1 /*
2 * Copyright (c) 1997, 2016, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "memory/resourceArea.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/arraycopynode.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/compile.hpp"
35 #include "opto/connode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/loopnode.hpp"
38 #include "opto/machnode.hpp"
39 #include "opto/matcher.hpp"
40 #include "opto/memnode.hpp"
41 #include "opto/mulnode.hpp"
42 #include "opto/narrowptrnode.hpp"
43 #include "opto/phaseX.hpp"
44 #include "opto/regmask.hpp"
45 #include "utilities/copy.hpp"
46
47 // Portions of code courtesy of Clifford Click
48
49 // Optimization - Graph Style
50
51 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
52
53 //=============================================================================
54 uint MemNode::size_of() const { return sizeof(*this); }
55
56 const TypePtr *MemNode::adr_type() const {
57 Node* adr = in(Address);
58 if (adr == NULL) return NULL; // node is dead
59 const TypePtr* cross_check = NULL;
60 DEBUG_ONLY(cross_check = _adr_type);
61 return calculate_adr_type(adr->bottom_type(), cross_check);
62 }
63
64 #ifndef PRODUCT
65 void MemNode::dump_spec(outputStream *st) const {
66 if (in(Address) == NULL) return; // node is dead
67 #ifndef ASSERT
68 // fake the missing field
69 const TypePtr* _adr_type = NULL;
70 if (in(Address) != NULL)
71 _adr_type = in(Address)->bottom_type()->isa_ptr();
72 #endif
73 dump_adr_type(this, _adr_type, st);
74
75 Compile* C = Compile::current();
76 if (C->alias_type(_adr_type)->is_volatile()) {
77 st->print(" Volatile!");
78 }
79 if (_unaligned_access) {
80 st->print(" unaligned");
81 }
82 if (_mismatched_access) {
83 st->print(" mismatched");
84 }
85 }
86
87 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
88 st->print(" @");
89 if (adr_type == NULL) {
90 st->print("NULL");
91 } else {
92 adr_type->dump_on(st);
93 Compile* C = Compile::current();
94 Compile::AliasType* atp = NULL;
95 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
96 if (atp == NULL)
97 st->print(", idx=?\?;");
98 else if (atp->index() == Compile::AliasIdxBot)
99 st->print(", idx=Bot;");
100 else if (atp->index() == Compile::AliasIdxTop)
101 st->print(", idx=Top;");
102 else if (atp->index() == Compile::AliasIdxRaw)
103 st->print(", idx=Raw;");
104 else {
105 ciField* field = atp->field();
106 if (field) {
107 st->print(", name=");
108 field->print_name_on(st);
109 }
110 st->print(", idx=%d;", atp->index());
111 }
112 }
113 }
114
115 extern void print_alias_types();
116
117 #endif
118
119 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
120 assert((t_oop != NULL), "sanity");
121 bool is_instance = t_oop->is_known_instance_field();
122 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
123 (load != NULL) && load->is_Load() &&
124 (phase->is_IterGVN() != NULL);
125 if (!(is_instance || is_boxed_value_load))
126 return mchain; // don't try to optimize non-instance types
127 uint instance_id = t_oop->instance_id();
128 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
129 Node *prev = NULL;
130 Node *result = mchain;
131 while (prev != result) {
132 prev = result;
133 if (result == start_mem)
134 break; // hit one of our sentinels
135 // skip over a call which does not affect this memory slice
136 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
137 Node *proj_in = result->in(0);
138 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
139 break; // hit one of our sentinels
140 } else if (proj_in->is_Call()) {
141 // ArrayCopyNodes processed here as well
142 CallNode *call = proj_in->as_Call();
143 if (!call->may_modify(t_oop, phase)) { // returns false for instances
144 result = call->in(TypeFunc::Memory);
145 }
146 } else if (proj_in->is_Initialize()) {
147 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
148 // Stop if this is the initialization for the object instance which
149 // contains this memory slice, otherwise skip over it.
150 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
151 break;
152 }
153 if (is_instance) {
154 result = proj_in->in(TypeFunc::Memory);
155 } else if (is_boxed_value_load) {
156 Node* klass = alloc->in(AllocateNode::KlassNode);
157 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
158 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
159 result = proj_in->in(TypeFunc::Memory); // not related allocation
160 }
161 }
162 } else if (proj_in->is_MemBar()) {
163 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase)) {
164 break;
165 }
166 result = proj_in->in(TypeFunc::Memory);
167 } else {
168 assert(false, "unexpected projection");
169 }
170 } else if (result->is_ClearArray()) {
171 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
172 // Can not bypass initialization of the instance
173 // we are looking for.
174 break;
175 }
176 // Otherwise skip it (the call updated 'result' value).
177 } else if (result->is_MergeMem()) {
178 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
179 }
180 }
181 return result;
182 }
183
184 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
185 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
186 if (t_oop == NULL)
187 return mchain; // don't try to optimize non-oop types
188 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
189 bool is_instance = t_oop->is_known_instance_field();
190 PhaseIterGVN *igvn = phase->is_IterGVN();
191 if (is_instance && igvn != NULL && result->is_Phi()) {
192 PhiNode *mphi = result->as_Phi();
193 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
194 const TypePtr *t = mphi->adr_type();
195 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
196 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
197 t->is_oopptr()->cast_to_exactness(true)
198 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
199 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
200 // clone the Phi with our address type
201 result = mphi->split_out_instance(t_adr, igvn);
202 } else {
203 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
204 }
205 }
206 return result;
207 }
208
209 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
210 uint alias_idx = phase->C->get_alias_index(tp);
211 Node *mem = mmem;
212 #ifdef ASSERT
213 {
214 // Check that current type is consistent with the alias index used during graph construction
215 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
216 bool consistent = adr_check == NULL || adr_check->empty() ||
217 phase->C->must_alias(adr_check, alias_idx );
218 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
219 if( !consistent && adr_check != NULL && !adr_check->empty() &&
220 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
221 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
222 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
223 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
224 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
225 // don't assert if it is dead code.
226 consistent = true;
227 }
228 if( !consistent ) {
229 st->print("alias_idx==%d, adr_check==", alias_idx);
230 if( adr_check == NULL ) {
231 st->print("NULL");
232 } else {
233 adr_check->dump();
234 }
235 st->cr();
236 print_alias_types();
237 assert(consistent, "adr_check must match alias idx");
238 }
239 }
240 #endif
241 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
242 // means an array I have not precisely typed yet. Do not do any
243 // alias stuff with it any time soon.
244 const TypeOopPtr *toop = tp->isa_oopptr();
245 if( tp->base() != Type::AnyPtr &&
246 !(toop &&
247 toop->klass() != NULL &&
248 toop->klass()->is_java_lang_Object() &&
249 toop->offset() == Type::OffsetBot) ) {
250 // compress paths and change unreachable cycles to TOP
251 // If not, we can update the input infinitely along a MergeMem cycle
252 // Equivalent code in PhiNode::Ideal
253 Node* m = phase->transform(mmem);
254 // If transformed to a MergeMem, get the desired slice
255 // Otherwise the returned node represents memory for every slice
256 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
257 // Update input if it is progress over what we have now
258 }
259 return mem;
260 }
261
262 //--------------------------Ideal_common---------------------------------------
263 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
264 // Unhook non-raw memories from complete (macro-expanded) initializations.
265 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
266 // If our control input is a dead region, kill all below the region
267 Node *ctl = in(MemNode::Control);
268 if (ctl && remove_dead_region(phase, can_reshape))
269 return this;
270 ctl = in(MemNode::Control);
271 // Don't bother trying to transform a dead node
272 if (ctl && ctl->is_top()) return NodeSentinel;
273
274 PhaseIterGVN *igvn = phase->is_IterGVN();
275 // Wait if control on the worklist.
276 if (ctl && can_reshape && igvn != NULL) {
277 Node* bol = NULL;
278 Node* cmp = NULL;
279 if (ctl->in(0)->is_If()) {
280 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
281 bol = ctl->in(0)->in(1);
282 if (bol->is_Bool())
283 cmp = ctl->in(0)->in(1)->in(1);
284 }
285 if (igvn->_worklist.member(ctl) ||
286 (bol != NULL && igvn->_worklist.member(bol)) ||
287 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
288 // This control path may be dead.
289 // Delay this memory node transformation until the control is processed.
290 phase->is_IterGVN()->_worklist.push(this);
291 return NodeSentinel; // caller will return NULL
292 }
293 }
294 // Ignore if memory is dead, or self-loop
295 Node *mem = in(MemNode::Memory);
296 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
297 assert(mem != this, "dead loop in MemNode::Ideal");
298
299 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
300 // This memory slice may be dead.
301 // Delay this mem node transformation until the memory is processed.
302 phase->is_IterGVN()->_worklist.push(this);
303 return NodeSentinel; // caller will return NULL
304 }
305
306 Node *address = in(MemNode::Address);
307 const Type *t_adr = phase->type(address);
308 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
309
310 if (can_reshape && igvn != NULL &&
311 (igvn->_worklist.member(address) ||
312 igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) {
313 // The address's base and type may change when the address is processed.
314 // Delay this mem node transformation until the address is processed.
315 phase->is_IterGVN()->_worklist.push(this);
316 return NodeSentinel; // caller will return NULL
317 }
318
319 // Do NOT remove or optimize the next lines: ensure a new alias index
320 // is allocated for an oop pointer type before Escape Analysis.
321 // Note: C++ will not remove it since the call has side effect.
322 if (t_adr->isa_oopptr()) {
323 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
324 }
325
326 Node* base = NULL;
327 if (address->is_AddP()) {
328 base = address->in(AddPNode::Base);
329 }
330 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
331 !t_adr->isa_rawptr()) {
332 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
333 // Skip this node optimization if its address has TOP base.
334 return NodeSentinel; // caller will return NULL
335 }
336
337 // Avoid independent memory operations
338 Node* old_mem = mem;
339
340 // The code which unhooks non-raw memories from complete (macro-expanded)
341 // initializations was removed. After macro-expansion all stores catched
342 // by Initialize node became raw stores and there is no information
343 // which memory slices they modify. So it is unsafe to move any memory
344 // operation above these stores. Also in most cases hooked non-raw memories
345 // were already unhooked by using information from detect_ptr_independence()
346 // and find_previous_store().
347
348 if (mem->is_MergeMem()) {
349 MergeMemNode* mmem = mem->as_MergeMem();
350 const TypePtr *tp = t_adr->is_ptr();
351
352 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
353 }
354
355 if (mem != old_mem) {
356 set_req(MemNode::Memory, mem);
357 if (can_reshape && old_mem->outcnt() == 0) {
358 igvn->_worklist.push(old_mem);
359 }
360 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
361 return this;
362 }
363
364 // let the subclass continue analyzing...
365 return NULL;
366 }
367
368 // Helper function for proving some simple control dominations.
369 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
370 // Already assumes that 'dom' is available at 'sub', and that 'sub'
371 // is not a constant (dominated by the method's StartNode).
372 // Used by MemNode::find_previous_store to prove that the
373 // control input of a memory operation predates (dominates)
374 // an allocation it wants to look past.
375 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
376 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
377 return false; // Conservative answer for dead code
378
379 // Check 'dom'. Skip Proj and CatchProj nodes.
380 dom = dom->find_exact_control(dom);
381 if (dom == NULL || dom->is_top())
382 return false; // Conservative answer for dead code
383
384 if (dom == sub) {
385 // For the case when, for example, 'sub' is Initialize and the original
386 // 'dom' is Proj node of the 'sub'.
387 return false;
388 }
389
390 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
391 return true;
392
393 // 'dom' dominates 'sub' if its control edge and control edges
394 // of all its inputs dominate or equal to sub's control edge.
395
396 // Currently 'sub' is either Allocate, Initialize or Start nodes.
397 // Or Region for the check in LoadNode::Ideal();
398 // 'sub' should have sub->in(0) != NULL.
399 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
400 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
401
402 // Get control edge of 'sub'.
403 Node* orig_sub = sub;
404 sub = sub->find_exact_control(sub->in(0));
405 if (sub == NULL || sub->is_top())
406 return false; // Conservative answer for dead code
407
408 assert(sub->is_CFG(), "expecting control");
409
410 if (sub == dom)
411 return true;
412
413 if (sub->is_Start() || sub->is_Root())
414 return false;
415
416 {
417 // Check all control edges of 'dom'.
418
419 ResourceMark rm;
420 Arena* arena = Thread::current()->resource_area();
421 Node_List nlist(arena);
422 Unique_Node_List dom_list(arena);
423
424 dom_list.push(dom);
425 bool only_dominating_controls = false;
426
427 for (uint next = 0; next < dom_list.size(); next++) {
428 Node* n = dom_list.at(next);
429 if (n == orig_sub)
430 return false; // One of dom's inputs dominated by sub.
431 if (!n->is_CFG() && n->pinned()) {
432 // Check only own control edge for pinned non-control nodes.
433 n = n->find_exact_control(n->in(0));
434 if (n == NULL || n->is_top())
435 return false; // Conservative answer for dead code
436 assert(n->is_CFG(), "expecting control");
437 dom_list.push(n);
438 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
439 only_dominating_controls = true;
440 } else if (n->is_CFG()) {
441 if (n->dominates(sub, nlist))
442 only_dominating_controls = true;
443 else
444 return false;
445 } else {
446 // First, own control edge.
447 Node* m = n->find_exact_control(n->in(0));
448 if (m != NULL) {
449 if (m->is_top())
450 return false; // Conservative answer for dead code
451 dom_list.push(m);
452 }
453 // Now, the rest of edges.
454 uint cnt = n->req();
455 for (uint i = 1; i < cnt; i++) {
456 m = n->find_exact_control(n->in(i));
457 if (m == NULL || m->is_top())
458 continue;
459 dom_list.push(m);
460 }
461 }
462 }
463 return only_dominating_controls;
464 }
465 }
466
467 //---------------------detect_ptr_independence---------------------------------
468 // Used by MemNode::find_previous_store to prove that two base
469 // pointers are never equal.
470 // The pointers are accompanied by their associated allocations,
471 // if any, which have been previously discovered by the caller.
472 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
473 Node* p2, AllocateNode* a2,
474 PhaseTransform* phase) {
475 // Attempt to prove that these two pointers cannot be aliased.
476 // They may both manifestly be allocations, and they should differ.
477 // Or, if they are not both allocations, they can be distinct constants.
478 // Otherwise, one is an allocation and the other a pre-existing value.
479 if (a1 == NULL && a2 == NULL) { // neither an allocation
480 return (p1 != p2) && p1->is_Con() && p2->is_Con();
481 } else if (a1 != NULL && a2 != NULL) { // both allocations
482 return (a1 != a2);
483 } else if (a1 != NULL) { // one allocation a1
484 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
485 return all_controls_dominate(p2, a1);
486 } else { //(a2 != NULL) // one allocation a2
487 return all_controls_dominate(p1, a2);
488 }
489 return false;
490 }
491
492
493 // Find an arraycopy that must have set (can_see_stored_value=true) or
494 // could have set (can_see_stored_value=false) the value for this load
495 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
496 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
497 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
498 Node* mb = mem->in(0);
499 if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
500 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
501 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy();
502 if (ac->is_clonebasic()) {
503 intptr_t offset;
504 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset);
505 assert(alloc != NULL && alloc->initialization()->is_complete_with_arraycopy(), "broken allocation");
506 if (alloc == ld_alloc) {
507 return ac;
508 }
509 }
510 }
511 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
512 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
513
514 if (ac->is_arraycopy_validated() ||
515 ac->is_copyof_validated() ||
516 ac->is_copyofrange_validated()) {
517 Node* ld_addp = in(MemNode::Address);
518 if (ld_addp->is_AddP()) {
519 Node* ld_base = ld_addp->in(AddPNode::Address);
520 Node* ld_offs = ld_addp->in(AddPNode::Offset);
521
522 Node* dest = ac->in(ArrayCopyNode::Dest);
523
524 if (dest == ld_base) {
525 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
526 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
527 return ac;
528 }
529 if (!can_see_stored_value) {
530 mem = ac->in(TypeFunc::Memory);
531 }
532 }
533 }
534 }
535 }
536 return NULL;
537 }
538
539 // The logic for reordering loads and stores uses four steps:
540 // (a) Walk carefully past stores and initializations which we
541 // can prove are independent of this load.
542 // (b) Observe that the next memory state makes an exact match
543 // with self (load or store), and locate the relevant store.
544 // (c) Ensure that, if we were to wire self directly to the store,
545 // the optimizer would fold it up somehow.
546 // (d) Do the rewiring, and return, depending on some other part of
547 // the optimizer to fold up the load.
548 // This routine handles steps (a) and (b). Steps (c) and (d) are
549 // specific to loads and stores, so they are handled by the callers.
550 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
551 //
552 Node* MemNode::find_previous_store(PhaseTransform* phase) {
553 Node* ctrl = in(MemNode::Control);
554 Node* adr = in(MemNode::Address);
555 intptr_t offset = 0;
556 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
557 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
558
559 if (offset == Type::OffsetBot)
560 return NULL; // cannot unalias unless there are precise offsets
561
562 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
563
564 intptr_t size_in_bytes = memory_size();
565
566 Node* mem = in(MemNode::Memory); // start searching here...
567
568 int cnt = 50; // Cycle limiter
569 for (;;) { // While we can dance past unrelated stores...
570 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
571
572 Node* prev = mem;
573 if (mem->is_Store()) {
574 Node* st_adr = mem->in(MemNode::Address);
575 intptr_t st_offset = 0;
576 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
577 if (st_base == NULL)
578 break; // inscrutable pointer
579 if (st_offset != offset && st_offset != Type::OffsetBot) {
580 const int MAX_STORE = BytesPerLong;
581 if (st_offset >= offset + size_in_bytes ||
582 st_offset <= offset - MAX_STORE ||
583 st_offset <= offset - mem->as_Store()->memory_size()) {
584 // Success: The offsets are provably independent.
585 // (You may ask, why not just test st_offset != offset and be done?
586 // The answer is that stores of different sizes can co-exist
587 // in the same sequence of RawMem effects. We sometimes initialize
588 // a whole 'tile' of array elements with a single jint or jlong.)
589 mem = mem->in(MemNode::Memory);
590 continue; // (a) advance through independent store memory
591 }
592 }
593 if (st_base != base &&
594 detect_ptr_independence(base, alloc,
595 st_base,
596 AllocateNode::Ideal_allocation(st_base, phase),
597 phase)) {
598 // Success: The bases are provably independent.
599 mem = mem->in(MemNode::Memory);
600 continue; // (a) advance through independent store memory
601 }
602
603 // (b) At this point, if the bases or offsets do not agree, we lose,
604 // since we have not managed to prove 'this' and 'mem' independent.
605 if (st_base == base && st_offset == offset) {
606 return mem; // let caller handle steps (c), (d)
607 }
608
609 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
610 InitializeNode* st_init = mem->in(0)->as_Initialize();
611 AllocateNode* st_alloc = st_init->allocation();
612 if (st_alloc == NULL)
613 break; // something degenerated
614 bool known_identical = false;
615 bool known_independent = false;
616 if (alloc == st_alloc)
617 known_identical = true;
618 else if (alloc != NULL)
619 known_independent = true;
620 else if (all_controls_dominate(this, st_alloc))
621 known_independent = true;
622
623 if (known_independent) {
624 // The bases are provably independent: Either they are
625 // manifestly distinct allocations, or else the control
626 // of this load dominates the store's allocation.
