1 /*
2 * Copyright (c) 1997, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "memory/allocation.inline.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/connode.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/loopnode.hpp"
40 #include "opto/machnode.hpp"
41 #include "opto/matcher.hpp"
42 #include "opto/memnode.hpp"
43 #include "opto/mulnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/phaseX.hpp"
46 #include "opto/regmask.hpp"
47 #include "opto/rootnode.hpp"
48 #include "utilities/align.hpp"
49 #include "utilities/copy.hpp"
50 #include "utilities/macros.hpp"
51 #include "utilities/vmError.hpp"
52 #if INCLUDE_ZGC
53 #include "gc/z/c2/zBarrierSetC2.hpp"
54 #endif
55
56 // Portions of code courtesy of Clifford Click
57
58 // Optimization - Graph Style
59
60 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
61
62 //=============================================================================
63 uint MemNode::size_of() const { return sizeof(*this); }
64
65 const TypePtr *MemNode::adr_type() const {
66 Node* adr = in(Address);
67 if (adr == NULL) return NULL; // node is dead
68 const TypePtr* cross_check = NULL;
69 DEBUG_ONLY(cross_check = _adr_type);
70 return calculate_adr_type(adr->bottom_type(), cross_check);
71 }
72
73 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
74 if (adr != NULL) {
75 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
76 return true;
77 }
78 }
79 return false;
80 }
81
82 #ifndef PRODUCT
83 void MemNode::dump_spec(outputStream *st) const {
84 if (in(Address) == NULL) return; // node is dead
85 #ifndef ASSERT
86 // fake the missing field
87 const TypePtr* _adr_type = NULL;
88 if (in(Address) != NULL)
89 _adr_type = in(Address)->bottom_type()->isa_ptr();
90 #endif
91 dump_adr_type(this, _adr_type, st);
92
93 Compile* C = Compile::current();
94 if (C->alias_type(_adr_type)->is_volatile()) {
95 st->print(" Volatile!");
96 }
97 if (_unaligned_access) {
98 st->print(" unaligned");
99 }
100 if (_mismatched_access) {
101 st->print(" mismatched");
102 }
103 if (_unsafe_access) {
104 st->print(" unsafe");
105 }
106 }
107
108 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
109 st->print(" @");
110 if (adr_type == NULL) {
111 st->print("NULL");
112 } else {
113 adr_type->dump_on(st);
114 Compile* C = Compile::current();
115 Compile::AliasType* atp = NULL;
116 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
117 if (atp == NULL)
118 st->print(", idx=?\?;");
119 else if (atp->index() == Compile::AliasIdxBot)
120 st->print(", idx=Bot;");
121 else if (atp->index() == Compile::AliasIdxTop)
122 st->print(", idx=Top;");
123 else if (atp->index() == Compile::AliasIdxRaw)
124 st->print(", idx=Raw;");
125 else {
126 ciField* field = atp->field();
127 if (field) {
128 st->print(", name=");
129 field->print_name_on(st);
130 }
131 st->print(", idx=%d;", atp->index());
132 }
133 }
134 }
135
136 extern void print_alias_types();
137
138 #endif
139
140 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
141 assert((t_oop != NULL), "sanity");
142 bool is_instance = t_oop->is_known_instance_field();
143 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
144 (load != NULL) && load->is_Load() &&
145 (phase->is_IterGVN() != NULL);
146 if (!(is_instance || is_boxed_value_load))
147 return mchain; // don't try to optimize non-instance types
148 uint instance_id = t_oop->instance_id();
149 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
150 Node *prev = NULL;
151 Node *result = mchain;
152 while (prev != result) {
153 prev = result;
154 if (result == start_mem)
155 break; // hit one of our sentinels
156 // skip over a call which does not affect this memory slice
157 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
158 Node *proj_in = result->in(0);
159 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
160 break; // hit one of our sentinels
161 } else if (proj_in->is_Call()) {
162 // ArrayCopyNodes processed here as well
163 CallNode *call = proj_in->as_Call();
164 if (!call->may_modify(t_oop, phase)) { // returns false for instances
165 result = call->in(TypeFunc::Memory);
166 }
167 } else if (proj_in->is_Initialize()) {
168 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
169 // Stop if this is the initialization for the object instance which
170 // contains this memory slice, otherwise skip over it.
171 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
172 break;
173 }
174 if (is_instance) {
175 result = proj_in->in(TypeFunc::Memory);
176 } else if (is_boxed_value_load) {
177 Node* klass = alloc->in(AllocateNode::KlassNode);
178 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
179 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
180 result = proj_in->in(TypeFunc::Memory); // not related allocation
181 }
182 }
183 } else if (proj_in->is_MemBar()) {
184 ArrayCopyNode* ac = NULL;
185 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
186 break;
187 }
188 result = proj_in->in(TypeFunc::Memory);
189 } else {
190 assert(false, "unexpected projection");
191 }
192 } else if (result->is_ClearArray()) {
193 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
194 // Can not bypass initialization of the instance
195 // we are looking for.
196 break;
197 }
198 // Otherwise skip it (the call updated 'result' value).
199 } else if (result->is_MergeMem()) {
200 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
201 }
202 }
203 return result;
204 }
205
206 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
207 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
208 if (t_oop == NULL)
209 return mchain; // don't try to optimize non-oop types
210 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
211 bool is_instance = t_oop->is_known_instance_field();
212 PhaseIterGVN *igvn = phase->is_IterGVN();
213 if (is_instance && igvn != NULL && result->is_Phi()) {
214 PhiNode *mphi = result->as_Phi();
215 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
216 const TypePtr *t = mphi->adr_type();
217 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
218 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
219 t->is_oopptr()->cast_to_exactness(true)
220 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
221 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) {
222 // clone the Phi with our address type
223 result = mphi->split_out_instance(t_adr, igvn);
224 } else {
225 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
226 }
227 }
228 return result;
229 }
230
231 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
232 uint alias_idx = phase->C->get_alias_index(tp);
233 Node *mem = mmem;
234 #ifdef ASSERT
235 {
236 // Check that current type is consistent with the alias index used during graph construction
237 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
238 bool consistent = adr_check == NULL || adr_check->empty() ||
239 phase->C->must_alias(adr_check, alias_idx );
240 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
241 if( !consistent && adr_check != NULL && !adr_check->empty() &&
242 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
243 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
244 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
245 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
246 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
247 // don't assert if it is dead code.
248 consistent = true;
249 }
250 if( !consistent ) {
251 st->print("alias_idx==%d, adr_check==", alias_idx);
252 if( adr_check == NULL ) {
253 st->print("NULL");
254 } else {
255 adr_check->dump();
256 }
257 st->cr();
258 print_alias_types();
259 assert(consistent, "adr_check must match alias idx");
260 }
261 }
262 #endif
263 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
264 // means an array I have not precisely typed yet. Do not do any
265 // alias stuff with it any time soon.
266 const TypeOopPtr *toop = tp->isa_oopptr();
267 if( tp->base() != Type::AnyPtr &&
268 !(toop &&
269 toop->klass() != NULL &&
270 toop->klass()->is_java_lang_Object() &&
271 toop->offset() == Type::OffsetBot) ) {
272 // compress paths and change unreachable cycles to TOP
273 // If not, we can update the input infinitely along a MergeMem cycle
274 // Equivalent code in PhiNode::Ideal
275 Node* m = phase->transform(mmem);
276 // If transformed to a MergeMem, get the desired slice
277 // Otherwise the returned node represents memory for every slice
278 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
279 // Update input if it is progress over what we have now
280 }
281 return mem;
282 }
283
284 //--------------------------Ideal_common---------------------------------------
285 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
286 // Unhook non-raw memories from complete (macro-expanded) initializations.
287 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
288 // If our control input is a dead region, kill all below the region
289 Node *ctl = in(MemNode::Control);
290 if (ctl && remove_dead_region(phase, can_reshape))
291 return this;
292 ctl = in(MemNode::Control);
293 // Don't bother trying to transform a dead node
294 if (ctl && ctl->is_top()) return NodeSentinel;
295
296 PhaseIterGVN *igvn = phase->is_IterGVN();
297 // Wait if control on the worklist.
298 if (ctl && can_reshape && igvn != NULL) {
299 Node* bol = NULL;
300 Node* cmp = NULL;
301 if (ctl->in(0)->is_If()) {
302 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
303 bol = ctl->in(0)->in(1);
304 if (bol->is_Bool())
305 cmp = ctl->in(0)->in(1)->in(1);
306 }
307 if (igvn->_worklist.member(ctl) ||
308 (bol != NULL && igvn->_worklist.member(bol)) ||
309 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
310 // This control path may be dead.
311 // Delay this memory node transformation until the control is processed.
312 phase->is_IterGVN()->_worklist.push(this);
313 return NodeSentinel; // caller will return NULL
314 }
315 }
316 // Ignore if memory is dead, or self-loop
317 Node *mem = in(MemNode::Memory);
318 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
319 assert(mem != this, "dead loop in MemNode::Ideal");
320
321 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
322 // This memory slice may be dead.
323 // Delay this mem node transformation until the memory is processed.
324 phase->is_IterGVN()->_worklist.push(this);
325 return NodeSentinel; // caller will return NULL
326 }
327
328 Node *address = in(MemNode::Address);
329 const Type *t_adr = phase->type(address);
330 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
331
332 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
333 // Unsafe off-heap access with zero address. Remove access and other control users
334 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
335 assert(ctl != NULL, "unsafe accesses should be control dependent");
336 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
337 Node* u = ctl->fast_out(i);
338 if (u != ctl) {
339 igvn->rehash_node_delayed(u);
340 int nb = u->replace_edge(ctl, phase->C->top());
341 --i, imax -= nb;
342 }
343 }
344 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
345 Node* halt = igvn->transform(new HaltNode(ctl, frame));
346 phase->C->root()->add_req(halt);
347 return this;
348 }
349
350 if (can_reshape && igvn != NULL &&
351 (igvn->_worklist.member(address) ||
352 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
353 // The address's base and type may change when the address is processed.
354 // Delay this mem node transformation until the address is processed.
355 phase->is_IterGVN()->_worklist.push(this);
356 return NodeSentinel; // caller will return NULL
357 }
358
359 // Do NOT remove or optimize the next lines: ensure a new alias index
360 // is allocated for an oop pointer type before Escape Analysis.
361 // Note: C++ will not remove it since the call has side effect.
362 if (t_adr->isa_oopptr()) {
363 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
364 }
365
366 Node* base = NULL;
367 if (address->is_AddP()) {
368 base = address->in(AddPNode::Base);
369 }
370 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
371 !t_adr->isa_rawptr()) {
372 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
373 // Skip this node optimization if its address has TOP base.
374 return NodeSentinel; // caller will return NULL
375 }
376
377 // Avoid independent memory operations
378 Node* old_mem = mem;
379
380 // The code which unhooks non-raw memories from complete (macro-expanded)
381 // initializations was removed. After macro-expansion all stores catched
382 // by Initialize node became raw stores and there is no information
383 // which memory slices they modify. So it is unsafe to move any memory
384 // operation above these stores. Also in most cases hooked non-raw memories
385 // were already unhooked by using information from detect_ptr_independence()
386 // and find_previous_store().
387
388 if (mem->is_MergeMem()) {
389 MergeMemNode* mmem = mem->as_MergeMem();
390 const TypePtr *tp = t_adr->is_ptr();
391
392 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
393 }
394
395 if (mem != old_mem) {
396 set_req(MemNode::Memory, mem);
397 if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) {
398 igvn->_worklist.push(old_mem);
399 }
400 if (phase->type(mem) == Type::TOP) return NodeSentinel;
401 return this;
402 }
403
404 // let the subclass continue analyzing...
405 return NULL;
406 }
407
408 // Helper function for proving some simple control dominations.
409 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
410 // Already assumes that 'dom' is available at 'sub', and that 'sub'
411 // is not a constant (dominated by the method's StartNode).
412 // Used by MemNode::find_previous_store to prove that the
413 // control input of a memory operation predates (dominates)
414 // an allocation it wants to look past.
415 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
416 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
417 return false; // Conservative answer for dead code
418
419 // Check 'dom'. Skip Proj and CatchProj nodes.
420 dom = dom->find_exact_control(dom);
421 if (dom == NULL || dom->is_top())
422 return false; // Conservative answer for dead code
423
424 if (dom == sub) {
425 // For the case when, for example, 'sub' is Initialize and the original
426 // 'dom' is Proj node of the 'sub'.
427 return false;
428 }
429
430 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
431 return true;
432
433 // 'dom' dominates 'sub' if its control edge and control edges
434 // of all its inputs dominate or equal to sub's control edge.
435
436 // Currently 'sub' is either Allocate, Initialize or Start nodes.
437 // Or Region for the check in LoadNode::Ideal();
438 // 'sub' should have sub->in(0) != NULL.
439 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
440 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
441
442 // Get control edge of 'sub'.
443 Node* orig_sub = sub;
444 sub = sub->find_exact_control(sub->in(0));
445 if (sub == NULL || sub->is_top())
446 return false; // Conservative answer for dead code
447
448 assert(sub->is_CFG(), "expecting control");
449
450 if (sub == dom)
451 return true;
452
453 if (sub->is_Start() || sub->is_Root())
454 return false;
455
456 {
457 // Check all control edges of 'dom'.
458
459 ResourceMark rm;
460 Arena* arena = Thread::current()->resource_area();
461 Node_List nlist(arena);
462 Unique_Node_List dom_list(arena);
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 that must have set (can_see_stored_value=true) or
534 // could have set (can_see_stored_value=false) the value for this load
535 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
536 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
537 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
538 Node* mb = mem->in(0);
539 if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
540 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
541 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy();
542 if (ac->is_clonebasic()) {
543 intptr_t offset;
544 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset);
545 if (alloc != NULL && alloc == ld_alloc) {
546 return ac;
547 }
548 }
549 }
550 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
551 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
552
553 if (ac->is_arraycopy_validated() ||
554 ac->is_copyof_validated() ||
555 ac->is_copyofrange_validated()) {
556 Node* ld_addp = in(MemNode::Address);
557 if (ld_addp->is_AddP()) {
558 Node* ld_base = ld_addp->in(AddPNode::Address);
559 Node* ld_offs = ld_addp->in(AddPNode::Offset);
560
561 Node* dest = ac->in(ArrayCopyNode::Dest);
562
563 if (dest == ld_base) {
564 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
565 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
566 return ac;
567 }
568 if (!can_see_stored_value) {
569 mem = ac->in(TypeFunc::Memory);
570 }
571 }
572 }
573 }
574 }
575 return NULL;
576 }
577
578 // The logic for reordering loads and stores uses four steps:
579 // (a) Walk carefully past stores and initializations which we
580 // can prove are independent of this load.
581 // (b) Observe that the next memory state makes an exact match
582 // with self (load or store), and locate the relevant store.
583 // (c) Ensure that, if we were to wire self directly to the store,
584 // the optimizer would fold it up somehow.
585 // (d) Do the rewiring, and return, depending on some other part of
586 // the optimizer to fold up the load.
587 // This routine handles steps (a) and (b). Steps (c) and (d) are
588 // specific to loads and stores, so they are handled by the callers.
589 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
590 //
591 Node* MemNode::find_previous_store(PhaseTransform* phase) {
592 Node* ctrl = in(MemNode::Control);
593 Node* adr = in(MemNode::Address);
594 intptr_t offset = 0;
595 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
596 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
597
598 if (offset == Type::OffsetBot)
599 return NULL; // cannot unalias unless there are precise offsets
600
601 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
602 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
603
604 intptr_t size_in_bytes = memory_size();
605
606 Node* mem = in(MemNode::Memory); // start searching here...
607
608 int cnt = 50; // Cycle limiter
609 for (;;) { // While we can dance past unrelated stores...
610 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
611
612 Node* prev = mem;
613 if (mem->is_Store()) {
614 Node* st_adr = mem->in(MemNode::Address);
615 intptr_t st_offset = 0;
616 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
617 if (st_base == NULL)
618 break; // inscrutable pointer
619
620 // For raw accesses it's not enough to prove that constant offsets don't intersect.
621 // We need the bases to be the equal in order for the offset check to make sense.
622 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
623 break;
624 }
625
626 if (st_offset != offset && st_offset != Type::OffsetBot) {
627 const int MAX_STORE = BytesPerLong;
628 if (st_offset >= offset + size_in_bytes ||
629 st_offset <= offset - MAX_STORE ||
630 st_offset <= offset - mem->as_Store()->memory_size()) {
631 // Success: The offsets are provably independent.
632 // (You may ask, why not just test st_offset != offset and be done?
633 // The answer is that stores of different sizes can co-exist
634 // in the same sequence of RawMem effects. We sometimes initialize
635 // a whole 'tile' of array elements with a single jint or jlong.)
636 mem = mem->in(MemNode::Memory);
637 continue; // (a) advance through independent store memory
638 }
639 }
640 if (st_base != base &&
641 detect_ptr_independence(base, alloc,
642 st_base,
643 AllocateNode::Ideal_allocation(st_base, phase),
644 phase)) {
645 // Success: The bases are provably independent.
646 mem = mem->in(MemNode::Memory);
647 continue; // (a) advance through independent store memory
648 }
649
650 // (b) At this point, if the bases or offsets do not agree, we lose,
651 // since we have not managed to prove 'this' and 'mem' independent.
652 if (st_base == base && st_offset == offset) {
653 return mem; // let caller handle steps (c), (d)
654 }
655
656 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
657 InitializeNode* st_init = mem->in(0)->as_Initialize();
658 AllocateNode* st_alloc = st_init->allocation();
659 if (st_alloc == NULL)
660 break; // something degenerated
661 bool known_identical = false;
662 bool known_independent = false;
663 if (alloc == st_alloc)
664 known_identical = true;
665 else if (alloc != NULL)
666 known_independent = true;
667 else if (all_controls_dominate(this, st_alloc))
668 known_independent = true;
669
670 if (known_independent) {
671 // The bases are provably independent: Either they are
672 // manifestly distinct allocations, or else the control
673 // of this load dominates the store's allocation.
674 int alias_idx = phase->C->get_alias_index(adr_type());
675 if (alias_idx == Compile::AliasIdxRaw) {
676 mem = st_alloc->in(TypeFunc::Memory);
677 } else {
678 mem = st_init->memory(alias_idx);
679 }
680 continue; // (a) advance through independent store memory
681 }
682
683 // (b) at this point, if we are not looking at a store initializing
684 // the same allocation we are loading from, we lose.
685 if (known_identical) {
686 // From caller, can_see_stored_value will consult find_captured_store.
687 return mem; // let caller handle steps (c), (d)
688 }
689
690 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
691 if (prev != mem) {
692 // Found an arraycopy but it doesn't affect that load
693 continue;
694 }
695 // Found an arraycopy that may affect that load
696 return mem;
697 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
698 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
699 if (mem->is_Proj() && mem->in(0)->is_Call()) {
700 // ArrayCopyNodes processed here as well.
701 CallNode *call = mem->in(0)->as_Call();
702 if (!call->may_modify(addr_t, phase)) {
703 mem = call->in(TypeFunc::Memory);
704 continue; // (a) advance through independent call memory
705 }
706 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
707 ArrayCopyNode* ac = NULL;
708 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
709 break;
710 }
711 mem = mem->in(0)->in(TypeFunc::Memory);
712 continue; // (a) advance through independent MemBar memory
713 } else if (mem->is_ClearArray()) {
714 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
715 // (the call updated 'mem' value)
716 continue; // (a) advance through independent allocation memory
717 } else {
718 // Can not bypass initialization of the instance
719 // we are looking for.
720 return mem;
721 }
722 } else if (mem->is_MergeMem()) {
723 int alias_idx = phase->C->get_alias_index(adr_type());
724 mem = mem->as_MergeMem()->memory_at(alias_idx);
725 continue; // (a) advance through independent MergeMem memory
726 }
727 }
728
729 // Unless there is an explicit 'continue', we must bail out here,
730 // because 'mem' is an inscrutable memory state (e.g., a call).
