannot create a void array. Since void is a primitive type return null
2310 // klass. Users of this result need to do a null check on the returned klass.
2311 return TypePtr::NULL_PTR;
2312 }
2313 return TypeKlassPtr::make(ciArrayKlass::make(t));
2314 }
2315 if (!t->is_klass()) {
2316 // a primitive Class (e.g., int.class) has NULL for a klass field
2317 return TypePtr::NULL_PTR;
2318 }
2319 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2320 return TypeKlassPtr::make(t->as_klass());
2321 }
2322 // non-constant mirror, so we can't tell what's going on
2323 }
2324 if( !ik->is_loaded() )
2325 return _type; // Bail out if not loaded
2326 if (offset == oopDesc::klass_offset_in_bytes()) {
2327 if (tinst->klass_is_exact()) {
2328 return TypeKlassPtr::make(ik);
2329 }
2330 // See if we can become precise: no subklasses and no interface
2331 // (Note: We need to support verified interfaces.)
2332 if (!ik->is_interface() && !ik->has_subklass()) {
2333 // Add a dependence; if any subclass added we need to recompile
2334 if (!ik->is_final()) {
2335 // %%% should use stronger assert_unique_concrete_subtype instead
2336 phase->C->dependencies()->assert_leaf_type(ik);
2337 }
2338 // Return precise klass
2339 return TypeKlassPtr::make(ik);
2340 }
2341
2342 // Return root of possible klass
2343 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2344 }
2345 }
2346
2347 // Check for loading klass from an array
2348 const TypeAryPtr *tary = tp->isa_aryptr();
2349 if( tary != NULL ) {
2350 ciKlass *tary_klass = tary->klass();
2351 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2352 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2353 if (tary->klass_is_exact()) {
2354 return TypeKlassPtr::make(tary_klass);
2355 }
2356 ciArrayKlass *ak = tary->klass()->as_array_klass();
2357 // If the klass is an object array, we defer the question to the
2358 // array component klass.
2359 if( ak->is_obj_array_klass() ) {
2360 assert( ak->is_loaded(), "" );
2361 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2362 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2363 ciInstanceKlass* ik = base_k->as_instance_klass();
2364 // See if we can become precise: no subklasses and no interface
2365 if (!ik->is_interface() && !ik->has_subklass()) {
2366 // Add a dependence; if any subclass added we need to recompile
2367 if (!ik->is_final()) {
2368 phase->C->dependencies()->assert_leaf_type(ik);
2369 }
2370 // Return precise array klass
2371 return TypeKlassPtr::make(ak);
2372 }
2373 }
2374 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2375 } else { // Found a type-array?
2376 assert( ak->is_type_array_klass(), "" );
2377 return TypeKlassPtr::make(ak); // These are always precise
2378 }
2379 }
2380 }
2381
2382 // Check for loading klass from an array klass
2383 const TypeKlassPtr *tkls = tp->isa_klassptr();
2384 if (tkls != NULL && !StressReflectiveCode) {
2385 ciKlass* klass = tkls->klass();
2386 if( !klass->is_loaded() )
2387 return _type; // Bail out if not loaded
2388 if( klass->is_obj_array_klass() &&
2389 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2390 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2391 // // Always returning precise element type is incorrect,
2392 // // e.g., element type could be object and array may contain strings
2393 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2394
2395 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2396 // according to the element type's subclassing.
2397 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2398 }
2399 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2400 tkls->offset() == in_bytes(Klass::super_offset())) {
2401 ciKlass* sup = klass->as_instance_klass()->super();
2402 // The field is Klass::_super. Return its (constant) value.
2403 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2404 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2405 }
2406 }
2407
2408 // Bailout case
2409 return LoadNode::Value(phase);
2410 }
2411
2412 //------------------------------Identity---------------------------------------
2413 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2414 // Also feed through the klass in Allocate(...klass...)._klass.
2415 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2416 return klass_identity_common(phase);
2417 }
2418
2419 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2420 Node* x = LoadNode::Identity(phase);
2421 if (x != this) return x;
2422
2423 // Take apart the address into an oop and and offset.
2424 // Return 'this' if we cannot.
2425 Node* adr = in(MemNode::Address);
2426 intptr_t offset = 0;
2427 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2428 if (base == NULL) return this;
2429 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2430 if (toop == NULL) return this;
2431
2432 // Step over potential GC barrier for OopHandle resolve
2433 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2434 if (bs->is_gc_barrier_node(base)) {
2435 base = bs->step_over_gc_barrier(base);
2436 }
2437
2438 // We can fetch the klass directly through an AllocateNode.
2439 // This works even if the klass is not constant (clone or newArray).
2440 if (offset == oopDesc::klass_offset_in_bytes()) {
2441 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2442 if (allocated_klass != NULL) {
2443 return allocated_klass;
2444 }
2445 }
2446
2447 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2448 // See inline_native_Class_query for occurrences of these patterns.
2449 // Java Example: x.getClass().isAssignableFrom(y)
2450 //
2451 // This improves reflective code, often making the Class
2452 // mirror go completely dead. (Current exception: Class
2453 // mirrors may appear in debug info, but we could clean them out by
2454 // introducing a new debug info operator for Klass.java_mirror).
2455
2456 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2457 && offset == java_lang_Class::klass_offset()) {
2458 if (base->is_Load()) {
2459 Node* base2 = base->in(MemNode::Address);
2460 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2461 Node* adr2 = base2->in(MemNode::Address);
2462 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2463 if (tkls != NULL && !tkls->empty()
2464 && (tkls->klass()->is_instance_klass() ||
2465 tkls->klass()->is_array_klass())
2466 && adr2->is_AddP()
2467 ) {
2468 int mirror_field = in_bytes(Klass::java_mirror_offset());
2469 if (tkls->offset() == mirror_field) {
2470 return adr2->in(AddPNode::Base);
2471 }
2472 }
2473 }
2474 }
2475 }
2476
2477 return this;
2478 }
2479
2480
2481 //------------------------------Value------------------------------------------
2482 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2483 const Type *t = klass_value_common(phase);
2484 if (t == Type::TOP)
2485 return t;
2486
2487 return t->make_narrowklass();
2488 }
2489
2490 //------------------------------Identity---------------------------------------
2491 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2492 // Also feed through the klass in Allocate(...klass...)._klass.
2493 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2494 Node *x = klass_identity_common(phase);
2495
2496 const Type *t = phase->type( x );
2497 if( t == Type::TOP ) return x;
2498 if( t->isa_narrowklass()) return x;
2499 assert (!t->isa_narrowoop(), "no narrow oop here");
2500
2501 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2502 }
2503
2504 //------------------------------Value-----------------------------------------
2505 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2506 // Either input is TOP ==> the result is TOP
2507 const Type *t1 = phase->type( in(MemNode::Memory) );
2508 if( t1 == Type::TOP ) return Type::TOP;
2509 Node *adr = in(MemNode::Address);
2510 const Type *t2 = phase->type( adr );
2511 if( t2 == Type::TOP ) return Type::TOP;
2512 const TypePtr *tp = t2->is_ptr();
2513 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2514 const TypeAryPtr *tap = tp->isa_aryptr();
2515 if( !tap ) return _type;
2516 return tap->size();
2517 }
2518
2519 //-------------------------------Ideal---------------------------------------
2520 // Feed through the length in AllocateArray(...length...)._length.
2521 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2522 Node* p = MemNode::Ideal_common(phase, can_reshape);
2523 if (p) return (p == NodeSentinel) ? NULL : p;
2524
2525 // Take apart the address into an oop and and offset.
2526 // Return 'this' if we cannot.
2527 Node* adr = in(MemNode::Address);
2528 intptr_t offset = 0;
2529 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2530 if (base == NULL) return NULL;
2531 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2532 if (tary == NULL) return NULL;
2533
2534 // We can fetch the length directly through an AllocateArrayNode.
2535 // This works even if the length is not constant (clone or newArray).
2536 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2537 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2538 if (alloc != NULL) {
2539 Node* allocated_length = alloc->Ideal_length();
2540 Node* len = alloc->make_ideal_length(tary, phase);
2541 if (allocated_length != len) {
2542 // New CastII improves on this.
2543 return len;
2544 }
2545 }
2546 }
2547
2548 return NULL;
2549 }
2550
2551 //------------------------------Identity---------------------------------------
2552 // Feed through the length in AllocateArray(...length...)._length.
2553 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2554 Node* x = LoadINode::Identity(phase);
2555 if (x != this) return x;
2556
2557 // Take apart the address into an oop and and offset.
2558 // Return 'this' if we cannot.
2559 Node* adr = in(MemNode::Address);
2560 intptr_t offset = 0;
2561 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2562 if (base == NULL) return this;
2563 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2564 if (tary == NULL) return this;
2565
2566 // We can fetch the length directly through an AllocateArrayNode.
2567 // This works even if the length is not constant (clone or newArray).
2568 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2569 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2570 if (alloc != NULL) {
2571 Node* allocated_length = alloc->Ideal_length();
2572 // Do not allow make_ideal_length to allocate a CastII node.
2573 Node* len = alloc->make_ideal_length(tary, phase, false);
2574 if (allocated_length == len) {
2575 // Return allocated_length only if it would not be improved by a CastII.
2576 return allocated_length;
2577 }
2578 }
2579 }
2580
2581 return this;
2582
2583 }
2584
2585 //=============================================================================
2586 //---------------------------StoreNode::make-----------------------------------
2587 // Polymorphic factory method:
2588 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo, bool require_atomic_access) {
2589 assert((mo == unordered || mo == release), "unexpected");
2590 Compile* C = gvn.C;
2591 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2592 ctl != NULL, "raw memory operations should have control edge");
2593
2594 switch (bt) {
2595 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2596 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2597 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2598 case T_CHAR:
2599 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2600 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2601 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2602 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic_access);
2603 case T_METADATA:
2604 case T_ADDRESS:
2605 case T_OBJECT:
2606 #ifdef _LP64
2607 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2608 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2609 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2610 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2611 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2612 adr->bottom_type()->isa_rawptr())) {
2613 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2614 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2615 }
2616 #endif
2617 {
2618 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2619 }
2620 default:
2621 ShouldNotReachHere();
2622 return (StoreNode*)NULL;
2623 }
2624 }
2625
2626 //--------------------------bottom_type----------------------------------------
2627 const Type *StoreNode::bottom_type() const {
2628 return Type::MEMORY;
2629 }
2630
2631 //------------------------------hash-------------------------------------------
2632 uint StoreNode::hash() const {
2633 // unroll addition of interesting fields
2634 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2635
2636 // Since they are not commoned, do not hash them:
2637 return NO_HASH;
2638 }
2639
2640 //------------------------------Ideal------------------------------------------
2641 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2642 // When a store immediately follows a relevant allocation/initialization,
2643 // try to capture it into the initialization, or hoist it above.
2644 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2645 Node* p = MemNode::Ideal_common(phase, can_reshape);
2646 if (p) return (p == NodeSentinel) ? NULL : p;
2647
2648 Node* mem = in(MemNode::Memory);
2649 Node* address = in(MemNode::Address);
2650 Node* value = in(MemNode::ValueIn);
2651 // Back-to-back stores to same address? Fold em up. Generally
2652 // unsafe if I have intervening uses... Also disallowed for StoreCM
2653 // since they must follow each StoreP operation. Redundant StoreCMs
2654 // are eliminated just before matching in final_graph_reshape.
2655 {
2656 Node* st = mem;
2657 // If Store 'st' has more than one use, we cannot fold 'st' away.
2658 // For example, 'st' might be the final state at a conditional
2659 // return. Or, 'st' might be used by some node which is live at
2660 // the same time 'st' is live, which might be unschedulable. So,
2661 // require exactly ONE user until such time as we clone 'mem' for
2662 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2663 // true).
2664 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2665 // Looking at a dead closed cycle of memory?