627 int alias_idx = phase->C->get_alias_index(adr_type());
628 if (alias_idx == Compile::AliasIdxRaw) {
629 mem = st_alloc->in(TypeFunc::Memory);
630 } else {
631 mem = st_init->memory(alias_idx);
632 }
633 continue; // (a) advance through independent store memory
634 }
635
636 // (b) at this point, if we are not looking at a store initializing
637 // the same allocation we are loading from, we lose.
638 if (known_identical) {
639 // From caller, can_see_stored_value will consult find_captured_store.
640 return mem; // let caller handle steps (c), (d)
641 }
642
643 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
644 if (prev != mem) {
645 // Found an arraycopy but it doesn't affect that load
646 continue;
647 }
648 // Found an arraycopy that may affect that load
649 return mem;
650 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
651 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
652 if (mem->is_Proj() && mem->in(0)->is_Call()) {
653 // ArrayCopyNodes processed here as well.
654 CallNode *call = mem->in(0)->as_Call();
655 if (!call->may_modify(addr_t, phase)) {
656 mem = call->in(TypeFunc::Memory);
657 continue; // (a) advance through independent call memory
658 }
659 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
660 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase)) {
661 break;
662 }
663 mem = mem->in(0)->in(TypeFunc::Memory);
664 continue; // (a) advance through independent MemBar memory
665 } else if (mem->is_ClearArray()) {
666 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
667 // (the call updated 'mem' value)
668 continue; // (a) advance through independent allocation memory
669 } else {
670 // Can not bypass initialization of the instance
671 // we are looking for.
672 return mem;
673 }
674 } else if (mem->is_MergeMem()) {
675 int alias_idx = phase->C->get_alias_index(adr_type());
676 mem = mem->as_MergeMem()->memory_at(alias_idx);
677 continue; // (a) advance through independent MergeMem memory
678 }
679 }
680
681 // Unless there is an explicit 'continue', we must bail out here,
682 // because 'mem' is an inscrutable memory state (e.g., a call).
683 break;
684 }
685
686 return NULL; // bail out
687 }
688
689 //----------------------calculate_adr_type-------------------------------------
690 // Helper function. Notices when the given type of address hits top or bottom.
691 // Also, asserts a cross-check of the type against the expected address type.
692 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
693 if (t == Type::TOP) return NULL; // does not touch memory any more?
694 #ifdef PRODUCT
695 cross_check = NULL;
696 #else
697 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
698 #endif
699 const TypePtr* tp = t->isa_ptr();
700 if (tp == NULL) {
701 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
702 return TypePtr::BOTTOM; // touches lots of memory
703 } else {
704 #ifdef ASSERT
705 // %%%% [phh] We don't check the alias index if cross_check is
706 // TypeRawPtr::BOTTOM. Needs to be investigated.
707 if (cross_check != NULL &&
708 cross_check != TypePtr::BOTTOM &&
709 cross_check != TypeRawPtr::BOTTOM) {
710 // Recheck the alias index, to see if it has changed (due to a bug).
711 Compile* C = Compile::current();
712 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
713 "must stay in the original alias category");
714 // The type of the address must be contained in the adr_type,
715 // disregarding "null"-ness.
716 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
717 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
718 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
719 "real address must not escape from expected memory type");
720 }
721 #endif
722 return tp;
723 }
724 }
725
726 //=============================================================================
727 // Should LoadNode::Ideal() attempt to remove control edges?
728 bool LoadNode::can_remove_control() const {
729 return true;
730 }
731 uint LoadNode::size_of() const { return sizeof(*this); }
732 uint LoadNode::cmp( const Node &n ) const
733 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
734 const Type *LoadNode::bottom_type() const { return _type; }
735 uint LoadNode::ideal_reg() const {
736 return _type->ideal_reg();
737 }
738
739 #ifndef PRODUCT
740 void LoadNode::dump_spec(outputStream *st) const {
741 MemNode::dump_spec(st);
742 if( !Verbose && !WizardMode ) {
743 // standard dump does this in Verbose and WizardMode
744 st->print(" #"); _type->dump_on(st);
745 }
746 if (!_depends_only_on_test) {
747 st->print(" (does not depend only on test)");
748 }
749 }
750 #endif
751
752 #ifdef ASSERT
753 //----------------------------is_immutable_value-------------------------------
754 // Helper function to allow a raw load without control edge for some cases
755 bool LoadNode::is_immutable_value(Node* adr) {
756 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
757 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
758 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
759 in_bytes(JavaThread::osthread_offset())));
760 }
761 #endif
762
763 //----------------------------LoadNode::make-----------------------------------
764 // Polymorphic factory method:
765 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo,
766 ControlDependency control_dependency, bool unaligned, bool mismatched) {
767 Compile* C = gvn.C;
768
769 // sanity check the alias category against the created node type
770 assert(!(adr_type->isa_oopptr() &&
771 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
772 "use LoadKlassNode instead");
773 assert(!(adr_type->isa_aryptr() &&
774 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
775 "use LoadRangeNode instead");
776 // Check control edge of raw loads
777 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
778 // oop will be recorded in oop map if load crosses safepoint
779 rt->isa_oopptr() || is_immutable_value(adr),
780 "raw memory operations should have control edge");
781 LoadNode* load = NULL;
782 switch (bt) {
783 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
784 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
785 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
786 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
787 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
788 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break;
789 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
790 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
791 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
792 case T_OBJECT:
793 #ifdef _LP64
794 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
795 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
796 } else
797 #endif
798 {
799 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
800 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_oopptr(), mo, control_dependency);
801 }
802 break;
803 }
804 assert(load != NULL, "LoadNode should have been created");
805 if (unaligned) {
806 load->set_unaligned_access();
807 }
808 if (mismatched) {
809 load->set_mismatched_access();
810 }
811 if (load->Opcode() == Op_LoadN) {
812 Node* ld = gvn.transform(load);
813 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
814 }
815
816 return load;
817 }
818
819 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
820 ControlDependency control_dependency, bool unaligned, bool mismatched) {
821 bool require_atomic = true;
822 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic);
823 if (unaligned) {
824 load->set_unaligned_access();
825 }
826 if (mismatched) {
827 load->set_mismatched_access();
828 }
829 return load;
830 }
831
832 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
833 ControlDependency control_dependency, bool unaligned, bool mismatched) {
834 bool require_atomic = true;
835 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic);
836 if (unaligned) {
837 load->set_unaligned_access();
838 }
839 if (mismatched) {
840 load->set_mismatched_access();
841 }
842 return load;
843 }
844
845
846
847 //------------------------------hash-------------------------------------------
848 uint LoadNode::hash() const {
849 // unroll addition of interesting fields
850 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
851 }
852
853 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
854 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
855 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
856 bool is_stable_ary = FoldStableValues &&
857 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
858 tp->isa_aryptr()->is_stable();
859
860 return (eliminate_boxing && non_volatile) || is_stable_ary;
861 }
862
863 return false;
864 }
865
866 // Is the value loaded previously stored by an arraycopy? If so return
867 // a load node that reads from the source array so we may be able to
868 // optimize out the ArrayCopy node later.
869 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseTransform* phase) const {
870 Node* ld_adr = in(MemNode::Address);
871 intptr_t ld_off = 0;
872 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
873 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
874 if (ac != NULL) {
875 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
876
877 Node* ld = clone();
878 if (ac->as_ArrayCopy()->is_clonebasic()) {
879 assert(ld_alloc != NULL, "need an alloc");
880 Node* addp = in(MemNode::Address)->clone();
881 assert(addp->is_AddP(), "address must be addp");
882 assert(addp->in(AddPNode::Base) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base), "strange pattern");
883 assert(addp->in(AddPNode::Address) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address), "strange pattern");
884 addp->set_req(AddPNode::Base, ac->in(ArrayCopyNode::Src)->in(AddPNode::Base));
885 addp->set_req(AddPNode::Address, ac->in(ArrayCopyNode::Src)->in(AddPNode::Address));
886 ld->set_req(MemNode::Address, phase->transform(addp));
887 if (in(0) != NULL) {
888 assert(ld_alloc->in(0) != NULL, "alloc must have control");
889 ld->set_req(0, ld_alloc->in(0));
890 }
891 } else {
892 Node* addp = in(MemNode::Address)->clone();
893 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
894 addp->set_req(AddPNode::Base, ac->in(ArrayCopyNode::Src));
895 addp->set_req(AddPNode::Address, ac->in(ArrayCopyNode::Src));
896
897 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
898 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type();
899 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
900 uint shift = exact_log2(type2aelembytes(ary_elem));
901
902 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
903 #ifdef _LP64
904 diff = phase->transform(new ConvI2LNode(diff));
905 #endif
906 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
907
908 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
909 addp->set_req(AddPNode::Offset, offset);
910 ld->set_req(MemNode::Address, phase->transform(addp));
911
912 if (in(0) != NULL) {
913 assert(ac->in(0) != NULL, "alloc must have control");
914 ld->set_req(0, ac->in(0));
915 }
916 }
917 // load depends on the tests that validate the arraycopy
918 ld->as_Load()->_depends_only_on_test = Pinned;
919 return ld;
920 }
921 return NULL;
922 }
923
924
925 //---------------------------can_see_stored_value------------------------------
926 // This routine exists to make sure this set of tests is done the same
927 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
928 // will change the graph shape in a way which makes memory alive twice at the
929 // same time (uses the Oracle model of aliasing), then some
930 // LoadXNode::Identity will fold things back to the equivalence-class model
931 // of aliasing.
932 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
933 Node* ld_adr = in(MemNode::Address);
934 intptr_t ld_off = 0;
935 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
936 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
937 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
938 // This is more general than load from boxing objects.
939 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
940 uint alias_idx = atp->index();
941 bool final = !atp->is_rewritable();
942 Node* result = NULL;
943 Node* current = st;
944 // Skip through chains of MemBarNodes checking the MergeMems for
945 // new states for the slice of this load. Stop once any other
946 // kind of node is encountered. Loads from final memory can skip
947 // through any kind of MemBar but normal loads shouldn't skip
948 // through MemBarAcquire since the could allow them to move out of
949 // a synchronized region.
950 while (current->is_Proj()) {
951 int opc = current->in(0)->Opcode();
952 if ((final && (opc == Op_MemBarAcquire ||
953 opc == Op_MemBarAcquireLock ||
954 opc == Op_LoadFence)) ||
955 opc == Op_MemBarRelease ||
956 opc == Op_StoreFence ||
957 opc == Op_MemBarReleaseLock ||
958 opc == Op_MemBarStoreStore ||
959 opc == Op_MemBarCPUOrder) {
960 Node* mem = current->in(0)->in(TypeFunc::Memory);
961 if (mem->is_MergeMem()) {
962 MergeMemNode* merge = mem->as_MergeMem();
963 Node* new_st = merge->memory_at(alias_idx);
964 if (new_st == merge->base_memory()) {
965 // Keep searching
966 current = new_st;
967 continue;
968 }
969 // Save the new memory state for the slice and fall through
970 // to exit.
971 result = new_st;
972 }
973 }
974 break;
975 }
976 if (result != NULL) {
977 st = result;
978 }
979 }
980
981 // Loop around twice in the case Load -> Initialize -> Store.
982 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
983 for (int trip = 0; trip <= 1; trip++) {
984
985 if (st->is_Store()) {
986 Node* st_adr = st->in(MemNode::Address);
987 if (!phase->eqv(st_adr, ld_adr)) {
988 // Try harder before giving up... Match raw and non-raw pointers.
989 intptr_t st_off = 0;
990 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
991 if (alloc == NULL) return NULL;
992 if (alloc != ld_alloc) return NULL;
993 if (ld_off != st_off) return NULL;
994 // At this point we have proven something like this setup:
995 // A = Allocate(...)
996 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
997 // S = StoreQ(, AddP(, A.Parm , #Off), V)
998 // (Actually, we haven't yet proven the Q's are the same.)
999 // In other words, we are loading from a casted version of
1000 // the same pointer-and-offset that we stored to.
1001 // Thus, we are able to replace L by V.
1002 }
1003 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1004 if (store_Opcode() != st->Opcode())
1005 return NULL;
1006 return st->in(MemNode::ValueIn);
1007 }
1008
1009 // A load from a freshly-created object always returns zero.
1010 // (This can happen after LoadNode::Ideal resets the load's memory input
1011 // to find_captured_store, which returned InitializeNode::zero_memory.)
1012 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1013 (st->in(0) == ld_alloc) &&
1014 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1015 // return a zero value for the load's basic type
1016 // (This is one of the few places where a generic PhaseTransform
1017 // can create new nodes. Think of it as lazily manifesting
1018 // virtually pre-existing constants.)
1019 return phase->zerocon(memory_type());
1020 }
1021
1022 // A load from an initialization barrier can match a captured store.
1023 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1024 InitializeNode* init = st->in(0)->as_Initialize();
1025 AllocateNode* alloc = init->allocation();
1026 if ((alloc != NULL) && (alloc == ld_alloc)) {
1027 // examine a captured store value
1028 st = init->find_captured_store(ld_off, memory_size(), phase);
1029 if (st != NULL) {
1030 continue; // take one more trip around
1031 }
1032 }
1033 }
1034
1035 // Load boxed value from result of valueOf() call is input parameter.
1036 if (this->is_Load() && ld_adr->is_AddP() &&
1037 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1038 intptr_t ignore = 0;
1039 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1040 if (base != NULL && base->is_Proj() &&
1041 base->as_Proj()->_con == TypeFunc::Parms &&
1042 base->in(0)->is_CallStaticJava() &&
1043 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1044 return base->in(0)->in(TypeFunc::Parms);
1045 }
1046 }
1047
1048 break;
1049 }
1050
1051 return NULL;
1052 }
1053
1054 //----------------------is_instance_field_load_with_local_phi------------------
1055 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1056 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1057 in(Address)->is_AddP() ) {
1058 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1059 // Only instances and boxed values.
1060 if( t_oop != NULL &&
1061 (t_oop->is_ptr_to_boxed_value() ||
1062 t_oop->is_known_instance_field()) &&
1063 t_oop->offset() != Type::OffsetBot &&
1064 t_oop->offset() != Type::OffsetTop) {
1065 return true;
1066 }
1067 }
1068 return false;
1069 }
1070
1071 //------------------------------Identity---------------------------------------
1072 // Loads are identity if previous store is to same address
1073 Node* LoadNode::Identity(PhaseGVN* phase) {
1074 // If the previous store-maker is the right kind of Store, and the store is
1075 // to the same address, then we are equal to the value stored.
1076 Node* mem = in(Memory);
1077 Node* value = can_see_stored_value(mem, phase);
1078 if( value ) {
1079 // byte, short & char stores truncate naturally.
1080 // A load has to load the truncated value which requires
1081 // some sort of masking operation and that requires an
1082 // Ideal call instead of an Identity call.
1083 if (memory_size() < BytesPerInt) {
1084 // If the input to the store does not fit with the load's result type,
1085 // it must be truncated via an Ideal call.
1086 if (!phase->type(value)->higher_equal(phase->type(this)))
1087 return this;
1088 }
1089 // (This works even when value is a Con, but LoadNode::Value
1090 // usually runs first, producing the singleton type of the Con.)
1091 return value;
1092 }
1093
1094 // Search for an existing data phi which was generated before for the same
1095 // instance's field to avoid infinite generation of phis in a loop.
1096 Node *region = mem->in(0);
1097 if (is_instance_field_load_with_local_phi(region)) {
1098 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1099 int this_index = phase->C->get_alias_index(addr_t);
1100 int this_offset = addr_t->offset();
1101 int this_iid = addr_t->instance_id();
1102 if (!addr_t->is_known_instance() &&
1103 addr_t->is_ptr_to_boxed_value()) {
1104 // Use _idx of address base (could be Phi node) for boxed values.
1105 intptr_t ignore = 0;
1106 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1107 this_iid = base->_idx;
1108 }
1109 const Type* this_type = bottom_type();
1110 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1111 Node* phi = region->fast_out(i);
1112 if (phi->is_Phi() && phi != mem &&
1113 phi->as_Phi()->is_same_inst_field(this_type, this_iid, this_index, this_offset)) {
1114 return phi;
1115 }
1116 }
1117 }
1118
1119 return this;
1120 }
1121
1122 // We're loading from an object which has autobox behaviour.
1123 // If this object is result of a valueOf call we'll have a phi
1124 // merging a newly allocated object and a load from the cache.
1125 // We want to replace this load with the original incoming
1126 // argument to the valueOf call.
1127 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1128 assert(phase->C->eliminate_boxing(), "sanity");
1129 intptr_t ignore = 0;
1130 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1131 if ((base == NULL) || base->is_Phi()) {
1132 // Push the loads from the phi that comes from valueOf up
1133 // through it to allow elimination of the loads and the recovery
1134 // of the original value. It is done in split_through_phi().
1135 return NULL;
1136 } else if (base->is_Load() ||
1137 base->is_DecodeN() && base->in(1)->is_Load()) {
1138 // Eliminate the load of boxed value for integer types from the cache
1139 // array by deriving the value from the index into the array.
1140 // Capture the offset of the load and then reverse the computation.
1141
1142 // Get LoadN node which loads a boxing object from 'cache' array.
1143 if (base->is_DecodeN()) {
1144 base = base->in(1);
1145 }
1146 if (!base->in(Address)->is_AddP()) {
1147 return NULL; // Complex address
1148 }
1149 AddPNode* address = base->in(Address)->as_AddP();
1150 Node* cache_base = address->in(AddPNode::Base);
1151 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1152 // Get ConP node which is static 'cache' field.
1153 cache_base = cache_base->in(1);
1154 }
1155 if ((cache_base != NULL) && cache_base->is_Con()) {
1156 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1157 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1158 Node* elements[4];
1159 int shift = exact_log2(type2aelembytes(T_OBJECT));
1160 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1161 if ((count > 0) && elements[0]->is_Con() &&
1162 ((count == 1) ||
1163 (count == 2) && elements[1]->Opcode() == Op_LShiftX &&
1164 elements[1]->in(2) == phase->intcon(shift))) {
1165 ciObjArray* array = base_type->const_oop()->as_obj_array();
1166 // Fetch the box object cache[0] at the base of the array and get its value
1167 ciInstance* box = array->obj_at(0)->as_instance();
1168 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1169 assert(ik->is_box_klass(), "sanity");
1170 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1171 if (ik->nof_nonstatic_fields() == 1) {
1172 // This should be true nonstatic_field_at requires calling
1173 // nof_nonstatic_fields so check it anyway
1174 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1175 BasicType bt = c.basic_type();
1176 // Only integer types have boxing cache.
1177 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1178 bt == T_BYTE || bt == T_SHORT ||
1179 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1180 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1181 if (cache_low != (int)cache_low) {
1182 return NULL; // should not happen since cache is array indexed by value
1183 }
1184 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1185 if (offset != (int)offset) {
1186 return NULL; // should not happen since cache is array indexed by value
1187 }
1188 // Add up all the offsets making of the address of the load
1189 Node* result = elements[0];
1190 for (int i = 1; i < count; i++) {
1191 result = phase->transform(new AddXNode(result, elements[i]));
1192 }
1193 // Remove the constant offset from the address and then
1194 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset)));
1195 // remove the scaling of the offset to recover the original index.