731 break;
732 }
733
734 return NULL; // bail out
735 }
736
737 //----------------------calculate_adr_type-------------------------------------
738 // Helper function. Notices when the given type of address hits top or bottom.
739 // Also, asserts a cross-check of the type against the expected address type.
740 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
741 if (t == Type::TOP) return NULL; // does not touch memory any more?
742 #ifdef PRODUCT
743 cross_check = NULL;
744 #else
745 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL;
746 #endif
747 const TypePtr* tp = t->isa_ptr();
748 if (tp == NULL) {
749 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
750 return TypePtr::BOTTOM; // touches lots of memory
751 } else {
752 #ifdef ASSERT
753 // %%%% [phh] We don't check the alias index if cross_check is
754 // TypeRawPtr::BOTTOM. Needs to be investigated.
755 if (cross_check != NULL &&
756 cross_check != TypePtr::BOTTOM &&
757 cross_check != TypeRawPtr::BOTTOM) {
758 // Recheck the alias index, to see if it has changed (due to a bug).
759 Compile* C = Compile::current();
760 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
761 "must stay in the original alias category");
762 // The type of the address must be contained in the adr_type,
763 // disregarding "null"-ness.
764 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
765 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
766 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
767 "real address must not escape from expected memory type");
768 }
769 #endif
770 return tp;
771 }
772 }
773
774 //=============================================================================
775 // Should LoadNode::Ideal() attempt to remove control edges?
776 bool LoadNode::can_remove_control() const {
777 return true;
778 }
779 uint LoadNode::size_of() const { return sizeof(*this); }
780 bool LoadNode::cmp( const Node &n ) const
781 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
782 const Type *LoadNode::bottom_type() const { return _type; }
783 uint LoadNode::ideal_reg() const {
784 return _type->ideal_reg();
785 }
786
787 #ifndef PRODUCT
788 void LoadNode::dump_spec(outputStream *st) const {
789 MemNode::dump_spec(st);
790 if( !Verbose && !WizardMode ) {
791 // standard dump does this in Verbose and WizardMode
792 st->print(" #"); _type->dump_on(st);
793 }
794 if (!depends_only_on_test()) {
795 st->print(" (does not depend only on test)");
796 }
797 }
798 #endif
799
800 #ifdef ASSERT
801 //----------------------------is_immutable_value-------------------------------
802 // Helper function to allow a raw load without control edge for some cases
803 bool LoadNode::is_immutable_value(Node* adr) {
804 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
805 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
806 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
807 in_bytes(JavaThread::osthread_offset())));
808 }
809 #endif
810
811 //----------------------------LoadNode::make-----------------------------------
812 // Polymorphic factory method:
813 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo,
814 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) {
815 Compile* C = gvn.C;
816
817 // sanity check the alias category against the created node type
818 assert(!(adr_type->isa_oopptr() &&
819 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
820 "use LoadKlassNode instead");
821 assert(!(adr_type->isa_aryptr() &&
822 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
823 "use LoadRangeNode instead");
824 // Check control edge of raw loads
825 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
826 // oop will be recorded in oop map if load crosses safepoint
827 rt->isa_oopptr() || is_immutable_value(adr),
828 "raw memory operations should have control edge");
829 LoadNode* load = NULL;
830 switch (bt) {
831 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
832 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
833 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
834 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
835 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
836 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break;
837 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
838 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
839 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
840 case T_OBJECT:
841 #ifdef _LP64
842 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
843 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
844 } else
845 #endif
846 {
847 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
848 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
849 }
850 break;
851 default:
852 ShouldNotReachHere();
853 break;
854 }
855 assert(load != NULL, "LoadNode should have been created");
856 if (unaligned) {
857 load->set_unaligned_access();
858 }
859 if (mismatched) {
860 load->set_mismatched_access();
861 }
862 if (unsafe) {
863 load->set_unsafe_access();
864 }
865 if (load->Opcode() == Op_LoadN) {
866 Node* ld = gvn.transform(load);
867 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
868 }
869
870 return load;
871 }
872
873 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
874 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) {
875 bool require_atomic = true;
876 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic);
877 if (unaligned) {
878 load->set_unaligned_access();
879 }
880 if (mismatched) {
881 load->set_mismatched_access();
882 }
883 if (unsafe) {
884 load->set_unsafe_access();
885 }
886 return load;
887 }
888
889 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
890 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) {
891 bool require_atomic = true;
892 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic);
893 if (unaligned) {
894 load->set_unaligned_access();
895 }
896 if (mismatched) {
897 load->set_mismatched_access();
898 }
899 if (unsafe) {
900 load->set_unsafe_access();
901 }
902 return load;
903 }
904
905
906
907 //------------------------------hash-------------------------------------------
908 uint LoadNode::hash() const {
909 // unroll addition of interesting fields
910 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
911 }
912
913 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
914 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
915 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
916 bool is_stable_ary = FoldStableValues &&
917 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
918 tp->isa_aryptr()->is_stable();
919
920 return (eliminate_boxing && non_volatile) || is_stable_ary;
921 }
922
923 return false;
924 }
925
926 // Is the value loaded previously stored by an arraycopy? If so return
927 // a load node that reads from the source array so we may be able to
928 // optimize out the ArrayCopy node later.
929 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
930 Node* ld_adr = in(MemNode::Address);
931 intptr_t ld_off = 0;
932 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
933 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
934 if (ac != NULL) {
935 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
936
937 Node* mem = ac->in(TypeFunc::Memory);
938 Node* ctl = ac->in(0);
939 Node* src = ac->in(ArrayCopyNode::Src);
940
941 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
942 return NULL;
943 }
944
945 LoadNode* ld = clone()->as_Load();
946 Node* addp = in(MemNode::Address)->clone();
947 if (ac->as_ArrayCopy()->is_clonebasic()) {
948 assert(ld_alloc != NULL, "need an alloc");
949 assert(addp->is_AddP(), "address must be addp");
950 assert(ac->in(ArrayCopyNode::Dest)->is_AddP(), "dest must be an address");
951 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
952 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base)), "strange pattern");
953 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address)), "strange pattern");
954 addp->set_req(AddPNode::Base, src->in(AddPNode::Base));
955 addp->set_req(AddPNode::Address, src->in(AddPNode::Address));
956 } else {
957 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
958 ac->as_ArrayCopy()->is_copyof_validated() ||
959 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
960 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
961 addp->set_req(AddPNode::Base, src);
962 addp->set_req(AddPNode::Address, src);
963
964 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
965 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type();
966 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
967 uint shift = exact_log2(type2aelembytes(ary_elem));
968
969 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
970 #ifdef _LP64
971 diff = phase->transform(new ConvI2LNode(diff));
972 #endif
973 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
974
975 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
976 addp->set_req(AddPNode::Offset, offset);
977 }
978 addp = phase->transform(addp);
979 #ifdef ASSERT
980 const TypePtr* adr_type = phase->type(addp)->is_ptr();
981 ld->_adr_type = adr_type;
982 #endif
983 ld->set_req(MemNode::Address, addp);
984 ld->set_req(0, ctl);
985 ld->set_req(MemNode::Memory, mem);
986 // load depends on the tests that validate the arraycopy
987 ld->_control_dependency = Pinned;
988 return ld;
989 }
990 return NULL;
991 }
992
993
994 //---------------------------can_see_stored_value------------------------------
995 // This routine exists to make sure this set of tests is done the same
996 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
997 // will change the graph shape in a way which makes memory alive twice at the
998 // same time (uses the Oracle model of aliasing), then some
999 // LoadXNode::Identity will fold things back to the equivalence-class model
1000 // of aliasing.
1001 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
1002 Node* ld_adr = in(MemNode::Address);
1003 intptr_t ld_off = 0;
1004 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1005 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base, phase);
1006 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1007 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
1008 // This is more general than load from boxing objects.
1009 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1010 uint alias_idx = atp->index();
1011 bool final = !atp->is_rewritable();
1012 Node* result = NULL;
1013 Node* current = st;
1014 // Skip through chains of MemBarNodes checking the MergeMems for
1015 // new states for the slice of this load. Stop once any other
1016 // kind of node is encountered. Loads from final memory can skip
1017 // through any kind of MemBar but normal loads shouldn't skip
1018 // through MemBarAcquire since the could allow them to move out of
1019 // a synchronized region.
1020 while (current->is_Proj()) {
1021 int opc = current->in(0)->Opcode();
1022 if ((final && (opc == Op_MemBarAcquire ||
1023 opc == Op_MemBarAcquireLock ||
1024 opc == Op_LoadFence)) ||
1025 opc == Op_MemBarRelease ||
1026 opc == Op_StoreFence ||
1027 opc == Op_MemBarReleaseLock ||
1028 opc == Op_MemBarStoreStore ||
1029 opc == Op_MemBarCPUOrder) {
1030 Node* mem = current->in(0)->in(TypeFunc::Memory);
1031 if (mem->is_MergeMem()) {
1032 MergeMemNode* merge = mem->as_MergeMem();
1033 Node* new_st = merge->memory_at(alias_idx);
1034 if (new_st == merge->base_memory()) {
1035 // Keep searching
1036 current = new_st;
1037 continue;
1038 }
1039 // Save the new memory state for the slice and fall through
1040 // to exit.
1041 result = new_st;
1042 }
1043 }
1044 break;
1045 }
1046 if (result != NULL) {
1047 st = result;
1048 }
1049 }
1050
1051 // Loop around twice in the case Load -> Initialize -> Store.
1052 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1053 for (int trip = 0; trip <= 1; trip++) {
1054
1055 if (st->is_Store()) {
1056 Node* st_adr = st->in(MemNode::Address);
1057 if (!phase->eqv(st_adr, ld_adr)) {
1058 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1059 intptr_t st_off = 0;
1060 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1061 if (ld_base == NULL) return NULL;
1062 if (st_base == NULL) return NULL;
1063 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return NULL;
1064 if (ld_off != st_off) return NULL;
1065 if (ld_off == Type::OffsetBot) return NULL;
1066 // Same base, same offset.
1067 // Possible improvement for arrays: check index value instead of absolute offset.
1068
1069 // At this point we have proven something like this setup:
1070 // B = << base >>
1071 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1072 // S = StoreQ(AddP( B , #Off), V)
1073 // (Actually, we haven't yet proven the Q's are the same.)
1074 // In other words, we are loading from a casted version of
1075 // the same pointer-and-offset that we stored to.
1076 // Casted version may carry a dependency and it is respected.
1077 // Thus, we are able to replace L by V.
1078 }
1079 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1080 if (store_Opcode() != st->Opcode())
1081 return NULL;
1082 return st->in(MemNode::ValueIn);
1083 }
1084
1085 // A load from a freshly-created object always returns zero.
1086 // (This can happen after LoadNode::Ideal resets the load's memory input
1087 // to find_captured_store, which returned InitializeNode::zero_memory.)
1088 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1089 (st->in(0) == ld_alloc) &&
1090 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1091 // return a zero value for the load's basic type
1092 // (This is one of the few places where a generic PhaseTransform
1093 // can create new nodes. Think of it as lazily manifesting
1094 // virtually pre-existing constants.)
1095 return phase->zerocon(memory_type());
1096 }
1097
1098 // A load from an initialization barrier can match a captured store.
1099 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1100 InitializeNode* init = st->in(0)->as_Initialize();
1101 AllocateNode* alloc = init->allocation();
1102 if ((alloc != NULL) && (alloc == ld_alloc)) {
1103 // examine a captured store value
1104 st = init->find_captured_store(ld_off, memory_size(), phase);
1105 if (st != NULL) {
1106 continue; // take one more trip around
1107 }
1108 }
1109 }
1110
1111 // Load boxed value from result of valueOf() call is input parameter.
1112 if (this->is_Load() && ld_adr->is_AddP() &&
1113 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1114 intptr_t ignore = 0;
1115 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1116 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1117 base = bs->step_over_gc_barrier(base);
1118 if (base != NULL && base->is_Proj() &&
1119 base->as_Proj()->_con == TypeFunc::Parms &&
1120 base->in(0)->is_CallStaticJava() &&
1121 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1122 return base->in(0)->in(TypeFunc::Parms);
1123 }
1124 }
1125
1126 break;
1127 }
1128
1129 return NULL;
1130 }
1131
1132 //----------------------is_instance_field_load_with_local_phi------------------
1133 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1134 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1135 in(Address)->is_AddP() ) {
1136 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1137 // Only instances and boxed values.
1138 if( t_oop != NULL &&
1139 (t_oop->is_ptr_to_boxed_value() ||
1140 t_oop->is_known_instance_field()) &&
1141 t_oop->offset() != Type::OffsetBot &&
1142 t_oop->offset() != Type::OffsetTop) {
1143 return true;
1144 }
1145 }
1146 return false;
1147 }
1148
1149 //------------------------------Identity---------------------------------------
1150 // Loads are identity if previous store is to same address
1151 Node* LoadNode::Identity(PhaseGVN* phase) {
1152 // If the previous store-maker is the right kind of Store, and the store is
1153 // to the same address, then we are equal to the value stored.
1154 Node* mem = in(Memory);
1155 Node* value = can_see_stored_value(mem, phase);
1156 if( value ) {
1157 // byte, short & char stores truncate naturally.
1158 // A load has to load the truncated value which requires
1159 // some sort of masking operation and that requires an
1160 // Ideal call instead of an Identity call.
1161 if (memory_size() < BytesPerInt) {
1162 // If the input to the store does not fit with the load's result type,
1163 // it must be truncated via an Ideal call.
1164 if (!phase->type(value)->higher_equal(phase->type(this)))
1165 return this;
1166 }
1167 // (This works even when value is a Con, but LoadNode::Value
1168 // usually runs first, producing the singleton type of the Con.)
1169 return value;
1170 }
1171
1172 // Search for an existing data phi which was generated before for the same
1173 // instance's field to avoid infinite generation of phis in a loop.
1174 Node *region = mem->in(0);
1175 if (is_instance_field_load_with_local_phi(region)) {
1176 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1177 int this_index = phase->C->get_alias_index(addr_t);
1178 int this_offset = addr_t->offset();
1179 int this_iid = addr_t->instance_id();
1180 if (!addr_t->is_known_instance() &&
1181 addr_t->is_ptr_to_boxed_value()) {
1182 // Use _idx of address base (could be Phi node) for boxed values.
1183 intptr_t ignore = 0;
1184 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1185 if (base == NULL) {
1186 return this;
1187 }
1188 this_iid = base->_idx;
1189 }
1190 const Type* this_type = bottom_type();
1191 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1192 Node* phi = region->fast_out(i);
1193 if (phi->is_Phi() && phi != mem &&
1194 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1195 return phi;
1196 }
1197 }
1198 }
1199
1200 return this;
1201 }
1202
1203 // Construct an equivalent unsigned load.
1204 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1205 BasicType bt = T_ILLEGAL;
1206 const Type* rt = NULL;
1207 switch (Opcode()) {
1208 case Op_LoadUB: return this;
1209 case Op_LoadUS: return this;
1210 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1211 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1212 default:
1213 assert(false, "no unsigned variant: %s", Name());
1214 return NULL;
1215 }
1216 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1217 raw_adr_type(), rt, bt, _mo, _control_dependency,
1218 is_unaligned_access(), is_mismatched_access());
1219 }
1220
1221 // Construct an equivalent signed load.
1222 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1223 BasicType bt = T_ILLEGAL;
1224 const Type* rt = NULL;
1225 switch (Opcode()) {
1226 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1227 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1228 case Op_LoadB: // fall through
1229 case Op_LoadS: // fall through
1230 case Op_LoadI: // fall through
1231 case Op_LoadL: return this;
1232 default:
1233 assert(false, "no signed variant: %s", Name());
1234 return NULL;
1235 }
1236 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1237 raw_adr_type(), rt, bt, _mo, _control_dependency,
1238 is_unaligned_access(), is_mismatched_access());
1239 }
1240
1241 // We're loading from an object which has autobox behaviour.
1242 // If this object is result of a valueOf call we'll have a phi
1243 // merging a newly allocated object and a load from the cache.
1244 // We want to replace this load with the original incoming
1245 // argument to the valueOf call.
1246 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1247 assert(phase->C->eliminate_boxing(), "sanity");
1248 intptr_t ignore = 0;
1249 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1250 if ((base == NULL) || base->is_Phi()) {
1251 // Push the loads from the phi that comes from valueOf up
1252 // through it to allow elimination of the loads and the recovery
1253 // of the original value. It is done in split_through_phi().
1254 return NULL;
1255 } else if (base->is_Load() ||
1256 (base->is_DecodeN() && base->in(1)->is_Load())) {
1257 // Eliminate the load of boxed value for integer types from the cache
1258 // array by deriving the value from the index into the array.
1259 // Capture the offset of the load and then reverse the computation.
1260
1261 // Get LoadN node which loads a boxing object from 'cache' array.
1262 if (base->is_DecodeN()) {
1263 base = base->in(1);
1264 }
1265 if (!base->in(Address)->is_AddP()) {
1266 return NULL; // Complex address
1267 }
1268 AddPNode* address = base->in(Address)->as_AddP();
1269 Node* cache_base = address->in(AddPNode::Base);
1270 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1271 // Get ConP node which is static 'cache' field.
1272 cache_base = cache_base->in(1);
1273 }
1274 if ((cache_base != NULL) && cache_base->is_Con()) {
1275 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1276 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1277 Node* elements[4];
1278 int shift = exact_log2(type2aelembytes(T_OBJECT));
1279 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1280 if (count > 0 && elements[0]->is_Con() &&
1281 (count == 1 ||
1282 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1283 elements[1]->in(2) == phase->intcon(shift)))) {
1284 ciObjArray* array = base_type->const_oop()->as_obj_array();
1285 // Fetch the box object cache[0] at the base of the array and get its value
1286 ciInstance* box = array->obj_at(0)->as_instance();
1287 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1288 assert(ik->is_box_klass(), "sanity");
1289 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1290 if (ik->nof_nonstatic_fields() == 1) {
1291 // This should be true nonstatic_field_at requires calling
1292 // nof_nonstatic_fields so check it anyway
1293 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1294 BasicType bt = c.basic_type();
1295 // Only integer types have boxing cache.
1296 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1297 bt == T_BYTE || bt == T_SHORT ||
1298 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1299 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1300 if (cache_low != (int)cache_low) {
1301 return NULL; // should not happen since cache is array indexed by value
1302 }
1303 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1304 if (offset != (int)offset) {
1305 return NULL; // should not happen since cache is array indexed by value
1306 }
1307 // Add up all the offsets making of the address of the load
1308 Node* result = elements[0];
1309 for (int i = 1; i < count; i++) {
1310 result = phase->transform(new AddXNode(result, elements[i]));
1311 }
1312 // Remove the constant offset from the address and then
1313 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset)));
1314 // remove the scaling of the offset to recover the original index.
1315 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1316 // Peel the shift off directly but wrap it in a dummy node
1317 // since Ideal can't return existing nodes
1318 result = new RShiftXNode(result->in(1), phase->intcon(0));
1319 } else if (result->is_Add() && result->in(2)->is_Con() &&
1320 result->in(1)->Opcode() == Op_LShiftX &&
1321 result->in(1)->in(2) == phase->intcon(shift)) {
1322 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1323 // but for boxing cache access we know that X<<Z will not overflow
1324 // (there is range check) so we do this optimizatrion by hand here.
1325 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift));
1326 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con));
1327 } else {
1328 result = new RShiftXNode(result, phase->intcon(shift));
1329 }
1330 #ifdef _LP64
1331 if (bt != T_LONG) {
1332 result = new ConvL2INode(phase->transform(result));
1333 }
1334 #else
1335 if (bt == T_LONG) {
1336 result = new ConvI2LNode(phase->transform(result));
1337 }
1338 #endif
1339 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1340 // Need to preserve unboxing load type if it is unsigned.
