1 /*
2 * Copyright (c) 1997, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "oops/objArrayKlass.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/arraycopynode.hpp"
32 #include "opto/cfgnode.hpp"
33 #include "opto/compile.hpp"
34 #include "opto/connode.hpp"
35 #include "opto/convertnode.hpp"
36 #include "opto/loopnode.hpp"
37 #include "opto/machnode.hpp"
38 #include "opto/matcher.hpp"
39 #include "opto/memnode.hpp"
40 #include "opto/mulnode.hpp"
41 #include "opto/narrowptrnode.hpp"
42 #include "opto/phaseX.hpp"
43 #include "opto/regmask.hpp"
44 #include "utilities/copy.hpp"
45
46 // Portions of code courtesy of Clifford Click
47
48 // Optimization - Graph Style
49
50 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
51
52 //=============================================================================
53 uint MemNode::size_of() const { return sizeof(*this); }
54
55 const TypePtr *MemNode::adr_type() const {
56 Node* adr = in(Address);
57 if (adr == NULL) return NULL; // node is dead
58 const TypePtr* cross_check = NULL;
59 DEBUG_ONLY(cross_check = _adr_type);
60 return calculate_adr_type(adr->bottom_type(), cross_check);
61 }
62
63 #ifndef PRODUCT
64 void MemNode::dump_spec(outputStream *st) const {
65 if (in(Address) == NULL) return; // node is dead
66 #ifndef ASSERT
67 // fake the missing field
68 const TypePtr* _adr_type = NULL;
69 if (in(Address) != NULL)
70 _adr_type = in(Address)->bottom_type()->isa_ptr();
71 #endif
72 dump_adr_type(this, _adr_type, st);
73
74 Compile* C = Compile::current();
75 if (C->alias_type(_adr_type)->is_volatile()) {
76 st->print(" Volatile!");
77 }
78 if (_unaligned_access) {
79 st->print(" unaligned");
80 }
81 if (_mismatched_access) {
82 st->print(" mismatched");
83 }
84 }
85
86 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
87 st->print(" @");
88 if (adr_type == NULL) {
89 st->print("NULL");
90 } else {
91 adr_type->dump_on(st);
92 Compile* C = Compile::current();
93 Compile::AliasType* atp = NULL;
94 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
95 if (atp == NULL)
96 st->print(", idx=?\?;");
97 else if (atp->index() == Compile::AliasIdxBot)
98 st->print(", idx=Bot;");
99 else if (atp->index() == Compile::AliasIdxTop)
100 st->print(", idx=Top;");
101 else if (atp->index() == Compile::AliasIdxRaw)
102 st->print(", idx=Raw;");
103 else {
104 ciField* field = atp->field();
105 if (field) {
106 st->print(", name=");
107 field->print_name_on(st);
108 }
109 st->print(", idx=%d;", atp->index());
110 }
111 }
112 }
113
114 extern void print_alias_types();
115
116 #endif
117
118 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
119 assert((t_oop != NULL), "sanity");
120 bool is_instance = t_oop->is_known_instance_field();
121 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
122 (load != NULL) && load->is_Load() &&
123 (phase->is_IterGVN() != NULL);
124 if (!(is_instance || is_boxed_value_load))
125 return mchain; // don't try to optimize non-instance types
126 uint instance_id = t_oop->instance_id();
127 Node *start_mem = phase->C->start()->proj_out(TypeFunc::Memory);
128 Node *prev = NULL;
129 Node *result = mchain;
130 while (prev != result) {
131 prev = result;
132 if (result == start_mem)
133 break; // hit one of our sentinels
134 // skip over a call which does not affect this memory slice
135 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
136 Node *proj_in = result->in(0);
137 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
138 break; // hit one of our sentinels
139 } else if (proj_in->is_Call()) {
140 // ArrayCopyNodes processed here as well
141 CallNode *call = proj_in->as_Call();
142 if (!call->may_modify(t_oop, phase)) { // returns false for instances
143 result = call->in(TypeFunc::Memory);
144 }
145 } else if (proj_in->is_Initialize()) {
146 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
147 // Stop if this is the initialization for the object instance which
148 // contains this memory slice, otherwise skip over it.
149 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
150 break;
151 }
152 if (is_instance) {
153 result = proj_in->in(TypeFunc::Memory);
154 } else if (is_boxed_value_load) {
155 Node* klass = alloc->in(AllocateNode::KlassNode);
156 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
157 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
158 result = proj_in->in(TypeFunc::Memory); // not related allocation
159 }
160 }
161 } else if (proj_in->is_MemBar()) {
162 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase)) {
163 break;
164 }
165 result = proj_in->in(TypeFunc::Memory);
166 } else {
167 assert(false, "unexpected projection");
168 }
169 } else if (result->is_ClearArray()) {
170 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
171 // Can not bypass initialization of the instance
172 // we are looking for.
173 break;
174 }
175 // Otherwise skip it (the call updated 'result' value).
176 } else if (result->is_MergeMem()) {
177 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
178 }
179 }
180 return result;
181 }
182
183 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
184 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
185 if (t_oop == NULL)
186 return mchain; // don't try to optimize non-oop types
187 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
188 bool is_instance = t_oop->is_known_instance_field();
189 PhaseIterGVN *igvn = phase->is_IterGVN();
190 if (is_instance && igvn != NULL && result->is_Phi()) {
191 PhiNode *mphi = result->as_Phi();
192 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
193 const TypePtr *t = mphi->adr_type();
194 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
195 t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
196 t->is_oopptr()->cast_to_exactness(true)
197 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
198 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop) {
199 // clone the Phi with our address type
200 result = mphi->split_out_instance(t_adr, igvn);
201 } else {
202 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
203 }
204 }
205 return result;
206 }
207
208 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
209 uint alias_idx = phase->C->get_alias_index(tp);
210 Node *mem = mmem;
211 #ifdef ASSERT
212 {
213 // Check that current type is consistent with the alias index used during graph construction
214 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
215 bool consistent = adr_check == NULL || adr_check->empty() ||
216 phase->C->must_alias(adr_check, alias_idx );
217 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
218 if( !consistent && adr_check != NULL && !adr_check->empty() &&
219 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
220 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
221 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
222 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
223 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
224 // don't assert if it is dead code.
225 consistent = true;
226 }
227 if( !consistent ) {
228 st->print("alias_idx==%d, adr_check==", alias_idx);
229 if( adr_check == NULL ) {
230 st->print("NULL");
231 } else {
232 adr_check->dump();
233 }
234 st->cr();
235 print_alias_types();
236 assert(consistent, "adr_check must match alias idx");
237 }
238 }
239 #endif
240 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
241 // means an array I have not precisely typed yet. Do not do any
242 // alias stuff with it any time soon.
243 const TypeOopPtr *toop = tp->isa_oopptr();
244 if( tp->base() != Type::AnyPtr &&
245 !(toop &&
246 toop->klass() != NULL &&
247 toop->klass()->is_java_lang_Object() &&
248 toop->offset() == Type::OffsetBot) ) {
249 // compress paths and change unreachable cycles to TOP
250 // If not, we can update the input infinitely along a MergeMem cycle
251 // Equivalent code in PhiNode::Ideal
252 Node* m = phase->transform(mmem);
253 // If transformed to a MergeMem, get the desired slice
254 // Otherwise the returned node represents memory for every slice
255 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
256 // Update input if it is progress over what we have now
257 }
258 return mem;
259 }
260
261 //--------------------------Ideal_common---------------------------------------
262 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
263 // Unhook non-raw memories from complete (macro-expanded) initializations.
264 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
265 // If our control input is a dead region, kill all below the region
266 Node *ctl = in(MemNode::Control);
267 if (ctl && remove_dead_region(phase, can_reshape))
268 return this;
269 ctl = in(MemNode::Control);
270 // Don't bother trying to transform a dead node
271 if (ctl && ctl->is_top()) return NodeSentinel;
272
273 PhaseIterGVN *igvn = phase->is_IterGVN();
274 // Wait if control on the worklist.
275 if (ctl && can_reshape && igvn != NULL) {
276 Node* bol = NULL;
277 Node* cmp = NULL;
278 if (ctl->in(0)->is_If()) {
279 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
280 bol = ctl->in(0)->in(1);
281 if (bol->is_Bool())
282 cmp = ctl->in(0)->in(1)->in(1);
283 }
284 if (igvn->_worklist.member(ctl) ||
285 (bol != NULL && igvn->_worklist.member(bol)) ||
286 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
287 // This control path may be dead.
288 // Delay this memory node transformation until the control is processed.
289 phase->is_IterGVN()->_worklist.push(this);
290 return NodeSentinel; // caller will return NULL
291 }
292 }
293 // Ignore if memory is dead, or self-loop
294 Node *mem = in(MemNode::Memory);
295 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
296 assert(mem != this, "dead loop in MemNode::Ideal");
297
298 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
299 // This memory slice may be dead.
300 // Delay this mem node transformation until the memory is processed.
301 phase->is_IterGVN()->_worklist.push(this);
302 return NodeSentinel; // caller will return NULL
303 }
304
305 Node *address = in(MemNode::Address);
306 const Type *t_adr = phase->type(address);
307 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
308
309 if (can_reshape && igvn != NULL &&
310 (igvn->_worklist.member(address) ||
311 igvn->_worklist.size() > 0 && (t_adr != adr_type())) ) {
312 // The address's base and type may change when the address is processed.
313 // Delay this mem node transformation until the address is processed.
314 phase->is_IterGVN()->_worklist.push(this);
315 return NodeSentinel; // caller will return NULL
316 }
317
318 // Do NOT remove or optimize the next lines: ensure a new alias index
319 // is allocated for an oop pointer type before Escape Analysis.
320 // Note: C++ will not remove it since the call has side effect.
321 if (t_adr->isa_oopptr()) {
322 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
323 }
324
325 Node* base = NULL;
326 if (address->is_AddP()) {
327 base = address->in(AddPNode::Base);
328 }
329 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
330 !t_adr->isa_rawptr()) {
331 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
332 // Skip this node optimization if its address has TOP base.
333 return NodeSentinel; // caller will return NULL
334 }
335
336 // Avoid independent memory operations
337 Node* old_mem = mem;
338
339 // The code which unhooks non-raw memories from complete (macro-expanded)
340 // initializations was removed. After macro-expansion all stores catched
341 // by Initialize node became raw stores and there is no information
342 // which memory slices they modify. So it is unsafe to move any memory
343 // operation above these stores. Also in most cases hooked non-raw memories
344 // were already unhooked by using information from detect_ptr_independence()
345 // and find_previous_store().
346
347 if (mem->is_MergeMem()) {
348 MergeMemNode* mmem = mem->as_MergeMem();
349 const TypePtr *tp = t_adr->is_ptr();
350
351 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
352 }
353
354 if (mem != old_mem) {
355 set_req(MemNode::Memory, mem);
356 if (can_reshape && old_mem->outcnt() == 0) {
357 igvn->_worklist.push(old_mem);
358 }
359 if (phase->type( mem ) == Type::TOP) return NodeSentinel;
360 return this;
361 }
362
363 // let the subclass continue analyzing...