1196 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1197 // Peel the shift off directly but wrap it in a dummy node
1198 // since Ideal can't return existing nodes
1199 result = new RShiftXNode(result->in(1), phase->intcon(0));
1200 } else if (result->is_Add() && result->in(2)->is_Con() &&
1201 result->in(1)->Opcode() == Op_LShiftX &&
1202 result->in(1)->in(2) == phase->intcon(shift)) {
1203 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1204 // but for boxing cache access we know that X<<Z will not overflow
1205 // (there is range check) so we do this optimizatrion by hand here.
1206 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift));
1207 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con));
1208 } else {
1209 result = new RShiftXNode(result, phase->intcon(shift));
1210 }
1211 #ifdef _LP64
1212 if (bt != T_LONG) {
1213 result = new ConvL2INode(phase->transform(result));
1214 }
1215 #else
1216 if (bt == T_LONG) {
1217 result = new ConvI2LNode(phase->transform(result));
1218 }
1219 #endif
1220 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1221 // Need to preserve unboxing load type if it is unsigned.
1222 switch(this->Opcode()) {
1223 case Op_LoadUB:
1224 result = new AndINode(phase->transform(result), phase->intcon(0xFF));
1225 break;
1226 case Op_LoadUS:
1227 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF));
1228 break;
1229 }
1230 return result;
1231 }
1232 }
1233 }
1234 }
1235 }
1236 return NULL;
1237 }
1238
1239 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1240 Node* region = phi->in(0);
1241 if (region == NULL) {
1242 return false; // Wait stable graph
1243 }
1244 uint cnt = phi->req();
1245 for (uint i = 1; i < cnt; i++) {
1246 Node* rc = region->in(i);
1247 if (rc == NULL || phase->type(rc) == Type::TOP)
1248 return false; // Wait stable graph
1249 Node* in = phi->in(i);
1250 if (in == NULL || phase->type(in) == Type::TOP)
1251 return false; // Wait stable graph
1252 }
1253 return true;
1254 }
1255 //------------------------------split_through_phi------------------------------
1256 // Split instance or boxed field load through Phi.
1257 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1258 Node* mem = in(Memory);
1259 Node* address = in(Address);
1260 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1261
1262 assert((t_oop != NULL) &&
1263 (t_oop->is_known_instance_field() ||
1264 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1265
1266 Compile* C = phase->C;
1267 intptr_t ignore = 0;
1268 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1269 bool base_is_phi = (base != NULL) && base->is_Phi();
1270 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1271 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1272 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1273
1274 if (!((mem->is_Phi() || base_is_phi) &&
1275 (load_boxed_values || t_oop->is_known_instance_field()))) {
1276 return NULL; // memory is not Phi
1277 }
1278
1279 if (mem->is_Phi()) {
1280 if (!stable_phi(mem->as_Phi(), phase)) {
1281 return NULL; // Wait stable graph
1282 }
1283 uint cnt = mem->req();
1284 // Check for loop invariant memory.
1285 if (cnt == 3) {
1286 for (uint i = 1; i < cnt; i++) {
1287 Node* in = mem->in(i);
1288 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1289 if (m == mem) {
1290 set_req(Memory, mem->in(cnt - i));
1291 return this; // made change
1292 }
1293 }
1294 }
1295 }
1296 if (base_is_phi) {
1297 if (!stable_phi(base->as_Phi(), phase)) {
1298 return NULL; // Wait stable graph
1299 }
1300 uint cnt = base->req();
1301 // Check for loop invariant memory.
1302 if (cnt == 3) {
1303 for (uint i = 1; i < cnt; i++) {
1304 if (base->in(i) == base) {
1305 return NULL; // Wait stable graph
1306 }
1307 }
1308 }
1309 }
1310
1311 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1312
1313 // Split through Phi (see original code in loopopts.cpp).
1314 assert(C->have_alias_type(t_oop), "instance should have alias type");
1315
1316 // Do nothing here if Identity will find a value
1317 // (to avoid infinite chain of value phis generation).
1318 if (!phase->eqv(this, this->Identity(phase)))
1319 return NULL;
1320
1321 // Select Region to split through.
1322 Node* region;
1323 if (!base_is_phi) {
1324 assert(mem->is_Phi(), "sanity");
1325 region = mem->in(0);
1326 // Skip if the region dominates some control edge of the address.
1327 if (!MemNode::all_controls_dominate(address, region))
1328 return NULL;
1329 } else if (!mem->is_Phi()) {
1330 assert(base_is_phi, "sanity");
1331 region = base->in(0);
1332 // Skip if the region dominates some control edge of the memory.
1333 if (!MemNode::all_controls_dominate(mem, region))
1334 return NULL;
1335 } else if (base->in(0) != mem->in(0)) {
1336 assert(base_is_phi && mem->is_Phi(), "sanity");
1337 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1338 region = base->in(0);
1339 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1340 region = mem->in(0);
1341 } else {
1342 return NULL; // complex graph
1343 }
1344 } else {
1345 assert(base->in(0) == mem->in(0), "sanity");
1346 region = mem->in(0);
1347 }
1348
1349 const Type* this_type = this->bottom_type();
1350 int this_index = C->get_alias_index(t_oop);
1351 int this_offset = t_oop->offset();
1352 int this_iid = t_oop->instance_id();
1353 if (!t_oop->is_known_instance() && load_boxed_values) {
1354 // Use _idx of address base for boxed values.
1355 this_iid = base->_idx;
1356 }
1357 PhaseIterGVN* igvn = phase->is_IterGVN();
1358 Node* phi = new PhiNode(region, this_type, NULL, this_iid, this_index, this_offset);
1359 for (uint i = 1; i < region->req(); i++) {
1360 Node* x;
1361 Node* the_clone = NULL;
1362 if (region->in(i) == C->top()) {
1363 x = C->top(); // Dead path? Use a dead data op
1364 } else {
1365 x = this->clone(); // Else clone up the data op
1366 the_clone = x; // Remember for possible deletion.
1367 // Alter data node to use pre-phi inputs
1368 if (this->in(0) == region) {
1369 x->set_req(0, region->in(i));
1370 } else {
1371 x->set_req(0, NULL);
1372 }
1373 if (mem->is_Phi() && (mem->in(0) == region)) {
1374 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1375 }
1376 if (address->is_Phi() && address->in(0) == region) {
1377 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1378 }
1379 if (base_is_phi && (base->in(0) == region)) {
1380 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1381 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1382 x->set_req(Address, adr_x);
1383 }
1384 }
1385 // Check for a 'win' on some paths
1386 const Type *t = x->Value(igvn);
1387
1388 bool singleton = t->singleton();
1389
1390 // See comments in PhaseIdealLoop::split_thru_phi().
1391 if (singleton && t == Type::TOP) {
1392 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1393 }
1394
1395 if (singleton) {
1396 x = igvn->makecon(t);
1397 } else {
1398 // We now call Identity to try to simplify the cloned node.
1399 // Note that some Identity methods call phase->type(this).
1400 // Make sure that the type array is big enough for
1401 // our new node, even though we may throw the node away.
1402 // (This tweaking with igvn only works because x is a new node.)
1403 igvn->set_type(x, t);
1404 // If x is a TypeNode, capture any more-precise type permanently into Node
1405 // otherwise it will be not updated during igvn->transform since
1406 // igvn->type(x) is set to x->Value() already.
1407 x->raise_bottom_type(t);
1408 Node *y = x->Identity(igvn);
1409 if (y != x) {
1410 x = y;
1411 } else {
1412 y = igvn->hash_find_insert(x);
1413 if (y) {
1414 x = y;
1415 } else {
1416 // Else x is a new node we are keeping
1417 // We do not need register_new_node_with_optimizer
1418 // because set_type has already been called.
1419 igvn->_worklist.push(x);
1420 }
1421 }
1422 }
1423 if (x != the_clone && the_clone != NULL) {
1424 igvn->remove_dead_node(the_clone);
1425 }
1426 phi->set_req(i, x);
1427 }
1428 // Record Phi
1429 igvn->register_new_node_with_optimizer(phi);
1430 return phi;
1431 }
1432
1433 //------------------------------Ideal------------------------------------------
1434 // If the load is from Field memory and the pointer is non-null, it might be possible to
1435 // zero out the control input.
1436 // If the offset is constant and the base is an object allocation,
1437 // try to hook me up to the exact initializing store.
1438 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1439 Node* p = MemNode::Ideal_common(phase, can_reshape);
1440 if (p) return (p == NodeSentinel) ? NULL : p;
1441
1442 Node* ctrl = in(MemNode::Control);
1443 Node* address = in(MemNode::Address);
1444 bool progress = false;
1445
1446 // Skip up past a SafePoint control. Cannot do this for Stores because
1447 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1448 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1449 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1450 ctrl = ctrl->in(0);
1451 set_req(MemNode::Control,ctrl);
1452 progress = true;
1453 }
1454
1455 intptr_t ignore = 0;
1456 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1457 if (base != NULL
1458 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1459 // Check for useless control edge in some common special cases
1460 if (in(MemNode::Control) != NULL
1461 && can_remove_control()
1462 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1463 && all_controls_dominate(base, phase->C->start())) {
1464 // A method-invariant, non-null address (constant or 'this' argument).
1465 set_req(MemNode::Control, NULL);
1466 progress = true;
1467 }
1468 }
1469
1470 Node* mem = in(MemNode::Memory);
1471 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1472
1473 if (can_reshape && (addr_t != NULL)) {
1474 // try to optimize our memory input
1475 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1476 if (opt_mem != mem) {
1477 set_req(MemNode::Memory, opt_mem);
1478 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1479 return this;
1480 }
1481 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1482 if ((t_oop != NULL) &&
1483 (t_oop->is_known_instance_field() ||
1484 t_oop->is_ptr_to_boxed_value())) {
1485 PhaseIterGVN *igvn = phase->is_IterGVN();
1486 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1487 // Delay this transformation until memory Phi is processed.
1488 phase->is_IterGVN()->_worklist.push(this);
1489 return NULL;
1490 }
1491 // Split instance field load through Phi.
1492 Node* result = split_through_phi(phase);
1493 if (result != NULL) return result;
1494
1495 if (t_oop->is_ptr_to_boxed_value()) {
1496 Node* result = eliminate_autobox(phase);
1497 if (result != NULL) return result;
1498 }
1499 }
1500 }
1501
1502 // Is there a dominating load that loads the same value? Leave
1503 // anything that is not a load of a field/array element (like
1504 // barriers etc.) alone
1505 if (in(0) != NULL && adr_type() != TypeRawPtr::BOTTOM && can_reshape) {
1506 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1507 Node *use = mem->fast_out(i);
1508 if (use != this &&
1509 use->Opcode() == Opcode() &&
1510 use->in(0) != NULL &&
1511 use->in(0) != in(0) &&
1512 use->in(Address) == in(Address)) {
1513 Node* ctl = in(0);
1514 for (int i = 0; i < 10 && ctl != NULL; i++) {
1515 ctl = IfNode::up_one_dom(ctl);
1516 if (ctl == use->in(0)) {
1517 set_req(0, use->in(0));
1518 return this;
1519 }
1520 }
1521 }
1522 }
1523 }
1524
1525 // Check for prior store with a different base or offset; make Load
1526 // independent. Skip through any number of them. Bail out if the stores
1527 // are in an endless dead cycle and report no progress. This is a key
1528 // transform for Reflection. However, if after skipping through the Stores
1529 // we can't then fold up against a prior store do NOT do the transform as
1530 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1531 // array memory alive twice: once for the hoisted Load and again after the
1532 // bypassed Store. This situation only works if EVERYBODY who does
1533 // anti-dependence work knows how to bypass. I.e. we need all
1534 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1535 // the alias index stuff. So instead, peek through Stores and IFF we can
1536 // fold up, do so.
1537 Node* prev_mem = find_previous_store(phase);
1538 if (prev_mem != NULL) {
1539 Node* value = can_see_arraycopy_value(prev_mem, phase);
1540 if (value != NULL) {
1541 return value;
1542 }
1543 }
1544 // Steps (a), (b): Walk past independent stores to find an exact match.
1545 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1546 // (c) See if we can fold up on the spot, but don't fold up here.
1547 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1548 // just return a prior value, which is done by Identity calls.
1549 if (can_see_stored_value(prev_mem, phase)) {
1550 // Make ready for step (d):
1551 set_req(MemNode::Memory, prev_mem);
1552 return this;
1553 }
1554 }
1555
1556 return progress ? this : NULL;
1557 }
1558
1559 // Helper to recognize certain Klass fields which are invariant across
1560 // some group of array types (e.g., int[] or all T[] where T < Object).
1561 const Type*
1562 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1563 ciKlass* klass) const {
1564 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1565 // The field is Klass::_modifier_flags. Return its (constant) value.
1566 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1567 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1568 return TypeInt::make(klass->modifier_flags());
1569 }
1570 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1571 // The field is Klass::_access_flags. Return its (constant) value.
1572 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1573 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1574 return TypeInt::make(klass->access_flags());
1575 }
1576 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1577 // The field is Klass::_layout_helper. Return its constant value if known.
1578 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1579 return TypeInt::make(klass->layout_helper());
1580 }
1581
1582 // No match.
1583 return NULL;
1584 }
1585
1586 static bool is_mismatched_access(ciConstant con, BasicType loadbt) {
1587 BasicType conbt = con.basic_type();
1588 switch (conbt) {
1589 case T_BOOLEAN: conbt = T_BYTE; break;
1590 case T_ARRAY: conbt = T_OBJECT; break;
1591 }
1592 switch (loadbt) {
1593 case T_BOOLEAN: loadbt = T_BYTE; break;
1594 case T_NARROWOOP: loadbt = T_OBJECT; break;
1595 case T_ARRAY: loadbt = T_OBJECT; break;
1596 case T_ADDRESS: loadbt = T_OBJECT; break;
1597 }
1598 return (conbt != loadbt);
1599 }
1600
1601 // Try to constant-fold a stable array element.
1602 static const Type* fold_stable_ary_elem(const TypeAryPtr* ary, int off, BasicType loadbt) {
1603 assert(ary->const_oop(), "array should be constant");
1604 assert(ary->is_stable(), "array should be stable");
1605
1606 // Decode the results of GraphKit::array_element_address.
1607 ciArray* aobj = ary->const_oop()->as_array();
1608 ciConstant con = aobj->element_value_by_offset(off);
1609 if (con.basic_type() != T_ILLEGAL && !con.is_null_or_zero()) {
1610 bool is_mismatched = is_mismatched_access(con, loadbt);
1611 assert(!is_mismatched, "conbt=%s; loadbt=%s", type2name(con.basic_type()), type2name(loadbt));
1612 const Type* con_type = Type::make_from_constant(con);
1613 // Guard against erroneous constant folding.
1614 if (!is_mismatched && con_type != NULL) {
1615 if (con_type->isa_aryptr()) {
1616 // Join with the array element type, in case it is also stable.
1617 int dim = ary->stable_dimension();
1618 con_type = con_type->is_aryptr()->cast_to_stable(true, dim-1);
1619 }
1620 if (loadbt == T_NARROWOOP && con_type->isa_oopptr()) {
1621 con_type = con_type->make_narrowoop();
1622 }
1623 #ifndef PRODUCT
1624 if (TraceIterativeGVN) {
1625 tty->print("FoldStableValues: array element [off=%d]: con_type=", off);
1626 con_type->dump(); tty->cr();
1627 }
1628 #endif //PRODUCT
1629 return con_type;
1630 }
1631 }
1632 return NULL;
1633 }
1634
1635 //------------------------------Value-----------------------------------------
1636 const Type* LoadNode::Value(PhaseGVN* phase) const {
1637 // Either input is TOP ==> the result is TOP
1638 Node* mem = in(MemNode::Memory);
1639 const Type *t1 = phase->type(mem);
1640 if (t1 == Type::TOP) return Type::TOP;
1641 Node* adr = in(MemNode::Address);
1642 const TypePtr* tp = phase->type(adr)->isa_ptr();
1643 if (tp == NULL || tp->empty()) return Type::TOP;
1644 int off = tp->offset();
1645 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1646 Compile* C = phase->C;
1647
1648 // Try to guess loaded type from pointer type
1649 if (tp->isa_aryptr()) {
1650 const TypeAryPtr* ary = tp->is_aryptr();
1651 const Type* t = ary->elem();
1652
1653 // Determine whether the reference is beyond the header or not, by comparing
1654 // the offset against the offset of the start of the array's data.
1655 // Different array types begin at slightly different offsets (12 vs. 16).
1656 // We choose T_BYTE as an example base type that is least restrictive
1657 // as to alignment, which will therefore produce the smallest
1658 // possible base offset.
1659 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1660 const bool off_beyond_header = ((uint)off >= (uint)min_base_off);
1661
1662 // Try to constant-fold a stable array element.
1663 if (FoldStableValues && !is_mismatched_access() && ary->is_stable() && ary->const_oop() != NULL) {
1664 // Make sure the reference is not into the header and the offset is constant
1665 if (off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1666 const Type* con_type = fold_stable_ary_elem(ary, off, memory_type());
1667 if (con_type != NULL) {
1668 return con_type;
1669 }
1670 }
1671 }
1672
1673 // Don't do this for integer types. There is only potential profit if
1674 // the element type t is lower than _type; that is, for int types, if _type is
1675 // more restrictive than t. This only happens here if one is short and the other
1676 // char (both 16 bits), and in those cases we've made an intentional decision
1677 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1678 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1679 //
1680 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1681 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1682 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1683 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1684 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1685 // In fact, that could have been the original type of p1, and p1 could have
1686 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1687 // expression (LShiftL quux 3) independently optimized to the constant 8.
1688 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1689 && (_type->isa_vect() == NULL)
1690 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1691 // t might actually be lower than _type, if _type is a unique
1692 // concrete subclass of abstract class t.
1693 if (off_beyond_header) { // is the offset beyond the header?
1694 const Type* jt = t->join_speculative(_type);
1695 // In any case, do not allow the join, per se, to empty out the type.
1696 if (jt->empty() && !t->empty()) {
1697 // This can happen if a interface-typed array narrows to a class type.
1698 jt = _type;
1699 }
1700 #ifdef ASSERT
1701 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1702 // The pointers in the autobox arrays are always non-null
1703 Node* base = adr->in(AddPNode::Base);
1704 if ((base != NULL) && base->is_DecodeN()) {
1705 // Get LoadN node which loads IntegerCache.cache field
1706 base = base->in(1);
1707 }
1708 if ((base != NULL) && base->is_Con()) {
1709 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1710 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1711 // It could be narrow oop
1712 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1713 }
1714 }
1715 }
1716 #endif
1717 return jt;
1718 }
1719 }
1720 } else if (tp->base() == Type::InstPtr) {
1721 ciEnv* env = C->env();
1722 const TypeInstPtr* tinst = tp->is_instptr();
1723 ciKlass* klass = tinst->klass();
1724 assert( off != Type::OffsetBot ||
1725 // arrays can be cast to Objects
1726 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1727 // unsafe field access may not have a constant offset
1728 C->has_unsafe_access(),
1729 "Field accesses must be precise" );
1730 // For oop loads, we expect the _type to be precise.
1731 // Optimizations for constant objects
1732 ciObject* const_oop = tinst->const_oop();
1733 if (const_oop != NULL) {
1734 // For constant CallSites treat the target field as a compile time constant.