1341 switch(this->Opcode()) {
1342 case Op_LoadUB:
1343 result = new AndINode(phase->transform(result), phase->intcon(0xFF));
1344 break;
1345 case Op_LoadUS:
1346 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF));
1347 break;
1348 }
1349 return result;
1350 }
1351 }
1352 }
1353 }
1354 }
1355 return NULL;
1356 }
1357
1358 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1359 Node* region = phi->in(0);
1360 if (region == NULL) {
1361 return false; // Wait stable graph
1362 }
1363 uint cnt = phi->req();
1364 for (uint i = 1; i < cnt; i++) {
1365 Node* rc = region->in(i);
1366 if (rc == NULL || phase->type(rc) == Type::TOP)
1367 return false; // Wait stable graph
1368 Node* in = phi->in(i);
1369 if (in == NULL || phase->type(in) == Type::TOP)
1370 return false; // Wait stable graph
1371 }
1372 return true;
1373 }
1374 //------------------------------split_through_phi------------------------------
1375 // Split instance or boxed field load through Phi.
1376 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1377 Node* mem = in(Memory);
1378 Node* address = in(Address);
1379 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1380
1381 assert((t_oop != NULL) &&
1382 (t_oop->is_known_instance_field() ||
1383 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1384
1385 Compile* C = phase->C;
1386 intptr_t ignore = 0;
1387 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1388 bool base_is_phi = (base != NULL) && base->is_Phi();
1389 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1390 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1391 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1392
1393 if (!((mem->is_Phi() || base_is_phi) &&
1394 (load_boxed_values || t_oop->is_known_instance_field()))) {
1395 return NULL; // memory is not Phi
1396 }
1397
1398 if (mem->is_Phi()) {
1399 if (!stable_phi(mem->as_Phi(), phase)) {
1400 return NULL; // Wait stable graph
1401 }
1402 uint cnt = mem->req();
1403 // Check for loop invariant memory.
1404 if (cnt == 3) {
1405 for (uint i = 1; i < cnt; i++) {
1406 Node* in = mem->in(i);
1407 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1408 if (m == mem) {
1409 if (i == 1) {
1410 // if the first edge was a loop, check second edge too.
1411 // If both are replaceable - we are in an infinite loop
1412 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1413 if (n == mem) {
1414 break;
1415 }
1416 }
1417 set_req(Memory, mem->in(cnt - i));
1418 return this; // made change
1419 }
1420 }
1421 }
1422 }
1423 if (base_is_phi) {
1424 if (!stable_phi(base->as_Phi(), phase)) {
1425 return NULL; // Wait stable graph
1426 }
1427 uint cnt = base->req();
1428 // Check for loop invariant memory.
1429 if (cnt == 3) {
1430 for (uint i = 1; i < cnt; i++) {
1431 if (base->in(i) == base) {
1432 return NULL; // Wait stable graph
1433 }
1434 }
1435 }
1436 }
1437
1438 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1439
1440 // Split through Phi (see original code in loopopts.cpp).
1441 assert(C->have_alias_type(t_oop), "instance should have alias type");
1442
1443 // Do nothing here if Identity will find a value
1444 // (to avoid infinite chain of value phis generation).
1445 if (!phase->eqv(this, phase->apply_identity(this)))
1446 return NULL;
1447
1448 // Select Region to split through.
1449 Node* region;
1450 if (!base_is_phi) {
1451 assert(mem->is_Phi(), "sanity");
1452 region = mem->in(0);
1453 // Skip if the region dominates some control edge of the address.
1454 if (!MemNode::all_controls_dominate(address, region))
1455 return NULL;
1456 } else if (!mem->is_Phi()) {
1457 assert(base_is_phi, "sanity");
1458 region = base->in(0);
1459 // Skip if the region dominates some control edge of the memory.
1460 if (!MemNode::all_controls_dominate(mem, region))
1461 return NULL;
1462 } else if (base->in(0) != mem->in(0)) {
1463 assert(base_is_phi && mem->is_Phi(), "sanity");
1464 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1465 region = base->in(0);
1466 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1467 region = mem->in(0);
1468 } else {
1469 return NULL; // complex graph
1470 }
1471 } else {
1472 assert(base->in(0) == mem->in(0), "sanity");
1473 region = mem->in(0);
1474 }
1475
1476 const Type* this_type = this->bottom_type();
1477 int this_index = C->get_alias_index(t_oop);
1478 int this_offset = t_oop->offset();
1479 int this_iid = t_oop->instance_id();
1480 if (!t_oop->is_known_instance() && load_boxed_values) {
1481 // Use _idx of address base for boxed values.
1482 this_iid = base->_idx;
1483 }
1484 PhaseIterGVN* igvn = phase->is_IterGVN();
1485 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1486 for (uint i = 1; i < region->req(); i++) {
1487 Node* x;
1488 Node* the_clone = NULL;
1489 if (region->in(i) == C->top()) {
1490 x = C->top(); // Dead path? Use a dead data op
1491 } else {
1492 x = this->clone(); // Else clone up the data op
1493 the_clone = x; // Remember for possible deletion.
1494 // Alter data node to use pre-phi inputs
1495 if (this->in(0) == region) {
1496 x->set_req(0, region->in(i));
1497 } else {
1498 x->set_req(0, NULL);
1499 }
1500 if (mem->is_Phi() && (mem->in(0) == region)) {
1501 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1502 }
1503 if (address->is_Phi() && address->in(0) == region) {
1504 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1505 }
1506 if (base_is_phi && (base->in(0) == region)) {
1507 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1508 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1509 x->set_req(Address, adr_x);
1510 }
1511 }
1512 // Check for a 'win' on some paths
1513 const Type *t = x->Value(igvn);
1514
1515 bool singleton = t->singleton();
1516
1517 // See comments in PhaseIdealLoop::split_thru_phi().
1518 if (singleton && t == Type::TOP) {
1519 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1520 }
1521
1522 if (singleton) {
1523 x = igvn->makecon(t);
1524 } else {
1525 // We now call Identity to try to simplify the cloned node.
1526 // Note that some Identity methods call phase->type(this).
1527 // Make sure that the type array is big enough for
1528 // our new node, even though we may throw the node away.
1529 // (This tweaking with igvn only works because x is a new node.)
1530 igvn->set_type(x, t);
1531 // If x is a TypeNode, capture any more-precise type permanently into Node
1532 // otherwise it will be not updated during igvn->transform since
1533 // igvn->type(x) is set to x->Value() already.
1534 x->raise_bottom_type(t);
1535 Node *y = igvn->apply_identity(x);
1536 if (y != x) {
1537 x = y;
1538 } else {
1539 y = igvn->hash_find_insert(x);
1540 if (y) {
1541 x = y;
1542 } else {
1543 // Else x is a new node we are keeping
1544 // We do not need register_new_node_with_optimizer
1545 // because set_type has already been called.
1546 igvn->_worklist.push(x);
1547 }
1548 }
1549 }
1550 if (x != the_clone && the_clone != NULL) {
1551 igvn->remove_dead_node(the_clone);
1552 }
1553 phi->set_req(i, x);
1554 }
1555 // Record Phi
1556 igvn->register_new_node_with_optimizer(phi);
1557 return phi;
1558 }
1559
1560 //------------------------------Ideal------------------------------------------
1561 // If the load is from Field memory and the pointer is non-null, it might be possible to
1562 // zero out the control input.
1563 // If the offset is constant and the base is an object allocation,
1564 // try to hook me up to the exact initializing store.
1565 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1566 Node* p = MemNode::Ideal_common(phase, can_reshape);
1567 if (p) return (p == NodeSentinel) ? NULL : p;
1568
1569 Node* ctrl = in(MemNode::Control);
1570 Node* address = in(MemNode::Address);
1571 bool progress = false;
1572
1573 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1574 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1575
1576 // Skip up past a SafePoint control. Cannot do this for Stores because
1577 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1578 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1579 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1580 !addr_mark ) {
1581 ctrl = ctrl->in(0);
1582 set_req(MemNode::Control,ctrl);
1583 progress = true;
1584 }
1585
1586 intptr_t ignore = 0;
1587 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1588 if (base != NULL
1589 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1590 // Check for useless control edge in some common special cases
1591 if (in(MemNode::Control) != NULL
1592 && can_remove_control()
1593 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1594 && all_controls_dominate(base, phase->C->start())) {
1595 // A method-invariant, non-null address (constant or 'this' argument).
1596 set_req(MemNode::Control, NULL);
1597 progress = true;
1598 }
1599 }
1600
1601 Node* mem = in(MemNode::Memory);
1602 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1603
1604 if (can_reshape && (addr_t != NULL)) {
1605 // try to optimize our memory input
1606 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1607 if (opt_mem != mem) {
1608 set_req(MemNode::Memory, opt_mem);
1609 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1610 return this;
1611 }
1612 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1613 if ((t_oop != NULL) &&
1614 (t_oop->is_known_instance_field() ||
1615 t_oop->is_ptr_to_boxed_value())) {
1616 PhaseIterGVN *igvn = phase->is_IterGVN();
1617 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1618 // Delay this transformation until memory Phi is processed.
1619 phase->is_IterGVN()->_worklist.push(this);
1620 return NULL;
1621 }
1622 // Split instance field load through Phi.
1623 Node* result = split_through_phi(phase);
1624 if (result != NULL) return result;
1625
1626 if (t_oop->is_ptr_to_boxed_value()) {
1627 Node* result = eliminate_autobox(phase);
1628 if (result != NULL) return result;
1629 }
1630 }
1631 }
1632
1633 // Is there a dominating load that loads the same value? Leave
1634 // anything that is not a load of a field/array element (like
1635 // barriers etc.) alone
1636 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) {
1637 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1638 Node *use = mem->fast_out(i);
1639 if (use != this &&
1640 use->Opcode() == Opcode() &&
1641 use->in(0) != NULL &&
1642 use->in(0) != in(0) &&
1643 use->in(Address) == in(Address)) {
1644 Node* ctl = in(0);
1645 for (int i = 0; i < 10 && ctl != NULL; i++) {
1646 ctl = IfNode::up_one_dom(ctl);
1647 if (ctl == use->in(0)) {
1648 set_req(0, use->in(0));
1649 return this;
1650 }
1651 }
1652 }
1653 }
1654 }
1655
1656 // Check for prior store with a different base or offset; make Load
1657 // independent. Skip through any number of them. Bail out if the stores
1658 // are in an endless dead cycle and report no progress. This is a key
1659 // transform for Reflection. However, if after skipping through the Stores
1660 // we can't then fold up against a prior store do NOT do the transform as
1661 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1662 // array memory alive twice: once for the hoisted Load and again after the
1663 // bypassed Store. This situation only works if EVERYBODY who does
1664 // anti-dependence work knows how to bypass. I.e. we need all
1665 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1666 // the alias index stuff. So instead, peek through Stores and IFF we can
1667 // fold up, do so.
1668 Node* prev_mem = find_previous_store(phase);
1669 if (prev_mem != NULL) {
1670 Node* value = can_see_arraycopy_value(prev_mem, phase);
1671 if (value != NULL) {
1672 return value;
1673 }
1674 }
1675 // Steps (a), (b): Walk past independent stores to find an exact match.
1676 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1677 // (c) See if we can fold up on the spot, but don't fold up here.
1678 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1679 // just return a prior value, which is done by Identity calls.
1680 if (can_see_stored_value(prev_mem, phase)) {
1681 // Make ready for step (d):
1682 set_req(MemNode::Memory, prev_mem);
1683 return this;
1684 }
1685 }
1686
1687 return progress ? this : NULL;
1688 }
1689
1690 // Helper to recognize certain Klass fields which are invariant across
1691 // some group of array types (e.g., int[] or all T[] where T < Object).
1692 const Type*
1693 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1694 ciKlass* klass) const {
1695 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1696 // The field is Klass::_modifier_flags. Return its (constant) value.
1697 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1698 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1699 return TypeInt::make(klass->modifier_flags());
1700 }
1701 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1702 // The field is Klass::_access_flags. Return its (constant) value.
1703 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1704 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1705 return TypeInt::make(klass->access_flags());
1706 }
1707 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1708 // The field is Klass::_layout_helper. Return its constant value if known.
1709 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1710 return TypeInt::make(klass->layout_helper());
1711 }
1712
1713 // No match.
1714 return NULL;
1715 }
1716
1717 //------------------------------Value-----------------------------------------
1718 const Type* LoadNode::Value(PhaseGVN* phase) const {
1719 // Either input is TOP ==> the result is TOP
1720 Node* mem = in(MemNode::Memory);
1721 const Type *t1 = phase->type(mem);
1722 if (t1 == Type::TOP) return Type::TOP;
1723 Node* adr = in(MemNode::Address);
1724 const TypePtr* tp = phase->type(adr)->isa_ptr();
1725 if (tp == NULL || tp->empty()) return Type::TOP;
1726 int off = tp->offset();
1727 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1728 Compile* C = phase->C;
1729
1730 // Try to guess loaded type from pointer type
1731 if (tp->isa_aryptr()) {
1732 const TypeAryPtr* ary = tp->is_aryptr();
1733 const Type* t = ary->elem();
1734
1735 // Determine whether the reference is beyond the header or not, by comparing
1736 // the offset against the offset of the start of the array's data.
1737 // Different array types begin at slightly different offsets (12 vs. 16).
1738 // We choose T_BYTE as an example base type that is least restrictive
1739 // as to alignment, which will therefore produce the smallest
1740 // possible base offset.
1741 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1742 const bool off_beyond_header = (off >= min_base_off);
1743
1744 // Try to constant-fold a stable array element.
1745 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1746 // Make sure the reference is not into the header and the offset is constant
1747 ciObject* aobj = ary->const_oop();
1748 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1749 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1750 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1751 stable_dimension,
1752 memory_type(), is_unsigned());
1753 if (con_type != NULL) {
1754 return con_type;
1755 }
1756 }
1757 }
1758
1759 // Don't do this for integer types. There is only potential profit if
1760 // the element type t is lower than _type; that is, for int types, if _type is
1761 // more restrictive than t. This only happens here if one is short and the other
1762 // char (both 16 bits), and in those cases we've made an intentional decision
1763 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1764 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1765 //
1766 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1767 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1768 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1769 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1770 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1771 // In fact, that could have been the original type of p1, and p1 could have
1772 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1773 // expression (LShiftL quux 3) independently optimized to the constant 8.
1774 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1775 && (_type->isa_vect() == NULL)
1776 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1777 // t might actually be lower than _type, if _type is a unique
1778 // concrete subclass of abstract class t.
1779 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
1780 const Type* jt = t->join_speculative(_type);
1781 // In any case, do not allow the join, per se, to empty out the type.
1782 if (jt->empty() && !t->empty()) {
1783 // This can happen if a interface-typed array narrows to a class type.
1784 jt = _type;
1785 }
1786 #ifdef ASSERT
1787 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1788 // The pointers in the autobox arrays are always non-null
1789 Node* base = adr->in(AddPNode::Base);
1790 if ((base != NULL) && base->is_DecodeN()) {
1791 // Get LoadN node which loads IntegerCache.cache field
1792 base = base->in(1);
1793 }
1794 if ((base != NULL) && base->is_Con()) {
1795 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1796 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1797 // It could be narrow oop
1798 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1799 }
1800 }
1801 }
1802 #endif
1803 return jt;
1804 }
1805 }
1806 } else if (tp->base() == Type::InstPtr) {
1807 assert( off != Type::OffsetBot ||
1808 // arrays can be cast to Objects
1809 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1810 // unsafe field access may not have a constant offset
1811 C->has_unsafe_access(),
1812 "Field accesses must be precise" );
1813 // For oop loads, we expect the _type to be precise.
1814
1815 // Optimize loads from constant fields.
1816 const TypeInstPtr* tinst = tp->is_instptr();
1817 ciObject* const_oop = tinst->const_oop();
1818 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1819 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
1820 if (con_type != NULL) {
1821 return con_type;
1822 }
1823 }
1824 } else if (tp->base() == Type::KlassPtr) {
1825 assert( off != Type::OffsetBot ||
1826 // arrays can be cast to Objects
1827 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1828 // also allow array-loading from the primary supertype
1829 // array during subtype checks
1830 Opcode() == Op_LoadKlass,
1831 "Field accesses must be precise" );
1832 // For klass/static loads, we expect the _type to be precise
1833 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
1834 /* With mirrors being an indirect in the Klass*
1835 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
1836 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
1837 *
1838 * So check the type and klass of the node before the LoadP.
1839 */
1840 Node* adr2 = adr->in(MemNode::Address);
1841 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1842 if (tkls != NULL && !StressReflectiveCode) {
1843 ciKlass* klass = tkls->klass();
1844 if (klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1845 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1846 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1847 return TypeInstPtr::make(klass->java_mirror());
1848 }
1849 }
1850 }
1851
1852 const TypeKlassPtr *tkls = tp->isa_klassptr();
1853 if (tkls != NULL && !StressReflectiveCode) {
1854 ciKlass* klass = tkls->klass();
1855 if (klass->is_loaded() && tkls->klass_is_exact()) {
1856 // We are loading a field from a Klass metaobject whose identity
1857 // is known at compile time (the type is "exact" or "precise").
1858 // Check for fields we know are maintained as constants by the VM.
1859 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1860 // The field is Klass::_super_check_offset. Return its (constant) value.
1861 // (Folds up type checking code.)
1862 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1863 return TypeInt::make(klass->super_check_offset());
1864 }
1865 // Compute index into primary_supers array
1866 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1867 // Check for overflowing; use unsigned compare to handle the negative case.
1868 if( depth < ciKlass::primary_super_limit() ) {
1869 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1870 // (Folds up type checking code.)
1871 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1872 ciKlass *ss = klass->super_of_depth(depth);
1873 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1874 }
1875 const Type* aift = load_array_final_field(tkls, klass);
1876 if (aift != NULL) return aift;
1877 }
1878
1879 // We can still check if we are loading from the primary_supers array at a
1880 // shallow enough depth. Even though the klass is not exact, entries less
1881 // than or equal to its super depth are correct.
1882 if (klass->is_loaded() ) {
1883 ciType *inner = klass;
1884 while( inner->is_obj_array_klass() )
1885 inner = inner->as_obj_array_klass()->base_element_type();
1886 if( inner->is_instance_klass() &&
1887 !inner->as_instance_klass()->flags().is_interface() ) {
1888 // Compute index into primary_supers array
1889 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1890 // Check for overflowing; use unsigned compare to handle the negative case.
1891 if( depth < ciKlass::primary_super_limit() &&
1892 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1893 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1894 // (Folds up type checking code.)
1895 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1896 ciKlass *ss = klass->super_of_depth(depth);
1897 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1898 }
1899 }
1900 }
1901
1902 // If the type is enough to determine that the thing is not an array,
1903 // we can give the layout_helper a positive interval type.
1904 // This will help short-circuit some reflective code.
1905 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1906 && !klass->is_array_klass() // not directly typed as an array
1907 && !klass->is_interface() // specifically not Serializable & Cloneable
1908 && !klass->is_java_lang_Object() // not the supertype of all T[]
1909 ) {
1910 // Note: When interfaces are reliable, we can narrow the interface
1911 // test to (klass != Serializable && klass != Cloneable).
1912 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1913 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1914 // The key property of this type is that it folds up tests
1915 // for array-ness, since it proves that the layout_helper is positive.
1916 // Thus, a generic value like the basic object layout helper works fine.
1917 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1918 }
1919 }
1920
1921 // If we are loading from a freshly-allocated object, produce a zero,
1922 // if the load is provably beyond the header of the object.
1923 // (Also allow a variable load from a fresh array to produce zero.)