2666 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2667 assert(Opcode() == st->Opcode() ||
2668 st->Opcode() == Op_StoreVector ||
2669 Opcode() == Op_StoreVector ||
2670 st->Opcode() == Op_StoreVectorScatter ||
2671 Opcode() == Op_StoreVectorScatter ||
2672 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2673 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2674 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2675 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2676 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2677
2678 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2679 st->as_Store()->memory_size() <= this->memory_size()) {
2680 Node* use = st->raw_out(0);
2681 if (phase->is_IterGVN()) {
2682 phase->is_IterGVN()->rehash_node_delayed(use);
2683 }
2684 // It's OK to do this in the parser, since DU info is always accurate,
2685 // and the parser always refers to nodes via SafePointNode maps.
2686 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase);
2687 return this;
2688 }
2689 st = st->in(MemNode::Memory);
2690 }
2691 }
2692
2693
2694 // Capture an unaliased, unconditional, simple store into an initializer.
2695 // Or, if it is independent of the allocation, hoist it above the allocation.
2696 if (ReduceFieldZeroing && /*can_reshape &&*/
2697 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2698 InitializeNode* init = mem->in(0)->as_Initialize();
2699 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2700 if (offset > 0) {
2701 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2702 // If the InitializeNode captured me, it made a raw copy of me,
2703 // and I need to disappear.
2704 if (moved != NULL) {
2705 // %%% hack to ensure that Ideal returns a new node:
2706 mem = MergeMemNode::make(mem);
2707 return mem; // fold me away
2708 }
2709 }
2710 }
2711
2712 // Fold reinterpret cast into memory operation:
2713 // StoreX mem (MoveY2X v) => StoreY mem v
2714 if (value->is_Move()) {
2715 const Type* vt = value->in(1)->bottom_type();
2716 if (has_reinterpret_variant(vt)) {
2717 if (phase->C->post_loop_opts_phase()) {
2718 return convert_to_reinterpret_store(*phase, value->in(1), vt);
2719 } else {
2720 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
2721 }
2722 }
2723 }
2724
2725 return NULL; // No further progress
2726 }
2727
2728 //------------------------------Value-----------------------------------------
2729 const Type* StoreNode::Value(PhaseGVN* phase) const {
2730 // Either input is TOP ==> the result is TOP
2731 const Type *t1 = phase->type( in(MemNode::Memory) );
2732 if( t1 == Type::TOP ) return Type::TOP;
2733 const Type *t2 = phase->type( in(MemNode::Address) );
2734 if( t2 == Type::TOP ) return Type::TOP;
2735 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2736 if( t3 == Type::TOP ) return Type::TOP;
2737 return Type::MEMORY;
2738 }
2739
2740 //------------------------------Identity---------------------------------------
2741 // Remove redundant stores:
2742 // Store(m, p, Load(m, p)) changes to m.
2743 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2744 Node* StoreNode::Identity(PhaseGVN* phase) {
2745 Node* mem = in(MemNode::Memory);
2746 Node* adr = in(MemNode::Address);
2747 Node* val = in(MemNode::ValueIn);
2748
2749 Node* result = this;
2750
2751 // Load then Store? Then the Store is useless
2752 if (val->is_Load() &&
2753 val->in(MemNode::Address)->eqv_uncast(adr) &&
2754 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2755 val->as_Load()->store_Opcode() == Opcode()) {
2756 result = mem;
2757 }
2758
2759 // Two stores in a row of the same value?
2760 if (result == this &&
2761 mem->is_Store() &&
2762 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2763 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2764 mem->Opcode() == Opcode()) {
2765 result = mem;
2766 }
2767
2768 // Store of zero anywhere into a freshly-allocated object?
2769 // Then the store is useless.
2770 // (It must already have been captured by the InitializeNode.)
2771 if (result == this &&
2772 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2773 // a newly allocated object is already all-zeroes everywhere
2774 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2775 result = mem;
2776 }
2777
2778 if (result == this) {
2779 // the store may also apply to zero-bits in an earlier object
2780 Node* prev_mem = find_previous_store(phase);
2781 // Steps (a), (b): Walk past independent stores to find an exact match.
2782 if (prev_mem != NULL) {
2783 Node* prev_val = can_see_stored_value(prev_mem, phase);
2784 if (prev_val != NULL && prev_val == val) {
2785 // prev_val and val might differ by a cast; it would be good
2786 // to keep the more informative of the two.
2787 result = mem;
2788 }
2789 }
2790 }
2791 }
2792
2793 PhaseIterGVN* igvn = phase->is_IterGVN();
2794 if (result != this && igvn != NULL) {
2795 MemBarNode* trailing = trailing_membar();
2796 if (trailing != NULL) {
2797 #ifdef ASSERT
2798 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2799 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects");
2800 #endif
2801 trailing->remove(igvn);
2802 }
2803 }
2804
2805 return result;
2806 }
2807
2808 //------------------------------match_edge-------------------------------------
2809 // Do we Match on this edge index or not? Match only memory & value
2810 uint StoreNode::match_edge(uint idx) const {
2811 return idx == MemNode::Address || idx == MemNode::ValueIn;
2812 }
2813
2814 //------------------------------cmp--------------------------------------------
2815 // Do not common stores up together. They generally have to be split
2816 // back up anyways, so do not bother.
2817 bool StoreNode::cmp( const Node &n ) const {
2818 return (&n == this); // Always fail except on self
2819 }
2820
2821 //------------------------------Ideal_masked_input-----------------------------
2822 // Check for a useless mask before a partial-word store
2823 // (StoreB ... (AndI valIn conIa) )
2824 // If (conIa & mask == mask) this simplifies to
2825 // (StoreB ... (valIn) )
2826 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2827 Node *val = in(MemNode::ValueIn);
2828 if( val->Opcode() == Op_AndI ) {
2829 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2830 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2831 set_req_X(MemNode::ValueIn, val->in(1), phase);
2832 return this;
2833 }
2834 }
2835 return NULL;
2836 }
2837
2838
2839 //------------------------------Ideal_sign_extended_input----------------------
2840 // Check for useless sign-extension before a partial-word store
2841 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2842 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2843 // (StoreB ... (valIn) )
2844 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2845 Node *val = in(MemNode::ValueIn);
2846 if( val->Opcode() == Op_RShiftI ) {
2847 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2848 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2849 Node *shl = val->in(1);
2850 if( shl->Opcode() == Op_LShiftI ) {
2851 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2852 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2853 set_req_X(MemNode::ValueIn, shl->in(1), phase);
2854 return this;
2855 }
2856 }
2857 }
2858 }
2859 return NULL;
2860 }
2861
2862 //------------------------------value_never_loaded-----------------------------------
2863 // Determine whether there are any possible loads of the value stored.
2864 // For simplicity, we actually check if there are any loads from the
2865 // address stored to, not just for loads of the value stored by this node.
2866 //
2867 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2868 Node *adr = in(Address);
2869 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2870 if (adr_oop == NULL)
2871 return false;
2872 if (!adr_oop->is_known_instance_field())
2873 return false; // if not a distinct instance, there may be aliases of the address
2874 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2875 Node *use = adr->fast_out(i);
2876 if (use->is_Load() || use->is_LoadStore()) {
2877 return false;
2878 }
2879 }
2880 return true;
2881 }
2882
2883 MemBarNode* StoreNode::trailing_membar() const {
2884 if (is_release()) {
2885 MemBarNode* trailing_mb = NULL;
2886 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2887 Node* u = fast_out(i);
2888 if (u->is_MemBar()) {
2889 if (u->as_MemBar()->trailing_store()) {
2890 assert(u->Opcode() == Op_MemBarVolatile, "");
2891 assert(trailing_mb == NULL, "only one");
2892 trailing_mb = u->as_MemBar();
2893 #ifdef ASSERT
2894 Node* leading = u->as_MemBar()->leading_membar();
2895 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2896 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2897 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2898 #endif
2899 } else {
2900 assert(u->as_MemBar()->standalone(), "");
2901 }
2902 }
2903 }
2904 return trailing_mb;
2905 }
2906 return NULL;
2907 }
2908
2909
2910 //=============================================================================
2911 //------------------------------Ideal------------------------------------------
2912 // If the store is from an AND mask that leaves the low bits untouched, then
2913 // we can skip the AND operation. If the store is from a sign-extension
2914 // (a left shift, then right shift) we can skip both.
2915 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2916 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2917 if( progress != NULL ) return progress;
2918
2919 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2920 if( progress != NULL ) return progress;
2921
2922 // Finally check the default case
2923 return StoreNode::Ideal(phase, can_reshape);
2924 }
2925
2926 //=============================================================================
2927 //------------------------------Ideal------------------------------------------
2928 // If the store is from an AND mask that leaves the low bits untouched, then
2929 // we can skip the AND operation
2930 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2931 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2932 if( progress != NULL ) return progress;
2933
2934 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2935 if( progress != NULL ) return progress;
2936
2937 // Finally check the default case
2938 return StoreNode::Ideal(phase, can_reshape);
2939 }
2940
2941 //=============================================================================
2942 //------------------------------Identity---------------------------------------
2943 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2944 // No need to card mark when storing a null ptr
2945 Node* my_store = in(MemNode::OopStore);
2946 if (my_store->is_Store()) {
2947 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2948 if( t1 == TypePtr::NULL_PTR ) {
2949 return in(MemNode::Memory);
2950 }
2951 }
2952 return this;
2953 }
2954
2955 //=============================================================================
2956 //------------------------------Ideal---------------------------------------
2957 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2958 Node* progress = StoreNode::Ideal(phase, can_reshape);
2959 if (progress != NULL) return progress;
2960
2961 Node* my_store = in(MemNode::OopStore);
2962 if (my_store->is_MergeMem()) {
2963 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2964 set_req_X(MemNode::OopStore, mem, phase);
2965 return this;
2966 }
2967
2968 return NULL;
2969 }
2970
2971 //------------------------------Value-----------------------------------------
2972 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2973 // Either input is TOP ==> the result is TOP (checked in StoreNode::Value).