1735 if (const_oop->is_call_site()) {
1736 ciCallSite* call_site = const_oop->as_call_site();
1737 ciField* field = call_site->klass()->as_instance_klass()->get_field_by_offset(off, /*is_static=*/ false);
1738 if (field != NULL && field->is_call_site_target()) {
1739 ciMethodHandle* target = call_site->get_target();
1740 if (target != NULL) { // just in case
1741 ciConstant constant(T_OBJECT, target);
1742 const Type* t;
1743 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
1744 t = TypeNarrowOop::make_from_constant(constant.as_object(), true);
1745 } else {
1746 t = TypeOopPtr::make_from_constant(constant.as_object(), true);
1747 }
1748 // Add a dependence for invalidation of the optimization.
1749 if (!call_site->is_constant_call_site()) {
1750 C->dependencies()->assert_call_site_target_value(call_site, target);
1751 }
1752 return t;
1753 }
1754 }
1755 }
1756 }
1757 } else if (tp->base() == Type::KlassPtr) {
1758 assert( off != Type::OffsetBot ||
1759 // arrays can be cast to Objects
1760 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1761 // also allow array-loading from the primary supertype
1762 // array during subtype checks
1763 Opcode() == Op_LoadKlass,
1764 "Field accesses must be precise" );
1765 // For klass/static loads, we expect the _type to be precise
1766 }
1767
1768 const TypeKlassPtr *tkls = tp->isa_klassptr();
1769 if (tkls != NULL && !StressReflectiveCode) {
1770 ciKlass* klass = tkls->klass();
1771 if (klass->is_loaded() && tkls->klass_is_exact()) {
1772 // We are loading a field from a Klass metaobject whose identity
1773 // is known at compile time (the type is "exact" or "precise").
1774 // Check for fields we know are maintained as constants by the VM.
1775 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1776 // The field is Klass::_super_check_offset. Return its (constant) value.
1777 // (Folds up type checking code.)
1778 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1779 return TypeInt::make(klass->super_check_offset());
1780 }
1781 // Compute index into primary_supers array
1782 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1783 // Check for overflowing; use unsigned compare to handle the negative case.
1784 if( depth < ciKlass::primary_super_limit() ) {
1785 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1786 // (Folds up type checking code.)
1787 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1788 ciKlass *ss = klass->super_of_depth(depth);
1789 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1790 }
1791 const Type* aift = load_array_final_field(tkls, klass);
1792 if (aift != NULL) return aift;
1793 if (tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1794 // The field is Klass::_java_mirror. Return its (constant) value.
1795 // (Folds up the 2nd indirection in anObjConstant.getClass().)
1796 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1797 return TypeInstPtr::make(klass->java_mirror());
1798 }
1799 }
1800
1801 // We can still check if we are loading from the primary_supers array at a
1802 // shallow enough depth. Even though the klass is not exact, entries less
1803 // than or equal to its super depth are correct.
1804 if (klass->is_loaded() ) {
1805 ciType *inner = klass;
1806 while( inner->is_obj_array_klass() )
1807 inner = inner->as_obj_array_klass()->base_element_type();
1808 if( inner->is_instance_klass() &&
1809 !inner->as_instance_klass()->flags().is_interface() ) {
1810 // Compute index into primary_supers array
1811 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1812 // Check for overflowing; use unsigned compare to handle the negative case.
1813 if( depth < ciKlass::primary_super_limit() &&
1814 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1815 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1816 // (Folds up type checking code.)
1817 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1818 ciKlass *ss = klass->super_of_depth(depth);
1819 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1820 }
1821 }
1822 }
1823
1824 // If the type is enough to determine that the thing is not an array,
1825 // we can give the layout_helper a positive interval type.
1826 // This will help short-circuit some reflective code.
1827 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1828 && !klass->is_array_klass() // not directly typed as an array
1829 && !klass->is_interface() // specifically not Serializable & Cloneable
1830 && !klass->is_java_lang_Object() // not the supertype of all T[]
1831 ) {
1832 // Note: When interfaces are reliable, we can narrow the interface
1833 // test to (klass != Serializable && klass != Cloneable).
1834 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1835 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1836 // The key property of this type is that it folds up tests
1837 // for array-ness, since it proves that the layout_helper is positive.
1838 // Thus, a generic value like the basic object layout helper works fine.
1839 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1840 }
1841 }
1842
1843 // If we are loading from a freshly-allocated object, produce a zero,
1844 // if the load is provably beyond the header of the object.
1845 // (Also allow a variable load from a fresh array to produce zero.)
1846 const TypeOopPtr *tinst = tp->isa_oopptr();
1847 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1848 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1849 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1850 Node* value = can_see_stored_value(mem,phase);
1851 if (value != NULL && value->is_Con()) {
1852 assert(value->bottom_type()->higher_equal(_type),"sanity");
1853 return value->bottom_type();
1854 }
1855 }
1856
1857 if (is_instance) {
1858 // If we have an instance type and our memory input is the
1859 // programs's initial memory state, there is no matching store,
1860 // so just return a zero of the appropriate type
1861 Node *mem = in(MemNode::Memory);
1862 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1863 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1864 return Type::get_zero_type(_type->basic_type());
1865 }
1866 }
1867 return _type;
1868 }
1869
1870 //------------------------------match_edge-------------------------------------
1871 // Do we Match on this edge index or not? Match only the address.
1872 uint LoadNode::match_edge(uint idx) const {
1873 return idx == MemNode::Address;
1874 }
1875
1876 //--------------------------LoadBNode::Ideal--------------------------------------
1877 //
1878 // If the previous store is to the same address as this load,
1879 // and the value stored was larger than a byte, replace this load
1880 // with the value stored truncated to a byte. If no truncation is
1881 // needed, the replacement is done in LoadNode::Identity().
1882 //
1883 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1884 Node* mem = in(MemNode::Memory);
1885 Node* value = can_see_stored_value(mem,phase);
1886 if( value && !phase->type(value)->higher_equal( _type ) ) {
1887 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
1888 return new RShiftINode(result, phase->intcon(24));
1889 }
1890 // Identity call will handle the case where truncation is not needed.
1891 return LoadNode::Ideal(phase, can_reshape);
1892 }
1893
1894 const Type* LoadBNode::Value(PhaseGVN* phase) const {
1895 Node* mem = in(MemNode::Memory);
1896 Node* value = can_see_stored_value(mem,phase);
1897 if (value != NULL && value->is_Con() &&
1898 !value->bottom_type()->higher_equal(_type)) {
1899 // If the input to the store does not fit with the load's result type,
1900 // it must be truncated. We can't delay until Ideal call since
1901 // a singleton Value is needed for split_thru_phi optimization.
1902 int con = value->get_int();
1903 return TypeInt::make((con << 24) >> 24);
1904 }
1905 return LoadNode::Value(phase);
1906 }
1907
1908 //--------------------------LoadUBNode::Ideal-------------------------------------
1909 //
1910 // If the previous store is to the same address as this load,
1911 // and the value stored was larger than a byte, replace this load
1912 // with the value stored truncated to a byte. If no truncation is
1913 // needed, the replacement is done in LoadNode::Identity().
1914 //
1915 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1916 Node* mem = in(MemNode::Memory);
1917 Node* value = can_see_stored_value(mem, phase);
1918 if (value && !phase->type(value)->higher_equal(_type))
1919 return new AndINode(value, phase->intcon(0xFF));
1920 // Identity call will handle the case where truncation is not needed.
1921 return LoadNode::Ideal(phase, can_reshape);
1922 }
1923
1924 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
1925 Node* mem = in(MemNode::Memory);
1926 Node* value = can_see_stored_value(mem,phase);
1927 if (value != NULL && value->is_Con() &&
1928 !value->bottom_type()->higher_equal(_type)) {
1929 // If the input to the store does not fit with the load's result type,
1930 // it must be truncated. We can't delay until Ideal call since
1931 // a singleton Value is needed for split_thru_phi optimization.
1932 int con = value->get_int();
1933 return TypeInt::make(con & 0xFF);
1934 }
1935 return LoadNode::Value(phase);
1936 }
1937
1938 //--------------------------LoadUSNode::Ideal-------------------------------------
1939 //
1940 // If the previous store is to the same address as this load,
1941 // and the value stored was larger than a char, replace this load
1942 // with the value stored truncated to a char. If no truncation is
1943 // needed, the replacement is done in LoadNode::Identity().
1944 //
1945 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1946 Node* mem = in(MemNode::Memory);
1947 Node* value = can_see_stored_value(mem,phase);
1948 if( value && !phase->type(value)->higher_equal( _type ) )
1949 return new AndINode(value,phase->intcon(0xFFFF));
1950 // Identity call will handle the case where truncation is not needed.
1951 return LoadNode::Ideal(phase, can_reshape);
1952 }
1953
1954 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
1955 Node* mem = in(MemNode::Memory);
1956 Node* value = can_see_stored_value(mem,phase);
1957 if (value != NULL && value->is_Con() &&
1958 !value->bottom_type()->higher_equal(_type)) {
1959 // If the input to the store does not fit with the load's result type,
1960 // it must be truncated. We can't delay until Ideal call since
1961 // a singleton Value is needed for split_thru_phi optimization.
1962 int con = value->get_int();
1963 return TypeInt::make(con & 0xFFFF);
1964 }
1965 return LoadNode::Value(phase);
1966 }
1967
1968 //--------------------------LoadSNode::Ideal--------------------------------------
1969 //
1970 // If the previous store is to the same address as this load,
1971 // and the value stored was larger than a short, replace this load
1972 // with the value stored truncated to a short. If no truncation is
1973 // needed, the replacement is done in LoadNode::Identity().
1974 //
1975 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1976 Node* mem = in(MemNode::Memory);
1977 Node* value = can_see_stored_value(mem,phase);
1978 if( value && !phase->type(value)->higher_equal( _type ) ) {
1979 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
1980 return new RShiftINode(result, phase->intcon(16));
1981 }
1982 // Identity call will handle the case where truncation is not needed.
1983 return LoadNode::Ideal(phase, can_reshape);
1984 }
1985
1986 const Type* LoadSNode::Value(PhaseGVN* phase) const {
1987 Node* mem = in(MemNode::Memory);
1988 Node* value = can_see_stored_value(mem,phase);
1989 if (value != NULL && value->is_Con() &&
1990 !value->bottom_type()->higher_equal(_type)) {
1991 // If the input to the store does not fit with the load's result type,
1992 // it must be truncated. We can't delay until Ideal call since
1993 // a singleton Value is needed for split_thru_phi optimization.
1994 int con = value->get_int();
1995 return TypeInt::make((con << 16) >> 16);
1996 }
1997 return LoadNode::Value(phase);
1998 }
1999
2000 //=============================================================================
2001 //----------------------------LoadKlassNode::make------------------------------
2002 // Polymorphic factory method:
2003 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2004 // sanity check the alias category against the created node type
2005 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2006 assert(adr_type != NULL, "expecting TypeKlassPtr");
2007 #ifdef _LP64
2008 if (adr_type->is_ptr_to_narrowklass()) {
2009 assert(UseCompressedClassPointers, "no compressed klasses");
2010 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2011 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2012 }
2013 #endif
2014 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2015 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2016 }
2017
2018 //------------------------------Value------------------------------------------
2019 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2020 return klass_value_common(phase);
2021 }
2022
2023 // In most cases, LoadKlassNode does not have the control input set. If the control
2024 // input is set, it must not be removed (by LoadNode::Ideal()).
2025 bool LoadKlassNode::can_remove_control() const {
2026 return false;
2027 }
2028
2029 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2030 // Either input is TOP ==> the result is TOP
2031 const Type *t1 = phase->type( in(MemNode::Memory) );
2032 if (t1 == Type::TOP) return Type::TOP;
2033 Node *adr = in(MemNode::Address);
2034 const Type *t2 = phase->type( adr );
2035 if (t2 == Type::TOP) return Type::TOP;
2036 const TypePtr *tp = t2->is_ptr();
2037 if (TypePtr::above_centerline(tp->ptr()) ||
2038 tp->ptr() == TypePtr::Null) return Type::TOP;
2039
2040 // Return a more precise klass, if possible
2041 const TypeInstPtr *tinst = tp->isa_instptr();
2042 if (tinst != NULL) {
2043 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2044 int offset = tinst->offset();
2045 if (ik == phase->C->env()->Class_klass()
2046 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2047 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2048 // We are loading a special hidden field from a Class mirror object,
2049 // the field which points to the VM's Klass metaobject.
2050 ciType* t = tinst->java_mirror_type();
2051 // java_mirror_type returns non-null for compile-time Class constants.
2052 if (t != NULL) {
2053 // constant oop => constant klass
2054 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2055 if (t->is_void()) {
2056 // We cannot create a void array. Since void is a primitive type return null
2057 // klass. Users of this result need to do a null check on the returned klass.
2058 return TypePtr::NULL_PTR;
2059 }
2060 return TypeKlassPtr::make(ciArrayKlass::make(t));
2061 }
2062 if (!t->is_klass()) {
2063 // a primitive Class (e.g., int.class) has NULL for a klass field
2064 return TypePtr::NULL_PTR;
2065 }
2066 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2067 return TypeKlassPtr::make(t->as_klass());
2068 }
2069 // non-constant mirror, so we can't tell what's going on
2070 }
2071 if( !ik->is_loaded() )
2072 return _type; // Bail out if not loaded
2073 if (offset == oopDesc::klass_offset_in_bytes()) {
2074 if (tinst->klass_is_exact()) {
2075 return TypeKlassPtr::make(ik);
2076 }
2077 // See if we can become precise: no subklasses and no interface
2078 // (Note: We need to support verified interfaces.)
2079 if (!ik->is_interface() && !ik->has_subklass()) {
2080 //assert(!UseExactTypes, "this code should be useless with exact types");
2081 // Add a dependence; if any subclass added we need to recompile
2082 if (!ik->is_final()) {
2083 // %%% should use stronger assert_unique_concrete_subtype instead
2084 phase->C->dependencies()->assert_leaf_type(ik);
2085 }
2086 // Return precise klass
2087 return TypeKlassPtr::make(ik);
2088 }
2089
2090 // Return root of possible klass
2091 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2092 }
2093 }
2094
2095 // Check for loading klass from an array
2096 const TypeAryPtr *tary = tp->isa_aryptr();
2097 if( tary != NULL ) {
2098 ciKlass *tary_klass = tary->klass();
2099 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2100 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2101 if (tary->klass_is_exact()) {
2102 return TypeKlassPtr::make(tary_klass);
2103 }
2104 ciArrayKlass *ak = tary->klass()->as_array_klass();
2105 // If the klass is an object array, we defer the question to the
2106 // array component klass.
2107 if( ak->is_obj_array_klass() ) {
2108 assert( ak->is_loaded(), "" );
2109 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2110 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2111 ciInstanceKlass* ik = base_k->as_instance_klass();
2112 // See if we can become precise: no subklasses and no interface
2113 if (!ik->is_interface() && !ik->has_subklass()) {
2114 //assert(!UseExactTypes, "this code should be useless with exact types");
2115 // Add a dependence; if any subclass added we need to recompile
2116 if (!ik->is_final()) {
2117 phase->C->dependencies()->assert_leaf_type(ik);
2118 }
2119 // Return precise array klass
2120 return TypeKlassPtr::make(ak);
2121 }
2122 }
2123 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2124 } else { // Found a type-array?
2125 //assert(!UseExactTypes, "this code should be useless with exact types");
2126 assert( ak->is_type_array_klass(), "" );
2127 return TypeKlassPtr::make(ak); // These are always precise
2128 }
2129 }
2130 }
2131
2132 // Check for loading klass from an array klass
2133 const TypeKlassPtr *tkls = tp->isa_klassptr();
2134 if (tkls != NULL && !StressReflectiveCode) {
2135 ciKlass* klass = tkls->klass();
2136 if( !klass->is_loaded() )
2137 return _type; // Bail out if not loaded
2138 if( klass->is_obj_array_klass() &&
2139 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2140 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2141 // // Always returning precise element type is incorrect,
2142 // // e.g., element type could be object and array may contain strings
2143 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2144
2145 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2146 // according to the element type's subclassing.
2147 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2148 }
2149 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2150 tkls->offset() == in_bytes(Klass::super_offset())) {
2151 ciKlass* sup = klass->as_instance_klass()->super();
2152 // The field is Klass::_super. Return its (constant) value.
2153 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2154 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2155 }
2156 }
2157
2158 // Bailout case
2159 return LoadNode::Value(phase);
2160 }
2161
2162 //------------------------------Identity---------------------------------------
2163 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2164 // Also feed through the klass in Allocate(...klass...)._klass.
2165 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2166 return klass_identity_common(phase);
2167 }
2168
2169 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2170 Node* x = LoadNode::Identity(phase);
2171 if (x != this) return x;
2172
2173 // Take apart the address into an oop and and offset.
2174 // Return 'this' if we cannot.
2175 Node* adr = in(MemNode::Address);
2176 intptr_t offset = 0;
2177 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2178 if (base == NULL) return this;
2179 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2180 if (toop == NULL) return this;
2181
2182 // We can fetch the klass directly through an AllocateNode.
2183 // This works even if the klass is not constant (clone or newArray).
2184 if (offset == oopDesc::klass_offset_in_bytes()) {
2185 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2186 if (allocated_klass != NULL) {
2187 return allocated_klass;
2188 }
2189 }
2190
2191 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2192 // See inline_native_Class_query for occurrences of these patterns.
2193 // Java Example: x.getClass().isAssignableFrom(y)
2194 //
2195 // This improves reflective code, often making the Class
2196 // mirror go completely dead. (Current exception: Class
2197 // mirrors may appear in debug info, but we could clean them out by
2198 // introducing a new debug info operator for Klass*.java_mirror).
2199 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2200 && offset == java_lang_Class::klass_offset_in_bytes()) {
2201 // We are loading a special hidden field from a Class mirror,
2202 // the field which points to its Klass or ArrayKlass metaobject.
2203 if (base->is_Load()) {
2204 Node* adr2 = base->in(MemNode::Address);
2205 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2206 if (tkls != NULL && !tkls->empty()
2207 && (tkls->klass()->is_instance_klass() ||
2208 tkls->klass()->is_array_klass())
2209 && adr2->is_AddP()
2210 ) {
2211 int mirror_field = in_bytes(Klass::java_mirror_offset());
2212 if (tkls->offset() == mirror_field) {
2213 return adr2->in(AddPNode::Base);
2214 }
2215 }
2216 }
2217 }
2218
2219 return this;
2220 }
2221
2222
2223 //------------------------------Value------------------------------------------
2224 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2225 const Type *t = klass_value_common(phase);
2226 if (t == Type::TOP)
2227 return t;
2228
2229 return t->make_narrowklass();
2230 }
2231
2232 //------------------------------Identity---------------------------------------
2233 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2234 // Also feed through the klass in Allocate(...klass...)._klass.
2235 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2236 Node *x = klass_identity_common(phase);
2237
2238 const Type *t = phase->type( x );
2239 if( t == Type::TOP ) return x;
2240 if( t->isa_narrowklass()) return x;
2241 assert (!t->isa_narrowoop(), "no narrow oop here");
2242
2243 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2244 }
2245
2246 //------------------------------Value-----------------------------------------
2247 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2248 // Either input is TOP ==> the result is TOP
2249 const Type *t1 = phase->type( in(MemNode::Memory) );
2250 if( t1 == Type::TOP ) return Type::TOP;
2251 Node *adr = in(MemNode::Address);
2252 const Type *t2 = phase->type( adr );
2253 if( t2 == Type::TOP ) return Type::TOP;
2254 const TypePtr *tp = t2->is_ptr();
2255 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2256 const TypeAryPtr *tap = tp->isa_aryptr();
2257 if( !tap ) return _type;
2258 return tap->size();
2259 }
2260
2261 //-------------------------------Ideal---------------------------------------
2262 // Feed through the length in AllocateArray(...length...)._length.