1924 const TypeOopPtr *tinst = tp->isa_oopptr();
1925 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1926 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1927 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1928 Node* value = can_see_stored_value(mem,phase);
1929 if (value != NULL && value->is_Con()) {
1930 assert(value->bottom_type()->higher_equal(_type),"sanity");
1931 return value->bottom_type();
1932 }
1933 }
1934
1935 if (is_instance) {
1936 // If we have an instance type and our memory input is the
1937 // programs's initial memory state, there is no matching store,
1938 // so just return a zero of the appropriate type
1939 Node *mem = in(MemNode::Memory);
1940 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1941 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1942 return Type::get_zero_type(_type->basic_type());
1943 }
1944 }
1945 return _type;
1946 }
1947
1948 //------------------------------match_edge-------------------------------------
1949 // Do we Match on this edge index or not? Match only the address.
1950 uint LoadNode::match_edge(uint idx) const {
1951 return idx == MemNode::Address;
1952 }
1953
1954 //--------------------------LoadBNode::Ideal--------------------------------------
1955 //
1956 // If the previous store is to the same address as this load,
1957 // and the value stored was larger than a byte, replace this load
1958 // with the value stored truncated to a byte. If no truncation is
1959 // needed, the replacement is done in LoadNode::Identity().
1960 //
1961 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1962 Node* mem = in(MemNode::Memory);
1963 Node* value = can_see_stored_value(mem,phase);
1964 if( value && !phase->type(value)->higher_equal( _type ) ) {
1965 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
1966 return new RShiftINode(result, phase->intcon(24));
1967 }
1968 // Identity call will handle the case where truncation is not needed.
1969 return LoadNode::Ideal(phase, can_reshape);
1970 }
1971
1972 const Type* LoadBNode::Value(PhaseGVN* phase) const {
1973 Node* mem = in(MemNode::Memory);
1974 Node* value = can_see_stored_value(mem,phase);
1975 if (value != NULL && value->is_Con() &&
1976 !value->bottom_type()->higher_equal(_type)) {
1977 // If the input to the store does not fit with the load's result type,
1978 // it must be truncated. We can't delay until Ideal call since
1979 // a singleton Value is needed for split_thru_phi optimization.
1980 int con = value->get_int();
1981 return TypeInt::make((con << 24) >> 24);
1982 }
1983 return LoadNode::Value(phase);
1984 }
1985
1986 //--------------------------LoadUBNode::Ideal-------------------------------------
1987 //
1988 // If the previous store is to the same address as this load,
1989 // and the value stored was larger than a byte, replace this load
1990 // with the value stored truncated to a byte. If no truncation is
1991 // needed, the replacement is done in LoadNode::Identity().
1992 //
1993 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1994 Node* mem = in(MemNode::Memory);
1995 Node* value = can_see_stored_value(mem, phase);
1996 if (value && !phase->type(value)->higher_equal(_type))
1997 return new AndINode(value, phase->intcon(0xFF));
1998 // Identity call will handle the case where truncation is not needed.
1999 return LoadNode::Ideal(phase, can_reshape);
2000 }
2001
2002 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2003 Node* mem = in(MemNode::Memory);
2004 Node* value = can_see_stored_value(mem,phase);
2005 if (value != NULL && value->is_Con() &&
2006 !value->bottom_type()->higher_equal(_type)) {
2007 // If the input to the store does not fit with the load's result type,
2008 // it must be truncated. We can't delay until Ideal call since
2009 // a singleton Value is needed for split_thru_phi optimization.
2010 int con = value->get_int();
2011 return TypeInt::make(con & 0xFF);
2012 }
2013 return LoadNode::Value(phase);
2014 }
2015
2016 //--------------------------LoadUSNode::Ideal-------------------------------------
2017 //
2018 // If the previous store is to the same address as this load,
2019 // and the value stored was larger than a char, replace this load
2020 // with the value stored truncated to a char. If no truncation is
2021 // needed, the replacement is done in LoadNode::Identity().
2022 //
2023 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2024 Node* mem = in(MemNode::Memory);
2025 Node* value = can_see_stored_value(mem,phase);
2026 if( value && !phase->type(value)->higher_equal( _type ) )
2027 return new AndINode(value,phase->intcon(0xFFFF));
2028 // Identity call will handle the case where truncation is not needed.
2029 return LoadNode::Ideal(phase, can_reshape);
2030 }
2031
2032 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2033 Node* mem = in(MemNode::Memory);
2034 Node* value = can_see_stored_value(mem,phase);
2035 if (value != NULL && value->is_Con() &&
2036 !value->bottom_type()->higher_equal(_type)) {
2037 // If the input to the store does not fit with the load's result type,
2038 // it must be truncated. We can't delay until Ideal call since
2039 // a singleton Value is needed for split_thru_phi optimization.
2040 int con = value->get_int();
2041 return TypeInt::make(con & 0xFFFF);
2042 }
2043 return LoadNode::Value(phase);
2044 }
2045
2046 //--------------------------LoadSNode::Ideal--------------------------------------
2047 //
2048 // If the previous store is to the same address as this load,
2049 // and the value stored was larger than a short, replace this load
2050 // with the value stored truncated to a short. If no truncation is
2051 // needed, the replacement is done in LoadNode::Identity().
2052 //
2053 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2054 Node* mem = in(MemNode::Memory);
2055 Node* value = can_see_stored_value(mem,phase);
2056 if( value && !phase->type(value)->higher_equal( _type ) ) {
2057 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
2058 return new RShiftINode(result, phase->intcon(16));
2059 }
2060 // Identity call will handle the case where truncation is not needed.
2061 return LoadNode::Ideal(phase, can_reshape);
2062 }
2063
2064 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2065 Node* mem = in(MemNode::Memory);
2066 Node* value = can_see_stored_value(mem,phase);
2067 if (value != NULL && value->is_Con() &&
2068 !value->bottom_type()->higher_equal(_type)) {
2069 // If the input to the store does not fit with the load's result type,
2070 // it must be truncated. We can't delay until Ideal call since
2071 // a singleton Value is needed for split_thru_phi optimization.
2072 int con = value->get_int();
2073 return TypeInt::make((con << 16) >> 16);
2074 }
2075 return LoadNode::Value(phase);
2076 }
2077
2078 //=============================================================================
2079 //----------------------------LoadKlassNode::make------------------------------
2080 // Polymorphic factory method:
2081 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2082 // sanity check the alias category against the created node type
2083 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2084 assert(adr_type != NULL, "expecting TypeKlassPtr");
2085 #ifdef _LP64
2086 if (adr_type->is_ptr_to_narrowklass()) {
2087 assert(UseCompressedClassPointers, "no compressed klasses");
2088 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2089 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2090 }
2091 #endif
2092 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2093 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2094 }
2095
2096 //------------------------------Value------------------------------------------
2097 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2098 return klass_value_common(phase);
2099 }
2100
2101 // In most cases, LoadKlassNode does not have the control input set. If the control
2102 // input is set, it must not be removed (by LoadNode::Ideal()).
2103 bool LoadKlassNode::can_remove_control() const {
2104 return false;
2105 }
2106
2107 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2108 // Either input is TOP ==> the result is TOP
2109 const Type *t1 = phase->type( in(MemNode::Memory) );
2110 if (t1 == Type::TOP) return Type::TOP;
2111 Node *adr = in(MemNode::Address);
2112 const Type *t2 = phase->type( adr );
2113 if (t2 == Type::TOP) return Type::TOP;
2114 const TypePtr *tp = t2->is_ptr();
2115 if (TypePtr::above_centerline(tp->ptr()) ||
2116 tp->ptr() == TypePtr::Null) return Type::TOP;
2117
2118 // Return a more precise klass, if possible
2119 const TypeInstPtr *tinst = tp->isa_instptr();
2120 if (tinst != NULL) {
2121 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2122 int offset = tinst->offset();
2123 if (ik == phase->C->env()->Class_klass()
2124 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2125 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2126 // We are loading a special hidden field from a Class mirror object,
2127 // the field which points to the VM's Klass metaobject.
2128 ciType* t = tinst->java_mirror_type();
2129 // java_mirror_type returns non-null for compile-time Class constants.
2130 if (t != NULL) {
2131 // constant oop => constant klass
2132 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2133 if (t->is_void()) {
2134 // We cannot create a void array. Since void is a primitive type return null
2135 // klass. Users of this result need to do a null check on the returned klass.
2136 return TypePtr::NULL_PTR;
2137 }
2138 return TypeKlassPtr::make(ciArrayKlass::make(t));
2139 }
2140 if (!t->is_klass()) {
2141 // a primitive Class (e.g., int.class) has NULL for a klass field
2142 return TypePtr::NULL_PTR;
2143 }
2144 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2145 return TypeKlassPtr::make(t->as_klass());
2146 }
2147 // non-constant mirror, so we can't tell what's going on
2148 }
2149 if( !ik->is_loaded() )
2150 return _type; // Bail out if not loaded
2151 if (offset == oopDesc::klass_offset_in_bytes()) {
2152 if (tinst->klass_is_exact()) {
2153 return TypeKlassPtr::make(ik);
2154 }
2155 // See if we can become precise: no subklasses and no interface
2156 // (Note: We need to support verified interfaces.)
2157 if (!ik->is_interface() && !ik->has_subklass()) {
2158 //assert(!UseExactTypes, "this code should be useless with exact types");
2159 // Add a dependence; if any subclass added we need to recompile
2160 if (!ik->is_final()) {
2161 // %%% should use stronger assert_unique_concrete_subtype instead
2162 phase->C->dependencies()->assert_leaf_type(ik);
2163 }
2164 // Return precise klass
2165 return TypeKlassPtr::make(ik);
2166 }
2167
2168 // Return root of possible klass
2169 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2170 }
2171 }
2172
2173 // Check for loading klass from an array
2174 const TypeAryPtr *tary = tp->isa_aryptr();
2175 if( tary != NULL ) {
2176 ciKlass *tary_klass = tary->klass();
2177 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2178 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2179 if (tary->klass_is_exact()) {
2180 return TypeKlassPtr::make(tary_klass);
2181 }
2182 ciArrayKlass *ak = tary->klass()->as_array_klass();
2183 // If the klass is an object array, we defer the question to the
2184 // array component klass.
2185 if( ak->is_obj_array_klass() ) {
2186 assert( ak->is_loaded(), "" );
2187 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2188 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2189 ciInstanceKlass* ik = base_k->as_instance_klass();
2190 // See if we can become precise: no subklasses and no interface
2191 if (!ik->is_interface() && !ik->has_subklass()) {
2192 //assert(!UseExactTypes, "this code should be useless with exact types");
2193 // Add a dependence; if any subclass added we need to recompile
2194 if (!ik->is_final()) {
2195 phase->C->dependencies()->assert_leaf_type(ik);
2196 }
2197 // Return precise array klass
2198 return TypeKlassPtr::make(ak);
2199 }
2200 }
2201 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2202 } else { // Found a type-array?
2203 //assert(!UseExactTypes, "this code should be useless with exact types");
2204 assert( ak->is_type_array_klass(), "" );
2205 return TypeKlassPtr::make(ak); // These are always precise
2206 }
2207 }
2208 }
2209
2210 // Check for loading klass from an array klass
2211 const TypeKlassPtr *tkls = tp->isa_klassptr();
2212 if (tkls != NULL && !StressReflectiveCode) {
2213 ciKlass* klass = tkls->klass();
2214 if( !klass->is_loaded() )
2215 return _type; // Bail out if not loaded
2216 if( klass->is_obj_array_klass() &&
2217 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2218 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2219 // // Always returning precise element type is incorrect,
2220 // // e.g., element type could be object and array may contain strings
2221 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2222
2223 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2224 // according to the element type's subclassing.
2225 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2226 }
2227 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2228 tkls->offset() == in_bytes(Klass::super_offset())) {
2229 ciKlass* sup = klass->as_instance_klass()->super();
2230 // The field is Klass::_super. Return its (constant) value.
2231 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2232 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2233 }
2234 }
2235
2236 // Bailout case
2237 return LoadNode::Value(phase);
2238 }
2239
2240 //------------------------------Identity---------------------------------------
2241 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2242 // Also feed through the klass in Allocate(...klass...)._klass.
2243 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2244 return klass_identity_common(phase);
2245 }
2246
2247 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2248 Node* x = LoadNode::Identity(phase);
2249 if (x != this) return x;
2250
2251 // Take apart the address into an oop and and offset.
2252 // Return 'this' if we cannot.
2253 Node* adr = in(MemNode::Address);
2254 intptr_t offset = 0;
2255 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2256 if (base == NULL) return this;
2257 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2258 if (toop == NULL) return this;
2259
2260 // Step over potential GC barrier for OopHandle resolve
2261 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2262 if (bs->is_gc_barrier_node(base)) {
2263 base = bs->step_over_gc_barrier(base);
2264 }
2265
2266 // We can fetch the klass directly through an AllocateNode.
2267 // This works even if the klass is not constant (clone or newArray).
2268 if (offset == oopDesc::klass_offset_in_bytes()) {
2269 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2270 if (allocated_klass != NULL) {
2271 return allocated_klass;
2272 }
2273 }
2274
2275 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2276 // See inline_native_Class_query for occurrences of these patterns.
2277 // Java Example: x.getClass().isAssignableFrom(y)
2278 //
2279 // This improves reflective code, often making the Class
2280 // mirror go completely dead. (Current exception: Class
2281 // mirrors may appear in debug info, but we could clean them out by
2282 // introducing a new debug info operator for Klass.java_mirror).
2283
2284 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2285 && offset == java_lang_Class::klass_offset_in_bytes()) {
2286 if (base->is_Load()) {
2287 Node* base2 = base->in(MemNode::Address);
2288 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2289 Node* adr2 = base2->in(MemNode::Address);
2290 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2291 if (tkls != NULL && !tkls->empty()
2292 && (tkls->klass()->is_instance_klass() ||
2293 tkls->klass()->is_array_klass())
2294 && adr2->is_AddP()
2295 ) {
2296 int mirror_field = in_bytes(Klass::java_mirror_offset());
2297 if (tkls->offset() == mirror_field) {
2298 return adr2->in(AddPNode::Base);
2299 }
2300 }
2301 }
2302 }
2303 }
2304
2305 return this;
2306 }
2307
2308
2309 //------------------------------Value------------------------------------------
2310 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2311 const Type *t = klass_value_common(phase);
2312 if (t == Type::TOP)
2313 return t;
2314
2315 return t->make_narrowklass();
2316 }
2317
2318 //------------------------------Identity---------------------------------------
2319 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2320 // Also feed through the klass in Allocate(...klass...)._klass.
2321 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2322 Node *x = klass_identity_common(phase);
2323
2324 const Type *t = phase->type( x );
2325 if( t == Type::TOP ) return x;
2326 if( t->isa_narrowklass()) return x;
2327 assert (!t->isa_narrowoop(), "no narrow oop here");
2328
2329 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2330 }
2331
2332 //------------------------------Value-----------------------------------------
2333 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2334 // Either input is TOP ==> the result is TOP
2335 const Type *t1 = phase->type( in(MemNode::Memory) );
2336 if( t1 == Type::TOP ) return Type::TOP;
2337 Node *adr = in(MemNode::Address);
2338 const Type *t2 = phase->type( adr );
2339 if( t2 == Type::TOP ) return Type::TOP;
2340 const TypePtr *tp = t2->is_ptr();
2341 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2342 const TypeAryPtr *tap = tp->isa_aryptr();
2343 if( !tap ) return _type;
2344 return tap->size();
2345 }
2346
2347 //-------------------------------Ideal---------------------------------------
2348 // Feed through the length in AllocateArray(...length...)._length.
2349 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2350 Node* p = MemNode::Ideal_common(phase, can_reshape);
2351 if (p) return (p == NodeSentinel) ? NULL : p;
2352
2353 // Take apart the address into an oop and and offset.
2354 // Return 'this' if we cannot.
2355 Node* adr = in(MemNode::Address);
2356 intptr_t offset = 0;
2357 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2358 if (base == NULL) return NULL;
2359 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2360 if (tary == NULL) return NULL;
2361
2362 // We can fetch the length directly through an AllocateArrayNode.
2363 // This works even if the length is not constant (clone or newArray).
2364 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2365 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2366 if (alloc != NULL) {
2367 Node* allocated_length = alloc->Ideal_length();
2368 Node* len = alloc->make_ideal_length(tary, phase);
2369 if (allocated_length != len) {
2370 // New CastII improves on this.
2371 return len;
2372 }
2373 }
2374 }
2375
2376 return NULL;
2377 }
2378
2379 //------------------------------Identity---------------------------------------
2380 // Feed through the length in AllocateArray(...length...)._length.
2381 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2382 Node* x = LoadINode::Identity(phase);
2383 if (x != this) return x;
2384
2385 // Take apart the address into an oop and and offset.
2386 // Return 'this' if we cannot.
2387 Node* adr = in(MemNode::Address);
2388 intptr_t offset = 0;
2389 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2390 if (base == NULL) return this;
2391 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2392 if (tary == NULL) return this;
2393
2394 // We can fetch the length directly through an AllocateArrayNode.
2395 // This works even if the length is not constant (clone or newArray).
2396 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2397 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2398 if (alloc != NULL) {
2399 Node* allocated_length = alloc->Ideal_length();
2400 // Do not allow make_ideal_length to allocate a CastII node.
2401 Node* len = alloc->make_ideal_length(tary, phase, false);
2402 if (allocated_length == len) {
2403 // Return allocated_length only if it would not be improved by a CastII.
2404 return allocated_length;
2405 }
2406 }
2407 }
2408
2409 return this;
2410
2411 }
2412
2413 //=============================================================================
2414 //---------------------------StoreNode::make-----------------------------------
2415 // Polymorphic factory method:
2416 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2417 assert((mo == unordered || mo == release), "unexpected");
2418 Compile* C = gvn.C;
2419 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2420 ctl != NULL, "raw memory operations should have control edge");
2421
2422 switch (bt) {
2423 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2424 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2425 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2426 case T_CHAR:
2427 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2428 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2429 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2430 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2431 case T_METADATA:
2432 case T_ADDRESS:
2433 case T_OBJECT:
2434 #ifdef _LP64
2435 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2436 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2437 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2438 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2439 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2440 adr->bottom_type()->isa_rawptr())) {
2441 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2442 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2443 }
2444 #endif
2445 {
2446 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2447 }
2448 default:
2449 ShouldNotReachHere();
2450 return (StoreNode*)NULL;
2451 }
2452 }
2453
2454 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2455 bool require_atomic = true;
2456 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2457 }
2458
2459 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2460 bool require_atomic = true;
2461 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2462 }
2463
2464
2465 //--------------------------bottom_type----------------------------------------
2466 const Type *StoreNode::bottom_type() const {
2467 return Type::MEMORY;
2468 }
2469
2470 //------------------------------hash-------------------------------------------
2471 uint StoreNode::hash() const {
2472 // unroll addition of interesting fields
2473 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2474
2475 // Since they are not commoned, do not hash them:
2476 return NO_HASH;
2477 }
2478
2479 //------------------------------Ideal------------------------------------------
2480 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2481 // When a store immediately follows a relevant allocation/initialization,
2482 // try to capture it into the initialization, or hoist it above.
2483 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2484 Node* p = MemNode::Ideal_common(phase, can_reshape);
2485 if (p) return (p == NodeSentinel) ? NULL : p;
2486
2487 Node* mem = in(MemNode::Memory);
2488 Node* address = in(MemNode::Address);
2489 // Back-to-back stores to same address? Fold em up. Generally
2490 // unsafe if I have intervening uses... Also disallowed for StoreCM
2491 // since they must follow each StoreP operation. Redundant StoreCMs
2492 // are eliminated just before matching in final_graph_reshape.
2493 {
2494 Node* st = mem;
2495 // If Store 'st' has more than one use, we cannot fold 'st' away.
2496 // For example, 'st' might be the final state at a conditional
2497 // return. Or, 'st' might be used by some node which is live at
2498 // the same time 'st' is live, which might be unschedulable. So,
2499 // require exactly ONE user until such time as we clone 'mem' for
2500 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2501 // true).
2502 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2503 // Looking at a dead closed cycle of memory?