2974 // If extra input is TOP ==> the result is TOP
2975 const Type* t = phase->type(in(MemNode::OopStore));
2976 if (t == Type::TOP) {
2977 return Type::TOP;
2978 }
2979 return StoreNode::Value(phase);
2980 }
2981
2982
2983 //=============================================================================
2984 //----------------------------------SCMemProjNode------------------------------
2985 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2986 {
2987 if (in(0) == NULL || phase->type(in(0)) == Type::TOP) {
2988 return Type::TOP;
2989 }
2990 return bottom_type();
2991 }
2992
2993 //=============================================================================
2994 //----------------------------------LoadStoreNode------------------------------
2995 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2996 : Node(required),
2997 _type(rt),
2998 _adr_type(at),
2999 _barrier_data(0)
3000 {
3001 init_req(MemNode::Control, c );
3002 init_req(MemNode::Memory , mem);
3003 init_req(MemNode::Address, adr);
3004 init_req(MemNode::ValueIn, val);
3005 init_class_id(Class_LoadStore);
3006 }
3007
3008 //------------------------------Value-----------------------------------------
3009 const Type* LoadStoreNode::Value(PhaseGVN* phase) const {
3010 // Either input is TOP ==> the result is TOP
3011 if (!in(MemNode::Control) || phase->type(in(MemNode::Control)) == Type::TOP) {
3012 return Type::TOP;
3013 }
3014 const Type* t = phase->type(in(MemNode::Memory));
3015 if (t == Type::TOP) {
3016 return Type::TOP;
3017 }
3018 t = phase->type(in(MemNode::Address));
3019 if (t == Type::TOP) {
3020 return Type::TOP;
3021 }
3022 t = phase->type(in(MemNode::ValueIn));
3023 if (t == Type::TOP) {
3024 return Type::TOP;
3025 }
3026 return bottom_type();
3027 }
3028
3029 uint LoadStoreNode::ideal_reg() const {
3030 return _type->ideal_reg();
3031 }
3032
3033 bool LoadStoreNode::result_not_used() const {
3034 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
3035 Node *x = fast_out(i);
3036 if (x->Opcode() == Op_SCMemProj) continue;
3037 return false;
3038 }
3039 return true;
3040 }
3041
3042 MemBarNode* LoadStoreNode::trailing_membar() const {
3043 MemBarNode* trailing = NULL;
3044 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3045 Node* u = fast_out(i);
3046 if (u->is_MemBar()) {
3047 if (u->as_MemBar()->trailing_load_store()) {
3048 assert(u->Opcode() == Op_MemBarAcquire, "");
3049 assert(trailing == NULL, "only one");
3050 trailing = u->as_MemBar();
3051 #ifdef ASSERT
3052 Node* leading = trailing->leading_membar();
3053 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3054 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
3055 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
3056 #endif
3057 } else {
3058 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
3059 }
3060 }
3061 }
3062
3063 return trailing;
3064 }
3065
3066 uint LoadStoreNode::size_of() const { return sizeof(*this); }
3067
3068 //=============================================================================
3069 //----------------------------------LoadStoreConditionalNode--------------------
3070 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
3071 init_req(ExpectedIn, ex );
3072 }
3073
3074 const Type* LoadStoreConditionalNode::Value(PhaseGVN* phase) const {
3075 // Either input is TOP ==> the result is TOP
3076 const Type* t = phase->type(in(ExpectedIn));
3077 if (t == Type::TOP) {
3078 return Type::TOP;
3079 }
3080 return LoadStoreNode::Value(phase);
3081 }
3082
3083 //=============================================================================
3084 //-------------------------------adr_type--------------------------------------
3085 const TypePtr* ClearArrayNode::adr_type() const {
3086 Node *adr = in(3);
3087 if (adr == NULL) return NULL; // node is dead
3088 return MemNode::calculate_adr_type(adr->bottom_type());
3089 }
3090
3091 //------------------------------match_edge-------------------------------------
3092 // Do we Match on this edge index or not? Do not match memory
3093 uint ClearArrayNode::match_edge(uint idx) const {
3094 return idx > 1;
3095 }
3096
3097 //------------------------------Identity---------------------------------------
3098 // Clearing a zero length array does nothing
3099 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
3100 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
3101 }
3102
3103 //------------------------------Idealize---------------------------------------
3104 // Clearing a short array is faster with stores
3105 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3106 // Already know this is a large node, do not try to ideal it
3107 if (_is_large) return NULL;
3108
3109 const int unit = BytesPerLong;
3110 const TypeX* t = phase->type(in(2))->isa_intptr_t();
3111 if (!t) return NULL;
3112 if (!t->is_con()) return NULL;
3113 intptr_t raw_count = t->get_con();
3114 intptr_t size = raw_count;
3115 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
3116 // Clearing nothing uses the Identity call.
3117 // Negative clears are possible on dead ClearArrays
3118 // (see jck test stmt114.stmt11402.val).
3119 if (size <= 0 || size % unit != 0) return NULL;
3120 intptr_t count = size / unit;
3121 // Length too long; communicate this to matchers and assemblers.
3122 // Assemblers are responsible to produce fast hardware clears for it.
3123 if (size > InitArrayShortSize) {
3124 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
3125 } else if (size > 2 && Matcher::match_rule_supported_vector(Op_ClearArray, 4, T_LONG)) {
3126 return NULL;
3127 }
3128 if (!IdealizeClearArrayNode) return NULL;
3129 Node *mem = in(1);
3130 if( phase->type(mem)==Type::TOP ) return NULL;
3131 Node *adr = in(3);
3132 const Type* at = phase->type(adr);
3133 if( at==Type::TOP ) return NULL;
3134 const TypePtr* atp = at->isa_ptr();
3135 // adjust atp to be the correct array element address type
3136 if (atp == NULL) atp = TypePtr::BOTTOM;
3137 else atp = atp->add_offset(Type::OffsetBot);
3138 // Get base for derived pointer purposes
3139 if( adr->Opcode() != Op_AddP ) Unimplemented();
3140 Node *base = adr->in(1);
3141
3142 Node *zero = phase->makecon(TypeLong::ZERO);
3143 Node *off = phase->MakeConX(BytesPerLong);
3144 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3145 count--;
3146 while( count-- ) {
3147 mem = phase->transform(mem);
3148 adr = phase->transform(new AddPNode(base,adr,off));
3149 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3150 }
3151 return mem;
3152 }
3153
3154 //----------------------------step_through----------------------------------
3155 // Return allocation input memory edge if it is different instance
3156 // or itself if it is the one we are looking for.
3157 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
3158 Node* n = *np;
3159 assert(n->is_ClearArray(), "sanity");
3160 intptr_t offset;
3161 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
3162 // This method is called only before Allocate nodes are expanded
3163 // during macro nodes expansion. Before that ClearArray nodes are
3164 // only generated in PhaseMacroExpand::generate_arraycopy() (before
3165 // Allocate nodes are expanded) which follows allocations.
3166 assert(alloc != NULL, "should have allocation");
3167 if (alloc->_idx == instance_id) {
3168 // Can not bypass initialization of the instance we are looking for.
3169 return false;
3170 }
3171 // Otherwise skip it.
3172 InitializeNode* init = alloc->initialization();
3173 if (init != NULL)
3174 *np = init->in(TypeFunc::Memory);
3175 else
3176 *np = alloc->in(TypeFunc::Memory);
3177 return true;
3178 }
3179
3180 //----------------------------clear_memory-------------------------------------
3181 // Generate code to initialize object storage to zero.
3182 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3183 intptr_t start_offset,
3184 Node* end_offset,
3185 PhaseGVN* phase) {
3186 intptr_t offset = start_offset;
3187
3188 int unit = BytesPerLong;
3189 if ((offset % unit) != 0) {
3190 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
3191 adr = phase->transform(adr);
3192 const TypePtr* atp = TypeRawPtr::BOTTOM;
3193 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3194 mem = phase->transform(mem);
3195 offset += BytesPerInt;
3196 }
3197 assert((offset % unit) == 0, "");
3198
3199 // Initialize the remaining stuff, if any, with a ClearArray.
3200 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
3201 }
3202
3203 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3204 Node* start_offset,
3205 Node* end_offset,
3206 PhaseGVN* phase) {
3207 if (start_offset == end_offset) {
3208 // nothing to do
3209 return mem;
3210 }
3211
3212 int unit = BytesPerLong;
3213 Node* zbase = start_offset;
3214 Node* zend = end_offset;
3215
3216 // Scale to the unit required by the CPU:
3217 if (!Matcher::init_array_count_is_in_bytes) {
3218 Node* shift = phase->intcon(exact_log2(unit));
3219 zbase = phase->transform(new URShiftXNode(zbase, shift) );
3220 zend = phase->transform(new URShiftXNode(zend, shift) );
3221 }
3222
3223 // Bulk clear double-words
3224 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
3225 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
3226 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
3227 return phase->transform(mem);
3228 }
3229
3230 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3231 intptr_t start_offset,
3232 intptr_t end_offset,
3233 PhaseGVN* phase) {
3234 if (start_offset == end_offset) {
3235 // nothing to do
3236 return mem;
3237 }
3238
3239 assert((end_offset % BytesPerInt) == 0, "odd end offset");
3240 intptr_t done_offset = end_offset;
3241 if ((done_offset % BytesPerLong) != 0) {
3242 done_offset -= BytesPerInt;
3243 }
3244 if (done_offset > start_offset) {
3245 mem = clear_memory(ctl, mem, dest,
3246 start_offset, phase->MakeConX(done_offset), phase);
3247 }
3248 if (done_offset < end_offset) { // emit the final 32-bit store
3249 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3250 adr = phase->transform(adr);
3251 const TypePtr* atp = TypeRawPtr::BOTTOM;
3252 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3253 mem = phase->transform(mem);
3254 done_offset += BytesPerInt;
3255 }
3256 assert(done_offset == end_offset, "");
3257 return mem;
3258 }
3259
3260 //=============================================================================
3261 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3262 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
3263 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3264 #ifdef ASSERT
3265 , _pair_idx(0)
3266 #endif
3267 {
3268 init_class_id(Class_MemBar);
3269 Node* top = C->top();
3270 init_req(TypeFunc::I_O,top);
3271 init_req(TypeFunc::FramePtr,top);
3272 init_req(TypeFunc::ReturnAdr,top);
3273 if (precedent != NULL)
3274 init_req(TypeFunc::Parms, precedent);
3275 }
3276
3277 //------------------------------cmp--------------------------------------------
3278 uint MemBarNode::hash() const { return NO_HASH; }
3279 bool MemBarNode::cmp( const Node &n ) const {
3280 return (&n == this); // Always fail except on self
3281 }
3282
3283 //------------------------------make-------------------------------------------
3284 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3285 switch (opcode) {
3286 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
3287 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
3288 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
3289 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
3290 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
3291 case Op_StoreStoreFence: return new StoreStoreFenceNode(C, atp, pn);
3292 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3293 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3294 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
3295 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
3296 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
3297 case Op_Initialize: return new InitializeNode(C, atp, pn);
3298 default: ShouldNotReachHere(); return NULL;
3299 }
3300 }
3301
3302 void MemBarNode::remove(PhaseIterGVN *igvn) {
3303 if (outcnt() != 2) {
3304 assert(Opcode() == Op_Initialize, "Only seen when there are no use of init memory");
3305 assert(outcnt() == 1, "Only control then");
3306 }
3307 if (trailing_store() || trailing_load_store()) {
3308 MemBarNode* leading = leading_membar();
3309 if (leading != NULL) {
3310 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3311 leading->remove(igvn);
3312 }
3313 }
3314 if (proj_out_or_null(TypeFunc::Memory) != NULL) {
3315 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3316 }
3317 if (proj_out_or_null(TypeFunc::Control) != NULL) {
3318 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3319 }
3320 }
3321
3322 //------------------------------Ideal------------------------------------------
3323 // Return a node which is more "ideal" than the current node. Strip out
3324 // control copies
3325 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3326 if (remove_dead_region(phase, can_reshape)) return this;
3327 // Don't bother trying to transform a dead node
3328 if (in(0) && in(0)->is_top()) {
3329 return NULL;
3330 }
3331
3332 bool progress = false;
3333 // Eliminate volatile MemBars for scalar replaced objects.
3334 if (can_reshape && req() == (Precedent+1)) {
3335 bool eliminate = false;
3336 int opc = Opcode();
3337 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3338 // Volatile field loads and stores.
3339 Node* my_mem = in(MemBarNode::Precedent);
3340 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3341 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3342 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3343 // replace this Precedent (decodeN) with the Load instead.
3344 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3345 Node* load_node = my_mem->in(1);
3346 set_req(MemBarNode::Precedent, load_node);
3347 phase->is_IterGVN()->_worklist.push(my_mem);
3348 my_mem = load_node;
3349 } else {
3350 assert(my_mem->unique_out() == this, "sanity");
3351 del_req(Precedent);
3352 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3353 my_mem = NULL;
3354 }
3355 progress = true;
3356 }
3357 if (my_mem != NULL && my_mem->is_Mem()) {
3358 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3359 // Check for scalar replaced object reference.
3360 if( t_oop != NULL && t_oop->is_known_instance_field() &&
3361 t_oop->offset() != Type::OffsetBot &&
3362 t_oop->offset() != Type::OffsetTop) {
3363 eliminate = true;
3364 }
3365 }
3366 } else if (opc == Op_MemBarRelease) {
3367 // Final field stores.
3368 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3369 if ((alloc != NULL) && alloc->is_Allocate() &&
3370 alloc->as_Allocate()->does_not_escape_thread()) {
3371 // The allocated object does not escape.