364 return NULL;
365 }
366
367 // Helper function for proving some simple control dominations.
368 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
369 // Already assumes that 'dom' is available at 'sub', and that 'sub'
370 // is not a constant (dominated by the method's StartNode).
371 // Used by MemNode::find_previous_store to prove that the
372 // control input of a memory operation predates (dominates)
373 // an allocation it wants to look past.
374 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
375 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
376 return false; // Conservative answer for dead code
377
378 // Check 'dom'. Skip Proj and CatchProj nodes.
379 dom = dom->find_exact_control(dom);
380 if (dom == NULL || dom->is_top())
381 return false; // Conservative answer for dead code
382
383 if (dom == sub) {
384 // For the case when, for example, 'sub' is Initialize and the original
385 // 'dom' is Proj node of the 'sub'.
386 return false;
387 }
388
389 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
390 return true;
391
392 // 'dom' dominates 'sub' if its control edge and control edges
393 // of all its inputs dominate or equal to sub's control edge.
394
395 // Currently 'sub' is either Allocate, Initialize or Start nodes.
396 // Or Region for the check in LoadNode::Ideal();
397 // 'sub' should have sub->in(0) != NULL.
398 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
399 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
400
401 // Get control edge of 'sub'.
402 Node* orig_sub = sub;
403 sub = sub->find_exact_control(sub->in(0));
404 if (sub == NULL || sub->is_top())
405 return false; // Conservative answer for dead code
406
407 assert(sub->is_CFG(), "expecting control");
408
409 if (sub == dom)
410 return true;
411
412 if (sub->is_Start() || sub->is_Root())
413 return false;
414
415 {
416 // Check all control edges of 'dom'.
417
418 ResourceMark rm;
419 Arena* arena = Thread::current()->resource_area();
420 Node_List nlist(arena);
421 Unique_Node_List dom_list(arena);
422
423 dom_list.push(dom);
424 bool only_dominating_controls = false;
425
426 for (uint next = 0; next < dom_list.size(); next++) {
427 Node* n = dom_list.at(next);
428 if (n == orig_sub)
429 return false; // One of dom's inputs dominated by sub.
430 if (!n->is_CFG() && n->pinned()) {
431 // Check only own control edge for pinned non-control nodes.
432 n = n->find_exact_control(n->in(0));
433 if (n == NULL || n->is_top())
434 return false; // Conservative answer for dead code
435 assert(n->is_CFG(), "expecting control");
436 dom_list.push(n);
437 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
438 only_dominating_controls = true;
439 } else if (n->is_CFG()) {
440 if (n->dominates(sub, nlist))
441 only_dominating_controls = true;
442 else
443 return false;
444 } else {
445 // First, own control edge.
446 Node* m = n->find_exact_control(n->in(0));
447 if (m != NULL) {
448 if (m->is_top())
449 return false; // Conservative answer for dead code
450 dom_list.push(m);
451 }
452 // Now, the rest of edges.
453 uint cnt = n->req();
454 for (uint i = 1; i < cnt; i++) {
455 m = n->find_exact_control(n->in(i));
456 if (m == NULL || m->is_top())
457 continue;
458 dom_list.push(m);
459 }
460 }
461 }
462 return only_dominating_controls;
463 }
464 }
465
466 //---------------------detect_ptr_independence---------------------------------
467 // Used by MemNode::find_previous_store to prove that two base
468 // pointers are never equal.
469 // The pointers are accompanied by their associated allocations,
470 // if any, which have been previously discovered by the caller.
471 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
472 Node* p2, AllocateNode* a2,
473 PhaseTransform* phase) {
474 // Attempt to prove that these two pointers cannot be aliased.
475 // They may both manifestly be allocations, and they should differ.
476 // Or, if they are not both allocations, they can be distinct constants.
477 // Otherwise, one is an allocation and the other a pre-existing value.
478 if (a1 == NULL && a2 == NULL) { // neither an allocation
479 return (p1 != p2) && p1->is_Con() && p2->is_Con();
480 } else if (a1 != NULL && a2 != NULL) { // both allocations
481 return (a1 != a2);
482 } else if (a1 != NULL) { // one allocation a1
483 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
484 return all_controls_dominate(p2, a1);
485 } else { //(a2 != NULL) // one allocation a2
486 return all_controls_dominate(p1, a2);
487 }
488 return false;
489 }
490
491
492 // Find an arraycopy that must have set (can_see_stored_value=true) or
493 // could have set (can_see_stored_value=false) the value for this load
494 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
495 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
496 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
497 Node* mb = mem->in(0);
498 if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
499 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
500 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy();
501 if (ac->is_clonebasic()) {
502 intptr_t offset;
503 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset);
504 assert(alloc != NULL && alloc->initialization()->is_complete_with_arraycopy(), "broken allocation");
505 if (alloc == ld_alloc) {
506 return ac;
507 }
508 }
509 }
510 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
511 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
512
513 if (ac->is_arraycopy_validated() ||
514 ac->is_copyof_validated() ||
515 ac->is_copyofrange_validated()) {
516 Node* ld_addp = in(MemNode::Address);
517 if (ld_addp->is_AddP()) {
518 Node* ld_base = ld_addp->in(AddPNode::Address);
519 Node* ld_offs = ld_addp->in(AddPNode::Offset);
520
521 Node* dest = ac->in(ArrayCopyNode::Dest);
522
523 if (dest == ld_base) {
524 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
525 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
526 return ac;
527 }
528 if (!can_see_stored_value) {
529 mem = ac->in(TypeFunc::Memory);
530 }
531 }
532 }
533 }
534 }
535 return NULL;
536 }
537
538 // The logic for reordering loads and stores uses four steps:
539 // (a) Walk carefully past stores and initializations which we
540 // can prove are independent of this load.
541 // (b) Observe that the next memory state makes an exact match
542 // with self (load or store), and locate the relevant store.
543 // (c) Ensure that, if we were to wire self directly to the store,
544 // the optimizer would fold it up somehow.
545 // (d) Do the rewiring, and return, depending on some other part of
546 // the optimizer to fold up the load.
547 // This routine handles steps (a) and (b). Steps (c) and (d) are
548 // specific to loads and stores, so they are handled by the callers.
549 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
550 //
551 Node* MemNode::find_previous_store(PhaseTransform* phase) {
552 Node* ctrl = in(MemNode::Control);
553 Node* adr = in(MemNode::Address);
554 intptr_t offset = 0;
555 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
556 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
557
558 if (offset == Type::OffsetBot)
559 return NULL; // cannot unalias unless there are precise offsets
560
561 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
562
563 intptr_t size_in_bytes = memory_size();
564
565 Node* mem = in(MemNode::Memory); // start searching here...
566
567 int cnt = 50; // Cycle limiter
568 for (;;) { // While we can dance past unrelated stores...
569 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
570
571 Node* prev = mem;
572 if (mem->is_Store()) {
573 Node* st_adr = mem->in(MemNode::Address);
574 intptr_t st_offset = 0;
575 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
576 if (st_base == NULL)
577 break; // inscrutable pointer
578 if (st_offset != offset && st_offset != Type::OffsetBot) {
579 const int MAX_STORE = BytesPerLong;
580 if (st_offset >= offset + size_in_bytes ||
581 st_offset <= offset - MAX_STORE ||
582 st_offset <= offset - mem->as_Store()->memory_size()) {
583 // Success: The offsets are provably independent.
584 // (You may ask, why not just test st_offset != offset and be done?
585 // The answer is that stores of different sizes can co-exist
586 // in the same sequence of RawMem effects. We sometimes initialize
587 // a whole 'tile' of array elements with a single jint or jlong.)
588 mem = mem->in(MemNode::Memory);
589 continue; // (a) advance through independent store memory
590 }
591 }
592 if (st_base != base &&
593 detect_ptr_independence(base, alloc,
594 st_base,
595 AllocateNode::Ideal_allocation(st_base, phase),
596 phase)) {
597 // Success: The bases are provably independent.
598 mem = mem->in(MemNode::Memory);
599 continue; // (a) advance through independent store memory
600 }
601
602 // (b) At this point, if the bases or offsets do not agree, we lose,
603 // since we have not managed to prove 'this' and 'mem' independent.
604 if (st_base == base && st_offset == offset) {
605 return mem; // let caller handle steps (c), (d)
606 }
607
608 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
609 InitializeNode* st_init = mem->in(0)->as_Initialize();
610 AllocateNode* st_alloc = st_init->allocation();
611 if (st_alloc == NULL)
612 break; // something degenerated
613 bool known_identical = false;
614 bool known_independent = false;
615 if (alloc == st_alloc)
616 known_identical = true;
617 else if (alloc != NULL)
618 known_independent = true;
619 else if (all_controls_dominate(this, st_alloc))
620 known_independent = true;
621
622 if (known_independent) {
623 // The bases are provably independent: Either they are
624 // manifestly distinct allocations, or else the control
625 // of this load dominates the store's allocation.
626 int alias_idx = phase->C->get_alias_index(adr_type());
627 if (alias_idx == Compile::AliasIdxRaw) {
628 mem = st_alloc->in(TypeFunc::Memory);
629 } else {
630 mem = st_init->memory(alias_idx);
631 }
632 continue; // (a) advance through independent store memory
633 }
634
635 // (b) at this point, if we are not looking at a store initializing
636 // the same allocation we are loading from, we lose.
637 if (known_identical) {
638 // From caller, can_see_stored_value will consult find_captured_store.
639 return mem; // let caller handle steps (c), (d)
640 }
641
642 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
643 if (prev != mem) {
644 // Found an arraycopy but it doesn't affect that load
645 continue;
646 }
647 // Found an arraycopy that may affect that load
648 return mem;
649 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
650 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
651 if (mem->is_Proj() && mem->in(0)->is_Call()) {
652 // ArrayCopyNodes processed here as well.
653 CallNode *call = mem->in(0)->as_Call();
654 if (!call->may_modify(addr_t, phase)) {
655 mem = call->in(TypeFunc::Memory);
656 continue; // (a) advance through independent call memory
657 }
658 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
659 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase)) {
660 break;
661 }
662 mem = mem->in(0)->in(TypeFunc::Memory);
663 continue; // (a) advance through independent MemBar memory
664 } else if (mem->is_ClearArray()) {
665 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
666 // (the call updated 'mem' value)
667 continue; // (a) advance through independent allocation memory
668 } else {
669 // Can not bypass initialization of the instance
670 // we are looking for.
671 return mem;
672 }
673 } else if (mem->is_MergeMem()) {
674 int alias_idx = phase->C->get_alias_index(adr_type());
675 mem = mem->as_MergeMem()->memory_at(alias_idx);
676 continue; // (a) advance through independent MergeMem memory
677 }
678 }
679
680 // Unless there is an explicit 'continue', we must bail out here,
681 // because 'mem' is an inscrutable memory state (e.g., a call).