2504 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2505 assert(Opcode() == st->Opcode() ||
2506 st->Opcode() == Op_StoreVector ||
2507 Opcode() == Op_StoreVector ||
2508 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2509 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2510 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2511 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2512 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2513
2514 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2515 st->as_Store()->memory_size() <= this->memory_size()) {
2516 Node* use = st->raw_out(0);
2517 phase->igvn_rehash_node_delayed(use);
2518 if (can_reshape) {
2519 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2520 } else {
2521 // It's OK to do this in the parser, since DU info is always accurate,
2522 // and the parser always refers to nodes via SafePointNode maps.
2523 use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2524 }
2525 return this;
2526 }
2527 st = st->in(MemNode::Memory);
2528 }
2529 }
2530
2531
2532 // Capture an unaliased, unconditional, simple store into an initializer.
2533 // Or, if it is independent of the allocation, hoist it above the allocation.
2534 if (ReduceFieldZeroing && /*can_reshape &&*/
2535 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2536 InitializeNode* init = mem->in(0)->as_Initialize();
2537 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2538 if (offset > 0) {
2539 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2540 // If the InitializeNode captured me, it made a raw copy of me,
2541 // and I need to disappear.
2542 if (moved != NULL) {
2543 // %%% hack to ensure that Ideal returns a new node:
2544 mem = MergeMemNode::make(mem);
2545 return mem; // fold me away
2546 }
2547 }
2548 }
2549
2550 return NULL; // No further progress
2551 }
2552
2553 //------------------------------Value-----------------------------------------
2554 const Type* StoreNode::Value(PhaseGVN* phase) const {
2555 // Either input is TOP ==> the result is TOP
2556 const Type *t1 = phase->type( in(MemNode::Memory) );
2557 if( t1 == Type::TOP ) return Type::TOP;
2558 const Type *t2 = phase->type( in(MemNode::Address) );
2559 if( t2 == Type::TOP ) return Type::TOP;
2560 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2561 if( t3 == Type::TOP ) return Type::TOP;
2562 return Type::MEMORY;
2563 }
2564
2565 //------------------------------Identity---------------------------------------
2566 // Remove redundant stores:
2567 // Store(m, p, Load(m, p)) changes to m.
2568 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2569 Node* StoreNode::Identity(PhaseGVN* phase) {
2570 Node* mem = in(MemNode::Memory);
2571 Node* adr = in(MemNode::Address);
2572 Node* val = in(MemNode::ValueIn);
2573
2574 Node* result = this;
2575
2576 // Load then Store? Then the Store is useless
2577 if (val->is_Load() &&
2578 val->in(MemNode::Address)->eqv_uncast(adr) &&
2579 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2580 val->as_Load()->store_Opcode() == Opcode()) {
2581 result = mem;
2582 }
2583
2584 // Two stores in a row of the same value?
2585 if (result == this &&
2586 mem->is_Store() &&
2587 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2588 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2589 mem->Opcode() == Opcode()) {
2590 result = mem;
2591 }
2592
2593 // Store of zero anywhere into a freshly-allocated object?
2594 // Then the store is useless.
2595 // (It must already have been captured by the InitializeNode.)
2596 if (result == this &&
2597 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2598 // a newly allocated object is already all-zeroes everywhere
2599 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2600 result = mem;
2601 }
2602
2603 if (result == this) {
2604 // the store may also apply to zero-bits in an earlier object
2605 Node* prev_mem = find_previous_store(phase);
2606 // Steps (a), (b): Walk past independent stores to find an exact match.
2607 if (prev_mem != NULL) {
2608 Node* prev_val = can_see_stored_value(prev_mem, phase);
2609 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2610 // prev_val and val might differ by a cast; it would be good
2611 // to keep the more informative of the two.
2612 result = mem;
2613 }
2614 }
2615 }
2616 }
2617
2618 if (result != this && phase->is_IterGVN() != NULL) {
2619 MemBarNode* trailing = trailing_membar();
2620 if (trailing != NULL) {
2621 #ifdef ASSERT
2622 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2623 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects");
2624 #endif
2625 PhaseIterGVN* igvn = phase->is_IterGVN();
2626 trailing->remove(igvn);
2627 }
2628 }
2629
2630 return result;
2631 }
2632
2633 //------------------------------match_edge-------------------------------------
2634 // Do we Match on this edge index or not? Match only memory & value
2635 uint StoreNode::match_edge(uint idx) const {
2636 return idx == MemNode::Address || idx == MemNode::ValueIn;
2637 }
2638
2639 //------------------------------cmp--------------------------------------------
2640 // Do not common stores up together. They generally have to be split
2641 // back up anyways, so do not bother.
2642 bool StoreNode::cmp( const Node &n ) const {
2643 return (&n == this); // Always fail except on self
2644 }
2645
2646 //------------------------------Ideal_masked_input-----------------------------
2647 // Check for a useless mask before a partial-word store
2648 // (StoreB ... (AndI valIn conIa) )
2649 // If (conIa & mask == mask) this simplifies to
2650 // (StoreB ... (valIn) )
2651 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2652 Node *val = in(MemNode::ValueIn);
2653 if( val->Opcode() == Op_AndI ) {
2654 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2655 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2656 set_req(MemNode::ValueIn, val->in(1));
2657 return this;
2658 }
2659 }
2660 return NULL;
2661 }
2662
2663
2664 //------------------------------Ideal_sign_extended_input----------------------
2665 // Check for useless sign-extension before a partial-word store
2666 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2667 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2668 // (StoreB ... (valIn) )
2669 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2670 Node *val = in(MemNode::ValueIn);
2671 if( val->Opcode() == Op_RShiftI ) {
2672 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2673 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2674 Node *shl = val->in(1);
2675 if( shl->Opcode() == Op_LShiftI ) {
2676 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2677 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2678 set_req(MemNode::ValueIn, shl->in(1));
2679 return this;
2680 }
2681 }
2682 }
2683 }
2684 return NULL;
2685 }
2686
2687 //------------------------------value_never_loaded-----------------------------------
2688 // Determine whether there are any possible loads of the value stored.
2689 // For simplicity, we actually check if there are any loads from the
2690 // address stored to, not just for loads of the value stored by this node.
2691 //
2692 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2693 Node *adr = in(Address);
2694 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2695 if (adr_oop == NULL)
2696 return false;
2697 if (!adr_oop->is_known_instance_field())
2698 return false; // if not a distinct instance, there may be aliases of the address
2699 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2700 Node *use = adr->fast_out(i);
2701 if (use->is_Load() || use->is_LoadStore()) {
2702 return false;
2703 }
2704 }
2705 return true;
2706 }
2707
2708 MemBarNode* StoreNode::trailing_membar() const {
2709 if (is_release()) {
2710 MemBarNode* trailing_mb = NULL;
2711 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2712 Node* u = fast_out(i);
2713 if (u->is_MemBar()) {
2714 if (u->as_MemBar()->trailing_store()) {
2715 assert(u->Opcode() == Op_MemBarVolatile, "");
2716 assert(trailing_mb == NULL, "only one");
2717 trailing_mb = u->as_MemBar();
2718 #ifdef ASSERT
2719 Node* leading = u->as_MemBar()->leading_membar();
2720 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2721 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2722 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2723 #endif
2724 } else {
2725 assert(u->as_MemBar()->standalone(), "");
2726 }
2727 }
2728 }
2729 return trailing_mb;
2730 }
2731 return NULL;
2732 }
2733
2734
2735 //=============================================================================
2736 //------------------------------Ideal------------------------------------------
2737 // If the store is from an AND mask that leaves the low bits untouched, then
2738 // we can skip the AND operation. If the store is from a sign-extension
2739 // (a left shift, then right shift) we can skip both.
2740 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2741 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2742 if( progress != NULL ) return progress;
2743
2744 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2745 if( progress != NULL ) return progress;
2746
2747 // Finally check the default case
2748 return StoreNode::Ideal(phase, can_reshape);
2749 }
2750
2751 //=============================================================================
2752 //------------------------------Ideal------------------------------------------
2753 // If the store is from an AND mask that leaves the low bits untouched, then
2754 // we can skip the AND operation
2755 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2756 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2757 if( progress != NULL ) return progress;
2758
2759 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2760 if( progress != NULL ) return progress;
2761
2762 // Finally check the default case
2763 return StoreNode::Ideal(phase, can_reshape);
2764 }
2765
2766 //=============================================================================
2767 //------------------------------Identity---------------------------------------
2768 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2769 // No need to card mark when storing a null ptr
2770 Node* my_store = in(MemNode::OopStore);
2771 if (my_store->is_Store()) {
2772 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2773 if( t1 == TypePtr::NULL_PTR ) {
2774 return in(MemNode::Memory);
2775 }
2776 }
2777 return this;
2778 }
2779
2780 //=============================================================================
2781 //------------------------------Ideal---------------------------------------
2782 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2783 Node* progress = StoreNode::Ideal(phase, can_reshape);
2784 if (progress != NULL) return progress;
2785
2786 Node* my_store = in(MemNode::OopStore);
2787 if (my_store->is_MergeMem()) {
2788 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2789 set_req(MemNode::OopStore, mem);
2790 return this;
2791 }
2792
2793 return NULL;
2794 }
2795
2796 //------------------------------Value-----------------------------------------
2797 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2798 // Either input is TOP ==> the result is TOP
2799 const Type *t = phase->type( in(MemNode::Memory) );
2800 if( t == Type::TOP ) return Type::TOP;
2801 t = phase->type( in(MemNode::Address) );
2802 if( t == Type::TOP ) return Type::TOP;
2803 t = phase->type( in(MemNode::ValueIn) );
2804 if( t == Type::TOP ) return Type::TOP;
2805 // If extra input is TOP ==> the result is TOP
2806 t = phase->type( in(MemNode::OopStore) );
2807 if( t == Type::TOP ) return Type::TOP;
2808
2809 return StoreNode::Value( phase );
2810 }
2811
2812
2813 //=============================================================================
2814 //----------------------------------SCMemProjNode------------------------------
2815 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2816 {
2817 return bottom_type();
2818 }
2819
2820 //=============================================================================
2821 //----------------------------------LoadStoreNode------------------------------
2822 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2823 : Node(required),
2824 _type(rt),
2825 _adr_type(at),
2826 _has_barrier(false)
2827 {
2828 init_req(MemNode::Control, c );
2829 init_req(MemNode::Memory , mem);
2830 init_req(MemNode::Address, adr);
2831 init_req(MemNode::ValueIn, val);
2832 init_class_id(Class_LoadStore);
2833 }
2834
2835 uint LoadStoreNode::ideal_reg() const {
2836 return _type->ideal_reg();
2837 }
2838
2839 bool LoadStoreNode::result_not_used() const {
2840 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2841 Node *x = fast_out(i);
2842 if (x->Opcode() == Op_SCMemProj) continue;
2843 return false;
2844 }
2845 return true;
2846 }
2847
2848 MemBarNode* LoadStoreNode::trailing_membar() const {
2849 MemBarNode* trailing = NULL;
2850 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2851 Node* u = fast_out(i);
2852 if (u->is_MemBar()) {
2853 if (u->as_MemBar()->trailing_load_store()) {
2854 assert(u->Opcode() == Op_MemBarAcquire, "");
2855 assert(trailing == NULL, "only one");
2856 trailing = u->as_MemBar();
2857 #ifdef ASSERT
2858 Node* leading = trailing->leading_membar();
2859 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2860 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
2861 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
2862 #endif
2863 } else {
2864 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
2865 }
2866 }
2867 }
2868
2869 return trailing;
2870 }
2871
2872 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2873
2874 //=============================================================================
2875 //----------------------------------LoadStoreConditionalNode--------------------
2876 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2877 init_req(ExpectedIn, ex );
2878 }
2879
2880 //=============================================================================
2881 //-------------------------------adr_type--------------------------------------
2882 const TypePtr* ClearArrayNode::adr_type() const {
2883 Node *adr = in(3);
2884 if (adr == NULL) return NULL; // node is dead
2885 return MemNode::calculate_adr_type(adr->bottom_type());
2886 }
2887
2888 //------------------------------match_edge-------------------------------------
2889 // Do we Match on this edge index or not? Do not match memory
2890 uint ClearArrayNode::match_edge(uint idx) const {
2891 return idx > 1;
2892 }
2893
2894 //------------------------------Identity---------------------------------------
2895 // Clearing a zero length array does nothing
2896 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
2897 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2898 }
2899
2900 //------------------------------Idealize---------------------------------------
2901 // Clearing a short array is faster with stores
2902 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2903 // Already know this is a large node, do not try to ideal it
2904 if (!IdealizeClearArrayNode || _is_large) return NULL;
2905
2906 const int unit = BytesPerLong;
2907 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2908 if (!t) return NULL;
2909 if (!t->is_con()) return NULL;
2910 intptr_t raw_count = t->get_con();
2911 intptr_t size = raw_count;
2912 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2913 // Clearing nothing uses the Identity call.
2914 // Negative clears are possible on dead ClearArrays
2915 // (see jck test stmt114.stmt11402.val).
2916 if (size <= 0 || size % unit != 0) return NULL;
2917 intptr_t count = size / unit;
2918 // Length too long; communicate this to matchers and assemblers.
2919 // Assemblers are responsible to produce fast hardware clears for it.
2920 if (size > InitArrayShortSize) {
2921 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
2922 }
2923 Node *mem = in(1);
2924 if( phase->type(mem)==Type::TOP ) return NULL;
2925 Node *adr = in(3);
2926 const Type* at = phase->type(adr);
2927 if( at==Type::TOP ) return NULL;
2928 const TypePtr* atp = at->isa_ptr();
2929 // adjust atp to be the correct array element address type
2930 if (atp == NULL) atp = TypePtr::BOTTOM;
2931 else atp = atp->add_offset(Type::OffsetBot);
2932 // Get base for derived pointer purposes
2933 if( adr->Opcode() != Op_AddP ) Unimplemented();
2934 Node *base = adr->in(1);
2935
2936 Node *zero = phase->makecon(TypeLong::ZERO);
2937 Node *off = phase->MakeConX(BytesPerLong);
2938 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2939 count--;
2940 while( count-- ) {
2941 mem = phase->transform(mem);
2942 adr = phase->transform(new AddPNode(base,adr,off));
2943 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2944 }
2945 return mem;
2946 }
2947
2948 //----------------------------step_through----------------------------------
2949 // Return allocation input memory edge if it is different instance
2950 // or itself if it is the one we are looking for.
2951 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2952 Node* n = *np;
2953 assert(n->is_ClearArray(), "sanity");
2954 intptr_t offset;
2955 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2956 // This method is called only before Allocate nodes are expanded
2957 // during macro nodes expansion. Before that ClearArray nodes are
2958 // only generated in PhaseMacroExpand::generate_arraycopy() (before
2959 // Allocate nodes are expanded) which follows allocations.
2960 assert(alloc != NULL, "should have allocation");
2961 if (alloc->_idx == instance_id) {
2962 // Can not bypass initialization of the instance we are looking for.
2963 return false;
2964 }
2965 // Otherwise skip it.
2966 InitializeNode* init = alloc->initialization();
2967 if (init != NULL)
2968 *np = init->in(TypeFunc::Memory);
2969 else
2970 *np = alloc->in(TypeFunc::Memory);
2971 return true;
2972 }
2973
2974 //----------------------------clear_memory-------------------------------------
2975 // Generate code to initialize object storage to zero.
2976 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2977 intptr_t start_offset,
2978 Node* end_offset,
2979 PhaseGVN* phase) {
2980 intptr_t offset = start_offset;
2981
2982 int unit = BytesPerLong;
2983 if ((offset % unit) != 0) {
2984 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
2985 adr = phase->transform(adr);
2986 const TypePtr* atp = TypeRawPtr::BOTTOM;
2987 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2988 mem = phase->transform(mem);
2989 offset += BytesPerInt;
2990 }
2991 assert((offset % unit) == 0, "");
2992
2993 // Initialize the remaining stuff, if any, with a ClearArray.
2994 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2995 }
2996
2997 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2998 Node* start_offset,
2999 Node* end_offset,
3000 PhaseGVN* phase) {
3001 if (start_offset == end_offset) {
3002 // nothing to do
3003 return mem;
3004 }
3005
3006 int unit = BytesPerLong;
3007 Node* zbase = start_offset;
3008 Node* zend = end_offset;
3009
3010 // Scale to the unit required by the CPU:
3011 if (!Matcher::init_array_count_is_in_bytes) {
3012 Node* shift = phase->intcon(exact_log2(unit));
3013 zbase = phase->transform(new URShiftXNode(zbase, shift) );
3014 zend = phase->transform(new URShiftXNode(zend, shift) );
3015 }
3016
3017 // Bulk clear double-words
3018 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
3019 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
3020 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
3021 return phase->transform(mem);
3022 }
3023
3024 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3025 intptr_t start_offset,
3026 intptr_t end_offset,
3027 PhaseGVN* phase) {
3028 if (start_offset == end_offset) {
3029 // nothing to do
3030 return mem;
3031 }
3032
3033 assert((end_offset % BytesPerInt) == 0, "odd end offset");
3034 intptr_t done_offset = end_offset;
3035 if ((done_offset % BytesPerLong) != 0) {
3036 done_offset -= BytesPerInt;
3037 }
3038 if (done_offset > start_offset) {
3039 mem = clear_memory(ctl, mem, dest,
3040 start_offset, phase->MakeConX(done_offset), phase);
3041 }
3042 if (done_offset < end_offset) { // emit the final 32-bit store
3043 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3044 adr = phase->transform(adr);
3045 const TypePtr* atp = TypeRawPtr::BOTTOM;
3046 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3047 mem = phase->transform(mem);
3048 done_offset += BytesPerInt;
3049 }
3050 assert(done_offset == end_offset, "");
3051 return mem;
3052 }
3053
3054 //=============================================================================
3055 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3056 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
3057 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3058 #ifdef ASSERT
3059 , _pair_idx(0)
3060 #endif
3061 {
3062 init_class_id(Class_MemBar);
3063 Node* top = C->top();
3064 init_req(TypeFunc::I_O,top);
3065 init_req(TypeFunc::FramePtr,top);
3066 init_req(TypeFunc::ReturnAdr,top);
3067 if (precedent != NULL)
3068 init_req(TypeFunc::Parms, precedent);
3069 }
3070
3071 //------------------------------cmp--------------------------------------------
3072 uint MemBarNode::hash() const { return NO_HASH; }
3073 bool MemBarNode::cmp( const Node &n ) const {
3074 return (&n == this); // Always fail except on self
3075 }
3076
3077 //------------------------------make-------------------------------------------
3078 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3079 switch (opcode) {
3080 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
3081 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
3082 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
3083 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
3084 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3085 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3086 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
3087 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
3088 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
3089 case Op_Initialize: return new InitializeNode(C, atp, pn);
3090 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
3091 default: ShouldNotReachHere(); return NULL;
3092 }
3093 }
3094
3095 void MemBarNode::remove(PhaseIterGVN *igvn) {
3096 if (outcnt() != 2) {
3097 return;
3098 }
3099 if (trailing_store() || trailing_load_store()) {
3100 MemBarNode* leading = leading_membar();
3101 if (leading != NULL) {
3102 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3103 leading->remove(igvn);
3104 }
3105 }
3106 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3107 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3108 }
3109
3110 //------------------------------Ideal------------------------------------------
3111 // Return a node which is more "ideal" than the current node. Strip out
3112 // control copies
3113 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3114 if (remove_dead_region(phase, can_reshape)) return this;
3115 // Don't bother trying to transform a dead node
3116 if (in(0) && in(0)->is_top()) {
3117 return NULL;
3118 }
3119
3120 bool progress = false;
3121 // Eliminate volatile MemBars for scalar replaced objects.
3122 if (can_reshape && req() == (Precedent+1)) {
3123 bool eliminate = false;
3124 int opc = Opcode();
3125 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3126 // Volatile field loads and stores.
3127 Node* my_mem = in(MemBarNode::Precedent);
3128 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3129 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3130 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3131 // replace this Precedent (decodeN) with the Load instead.