3372 eliminate = true;
3373 }
3374 }
3375 if (eliminate) {
3376 // Replace MemBar projections by its inputs.
3377 PhaseIterGVN* igvn = phase->is_IterGVN();
3378 remove(igvn);
3379 // Must return either the original node (now dead) or a new node
3380 // (Do not return a top here, since that would break the uniqueness of top.)
3381 return new ConINode(TypeInt::ZERO);
3382 }
3383 }
3384 return progress ? this : NULL;
3385 }
3386
3387 //------------------------------Value------------------------------------------
3388 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3389 if( !in(0) ) return Type::TOP;
3390 if( phase->type(in(0)) == Type::TOP )
3391 return Type::TOP;
3392 return TypeTuple::MEMBAR;
3393 }
3394
3395 //------------------------------match------------------------------------------
3396 // Construct projections for memory.
3397 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3398 switch (proj->_con) {
3399 case TypeFunc::Control:
3400 case TypeFunc::Memory:
3401 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3402 }
3403 ShouldNotReachHere();
3404 return NULL;
3405 }
3406
3407 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3408 trailing->_kind = TrailingStore;
3409 leading->_kind = LeadingStore;
3410 #ifdef ASSERT
3411 trailing->_pair_idx = leading->_idx;
3412 leading->_pair_idx = leading->_idx;
3413 #endif
3414 }
3415
3416 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3417 trailing->_kind = TrailingLoadStore;
3418 leading->_kind = LeadingLoadStore;
3419 #ifdef ASSERT
3420 trailing->_pair_idx = leading->_idx;
3421 leading->_pair_idx = leading->_idx;
3422 #endif
3423 }
3424
3425 MemBarNode* MemBarNode::trailing_membar() const {
3426 ResourceMark rm;
3427 Node* trailing = (Node*)this;
3428 VectorSet seen;
3429 Node_Stack multis(0);
3430 do {
3431 Node* c = trailing;
3432 uint i = 0;
3433 do {
3434 trailing = NULL;
3435 for (; i < c->outcnt(); i++) {
3436 Node* next = c->raw_out(i);
3437 if (next != c && next->is_CFG()) {
3438 if (c->is_MultiBranch()) {
3439 if (multis.node() == c) {
3440 multis.set_index(i+1);
3441 } else {
3442 multis.push(c, i+1);
3443 }
3444 }
3445 trailing = next;
3446 break;
3447 }
3448 }
3449 if (trailing != NULL && !seen.test_set(trailing->_idx)) {
3450 break;
3451 }
3452 while (multis.size() > 0) {
3453 c = multis.node();
3454 i = multis.index();
3455 if (i < c->req()) {
3456 break;
3457 }
3458 multis.pop();
3459 }
3460 } while (multis.size() > 0);
3461 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
3462
3463 MemBarNode* mb = trailing->as_MemBar();
3464 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3465 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3466 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3467 return mb;
3468 }
3469
3470 MemBarNode* MemBarNode::leading_membar() const {
3471 ResourceMark rm;
3472 VectorSet seen;
3473 Node_Stack regions(0);
3474 Node* leading = in(0);
3475 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3476 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) {
3477 leading = NULL;
3478 while (regions.size() > 0 && leading == NULL) {
3479 Node* r = regions.node();
3480 uint i = regions.index();
3481 if (i < r->req()) {
3482 leading = r->in(i);
3483 regions.set_index(i+1);
3484 } else {
3485 regions.pop();
3486 }
3487 }
3488 if (leading == NULL) {
3489 assert(regions.size() == 0, "all paths should have been tried");
3490 return NULL;
3491 }
3492 }
3493 if (leading->is_Region()) {
3494 regions.push(leading, 2);
3495 leading = leading->in(1);
3496 } else {
3497 leading = leading->in(0);
3498 }
3499 }
3500 #ifdef ASSERT
3501 Unique_Node_List wq;
3502 wq.push((Node*)this);
3503 uint found = 0;
3504 for (uint i = 0; i < wq.size(); i++) {
3505 Node* n = wq.at(i);
3506 if (n->is_Region()) {
3507 for (uint j = 1; j < n->req(); j++) {
3508 Node* in = n->in(j);
3509 if (in != NULL && !in->is_top()) {
3510 wq.push(in);
3511 }
3512 }
3513 } else {
3514 if (n->is_MemBar() && n->as_MemBar()->leading()) {
3515 assert(n == leading, "consistency check failed");
3516 found++;
3517 } else {
3518 Node* in = n->in(0);
3519 if (in != NULL && !in->is_top()) {
3520 wq.push(in);
3521 }
3522 }
3523 }
3524 }
3525 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed");
3526 #endif
3527 if (leading == NULL) {
3528 return NULL;
3529 }
3530 MemBarNode* mb = leading->as_MemBar();
3531 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3532 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3533 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3534 return mb;
3535 }
3536
3537
3538 //===========================InitializeNode====================================
3539 // SUMMARY:
3540 // This node acts as a memory barrier on raw memory, after some raw stores.
3541 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3542 // The Initialize can 'capture' suitably constrained stores as raw inits.
3543 // It can coalesce related raw stores into larger units (called 'tiles').
3544 // It can avoid zeroing new storage for memory units which have raw inits.
3545 // At macro-expansion, it is marked 'complete', and does not optimize further.
3546 //
3547 // EXAMPLE:
3548 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3549 // ctl = incoming control; mem* = incoming memory
3550 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3551 // First allocate uninitialized memory and fill in the header:
3552 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3553 // ctl := alloc.Control; mem* := alloc.Memory*
3554 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3555 // Then initialize to zero the non-header parts of the raw memory block:
3556 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3557 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3558 // After the initialize node executes, the object is ready for service:
3559 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3560 // Suppose its body is immediately initialized as {1,2}:
3561 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3562 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3563 // mem.SLICE(#short[*]) := store2
3564 //
3565 // DETAILS:
3566 // An InitializeNode collects and isolates object initialization after
3567 // an AllocateNode and before the next possible safepoint. As a
3568 // memory barrier (MemBarNode), it keeps critical stores from drifting
3569 // down past any safepoint or any publication of the allocation.
3570 // Before this barrier, a newly-allocated object may have uninitialized bits.
3571 // After this barrier, it may be treated as a real oop, and GC is allowed.
3572 //
3573 // The semantics of the InitializeNode include an implicit zeroing of
3574 // the new object from object header to the end of the object.
3575 // (The object header and end are determined by the AllocateNode.)
3576 //
3577 // Certain stores may be added as direct inputs to the InitializeNode.
3578 // These stores must update raw memory, and they must be to addresses
3579 // derived from the raw address produced by AllocateNode, and with
3580 // a constant offset. They must be ordered by increasing offset.
3581 // The first one is at in(RawStores), the last at in(req()-1).
3582 // Unlike most memory operations, they are not linked in a chain,
3583 // but are displayed in parallel as users of the rawmem output of
3584 // the allocation.
3585 //
3586 // (See comments in InitializeNode::capture_store, which continue
3587 // the example given above.)
3588 //
3589 // When the associated Allocate is macro-expanded, the InitializeNode
3590 // may be rewritten to optimize collected stores. A ClearArrayNode
3591 // may also be created at that point to represent any required zeroing.
3592 // The InitializeNode is then marked 'complete', prohibiting further
3593 // capturing of nearby memory operations.
3594 //
3595 // During macro-expansion, all captured initializations which store
3596 // constant values of 32 bits or smaller are coalesced (if advantageous)
3597 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3598 // initialized in fewer memory operations. Memory words which are
3599 // covered by neither tiles nor non-constant stores are pre-zeroed
3600 // by explicit stores of zero. (The code shape happens to do all
3601 // zeroing first, then all other stores, with both sequences occurring
3602 // in order of ascending offsets.)
3603 //
3604 // Alternatively, code may be inserted between an AllocateNode and its
3605 // InitializeNode, to perform arbitrary initialization of the new object.
3606 // E.g., the object copying intrinsics insert complex data transfers here.
3607 // The initialization must then be marked as 'complete' disable the
3608 // built-in zeroing semantics and the collection of initializing stores.
3609 //
3610 // While an InitializeNode is incomplete, reads from the memory state
3611 // produced by it are optimizable if they match the control edge and
3612 // new oop address associated with the allocation/initialization.
3613 // They return a stored value (if the offset matches) or else zero.
3614 // A write to the memory state, if it matches control and address,
3615 // and if it is to a constant offset, may be 'captured' by the
3616 // InitializeNode. It is cloned as a raw memory operation and rewired
3617 // inside the initialization, to the raw oop produced by the allocation.
3618 // Operations on addresses which are provably distinct (e.g., to
3619 // other AllocateNodes) are allowed to bypass the initialization.
3620 //
3621 // The effect of all this is to consolidate object initialization
3622 // (both arrays and non-arrays, both piecewise and bulk) into a
3623 // single location, where it can be optimized as a unit.
3624 //
3625 // Only stores with an offset less than TrackedInitializationLimit words
3626 // will be considered for capture by an InitializeNode. This puts a
3627 // reasonable limit on the complexity of optimized initializations.
3628
3629 //---------------------------InitializeNode------------------------------------
3630 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3631 : MemBarNode(C, adr_type, rawoop),
3632 _is_complete(Incomplete), _does_not_escape(false)
3633 {
3634 init_class_id(Class_Initialize);
3635
3636 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3637 assert(in(RawAddress) == rawoop, "proper init");
3638 // Note: allocation() can be NULL, for secondary initialization barriers
3639 }
3640
3641 // Since this node is not matched, it will be processed by the
3642 // register allocator. Declare that there are no constraints
3643 // on the allocation of the RawAddress edge.
3644 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3645 // This edge should be set to top, by the set_complete. But be conservative.
3646 if (idx == InitializeNode::RawAddress)
3647 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3648 return RegMask::Empty;
3649 }
3650
3651 Node* InitializeNode::memory(uint alias_idx) {
3652 Node* mem = in(Memory);
3653 if (mem->is_MergeMem()) {
3654 return mem->as_MergeMem()->memory_at(alias_idx);
3655 } else {
3656 // incoming raw memory is not split
3657 return mem;
3658 }
3659 }
3660
3661 bool InitializeNode::is_non_zero() {
3662 if (is_complete()) return false;
3663 remove_extra_zeroes();
3664 return (req() > RawStores);
3665 }
3666
3667 void InitializeNode::set_complete(PhaseGVN* phase) {
3668 assert(!is_complete(), "caller responsibility");
3669 _is_complete = Complete;
3670
3671 // After this node is complete, it contains a bunch of
3672 // raw-memory initializations. There is no need for
3673 // it to have anything to do with non-raw memory effects.
3674 // Therefore, tell all non-raw users to re-optimize themselves,
3675 // after skipping the memory effects of this initialization.
3676 PhaseIterGVN* igvn = phase->is_IterGVN();
3677 if (igvn) igvn->add_users_to_worklist(this);
3678 }
3679
3680 // convenience function
3681 // return false if the init contains any stores already
3682 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3683 InitializeNode* init = initialization();
3684 if (init == NULL || init->is_complete()) return false;
3685 init->remove_extra_zeroes();
3686 // for now, if this allocation has already collected any inits, bail:
3687 if (init->is_non_zero()) return false;
3688 init->set_complete(phase);
3689 return true;
3690 }
3691
3692 void InitializeNode::remove_extra_zeroes() {
3693 if (req() == RawStores) return;
3694 Node* zmem = zero_memory();
3695 uint fill = RawStores;
3696 for (uint i = fill; i < req(); i++) {
3697 Node* n = in(i);
3698 if (n->is_top() || n == zmem) continue; // skip
3699 if (fill < i) set_req(fill, n); // compact
3700 ++fill;
3701 }
3702 // delete any empty spaces created:
3703 while (fill < req()) {
3704 del_req(fill);
3705 }
3706 }
3707
3708 // Helper for remembering which stores go with which offsets.