2263 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2264 Node* p = MemNode::Ideal_common(phase, can_reshape);
2265 if (p) return (p == NodeSentinel) ? NULL : p;
2266
2267 // Take apart the address into an oop and and offset.
2268 // Return 'this' if we cannot.
2269 Node* adr = in(MemNode::Address);
2270 intptr_t offset = 0;
2271 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2272 if (base == NULL) return NULL;
2273 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2274 if (tary == NULL) return NULL;
2275
2276 // We can fetch the length directly through an AllocateArrayNode.
2277 // This works even if the length is not constant (clone or newArray).
2278 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2279 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2280 if (alloc != NULL) {
2281 Node* allocated_length = alloc->Ideal_length();
2282 Node* len = alloc->make_ideal_length(tary, phase);
2283 if (allocated_length != len) {
2284 // New CastII improves on this.
2285 return len;
2286 }
2287 }
2288 }
2289
2290 return NULL;
2291 }
2292
2293 //------------------------------Identity---------------------------------------
2294 // Feed through the length in AllocateArray(...length...)._length.
2295 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2296 Node* x = LoadINode::Identity(phase);
2297 if (x != this) return x;
2298
2299 // Take apart the address into an oop and and offset.
2300 // Return 'this' if we cannot.
2301 Node* adr = in(MemNode::Address);
2302 intptr_t offset = 0;
2303 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2304 if (base == NULL) return this;
2305 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2306 if (tary == NULL) return this;
2307
2308 // We can fetch the length directly through an AllocateArrayNode.
2309 // This works even if the length is not constant (clone or newArray).
2310 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2311 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2312 if (alloc != NULL) {
2313 Node* allocated_length = alloc->Ideal_length();
2314 // Do not allow make_ideal_length to allocate a CastII node.
2315 Node* len = alloc->make_ideal_length(tary, phase, false);
2316 if (allocated_length == len) {
2317 // Return allocated_length only if it would not be improved by a CastII.
2318 return allocated_length;
2319 }
2320 }
2321 }
2322
2323 return this;
2324
2325 }
2326
2327 //=============================================================================
2328 //---------------------------StoreNode::make-----------------------------------
2329 // Polymorphic factory method:
2330 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2331 assert((mo == unordered || mo == release), "unexpected");
2332 Compile* C = gvn.C;
2333 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2334 ctl != NULL, "raw memory operations should have control edge");
2335
2336 switch (bt) {
2337 case T_BOOLEAN:
2338 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2339 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2340 case T_CHAR:
2341 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2342 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2343 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2344 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2345 case T_METADATA:
2346 case T_ADDRESS:
2347 case T_OBJECT:
2348 #ifdef _LP64
2349 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2350 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2351 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2352 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2353 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2354 adr->bottom_type()->isa_rawptr())) {
2355 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2356 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2357 }
2358 #endif
2359 {
2360 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2361 }
2362 }
2363 ShouldNotReachHere();
2364 return (StoreNode*)NULL;
2365 }
2366
2367 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2368 bool require_atomic = true;
2369 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2370 }
2371
2372 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2373 bool require_atomic = true;
2374 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2375 }
2376
2377
2378 //--------------------------bottom_type----------------------------------------
2379 const Type *StoreNode::bottom_type() const {
2380 return Type::MEMORY;
2381 }
2382
2383 //------------------------------hash-------------------------------------------
2384 uint StoreNode::hash() const {
2385 // unroll addition of interesting fields
2386 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2387
2388 // Since they are not commoned, do not hash them:
2389 return NO_HASH;
2390 }
2391
2392 //------------------------------Ideal------------------------------------------
2393 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2394 // When a store immediately follows a relevant allocation/initialization,
2395 // try to capture it into the initialization, or hoist it above.
2396 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2397 Node* p = MemNode::Ideal_common(phase, can_reshape);
2398 if (p) return (p == NodeSentinel) ? NULL : p;
2399
2400 Node* mem = in(MemNode::Memory);
2401 Node* address = in(MemNode::Address);
2402 // Back-to-back stores to same address? Fold em up. Generally
2403 // unsafe if I have intervening uses... Also disallowed for StoreCM
2404 // since they must follow each StoreP operation. Redundant StoreCMs
2405 // are eliminated just before matching in final_graph_reshape.
2406 {
2407 Node* st = mem;
2408 // If Store 'st' has more than one use, we cannot fold 'st' away.
2409 // For example, 'st' might be the final state at a conditional
2410 // return. Or, 'st' might be used by some node which is live at
2411 // the same time 'st' is live, which might be unschedulable. So,
2412 // require exactly ONE user until such time as we clone 'mem' for
2413 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2414 // true).
2415 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2416 // Looking at a dead closed cycle of memory?
2417 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2418 assert(Opcode() == st->Opcode() ||
2419 st->Opcode() == Op_StoreVector ||
2420 Opcode() == Op_StoreVector ||
2421 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2422 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2423 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2424 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2425
2426 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2427 st->as_Store()->memory_size() <= this->memory_size()) {
2428 Node* use = st->raw_out(0);
2429 phase->igvn_rehash_node_delayed(use);
2430 if (can_reshape) {
2431 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2432 } else {
2433 // It's OK to do this in the parser, since DU info is always accurate,
2434 // and the parser always refers to nodes via SafePointNode maps.
2435 use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2436 }
2437 return this;
2438 }
2439 st = st->in(MemNode::Memory);
2440 }
2441 }
2442
2443
2444 // Capture an unaliased, unconditional, simple store into an initializer.
2445 // Or, if it is independent of the allocation, hoist it above the allocation.
2446 if (ReduceFieldZeroing && /*can_reshape &&*/
2447 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2448 InitializeNode* init = mem->in(0)->as_Initialize();
2449 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2450 if (offset > 0) {
2451 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2452 // If the InitializeNode captured me, it made a raw copy of me,
2453 // and I need to disappear.
2454 if (moved != NULL) {
2455 // %%% hack to ensure that Ideal returns a new node:
2456 mem = MergeMemNode::make(mem);
2457 return mem; // fold me away
2458 }
2459 }
2460 }
2461
2462 return NULL; // No further progress
2463 }
2464
2465 //------------------------------Value-----------------------------------------
2466 const Type* StoreNode::Value(PhaseGVN* phase) const {
2467 // Either input is TOP ==> the result is TOP
2468 const Type *t1 = phase->type( in(MemNode::Memory) );
2469 if( t1 == Type::TOP ) return Type::TOP;
2470 const Type *t2 = phase->type( in(MemNode::Address) );
2471 if( t2 == Type::TOP ) return Type::TOP;
2472 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2473 if( t3 == Type::TOP ) return Type::TOP;
2474 return Type::MEMORY;
2475 }
2476
2477 //------------------------------Identity---------------------------------------
2478 // Remove redundant stores:
2479 // Store(m, p, Load(m, p)) changes to m.
2480 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2481 Node* StoreNode::Identity(PhaseGVN* phase) {
2482 Node* mem = in(MemNode::Memory);
2483 Node* adr = in(MemNode::Address);
2484 Node* val = in(MemNode::ValueIn);
2485
2486 // Load then Store? Then the Store is useless
2487 if (val->is_Load() &&
2488 val->in(MemNode::Address)->eqv_uncast(adr) &&
2489 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2490 val->as_Load()->store_Opcode() == Opcode()) {
2491 return mem;
2492 }
2493
2494 // Two stores in a row of the same value?
2495 if (mem->is_Store() &&
2496 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2497 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2498 mem->Opcode() == Opcode()) {
2499 return mem;
2500 }
2501
2502 // Store of zero anywhere into a freshly-allocated object?
2503 // Then the store is useless.
2504 // (It must already have been captured by the InitializeNode.)
2505 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2506 // a newly allocated object is already all-zeroes everywhere
2507 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2508 return mem;
2509 }
2510
2511 // the store may also apply to zero-bits in an earlier object
2512 Node* prev_mem = find_previous_store(phase);
2513 // Steps (a), (b): Walk past independent stores to find an exact match.
2514 if (prev_mem != NULL) {
2515 Node* prev_val = can_see_stored_value(prev_mem, phase);
2516 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2517 // prev_val and val might differ by a cast; it would be good
2518 // to keep the more informative of the two.
2519 return mem;
2520 }
2521 }
2522 }
2523
2524 return this;
2525 }
2526
2527 //------------------------------match_edge-------------------------------------
2528 // Do we Match on this edge index or not? Match only memory & value
2529 uint StoreNode::match_edge(uint idx) const {
2530 return idx == MemNode::Address || idx == MemNode::ValueIn;
2531 }
2532
2533 //------------------------------cmp--------------------------------------------
2534 // Do not common stores up together. They generally have to be split
2535 // back up anyways, so do not bother.
2536 uint StoreNode::cmp( const Node &n ) const {
2537 return (&n == this); // Always fail except on self
2538 }
2539
2540 //------------------------------Ideal_masked_input-----------------------------
2541 // Check for a useless mask before a partial-word store
2542 // (StoreB ... (AndI valIn conIa) )
2543 // If (conIa & mask == mask) this simplifies to
2544 // (StoreB ... (valIn) )
2545 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2546 Node *val = in(MemNode::ValueIn);
2547 if( val->Opcode() == Op_AndI ) {
2548 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2549 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2550 set_req(MemNode::ValueIn, val->in(1));
2551 return this;
2552 }
2553 }
2554 return NULL;
2555 }
2556
2557
2558 //------------------------------Ideal_sign_extended_input----------------------
2559 // Check for useless sign-extension before a partial-word store
2560 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2561 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2562 // (StoreB ... (valIn) )
2563 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2564 Node *val = in(MemNode::ValueIn);
2565 if( val->Opcode() == Op_RShiftI ) {
2566 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2567 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2568 Node *shl = val->in(1);
2569 if( shl->Opcode() == Op_LShiftI ) {
2570 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2571 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2572 set_req(MemNode::ValueIn, shl->in(1));
2573 return this;
2574 }
2575 }
2576 }
2577 }
2578 return NULL;
2579 }
2580
2581 //------------------------------value_never_loaded-----------------------------------
2582 // Determine whether there are any possible loads of the value stored.
2583 // For simplicity, we actually check if there are any loads from the
2584 // address stored to, not just for loads of the value stored by this node.
2585 //
2586 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2587 Node *adr = in(Address);
2588 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2589 if (adr_oop == NULL)
2590 return false;
2591 if (!adr_oop->is_known_instance_field())
2592 return false; // if not a distinct instance, there may be aliases of the address
2593 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2594 Node *use = adr->fast_out(i);
2595 if (use->is_Load() || use->is_LoadStore()) {
2596 return false;
2597 }
2598 }
2599 return true;
2600 }
2601
2602 //=============================================================================
2603 //------------------------------Ideal------------------------------------------
2604 // If the store is from an AND mask that leaves the low bits untouched, then
2605 // we can skip the AND operation. If the store is from a sign-extension
2606 // (a left shift, then right shift) we can skip both.
2607 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2608 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2609 if( progress != NULL ) return progress;
2610
2611 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2612 if( progress != NULL ) return progress;
2613
2614 // Finally check the default case
2615 return StoreNode::Ideal(phase, can_reshape);
2616 }
2617
2618 //=============================================================================
2619 //------------------------------Ideal------------------------------------------
2620 // If the store is from an AND mask that leaves the low bits untouched, then
2621 // we can skip the AND operation
2622 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2623 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2624 if( progress != NULL ) return progress;
2625
2626 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2627 if( progress != NULL ) return progress;
2628
2629 // Finally check the default case
2630 return StoreNode::Ideal(phase, can_reshape);
2631 }
2632
2633 //=============================================================================
2634 //------------------------------Identity---------------------------------------
2635 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2636 // No need to card mark when storing a null ptr
2637 Node* my_store = in(MemNode::OopStore);
2638 if (my_store->is_Store()) {
2639 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2640 if( t1 == TypePtr::NULL_PTR ) {
2641 return in(MemNode::Memory);
2642 }
2643 }
2644 return this;
2645 }
2646
2647 //=============================================================================
2648 //------------------------------Ideal---------------------------------------
2649 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2650 Node* progress = StoreNode::Ideal(phase, can_reshape);
2651 if (progress != NULL) return progress;
2652
2653 Node* my_store = in(MemNode::OopStore);
2654 if (my_store->is_MergeMem()) {
2655 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2656 set_req(MemNode::OopStore, mem);
2657 return this;
2658 }
2659
2660 return NULL;
2661 }
2662
2663 //------------------------------Value-----------------------------------------
2664 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2665 // Either input is TOP ==> the result is TOP
2666 const Type *t = phase->type( in(MemNode::Memory) );
2667 if( t == Type::TOP ) return Type::TOP;
2668 t = phase->type( in(MemNode::Address) );
2669 if( t == Type::TOP ) return Type::TOP;
2670 t = phase->type( in(MemNode::ValueIn) );
2671 if( t == Type::TOP ) return Type::TOP;
2672 // If extra input is TOP ==> the result is TOP
2673 t = phase->type( in(MemNode::OopStore) );
2674 if( t == Type::TOP ) return Type::TOP;
2675
2676 return StoreNode::Value( phase );
2677 }
2678
2679
2680 //=============================================================================
2681 //----------------------------------SCMemProjNode------------------------------
2682 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2683 {
2684 return bottom_type();
2685 }
2686
2687 //=============================================================================
2688 //----------------------------------LoadStoreNode------------------------------
2689 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2690 : Node(required),
2691 _type(rt),
2692 _adr_type(at)
2693 {
2694 init_req(MemNode::Control, c );
2695 init_req(MemNode::Memory , mem);
2696 init_req(MemNode::Address, adr);
2697 init_req(MemNode::ValueIn, val);
2698 init_class_id(Class_LoadStore);
2699 }
2700
2701 uint LoadStoreNode::ideal_reg() const {
2702 return _type->ideal_reg();
2703 }
2704
2705 bool LoadStoreNode::result_not_used() const {
2706 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2707 Node *x = fast_out(i);
2708 if (x->Opcode() == Op_SCMemProj) continue;
2709 return false;
2710 }
2711 return true;
2712 }
2713
2714 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2715
2716 //=============================================================================
2717 //----------------------------------LoadStoreConditionalNode--------------------
2718 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2719 init_req(ExpectedIn, ex );
2720 }
2721
2722 //=============================================================================
2723 //-------------------------------adr_type--------------------------------------
2724 const TypePtr* ClearArrayNode::adr_type() const {
2725 Node *adr = in(3);
2726 if (adr == NULL) return NULL; // node is dead
2727 return MemNode::calculate_adr_type(adr->bottom_type());
2728 }
2729
2730 //------------------------------match_edge-------------------------------------
2731 // Do we Match on this edge index or not? Do not match memory
2732 uint ClearArrayNode::match_edge(uint idx) const {
2733 return idx > 1;
2734 }
2735
2736 //------------------------------Identity---------------------------------------
2737 // Clearing a zero length array does nothing
2738 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
2739 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2740 }
2741
2742 //------------------------------Idealize---------------------------------------
2743 // Clearing a short array is faster with stores
2744 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2745 const int unit = BytesPerLong;
2746 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2747 if (!t) return NULL;
2748 if (!t->is_con()) return NULL;
2749 intptr_t raw_count = t->get_con();
2750 intptr_t size = raw_count;
2751 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2752 // Clearing nothing uses the Identity call.
2753 // Negative clears are possible on dead ClearArrays
2754 // (see jck test stmt114.stmt11402.val).
2755 if (size <= 0 || size % unit != 0) return NULL;
2756 intptr_t count = size / unit;
2757 // Length too long; use fast hardware clear
2758 if (size > Matcher::init_array_short_size) return NULL;
2759 Node *mem = in(1);
2760 if( phase->type(mem)==Type::TOP ) return NULL;
2761 Node *adr = in(3);
2762 const Type* at = phase->type(adr);
2763 if( at==Type::TOP ) return NULL;
2764 const TypePtr* atp = at->isa_ptr();
2765 // adjust atp to be the correct array element address type
2766 if (atp == NULL) atp = TypePtr::BOTTOM;
2767 else atp = atp->add_offset(Type::OffsetBot);
2768 // Get base for derived pointer purposes
2769 if( adr->Opcode() != Op_AddP ) Unimplemented();
2770 Node *base = adr->in(1);
2771
2772 Node *zero = phase->makecon(TypeLong::ZERO);
2773 Node *off = phase->MakeConX(BytesPerLong);
2774 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2775 count--;
2776 while( count-- ) {
2777 mem = phase->transform(mem);
2778 adr = phase->transform(new AddPNode(base,adr,off));
2779 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2780 }
2781 return mem;
2782 }
2783
2784 //----------------------------step_through----------------------------------
2785 // Return allocation input memory edge if it is different instance
2786 // or itself if it is the one we are looking for.
2787 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2788 Node* n = *np;
2789 assert(n->is_ClearArray(), "sanity");
2790 intptr_t offset;
2791 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2792 // This method is called only before Allocate nodes are expanded
2793 // during macro nodes expansion. Before that ClearArray nodes are
2794 // only generated in PhaseMacroExpand::generate_arraycopy() (before
2795 // Allocate nodes are expanded) which follows allocations.
2796 assert(alloc != NULL, "should have allocation");
2797 if (alloc->_idx == instance_id) {
2798 // Can not bypass initialization of the instance we are looking for.
2799 return false;
2800 }
2801 // Otherwise skip it.
2802 InitializeNode* init = alloc->initialization();
2803 if (init != NULL)
2804 *np = init->in(TypeFunc::Memory);
2805 else
2806 *np = alloc->in(TypeFunc::Memory);
2807 return true;
2808 }
2809
2810 //----------------------------clear_memory-------------------------------------
2811 // Generate code to initialize object storage to zero.
2812 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2813 intptr_t start_offset,
2814 Node* end_offset,
2815 PhaseGVN* phase) {
2816 intptr_t offset = start_offset;
2817
2818 int unit = BytesPerLong;
2819 if ((offset % unit) != 0) {
2820 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
2821 adr = phase->transform(adr);
2822 const TypePtr* atp = TypeRawPtr::BOTTOM;
2823 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2824 mem = phase->transform(mem);
2825 offset += BytesPerInt;
2826 }
2827 assert((offset % unit) == 0, "");
2828
2829 // Initialize the remaining stuff, if any, with a ClearArray.