3132 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3133 Node* load_node = my_mem->in(1);
3134 set_req(MemBarNode::Precedent, load_node);
3135 phase->is_IterGVN()->_worklist.push(my_mem);
3136 my_mem = load_node;
3137 } else {
3138 assert(my_mem->unique_out() == this, "sanity");
3139 del_req(Precedent);
3140 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3141 my_mem = NULL;
3142 }
3143 progress = true;
3144 }
3145 if (my_mem != NULL && my_mem->is_Mem()) {
3146 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3147 // Check for scalar replaced object reference.
3148 if( t_oop != NULL && t_oop->is_known_instance_field() &&
3149 t_oop->offset() != Type::OffsetBot &&
3150 t_oop->offset() != Type::OffsetTop) {
3151 eliminate = true;
3152 }
3153 }
3154 } else if (opc == Op_MemBarRelease) {
3155 // Final field stores.
3156 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3157 if ((alloc != NULL) && alloc->is_Allocate() &&
3158 alloc->as_Allocate()->does_not_escape_thread()) {
3159 // The allocated object does not escape.
3160 eliminate = true;
3161 }
3162 }
3163 if (eliminate) {
3164 // Replace MemBar projections by its inputs.
3165 PhaseIterGVN* igvn = phase->is_IterGVN();
3166 remove(igvn);
3167 // Must return either the original node (now dead) or a new node
3168 // (Do not return a top here, since that would break the uniqueness of top.)
3169 return new ConINode(TypeInt::ZERO);
3170 }
3171 }
3172 return progress ? this : NULL;
3173 }
3174
3175 //------------------------------Value------------------------------------------
3176 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3177 if( !in(0) ) return Type::TOP;
3178 if( phase->type(in(0)) == Type::TOP )
3179 return Type::TOP;
3180 return TypeTuple::MEMBAR;
3181 }
3182
3183 //------------------------------match------------------------------------------
3184 // Construct projections for memory.
3185 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3186 switch (proj->_con) {
3187 case TypeFunc::Control:
3188 case TypeFunc::Memory:
3189 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3190 }
3191 ShouldNotReachHere();
3192 return NULL;
3193 }
3194
3195 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3196 trailing->_kind = TrailingStore;
3197 leading->_kind = LeadingStore;
3198 #ifdef ASSERT
3199 trailing->_pair_idx = leading->_idx;
3200 leading->_pair_idx = leading->_idx;
3201 #endif
3202 }
3203
3204 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3205 trailing->_kind = TrailingLoadStore;
3206 leading->_kind = LeadingLoadStore;
3207 #ifdef ASSERT
3208 trailing->_pair_idx = leading->_idx;
3209 leading->_pair_idx = leading->_idx;
3210 #endif
3211 }
3212
3213 MemBarNode* MemBarNode::trailing_membar() const {
3214 ResourceMark rm;
3215 Node* trailing = (Node*)this;
3216 VectorSet seen(Thread::current()->resource_area());
3217 Node_Stack multis(0);
3218 do {
3219 Node* c = trailing;
3220 uint i = 0;
3221 do {
3222 trailing = NULL;
3223 for (; i < c->outcnt(); i++) {
3224 Node* next = c->raw_out(i);
3225 if (next != c && next->is_CFG()) {
3226 if (c->is_MultiBranch()) {
3227 if (multis.node() == c) {
3228 multis.set_index(i+1);
3229 } else {
3230 multis.push(c, i+1);
3231 }
3232 }
3233 trailing = next;
3234 break;
3235 }
3236 }
3237 if (trailing != NULL && !seen.test_set(trailing->_idx)) {
3238 break;
3239 }
3240 while (multis.size() > 0) {
3241 c = multis.node();
3242 i = multis.index();
3243 if (i < c->req()) {
3244 break;
3245 }
3246 multis.pop();
3247 }
3248 } while (multis.size() > 0);
3249 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
3250
3251 MemBarNode* mb = trailing->as_MemBar();
3252 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3253 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3254 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3255 return mb;
3256 }
3257
3258 MemBarNode* MemBarNode::leading_membar() const {
3259 ResourceMark rm;
3260 VectorSet seen(Thread::current()->resource_area());
3261 Node_Stack regions(0);
3262 Node* leading = in(0);
3263 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3264 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) {
3265 leading = NULL;
3266 while (regions.size() > 0 && leading == NULL) {
3267 Node* r = regions.node();
3268 uint i = regions.index();
3269 if (i < r->req()) {
3270 leading = r->in(i);
3271 regions.set_index(i+1);
3272 } else {
3273 regions.pop();
3274 }
3275 }
3276 if (leading == NULL) {
3277 assert(regions.size() == 0, "all paths should have been tried");
3278 return NULL;
3279 }
3280 }
3281 if (leading->is_Region()) {
3282 regions.push(leading, 2);
3283 leading = leading->in(1);
3284 } else {
3285 leading = leading->in(0);
3286 }
3287 }
3288 #ifdef ASSERT
3289 Unique_Node_List wq;
3290 wq.push((Node*)this);
3291 uint found = 0;
3292 for (uint i = 0; i < wq.size(); i++) {
3293 Node* n = wq.at(i);
3294 if (n->is_Region()) {
3295 for (uint j = 1; j < n->req(); j++) {
3296 Node* in = n->in(j);
3297 if (in != NULL && !in->is_top()) {
3298 wq.push(in);
3299 }
3300 }
3301 } else {
3302 if (n->is_MemBar() && n->as_MemBar()->leading()) {
3303 assert(n == leading, "consistency check failed");
3304 found++;
3305 } else {
3306 Node* in = n->in(0);
3307 if (in != NULL && !in->is_top()) {
3308 wq.push(in);
3309 }
3310 }
3311 }
3312 }
3313 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed");
3314 #endif
3315 if (leading == NULL) {
3316 return NULL;
3317 }
3318 MemBarNode* mb = leading->as_MemBar();
3319 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3320 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3321 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3322 return mb;
3323 }
3324
3325 //===========================InitializeNode====================================
3326 // SUMMARY:
3327 // This node acts as a memory barrier on raw memory, after some raw stores.
3328 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3329 // The Initialize can 'capture' suitably constrained stores as raw inits.
3330 // It can coalesce related raw stores into larger units (called 'tiles').
3331 // It can avoid zeroing new storage for memory units which have raw inits.
3332 // At macro-expansion, it is marked 'complete', and does not optimize further.
3333 //
3334 // EXAMPLE:
3335 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3336 // ctl = incoming control; mem* = incoming memory
3337 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3338 // First allocate uninitialized memory and fill in the header:
3339 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3340 // ctl := alloc.Control; mem* := alloc.Memory*
3341 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3342 // Then initialize to zero the non-header parts of the raw memory block:
3343 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3344 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3345 // After the initialize node executes, the object is ready for service:
3346 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3347 // Suppose its body is immediately initialized as {1,2}:
3348 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3349 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3350 // mem.SLICE(#short[*]) := store2
3351 //
3352 // DETAILS:
3353 // An InitializeNode collects and isolates object initialization after
3354 // an AllocateNode and before the next possible safepoint. As a
3355 // memory barrier (MemBarNode), it keeps critical stores from drifting
3356 // down past any safepoint or any publication of the allocation.
3357 // Before this barrier, a newly-allocated object may have uninitialized bits.
3358 // After this barrier, it may be treated as a real oop, and GC is allowed.
3359 //
3360 // The semantics of the InitializeNode include an implicit zeroing of
3361 // the new object from object header to the end of the object.
3362 // (The object header and end are determined by the AllocateNode.)
3363 //
3364 // Certain stores may be added as direct inputs to the InitializeNode.
3365 // These stores must update raw memory, and they must be to addresses
3366 // derived from the raw address produced by AllocateNode, and with
3367 // a constant offset. They must be ordered by increasing offset.
3368 // The first one is at in(RawStores), the last at in(req()-1).
3369 // Unlike most memory operations, they are not linked in a chain,
3370 // but are displayed in parallel as users of the rawmem output of
3371 // the allocation.
3372 //
3373 // (See comments in InitializeNode::capture_store, which continue
3374 // the example given above.)
3375 //
3376 // When the associated Allocate is macro-expanded, the InitializeNode
3377 // may be rewritten to optimize collected stores. A ClearArrayNode
3378 // may also be created at that point to represent any required zeroing.
3379 // The InitializeNode is then marked 'complete', prohibiting further
3380 // capturing of nearby memory operations.
3381 //
3382 // During macro-expansion, all captured initializations which store
3383 // constant values of 32 bits or smaller are coalesced (if advantageous)
3384 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3385 // initialized in fewer memory operations. Memory words which are
3386 // covered by neither tiles nor non-constant stores are pre-zeroed
3387 // by explicit stores of zero. (The code shape happens to do all
3388 // zeroing first, then all other stores, with both sequences occurring
3389 // in order of ascending offsets.)
3390 //
3391 // Alternatively, code may be inserted between an AllocateNode and its
3392 // InitializeNode, to perform arbitrary initialization of the new object.
3393 // E.g., the object copying intrinsics insert complex data transfers here.
3394 // The initialization must then be marked as 'complete' disable the
3395 // built-in zeroing semantics and the collection of initializing stores.
3396 //
3397 // While an InitializeNode is incomplete, reads from the memory state
3398 // produced by it are optimizable if they match the control edge and
3399 // new oop address associated with the allocation/initialization.
3400 // They return a stored value (if the offset matches) or else zero.
3401 // A write to the memory state, if it matches control and address,
3402 // and if it is to a constant offset, may be 'captured' by the
3403 // InitializeNode. It is cloned as a raw memory operation and rewired
3404 // inside the initialization, to the raw oop produced by the allocation.
3405 // Operations on addresses which are provably distinct (e.g., to
3406 // other AllocateNodes) are allowed to bypass the initialization.
3407 //
3408 // The effect of all this is to consolidate object initialization
3409 // (both arrays and non-arrays, both piecewise and bulk) into a
3410 // single location, where it can be optimized as a unit.
3411 //
3412 // Only stores with an offset less than TrackedInitializationLimit words
3413 // will be considered for capture by an InitializeNode. This puts a
3414 // reasonable limit on the complexity of optimized initializations.
3415
3416 //---------------------------InitializeNode------------------------------------
3417 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3418 : MemBarNode(C, adr_type, rawoop),
3419 _is_complete(Incomplete), _does_not_escape(false)
3420 {
3421 init_class_id(Class_Initialize);
3422
3423 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3424 assert(in(RawAddress) == rawoop, "proper init");
3425 // Note: allocation() can be NULL, for secondary initialization barriers
3426 }
3427
3428 // Since this node is not matched, it will be processed by the
3429 // register allocator. Declare that there are no constraints
3430 // on the allocation of the RawAddress edge.
3431 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3432 // This edge should be set to top, by the set_complete. But be conservative.
3433 if (idx == InitializeNode::RawAddress)
3434 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3435 return RegMask::Empty;
3436 }
3437
3438 Node* InitializeNode::memory(uint alias_idx) {
3439 Node* mem = in(Memory);
3440 if (mem->is_MergeMem()) {
3441 return mem->as_MergeMem()->memory_at(alias_idx);
3442 } else {
3443 // incoming raw memory is not split
3444 return mem;
3445 }
3446 }
3447
3448 bool InitializeNode::is_non_zero() {
3449 if (is_complete()) return false;
3450 remove_extra_zeroes();
3451 return (req() > RawStores);
3452 }
3453
3454 void InitializeNode::set_complete(PhaseGVN* phase) {
3455 assert(!is_complete(), "caller responsibility");
3456 _is_complete = Complete;
3457
3458 // After this node is complete, it contains a bunch of
3459 // raw-memory initializations. There is no need for
3460 // it to have anything to do with non-raw memory effects.
3461 // Therefore, tell all non-raw users to re-optimize themselves,
3462 // after skipping the memory effects of this initialization.
3463 PhaseIterGVN* igvn = phase->is_IterGVN();
3464 if (igvn) igvn->add_users_to_worklist(this);
3465 }
3466
3467 // convenience function
3468 // return false if the init contains any stores already
3469 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3470 InitializeNode* init = initialization();
3471 if (init == NULL || init->is_complete()) return false;
3472 init->remove_extra_zeroes();
3473 // for now, if this allocation has already collected any inits, bail:
3474 if (init->is_non_zero()) return false;
3475 init->set_complete(phase);
3476 return true;
3477 }
3478
3479 void InitializeNode::remove_extra_zeroes() {
3480 if (req() == RawStores) return;
3481 Node* zmem = zero_memory();
3482 uint fill = RawStores;
3483 for (uint i = fill; i < req(); i++) {
3484 Node* n = in(i);
3485 if (n->is_top() || n == zmem) continue; // skip
3486 if (fill < i) set_req(fill, n); // compact
3487 ++fill;
3488 }
3489 // delete any empty spaces created:
3490 while (fill < req()) {
3491 del_req(fill);
3492 }
3493 }
3494
3495 // Helper for remembering which stores go with which offsets.
3496 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3497 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3498 intptr_t offset = -1;
3499 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3500 phase, offset);
3501 if (base == NULL) return -1; // something is dead,
3502 if (offset < 0) return -1; // dead, dead
3503 return offset;
3504 }
3505
3506 // Helper for proving that an initialization expression is
3507 // "simple enough" to be folded into an object initialization.
3508 // Attempts to prove that a store's initial value 'n' can be captured
3509 // within the initialization without creating a vicious cycle, such as:
3510 // { Foo p = new Foo(); p.next = p; }
3511 // True for constants and parameters and small combinations thereof.
3512 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3513 if (n == NULL) return true; // (can this really happen?)
3514 if (n->is_Proj()) n = n->in(0);
3515 if (n == this) return false; // found a cycle
3516 if (n->is_Con()) return true;
3517 if (n->is_Start()) return true; // params, etc., are OK
3518 if (n->is_Root()) return true; // even better
3519
3520 Node* ctl = n->in(0);
3521 if (ctl != NULL && !ctl->is_top()) {
3522 if (ctl->is_Proj()) ctl = ctl->in(0);
3523 if (ctl == this) return false;
3524
3525 // If we already know that the enclosing memory op is pinned right after
3526 // the init, then any control flow that the store has picked up
3527 // must have preceded the init, or else be equal to the init.
3528 // Even after loop optimizations (which might change control edges)
3529 // a store is never pinned *before* the availability of its inputs.
3530 if (!MemNode::all_controls_dominate(n, this))
3531 return false; // failed to prove a good control
3532 }
3533
3534 // Check data edges for possible dependencies on 'this'.
3535 if ((count += 1) > 20) return false; // complexity limit
3536 for (uint i = 1; i < n->req(); i++) {
3537 Node* m = n->in(i);
3538 if (m == NULL || m == n || m->is_top()) continue;
3539 uint first_i = n->find_edge(m);
3540 if (i != first_i) continue; // process duplicate edge just once
3541 if (!detect_init_independence(m, count)) {
3542 return false;
3543 }
3544 }
3545
3546 return true;
3547 }
3548
3549 // Here are all the checks a Store must pass before it can be moved into
3550 // an initialization. Returns zero if a check fails.
3551 // On success, returns the (constant) offset to which the store applies,
3552 // within the initialized memory.
3553 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3554 const int FAIL = 0;
3555 if (st->req() != MemNode::ValueIn + 1)
3556 return FAIL; // an inscrutable StoreNode (card mark?)
3557 Node* ctl = st->in(MemNode::Control);
3558 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3559 return FAIL; // must be unconditional after the initialization
3560 Node* mem = st->in(MemNode::Memory);
3561 if (!(mem->is_Proj() && mem->in(0) == this))
3562 return FAIL; // must not be preceded by other stores
3563 Node* adr = st->in(MemNode::Address);
3564 intptr_t offset;
3565 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3566 if (alloc == NULL)
3567 return FAIL; // inscrutable address
3568 if (alloc != allocation())
3569 return FAIL; // wrong allocation! (store needs to float up)
3570 int size_in_bytes = st->memory_size();
3571 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
3572 return FAIL; // mismatched access
3573 }
3574 Node* val = st->in(MemNode::ValueIn);
3575 int complexity_count = 0;
3576 if (!detect_init_independence(val, complexity_count))
3577 return FAIL; // stored value must be 'simple enough'
3578
3579 // The Store can be captured only if nothing after the allocation
3580 // and before the Store is using the memory location that the store
3581 // overwrites.
3582 bool failed = false;
3583 // If is_complete_with_arraycopy() is true the shape of the graph is
3584 // well defined and is safe so no need for extra checks.
3585 if (!is_complete_with_arraycopy()) {
3586 // We are going to look at each use of the memory state following
3587 // the allocation to make sure nothing reads the memory that the
3588 // Store writes.
3589 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3590 int alias_idx = phase->C->get_alias_index(t_adr);
3591 ResourceMark rm;
3592 Unique_Node_List mems;
3593 mems.push(mem);
3594 Node* unique_merge = NULL;
3595 for (uint next = 0; next < mems.size(); ++next) {
3596 Node *m = mems.at(next);
3597 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3598 Node *n = m->fast_out(j);
3599 if (n->outcnt() == 0) {
3600 continue;
3601 }
3602 if (n == st) {
3603 continue;
3604 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3605 // If the control of this use is different from the control
3606 // of the Store which is right after the InitializeNode then
3607 // this node cannot be between the InitializeNode and the
3608 // Store.
3609 continue;
3610 } else if (n->is_MergeMem()) {
3611 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3612 // We can hit a MergeMemNode (that will likely go away
3613 // later) that is a direct use of the memory state
3614 // following the InitializeNode on the same slice as the
3615 // store node that we'd like to capture. We need to check
3616 // the uses of the MergeMemNode.
3617 mems.push(n);
3618 }
3619 } else if (n->is_Mem()) {
3620 Node* other_adr = n->in(MemNode::Address);
3621 if (other_adr == adr) {
3622 failed = true;
3623 break;
3624 } else {
3625 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3626 if (other_t_adr != NULL) {
3627 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3628 if (other_alias_idx == alias_idx) {
3629 // A load from the same memory slice as the store right
3630 // after the InitializeNode. We check the control of the
3631 // object/array that is loaded from. If it's the same as
3632 // the store control then we cannot capture the store.
3633 assert(!n->is_Store(), "2 stores to same slice on same control?");
3634 Node* base = other_adr;
3635 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3636 base = base->in(AddPNode::Base);
3637 if (base != NULL) {
3638 base = base->uncast();
3639 if (base->is_Proj() && base->in(0) == alloc) {
3640 failed = true;
3641 break;
3642 }
3643 }
3644 }
3645 }
3646 }
3647 } else {
3648 failed = true;
3649 break;
3650 }
3651 }
3652 }
3653 }
3654 if (failed) {
3655 if (!can_reshape) {
3656 // We decided we couldn't capture the store during parsing. We
3657 // should try again during the next IGVN once the graph is
3658 // cleaner.
3659 phase->C->record_for_igvn(st);
3660 }
3661 return FAIL;
3662 }
3663
3664 return offset; // success
3665 }
3666
3667 // Find the captured store in(i) which corresponds to the range
3668 // [start..start+size) in the initialized object.
3669 // If there is one, return its index i. If there isn't, return the
3670 // negative of the index where it should be inserted.
3671 // Return 0 if the queried range overlaps an initialization boundary
3672 // or if dead code is encountered.
3673 // If size_in_bytes is zero, do not bother with overlap checks.