3709 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3710 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3711 intptr_t offset = -1;
3712 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3713 phase, offset);
3714 if (base == NULL) return -1; // something is dead,
3715 if (offset < 0) return -1; // dead, dead
3716 return offset;
3717 }
3718
3719 // Helper for proving that an initialization expression is
3720 // "simple enough" to be folded into an object initialization.
3721 // Attempts to prove that a store's initial value 'n' can be captured
3722 // within the initialization without creating a vicious cycle, such as:
3723 // { Foo p = new Foo(); p.next = p; }
3724 // True for constants and parameters and small combinations thereof.
3725 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
3726 ResourceMark rm;
3727 Unique_Node_List worklist;
3728 worklist.push(value);
3729
3730 uint complexity_limit = 20;
3731 for (uint j = 0; j < worklist.size(); j++) {
3732 if (j >= complexity_limit) {
3733 return false; // Bail out if processed too many nodes
3734 }
3735
3736 Node* n = worklist.at(j);
3737 if (n == NULL) continue; // (can this really happen?)
3738 if (n->is_Proj()) n = n->in(0);
3739 if (n == this) return false; // found a cycle
3740 if (n->is_Con()) continue;
3741 if (n->is_Start()) continue; // params, etc., are OK
3742 if (n->is_Root()) continue; // even better
3743
3744 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
3745 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
3746 continue;
3747 }
3748
3749 Node* ctl = n->in(0);
3750 if (ctl != NULL && !ctl->is_top()) {
3751 if (ctl->is_Proj()) ctl = ctl->in(0);
3752 if (ctl == this) return false;
3753
3754 // If we already know that the enclosing memory op is pinned right after
3755 // the init, then any control flow that the store has picked up
3756 // must have preceded the init, or else be equal to the init.
3757 // Even after loop optimizations (which might change control edges)
3758 // a store is never pinned *before* the availability of its inputs.
3759 if (!MemNode::all_controls_dominate(n, this))
3760 return false; // failed to prove a good control
3761 }
3762
3763 // Check data edges for possible dependencies on 'this'.
3764 for (uint i = 1; i < n->req(); i++) {
3765 Node* m = n->in(i);
3766 if (m == NULL || m == n || m->is_top()) continue;
3767
3768 // Only process data inputs once
3769 worklist.push(m);
3770 }
3771 }
3772
3773 return true;
3774 }
3775
3776 // Here are all the checks a Store must pass before it can be moved into
3777 // an initialization. Returns zero if a check fails.
3778 // On success, returns the (constant) offset to which the store applies,
3779 // within the initialized memory.
3780 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
3781 const int FAIL = 0;
3782 if (st->req() != MemNode::ValueIn + 1)
3783 return FAIL; // an inscrutable StoreNode (card mark?)
3784 Node* ctl = st->in(MemNode::Control);
3785 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3786 return FAIL; // must be unconditional after the initialization
3787 Node* mem = st->in(MemNode::Memory);
3788 if (!(mem->is_Proj() && mem->in(0) == this))
3789 return FAIL; // must not be preceded by other stores
3790 Node* adr = st->in(MemNode::Address);
3791 intptr_t offset;
3792 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3793 if (alloc == NULL)
3794 return FAIL; // inscrutable address
3795 if (alloc != allocation())
3796 return FAIL; // wrong allocation! (store needs to float up)
3797 int size_in_bytes = st->memory_size();
3798 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
3799 return FAIL; // mismatched access
3800 }
3801 Node* val = st->in(MemNode::ValueIn);
3802
3803 if (!detect_init_independence(val, phase))
3804 return FAIL; // stored value must be 'simple enough'
3805
3806 // The Store can be captured only if nothing after the allocation
3807 // and before the Store is using the memory location that the store
3808 // overwrites.
3809 bool failed = false;
3810 // If is_complete_with_arraycopy() is true the shape of the graph is
3811 // well defined and is safe so no need for extra checks.
3812 if (!is_complete_with_arraycopy()) {
3813 // We are going to look at each use of the memory state following
3814 // the allocation to make sure nothing reads the memory that the
3815 // Store writes.
3816 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3817 int alias_idx = phase->C->get_alias_index(t_adr);
3818 ResourceMark rm;
3819 Unique_Node_List mems;
3820 mems.push(mem);
3821 Node* unique_merge = NULL;
3822 for (uint next = 0; next < mems.size(); ++next) {
3823 Node *m = mems.at(next);
3824 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3825 Node *n = m->fast_out(j);
3826 if (n->outcnt() == 0) {
3827 continue;
3828 }
3829 if (n == st) {
3830 continue;
3831 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3832 // If the control of this use is different from the control
3833 // of the Store which is right after the InitializeNode then
3834 // this node cannot be between the InitializeNode and the
3835 // Store.
3836 continue;
3837 } else if (n->is_MergeMem()) {
3838 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3839 // We can hit a MergeMemNode (that will likely go away
3840 // later) that is a direct use of the memory state
3841 // following the InitializeNode on the same slice as the
3842 // store node that we'd like to capture. We need to check
3843 // the uses of the MergeMemNode.
3844 mems.push(n);
3845 }
3846 } else if (n->is_Mem()) {
3847 Node* other_adr = n->in(MemNode::Address);
3848 if (other_adr == adr) {
3849 failed = true;
3850 break;
3851 } else {
3852 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3853 if (other_t_adr != NULL) {
3854 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3855 if (other_alias_idx == alias_idx) {
3856 // A load from the same memory slice as the store right
3857 // after the InitializeNode. We check the control of the
3858 // object/array that is loaded from. If it's the same as
3859 // the store control then we cannot capture the store.
3860 assert(!n->is_Store(), "2 stores to same slice on same control?");
3861 Node* base = other_adr;
3862 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3863 base = base->in(AddPNode::Base);
3864 if (base != NULL) {
3865 base = base->uncast();
3866 if (base->is_Proj() && base->in(0) == alloc) {
3867 failed = true;
3868 break;
3869 }
3870 }
3871 }
3872 }
3873 }
3874 } else {
3875 failed = true;
3876 break;
3877 }
3878 }
3879 }
3880 }
3881 if (failed) {
3882 if (!can_reshape) {
3883 // We decided we couldn't capture the store during parsing. We
3884 // should try again during the next IGVN once the graph is
3885 // cleaner.
3886 phase->C->record_for_igvn(st);
3887 }
3888 return FAIL;
3889 }
3890
3891 return offset; // success
3892 }
3893
3894 // Find the captured store in(i) which corresponds to the range
3895 // [start..start+size) in the initialized object.
3896 // If there is one, return its index i. If there isn't, return the
3897 // negative of the index where it should be inserted.
3898 // Return 0 if the queried range overlaps an initialization boundary
3899 // or if dead code is encountered.
3900 // If size_in_bytes is zero, do not bother with overlap checks.
3901 int InitializeNode::captured_store_insertion_point(intptr_t start,
3902 int size_in_bytes,
3903 PhaseTransform* phase) {
3904 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
3905
3906 if (is_complete())
3907 return FAIL; // arraycopy got here first; punt
3908
3909 assert(allocation() != NULL, "must be present");
3910
3911 // no negatives, no header fields:
3912 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3913
3914 // after a certain size, we bail out on tracking all the stores:
3915 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3916 if (start >= ti_limit) return FAIL;
3917
3918 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3919 if (i >= limit) return -(int)i; // not found; here is where to put it
3920
3921 Node* st = in(i);
3922 intptr_t st_off = get_store_offset(st, phase);
3923 if (st_off < 0) {
3924 if (st != zero_memory()) {
3925 return FAIL; // bail out if there is dead garbage
3926 }
3927 } else if (st_off > start) {
3928 // ...we are done, since stores are ordered
3929 if (st_off < start + size_in_bytes) {
3930 return FAIL; // the next store overlaps
3931 }
3932 return -(int)i; // not found; here is where to put it
3933 } else if (st_off < start) {
3934 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
3935 if (size_in_bytes != 0 &&
3936 start < st_off + MAX_STORE &&
3937 start < st_off + st->as_Store()->memory_size()) {
3938 return FAIL; // the previous store overlaps
3939 }
3940 } else {
3941 if (size_in_bytes != 0 &&
3942 st->as_Store()->memory_size() != size_in_bytes) {
3943 return FAIL; // mismatched store size
3944 }
3945 return i;
3946 }
3947
3948 ++i;
3949 }
3950 }
3951
3952 // Look for a captured store which initializes at the offset 'start'
3953 // with the given size. If there is no such store, and no other
3954 // initialization interferes, then return zero_memory (the memory
3955 // projection of the AllocateNode).
3956 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3957 PhaseTransform* phase) {
3958 assert(stores_are_sane(phase), "");
3959 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3960 if (i == 0) {
3961 return NULL; // something is dead
3962 } else if (i < 0) {
3963 return zero_memory(); // just primordial zero bits here
3964 } else {
3965 Node* st = in(i); // here is the store at this position
3966 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3967 return st;
3968 }
3969 }
3970
3971 // Create, as a raw pointer, an address within my new object at 'offset'.
3972 Node* InitializeNode::make_raw_address(intptr_t offset,
3973 PhaseTransform* phase) {
3974 Node* addr = in(RawAddress);
3975 if (offset != 0) {
3976 Compile* C = phase->C;
3977 addr = phase->transform( new AddPNode(C->top(), addr,
3978 phase->MakeConX(offset)) );
3979 }
3980 return addr;
3981 }
3982
3983 // Clone the given store, converting it into a raw store
3984 // initializing a field or element of my new object.
3985 // Caller is responsible for retiring the original store,
3986 // with subsume_node or the like.
3987 //
3988 // From the example above InitializeNode::InitializeNode,
3989 // here are the old stores to be captured:
3990 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3991 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3992 //
3993 // Here is the changed code; note the extra edges on init:
3994 // alloc = (Allocate ...)
3995 // rawoop = alloc.RawAddress
3996 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3997 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3998 // init = (Initialize alloc.Control alloc.Memory rawoop
3999 // rawstore1 rawstore2)
4000 //
4001 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
4002 PhaseGVN* phase, bool can_reshape) {
4003 assert(stores_are_sane(phase), "");
4004
4005 if (start < 0) return NULL;
4006 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
4007
4008 Compile* C = phase->C;
4009 int size_in_bytes = st->memory_size();
4010 int i = captured_store_insertion_point(start, size_in_bytes, phase);
4011 if (i == 0) return NULL; // bail out
4012 Node* prev_mem = NULL; // raw memory for the captured store
4013 if (i > 0) {
4014 prev_mem = in(i); // there is a pre-existing store under this one
4015 set_req(i, C->top()); // temporarily disconnect it
4016 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
4017 } else {
4018 i = -i; // no pre-existing store
4019 prev_mem = zero_memory(); // a slice of the newly allocated object
4020 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
4021 set_req(--i, C->top()); // reuse this edge; it has been folded away
4022 else
4023 ins_req(i, C->top()); // build a new edge
4024 }
4025 Node* new_st = st->clone();
4026 new_st->set_req(MemNode::Control, in(Control));
4027 new_st->set_req(MemNode::Memory, prev_mem);
4028 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
4029 new_st = phase->transform(new_st);
4030
4031 // At this point, new_st might have swallowed a pre-existing store
4032 // at the same offset, or perhaps new_st might have disappeared,
4033 // if it redundantly stored the same value (or zero to fresh memory).
4034
4035 // In any case, wire it in:
4036 PhaseIterGVN* igvn = phase->is_IterGVN();
4037 if (igvn) {
4038 igvn->rehash_node_delayed(this);
4039 }
4040 set_req(i, new_st);
4041
4042 // The caller may now kill the old guy.