682 break;
683 }
684
685 return NULL; // bail out
686 }
687
688 //----------------------calculate_adr_type-------------------------------------
689 // Helper function. Notices when the given type of address hits top or bottom.
690 // Also, asserts a cross-check of the type against the expected address type.
691 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
692 if (t == Type::TOP) return NULL; // does not touch memory any more?
693 #ifdef PRODUCT
694 cross_check = NULL;
695 #else
696 if (!VerifyAliases || is_error_reported() || Node::in_dump()) cross_check = NULL;
697 #endif
698 const TypePtr* tp = t->isa_ptr();
699 if (tp == NULL) {
700 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
701 return TypePtr::BOTTOM; // touches lots of memory
702 } else {
703 #ifdef ASSERT
704 // %%%% [phh] We don't check the alias index if cross_check is
705 // TypeRawPtr::BOTTOM. Needs to be investigated.
706 if (cross_check != NULL &&
707 cross_check != TypePtr::BOTTOM &&
708 cross_check != TypeRawPtr::BOTTOM) {
709 // Recheck the alias index, to see if it has changed (due to a bug).
710 Compile* C = Compile::current();
711 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
712 "must stay in the original alias category");
713 // The type of the address must be contained in the adr_type,
714 // disregarding "null"-ness.
715 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
716 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
717 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
718 "real address must not escape from expected memory type");
719 }
720 #endif
721 return tp;
722 }
723 }
724
725 //=============================================================================
726 // Should LoadNode::Ideal() attempt to remove control edges?
727 bool LoadNode::can_remove_control() const {
728 return true;
729 }
730 uint LoadNode::size_of() const { return sizeof(*this); }
731 uint LoadNode::cmp( const Node &n ) const
732 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
733 const Type *LoadNode::bottom_type() const { return _type; }
734 uint LoadNode::ideal_reg() const {
735 return _type->ideal_reg();
736 }
737
738 #ifndef PRODUCT
739 void LoadNode::dump_spec(outputStream *st) const {
740 MemNode::dump_spec(st);
741 if( !Verbose && !WizardMode ) {
742 // standard dump does this in Verbose and WizardMode
743 st->print(" #"); _type->dump_on(st);
744 }
745 if (!_depends_only_on_test) {
746 st->print(" (does not depend only on test)");
747 }
748 }
749 #endif
750
751 #ifdef ASSERT
752 //----------------------------is_immutable_value-------------------------------
753 // Helper function to allow a raw load without control edge for some cases
754 bool LoadNode::is_immutable_value(Node* adr) {
755 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
756 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
757 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
758 in_bytes(JavaThread::osthread_offset())));
759 }
760 #endif
761
762 //----------------------------LoadNode::make-----------------------------------
763 // Polymorphic factory method:
764 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo,
765 ControlDependency control_dependency, bool unaligned, bool mismatched) {
766 Compile* C = gvn.C;
767
768 // sanity check the alias category against the created node type
769 assert(!(adr_type->isa_oopptr() &&
770 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
771 "use LoadKlassNode instead");
772 assert(!(adr_type->isa_aryptr() &&
773 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
774 "use LoadRangeNode instead");
775 // Check control edge of raw loads
776 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
777 // oop will be recorded in oop map if load crosses safepoint
778 rt->isa_oopptr() || is_immutable_value(adr),
779 "raw memory operations should have control edge");
780 LoadNode* load = NULL;
781 switch (bt) {
782 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
783 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
784 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
785 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
786 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
787 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break;
788 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
789 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
790 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
791 case T_VALUETYPE:
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_VALUETYPE:
2348 case T_OBJECT:
2349 #ifdef _LP64
2350 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2351 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2352 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2353 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2354 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2355 adr->bottom_type()->isa_rawptr())) {
2356 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2357 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2358 }
2359 #endif
2360 {
2361 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2362 }
2363 }
2364 ShouldNotReachHere();
2365 return (StoreNode*)NULL;
2366 }
2367
2368 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2369 bool require_atomic = true;
2370 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2371 }
2372
2373 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2374 bool require_atomic = true;
2375 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2376 }
2377
2378
2379 //--------------------------bottom_type----------------------------------------
2380 const Type *StoreNode::bottom_type() const {
2381 return Type::MEMORY;
2382 }
2383
2384 //------------------------------hash-------------------------------------------
2385 uint StoreNode::hash() const {
2386 // unroll addition of interesting fields
2387 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2388
2389 // Since they are not commoned, do not hash them:
2390 return NO_HASH;
2391 }
2392
2393 //------------------------------Ideal------------------------------------------
2394 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2395 // When a store immediately follows a relevant allocation/initialization,
2396 // try to capture it into the initialization, or hoist it above.
2397 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2398 Node* p = MemNode::Ideal_common(phase, can_reshape);
2399 if (p) return (p == NodeSentinel) ? NULL : p;
2400
2401 Node* mem = in(MemNode::Memory);
2402 Node* address = in(MemNode::Address);
2403 // Back-to-back stores to same address? Fold em up. Generally
2404 // unsafe if I have intervening uses... Also disallowed for StoreCM
2405 // since they must follow each StoreP operation. Redundant StoreCMs
2406 // are eliminated just before matching in final_graph_reshape.
2407 {
2408 Node* st = mem;
2409 // If Store 'st' has more than one use, we cannot fold 'st' away.
2410 // For example, 'st' might be the final state at a conditional
2411 // return. Or, 'st' might be used by some node which is live at
2412 // the same time 'st' is live, which might be unschedulable. So,
2413 // require exactly ONE user until such time as we clone 'mem' for
2414 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2415 // true).
2416 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2417 // Looking at a dead closed cycle of memory?
2418 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2419 assert(Opcode() == st->Opcode() ||
2420 st->Opcode() == Op_StoreVector ||
2421 Opcode() == Op_StoreVector ||
2422 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2423 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2424 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2425 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2426
2427 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2428 st->as_Store()->memory_size() <= this->memory_size()) {
2429 Node* use = st->raw_out(0);
2430 phase->igvn_rehash_node_delayed(use);
2431 if (can_reshape) {
2432 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2433 } else {
2434 // It's OK to do this in the parser, since DU info is always accurate,
2435 // and the parser always refers to nodes via SafePointNode maps.
2436 use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2437 }
2438 return this;
2439 }
2440 st = st->in(MemNode::Memory);
2441 }
2442 }
2443
2444
2445 // Capture an unaliased, unconditional, simple store into an initializer.
2446 // Or, if it is independent of the allocation, hoist it above the allocation.
2447 if (ReduceFieldZeroing && /*can_reshape &&*/
2448 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2449 InitializeNode* init = mem->in(0)->as_Initialize();
2450 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2451 if (offset > 0) {
2452 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2453 // If the InitializeNode captured me, it made a raw copy of me,
2454 // and I need to disappear.
2455 if (moved != NULL) {
2456 // %%% hack to ensure that Ideal returns a new node:
2457 mem = MergeMemNode::make(mem);
2458 return mem; // fold me away
2459 }
2460 }
2461 }
2462
2463 return NULL; // No further progress
2464 }
2465
2466 //------------------------------Value-----------------------------------------
2467 const Type* StoreNode::Value(PhaseGVN* phase) const {
2468 // Either input is TOP ==> the result is TOP
2469 const Type *t1 = phase->type( in(MemNode::Memory) );
2470 if( t1 == Type::TOP ) return Type::TOP;
2471 const Type *t2 = phase->type( in(MemNode::Address) );
2472 if( t2 == Type::TOP ) return Type::TOP;
2473 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2474 if( t3 == Type::TOP ) return Type::TOP;
2475 return Type::MEMORY;
2476 }
2477
2478 //------------------------------Identity---------------------------------------
2479 // Remove redundant stores:
2480 // Store(m, p, Load(m, p)) changes to m.
2481 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2482 Node* StoreNode::Identity(PhaseGVN* phase) {
2483 Node* mem = in(MemNode::Memory);
2484 Node* adr = in(MemNode::Address);
2485 Node* val = in(MemNode::ValueIn);
2486
2487 // Load then Store? Then the Store is useless
2488 if (val->is_Load() &&
2489 val->in(MemNode::Address)->eqv_uncast(adr) &&
2490 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2491 val->as_Load()->store_Opcode() == Opcode()) {
2492 return mem;
2493 }
2494
2495 // Two stores in a row of the same value?
2496 if (mem->is_Store() &&
2497 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2498 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2499 mem->Opcode() == Opcode()) {
2500 return mem;
2501 }
2502
2503 // Store of zero anywhere into a freshly-allocated object?
2504 // Then the store is useless.
2505 // (It must already have been captured by the InitializeNode.)
2506 if (ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2507 // a newly allocated object is already all-zeroes everywhere
2508 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2509 return mem;
2510 }
2511
2512 // the store may also apply to zero-bits in an earlier object
2513 Node* prev_mem = find_previous_store(phase);
2514 // Steps (a), (b): Walk past independent stores to find an exact match.
2515 if (prev_mem != NULL) {
2516 Node* prev_val = can_see_stored_value(prev_mem, phase);
2517 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2518 // prev_val and val might differ by a cast; it would be good
2519 // to keep the more informative of the two.
2520 return mem;
2521 }
2522 }
2523 }
2524
2525 return this;
2526 }
2527
2528 //------------------------------match_edge-------------------------------------
2529 // Do we Match on this edge index or not? Match only memory & value
2530 uint StoreNode::match_edge(uint idx) const {
2531 return idx == MemNode::Address || idx == MemNode::ValueIn;
2532 }
2533
2534 //------------------------------cmp--------------------------------------------
2535 // Do not common stores up together. They generally have to be split
2536 // back up anyways, so do not bother.
2537 uint StoreNode::cmp( const Node &n ) const {
2538 return (&n == this); // Always fail except on self
2539 }
2540
2541 //------------------------------Ideal_masked_input-----------------------------
2542 // Check for a useless mask before a partial-word store
2543 // (StoreB ... (AndI valIn conIa) )
2544 // If (conIa & mask == mask) this simplifies to
2545 // (StoreB ... (valIn) )
2546 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2547 Node *val = in(MemNode::ValueIn);
2548 if( val->Opcode() == Op_AndI ) {
2549 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2550 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2551 set_req(MemNode::ValueIn, val->in(1));
2552 return this;
2553 }
2554 }
2555 return NULL;
2556 }
2557
2558
2559 //------------------------------Ideal_sign_extended_input----------------------
2560 // Check for useless sign-extension before a partial-word store
2561 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2562 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2563 // (StoreB ... (valIn) )
2564 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2565 Node *val = in(MemNode::ValueIn);
2566 if( val->Opcode() == Op_RShiftI ) {
2567 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2568 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2569 Node *shl = val->in(1);
2570 if( shl->Opcode() == Op_LShiftI ) {
2571 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2572 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2573 set_req(MemNode::ValueIn, shl->in(1));
2574 return this;
2575 }
2576 }
2577 }
2578 }
2579 return NULL;
2580 }
2581
2582 //------------------------------value_never_loaded-----------------------------------
2583 // Determine whether there are any possible loads of the value stored.