2830 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2831 }
2832
2833 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2834 Node* start_offset,
2835 Node* end_offset,
2836 PhaseGVN* phase) {
2837 if (start_offset == end_offset) {
2838 // nothing to do
2839 return mem;
2840 }
2841
2842 int unit = BytesPerLong;
2843 Node* zbase = start_offset;
2844 Node* zend = end_offset;
2845
2846 // Scale to the unit required by the CPU:
2847 if (!Matcher::init_array_count_is_in_bytes) {
2848 Node* shift = phase->intcon(exact_log2(unit));
2849 zbase = phase->transform(new URShiftXNode(zbase, shift) );
2850 zend = phase->transform(new URShiftXNode(zend, shift) );
2851 }
2852
2853 // Bulk clear double-words
2854 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
2855 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
2856 mem = new ClearArrayNode(ctl, mem, zsize, adr);
2857 return phase->transform(mem);
2858 }
2859
2860 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2861 intptr_t start_offset,
2862 intptr_t end_offset,
2863 PhaseGVN* phase) {
2864 if (start_offset == end_offset) {
2865 // nothing to do
2866 return mem;
2867 }
2868
2869 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2870 intptr_t done_offset = end_offset;
2871 if ((done_offset % BytesPerLong) != 0) {
2872 done_offset -= BytesPerInt;
2873 }
2874 if (done_offset > start_offset) {
2875 mem = clear_memory(ctl, mem, dest,
2876 start_offset, phase->MakeConX(done_offset), phase);
2877 }
2878 if (done_offset < end_offset) { // emit the final 32-bit store
2879 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
2880 adr = phase->transform(adr);
2881 const TypePtr* atp = TypeRawPtr::BOTTOM;
2882 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2883 mem = phase->transform(mem);
2884 done_offset += BytesPerInt;
2885 }
2886 assert(done_offset == end_offset, "");
2887 return mem;
2888 }
2889
2890 //=============================================================================
2891 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2892 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2893 _adr_type(C->get_adr_type(alias_idx))
2894 {
2895 init_class_id(Class_MemBar);
2896 Node* top = C->top();
2897 init_req(TypeFunc::I_O,top);
2898 init_req(TypeFunc::FramePtr,top);
2899 init_req(TypeFunc::ReturnAdr,top);
2900 if (precedent != NULL)
2901 init_req(TypeFunc::Parms, precedent);
2902 }
2903
2904 //------------------------------cmp--------------------------------------------
2905 uint MemBarNode::hash() const { return NO_HASH; }
2906 uint MemBarNode::cmp( const Node &n ) const {
2907 return (&n == this); // Always fail except on self
2908 }
2909
2910 //------------------------------make-------------------------------------------
2911 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2912 switch (opcode) {
2913 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
2914 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
2915 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
2916 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
2917 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
2918 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
2919 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
2920 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
2921 case Op_Initialize: return new InitializeNode(C, atp, pn);
2922 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
2923 default: ShouldNotReachHere(); return NULL;
2924 }
2925 }
2926
2927 //------------------------------Ideal------------------------------------------
2928 // Return a node which is more "ideal" than the current node. Strip out
2929 // control copies
2930 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2931 if (remove_dead_region(phase, can_reshape)) return this;
2932 // Don't bother trying to transform a dead node
2933 if (in(0) && in(0)->is_top()) {
2934 return NULL;
2935 }
2936
2937 bool progress = false;
2938 // Eliminate volatile MemBars for scalar replaced objects.
2939 if (can_reshape && req() == (Precedent+1)) {
2940 bool eliminate = false;
2941 int opc = Opcode();
2942 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
2943 // Volatile field loads and stores.
2944 Node* my_mem = in(MemBarNode::Precedent);
2945 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
2946 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
2947 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
2948 // replace this Precedent (decodeN) with the Load instead.
2949 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
2950 Node* load_node = my_mem->in(1);
2951 set_req(MemBarNode::Precedent, load_node);
2952 phase->is_IterGVN()->_worklist.push(my_mem);
2953 my_mem = load_node;
2954 } else {
2955 assert(my_mem->unique_out() == this, "sanity");
2956 del_req(Precedent);
2957 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
2958 my_mem = NULL;
2959 }
2960 progress = true;
2961 }
2962 if (my_mem != NULL && my_mem->is_Mem()) {
2963 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
2964 // Check for scalar replaced object reference.
2965 if( t_oop != NULL && t_oop->is_known_instance_field() &&
2966 t_oop->offset() != Type::OffsetBot &&
2967 t_oop->offset() != Type::OffsetTop) {
2968 eliminate = true;
2969 }
2970 }
2971 } else if (opc == Op_MemBarRelease) {
2972 // Final field stores.
2973 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
2974 if ((alloc != NULL) && alloc->is_Allocate() &&
2975 alloc->as_Allocate()->does_not_escape_thread()) {
2976 // The allocated object does not escape.
2977 eliminate = true;
2978 }
2979 }
2980 if (eliminate) {
2981 // Replace MemBar projections by its inputs.
2982 PhaseIterGVN* igvn = phase->is_IterGVN();
2983 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
2984 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
2985 // Must return either the original node (now dead) or a new node
2986 // (Do not return a top here, since that would break the uniqueness of top.)
2987 return new ConINode(TypeInt::ZERO);
2988 }
2989 }
2990 return progress ? this : NULL;
2991 }
2992
2993 //------------------------------Value------------------------------------------
2994 const Type* MemBarNode::Value(PhaseGVN* phase) const {
2995 if( !in(0) ) return Type::TOP;
2996 if( phase->type(in(0)) == Type::TOP )
2997 return Type::TOP;
2998 return TypeTuple::MEMBAR;
2999 }
3000
3001 //------------------------------match------------------------------------------
3002 // Construct projections for memory.
3003 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3004 switch (proj->_con) {
3005 case TypeFunc::Control:
3006 case TypeFunc::Memory:
3007 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3008 }
3009 ShouldNotReachHere();
3010 return NULL;
3011 }
3012
3013 //===========================InitializeNode====================================
3014 // SUMMARY:
3015 // This node acts as a memory barrier on raw memory, after some raw stores.
3016 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3017 // The Initialize can 'capture' suitably constrained stores as raw inits.
3018 // It can coalesce related raw stores into larger units (called 'tiles').
3019 // It can avoid zeroing new storage for memory units which have raw inits.
3020 // At macro-expansion, it is marked 'complete', and does not optimize further.
3021 //
3022 // EXAMPLE:
3023 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3024 // ctl = incoming control; mem* = incoming memory
3025 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3026 // First allocate uninitialized memory and fill in the header:
3027 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3028 // ctl := alloc.Control; mem* := alloc.Memory*
3029 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3030 // Then initialize to zero the non-header parts of the raw memory block:
3031 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3032 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3033 // After the initialize node executes, the object is ready for service:
3034 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3035 // Suppose its body is immediately initialized as {1,2}:
3036 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3037 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3038 // mem.SLICE(#short[*]) := store2
3039 //
3040 // DETAILS:
3041 // An InitializeNode collects and isolates object initialization after
3042 // an AllocateNode and before the next possible safepoint. As a
3043 // memory barrier (MemBarNode), it keeps critical stores from drifting
3044 // down past any safepoint or any publication of the allocation.
3045 // Before this barrier, a newly-allocated object may have uninitialized bits.
3046 // After this barrier, it may be treated as a real oop, and GC is allowed.
3047 //
3048 // The semantics of the InitializeNode include an implicit zeroing of
3049 // the new object from object header to the end of the object.
3050 // (The object header and end are determined by the AllocateNode.)
3051 //
3052 // Certain stores may be added as direct inputs to the InitializeNode.
3053 // These stores must update raw memory, and they must be to addresses
3054 // derived from the raw address produced by AllocateNode, and with
3055 // a constant offset. They must be ordered by increasing offset.
3056 // The first one is at in(RawStores), the last at in(req()-1).
3057 // Unlike most memory operations, they are not linked in a chain,
3058 // but are displayed in parallel as users of the rawmem output of
3059 // the allocation.
3060 //
3061 // (See comments in InitializeNode::capture_store, which continue
3062 // the example given above.)
3063 //
3064 // When the associated Allocate is macro-expanded, the InitializeNode
3065 // may be rewritten to optimize collected stores. A ClearArrayNode
3066 // may also be created at that point to represent any required zeroing.
3067 // The InitializeNode is then marked 'complete', prohibiting further
3068 // capturing of nearby memory operations.
3069 //
3070 // During macro-expansion, all captured initializations which store
3071 // constant values of 32 bits or smaller are coalesced (if advantageous)
3072 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3073 // initialized in fewer memory operations. Memory words which are
3074 // covered by neither tiles nor non-constant stores are pre-zeroed
3075 // by explicit stores of zero. (The code shape happens to do all
3076 // zeroing first, then all other stores, with both sequences occurring
3077 // in order of ascending offsets.)
3078 //
3079 // Alternatively, code may be inserted between an AllocateNode and its
3080 // InitializeNode, to perform arbitrary initialization of the new object.
3081 // E.g., the object copying intrinsics insert complex data transfers here.
3082 // The initialization must then be marked as 'complete' disable the
3083 // built-in zeroing semantics and the collection of initializing stores.
3084 //
3085 // While an InitializeNode is incomplete, reads from the memory state
3086 // produced by it are optimizable if they match the control edge and
3087 // new oop address associated with the allocation/initialization.
3088 // They return a stored value (if the offset matches) or else zero.
3089 // A write to the memory state, if it matches control and address,
3090 // and if it is to a constant offset, may be 'captured' by the
3091 // InitializeNode. It is cloned as a raw memory operation and rewired
3092 // inside the initialization, to the raw oop produced by the allocation.
3093 // Operations on addresses which are provably distinct (e.g., to
3094 // other AllocateNodes) are allowed to bypass the initialization.
3095 //
3096 // The effect of all this is to consolidate object initialization
3097 // (both arrays and non-arrays, both piecewise and bulk) into a
3098 // single location, where it can be optimized as a unit.
3099 //
3100 // Only stores with an offset less than TrackedInitializationLimit words
3101 // will be considered for capture by an InitializeNode. This puts a
3102 // reasonable limit on the complexity of optimized initializations.
3103
3104 //---------------------------InitializeNode------------------------------------
3105 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3106 : _is_complete(Incomplete), _does_not_escape(false),
3107 MemBarNode(C, adr_type, rawoop)
3108 {
3109 init_class_id(Class_Initialize);
3110
3111 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3112 assert(in(RawAddress) == rawoop, "proper init");
3113 // Note: allocation() can be NULL, for secondary initialization barriers
3114 }
3115
3116 // Since this node is not matched, it will be processed by the
3117 // register allocator. Declare that there are no constraints
3118 // on the allocation of the RawAddress edge.
3119 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3120 // This edge should be set to top, by the set_complete. But be conservative.
3121 if (idx == InitializeNode::RawAddress)
3122 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3123 return RegMask::Empty;
3124 }
3125
3126 Node* InitializeNode::memory(uint alias_idx) {
3127 Node* mem = in(Memory);
3128 if (mem->is_MergeMem()) {
3129 return mem->as_MergeMem()->memory_at(alias_idx);
3130 } else {
3131 // incoming raw memory is not split
3132 return mem;
3133 }
3134 }
3135
3136 bool InitializeNode::is_non_zero() {
3137 if (is_complete()) return false;
3138 remove_extra_zeroes();
3139 return (req() > RawStores);
3140 }
3141
3142 void InitializeNode::set_complete(PhaseGVN* phase) {
3143 assert(!is_complete(), "caller responsibility");
3144 _is_complete = Complete;
3145
3146 // After this node is complete, it contains a bunch of
3147 // raw-memory initializations. There is no need for
3148 // it to have anything to do with non-raw memory effects.
3149 // Therefore, tell all non-raw users to re-optimize themselves,
3150 // after skipping the memory effects of this initialization.
3151 PhaseIterGVN* igvn = phase->is_IterGVN();
3152 if (igvn) igvn->add_users_to_worklist(this);
3153 }
3154
3155 // convenience function
3156 // return false if the init contains any stores already
3157 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3158 InitializeNode* init = initialization();
3159 if (init == NULL || init->is_complete()) return false;
3160 init->remove_extra_zeroes();
3161 // for now, if this allocation has already collected any inits, bail:
3162 if (init->is_non_zero()) return false;
3163 init->set_complete(phase);
3164 return true;
3165 }
3166
3167 void InitializeNode::remove_extra_zeroes() {
3168 if (req() == RawStores) return;
3169 Node* zmem = zero_memory();
3170 uint fill = RawStores;
3171 for (uint i = fill; i < req(); i++) {
3172 Node* n = in(i);
3173 if (n->is_top() || n == zmem) continue; // skip
3174 if (fill < i) set_req(fill, n); // compact
3175 ++fill;
3176 }
3177 // delete any empty spaces created:
3178 while (fill < req()) {
3179 del_req(fill);
3180 }
3181 }
3182
3183 // Helper for remembering which stores go with which offsets.
3184 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3185 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3186 intptr_t offset = -1;
3187 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3188 phase, offset);
3189 if (base == NULL) return -1; // something is dead,
3190 if (offset < 0) return -1; // dead, dead
3191 return offset;
3192 }
3193
3194 // Helper for proving that an initialization expression is
3195 // "simple enough" to be folded into an object initialization.
3196 // Attempts to prove that a store's initial value 'n' can be captured
3197 // within the initialization without creating a vicious cycle, such as:
3198 // { Foo p = new Foo(); p.next = p; }
3199 // True for constants and parameters and small combinations thereof.
3200 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3201 if (n == NULL) return true; // (can this really happen?)
3202 if (n->is_Proj()) n = n->in(0);
3203 if (n == this) return false; // found a cycle
3204 if (n->is_Con()) return true;
3205 if (n->is_Start()) return true; // params, etc., are OK
3206 if (n->is_Root()) return true; // even better
3207
3208 Node* ctl = n->in(0);
3209 if (ctl != NULL && !ctl->is_top()) {
3210 if (ctl->is_Proj()) ctl = ctl->in(0);
3211 if (ctl == this) return false;
3212
3213 // If we already know that the enclosing memory op is pinned right after
3214 // the init, then any control flow that the store has picked up
3215 // must have preceded the init, or else be equal to the init.
3216 // Even after loop optimizations (which might change control edges)
3217 // a store is never pinned *before* the availability of its inputs.
3218 if (!MemNode::all_controls_dominate(n, this))
3219 return false; // failed to prove a good control
3220 }
3221
3222 // Check data edges for possible dependencies on 'this'.
3223 if ((count += 1) > 20) return false; // complexity limit
3224 for (uint i = 1; i < n->req(); i++) {
3225 Node* m = n->in(i);
3226 if (m == NULL || m == n || m->is_top()) continue;
3227 uint first_i = n->find_edge(m);
3228 if (i != first_i) continue; // process duplicate edge just once
3229 if (!detect_init_independence(m, count)) {
3230 return false;
3231 }
3232 }
3233
3234 return true;
3235 }
3236
3237 // Here are all the checks a Store must pass before it can be moved into
3238 // an initialization. Returns zero if a check fails.
3239 // On success, returns the (constant) offset to which the store applies,
3240 // within the initialized memory.
3241 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3242 const int FAIL = 0;
3243 if (st->is_unaligned_access()) {
3244 return FAIL;
3245 }
3246 if (st->req() != MemNode::ValueIn + 1)
3247 return FAIL; // an inscrutable StoreNode (card mark?)
3248 Node* ctl = st->in(MemNode::Control);
3249 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3250 return FAIL; // must be unconditional after the initialization
3251 Node* mem = st->in(MemNode::Memory);
3252 if (!(mem->is_Proj() && mem->in(0) == this))
3253 return FAIL; // must not be preceded by other stores
3254 Node* adr = st->in(MemNode::Address);
3255 intptr_t offset;
3256 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3257 if (alloc == NULL)
3258 return FAIL; // inscrutable address
3259 if (alloc != allocation())
3260 return FAIL; // wrong allocation! (store needs to float up)
3261 Node* val = st->in(MemNode::ValueIn);
3262 int complexity_count = 0;
3263 if (!detect_init_independence(val, complexity_count))
3264 return FAIL; // stored value must be 'simple enough'
3265
3266 // The Store can be captured only if nothing after the allocation
3267 // and before the Store is using the memory location that the store
3268 // overwrites.
3269 bool failed = false;
3270 // If is_complete_with_arraycopy() is true the shape of the graph is
3271 // well defined and is safe so no need for extra checks.
3272 if (!is_complete_with_arraycopy()) {
3273 // We are going to look at each use of the memory state following
3274 // the allocation to make sure nothing reads the memory that the
3275 // Store writes.
3276 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3277 int alias_idx = phase->C->get_alias_index(t_adr);
3278 ResourceMark rm;
3279 Unique_Node_List mems;
3280 mems.push(mem);
3281 Node* unique_merge = NULL;
3282 for (uint next = 0; next < mems.size(); ++next) {
3283 Node *m = mems.at(next);
3284 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3285 Node *n = m->fast_out(j);
3286 if (n->outcnt() == 0) {
3287 continue;
3288 }
3289 if (n == st) {
3290 continue;
3291 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3292 // If the control of this use is different from the control
3293 // of the Store which is right after the InitializeNode then
3294 // this node cannot be between the InitializeNode and the
3295 // Store.
3296 continue;
3297 } else if (n->is_MergeMem()) {
3298 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3299 // We can hit a MergeMemNode (that will likely go away
3300 // later) that is a direct use of the memory state
3301 // following the InitializeNode on the same slice as the
3302 // store node that we'd like to capture. We need to check
3303 // the uses of the MergeMemNode.
3304 mems.push(n);
3305 }
3306 } else if (n->is_Mem()) {
3307 Node* other_adr = n->in(MemNode::Address);
3308 if (other_adr == adr) {
3309 failed = true;
3310 break;
3311 } else {
3312 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3313 if (other_t_adr != NULL) {
3314 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3315 if (other_alias_idx == alias_idx) {
3316 // A load from the same memory slice as the store right
3317 // after the InitializeNode. We check the control of the
3318 // object/array that is loaded from. If it's the same as
3319 // the store control then we cannot capture the store.
3320 assert(!n->is_Store(), "2 stores to same slice on same control?");
3321 Node* base = other_adr;
3322 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3323 base = base->in(AddPNode::Base);
3324 if (base != NULL) {
3325 base = base->uncast();
3326 if (base->is_Proj() && base->in(0) == alloc) {
3327 failed = true;
3328 break;
3329 }
3330 }
3331 }
3332 }
3333 }
3334 } else {
3335 failed = true;
3336 break;
3337 }
3338 }
3339 }
3340 }
3341 if (failed) {
3342 if (!can_reshape) {
3343 // We decided we couldn't capture the store during parsing. We
3344 // should try again during the next IGVN once the graph is
3345 // cleaner.
3346 phase->C->record_for_igvn(st);
3347 }
3348 return FAIL;
3349 }
3350
3351 return offset; // success
3352 }
3353
3354 // Find the captured store in(i) which corresponds to the range
3355 // [start..start+size) in the initialized object.
3356 // If there is one, return its index i. If there isn't, return the
3357 // negative of the index where it should be inserted.
3358 // Return 0 if the queried range overlaps an initialization boundary
3359 // or if dead code is encountered.
3360 // If size_in_bytes is zero, do not bother with overlap checks.
3361 int InitializeNode::captured_store_insertion_point(intptr_t start,
3362 int size_in_bytes,
3363 PhaseTransform* phase) {
3364 const int FAIL = 0, MAX_STORE = BytesPerLong;
3365
3366 if (is_complete())
3367 return FAIL; // arraycopy got here first; punt
3368
3369 assert(allocation() != NULL, "must be present");
3370
3371 // no negatives, no header fields:
3372 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3373
3374 // after a certain size, we bail out on tracking all the stores:
3375 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3376 if (start >= ti_limit) return FAIL;
3377
3378 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3379 if (i >= limit) return -(int)i; // not found; here is where to put it
3380
3381 Node* st = in(i);
3382 intptr_t st_off = get_store_offset(st, phase);
3383 if (st_off < 0) {
3384 if (st != zero_memory()) {
3385 return FAIL; // bail out if there is dead garbage
3386 }
3387 } else if (st_off > start) {
3388 // ...we are done, since stores are ordered
3389 if (st_off < start + size_in_bytes) {
3390 return FAIL; // the next store overlaps
3391 }
3392 return -(int)i; // not found; here is where to put it
3393 } else if (st_off < start) {
3394 if (size_in_bytes != 0 &&
3395 start < st_off + MAX_STORE &&
3396 start < st_off + st->as_Store()->memory_size()) {
3397 return FAIL; // the previous store overlaps
3398 }
3399 } else {
3400 if (size_in_bytes != 0 &&
3401 st->as_Store()->memory_size() != size_in_bytes) {
3402 return FAIL; // mismatched store size
3403 }
3404 return i;
3405 }
3406
3407 ++i;
3408 }
3409 }
3410
3411 // Look for a captured store which initializes at the offset 'start'
3412 // with the given size. If there is no such store, and no other
3413 // initialization interferes, then return zero_memory (the memory
3414 // projection of the AllocateNode).