3674 int InitializeNode::captured_store_insertion_point(intptr_t start,
3675 int size_in_bytes,
3676 PhaseTransform* phase) {
3677 const int FAIL = 0, MAX_STORE = BytesPerLong;
3678
3679 if (is_complete())
3680 return FAIL; // arraycopy got here first; punt
3681
3682 assert(allocation() != NULL, "must be present");
3683
3684 // no negatives, no header fields:
3685 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3686
3687 // after a certain size, we bail out on tracking all the stores:
3688 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3689 if (start >= ti_limit) return FAIL;
3690
3691 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3692 if (i >= limit) return -(int)i; // not found; here is where to put it
3693
3694 Node* st = in(i);
3695 intptr_t st_off = get_store_offset(st, phase);
3696 if (st_off < 0) {
3697 if (st != zero_memory()) {
3698 return FAIL; // bail out if there is dead garbage
3699 }
3700 } else if (st_off > start) {
3701 // ...we are done, since stores are ordered
3702 if (st_off < start + size_in_bytes) {
3703 return FAIL; // the next store overlaps
3704 }
3705 return -(int)i; // not found; here is where to put it
3706 } else if (st_off < start) {
3707 if (size_in_bytes != 0 &&
3708 start < st_off + MAX_STORE &&
3709 start < st_off + st->as_Store()->memory_size()) {
3710 return FAIL; // the previous store overlaps
3711 }
3712 } else {
3713 if (size_in_bytes != 0 &&
3714 st->as_Store()->memory_size() != size_in_bytes) {
3715 return FAIL; // mismatched store size
3716 }
3717 return i;
3718 }
3719
3720 ++i;
3721 }
3722 }
3723
3724 // Look for a captured store which initializes at the offset 'start'
3725 // with the given size. If there is no such store, and no other
3726 // initialization interferes, then return zero_memory (the memory
3727 // projection of the AllocateNode).
3728 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3729 PhaseTransform* phase) {
3730 assert(stores_are_sane(phase), "");
3731 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3732 if (i == 0) {
3733 return NULL; // something is dead
3734 } else if (i < 0) {
3735 return zero_memory(); // just primordial zero bits here
3736 } else {
3737 Node* st = in(i); // here is the store at this position
3738 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3739 return st;
3740 }
3741 }
3742
3743 // Create, as a raw pointer, an address within my new object at 'offset'.
3744 Node* InitializeNode::make_raw_address(intptr_t offset,
3745 PhaseTransform* phase) {
3746 Node* addr = in(RawAddress);
3747 if (offset != 0) {
3748 Compile* C = phase->C;
3749 addr = phase->transform( new AddPNode(C->top(), addr,
3750 phase->MakeConX(offset)) );
3751 }
3752 return addr;
3753 }
3754
3755 // Clone the given store, converting it into a raw store
3756 // initializing a field or element of my new object.
3757 // Caller is responsible for retiring the original store,
3758 // with subsume_node or the like.
3759 //
3760 // From the example above InitializeNode::InitializeNode,
3761 // here are the old stores to be captured:
3762 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3763 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3764 //
3765 // Here is the changed code; note the extra edges on init:
3766 // alloc = (Allocate ...)
3767 // rawoop = alloc.RawAddress
3768 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3769 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3770 // init = (Initialize alloc.Control alloc.Memory rawoop
3771 // rawstore1 rawstore2)
3772 //
3773 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3774 PhaseTransform* phase, bool can_reshape) {
3775 assert(stores_are_sane(phase), "");
3776
3777 if (start < 0) return NULL;
3778 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3779
3780 Compile* C = phase->C;
3781 int size_in_bytes = st->memory_size();
3782 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3783 if (i == 0) return NULL; // bail out
3784 Node* prev_mem = NULL; // raw memory for the captured store
3785 if (i > 0) {
3786 prev_mem = in(i); // there is a pre-existing store under this one
3787 set_req(i, C->top()); // temporarily disconnect it
3788 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3789 } else {
3790 i = -i; // no pre-existing store
3791 prev_mem = zero_memory(); // a slice of the newly allocated object
3792 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3793 set_req(--i, C->top()); // reuse this edge; it has been folded away
3794 else
3795 ins_req(i, C->top()); // build a new edge
3796 }
3797 Node* new_st = st->clone();
3798 new_st->set_req(MemNode::Control, in(Control));
3799 new_st->set_req(MemNode::Memory, prev_mem);
3800 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3801 new_st = phase->transform(new_st);
3802
3803 // At this point, new_st might have swallowed a pre-existing store
3804 // at the same offset, or perhaps new_st might have disappeared,
3805 // if it redundantly stored the same value (or zero to fresh memory).
3806
3807 // In any case, wire it in:
3808 phase->igvn_rehash_node_delayed(this);
3809 set_req(i, new_st);
3810
3811 // The caller may now kill the old guy.
3812 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3813 assert(check_st == new_st || check_st == NULL, "must be findable");
3814 assert(!is_complete(), "");
3815 return new_st;
3816 }
3817
3818 static bool store_constant(jlong* tiles, int num_tiles,
3819 intptr_t st_off, int st_size,
3820 jlong con) {
3821 if ((st_off & (st_size-1)) != 0)
3822 return false; // strange store offset (assume size==2**N)
3823 address addr = (address)tiles + st_off;
3824 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3825 switch (st_size) {
3826 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3827 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3828 case sizeof(jint): *(jint*) addr = (jint) con; break;
3829 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3830 default: return false; // strange store size (detect size!=2**N here)
3831 }
3832 return true; // return success to caller
3833 }
3834
3835 // Coalesce subword constants into int constants and possibly
3836 // into long constants. The goal, if the CPU permits,
3837 // is to initialize the object with a small number of 64-bit tiles.
3838 // Also, convert floating-point constants to bit patterns.
3839 // Non-constants are not relevant to this pass.
3840 //
3841 // In terms of the running example on InitializeNode::InitializeNode
3842 // and InitializeNode::capture_store, here is the transformation
3843 // of rawstore1 and rawstore2 into rawstore12:
3844 // alloc = (Allocate ...)
3845 // rawoop = alloc.RawAddress
3846 // tile12 = 0x00010002
3847 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3848 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3849 //
3850 void
3851 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3852 Node* size_in_bytes,
3853 PhaseGVN* phase) {
3854 Compile* C = phase->C;
3855
3856 assert(stores_are_sane(phase), "");
3857 // Note: After this pass, they are not completely sane,
3858 // since there may be some overlaps.
3859
3860 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3861
3862 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3863 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3864 size_limit = MIN2(size_limit, ti_limit);
3865 size_limit = align_up(size_limit, BytesPerLong);
3866 int num_tiles = size_limit / BytesPerLong;
3867
3868 // allocate space for the tile map:
3869 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3870 jlong tiles_buf[small_len];
3871 Node* nodes_buf[small_len];
3872 jlong inits_buf[small_len];
3873 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3874 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3875 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3876 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3877 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3878 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3879 // tiles: exact bitwise model of all primitive constants
3880 // nodes: last constant-storing node subsumed into the tiles model
3881 // inits: which bytes (in each tile) are touched by any initializations
3882
3883 //// Pass A: Fill in the tile model with any relevant stores.
3884
3885 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3886 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3887 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3888 Node* zmem = zero_memory(); // initially zero memory state
3889 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3890 Node* st = in(i);
3891 intptr_t st_off = get_store_offset(st, phase);
3892
3893 // Figure out the store's offset and constant value:
3894 if (st_off < header_size) continue; //skip (ignore header)
3895 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3896 int st_size = st->as_Store()->memory_size();
3897 if (st_off + st_size > size_limit) break;
3898
3899 // Record which bytes are touched, whether by constant or not.
3900 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3901 continue; // skip (strange store size)
3902
3903 const Type* val = phase->type(st->in(MemNode::ValueIn));
3904 if (!val->singleton()) continue; //skip (non-con store)
3905 BasicType type = val->basic_type();
3906
3907 jlong con = 0;
3908 switch (type) {
3909 case T_INT: con = val->is_int()->get_con(); break;
3910 case T_LONG: con = val->is_long()->get_con(); break;
3911 case T_FLOAT: con = jint_cast(val->getf()); break;
3912 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3913 default: continue; //skip (odd store type)
3914 }
3915
3916 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3917 st->Opcode() == Op_StoreL) {
3918 continue; // This StoreL is already optimal.
3919 }
3920
3921 // Store down the constant.
3922 store_constant(tiles, num_tiles, st_off, st_size, con);
3923
3924 intptr_t j = st_off >> LogBytesPerLong;
3925
3926 if (type == T_INT && st_size == BytesPerInt
3927 && (st_off & BytesPerInt) == BytesPerInt) {
3928 jlong lcon = tiles[j];
3929 if (!Matcher::isSimpleConstant64(lcon) &&
3930 st->Opcode() == Op_StoreI) {
3931 // This StoreI is already optimal by itself.
3932 jint* intcon = (jint*) &tiles[j];
3933 intcon[1] = 0; // undo the store_constant()
3934
3935 // If the previous store is also optimal by itself, back up and
3936 // undo the action of the previous loop iteration... if we can.
3937 // But if we can't, just let the previous half take care of itself.
3938 st = nodes[j];
3939 st_off -= BytesPerInt;
3940 con = intcon[0];
3941 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3942 assert(st_off >= header_size, "still ignoring header");
3943 assert(get_store_offset(st, phase) == st_off, "must be");
3944 assert(in(i-1) == zmem, "must be");
3945 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3946 assert(con == tcon->is_int()->get_con(), "must be");
3947 // Undo the effects of the previous loop trip, which swallowed st:
3948 intcon[0] = 0; // undo store_constant()
3949 set_req(i-1, st); // undo set_req(i, zmem)
3950 nodes[j] = NULL; // undo nodes[j] = st
3951 --old_subword; // undo ++old_subword
3952 }
3953 continue; // This StoreI is already optimal.
3954 }
3955 }
3956
3957 // This store is not needed.
3958 set_req(i, zmem);
3959 nodes[j] = st; // record for the moment
3960 if (st_size < BytesPerLong) // something has changed
3961 ++old_subword; // includes int/float, but who's counting...
3962 else ++old_long;
3963 }
3964
3965 if ((old_subword + old_long) == 0)
3966 return; // nothing more to do
3967
3968 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3969 // Be sure to insert them before overlapping non-constant stores.
3970 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3971 for (int j = 0; j < num_tiles; j++) {
3972 jlong con = tiles[j];
3973 jlong init = inits[j];
3974 if (con == 0) continue;
3975 jint con0, con1; // split the constant, address-wise
3976 jint init0, init1; // split the init map, address-wise
3977 { union { jlong con; jint intcon[2]; } u;
3978 u.con = con;
3979 con0 = u.intcon[0];
3980 con1 = u.intcon[1];
3981 u.con = init;
3982 init0 = u.intcon[0];
3983 init1 = u.intcon[1];
3984 }
3985
3986 Node* old = nodes[j];
3987 assert(old != NULL, "need the prior store");
3988 intptr_t offset = (j * BytesPerLong);
3989
3990 bool split = !Matcher::isSimpleConstant64(con);
3991
3992 if (offset < header_size) {
3993 assert(offset + BytesPerInt >= header_size, "second int counts");
3994 assert(*(jint*)&tiles[j] == 0, "junk in header");
3995 split = true; // only the second word counts
3996 // Example: int a[] = { 42 ... }
3997 } else if (con0 == 0 && init0 == -1) {
3998 split = true; // first word is covered by full inits
3999 // Example: int a[] = { ... foo(), 42 ... }
4000 } else if (con1 == 0 && init1 == -1) {
4001 split = true; // second word is covered by full inits
4002 // Example: int a[] = { ... 42, foo() ... }
4003 }
4004
4005 // Here's a case where init0 is neither 0 nor -1:
4006 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
4007 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
4008 // In this case the tile is not split; it is (jlong)42.
4009 // The big tile is stored down, and then the foo() value is inserted.
4010 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
4011
4012 Node* ctl = old->in(MemNode::Control);
4013 Node* adr = make_raw_address(offset, phase);
4014 const TypePtr* atp = TypeRawPtr::BOTTOM;
4015
4016 // One or two coalesced stores to plop down.
4017 Node* st[2];
4018 intptr_t off[2];
4019 int nst = 0;
4020 if (!split) {
4021 ++new_long;
4022 off[nst] = offset;
4023 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4024 phase->longcon(con), T_LONG, MemNode::unordered);
4025 } else {
4026 // Omit either if it is a zero.
4027 if (con0 != 0) {
4028 ++new_int;
4029 off[nst] = offset;
4030 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4031 phase->intcon(con0), T_INT, MemNode::unordered);
4032 }
4033 if (con1 != 0) {
4034 ++new_int;
4035 offset += BytesPerInt;
4036 adr = make_raw_address(offset, phase);
4037 off[nst] = offset;
4038 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4039 phase->intcon(con1), T_INT, MemNode::unordered);
4040 }
4041 }
4042
4043 // Insert second store first, then the first before the second.
4044 // Insert each one just before any overlapping non-constant stores.
4045 while (nst > 0) {
4046 Node* st1 = st[--nst];
4047 C->copy_node_notes_to(st1, old);
4048 st1 = phase->transform(st1);
4049 offset = off[nst];
4050 assert(offset >= header_size, "do not smash header");
4051 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4052 guarantee(ins_idx != 0, "must re-insert constant store");
4053 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
4054 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4055 set_req(--ins_idx, st1);
4056 else
4057 ins_req(ins_idx, st1);
4058 }
4059 }
4060
4061 if (PrintCompilation && WizardMode)
4062 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4063 old_subword, old_long, new_int, new_long);
4064 if (C->log() != NULL)
4065 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4066 old_subword, old_long, new_int, new_long);
4067
4068 // Clean up any remaining occurrences of zmem:
4069 remove_extra_zeroes();
4070 }
4071
4072 // Explore forward from in(start) to find the first fully initialized
4073 // word, and return its offset. Skip groups of subword stores which
4074 // together initialize full words. If in(start) is itself part of a
4075 // fully initialized word, return the offset of in(start). If there
4076 // are no following full-word stores, or if something is fishy, return
4077 // a negative value.
4078 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4079 int int_map = 0;
4080 intptr_t int_map_off = 0;
4081 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
4082
4083 for (uint i = start, limit = req(); i < limit; i++) {
4084 Node* st = in(i);
4085
4086 intptr_t st_off = get_store_offset(st, phase);
4087 if (st_off < 0) break; // return conservative answer
4088
4089 int st_size = st->as_Store()->memory_size();
4090 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4091 return st_off; // we found a complete word init
4092 }
4093
4094 // update the map:
4095
4096 intptr_t this_int_off = align_down(st_off, BytesPerInt);
4097 if (this_int_off != int_map_off) {
4098 // reset the map:
4099 int_map = 0;
4100 int_map_off = this_int_off;
4101 }
4102
4103 int subword_off = st_off - this_int_off;
4104 int_map |= right_n_bits(st_size) << subword_off;
4105 if ((int_map & FULL_MAP) == FULL_MAP) {
4106 return this_int_off; // we found a complete word init
4107 }
4108
4109 // Did this store hit or cross the word boundary?
4110 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4111 if (next_int_off == this_int_off + BytesPerInt) {
4112 // We passed the current int, without fully initializing it.
4113 int_map_off = next_int_off;
4114 int_map >>= BytesPerInt;
4115 } else if (next_int_off > this_int_off + BytesPerInt) {
4116 // We passed the current and next int.
4117 return this_int_off + BytesPerInt;
4118 }
4119 }
4120
4121 return -1;
4122 }
4123
4124
4125 // Called when the associated AllocateNode is expanded into CFG.
4126 // At this point, we may perform additional optimizations.
4127 // Linearize the stores by ascending offset, to make memory
4128 // activity as coherent as possible.
4129 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4130 intptr_t header_size,
4131 Node* size_in_bytes,
4132 PhaseGVN* phase) {
4133 assert(!is_complete(), "not already complete");
4134 assert(stores_are_sane(phase), "");
4135 assert(allocation() != NULL, "must be present");
4136
4137 remove_extra_zeroes();
4138
4139 if (ReduceFieldZeroing || ReduceBulkZeroing)
4140 // reduce instruction count for common initialization patterns
4141 coalesce_subword_stores(header_size, size_in_bytes, phase);
4142
4143 Node* zmem = zero_memory(); // initially zero memory state
4144 Node* inits = zmem; // accumulating a linearized chain of inits
4145 #ifdef ASSERT
4146 intptr_t first_offset = allocation()->minimum_header_size();
4147 intptr_t last_init_off = first_offset; // previous init offset
4148 intptr_t last_init_end = first_offset; // previous init offset+size
4149 intptr_t last_tile_end = first_offset; // previous tile offset+size
4150 #endif
4151 intptr_t zeroes_done = header_size;
4152
4153 bool do_zeroing = true; // we might give up if inits are very sparse
4154 int big_init_gaps = 0; // how many large gaps have we seen?
4155
4156 if (UseTLAB && ZeroTLAB) do_zeroing = false;
4157 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
4158
4159 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4160 Node* st = in(i);
4161 intptr_t st_off = get_store_offset(st, phase);
4162 if (st_off < 0)
4163 break; // unknown junk in the inits
4164 if (st->in(MemNode::Memory) != zmem)
4165 break; // complicated store chains somehow in list
4166
4167 int st_size = st->as_Store()->memory_size();
4168 intptr_t next_init_off = st_off + st_size;
4169
4170 if (do_zeroing && zeroes_done < next_init_off) {
4171 // See if this store needs a zero before it or under it.
4172 intptr_t zeroes_needed = st_off;
4173
4174 if (st_size < BytesPerInt) {
4175 // Look for subword stores which only partially initialize words.
4176 // If we find some, we must lay down some word-level zeroes first,
4177 // underneath the subword stores.
4178 //
4179 // Examples:
4180 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
4181 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
4182 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
4183 //
4184 // Note: coalesce_subword_stores may have already done this,
4185 // if it was prompted by constant non-zero subword initializers.
4186 // But this case can still arise with non-constant stores.
4187
4188 intptr_t next_full_store = find_next_fullword_store(i, phase);
4189
4190 // In the examples above:
4191 // in(i) p q r s x y z
4192 // st_off 12 13 14 15 12 13 14
4193 // st_size 1 1 1 1 1 1 1
4194 // next_full_s. 12 16 16 16 16 16 16
4195 // z's_done 12 16 16 16 12 16 12
4196 // z's_needed 12 16 16 16 16 16 16
4197 // zsize 0 0 0 0 4 0 4
4198 if (next_full_store < 0) {
4199 // Conservative tack: Zero to end of current word.
4200 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4201 } else {
4202 // Zero to beginning of next fully initialized word.
4203 // Or, don't zero at all, if we are already in that word.
4204 assert(next_full_store >= zeroes_needed, "must go forward");
4205 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4206 zeroes_needed = next_full_store;
4207 }
4208 }
4209
4210 if (zeroes_needed > zeroes_done) {
4211 intptr_t zsize = zeroes_needed - zeroes_done;
4212 // Do some incremental zeroing on rawmem, in parallel with inits.
4213 zeroes_done = align_down(zeroes_done, BytesPerInt);
4214 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4215 zeroes_done, zeroes_needed,
4216 phase);
4217 zeroes_done = zeroes_needed;
4218 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4219 do_zeroing = false; // leave the hole, next time
4220 }
4221 }
4222
4223 // Collect the store and move on:
4224 st->set_req(MemNode::Memory, inits);
4225 inits = st; // put it on the linearized chain
4226 set_req(i, zmem); // unhook from previous position
4227
4228 if (zeroes_done == st_off)
4229 zeroes_done = next_init_off;
4230
4231 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4232
4233 #ifdef ASSERT
4234 // Various order invariants. Weaker than stores_are_sane because
4235 // a large constant tile can be filled in by smaller non-constant stores.
4236 assert(st_off >= last_init_off, "inits do not reverse");
4237 last_init_off = st_off;
4238 const Type* val = NULL;
4239 if (st_size >= BytesPerInt &&
4240 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4241 (int)val->basic_type() < (int)T_OBJECT) {
4242 assert(st_off >= last_tile_end, "tiles do not overlap");
4243 assert(st_off >= last_init_end, "tiles do not overwrite inits");
4244 last_tile_end = MAX2(last_tile_end, next_init_off);
4245 } else {
4246 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4247 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4248 assert(st_off >= last_init_end, "inits do not overlap");
4249 last_init_end = next_init_off; // it's a non-tile
4250 }
4251 #endif //ASSERT
4252 }
4253
4254 remove_extra_zeroes(); // clear out all the zmems left over
4255 add_req(inits);
4256
4257 if (!(UseTLAB && ZeroTLAB)) {
4258 // If anything remains to be zeroed, zero it all now.