2584 // For simplicity, we actually check if there are any loads from the
2585 // address stored to, not just for loads of the value stored by this node.
2586 //
2587 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2588 Node *adr = in(Address);
2589 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2590 if (adr_oop == NULL)
2591 return false;
2592 if (!adr_oop->is_known_instance_field())
2593 return false; // if not a distinct instance, there may be aliases of the address
2594 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2595 Node *use = adr->fast_out(i);
2596 if (use->is_Load() || use->is_LoadStore()) {
2597 return false;
2598 }
2599 }
2600 return true;
2601 }
2602
2603 //=============================================================================
2604 //------------------------------Ideal------------------------------------------
2605 // If the store is from an AND mask that leaves the low bits untouched, then
2606 // we can skip the AND operation. If the store is from a sign-extension
2607 // (a left shift, then right shift) we can skip both.
2608 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2609 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2610 if( progress != NULL ) return progress;
2611
2612 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2613 if( progress != NULL ) return progress;
2614
2615 // Finally check the default case
2616 return StoreNode::Ideal(phase, can_reshape);
2617 }
2618
2619 //=============================================================================
2620 //------------------------------Ideal------------------------------------------
2621 // If the store is from an AND mask that leaves the low bits untouched, then
2622 // we can skip the AND operation
2623 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2624 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2625 if( progress != NULL ) return progress;
2626
2627 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2628 if( progress != NULL ) return progress;
2629
2630 // Finally check the default case
2631 return StoreNode::Ideal(phase, can_reshape);
2632 }
2633
2634 //=============================================================================
2635 //------------------------------Identity---------------------------------------
2636 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2637 // No need to card mark when storing a null ptr
2638 Node* my_store = in(MemNode::OopStore);
2639 if (my_store->is_Store()) {
2640 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2641 if( t1 == TypePtr::NULL_PTR ) {
2642 return in(MemNode::Memory);
2643 }
2644 }
2645 return this;
2646 }
2647
2648 //=============================================================================
2649 //------------------------------Ideal---------------------------------------
2650 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2651 Node* progress = StoreNode::Ideal(phase, can_reshape);
2652 if (progress != NULL) return progress;
2653
2654 Node* my_store = in(MemNode::OopStore);
2655 if (my_store->is_MergeMem()) {
2656 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2657 set_req(MemNode::OopStore, mem);
2658 return this;
2659 }
2660
2661 return NULL;
2662 }
2663
2664 //------------------------------Value-----------------------------------------
2665 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2666 // Either input is TOP ==> the result is TOP
2667 const Type *t = phase->type( in(MemNode::Memory) );
2668 if( t == Type::TOP ) return Type::TOP;
2669 t = phase->type( in(MemNode::Address) );
2670 if( t == Type::TOP ) return Type::TOP;
2671 t = phase->type( in(MemNode::ValueIn) );
2672 if( t == Type::TOP ) return Type::TOP;
2673 // If extra input is TOP ==> the result is TOP
2674 t = phase->type( in(MemNode::OopStore) );
2675 if( t == Type::TOP ) return Type::TOP;
2676
2677 return StoreNode::Value( phase );
2678 }
2679
2680
2681 //=============================================================================
2682 //----------------------------------SCMemProjNode------------------------------
2683 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2684 {
2685 return bottom_type();
2686 }
2687
2688 //=============================================================================
2689 //----------------------------------LoadStoreNode------------------------------
2690 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2691 : Node(required),
2692 _type(rt),
2693 _adr_type(at)
2694 {
2695 init_req(MemNode::Control, c );
2696 init_req(MemNode::Memory , mem);
2697 init_req(MemNode::Address, adr);
2698 init_req(MemNode::ValueIn, val);
2699 init_class_id(Class_LoadStore);
2700 }
2701
2702 uint LoadStoreNode::ideal_reg() const {
2703 return _type->ideal_reg();
2704 }
2705
2706 bool LoadStoreNode::result_not_used() const {
2707 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2708 Node *x = fast_out(i);
2709 if (x->Opcode() == Op_SCMemProj) continue;
2710 return false;
2711 }
2712 return true;
2713 }
2714
2715 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2716
2717 //=============================================================================
2718 //----------------------------------LoadStoreConditionalNode--------------------
2719 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2720 init_req(ExpectedIn, ex );
2721 }
2722
2723 //=============================================================================
2724 //-------------------------------adr_type--------------------------------------
2725 const TypePtr* ClearArrayNode::adr_type() const {
2726 Node *adr = in(3);
2727 if (adr == NULL) return NULL; // node is dead
2728 return MemNode::calculate_adr_type(adr->bottom_type());
2729 }
2730
2731 //------------------------------match_edge-------------------------------------
2732 // Do we Match on this edge index or not? Do not match memory
2733 uint ClearArrayNode::match_edge(uint idx) const {
2734 return idx > 1;
2735 }
2736
2737 //------------------------------Identity---------------------------------------
2738 // Clearing a zero length array does nothing
2739 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
2740 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2741 }
2742
2743 //------------------------------Idealize---------------------------------------
2744 // Clearing a short array is faster with stores
2745 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape){
2746 const int unit = BytesPerLong;
2747 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2748 if (!t) return NULL;
2749 if (!t->is_con()) return NULL;
2750 intptr_t raw_count = t->get_con();
2751 intptr_t size = raw_count;
2752 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2753 // Clearing nothing uses the Identity call.
2754 // Negative clears are possible on dead ClearArrays
2755 // (see jck test stmt114.stmt11402.val).
2756 if (size <= 0 || size % unit != 0) return NULL;
2757 intptr_t count = size / unit;
2758 // Length too long; use fast hardware clear
2759 if (size > Matcher::init_array_short_size) return NULL;
2760 Node *mem = in(1);
2761 if( phase->type(mem)==Type::TOP ) return NULL;
2762 Node *adr = in(3);
2763 const Type* at = phase->type(adr);
2764 if( at==Type::TOP ) return NULL;
2765 const TypePtr* atp = at->isa_ptr();
2766 // adjust atp to be the correct array element address type
2767 if (atp == NULL) atp = TypePtr::BOTTOM;
2768 else atp = atp->add_offset(Type::OffsetBot);
2769 // Get base for derived pointer purposes
2770 if( adr->Opcode() != Op_AddP ) Unimplemented();
2771 Node *base = adr->in(1);
2772
2773 Node *zero = phase->makecon(TypeLong::ZERO);
2774 Node *off = phase->MakeConX(BytesPerLong);
2775 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2776 count--;
2777 while( count-- ) {
2778 mem = phase->transform(mem);
2779 adr = phase->transform(new AddPNode(base,adr,off));
2780 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2781 }
2782 return mem;
2783 }
2784
2785 //----------------------------step_through----------------------------------
2786 // Return allocation input memory edge if it is different instance
2787 // or itself if it is the one we are looking for.
2788 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2789 Node* n = *np;
2790 assert(n->is_ClearArray(), "sanity");
2791 intptr_t offset;
2792 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2793 // This method is called only before Allocate nodes are expanded
2794 // during macro nodes expansion. Before that ClearArray nodes are
2795 // only generated in PhaseMacroExpand::generate_arraycopy() (before
2796 // Allocate nodes are expanded) which follows allocations.
2797 assert(alloc != NULL, "should have allocation");
2798 if (alloc->_idx == instance_id) {
2799 // Can not bypass initialization of the instance we are looking for.
2800 return false;
2801 }
2802 // Otherwise skip it.
2803 InitializeNode* init = alloc->initialization();
2804 if (init != NULL)
2805 *np = init->in(TypeFunc::Memory);
2806 else
2807 *np = alloc->in(TypeFunc::Memory);
2808 return true;
2809 }
2810
2811 //----------------------------clear_memory-------------------------------------
2812 // Generate code to initialize object storage to zero.
2813 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2814 intptr_t start_offset,
2815 Node* end_offset,
2816 PhaseGVN* phase) {
2817 intptr_t offset = start_offset;
2818
2819 int unit = BytesPerLong;
2820 if ((offset % unit) != 0) {
2821 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
2822 adr = phase->transform(adr);
2823 const TypePtr* atp = TypeRawPtr::BOTTOM;
2824 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2825 mem = phase->transform(mem);
2826 offset += BytesPerInt;
2827 }
2828 assert((offset % unit) == 0, "");
2829
2830 // Initialize the remaining stuff, if any, with a ClearArray.
2831 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2832 }
2833
2834 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2835 Node* start_offset,
2836 Node* end_offset,
2837 PhaseGVN* phase) {
2838 if (start_offset == end_offset) {
2839 // nothing to do
2840 return mem;
2841 }
2842
2843 int unit = BytesPerLong;
2844 Node* zbase = start_offset;
2845 Node* zend = end_offset;
2846
2847 // Scale to the unit required by the CPU:
2848 if (!Matcher::init_array_count_is_in_bytes) {
2849 Node* shift = phase->intcon(exact_log2(unit));
2850 zbase = phase->transform(new URShiftXNode(zbase, shift) );
2851 zend = phase->transform(new URShiftXNode(zend, shift) );
2852 }
2853
2854 // Bulk clear double-words
2855 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
2856 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
2857 mem = new ClearArrayNode(ctl, mem, zsize, adr);
2858 return phase->transform(mem);
2859 }
2860
2861 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2862 intptr_t start_offset,
2863 intptr_t end_offset,
2864 PhaseGVN* phase) {
2865 if (start_offset == end_offset) {
2866 // nothing to do
2867 return mem;
2868 }
2869
2870 assert((end_offset % BytesPerInt) == 0, "odd end offset");
2871 intptr_t done_offset = end_offset;
2872 if ((done_offset % BytesPerLong) != 0) {
2873 done_offset -= BytesPerInt;
2874 }
2875 if (done_offset > start_offset) {
2876 mem = clear_memory(ctl, mem, dest,
2877 start_offset, phase->MakeConX(done_offset), phase);
2878 }
2879 if (done_offset < end_offset) { // emit the final 32-bit store
2880 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
2881 adr = phase->transform(adr);
2882 const TypePtr* atp = TypeRawPtr::BOTTOM;
2883 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2884 mem = phase->transform(mem);
2885 done_offset += BytesPerInt;
2886 }
2887 assert(done_offset == end_offset, "");
2888 return mem;
2889 }
2890
2891 //=============================================================================
2892 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
2893 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
2894 _adr_type(C->get_adr_type(alias_idx))
2895 {
2896 init_class_id(Class_MemBar);
2897 Node* top = C->top();
2898 init_req(TypeFunc::I_O,top);
2899 init_req(TypeFunc::FramePtr,top);
2900 init_req(TypeFunc::ReturnAdr,top);
2901 if (precedent != NULL)
2902 init_req(TypeFunc::Parms, precedent);
2903 }
2904
2905 //------------------------------cmp--------------------------------------------
2906 uint MemBarNode::hash() const { return NO_HASH; }
2907 uint MemBarNode::cmp( const Node &n ) const {
2908 return (&n == this); // Always fail except on self
2909 }
2910
2911 //------------------------------make-------------------------------------------
2912 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
2913 switch (opcode) {
2914 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
2915 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
2916 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
2917 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
2918 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
2919 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
2920 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
2921 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
2922 case Op_Initialize: return new InitializeNode(C, atp, pn);
2923 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
2924 default: ShouldNotReachHere(); return NULL;
2925 }
2926 }
2927
2928 //------------------------------Ideal------------------------------------------
2929 // Return a node which is more "ideal" than the current node. Strip out
2930 // control copies
2931 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2932 if (remove_dead_region(phase, can_reshape)) return this;
2933 // Don't bother trying to transform a dead node
2934 if (in(0) && in(0)->is_top()) {
2935 return NULL;
2936 }
2937
2938 bool progress = false;
2939 // Eliminate volatile MemBars for scalar replaced objects.