3415 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3416 PhaseTransform* phase) {
3417 assert(stores_are_sane(phase), "");
3418 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3419 if (i == 0) {
3420 return NULL; // something is dead
3421 } else if (i < 0) {
3422 return zero_memory(); // just primordial zero bits here
3423 } else {
3424 Node* st = in(i); // here is the store at this position
3425 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3426 return st;
3427 }
3428 }
3429
3430 // Create, as a raw pointer, an address within my new object at 'offset'.
3431 Node* InitializeNode::make_raw_address(intptr_t offset,
3432 PhaseTransform* phase) {
3433 Node* addr = in(RawAddress);
3434 if (offset != 0) {
3435 Compile* C = phase->C;
3436 addr = phase->transform( new AddPNode(C->top(), addr,
3437 phase->MakeConX(offset)) );
3438 }
3439 return addr;
3440 }
3441
3442 // Clone the given store, converting it into a raw store
3443 // initializing a field or element of my new object.
3444 // Caller is responsible for retiring the original store,
3445 // with subsume_node or the like.
3446 //
3447 // From the example above InitializeNode::InitializeNode,
3448 // here are the old stores to be captured:
3449 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3450 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3451 //
3452 // Here is the changed code; note the extra edges on init:
3453 // alloc = (Allocate ...)
3454 // rawoop = alloc.RawAddress
3455 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3456 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3457 // init = (Initialize alloc.Control alloc.Memory rawoop
3458 // rawstore1 rawstore2)
3459 //
3460 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3461 PhaseTransform* phase, bool can_reshape) {
3462 assert(stores_are_sane(phase), "");
3463
3464 if (start < 0) return NULL;
3465 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3466
3467 Compile* C = phase->C;
3468 int size_in_bytes = st->memory_size();
3469 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3470 if (i == 0) return NULL; // bail out
3471 Node* prev_mem = NULL; // raw memory for the captured store
3472 if (i > 0) {
3473 prev_mem = in(i); // there is a pre-existing store under this one
3474 set_req(i, C->top()); // temporarily disconnect it
3475 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3476 } else {
3477 i = -i; // no pre-existing store
3478 prev_mem = zero_memory(); // a slice of the newly allocated object
3479 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3480 set_req(--i, C->top()); // reuse this edge; it has been folded away
3481 else
3482 ins_req(i, C->top()); // build a new edge
3483 }
3484 Node* new_st = st->clone();
3485 new_st->set_req(MemNode::Control, in(Control));
3486 new_st->set_req(MemNode::Memory, prev_mem);
3487 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3488 new_st = phase->transform(new_st);
3489
3490 // At this point, new_st might have swallowed a pre-existing store
3491 // at the same offset, or perhaps new_st might have disappeared,
3492 // if it redundantly stored the same value (or zero to fresh memory).
3493
3494 // In any case, wire it in:
3495 phase->igvn_rehash_node_delayed(this);
3496 set_req(i, new_st);
3497
3498 // The caller may now kill the old guy.
3499 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3500 assert(check_st == new_st || check_st == NULL, "must be findable");
3501 assert(!is_complete(), "");
3502 return new_st;
3503 }
3504
3505 static bool store_constant(jlong* tiles, int num_tiles,
3506 intptr_t st_off, int st_size,
3507 jlong con) {
3508 if ((st_off & (st_size-1)) != 0)
3509 return false; // strange store offset (assume size==2**N)
3510 address addr = (address)tiles + st_off;
3511 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3512 switch (st_size) {
3513 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3514 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3515 case sizeof(jint): *(jint*) addr = (jint) con; break;
3516 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3517 default: return false; // strange store size (detect size!=2**N here)
3518 }
3519 return true; // return success to caller
3520 }
3521
3522 // Coalesce subword constants into int constants and possibly
3523 // into long constants. The goal, if the CPU permits,
3524 // is to initialize the object with a small number of 64-bit tiles.
3525 // Also, convert floating-point constants to bit patterns.
3526 // Non-constants are not relevant to this pass.
3527 //
3528 // In terms of the running example on InitializeNode::InitializeNode
3529 // and InitializeNode::capture_store, here is the transformation
3530 // of rawstore1 and rawstore2 into rawstore12:
3531 // alloc = (Allocate ...)
3532 // rawoop = alloc.RawAddress
3533 // tile12 = 0x00010002
3534 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3535 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3536 //
3537 void
3538 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3539 Node* size_in_bytes,
3540 PhaseGVN* phase) {
3541 Compile* C = phase->C;
3542
3543 assert(stores_are_sane(phase), "");
3544 // Note: After this pass, they are not completely sane,
3545 // since there may be some overlaps.
3546
3547 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3548
3549 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3550 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3551 size_limit = MIN2(size_limit, ti_limit);
3552 size_limit = align_size_up(size_limit, BytesPerLong);
3553 int num_tiles = size_limit / BytesPerLong;
3554
3555 // allocate space for the tile map:
3556 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3557 jlong tiles_buf[small_len];
3558 Node* nodes_buf[small_len];
3559 jlong inits_buf[small_len];
3560 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3561 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3562 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3563 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3564 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3565 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3566 // tiles: exact bitwise model of all primitive constants
3567 // nodes: last constant-storing node subsumed into the tiles model
3568 // inits: which bytes (in each tile) are touched by any initializations
3569
3570 //// Pass A: Fill in the tile model with any relevant stores.
3571
3572 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3573 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3574 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3575 Node* zmem = zero_memory(); // initially zero memory state
3576 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3577 Node* st = in(i);
3578 intptr_t st_off = get_store_offset(st, phase);
3579
3580 // Figure out the store's offset and constant value:
3581 if (st_off < header_size) continue; //skip (ignore header)
3582 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3583 int st_size = st->as_Store()->memory_size();
3584 if (st_off + st_size > size_limit) break;
3585
3586 // Record which bytes are touched, whether by constant or not.
3587 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3588 continue; // skip (strange store size)
3589
3590 const Type* val = phase->type(st->in(MemNode::ValueIn));
3591 if (!val->singleton()) continue; //skip (non-con store)
3592 BasicType type = val->basic_type();
3593
3594 jlong con = 0;
3595 switch (type) {
3596 case T_INT: con = val->is_int()->get_con(); break;
3597 case T_LONG: con = val->is_long()->get_con(); break;
3598 case T_FLOAT: con = jint_cast(val->getf()); break;
3599 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3600 default: continue; //skip (odd store type)
3601 }
3602
3603 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3604 st->Opcode() == Op_StoreL) {
3605 continue; // This StoreL is already optimal.
3606 }
3607
3608 // Store down the constant.
3609 store_constant(tiles, num_tiles, st_off, st_size, con);
3610
3611 intptr_t j = st_off >> LogBytesPerLong;
3612
3613 if (type == T_INT && st_size == BytesPerInt
3614 && (st_off & BytesPerInt) == BytesPerInt) {
3615 jlong lcon = tiles[j];
3616 if (!Matcher::isSimpleConstant64(lcon) &&
3617 st->Opcode() == Op_StoreI) {
3618 // This StoreI is already optimal by itself.
3619 jint* intcon = (jint*) &tiles[j];
3620 intcon[1] = 0; // undo the store_constant()
3621
3622 // If the previous store is also optimal by itself, back up and
3623 // undo the action of the previous loop iteration... if we can.
3624 // But if we can't, just let the previous half take care of itself.
3625 st = nodes[j];
3626 st_off -= BytesPerInt;
3627 con = intcon[0];
3628 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3629 assert(st_off >= header_size, "still ignoring header");
3630 assert(get_store_offset(st, phase) == st_off, "must be");
3631 assert(in(i-1) == zmem, "must be");
3632 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3633 assert(con == tcon->is_int()->get_con(), "must be");
3634 // Undo the effects of the previous loop trip, which swallowed st:
3635 intcon[0] = 0; // undo store_constant()
3636 set_req(i-1, st); // undo set_req(i, zmem)
3637 nodes[j] = NULL; // undo nodes[j] = st
3638 --old_subword; // undo ++old_subword
3639 }
3640 continue; // This StoreI is already optimal.
3641 }
3642 }
3643
3644 // This store is not needed.
3645 set_req(i, zmem);
3646 nodes[j] = st; // record for the moment
3647 if (st_size < BytesPerLong) // something has changed
3648 ++old_subword; // includes int/float, but who's counting...
3649 else ++old_long;
3650 }
3651
3652 if ((old_subword + old_long) == 0)
3653 return; // nothing more to do
3654
3655 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3656 // Be sure to insert them before overlapping non-constant stores.
3657 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3658 for (int j = 0; j < num_tiles; j++) {
3659 jlong con = tiles[j];
3660 jlong init = inits[j];
3661 if (con == 0) continue;
3662 jint con0, con1; // split the constant, address-wise
3663 jint init0, init1; // split the init map, address-wise
3664 { union { jlong con; jint intcon[2]; } u;
3665 u.con = con;
3666 con0 = u.intcon[0];
3667 con1 = u.intcon[1];
3668 u.con = init;
3669 init0 = u.intcon[0];
3670 init1 = u.intcon[1];
3671 }
3672
3673 Node* old = nodes[j];
3674 assert(old != NULL, "need the prior store");
3675 intptr_t offset = (j * BytesPerLong);
3676
3677 bool split = !Matcher::isSimpleConstant64(con);
3678
3679 if (offset < header_size) {
3680 assert(offset + BytesPerInt >= header_size, "second int counts");
3681 assert(*(jint*)&tiles[j] == 0, "junk in header");
3682 split = true; // only the second word counts
3683 // Example: int a[] = { 42 ... }
3684 } else if (con0 == 0 && init0 == -1) {
3685 split = true; // first word is covered by full inits
3686 // Example: int a[] = { ... foo(), 42 ... }
3687 } else if (con1 == 0 && init1 == -1) {
3688 split = true; // second word is covered by full inits
3689 // Example: int a[] = { ... 42, foo() ... }
3690 }
3691
3692 // Here's a case where init0 is neither 0 nor -1:
3693 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3694 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3695 // In this case the tile is not split; it is (jlong)42.
3696 // The big tile is stored down, and then the foo() value is inserted.
3697 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3698
3699 Node* ctl = old->in(MemNode::Control);
3700 Node* adr = make_raw_address(offset, phase);
3701 const TypePtr* atp = TypeRawPtr::BOTTOM;
3702
3703 // One or two coalesced stores to plop down.
3704 Node* st[2];
3705 intptr_t off[2];
3706 int nst = 0;
3707 if (!split) {
3708 ++new_long;
3709 off[nst] = offset;
3710 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3711 phase->longcon(con), T_LONG, MemNode::unordered);
3712 } else {
3713 // Omit either if it is a zero.
3714 if (con0 != 0) {
3715 ++new_int;
3716 off[nst] = offset;
3717 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3718 phase->intcon(con0), T_INT, MemNode::unordered);
3719 }
3720 if (con1 != 0) {
3721 ++new_int;
3722 offset += BytesPerInt;
3723 adr = make_raw_address(offset, phase);
3724 off[nst] = offset;
3725 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3726 phase->intcon(con1), T_INT, MemNode::unordered);
3727 }
3728 }
3729
3730 // Insert second store first, then the first before the second.
3731 // Insert each one just before any overlapping non-constant stores.
3732 while (nst > 0) {
3733 Node* st1 = st[--nst];
3734 C->copy_node_notes_to(st1, old);
3735 st1 = phase->transform(st1);
3736 offset = off[nst];
3737 assert(offset >= header_size, "do not smash header");
3738 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3739 guarantee(ins_idx != 0, "must re-insert constant store");
3740 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3741 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3742 set_req(--ins_idx, st1);
3743 else
3744 ins_req(ins_idx, st1);
3745 }
3746 }
3747
3748 if (PrintCompilation && WizardMode)
3749 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3750 old_subword, old_long, new_int, new_long);
3751 if (C->log() != NULL)
3752 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3753 old_subword, old_long, new_int, new_long);
3754
3755 // Clean up any remaining occurrences of zmem:
3756 remove_extra_zeroes();
3757 }
3758
3759 // Explore forward from in(start) to find the first fully initialized
3760 // word, and return its offset. Skip groups of subword stores which
3761 // together initialize full words. If in(start) is itself part of a
3762 // fully initialized word, return the offset of in(start). If there
3763 // are no following full-word stores, or if something is fishy, return
3764 // a negative value.
3765 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3766 int int_map = 0;
3767 intptr_t int_map_off = 0;
3768 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3769
3770 for (uint i = start, limit = req(); i < limit; i++) {
3771 Node* st = in(i);
3772
3773 intptr_t st_off = get_store_offset(st, phase);
3774 if (st_off < 0) break; // return conservative answer
3775
3776 int st_size = st->as_Store()->memory_size();
3777 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3778 return st_off; // we found a complete word init
3779 }
3780
3781 // update the map:
3782
3783 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3784 if (this_int_off != int_map_off) {
3785 // reset the map:
3786 int_map = 0;
3787 int_map_off = this_int_off;
3788 }
3789
3790 int subword_off = st_off - this_int_off;
3791 int_map |= right_n_bits(st_size) << subword_off;
3792 if ((int_map & FULL_MAP) == FULL_MAP) {
3793 return this_int_off; // we found a complete word init
3794 }
3795
3796 // Did this store hit or cross the word boundary?
3797 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3798 if (next_int_off == this_int_off + BytesPerInt) {
3799 // We passed the current int, without fully initializing it.
3800 int_map_off = next_int_off;
3801 int_map >>= BytesPerInt;
3802 } else if (next_int_off > this_int_off + BytesPerInt) {
3803 // We passed the current and next int.
3804 return this_int_off + BytesPerInt;
3805 }
3806 }
3807
3808 return -1;
3809 }
3810
3811
3812 // Called when the associated AllocateNode is expanded into CFG.
3813 // At this point, we may perform additional optimizations.
3814 // Linearize the stores by ascending offset, to make memory
3815 // activity as coherent as possible.
3816 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3817 intptr_t header_size,
3818 Node* size_in_bytes,
3819 PhaseGVN* phase) {
3820 assert(!is_complete(), "not already complete");
3821 assert(stores_are_sane(phase), "");
3822 assert(allocation() != NULL, "must be present");
3823
3824 remove_extra_zeroes();
3825
3826 if (ReduceFieldZeroing || ReduceBulkZeroing)
3827 // reduce instruction count for common initialization patterns
3828 coalesce_subword_stores(header_size, size_in_bytes, phase);
3829
3830 Node* zmem = zero_memory(); // initially zero memory state
3831 Node* inits = zmem; // accumulating a linearized chain of inits
3832 #ifdef ASSERT
3833 intptr_t first_offset = allocation()->minimum_header_size();
3834 intptr_t last_init_off = first_offset; // previous init offset
3835 intptr_t last_init_end = first_offset; // previous init offset+size
3836 intptr_t last_tile_end = first_offset; // previous tile offset+size
3837 #endif
3838 intptr_t zeroes_done = header_size;
3839
3840 bool do_zeroing = true; // we might give up if inits are very sparse
3841 int big_init_gaps = 0; // how many large gaps have we seen?
3842
3843 if (UseTLAB && ZeroTLAB) do_zeroing = false;
3844 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3845
3846 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3847 Node* st = in(i);
3848 intptr_t st_off = get_store_offset(st, phase);
3849 if (st_off < 0)
3850 break; // unknown junk in the inits
3851 if (st->in(MemNode::Memory) != zmem)
3852 break; // complicated store chains somehow in list
3853
3854 int st_size = st->as_Store()->memory_size();
3855 intptr_t next_init_off = st_off + st_size;
3856
3857 if (do_zeroing && zeroes_done < next_init_off) {
3858 // See if this store needs a zero before it or under it.
3859 intptr_t zeroes_needed = st_off;
3860
3861 if (st_size < BytesPerInt) {
3862 // Look for subword stores which only partially initialize words.
3863 // If we find some, we must lay down some word-level zeroes first,
3864 // underneath the subword stores.
3865 //
3866 // Examples:
3867 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3868 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3869 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3870 //
3871 // Note: coalesce_subword_stores may have already done this,
3872 // if it was prompted by constant non-zero subword initializers.
3873 // But this case can still arise with non-constant stores.
3874
3875 intptr_t next_full_store = find_next_fullword_store(i, phase);
3876
3877 // In the examples above:
3878 // in(i) p q r s x y z
3879 // st_off 12 13 14 15 12 13 14
3880 // st_size 1 1 1 1 1 1 1
3881 // next_full_s. 12 16 16 16 16 16 16
3882 // z's_done 12 16 16 16 12 16 12
3883 // z's_needed 12 16 16 16 16 16 16
3884 // zsize 0 0 0 0 4 0 4
3885 if (next_full_store < 0) {
3886 // Conservative tack: Zero to end of current word.
3887 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3888 } else {
3889 // Zero to beginning of next fully initialized word.
3890 // Or, don't zero at all, if we are already in that word.
3891 assert(next_full_store >= zeroes_needed, "must go forward");
3892 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3893 zeroes_needed = next_full_store;
3894 }
3895 }
3896
3897 if (zeroes_needed > zeroes_done) {
3898 intptr_t zsize = zeroes_needed - zeroes_done;
3899 // Do some incremental zeroing on rawmem, in parallel with inits.
3900 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3901 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3902 zeroes_done, zeroes_needed,
3903 phase);
3904 zeroes_done = zeroes_needed;
3905 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3906 do_zeroing = false; // leave the hole, next time
3907 }
3908 }
3909
3910 // Collect the store and move on:
3911 st->set_req(MemNode::Memory, inits);
3912 inits = st; // put it on the linearized chain
3913 set_req(i, zmem); // unhook from previous position
3914
3915 if (zeroes_done == st_off)
3916 zeroes_done = next_init_off;
3917
3918 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3919
3920 #ifdef ASSERT
3921 // Various order invariants. Weaker than stores_are_sane because
3922 // a large constant tile can be filled in by smaller non-constant stores.
3923 assert(st_off >= last_init_off, "inits do not reverse");
3924 last_init_off = st_off;
3925 const Type* val = NULL;
3926 if (st_size >= BytesPerInt &&
3927 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3928 (int)val->basic_type() < (int)T_OBJECT) {
3929 assert(st_off >= last_tile_end, "tiles do not overlap");
3930 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3931 last_tile_end = MAX2(last_tile_end, next_init_off);
3932 } else {
3933 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3934 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3935 assert(st_off >= last_init_end, "inits do not overlap");
3936 last_init_end = next_init_off; // it's a non-tile
3937 }
3938 #endif //ASSERT
3939 }
3940
3941 remove_extra_zeroes(); // clear out all the zmems left over
3942 add_req(inits);
3943
3944 if (!(UseTLAB && ZeroTLAB)) {
3945 // If anything remains to be zeroed, zero it all now.