4043 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
4044 assert(check_st == new_st || check_st == NULL, "must be findable");
4045 assert(!is_complete(), "");
4046 return new_st;
4047 }
4048
4049 static bool store_constant(jlong* tiles, int num_tiles,
4050 intptr_t st_off, int st_size,
4051 jlong con) {
4052 if ((st_off & (st_size-1)) != 0)
4053 return false; // strange store offset (assume size==2**N)
4054 address addr = (address)tiles + st_off;
4055 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
4056 switch (st_size) {
4057 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
4058 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
4059 case sizeof(jint): *(jint*) addr = (jint) con; break;
4060 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
4061 default: return false; // strange store size (detect size!=2**N here)
4062 }
4063 return true; // return success to caller
4064 }
4065
4066 // Coalesce subword constants into int constants and possibly
4067 // into long constants. The goal, if the CPU permits,
4068 // is to initialize the object with a small number of 64-bit tiles.
4069 // Also, convert floating-point constants to bit patterns.
4070 // Non-constants are not relevant to this pass.
4071 //
4072 // In terms of the running example on InitializeNode::InitializeNode
4073 // and InitializeNode::capture_store, here is the transformation
4074 // of rawstore1 and rawstore2 into rawstore12:
4075 // alloc = (Allocate ...)
4076 // rawoop = alloc.RawAddress
4077 // tile12 = 0x00010002
4078 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
4079 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
4080 //
4081 void
4082 InitializeNode::coalesce_subword_stores(intptr_t header_size,
4083 Node* size_in_bytes,
4084 PhaseGVN* phase) {
4085 Compile* C = phase->C;
4086
4087 assert(stores_are_sane(phase), "");
4088 // Note: After this pass, they are not completely sane,
4089 // since there may be some overlaps.
4090
4091 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
4092
4093 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4094 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
4095 size_limit = MIN2(size_limit, ti_limit);
4096 size_limit = align_up(size_limit, BytesPerLong);
4097 int num_tiles = size_limit / BytesPerLong;
4098
4099 // allocate space for the tile map:
4100 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
4101 jlong tiles_buf[small_len];
4102 Node* nodes_buf[small_len];
4103 jlong inits_buf[small_len];
4104 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
4105 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4106 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
4107 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
4108 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
4109 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4110 // tiles: exact bitwise model of all primitive constants
4111 // nodes: last constant-storing node subsumed into the tiles model
4112 // inits: which bytes (in each tile) are touched by any initializations
4113
4114 //// Pass A: Fill in the tile model with any relevant stores.
4115
4116 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
4117 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
4118 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
4119 Node* zmem = zero_memory(); // initially zero memory state
4120 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4121 Node* st = in(i);
4122 intptr_t st_off = get_store_offset(st, phase);
4123
4124 // Figure out the store's offset and constant value:
4125 if (st_off < header_size) continue; //skip (ignore header)
4126 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
4127 int st_size = st->as_Store()->memory_size();
4128 if (st_off + st_size > size_limit) break;
4129
4130 // Record which bytes are touched, whether by constant or not.
4131 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
4132 continue; // skip (strange store size)
4133
4134 const Type* val = phase->type(st->in(MemNode::ValueIn));
4135 if (!val->singleton()) continue; //skip (non-con store)
4136 BasicType type = val->basic_type();
4137
4138 jlong con = 0;
4139 switch (type) {
4140 case T_INT: con = val->is_int()->get_con(); break;
4141 case T_LONG: con = val->is_long()->get_con(); break;
4142 case T_FLOAT: con = jint_cast(val->getf()); break;
4143 case T_DOUBLE: con = jlong_cast(val->getd()); break;
4144 default: continue; //skip (odd store type)
4145 }
4146
4147 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
4148 st->Opcode() == Op_StoreL) {
4149 continue; // This StoreL is already optimal.
4150 }
4151
4152 // Store down the constant.
4153 store_constant(tiles, num_tiles, st_off, st_size, con);
4154
4155 intptr_t j = st_off >> LogBytesPerLong;
4156
4157 if (type == T_INT && st_size == BytesPerInt
4158 && (st_off & BytesPerInt) == BytesPerInt) {
4159 jlong lcon = tiles[j];
4160 if (!Matcher::isSimpleConstant64(lcon) &&
4161 st->Opcode() == Op_StoreI) {
4162 // This StoreI is already optimal by itself.
4163 jint* intcon = (jint*) &tiles[j];
4164 intcon[1] = 0; // undo the store_constant()
4165
4166 // If the previous store is also optimal by itself, back up and
4167 // undo the action of the previous loop iteration... if we can.
4168 // But if we can't, just let the previous half take care of itself.
4169 st = nodes[j];
4170 st_off -= BytesPerInt;
4171 con = intcon[0];
4172 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
4173 assert(st_off >= header_size, "still ignoring header");
4174 assert(get_store_offset(st, phase) == st_off, "must be");
4175 assert(in(i-1) == zmem, "must be");
4176 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
4177 assert(con == tcon->is_int()->get_con(), "must be");
4178 // Undo the effects of the previous loop trip, which swallowed st:
4179 intcon[0] = 0; // undo store_constant()
4180 set_req(i-1, st); // undo set_req(i, zmem)
4181 nodes[j] = NULL; // undo nodes[j] = st
4182 --old_subword; // undo ++old_subword
4183 }
4184 continue; // This StoreI is already optimal.
4185 }
4186 }
4187
4188 // This store is not needed.
4189 set_req(i, zmem);
4190 nodes[j] = st; // record for the moment
4191 if (st_size < BytesPerLong) // something has changed
4192 ++old_subword; // includes int/float, but who's counting...
4193 else ++old_long;
4194 }
4195
4196 if ((old_subword + old_long) == 0)
4197 return; // nothing more to do
4198
4199 //// Pass B: Convert any non-zero tiles into optimal constant stores.
4200 // Be sure to insert them before overlapping non-constant stores.
4201 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
4202 for (int j = 0; j < num_tiles; j++) {
4203 jlong con = tiles[j];
4204 jlong init = inits[j];
4205 if (con == 0) continue;
4206 jint con0, con1; // split the constant, address-wise
4207 jint init0, init1; // split the init map, address-wise
4208 { union { jlong con; jint intcon[2]; } u;
4209 u.con = con;
4210 con0 = u.intcon[0];
4211 con1 = u.intcon[1];
4212 u.con = init;
4213 init0 = u.intcon[0];
4214 init1 = u.intcon[1];
4215 }
4216
4217 Node* old = nodes[j];
4218 assert(old != NULL, "need the prior store");
4219 intptr_t offset = (j * BytesPerLong);
4220
4221 bool split = !Matcher::isSimpleConstant64(con);
4222
4223 if (offset < header_size) {
4224 assert(offset + BytesPerInt >= header_size, "second int counts");
4225 assert(*(jint*)&tiles[j] == 0, "junk in header");
4226 split = true; // only the second word counts
4227 // Example: int a[] = { 42 ... }
4228 } else if (con0 == 0 && init0 == -1) {
4229 split = true; // first word is covered by full inits
4230 // Example: int a[] = { ... foo(), 42 ... }
4231 } else if (con1 == 0 && init1 == -1) {
4232 split = true; // second word is covered by full inits
4233 // Example: int a[] = { ... 42, foo() ... }
4234 }
4235
4236 // Here's a case where init0 is neither 0 nor -1:
4237 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
4238 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
4239 // In this case the tile is not split; it is (jlong)42.
4240 // The big tile is stored down, and then the foo() value is inserted.
4241 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
4242
4243 Node* ctl = old->in(MemNode::Control);
4244 Node* adr = make_raw_address(offset, phase);
4245 const TypePtr* atp = TypeRawPtr::BOTTOM;
4246
4247 // One or two coalesced stores to plop down.
4248 Node* st[2];
4249 intptr_t off[2];
4250 int nst = 0;
4251 if (!split) {
4252 ++new_long;
4253 off[nst] = offset;
4254 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4255 phase->longcon(con), T_LONG, MemNode::unordered);
4256 } else {
4257 // Omit either if it is a zero.
4258 if (con0 != 0) {
4259 ++new_int;
4260 off[nst] = offset;
4261 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4262 phase->intcon(con0), T_INT, MemNode::unordered);
4263 }
4264 if (con1 != 0) {
4265 ++new_int;
4266 offset += BytesPerInt;
4267 adr = make_raw_address(offset, phase);
4268 off[nst] = offset;
4269 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4270 phase->intcon(con1), T_INT, MemNode::unordered);
4271 }
4272 }
4273
4274 // Insert second store first, then the first before the second.
4275 // Insert each one just before any overlapping non-constant stores.
4276 while (nst > 0) {
4277 Node* st1 = st[--nst];
4278 C->copy_node_notes_to(st1, old);
4279 st1 = phase->transform(st1);
4280 offset = off[nst];
4281 assert(offset >= header_size, "do not smash header");
4282 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4283 guarantee(ins_idx != 0, "must re-insert constant store");
4284 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
4285 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4286 set_req(--ins_idx, st1);
4287 else
4288 ins_req(ins_idx, st1);
4289 }
4290 }
4291
4292 if (PrintCompilation && WizardMode)
4293 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4294 old_subword, old_long, new_int, new_long);
4295 if (C->log() != NULL)
4296 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4297 old_subword, old_long, new_int, new_long);
4298
4299 // Clean up any remaining occurrences of zmem:
4300 remove_extra_zeroes();
4301 }
4302
4303 // Explore forward from in(start) to find the first fully initialized
4304 // word, and return its offset. Skip groups of subword stores which
4305 // together initialize full words. If in(start) is itself part of a
4306 // fully initialized word, return the offset of in(start). If there
4307 // are no following full-word stores, or if something is fishy, return
4308 // a negative value.
4309 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4310 int int_map = 0;
4311 intptr_t int_map_off = 0;
4312 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
4313
4314 for (uint i = start, limit = req(); i < limit; i++) {
4315 Node* st = in(i);
4316
4317 intptr_t st_off = get_store_offset(st, phase);
4318 if (st_off < 0) break; // return conservative answer
4319
4320 int st_size = st->as_Store()->memory_size();
4321 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4322 return st_off; // we found a complete word init
4323 }
4324
4325 // update the map:
4326
4327 intptr_t this_int_off = align_down(st_off, BytesPerInt);
4328 if (this_int_off != int_map_off) {
4329 // reset the map:
4330 int_map = 0;
4331 int_map_off = this_int_off;
4332 }
4333
4334 int subword_off = st_off - this_int_off;
4335 int_map |= right_n_bits(st_size) << subword_off;
4336 if ((int_map & FULL_MAP) == FULL_MAP) {
4337 return this_int_off; // we found a complete word init
4338 }
4339
4340 // Did this store hit or cross the word boundary?
4341 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4342 if (next_int_off == this_int_off + BytesPerInt) {
4343 // We passed the current int, without fully initializing it.
4344 int_map_off = next_int_off;
4345 int_map >>= BytesPerInt;
4346 } else if (next_int_off > this_int_off + BytesPerInt) {
4347 // We passed the current and next int.
4348 return this_int_off + BytesPerInt;
4349 }
4350 }
4351
4352 return -1;
4353 }
4354
4355
4356 // Called when the associated AllocateNode is expanded into CFG.
4357 // At this point, we may perform additional optimizations.
4358 // Linearize the stores by ascending offset, to make memory
4359 // activity as coherent as possible.