2940 if (can_reshape && req() == (Precedent+1)) {
2941 bool eliminate = false;
2942 int opc = Opcode();
2943 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
2944 // Volatile field loads and stores.
2945 Node* my_mem = in(MemBarNode::Precedent);
2946 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
2947 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
2948 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
2949 // replace this Precedent (decodeN) with the Load instead.
2950 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
2951 Node* load_node = my_mem->in(1);
2952 set_req(MemBarNode::Precedent, load_node);
2953 phase->is_IterGVN()->_worklist.push(my_mem);
2954 my_mem = load_node;
2955 } else {
2956 assert(my_mem->unique_out() == this, "sanity");
2957 del_req(Precedent);
2958 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
2959 my_mem = NULL;
2960 }
2961 progress = true;
2962 }
2963 if (my_mem != NULL && my_mem->is_Mem()) {
2964 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
2965 // Check for scalar replaced object reference.
2966 if( t_oop != NULL && t_oop->is_known_instance_field() &&
2967 t_oop->offset() != Type::OffsetBot &&
2968 t_oop->offset() != Type::OffsetTop) {
2969 eliminate = true;
2970 }
2971 }
2972 } else if (opc == Op_MemBarRelease) {
2973 // Final field stores.
2974 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
2975 if ((alloc != NULL) && alloc->is_Allocate() &&
2976 alloc->as_Allocate()->does_not_escape_thread()) {
2977 // The allocated object does not escape.
2978 eliminate = true;
2979 }
2980 }
2981 if (eliminate) {
2982 // Replace MemBar projections by its inputs.
2983 PhaseIterGVN* igvn = phase->is_IterGVN();
2984 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
2985 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
2986 // Must return either the original node (now dead) or a new node
2987 // (Do not return a top here, since that would break the uniqueness of top.)
2988 return new ConINode(TypeInt::ZERO);
2989 }
2990 }
2991 return progress ? this : NULL;
2992 }
2993
2994 //------------------------------Value------------------------------------------
2995 const Type* MemBarNode::Value(PhaseGVN* phase) const {
2996 if( !in(0) ) return Type::TOP;
2997 if( phase->type(in(0)) == Type::TOP )
2998 return Type::TOP;
2999 return TypeTuple::MEMBAR;
3000 }
3001
3002 //------------------------------match------------------------------------------
3003 // Construct projections for memory.
3004 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3005 switch (proj->_con) {
3006 case TypeFunc::Control:
3007 case TypeFunc::Memory:
3008 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3009 }
3010 ShouldNotReachHere();
3011 return NULL;
3012 }
3013
3014 //===========================InitializeNode====================================
3015 // SUMMARY:
3016 // This node acts as a memory barrier on raw memory, after some raw stores.
3017 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3018 // The Initialize can 'capture' suitably constrained stores as raw inits.
3019 // It can coalesce related raw stores into larger units (called 'tiles').
3020 // It can avoid zeroing new storage for memory units which have raw inits.
3021 // At macro-expansion, it is marked 'complete', and does not optimize further.
3022 //
3023 // EXAMPLE:
3024 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3025 // ctl = incoming control; mem* = incoming memory
3026 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3027 // First allocate uninitialized memory and fill in the header:
3028 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3029 // ctl := alloc.Control; mem* := alloc.Memory*
3030 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3031 // Then initialize to zero the non-header parts of the raw memory block:
3032 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3033 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3034 // After the initialize node executes, the object is ready for service:
3035 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3036 // Suppose its body is immediately initialized as {1,2}:
3037 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3038 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3039 // mem.SLICE(#short[*]) := store2
3040 //
3041 // DETAILS:
3042 // An InitializeNode collects and isolates object initialization after
3043 // an AllocateNode and before the next possible safepoint. As a
3044 // memory barrier (MemBarNode), it keeps critical stores from drifting
3045 // down past any safepoint or any publication of the allocation.
3046 // Before this barrier, a newly-allocated object may have uninitialized bits.
3047 // After this barrier, it may be treated as a real oop, and GC is allowed.
3048 //
3049 // The semantics of the InitializeNode include an implicit zeroing of
3050 // the new object from object header to the end of the object.
3051 // (The object header and end are determined by the AllocateNode.)
3052 //
3053 // Certain stores may be added as direct inputs to the InitializeNode.
3054 // These stores must update raw memory, and they must be to addresses
3055 // derived from the raw address produced by AllocateNode, and with
3056 // a constant offset. They must be ordered by increasing offset.
3057 // The first one is at in(RawStores), the last at in(req()-1).
3058 // Unlike most memory operations, they are not linked in a chain,
3059 // but are displayed in parallel as users of the rawmem output of
3060 // the allocation.
3061 //
3062 // (See comments in InitializeNode::capture_store, which continue
3063 // the example given above.)
3064 //
3065 // When the associated Allocate is macro-expanded, the InitializeNode
3066 // may be rewritten to optimize collected stores. A ClearArrayNode
3067 // may also be created at that point to represent any required zeroing.
3068 // The InitializeNode is then marked 'complete', prohibiting further
3069 // capturing of nearby memory operations.
3070 //
3071 // During macro-expansion, all captured initializations which store
3072 // constant values of 32 bits or smaller are coalesced (if advantageous)
3073 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3074 // initialized in fewer memory operations. Memory words which are
3075 // covered by neither tiles nor non-constant stores are pre-zeroed
3076 // by explicit stores of zero. (The code shape happens to do all
3077 // zeroing first, then all other stores, with both sequences occurring
3078 // in order of ascending offsets.)
3079 //
3080 // Alternatively, code may be inserted between an AllocateNode and its
3081 // InitializeNode, to perform arbitrary initialization of the new object.
3082 // E.g., the object copying intrinsics insert complex data transfers here.
3083 // The initialization must then be marked as 'complete' disable the
3084 // built-in zeroing semantics and the collection of initializing stores.
3085 //
3086 // While an InitializeNode is incomplete, reads from the memory state
3087 // produced by it are optimizable if they match the control edge and
3088 // new oop address associated with the allocation/initialization.
3089 // They return a stored value (if the offset matches) or else zero.
3090 // A write to the memory state, if it matches control and address,
3091 // and if it is to a constant offset, may be 'captured' by the
3092 // InitializeNode. It is cloned as a raw memory operation and rewired
3093 // inside the initialization, to the raw oop produced by the allocation.
3094 // Operations on addresses which are provably distinct (e.g., to
3095 // other AllocateNodes) are allowed to bypass the initialization.
3096 //
3097 // The effect of all this is to consolidate object initialization
3098 // (both arrays and non-arrays, both piecewise and bulk) into a
3099 // single location, where it can be optimized as a unit.
3100 //
3101 // Only stores with an offset less than TrackedInitializationLimit words
3102 // will be considered for capture by an InitializeNode. This puts a
3103 // reasonable limit on the complexity of optimized initializations.
3104
3105 //---------------------------InitializeNode------------------------------------
3106 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3107 : _is_complete(Incomplete), _does_not_escape(false),
3108 MemBarNode(C, adr_type, rawoop)
3109 {
3110 init_class_id(Class_Initialize);
3111
3112 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3113 assert(in(RawAddress) == rawoop, "proper init");
3114 // Note: allocation() can be NULL, for secondary initialization barriers
3115 }
3116
3117 // Since this node is not matched, it will be processed by the
3118 // register allocator. Declare that there are no constraints
3119 // on the allocation of the RawAddress edge.
3120 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3121 // This edge should be set to top, by the set_complete. But be conservative.
3122 if (idx == InitializeNode::RawAddress)
3123 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3124 return RegMask::Empty;
3125 }
3126
3127 Node* InitializeNode::memory(uint alias_idx) {
3128 Node* mem = in(Memory);
3129 if (mem->is_MergeMem()) {
3130 return mem->as_MergeMem()->memory_at(alias_idx);
3131 } else {
3132 // incoming raw memory is not split
3133 return mem;
3134 }
3135 }
3136
3137 bool InitializeNode::is_non_zero() {
3138 if (is_complete()) return false;
3139 remove_extra_zeroes();
3140 return (req() > RawStores);
3141 }
3142
3143 void InitializeNode::set_complete(PhaseGVN* phase) {
3144 assert(!is_complete(), "caller responsibility");
3145 _is_complete = Complete;
3146
3147 // After this node is complete, it contains a bunch of
3148 // raw-memory initializations. There is no need for
3149 // it to have anything to do with non-raw memory effects.
3150 // Therefore, tell all non-raw users to re-optimize themselves,
3151 // after skipping the memory effects of this initialization.
3152 PhaseIterGVN* igvn = phase->is_IterGVN();
3153 if (igvn) igvn->add_users_to_worklist(this);
3154 }
3155
3156 // convenience function
3157 // return false if the init contains any stores already
3158 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3159 InitializeNode* init = initialization();
3160 if (init == NULL || init->is_complete()) return false;
3161 init->remove_extra_zeroes();
3162 // for now, if this allocation has already collected any inits, bail:
3163 if (init->is_non_zero()) return false;
3164 init->set_complete(phase);
3165 return true;
3166 }
3167
3168 void InitializeNode::remove_extra_zeroes() {
3169 if (req() == RawStores) return;
3170 Node* zmem = zero_memory();
3171 uint fill = RawStores;
3172 for (uint i = fill; i < req(); i++) {
3173 Node* n = in(i);
3174 if (n->is_top() || n == zmem) continue; // skip
3175 if (fill < i) set_req(fill, n); // compact
3176 ++fill;
3177 }
3178 // delete any empty spaces created:
3179 while (fill < req()) {
3180 del_req(fill);
3181 }
3182 }
3183
3184 // Helper for remembering which stores go with which offsets.
3185 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3186 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3187 intptr_t offset = -1;
3188 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3189 phase, offset);
3190 if (base == NULL) return -1; // something is dead,
3191 if (offset < 0) return -1; // dead, dead
3192 return offset;
3193 }
3194
3195 // Helper for proving that an initialization expression is
3196 // "simple enough" to be folded into an object initialization.