3946 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3947 // if it is the last unused 4 bytes of an instance, forget about it
3948 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3949 if (zeroes_done + BytesPerLong >= size_limit) {
3950 assert(allocation() != NULL, "");
3951 if (allocation()->Opcode() == Op_Allocate) {
3952 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3953 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3954 if (zeroes_done == k->layout_helper())
3955 zeroes_done = size_limit;
3956 }
3957 }
3958 if (zeroes_done < size_limit) {
3959 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3960 zeroes_done, size_in_bytes, phase);
3961 }
3962 }
3963
3964 set_complete(phase);
3965 return rawmem;
3966 }
3967
3968
3969 #ifdef ASSERT
3970 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3971 if (is_complete())
3972 return true; // stores could be anything at this point
3973 assert(allocation() != NULL, "must be present");
3974 intptr_t last_off = allocation()->minimum_header_size();
3975 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3976 Node* st = in(i);
3977 intptr_t st_off = get_store_offset(st, phase);
3978 if (st_off < 0) continue; // ignore dead garbage
3979 if (last_off > st_off) {
3980 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
3981 this->dump(2);
3982 assert(false, "ascending store offsets");
3983 return false;
3984 }
3985 last_off = st_off + st->as_Store()->memory_size();
3986 }
3987 return true;
3988 }
3989 #endif //ASSERT
3990
3991
3992
3993
3994 //============================MergeMemNode=====================================
3995 //
3996 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3997 // contributing store or call operations. Each contributor provides the memory
3998 // state for a particular "alias type" (see Compile::alias_type). For example,
3999 // if a MergeMem has an input X for alias category #6, then any memory reference
4000 // to alias category #6 may use X as its memory state input, as an exact equivalent
4001 // to using the MergeMem as a whole.
4002 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4003 //
4004 // (Here, the <N> notation gives the index of the relevant adr_type.)
4005 //
4006 // In one special case (and more cases in the future), alias categories overlap.
4007 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4008 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4009 // it is exactly equivalent to that state W:
4010 // MergeMem(<Bot>: W) <==> W
4011 //
4012 // Usually, the merge has more than one input. In that case, where inputs
4013 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4014 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4015 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4016 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4017 //
4018 // A merge can take a "wide" memory state as one of its narrow inputs.
4019 // This simply means that the merge observes out only the relevant parts of
4020 // the wide input. That is, wide memory states arriving at narrow merge inputs
4021 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4022 //
4023 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4024 // and that memory slices "leak through":
4025 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4026 //
4027 // But, in such a cascade, repeated memory slices can "block the leak":
4028 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4029 //
4030 // In the last example, Y is not part of the combined memory state of the
4031 // outermost MergeMem. The system must, of course, prevent unschedulable
4032 // memory states from arising, so you can be sure that the state Y is somehow
4033 // a precursor to state Y'.
4034 //
4035 //
4036 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4037 // of each MergeMemNode array are exactly the numerical alias indexes, including
4038 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4039 // Compile::alias_type (and kin) produce and manage these indexes.
4040 //
4041 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4042 // (Note that this provides quick access to the top node inside MergeMem methods,
4043 // without the need to reach out via TLS to Compile::current.)
4044 //
4045 // As a consequence of what was just described, a MergeMem that represents a full
4046 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4047 // containing all alias categories.
4048 //
4049 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4050 //
4051 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4052 // a memory state for the alias type <N>, or else the top node, meaning that
4053 // there is no particular input for that alias type. Note that the length of
4054 // a MergeMem is variable, and may be extended at any time to accommodate new
4055 // memory states at larger alias indexes. When merges grow, they are of course
4056 // filled with "top" in the unused in() positions.
4057 //
4058 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4059 // (Top was chosen because it works smoothly with passes like GCM.)
4060 //
4061 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4062 // the type of random VM bits like TLS references.) Since it is always the
4063 // first non-Bot memory slice, some low-level loops use it to initialize an
4064 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4065 //
4066 //
4067 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4068 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4069 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4070 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4071 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4072 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4073 //
4074 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4075 // really that different from the other memory inputs. An abbreviation called
4076 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4077 //
4078 //
4079 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4080 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4081 // that "emerges though" the base memory will be marked as excluding the alias types
4082 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4083 //
4084 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4085 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4086 //
4087 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4088 // (It is currently unimplemented.) As you can see, the resulting merge is
4089 // actually a disjoint union of memory states, rather than an overlay.
4090 //
4091
4092 //------------------------------MergeMemNode-----------------------------------
4093 Node* MergeMemNode::make_empty_memory() {
4094 Node* empty_memory = (Node*) Compile::current()->top();
4095 assert(empty_memory->is_top(), "correct sentinel identity");
4096 return empty_memory;
4097 }
4098
4099 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4100 init_class_id(Class_MergeMem);
4101 // all inputs are nullified in Node::Node(int)
4102 // set_input(0, NULL); // no control input
4103
4104 // Initialize the edges uniformly to top, for starters.
4105 Node* empty_mem = make_empty_memory();
4106 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4107 init_req(i,empty_mem);
4108 }
4109 assert(empty_memory() == empty_mem, "");
4110
4111 if( new_base != NULL && new_base->is_MergeMem() ) {
4112 MergeMemNode* mdef = new_base->as_MergeMem();
4113 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4114 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4115 mms.set_memory(mms.memory2());
4116 }
4117 assert(base_memory() == mdef->base_memory(), "");
4118 } else {
4119 set_base_memory(new_base);
4120 }
4121 }
4122
4123 // Make a new, untransformed MergeMem with the same base as 'mem'.
4124 // If mem is itself a MergeMem, populate the result with the same edges.
4125 MergeMemNode* MergeMemNode::make(Node* mem) {
4126 return new MergeMemNode(mem);
4127 }
4128
4129 //------------------------------cmp--------------------------------------------
4130 uint MergeMemNode::hash() const { return NO_HASH; }
4131 uint MergeMemNode::cmp( const Node &n ) const {
4132 return (&n == this); // Always fail except on self
4133 }
4134
4135 //------------------------------Identity---------------------------------------
4136 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4137 // Identity if this merge point does not record any interesting memory
4138 // disambiguations.
4139 Node* base_mem = base_memory();
4140 Node* empty_mem = empty_memory();
4141 if (base_mem != empty_mem) { // Memory path is not dead?
4142 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4143 Node* mem = in(i);
4144 if (mem != empty_mem && mem != base_mem) {
4145 return this; // Many memory splits; no change
4146 }
4147 }
4148 }
4149 return base_mem; // No memory splits; ID on the one true input
4150 }
4151
4152 //------------------------------Ideal------------------------------------------
4153 // This method is invoked recursively on chains of MergeMem nodes
4154 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4155 // Remove chain'd MergeMems
4156 //
4157 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4158 // relative to the "in(Bot)". Since we are patching both at the same time,
4159 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4160 // but rewrite each "in(i)" relative to the new "in(Bot)".
4161 Node *progress = NULL;
4162
4163
4164 Node* old_base = base_memory();
4165 Node* empty_mem = empty_memory();
4166 if (old_base == empty_mem)
4167 return NULL; // Dead memory path.
4168
4169 MergeMemNode* old_mbase;
4170 if (old_base != NULL && old_base->is_MergeMem())
4171 old_mbase = old_base->as_MergeMem();
4172 else
4173 old_mbase = NULL;
4174 Node* new_base = old_base;
4175
4176 // simplify stacked MergeMems in base memory
4177 if (old_mbase) new_base = old_mbase->base_memory();
4178
4179 // the base memory might contribute new slices beyond my req()
4180 if (old_mbase) grow_to_match(old_mbase);
4181
4182 // Look carefully at the base node if it is a phi.
4183 PhiNode* phi_base;
4184 if (new_base != NULL && new_base->is_Phi())
4185 phi_base = new_base->as_Phi();
4186 else
4187 phi_base = NULL;
4188
4189 Node* phi_reg = NULL;
4190 uint phi_len = (uint)-1;
4191 if (phi_base != NULL && !phi_base->is_copy()) {
4192 // do not examine phi if degraded to a copy
4193 phi_reg = phi_base->region();
4194 phi_len = phi_base->req();
4195 // see if the phi is unfinished
4196 for (uint i = 1; i < phi_len; i++) {
4197 if (phi_base->in(i) == NULL) {
4198 // incomplete phi; do not look at it yet!
4199 phi_reg = NULL;
4200 phi_len = (uint)-1;
4201 break;
4202 }
4203 }
4204 }
4205
4206 // Note: We do not call verify_sparse on entry, because inputs
4207 // can normalize to the base_memory via subsume_node or similar
4208 // mechanisms. This method repairs that damage.
4209
4210 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4211
4212 // Look at each slice.
4213 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4214 Node* old_in = in(i);
4215 // calculate the old memory value
4216 Node* old_mem = old_in;
4217 if (old_mem == empty_mem) old_mem = old_base;
4218 assert(old_mem == memory_at(i), "");
4219
4220 // maybe update (reslice) the old memory value
4221
4222 // simplify stacked MergeMems
4223 Node* new_mem = old_mem;
4224 MergeMemNode* old_mmem;
4225 if (old_mem != NULL && old_mem->is_MergeMem())
4226 old_mmem = old_mem->as_MergeMem();
4227 else
4228 old_mmem = NULL;
4229 if (old_mmem == this) {
4230 // This can happen if loops break up and safepoints disappear.
4231 // A merge of BotPtr (default) with a RawPtr memory derived from a
4232 // safepoint can be rewritten to a merge of the same BotPtr with
4233 // the BotPtr phi coming into the loop. If that phi disappears
4234 // also, we can end up with a self-loop of the mergemem.
4235 // In general, if loops degenerate and memory effects disappear,
4236 // a mergemem can be left looking at itself. This simply means
4237 // that the mergemem's default should be used, since there is
4238 // no longer any apparent effect on this slice.
4239 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4240 // from start. Update the input to TOP.
4241 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4242 }
4243 else if (old_mmem != NULL) {
4244 new_mem = old_mmem->memory_at(i);
4245 }
4246 // else preceding memory was not a MergeMem
4247
4248 // replace equivalent phis (unfortunately, they do not GVN together)
4249 if (new_mem != NULL && new_mem != new_base &&
4250 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4251 if (new_mem->is_Phi()) {
4252 PhiNode* phi_mem = new_mem->as_Phi();
4253 for (uint i = 1; i < phi_len; i++) {
4254 if (phi_base->in(i) != phi_mem->in(i)) {
4255 phi_mem = NULL;
4256 break;
4257 }
4258 }
4259 if (phi_mem != NULL) {
4260 // equivalent phi nodes; revert to the def
4261 new_mem = new_base;
4262 }
4263 }
4264 }
4265
4266 // maybe store down a new value
4267 Node* new_in = new_mem;
4268 if (new_in == new_base) new_in = empty_mem;
4269
4270 if (new_in != old_in) {
4271 // Warning: Do not combine this "if" with the previous "if"
4272 // A memory slice might have be be rewritten even if it is semantically
4273 // unchanged, if the base_memory value has changed.
4274 set_req(i, new_in);
4275 progress = this; // Report progress
4276 }
4277 }
4278
4279 if (new_base != old_base) {
4280 set_req(Compile::AliasIdxBot, new_base);
4281 // Don't use set_base_memory(new_base), because we need to update du.
4282 assert(base_memory() == new_base, "");
4283 progress = this;
4284 }
4285
4286 if( base_memory() == this ) {
4287 // a self cycle indicates this memory path is dead
4288 set_req(Compile::AliasIdxBot, empty_mem);
4289 }
4290
4291 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4292 // Recursion must occur after the self cycle check above
4293 if( base_memory()->is_MergeMem() ) {
4294 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4295 Node *m = phase->transform(new_mbase); // Rollup any cycles
4296 if( m != NULL && (m->is_top() ||
4297 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
4298 // propagate rollup of dead cycle to self
4299 set_req(Compile::AliasIdxBot, empty_mem);
4300 }
4301 }
4302
4303 if( base_memory() == empty_mem ) {
4304 progress = this;
4305 // Cut inputs during Parse phase only.
4306 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4307 if( !can_reshape ) {
4308 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4309 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4310 }
4311 }
4312 }
4313
4314 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4315 // Check if PhiNode::Ideal's "Split phis through memory merges"
4316 // transform should be attempted. Look for this->phi->this cycle.
4317 uint merge_width = req();
4318 if (merge_width > Compile::AliasIdxRaw) {
4319 PhiNode* phi = base_memory()->as_Phi();
4320 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4321 if (phi->in(i) == this) {
4322 phase->is_IterGVN()->_worklist.push(phi);
4323 break;
4324 }
4325 }
4326 }
4327 }
4328
4329 assert(progress || verify_sparse(), "please, no dups of base");
4330 return progress;
4331 }
4332
4333 //-------------------------set_base_memory-------------------------------------
4334 void MergeMemNode::set_base_memory(Node *new_base) {
4335 Node* empty_mem = empty_memory();
4336 set_req(Compile::AliasIdxBot, new_base);
4337 assert(memory_at(req()) == new_base, "must set default memory");
4338 // Clear out other occurrences of new_base:
4339 if (new_base != empty_mem) {
4340 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4341 if (in(i) == new_base) set_req(i, empty_mem);
4342 }
4343 }
4344 }
4345
4346 //------------------------------out_RegMask------------------------------------
4347 const RegMask &MergeMemNode::out_RegMask() const {
4348 return RegMask::Empty;
4349 }
4350
4351 //------------------------------dump_spec--------------------------------------
4352 #ifndef PRODUCT
4353 void MergeMemNode::dump_spec(outputStream *st) const {
4354 st->print(" {");
4355 Node* base_mem = base_memory();
4356 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4357 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4358 if (mem == base_mem) { st->print(" -"); continue; }
4359 st->print( " N%d:", mem->_idx );
4360 Compile::current()->get_adr_type(i)->dump_on(st);
4361 }
4362 st->print(" }");
4363 }
4364 #endif // !PRODUCT
4365
4366
4367 #ifdef ASSERT
4368 static bool might_be_same(Node* a, Node* b) {
4369 if (a == b) return true;
4370 if (!(a->is_Phi() || b->is_Phi())) return false;
4371 // phis shift around during optimization
4372 return true; // pretty stupid...
4373 }
4374
4375 // verify a narrow slice (either incoming or outgoing)
4376 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4377 if (!VerifyAliases) return; // don't bother to verify unless requested
4378 if (is_error_reported()) return; // muzzle asserts when debugging an error
4379 if (Node::in_dump()) return; // muzzle asserts when printing
4380 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4381 assert(n != NULL, "");
4382 // Elide intervening MergeMem's
4383 while (n->is_MergeMem()) {
4384 n = n->as_MergeMem()->memory_at(alias_idx);
4385 }
4386 Compile* C = Compile::current();
4387 const TypePtr* n_adr_type = n->adr_type();
4388 if (n == m->empty_memory()) {
4389 // Implicit copy of base_memory()
4390 } else if (n_adr_type != TypePtr::BOTTOM) {
4391 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4392 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4393 } else {
4394 // A few places like make_runtime_call "know" that VM calls are narrow,
4395 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4396 bool expected_wide_mem = false;
4397 if (n == m->base_memory()) {
4398 expected_wide_mem = true;
4399 } else if (alias_idx == Compile::AliasIdxRaw ||
4400 n == m->memory_at(Compile::AliasIdxRaw)) {
4401 expected_wide_mem = true;
4402 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4403 // memory can "leak through" calls on channels that
4404 // are write-once. Allow this also.
4405 expected_wide_mem = true;
4406 }
4407 assert(expected_wide_mem, "expected narrow slice replacement");
4408 }
4409 }
4410 #else // !ASSERT
4411 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4412 #endif
4413
4414
4415 //-----------------------------memory_at---------------------------------------
4416 Node* MergeMemNode::memory_at(uint alias_idx) const {
4417 assert(alias_idx >= Compile::AliasIdxRaw ||
4418 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4419 "must avoid base_memory and AliasIdxTop");
4420
4421 // Otherwise, it is a narrow slice.
4422 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4423 Compile *C = Compile::current();
4424 if (is_empty_memory(n)) {
4425 // the array is sparse; empty slots are the "top" node
4426 n = base_memory();
4427 assert(Node::in_dump()
4428 || n == NULL || n->bottom_type() == Type::TOP
4429 || n->adr_type() == NULL // address is TOP
4430 || n->adr_type() == TypePtr::BOTTOM
4431 || n->adr_type() == TypeRawPtr::BOTTOM
4432 || Compile::current()->AliasLevel() == 0,
4433 "must be a wide memory");
4434 // AliasLevel == 0 if we are organizing the memory states manually.
4435 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4436 } else {
4437 // make sure the stored slice is sane
4438 #ifdef ASSERT
4439 if (is_error_reported() || Node::in_dump()) {
4440 } else if (might_be_same(n, base_memory())) {
4441 // Give it a pass: It is a mostly harmless repetition of the base.
4442 // This can arise normally from node subsumption during optimization.
4443 } else {
4444 verify_memory_slice(this, alias_idx, n);
4445 }
4446 #endif
4447 }
4448 return n;
4449 }
4450
4451 //---------------------------set_memory_at-------------------------------------
4452 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4453 verify_memory_slice(this, alias_idx, n);
4454 Node* empty_mem = empty_memory();
4455 if (n == base_memory()) n = empty_mem; // collapse default
4456 uint need_req = alias_idx+1;
4457 if (req() < need_req) {
4458 if (n == empty_mem) return; // already the default, so do not grow me
4459 // grow the sparse array
4460 do {
4461 add_req(empty_mem);
4462 } while (req() < need_req);
4463 }
4464 set_req( alias_idx, n );
4465 }
4466
4467
4468
4469 //--------------------------iteration_setup------------------------------------
4470 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4471 if (other != NULL) {
4472 grow_to_match(other);
4473 // invariant: the finite support of mm2 is within mm->req()
4474 #ifdef ASSERT
4475 for (uint i = req(); i < other->req(); i++) {
4476 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4477 }
4478 #endif
4479 }
4480 // Replace spurious copies of base_memory by top.
4481 Node* base_mem = base_memory();
4482 if (base_mem != NULL && !base_mem->is_top()) {
4483 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4484 if (in(i) == base_mem)
4485 set_req(i, empty_memory());
4486 }
4487 }
4488 }
4489
4490 //---------------------------grow_to_match-------------------------------------
4491 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4492 Node* empty_mem = empty_memory();
4493 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4494 // look for the finite support of the other memory
4495 for (uint i = other->req(); --i >= req(); ) {
4496 if (other->in(i) != empty_mem) {
4497 uint new_len = i+1;
4498 while (req() < new_len) add_req(empty_mem);
4499 break;
4500 }
4501 }
4502 }
4503
4504 //---------------------------verify_sparse-------------------------------------
4505 #ifndef PRODUCT
4506 bool MergeMemNode::verify_sparse() const {
4507 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4508 Node* base_mem = base_memory();
4509 // The following can happen in degenerate cases, since empty==top.
4510 if (is_empty_memory(base_mem)) return true;
4511 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4512 assert(in(i) != NULL, "sane slice");
4513 if (in(i) == base_mem) return false; // should have been the sentinel value!
4514 }
4515 return true;
4516 }
4517
4518 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4519 Node* n;
4520 n = mm->in(idx);
4521 if (mem == n) return true; // might be empty_memory()
4522 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4523 if (mem == n) return true;
4524 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4525 if (mem == n) return true;
4526 if (n == NULL) break;
4527 }
4528 return false;
4529 }
4530 #endif // !PRODUCT