4259 zeroes_done = align_down(zeroes_done, BytesPerInt);
4260 // if it is the last unused 4 bytes of an instance, forget about it
4261 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4262 if (zeroes_done + BytesPerLong >= size_limit) {
4263 AllocateNode* alloc = allocation();
4264 assert(alloc != NULL, "must be present");
4265 if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
4266 Node* klass_node = alloc->in(AllocateNode::KlassNode);
4267 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4268 if (zeroes_done == k->layout_helper())
4269 zeroes_done = size_limit;
4270 }
4271 }
4272 if (zeroes_done < size_limit) {
4273 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4274 zeroes_done, size_in_bytes, phase);
4275 }
4276 }
4277
4278 set_complete(phase);
4279 return rawmem;
4280 }
4281
4282
4283 #ifdef ASSERT
4284 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4285 if (is_complete())
4286 return true; // stores could be anything at this point
4287 assert(allocation() != NULL, "must be present");
4288 intptr_t last_off = allocation()->minimum_header_size();
4289 for (uint i = InitializeNode::RawStores; i < req(); i++) {
4290 Node* st = in(i);
4291 intptr_t st_off = get_store_offset(st, phase);
4292 if (st_off < 0) continue; // ignore dead garbage
4293 if (last_off > st_off) {
4294 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4295 this->dump(2);
4296 assert(false, "ascending store offsets");
4297 return false;
4298 }
4299 last_off = st_off + st->as_Store()->memory_size();
4300 }
4301 return true;
4302 }
4303 #endif //ASSERT
4304
4305
4306
4307
4308 //============================MergeMemNode=====================================
4309 //
4310 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4311 // contributing store or call operations. Each contributor provides the memory
4312 // state for a particular "alias type" (see Compile::alias_type). For example,
4313 // if a MergeMem has an input X for alias category #6, then any memory reference
4314 // to alias category #6 may use X as its memory state input, as an exact equivalent
4315 // to using the MergeMem as a whole.
4316 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4317 //
4318 // (Here, the <N> notation gives the index of the relevant adr_type.)
4319 //
4320 // In one special case (and more cases in the future), alias categories overlap.
4321 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4322 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4323 // it is exactly equivalent to that state W:
4324 // MergeMem(<Bot>: W) <==> W
4325 //
4326 // Usually, the merge has more than one input. In that case, where inputs
4327 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4328 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4329 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4330 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4331 //
4332 // A merge can take a "wide" memory state as one of its narrow inputs.
4333 // This simply means that the merge observes out only the relevant parts of
4334 // the wide input. That is, wide memory states arriving at narrow merge inputs
4335 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4336 //
4337 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4338 // and that memory slices "leak through":
4339 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4340 //
4341 // But, in such a cascade, repeated memory slices can "block the leak":
4342 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4343 //
4344 // In the last example, Y is not part of the combined memory state of the
4345 // outermost MergeMem. The system must, of course, prevent unschedulable
4346 // memory states from arising, so you can be sure that the state Y is somehow
4347 // a precursor to state Y'.
4348 //
4349 //
4350 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4351 // of each MergeMemNode array are exactly the numerical alias indexes, including
4352 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4353 // Compile::alias_type (and kin) produce and manage these indexes.
4354 //
4355 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4356 // (Note that this provides quick access to the top node inside MergeMem methods,
4357 // without the need to reach out via TLS to Compile::current.)
4358 //
4359 // As a consequence of what was just described, a MergeMem that represents a full
4360 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4361 // containing all alias categories.
4362 //
4363 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4364 //
4365 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4366 // a memory state for the alias type <N>, or else the top node, meaning that
4367 // there is no particular input for that alias type. Note that the length of
4368 // a MergeMem is variable, and may be extended at any time to accommodate new
4369 // memory states at larger alias indexes. When merges grow, they are of course
4370 // filled with "top" in the unused in() positions.
4371 //
4372 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4373 // (Top was chosen because it works smoothly with passes like GCM.)
4374 //
4375 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4376 // the type of random VM bits like TLS references.) Since it is always the
4377 // first non-Bot memory slice, some low-level loops use it to initialize an
4378 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4379 //
4380 //
4381 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4382 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4383 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4384 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4385 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4386 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4387 //
4388 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4389 // really that different from the other memory inputs. An abbreviation called
4390 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4391 //
4392 //
4393 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4394 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4395 // that "emerges though" the base memory will be marked as excluding the alias types
4396 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4397 //
4398 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4399 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4400 //
4401 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4402 // (It is currently unimplemented.) As you can see, the resulting merge is
4403 // actually a disjoint union of memory states, rather than an overlay.
4404 //
4405
4406 //------------------------------MergeMemNode-----------------------------------
4407 Node* MergeMemNode::make_empty_memory() {
4408 Node* empty_memory = (Node*) Compile::current()->top();
4409 assert(empty_memory->is_top(), "correct sentinel identity");
4410 return empty_memory;
4411 }
4412
4413 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4414 init_class_id(Class_MergeMem);
4415 // all inputs are nullified in Node::Node(int)
4416 // set_input(0, NULL); // no control input
4417
4418 // Initialize the edges uniformly to top, for starters.
4419 Node* empty_mem = make_empty_memory();
4420 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4421 init_req(i,empty_mem);
4422 }
4423 assert(empty_memory() == empty_mem, "");
4424
4425 if( new_base != NULL && new_base->is_MergeMem() ) {
4426 MergeMemNode* mdef = new_base->as_MergeMem();
4427 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4428 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4429 mms.set_memory(mms.memory2());
4430 }
4431 assert(base_memory() == mdef->base_memory(), "");
4432 } else {
4433 set_base_memory(new_base);
4434 }
4435 }
4436
4437 // Make a new, untransformed MergeMem with the same base as 'mem'.
4438 // If mem is itself a MergeMem, populate the result with the same edges.
4439 MergeMemNode* MergeMemNode::make(Node* mem) {
4440 return new MergeMemNode(mem);
4441 }
4442
4443 //------------------------------cmp--------------------------------------------
4444 uint MergeMemNode::hash() const { return NO_HASH; }
4445 bool MergeMemNode::cmp( const Node &n ) const {
4446 return (&n == this); // Always fail except on self
4447 }
4448
4449 //------------------------------Identity---------------------------------------
4450 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4451 // Identity if this merge point does not record any interesting memory
4452 // disambiguations.
4453 Node* base_mem = base_memory();
4454 Node* empty_mem = empty_memory();
4455 if (base_mem != empty_mem) { // Memory path is not dead?
4456 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4457 Node* mem = in(i);
4458 if (mem != empty_mem && mem != base_mem) {
4459 return this; // Many memory splits; no change
4460 }
4461 }
4462 }
4463 return base_mem; // No memory splits; ID on the one true input
4464 }
4465
4466 //------------------------------Ideal------------------------------------------
4467 // This method is invoked recursively on chains of MergeMem nodes
4468 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4469 // Remove chain'd MergeMems
4470 //
4471 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4472 // relative to the "in(Bot)". Since we are patching both at the same time,
4473 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4474 // but rewrite each "in(i)" relative to the new "in(Bot)".
4475 Node *progress = NULL;
4476
4477
4478 Node* old_base = base_memory();
4479 Node* empty_mem = empty_memory();
4480 if (old_base == empty_mem)
4481 return NULL; // Dead memory path.
4482
4483 MergeMemNode* old_mbase;
4484 if (old_base != NULL && old_base->is_MergeMem())
4485 old_mbase = old_base->as_MergeMem();
4486 else
4487 old_mbase = NULL;
4488 Node* new_base = old_base;
4489
4490 // simplify stacked MergeMems in base memory
4491 if (old_mbase) new_base = old_mbase->base_memory();
4492
4493 // the base memory might contribute new slices beyond my req()
4494 if (old_mbase) grow_to_match(old_mbase);
4495
4496 // Look carefully at the base node if it is a phi.
4497 PhiNode* phi_base;
4498 if (new_base != NULL && new_base->is_Phi())
4499 phi_base = new_base->as_Phi();
4500 else
4501 phi_base = NULL;
4502
4503 Node* phi_reg = NULL;
4504 uint phi_len = (uint)-1;
4505 if (phi_base != NULL && !phi_base->is_copy()) {
4506 // do not examine phi if degraded to a copy
4507 phi_reg = phi_base->region();
4508 phi_len = phi_base->req();
4509 // see if the phi is unfinished
4510 for (uint i = 1; i < phi_len; i++) {
4511 if (phi_base->in(i) == NULL) {
4512 // incomplete phi; do not look at it yet!
4513 phi_reg = NULL;
4514 phi_len = (uint)-1;
4515 break;
4516 }
4517 }
4518 }
4519
4520 // Note: We do not call verify_sparse on entry, because inputs
4521 // can normalize to the base_memory via subsume_node or similar
4522 // mechanisms. This method repairs that damage.
4523
4524 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4525
4526 // Look at each slice.
4527 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4528 Node* old_in = in(i);
4529 // calculate the old memory value
4530 Node* old_mem = old_in;
4531 if (old_mem == empty_mem) old_mem = old_base;
4532 assert(old_mem == memory_at(i), "");
4533
4534 // maybe update (reslice) the old memory value
4535
4536 // simplify stacked MergeMems
4537 Node* new_mem = old_mem;
4538 MergeMemNode* old_mmem;
4539 if (old_mem != NULL && old_mem->is_MergeMem())
4540 old_mmem = old_mem->as_MergeMem();
4541 else
4542 old_mmem = NULL;
4543 if (old_mmem == this) {
4544 // This can happen if loops break up and safepoints disappear.
4545 // A merge of BotPtr (default) with a RawPtr memory derived from a
4546 // safepoint can be rewritten to a merge of the same BotPtr with
4547 // the BotPtr phi coming into the loop. If that phi disappears
4548 // also, we can end up with a self-loop of the mergemem.
4549 // In general, if loops degenerate and memory effects disappear,
4550 // a mergemem can be left looking at itself. This simply means
4551 // that the mergemem's default should be used, since there is
4552 // no longer any apparent effect on this slice.
4553 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4554 // from start. Update the input to TOP.
4555 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4556 }
4557 else if (old_mmem != NULL) {
4558 new_mem = old_mmem->memory_at(i);
4559 }
4560 // else preceding memory was not a MergeMem
4561
4562 // replace equivalent phis (unfortunately, they do not GVN together)
4563 if (new_mem != NULL && new_mem != new_base &&
4564 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4565 if (new_mem->is_Phi()) {
4566 PhiNode* phi_mem = new_mem->as_Phi();
4567 for (uint i = 1; i < phi_len; i++) {
4568 if (phi_base->in(i) != phi_mem->in(i)) {
4569 phi_mem = NULL;
4570 break;
4571 }
4572 }
4573 if (phi_mem != NULL) {
4574 // equivalent phi nodes; revert to the def
4575 new_mem = new_base;
4576 }
4577 }
4578 }
4579
4580 // maybe store down a new value
4581 Node* new_in = new_mem;
4582 if (new_in == new_base) new_in = empty_mem;
4583
4584 if (new_in != old_in) {
4585 // Warning: Do not combine this "if" with the previous "if"
4586 // A memory slice might have be be rewritten even if it is semantically
4587 // unchanged, if the base_memory value has changed.
4588 set_req(i, new_in);
4589 progress = this; // Report progress
4590 }
4591 }
4592
4593 if (new_base != old_base) {
4594 set_req(Compile::AliasIdxBot, new_base);
4595 // Don't use set_base_memory(new_base), because we need to update du.
4596 assert(base_memory() == new_base, "");
4597 progress = this;
4598 }
4599
4600 if( base_memory() == this ) {
4601 // a self cycle indicates this memory path is dead
4602 set_req(Compile::AliasIdxBot, empty_mem);
4603 }
4604
4605 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4606 // Recursion must occur after the self cycle check above
4607 if( base_memory()->is_MergeMem() ) {
4608 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4609 Node *m = phase->transform(new_mbase); // Rollup any cycles
4610 if( m != NULL &&
4611 (m->is_top() ||
4612 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4613 // propagate rollup of dead cycle to self
4614 set_req(Compile::AliasIdxBot, empty_mem);
4615 }
4616 }
4617
4618 if( base_memory() == empty_mem ) {
4619 progress = this;
4620 // Cut inputs during Parse phase only.
4621 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4622 if( !can_reshape ) {
4623 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4624 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4625 }
4626 }
4627 }
4628
4629 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4630 // Check if PhiNode::Ideal's "Split phis through memory merges"
4631 // transform should be attempted. Look for this->phi->this cycle.
4632 uint merge_width = req();
4633 if (merge_width > Compile::AliasIdxRaw) {
4634 PhiNode* phi = base_memory()->as_Phi();
4635 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4636 if (phi->in(i) == this) {
4637 phase->is_IterGVN()->_worklist.push(phi);
4638 break;
4639 }
4640 }
4641 }
4642 }
4643
4644 assert(progress || verify_sparse(), "please, no dups of base");
4645 return progress;
4646 }
4647
4648 //-------------------------set_base_memory-------------------------------------
4649 void MergeMemNode::set_base_memory(Node *new_base) {
4650 Node* empty_mem = empty_memory();
4651 set_req(Compile::AliasIdxBot, new_base);
4652 assert(memory_at(req()) == new_base, "must set default memory");
4653 // Clear out other occurrences of new_base:
4654 if (new_base != empty_mem) {
4655 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4656 if (in(i) == new_base) set_req(i, empty_mem);
4657 }
4658 }
4659 }
4660
4661 //------------------------------out_RegMask------------------------------------
4662 const RegMask &MergeMemNode::out_RegMask() const {
4663 return RegMask::Empty;
4664 }
4665
4666 //------------------------------dump_spec--------------------------------------
4667 #ifndef PRODUCT
4668 void MergeMemNode::dump_spec(outputStream *st) const {
4669 st->print(" {");
4670 Node* base_mem = base_memory();
4671 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4672 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4673 if (mem == base_mem) { st->print(" -"); continue; }
4674 st->print( " N%d:", mem->_idx );
4675 Compile::current()->get_adr_type(i)->dump_on(st);
4676 }
4677 st->print(" }");
4678 }
4679 #endif // !PRODUCT
4680
4681
4682 #ifdef ASSERT
4683 static bool might_be_same(Node* a, Node* b) {
4684 if (a == b) return true;
4685 if (!(a->is_Phi() || b->is_Phi())) return false;
4686 // phis shift around during optimization
4687 return true; // pretty stupid...
4688 }
4689
4690 // verify a narrow slice (either incoming or outgoing)
4691 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4692 if (!VerifyAliases) return; // don't bother to verify unless requested
4693 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
4694 if (Node::in_dump()) return; // muzzle asserts when printing
4695 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4696 assert(n != NULL, "");
4697 // Elide intervening MergeMem's
4698 while (n->is_MergeMem()) {
4699 n = n->as_MergeMem()->memory_at(alias_idx);
4700 }
4701 Compile* C = Compile::current();
4702 const TypePtr* n_adr_type = n->adr_type();
4703 if (n == m->empty_memory()) {
4704 // Implicit copy of base_memory()
4705 } else if (n_adr_type != TypePtr::BOTTOM) {
4706 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4707 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4708 } else {
4709 // A few places like make_runtime_call "know" that VM calls are narrow,
4710 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4711 bool expected_wide_mem = false;
4712 if (n == m->base_memory()) {
4713 expected_wide_mem = true;
4714 } else if (alias_idx == Compile::AliasIdxRaw ||
4715 n == m->memory_at(Compile::AliasIdxRaw)) {
4716 expected_wide_mem = true;
4717 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4718 // memory can "leak through" calls on channels that
4719 // are write-once. Allow this also.
4720 expected_wide_mem = true;
4721 }
4722 assert(expected_wide_mem, "expected narrow slice replacement");
4723 }
4724 }
4725 #else // !ASSERT
4726 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4727 #endif
4728
4729
4730 //-----------------------------memory_at---------------------------------------
4731 Node* MergeMemNode::memory_at(uint alias_idx) const {
4732 assert(alias_idx >= Compile::AliasIdxRaw ||
4733 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4734 "must avoid base_memory and AliasIdxTop");
4735
4736 // Otherwise, it is a narrow slice.
4737 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4738 Compile *C = Compile::current();
4739 if (is_empty_memory(n)) {
4740 // the array is sparse; empty slots are the "top" node
4741 n = base_memory();
4742 assert(Node::in_dump()
4743 || n == NULL || n->bottom_type() == Type::TOP
4744 || n->adr_type() == NULL // address is TOP
4745 || n->adr_type() == TypePtr::BOTTOM
4746 || n->adr_type() == TypeRawPtr::BOTTOM
4747 || Compile::current()->AliasLevel() == 0,
4748 "must be a wide memory");
4749 // AliasLevel == 0 if we are organizing the memory states manually.
4750 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4751 } else {
4752 // make sure the stored slice is sane
4753 #ifdef ASSERT
4754 if (VMError::is_error_reported() || Node::in_dump()) {
4755 } else if (might_be_same(n, base_memory())) {
4756 // Give it a pass: It is a mostly harmless repetition of the base.
4757 // This can arise normally from node subsumption during optimization.
4758 } else {
4759 verify_memory_slice(this, alias_idx, n);
4760 }
4761 #endif
4762 }
4763 return n;
4764 }
4765
4766 //---------------------------set_memory_at-------------------------------------
4767 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4768 verify_memory_slice(this, alias_idx, n);
4769 Node* empty_mem = empty_memory();
4770 if (n == base_memory()) n = empty_mem; // collapse default
4771 uint need_req = alias_idx+1;
4772 if (req() < need_req) {
4773 if (n == empty_mem) return; // already the default, so do not grow me
4774 // grow the sparse array
4775 do {
4776 add_req(empty_mem);
4777 } while (req() < need_req);
4778 }
4779 set_req( alias_idx, n );
4780 }
4781
4782
4783
4784 //--------------------------iteration_setup------------------------------------
4785 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4786 if (other != NULL) {
4787 grow_to_match(other);
4788 // invariant: the finite support of mm2 is within mm->req()
4789 #ifdef ASSERT
4790 for (uint i = req(); i < other->req(); i++) {
4791 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4792 }
4793 #endif
4794 }
4795 // Replace spurious copies of base_memory by top.
4796 Node* base_mem = base_memory();
4797 if (base_mem != NULL && !base_mem->is_top()) {
4798 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4799 if (in(i) == base_mem)
4800 set_req(i, empty_memory());
4801 }
4802 }
4803 }
4804
4805 //---------------------------grow_to_match-------------------------------------
4806 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4807 Node* empty_mem = empty_memory();
4808 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4809 // look for the finite support of the other memory
4810 for (uint i = other->req(); --i >= req(); ) {
4811 if (other->in(i) != empty_mem) {
4812 uint new_len = i+1;
4813 while (req() < new_len) add_req(empty_mem);
4814 break;
4815 }
4816 }
4817 }
4818
4819 //---------------------------verify_sparse-------------------------------------
4820 #ifndef PRODUCT
4821 bool MergeMemNode::verify_sparse() const {
4822 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4823 Node* base_mem = base_memory();
4824 // The following can happen in degenerate cases, since empty==top.
4825 if (is_empty_memory(base_mem)) return true;
4826 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4827 assert(in(i) != NULL, "sane slice");
4828 if (in(i) == base_mem) return false; // should have been the sentinel value!
4829 }
4830 return true;
4831 }
4832
4833 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4834 Node* n;
4835 n = mm->in(idx);
4836 if (mem == n) return true; // might be empty_memory()
4837 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4838 if (mem == n) return true;
4839 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4840 if (mem == n) return true;
4841 if (n == NULL) break;
4842 }
4843 return false;
4844 }
4845 #endif // !PRODUCT