3197 // Attempts to prove that a store's initial value 'n' can be captured
3198 // within the initialization without creating a vicious cycle, such as:
3199 // { Foo p = new Foo(); p.next = p; }
3200 // True for constants and parameters and small combinations thereof.
3201 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3202 if (n == NULL) return true; // (can this really happen?)
3203 if (n->is_Proj()) n = n->in(0);
3204 if (n == this) return false; // found a cycle
3205 if (n->is_Con()) return true;
3206 if (n->is_Start()) return true; // params, etc., are OK
3207 if (n->is_Root()) return true; // even better
3208
3209 Node* ctl = n->in(0);
3210 if (ctl != NULL && !ctl->is_top()) {
3211 if (ctl->is_Proj()) ctl = ctl->in(0);
3212 if (ctl == this) return false;
3213
3214 // If we already know that the enclosing memory op is pinned right after
3215 // the init, then any control flow that the store has picked up
3216 // must have preceded the init, or else be equal to the init.
3217 // Even after loop optimizations (which might change control edges)
3218 // a store is never pinned *before* the availability of its inputs.
3219 if (!MemNode::all_controls_dominate(n, this))
3220 return false; // failed to prove a good control
3221 }
3222
3223 // Check data edges for possible dependencies on 'this'.
3224 if ((count += 1) > 20) return false; // complexity limit
3225 for (uint i = 1; i < n->req(); i++) {
3226 Node* m = n->in(i);
3227 if (m == NULL || m == n || m->is_top()) continue;
3228 uint first_i = n->find_edge(m);
3229 if (i != first_i) continue; // process duplicate edge just once
3230 if (!detect_init_independence(m, count)) {
3231 return false;
3232 }
3233 }
3234
3235 return true;
3236 }
3237
3238 // Here are all the checks a Store must pass before it can be moved into
3239 // an initialization. Returns zero if a check fails.
3240 // On success, returns the (constant) offset to which the store applies,
3241 // within the initialized memory.
3242 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3243 const int FAIL = 0;
3244 if (st->is_unaligned_access()) {
3245 return FAIL;
3246 }
3247 if (st->req() != MemNode::ValueIn + 1)
3248 return FAIL; // an inscrutable StoreNode (card mark?)
3249 Node* ctl = st->in(MemNode::Control);
3250 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3251 return FAIL; // must be unconditional after the initialization
3252 Node* mem = st->in(MemNode::Memory);
3253 if (!(mem->is_Proj() && mem->in(0) == this))
3254 return FAIL; // must not be preceded by other stores
3255 Node* adr = st->in(MemNode::Address);
3256 intptr_t offset;
3257 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3258 if (alloc == NULL)
3259 return FAIL; // inscrutable address
3260 if (alloc != allocation())
3261 return FAIL; // wrong allocation! (store needs to float up)
3262 Node* val = st->in(MemNode::ValueIn);
3263 int complexity_count = 0;
3264 if (!detect_init_independence(val, complexity_count))
3265 return FAIL; // stored value must be 'simple enough'
3266
3267 // The Store can be captured only if nothing after the allocation
3268 // and before the Store is using the memory location that the store
3269 // overwrites.
3270 bool failed = false;
3271 // If is_complete_with_arraycopy() is true the shape of the graph is
3272 // well defined and is safe so no need for extra checks.
3273 if (!is_complete_with_arraycopy()) {
3274 // We are going to look at each use of the memory state following
3275 // the allocation to make sure nothing reads the memory that the
3276 // Store writes.
3277 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3278 int alias_idx = phase->C->get_alias_index(t_adr);
3279 ResourceMark rm;
3280 Unique_Node_List mems;
3281 mems.push(mem);
3282 Node* unique_merge = NULL;
3283 for (uint next = 0; next < mems.size(); ++next) {
3284 Node *m = mems.at(next);
3285 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3286 Node *n = m->fast_out(j);
3287 if (n->outcnt() == 0) {
3288 continue;
3289 }
3290 if (n == st) {
3291 continue;
3292 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3293 // If the control of this use is different from the control
3294 // of the Store which is right after the InitializeNode then
3295 // this node cannot be between the InitializeNode and the
3296 // Store.
3297 continue;
3298 } else if (n->is_MergeMem()) {
3299 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3300 // We can hit a MergeMemNode (that will likely go away
3301 // later) that is a direct use of the memory state
3302 // following the InitializeNode on the same slice as the
3303 // store node that we'd like to capture. We need to check
3304 // the uses of the MergeMemNode.
3305 mems.push(n);
3306 }
3307 } else if (n->is_Mem()) {
3308 Node* other_adr = n->in(MemNode::Address);
3309 if (other_adr == adr) {
3310 failed = true;
3311 break;
3312 } else {
3313 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3314 if (other_t_adr != NULL) {
3315 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3316 if (other_alias_idx == alias_idx) {
3317 // A load from the same memory slice as the store right
3318 // after the InitializeNode. We check the control of the
3319 // object/array that is loaded from. If it's the same as
3320 // the store control then we cannot capture the store.
3321 assert(!n->is_Store(), "2 stores to same slice on same control?");
3322 Node* base = other_adr;
3323 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3324 base = base->in(AddPNode::Base);
3325 if (base != NULL) {
3326 base = base->uncast();
3327 if (base->is_Proj() && base->in(0) == alloc) {
3328 failed = true;
3329 break;
3330 }
3331 }
3332 }
3333 }
3334 }
3335 } else {
3336 failed = true;
3337 break;
3338 }
3339 }
3340 }
3341 }
3342 if (failed) {
3343 if (!can_reshape) {
3344 // We decided we couldn't capture the store during parsing. We
3345 // should try again during the next IGVN once the graph is
3346 // cleaner.
3347 phase->C->record_for_igvn(st);
3348 }
3349 return FAIL;
3350 }
3351
3352 return offset; // success
3353 }
3354
3355 // Find the captured store in(i) which corresponds to the range
3356 // [start..start+size) in the initialized object.
3357 // If there is one, return its index i. If there isn't, return the
3358 // negative of the index where it should be inserted.
3359 // Return 0 if the queried range overlaps an initialization boundary
3360 // or if dead code is encountered.
3361 // If size_in_bytes is zero, do not bother with overlap checks.
3362 int InitializeNode::captured_store_insertion_point(intptr_t start,
3363 int size_in_bytes,
3364 PhaseTransform* phase) {
3365 const int FAIL = 0, MAX_STORE = BytesPerLong;
3366
3367 if (is_complete())
3368 return FAIL; // arraycopy got here first; punt
3369
3370 assert(allocation() != NULL, "must be present");
3371
3372 // no negatives, no header fields:
3373 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3374
3375 // after a certain size, we bail out on tracking all the stores:
3376 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3377 if (start >= ti_limit) return FAIL;
3378
3379 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3380 if (i >= limit) return -(int)i; // not found; here is where to put it
3381
3382 Node* st = in(i);
3383 intptr_t st_off = get_store_offset(st, phase);
3384 if (st_off < 0) {
3385 if (st != zero_memory()) {
3386 return FAIL; // bail out if there is dead garbage
3387 }
3388 } else if (st_off > start) {
3389 // ...we are done, since stores are ordered
3390 if (st_off < start + size_in_bytes) {
3391 return FAIL; // the next store overlaps
3392 }
3393 return -(int)i; // not found; here is where to put it
3394 } else if (st_off < start) {
3395 if (size_in_bytes != 0 &&
3396 start < st_off + MAX_STORE &&
3397 start < st_off + st->as_Store()->memory_size()) {
3398 return FAIL; // the previous store overlaps
3399 }
3400 } else {
3401 if (size_in_bytes != 0 &&
3402 st->as_Store()->memory_size() != size_in_bytes) {
3403 return FAIL; // mismatched store size
3404 }
3405 return i;
3406 }
3407
3408 ++i;
3409 }
3410 }
3411
3412 // Look for a captured store which initializes at the offset 'start'
3413 // with the given size. If there is no such store, and no other
3414 // initialization interferes, then return zero_memory (the memory
3415 // projection of the AllocateNode).
3416 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3417 PhaseTransform* phase) {
3418 assert(stores_are_sane(phase), "");
3419 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3420 if (i == 0) {
3421 return NULL; // something is dead
3422 } else if (i < 0) {
3423 return zero_memory(); // just primordial zero bits here
3424 } else {
3425 Node* st = in(i); // here is the store at this position
3426 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3427 return st;
3428 }
3429 }
3430
3431 // Create, as a raw pointer, an address within my new object at 'offset'.
3432 Node* InitializeNode::make_raw_address(intptr_t offset,
3433 PhaseTransform* phase) {
3434 Node* addr = in(RawAddress);
3435 if (offset != 0) {
3436 Compile* C = phase->C;
3437 addr = phase->transform( new AddPNode(C->top(), addr,
3438 phase->MakeConX(offset)) );
3439 }
3440 return addr;
3441 }
3442
3443 // Clone the given store, converting it into a raw store
3444 // initializing a field or element of my new object.
3445 // Caller is responsible for retiring the original store,
3446 // with subsume_node or the like.
3447 //
3448 // From the example above InitializeNode::InitializeNode,
3449 // here are the old stores to be captured:
3450 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3451 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3452 //
3453 // Here is the changed code; note the extra edges on init:
3454 // alloc = (Allocate ...)
3455 // rawoop = alloc.RawAddress
3456 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3457 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3458 // init = (Initialize alloc.Control alloc.Memory rawoop
3459 // rawstore1 rawstore2)
3460 //
3461 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3462 PhaseTransform* phase, bool can_reshape) {
3463 assert(stores_are_sane(phase), "");
3464
3465 if (start < 0) return NULL;
3466 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3467
3468 Compile* C = phase->C;
3469 int size_in_bytes = st->memory_size();
3470 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3471 if (i == 0) return NULL; // bail out
3472 Node* prev_mem = NULL; // raw memory for the captured store
3473 if (i > 0) {
3474 prev_mem = in(i); // there is a pre-existing store under this one
3475 set_req(i, C->top()); // temporarily disconnect it
3476 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3477 } else {
3478 i = -i; // no pre-existing store
3479 prev_mem = zero_memory(); // a slice of the newly allocated object
3480 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3481 set_req(--i, C->top()); // reuse this edge; it has been folded away
3482 else
3483 ins_req(i, C->top()); // build a new edge
3484 }
3485 Node* new_st = st->clone();
3486 new_st->set_req(MemNode::Control, in(Control));
3487 new_st->set_req(MemNode::Memory, prev_mem);
3488 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3489 new_st = phase->transform(new_st);
3490
3491 // At this point, new_st might have swallowed a pre-existing store
3492 // at the same offset, or perhaps new_st might have disappeared,
3493 // if it redundantly stored the same value (or zero to fresh memory).
3494
3495 // In any case, wire it in:
3496 phase->igvn_rehash_node_delayed(this);
3497 set_req(i, new_st);
3498
3499 // The caller may now kill the old guy.