4360 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4361 intptr_t header_size,
4362 Node* size_in_bytes,
4363 PhaseIterGVN* phase) {
4364 assert(!is_complete(), "not already complete");
4365 assert(stores_are_sane(phase), "");
4366 assert(allocation() != NULL, "must be present");
4367
4368 remove_extra_zeroes();
4369
4370 if (ReduceFieldZeroing || ReduceBulkZeroing)
4371 // reduce instruction count for common initialization patterns
4372 coalesce_subword_stores(header_size, size_in_bytes, phase);
4373
4374 Node* zmem = zero_memory(); // initially zero memory state
4375 Node* inits = zmem; // accumulating a linearized chain of inits
4376 #ifdef ASSERT
4377 intptr_t first_offset = allocation()->minimum_header_size();
4378 intptr_t last_init_off = first_offset; // previous init offset
4379 intptr_t last_init_end = first_offset; // previous init offset+size
4380 intptr_t last_tile_end = first_offset; // previous tile offset+size
4381 #endif
4382 intptr_t zeroes_done = header_size;
4383
4384 bool do_zeroing = true; // we might give up if inits are very sparse
4385 int big_init_gaps = 0; // how many large gaps have we seen?
4386
4387 if (UseTLAB && ZeroTLAB) do_zeroing = false;
4388 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
4389
4390 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4391 Node* st = in(i);
4392 intptr_t st_off = get_store_offset(st, phase);
4393 if (st_off < 0)
4394 break; // unknown junk in the inits
4395 if (st->in(MemNode::Memory) != zmem)
4396 break; // complicated store chains somehow in list
4397
4398 int st_size = st->as_Store()->memory_size();
4399 intptr_t next_init_off = st_off + st_size;
4400
4401 if (do_zeroing && zeroes_done < next_init_off) {
4402 // See if this store needs a zero before it or under it.
4403 intptr_t zeroes_needed = st_off;
4404
4405 if (st_size < BytesPerInt) {
4406 // Look for subword stores which only partially initialize words.
4407 // If we find some, we must lay down some word-level zeroes first,
4408 // underneath the subword stores.
4409 //
4410 // Examples:
4411 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
4412 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
4413 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
4414 //
4415 // Note: coalesce_subword_stores may have already done this,
4416 // if it was prompted by constant non-zero subword initializers.
4417 // But this case can still arise with non-constant stores.
4418
4419 intptr_t next_full_store = find_next_fullword_store(i, phase);
4420
4421 // In the examples above:
4422 // in(i) p q r s x y z
4423 // st_off 12 13 14 15 12 13 14
4424 // st_size 1 1 1 1 1 1 1
4425 // next_full_s. 12 16 16 16 16 16 16
4426 // z's_done 12 16 16 16 12 16 12
4427 // z's_needed 12 16 16 16 16 16 16
4428 // zsize 0 0 0 0 4 0 4
4429 if (next_full_store < 0) {
4430 // Conservative tack: Zero to end of current word.
4431 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4432 } else {
4433 // Zero to beginning of next fully initialized word.
4434 // Or, don't zero at all, if we are already in that word.
4435 assert(next_full_store >= zeroes_needed, "must go forward");
4436 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4437 zeroes_needed = next_full_store;
4438 }
4439 }
4440
4441 if (zeroes_needed > zeroes_done) {
4442 intptr_t zsize = zeroes_needed - zeroes_done;
4443 // Do some incremental zeroing on rawmem, in parallel with inits.
4444 zeroes_done = align_down(zeroes_done, BytesPerInt);
4445 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4446 zeroes_done, zeroes_needed,
4447 phase);
4448 zeroes_done = zeroes_needed;
4449 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4450 do_zeroing = false; // leave the hole, next time
4451 }
4452 }
4453
4454 // Collect the store and move on:
4455 phase->replace_input_of(st, MemNode::Memory, inits);
4456 inits = st; // put it on the linearized chain
4457 set_req(i, zmem); // unhook from previous position
4458
4459 if (zeroes_done == st_off)
4460 zeroes_done = next_init_off;
4461
4462 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4463
4464 #ifdef ASSERT
4465 // Various order invariants. Weaker than stores_are_sane because
4466 // a large constant tile can be filled in by smaller non-constant stores.
4467 assert(st_off >= last_init_off, "inits do not reverse");
4468 last_init_off = st_off;
4469 const Type* val = NULL;
4470 if (st_size >= BytesPerInt &&
4471 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4472 (int)val->basic_type() < (int)T_OBJECT) {
4473 assert(st_off >= last_tile_end, "tiles do not overlap");
4474 assert(st_off >= last_init_end, "tiles do not overwrite inits");
4475 last_tile_end = MAX2(last_tile_end, next_init_off);
4476 } else {
4477 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4478 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4479 assert(st_off >= last_init_end, "inits do not overlap");
4480 last_init_end = next_init_off; // it's a non-tile
4481 }
4482 #endif //ASSERT
4483 }
4484
4485 remove_extra_zeroes(); // clear out all the zmems left over
4486 add_req(inits);
4487
4488 if (!(UseTLAB && ZeroTLAB)) {
4489 // If anything remains to be zeroed, zero it all now.
4490 zeroes_done = align_down(zeroes_done, BytesPerInt);
4491 // if it is the last unused 4 bytes of an instance, forget about it
4492 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4493 if (zeroes_done + BytesPerLong >= size_limit) {
4494 AllocateNode* alloc = allocation();
4495 assert(alloc != NULL, "must be present");
4496 if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
4497 Node* klass_node = alloc->in(AllocateNode::KlassNode);
4498 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4499 if (zeroes_done == k->layout_helper())
4500 zeroes_done = size_limit;
4501 }
4502 }
4503 if (zeroes_done < size_limit) {
4504 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4505 zeroes_done, size_in_bytes, phase);
4506 }
4507 }
4508
4509 set_complete(phase);
4510 return rawmem;
4511 }
4512
4513
4514 #ifdef ASSERT
4515 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4516 if (is_complete())
4517 return true; // stores could be anything at this point
4518 assert(allocation() != NULL, "must be present");
4519 intptr_t last_off = allocation()->minimum_header_size();
4520 for (uint i = InitializeNode::RawStores; i < req(); i++) {
4521 Node* st = in(i);
4522 intptr_t st_off = get_store_offset(st, phase);
4523 if (st_off < 0) continue; // ignore dead garbage
4524 if (last_off > st_off) {
4525 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4526 this->dump(2);
4527 assert(false, "ascending store offsets");
4528 return false;
4529 }
4530 last_off = st_off + st->as_Store()->memory_size();
4531 }
4532 return true;
4533 }
4534 #endif //ASSERT
4535
4536
4537
4538
4539 //============================MergeMemNode=====================================
4540 //
4541 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4542 // contributing store or call operations. Each contributor provides the memory
4543 // state for a particular "alias type" (see Compile::alias_type). For example,
4544 // if a MergeMem has an input X for alias category #6, then any memory reference
4545 // to alias category #6 may use X as its memory state input, as an exact equivalent
4546 // to using the MergeMem as a whole.
4547 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4548 //
4549 // (Here, the <N> notation gives the index of the relevant adr_type.)
4550 //
4551 // In one special case (and more cases in the future), alias categories overlap.
4552 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4553 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4554 // it is exactly equivalent to that state W:
4555 // MergeMem(<Bot>: W) <==> W
4556 //
4557 // Usually, the merge has more than one input. In that case, where inputs
4558 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4559 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4560 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4561 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4562 //
4563 // A merge can take a "wide" memory state as one of its narrow inputs.
4564 // This simply means that the merge observes out only the relevant parts of
4565 // the wide input. That is, wide memory states arriving at narrow merge inputs
4566 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4567 //
4568 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4569 // and that memory slices "leak through":
4570 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4571 //
4572 // But, in such a cascade, repeated memory slices can "block the leak":
4573 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4574 //
4575 // In the last example, Y is not part of the combined memory state of the
4576 // outermost MergeMem. The system must, of course, prevent unschedulable
4577 // memory states from arising, so you can be sure that the state Y is somehow
4578 // a precursor to state Y'.
4579 //
4580 //
4581 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4582 // of each MergeMemNode array are exactly the numerical alias indexes, including
4583 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4584 // Compile::alias_type (and kin) produce and manage these indexes.
4585 //
4586 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4587 // (Note that this provides quick access to the top node inside MergeMem methods,
4588 // without the need to reach out via TLS to Compile::current.)
4589 //
4590 // As a consequence of what was just described, a MergeMem that represents a full
4591 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4592 // containing all alias categories.
4593 //
4594 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4595 //
4596 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4597 // a memory state for the alias type <N>, or else the top node, meaning that
4598 // there is no particular input for that alias type. Note that the length of
4599 // a MergeMem is variable, and may be extended at any time to accommodate new
4600 // memory states at larger alias indexes. When merges grow, they are of course
4601 // filled with "top" in the unused in() positions.
4602 //
4603 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4604 // (Top was chosen because it works smoothly with passes like GCM.)
4605 //
4606 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4607 // the type of random VM bits like TLS references.) Since it is always the
4608 // first non-Bot memory slice, some low-level loops use it to initialize an
4609 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4610 //
4611 //
4612 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4613 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4614 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4615 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4616 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4617 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4618 //
4619 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4620 // really that different from the other memory inputs. An abbreviation called
4621 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4622 //
4623 //
4624 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4625 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4626 // that "emerges though" the base memory will be marked as excluding the alias types
4627 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4628 //
4629 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4630 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4631 //
4632 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4633 // (It is currently unimplemented.) As you can see, the resulting merge is
4634 // actually a disjoint union of memory states, rather than an overlay.
4635 //
4636
4637 //------------------------------MergeMemNode-----------------------------------
4638 Node* MergeMemNode::make_empty_memory() {
4639 Node* empty_memory = (Node*) Compile::current()->top();
4640 assert(empty_memory->is_top(), "correct sentinel identity");
4641 return empty_memory;
4642 }
4643
4644 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4645 init_class_id(Class_MergeMem);
4646 // all inputs are nullified in Node::Node(int)
4647 // set_input(0, NULL); // no control input
4648
4649 // Initialize the edges uniformly to top, for starters.
4650 Node* empty_mem = make_empty_memory();
4651 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4652 init_req(i,empty_mem);
4653 }
4654 assert(empty_memory() == empty_mem, "");
4655
4656 if( new_base != NULL && new_base->is_MergeMem() ) {
4657 MergeMemNode* mdef = new_base->as_MergeMem();
4658 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4659 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4660 mms.set_memory(mms.memory2());
4661 }
4662 assert(base_memory() == mdef->base_memory(), "");
4663 } else {
4664 set_base_memory(new_base);
4665 }
4666 }
4667
4668 // Make a new, untransformed MergeMem with the same base as 'mem'.
4669 // If mem is itself a MergeMem, populate the result with the same edges.
4670 MergeMemNode* MergeMemNode::make(Node* mem) {
4671 return new MergeMemNode(mem);
4672 }
4673
4674 //------------------------------cmp--------------------------------------------
4675 uint MergeMemNode::hash() const { return NO_HASH; }
4676 bool MergeMemNode::cmp( const Node &n ) const {
4677 return (&n == this); // Always fail except on self
4678 }
4679
4680 //------------------------------Identity---------------------------------------
4681 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4682 // Identity if this merge point does not record any interesting memory
4683 // disambiguations.
4684 Node* base_mem = base_memory();
4685 Node* empty_mem = empty_memory();
4686 if (base_mem != empty_mem) { // Memory path is not dead?
4687 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4688 Node* mem = in(i);
4689 if (mem != empty_mem && mem != base_mem) {
4690 return this; // Many memory splits; no change
4691 }
4692 }
4693 }
4694 return base_mem; // No memory splits; ID on the one true input
4695 }
4696
4697 //------------------------------Ideal------------------------------------------
4698 // This method is invoked recursively on chains of MergeMem nodes
4699 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4700 // Remove chain'd MergeMems
4701 //
4702 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4703 // relative to the "in(Bot)". Since we are patching both at the same time,
4704 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4705 // but rewrite each "in(i)" relative to the new "in(Bot)".
4706 Node *progress = NULL;
4707
4708
4709 Node* old_base = base_memory();
4710 Node* empty_mem = empty_memory();
4711 if (old_base == empty_mem)
4712 return NULL; // Dead memory path.