3500 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3501 assert(check_st == new_st || check_st == NULL, "must be findable");
3502 assert(!is_complete(), "");
3503 return new_st;
3504 }
3505
3506 static bool store_constant(jlong* tiles, int num_tiles,
3507 intptr_t st_off, int st_size,
3508 jlong con) {
3509 if ((st_off & (st_size-1)) != 0)
3510 return false; // strange store offset (assume size==2**N)
3511 address addr = (address)tiles + st_off;
3512 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3513 switch (st_size) {
3514 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3515 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3516 case sizeof(jint): *(jint*) addr = (jint) con; break;
3517 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3518 default: return false; // strange store size (detect size!=2**N here)
3519 }
3520 return true; // return success to caller
3521 }
3522
3523 // Coalesce subword constants into int constants and possibly
3524 // into long constants. The goal, if the CPU permits,
3525 // is to initialize the object with a small number of 64-bit tiles.
3526 // Also, convert floating-point constants to bit patterns.
3527 // Non-constants are not relevant to this pass.
3528 //
3529 // In terms of the running example on InitializeNode::InitializeNode
3530 // and InitializeNode::capture_store, here is the transformation
3531 // of rawstore1 and rawstore2 into rawstore12:
3532 // alloc = (Allocate ...)
3533 // rawoop = alloc.RawAddress
3534 // tile12 = 0x00010002
3535 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3536 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3537 //
3538 void
3539 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3540 Node* size_in_bytes,
3541 PhaseGVN* phase) {
3542 Compile* C = phase->C;
3543
3544 assert(stores_are_sane(phase), "");
3545 // Note: After this pass, they are not completely sane,
3546 // since there may be some overlaps.
3547
3548 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3549
3550 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3551 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3552 size_limit = MIN2(size_limit, ti_limit);
3553 size_limit = align_size_up(size_limit, BytesPerLong);
3554 int num_tiles = size_limit / BytesPerLong;
3555
3556 // allocate space for the tile map:
3557 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3558 jlong tiles_buf[small_len];
3559 Node* nodes_buf[small_len];
3560 jlong inits_buf[small_len];
3561 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3562 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3563 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3564 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3565 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3566 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3567 // tiles: exact bitwise model of all primitive constants
3568 // nodes: last constant-storing node subsumed into the tiles model
3569 // inits: which bytes (in each tile) are touched by any initializations
3570
3571 //// Pass A: Fill in the tile model with any relevant stores.
3572
3573 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3574 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3575 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3576 Node* zmem = zero_memory(); // initially zero memory state
3577 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3578 Node* st = in(i);
3579 intptr_t st_off = get_store_offset(st, phase);
3580
3581 // Figure out the store's offset and constant value:
3582 if (st_off < header_size) continue; //skip (ignore header)
3583 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3584 int st_size = st->as_Store()->memory_size();
3585 if (st_off + st_size > size_limit) break;
3586
3587 // Record which bytes are touched, whether by constant or not.
3588 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3589 continue; // skip (strange store size)
3590
3591 const Type* val = phase->type(st->in(MemNode::ValueIn));
3592 if (!val->singleton()) continue; //skip (non-con store)
3593 BasicType type = val->basic_type();
3594
3595 jlong con = 0;
3596 switch (type) {
3597 case T_INT: con = val->is_int()->get_con(); break;
3598 case T_LONG: con = val->is_long()->get_con(); break;
3599 case T_FLOAT: con = jint_cast(val->getf()); break;
3600 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3601 default: continue; //skip (odd store type)
3602 }
3603
3604 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3605 st->Opcode() == Op_StoreL) {
3606 continue; // This StoreL is already optimal.
3607 }
3608
3609 // Store down the constant.
3610 store_constant(tiles, num_tiles, st_off, st_size, con);
3611
3612 intptr_t j = st_off >> LogBytesPerLong;
3613
3614 if (type == T_INT && st_size == BytesPerInt
3615 && (st_off & BytesPerInt) == BytesPerInt) {
3616 jlong lcon = tiles[j];
3617 if (!Matcher::isSimpleConstant64(lcon) &&
3618 st->Opcode() == Op_StoreI) {
3619 // This StoreI is already optimal by itself.
3620 jint* intcon = (jint*) &tiles[j];
3621 intcon[1] = 0; // undo the store_constant()
3622
3623 // If the previous store is also optimal by itself, back up and
3624 // undo the action of the previous loop iteration... if we can.
3625 // But if we can't, just let the previous half take care of itself.
3626 st = nodes[j];
3627 st_off -= BytesPerInt;
3628 con = intcon[0];
3629 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3630 assert(st_off >= header_size, "still ignoring header");
3631 assert(get_store_offset(st, phase) == st_off, "must be");
3632 assert(in(i-1) == zmem, "must be");
3633 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3634 assert(con == tcon->is_int()->get_con(), "must be");
3635 // Undo the effects of the previous loop trip, which swallowed st:
3636 intcon[0] = 0; // undo store_constant()
3637 set_req(i-1, st); // undo set_req(i, zmem)
3638 nodes[j] = NULL; // undo nodes[j] = st
3639 --old_subword; // undo ++old_subword
3640 }
3641 continue; // This StoreI is already optimal.
3642 }
3643 }
3644
3645 // This store is not needed.
3646 set_req(i, zmem);
3647 nodes[j] = st; // record for the moment
3648 if (st_size < BytesPerLong) // something has changed
3649 ++old_subword; // includes int/float, but who's counting...
3650 else ++old_long;
3651 }
3652
3653 if ((old_subword + old_long) == 0)
3654 return; // nothing more to do
3655
3656 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3657 // Be sure to insert them before overlapping non-constant stores.
3658 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3659 for (int j = 0; j < num_tiles; j++) {
3660 jlong con = tiles[j];
3661 jlong init = inits[j];
3662 if (con == 0) continue;
3663 jint con0, con1; // split the constant, address-wise
3664 jint init0, init1; // split the init map, address-wise
3665 { union { jlong con; jint intcon[2]; } u;
3666 u.con = con;
3667 con0 = u.intcon[0];
3668 con1 = u.intcon[1];
3669 u.con = init;
3670 init0 = u.intcon[0];
3671 init1 = u.intcon[1];
3672 }
3673
3674 Node* old = nodes[j];
3675 assert(old != NULL, "need the prior store");
3676 intptr_t offset = (j * BytesPerLong);
3677
3678 bool split = !Matcher::isSimpleConstant64(con);
3679
3680 if (offset < header_size) {
3681 assert(offset + BytesPerInt >= header_size, "second int counts");
3682 assert(*(jint*)&tiles[j] == 0, "junk in header");
3683 split = true; // only the second word counts
3684 // Example: int a[] = { 42 ... }
3685 } else if (con0 == 0 && init0 == -1) {
3686 split = true; // first word is covered by full inits
3687 // Example: int a[] = { ... foo(), 42 ... }
3688 } else if (con1 == 0 && init1 == -1) {
3689 split = true; // second word is covered by full inits
3690 // Example: int a[] = { ... 42, foo() ... }
3691 }
3692
3693 // Here's a case where init0 is neither 0 nor -1:
3694 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
3695 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
3696 // In this case the tile is not split; it is (jlong)42.
3697 // The big tile is stored down, and then the foo() value is inserted.
3698 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
3699
3700 Node* ctl = old->in(MemNode::Control);
3701 Node* adr = make_raw_address(offset, phase);
3702 const TypePtr* atp = TypeRawPtr::BOTTOM;
3703
3704 // One or two coalesced stores to plop down.
3705 Node* st[2];
3706 intptr_t off[2];
3707 int nst = 0;
3708 if (!split) {
3709 ++new_long;
3710 off[nst] = offset;
3711 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3712 phase->longcon(con), T_LONG, MemNode::unordered);
3713 } else {
3714 // Omit either if it is a zero.
3715 if (con0 != 0) {
3716 ++new_int;
3717 off[nst] = offset;
3718 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3719 phase->intcon(con0), T_INT, MemNode::unordered);
3720 }
3721 if (con1 != 0) {
3722 ++new_int;
3723 offset += BytesPerInt;
3724 adr = make_raw_address(offset, phase);
3725 off[nst] = offset;
3726 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
3727 phase->intcon(con1), T_INT, MemNode::unordered);
3728 }
3729 }
3730
3731 // Insert second store first, then the first before the second.
3732 // Insert each one just before any overlapping non-constant stores.
3733 while (nst > 0) {
3734 Node* st1 = st[--nst];
3735 C->copy_node_notes_to(st1, old);
3736 st1 = phase->transform(st1);
3737 offset = off[nst];
3738 assert(offset >= header_size, "do not smash header");
3739 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
3740 guarantee(ins_idx != 0, "must re-insert constant store");
3741 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
3742 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
3743 set_req(--ins_idx, st1);
3744 else
3745 ins_req(ins_idx, st1);
3746 }
3747 }
3748
3749 if (PrintCompilation && WizardMode)
3750 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
3751 old_subword, old_long, new_int, new_long);
3752 if (C->log() != NULL)
3753 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
3754 old_subword, old_long, new_int, new_long);
3755
3756 // Clean up any remaining occurrences of zmem:
3757 remove_extra_zeroes();
3758 }
3759
3760 // Explore forward from in(start) to find the first fully initialized
3761 // word, and return its offset. Skip groups of subword stores which
3762 // together initialize full words. If in(start) is itself part of a
3763 // fully initialized word, return the offset of in(start). If there
3764 // are no following full-word stores, or if something is fishy, return
3765 // a negative value.
3766 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
3767 int int_map = 0;
3768 intptr_t int_map_off = 0;
3769 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
3770
3771 for (uint i = start, limit = req(); i < limit; i++) {
3772 Node* st = in(i);
3773
3774 intptr_t st_off = get_store_offset(st, phase);
3775 if (st_off < 0) break; // return conservative answer
3776
3777 int st_size = st->as_Store()->memory_size();
3778 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
3779 return st_off; // we found a complete word init
3780 }
3781
3782 // update the map:
3783
3784 intptr_t this_int_off = align_size_down(st_off, BytesPerInt);
3785 if (this_int_off != int_map_off) {
3786 // reset the map:
3787 int_map = 0;
3788 int_map_off = this_int_off;
3789 }
3790
3791 int subword_off = st_off - this_int_off;
3792 int_map |= right_n_bits(st_size) << subword_off;
3793 if ((int_map & FULL_MAP) == FULL_MAP) {
3794 return this_int_off; // we found a complete word init
3795 }
3796
3797 // Did this store hit or cross the word boundary?
3798 intptr_t next_int_off = align_size_down(st_off + st_size, BytesPerInt);
3799 if (next_int_off == this_int_off + BytesPerInt) {
3800 // We passed the current int, without fully initializing it.
3801 int_map_off = next_int_off;
3802 int_map >>= BytesPerInt;
3803 } else if (next_int_off > this_int_off + BytesPerInt) {
3804 // We passed the current and next int.
3805 return this_int_off + BytesPerInt;
3806 }
3807 }
3808
3809 return -1;
3810 }
3811
3812
3813 // Called when the associated AllocateNode is expanded into CFG.