4713
4714 MergeMemNode* old_mbase;
4715 if (old_base != NULL && old_base->is_MergeMem())
4716 old_mbase = old_base->as_MergeMem();
4717 else
4718 old_mbase = NULL;
4719 Node* new_base = old_base;
4720
4721 // simplify stacked MergeMems in base memory
4722 if (old_mbase) new_base = old_mbase->base_memory();
4723
4724 // the base memory might contribute new slices beyond my req()
4725 if (old_mbase) grow_to_match(old_mbase);
4726
4727 // Look carefully at the base node if it is a phi.
4728 PhiNode* phi_base;
4729 if (new_base != NULL && new_base->is_Phi())
4730 phi_base = new_base->as_Phi();
4731 else
4732 phi_base = NULL;
4733
4734 Node* phi_reg = NULL;
4735 uint phi_len = (uint)-1;
4736 if (phi_base != NULL) {
4737 phi_reg = phi_base->region();
4738 phi_len = phi_base->req();
4739 // see if the phi is unfinished
4740 for (uint i = 1; i < phi_len; i++) {
4741 if (phi_base->in(i) == NULL) {
4742 // incomplete phi; do not look at it yet!
4743 phi_reg = NULL;
4744 phi_len = (uint)-1;
4745 break;
4746 }
4747 }
4748 }
4749
4750 // Note: We do not call verify_sparse on entry, because inputs
4751 // can normalize to the base_memory via subsume_node or similar
4752 // mechanisms. This method repairs that damage.
4753
4754 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4755
4756 // Look at each slice.
4757 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4758 Node* old_in = in(i);
4759 // calculate the old memory value
4760 Node* old_mem = old_in;
4761 if (old_mem == empty_mem) old_mem = old_base;
4762 assert(old_mem == memory_at(i), "");
4763
4764 // maybe update (reslice) the old memory value
4765
4766 // simplify stacked MergeMems
4767 Node* new_mem = old_mem;
4768 MergeMemNode* old_mmem;
4769 if (old_mem != NULL && old_mem->is_MergeMem())
4770 old_mmem = old_mem->as_MergeMem();
4771 else
4772 old_mmem = NULL;
4773 if (old_mmem == this) {
4774 // This can happen if loops break up and safepoints disappear.
4775 // A merge of BotPtr (default) with a RawPtr memory derived from a
4776 // safepoint can be rewritten to a merge of the same BotPtr with
4777 // the BotPtr phi coming into the loop. If that phi disappears
4778 // also, we can end up with a self-loop of the mergemem.
4779 // In general, if loops degenerate and memory effects disappear,
4780 // a mergemem can be left looking at itself. This simply means
4781 // that the mergemem's default should be used, since there is
4782 // no longer any apparent effect on this slice.
4783 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4784 // from start. Update the input to TOP.
4785 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4786 }
4787 else if (old_mmem != NULL) {
4788 new_mem = old_mmem->memory_at(i);
4789 }
4790 // else preceding memory was not a MergeMem
4791
4792 // maybe store down a new value
4793 Node* new_in = new_mem;
4794 if (new_in == new_base) new_in = empty_mem;
4795
4796 if (new_in != old_in) {
4797 // Warning: Do not combine this "if" with the previous "if"
4798 // A memory slice might have be be rewritten even if it is semantically
4799 // unchanged, if the base_memory value has changed.
4800 set_req_X(i, new_in, phase);
4801 progress = this; // Report progress
4802 }
4803 }
4804
4805 if (new_base != old_base) {
4806 set_req_X(Compile::AliasIdxBot, new_base, phase);
4807 // Don't use set_base_memory(new_base), because we need to update du.
4808 assert(base_memory() == new_base, "");
4809 progress = this;
4810 }
4811
4812 if( base_memory() == this ) {
4813 // a self cycle indicates this memory path is dead
4814 set_req(Compile::AliasIdxBot, empty_mem);
4815 }
4816
4817 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4818 // Recursion must occur after the self cycle check above
4819 if( base_memory()->is_MergeMem() ) {
4820 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4821 Node *m = phase->transform(new_mbase); // Rollup any cycles
4822 if( m != NULL &&
4823 (m->is_top() ||
4824 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4825 // propagate rollup of dead cycle to self
4826 set_req(Compile::AliasIdxBot, empty_mem);
4827 }
4828 }
4829
4830 if( base_memory() == empty_mem ) {
4831 progress = this;
4832 // Cut inputs during Parse phase only.
4833 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4834 if( !can_reshape ) {
4835 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4836 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4837 }
4838 }
4839 }
4840
4841 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4842 // Check if PhiNode::Ideal's "Split phis through memory merges"
4843 // transform should be attempted. Look for this->phi->this cycle.
4844 uint merge_width = req();
4845 if (merge_width > Compile::AliasIdxRaw) {
4846 PhiNode* phi = base_memory()->as_Phi();
4847 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4848 if (phi->in(i) == this) {
4849 phase->is_IterGVN()->_worklist.push(phi);
4850 break;
4851 }
4852 }
4853 }
4854 }
4855
4856 assert(progress || verify_sparse(), "please, no dups of base");
4857 return progress;
4858 }
4859
4860 //-------------------------set_base_memory-------------------------------------
4861 void MergeMemNode::set_base_memory(Node *new_base) {
4862 Node* empty_mem = empty_memory();
4863 set_req(Compile::AliasIdxBot, new_base);
4864 assert(memory_at(req()) == new_base, "must set default memory");
4865 // Clear out other occurrences of new_base:
4866 if (new_base != empty_mem) {
4867 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4868 if (in(i) == new_base) set_req(i, empty_mem);
4869 }
4870 }
4871 }
4872
4873 //------------------------------out_RegMask------------------------------------
4874 const RegMask &MergeMemNode::out_RegMask() const {
4875 return RegMask::Empty;
4876 }
4877
4878 //------------------------------dump_spec--------------------------------------
4879 #ifndef PRODUCT
4880 void MergeMemNode::dump_spec(outputStream *st) const {
4881 st->print(" {");
4882 Node* base_mem = base_memory();
4883 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4884 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4885 if (mem == base_mem) { st->print(" -"); continue; }
4886 st->print( " N%d:", mem->_idx );
4887 Compile::current()->get_adr_type(i)->dump_on(st);
4888 }
4889 st->print(" }");
4890 }
4891 #endif // !PRODUCT
4892
4893
4894 #ifdef ASSERT
4895 static bool might_be_same(Node* a, Node* b) {
4896 if (a == b) return true;
4897 if (!(a->is_Phi() || b->is_Phi())) return false;
4898 // phis shift around during optimization
4899 return true; // pretty stupid...
4900 }
4901
4902 // verify a narrow slice (either incoming or outgoing)
4903 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4904 if (!VerifyAliases) return; // don't bother to verify unless requested
4905 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
4906 if (Node::in_dump()) return; // muzzle asserts when printing
4907 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4908 assert(n != NULL, "");
4909 // Elide intervening MergeMem's
4910 while (n->is_MergeMem()) {
4911 n = n->as_MergeMem()->memory_at(alias_idx);
4912 }
4913 Compile* C = Compile::current();
4914 const TypePtr* n_adr_type = n->adr_type();
4915 if (n == m->empty_memory()) {
4916 // Implicit copy of base_memory()
4917 } else if (n_adr_type != TypePtr::BOTTOM) {
4918 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4919 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4920 } else {
4921 // A few places like make_runtime_call "know" that VM calls are narrow,
4922 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4923 bool expected_wide_mem = false;
4924 if (n == m->base_memory()) {
4925 expected_wide_mem = true;
4926 } else if (alias_idx == Compile::AliasIdxRaw ||
4927 n == m->memory_at(Compile::AliasIdxRaw)) {
4928 expected_wide_mem = true;
4929 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4930 // memory can "leak through" calls on channels that
4931 // are write-once. Allow this also.
4932 expected_wide_mem = true;
4933 }
4934 assert(expected_wide_mem, "expected narrow slice replacement");
4935 }
4936 }
4937 #else // !ASSERT
4938 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4939 #endif
4940
4941
4942 //-----------------------------memory_at---------------------------------------
4943 Node* MergeMemNode::memory_at(uint alias_idx) const {
4944 assert(alias_idx >= Compile::AliasIdxRaw ||
4945 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4946 "must avoid base_memory and AliasIdxTop");
4947
4948 // Otherwise, it is a narrow slice.
4949 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4950 Compile *C = Compile::current();
4951 if (is_empty_memory(n)) {
4952 // the array is sparse; empty slots are the "top" node
4953 n = base_memory();
4954 assert(Node::in_dump()
4955 || n == NULL || n->bottom_type() == Type::TOP
4956 || n->adr_type() == NULL // address is TOP
4957 || n->adr_type() == TypePtr::BOTTOM
4958 || n->adr_type() == TypeRawPtr::BOTTOM
4959 || Compile::current()->AliasLevel() == 0,
4960 "must be a wide memory");
4961 // AliasLevel == 0 if we are organizing the memory states manually.
4962 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4963 } else {
4964 // make sure the stored slice is sane
4965 #ifdef ASSERT
4966 if (VMError::is_error_reported() || Node::in_dump()) {
4967 } else if (might_be_same(n, base_memory())) {
4968 // Give it a pass: It is a mostly harmless repetition of the base.
4969 // This can arise normally from node subsumption during optimization.
4970 } else {
4971 verify_memory_slice(this, alias_idx, n);
4972 }
4973 #endif
4974 }
4975 return n;
4976 }
4977
4978 //---------------------------set_memory_at-------------------------------------
4979 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4980 verify_memory_slice(this, alias_idx, n);
4981 Node* empty_mem = empty_memory();
4982 if (n == base_memory()) n = empty_mem; // collapse default
4983 uint need_req = alias_idx+1;
4984 if (req() < need_req) {
4985 if (n == empty_mem) return; // already the default, so do not grow me
4986 // grow the sparse array
4987 do {
4988 add_req(empty_mem);
4989 } while (req() < need_req);
4990 }
4991 set_req( alias_idx, n );
4992 }
4993
4994
4995
4996 //--------------------------iteration_setup------------------------------------
4997 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4998 if (other != NULL) {
4999 grow_to_match(other);
5000 // invariant: the finite support of mm2 is within mm->req()
5001 #ifdef ASSERT
5002 for (uint i = req(); i < other->req(); i++) {
5003 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
5004 }
5005 #endif
5006 }
5007 // Replace spurious copies of base_memory by top.
5008 Node* base_mem = base_memory();
5009 if (base_mem != NULL && !base_mem->is_top()) {
5010 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
5011 if (in(i) == base_mem)
5012 set_req(i, empty_memory());
5013 }
5014 }
5015 }
5016
5017 //---------------------------grow_to_match-------------------------------------
5018 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
5019 Node* empty_mem = empty_memory();
5020 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
5021 // look for the finite support of the other memory
5022 for (uint i = other->req(); --i >= req(); ) {
5023 if (other->in(i) != empty_mem) {
5024 uint new_len = i+1;
5025 while (req() < new_len) add_req(empty_mem);
5026 break;
5027 }
5028 }
5029 }
5030
5031 //---------------------------verify_sparse-------------------------------------
5032 #ifndef PRODUCT
5033 bool MergeMemNode::verify_sparse() const {
5034 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
5035 Node* base_mem = base_memory();
5036 // The following can happen in degenerate cases, since empty==top.
5037 if (is_empty_memory(base_mem)) return true;
5038 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
5039 assert(in(i) != NULL, "sane slice");
5040 if (in(i) == base_mem) return false; // should have been the sentinel value!
5041 }
5042 return true;
5043 }
5044
5045 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
5046 Node* n;
5047 n = mm->in(idx);
5048 if (mem == n) return true; // might be empty_memory()
5049 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
5050 if (mem == n) return true;
5051 return false;
5052 }
5053 #endif // !PRODUCT