3814 // At this point, we may perform additional optimizations.
3815 // Linearize the stores by ascending offset, to make memory
3816 // activity as coherent as possible.
3817 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
3818 intptr_t header_size,
3819 Node* size_in_bytes,
3820 PhaseGVN* phase) {
3821 assert(!is_complete(), "not already complete");
3822 assert(stores_are_sane(phase), "");
3823 assert(allocation() != NULL, "must be present");
3824
3825 remove_extra_zeroes();
3826
3827 if (ReduceFieldZeroing || ReduceBulkZeroing)
3828 // reduce instruction count for common initialization patterns
3829 coalesce_subword_stores(header_size, size_in_bytes, phase);
3830
3831 Node* zmem = zero_memory(); // initially zero memory state
3832 Node* inits = zmem; // accumulating a linearized chain of inits
3833 #ifdef ASSERT
3834 intptr_t first_offset = allocation()->minimum_header_size();
3835 intptr_t last_init_off = first_offset; // previous init offset
3836 intptr_t last_init_end = first_offset; // previous init offset+size
3837 intptr_t last_tile_end = first_offset; // previous tile offset+size
3838 #endif
3839 intptr_t zeroes_done = header_size;
3840
3841 bool do_zeroing = true; // we might give up if inits are very sparse
3842 int big_init_gaps = 0; // how many large gaps have we seen?
3843
3844 if (UseTLAB && ZeroTLAB) do_zeroing = false;
3845 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
3846
3847 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3848 Node* st = in(i);
3849 intptr_t st_off = get_store_offset(st, phase);
3850 if (st_off < 0)
3851 break; // unknown junk in the inits
3852 if (st->in(MemNode::Memory) != zmem)
3853 break; // complicated store chains somehow in list
3854
3855 int st_size = st->as_Store()->memory_size();
3856 intptr_t next_init_off = st_off + st_size;
3857
3858 if (do_zeroing && zeroes_done < next_init_off) {
3859 // See if this store needs a zero before it or under it.
3860 intptr_t zeroes_needed = st_off;
3861
3862 if (st_size < BytesPerInt) {
3863 // Look for subword stores which only partially initialize words.
3864 // If we find some, we must lay down some word-level zeroes first,
3865 // underneath the subword stores.
3866 //
3867 // Examples:
3868 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
3869 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
3870 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
3871 //
3872 // Note: coalesce_subword_stores may have already done this,
3873 // if it was prompted by constant non-zero subword initializers.
3874 // But this case can still arise with non-constant stores.
3875
3876 intptr_t next_full_store = find_next_fullword_store(i, phase);
3877
3878 // In the examples above:
3879 // in(i) p q r s x y z
3880 // st_off 12 13 14 15 12 13 14
3881 // st_size 1 1 1 1 1 1 1
3882 // next_full_s. 12 16 16 16 16 16 16
3883 // z's_done 12 16 16 16 12 16 12
3884 // z's_needed 12 16 16 16 16 16 16
3885 // zsize 0 0 0 0 4 0 4
3886 if (next_full_store < 0) {
3887 // Conservative tack: Zero to end of current word.
3888 zeroes_needed = align_size_up(zeroes_needed, BytesPerInt);
3889 } else {
3890 // Zero to beginning of next fully initialized word.
3891 // Or, don't zero at all, if we are already in that word.
3892 assert(next_full_store >= zeroes_needed, "must go forward");
3893 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
3894 zeroes_needed = next_full_store;
3895 }
3896 }
3897
3898 if (zeroes_needed > zeroes_done) {
3899 intptr_t zsize = zeroes_needed - zeroes_done;
3900 // Do some incremental zeroing on rawmem, in parallel with inits.
3901 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3902 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3903 zeroes_done, zeroes_needed,
3904 phase);
3905 zeroes_done = zeroes_needed;
3906 if (zsize > Matcher::init_array_short_size && ++big_init_gaps > 2)
3907 do_zeroing = false; // leave the hole, next time
3908 }
3909 }
3910
3911 // Collect the store and move on:
3912 st->set_req(MemNode::Memory, inits);
3913 inits = st; // put it on the linearized chain
3914 set_req(i, zmem); // unhook from previous position
3915
3916 if (zeroes_done == st_off)
3917 zeroes_done = next_init_off;
3918
3919 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
3920
3921 #ifdef ASSERT
3922 // Various order invariants. Weaker than stores_are_sane because
3923 // a large constant tile can be filled in by smaller non-constant stores.
3924 assert(st_off >= last_init_off, "inits do not reverse");
3925 last_init_off = st_off;
3926 const Type* val = NULL;
3927 if (st_size >= BytesPerInt &&
3928 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
3929 (int)val->basic_type() < (int)T_OBJECT) {
3930 assert(st_off >= last_tile_end, "tiles do not overlap");
3931 assert(st_off >= last_init_end, "tiles do not overwrite inits");
3932 last_tile_end = MAX2(last_tile_end, next_init_off);
3933 } else {
3934 intptr_t st_tile_end = align_size_up(next_init_off, BytesPerLong);
3935 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
3936 assert(st_off >= last_init_end, "inits do not overlap");
3937 last_init_end = next_init_off; // it's a non-tile
3938 }
3939 #endif //ASSERT
3940 }
3941
3942 remove_extra_zeroes(); // clear out all the zmems left over
3943 add_req(inits);
3944
3945 if (!(UseTLAB && ZeroTLAB)) {
3946 // If anything remains to be zeroed, zero it all now.
3947 zeroes_done = align_size_down(zeroes_done, BytesPerInt);
3948 // if it is the last unused 4 bytes of an instance, forget about it
3949 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
3950 if (zeroes_done + BytesPerLong >= size_limit) {
3951 assert(allocation() != NULL, "");
3952 if (allocation()->Opcode() == Op_Allocate) {
3953 Node* klass_node = allocation()->in(AllocateNode::KlassNode);
3954 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
3955 if (zeroes_done == k->layout_helper())
3956 zeroes_done = size_limit;
3957 }
3958 }
3959 if (zeroes_done < size_limit) {
3960 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
3961 zeroes_done, size_in_bytes, phase);
3962 }
3963 }
3964
3965 set_complete(phase);
3966 return rawmem;
3967 }
3968
3969
3970 #ifdef ASSERT
3971 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
3972 if (is_complete())
3973 return true; // stores could be anything at this point
3974 assert(allocation() != NULL, "must be present");
3975 intptr_t last_off = allocation()->minimum_header_size();
3976 for (uint i = InitializeNode::RawStores; i < req(); i++) {
3977 Node* st = in(i);
3978 intptr_t st_off = get_store_offset(st, phase);
3979 if (st_off < 0) continue; // ignore dead garbage
3980 if (last_off > st_off) {
3981 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
3982 this->dump(2);
3983 assert(false, "ascending store offsets");
3984 return false;
3985 }
3986 last_off = st_off + st->as_Store()->memory_size();
3987 }
3988 return true;
3989 }
3990 #endif //ASSERT
3991
3992
3993
3994
3995 //============================MergeMemNode=====================================
3996 //
3997 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
3998 // contributing store or call operations. Each contributor provides the memory
3999 // state for a particular "alias type" (see Compile::alias_type). For example,
4000 // if a MergeMem has an input X for alias category #6, then any memory reference
4001 // to alias category #6 may use X as its memory state input, as an exact equivalent
4002 // to using the MergeMem as a whole.
4003 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4004 //
4005 // (Here, the <N> notation gives the index of the relevant adr_type.)
4006 //
4007 // In one special case (and more cases in the future), alias categories overlap.
4008 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4009 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4010 // it is exactly equivalent to that state W:
4011 // MergeMem(<Bot>: W) <==> W
4012 //
4013 // Usually, the merge has more than one input. In that case, where inputs
4014 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4015 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4016 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4017 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4018 //
4019 // A merge can take a "wide" memory state as one of its narrow inputs.
4020 // This simply means that the merge observes out only the relevant parts of
4021 // the wide input. That is, wide memory states arriving at narrow merge inputs
4022 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4023 //
4024 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4025 // and that memory slices "leak through":
4026 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4027 //
4028 // But, in such a cascade, repeated memory slices can "block the leak":
4029 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4030 //
4031 // In the last example, Y is not part of the combined memory state of the
4032 // outermost MergeMem. The system must, of course, prevent unschedulable
4033 // memory states from arising, so you can be sure that the state Y is somehow
4034 // a precursor to state Y'.
4035 //
4036 //
4037 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4038 // of each MergeMemNode array are exactly the numerical alias indexes, including
4039 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4040 // Compile::alias_type (and kin) produce and manage these indexes.
4041 //
4042 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4043 // (Note that this provides quick access to the top node inside MergeMem methods,
4044 // without the need to reach out via TLS to Compile::current.)
4045 //
4046 // As a consequence of what was just described, a MergeMem that represents a full
4047 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4048 // containing all alias categories.
4049 //
4050 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4051 //
4052 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4053 // a memory state for the alias type <N>, or else the top node, meaning that
4054 // there is no particular input for that alias type. Note that the length of
4055 // a MergeMem is variable, and may be extended at any time to accommodate new
4056 // memory states at larger alias indexes. When merges grow, they are of course
4057 // filled with "top" in the unused in() positions.
4058 //
4059 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4060 // (Top was chosen because it works smoothly with passes like GCM.)
4061 //
4062 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4063 // the type of random VM bits like TLS references.) Since it is always the
4064 // first non-Bot memory slice, some low-level loops use it to initialize an
4065 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4066 //
4067 //
4068 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4069 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4070 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4071 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4072 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4073 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4074 //
4075 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4076 // really that different from the other memory inputs. An abbreviation called
4077 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4078 //
4079 //
4080 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4081 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4082 // that "emerges though" the base memory will be marked as excluding the alias types
4083 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4084 //
4085 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4086 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4087 //
4088 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4089 // (It is currently unimplemented.) As you can see, the resulting merge is
4090 // actually a disjoint union of memory states, rather than an overlay.
4091 //
4092
4093 //------------------------------MergeMemNode-----------------------------------
4094 Node* MergeMemNode::make_empty_memory() {
4095 Node* empty_memory = (Node*) Compile::current()->top();
4096 assert(empty_memory->is_top(), "correct sentinel identity");
4097 return empty_memory;
4098 }
4099
4100 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4101 init_class_id(Class_MergeMem);
4102 // all inputs are nullified in Node::Node(int)
4103 // set_input(0, NULL); // no control input
4104
4105 // Initialize the edges uniformly to top, for starters.
4106 Node* empty_mem = make_empty_memory();
4107 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4108 init_req(i,empty_mem);
4109 }
4110 assert(empty_memory() == empty_mem, "");
4111
4112 if( new_base != NULL && new_base->is_MergeMem() ) {
4113 MergeMemNode* mdef = new_base->as_MergeMem();
4114 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4115 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4116 mms.set_memory(mms.memory2());
4117 }
4118 assert(base_memory() == mdef->base_memory(), "");
4119 } else {
4120 set_base_memory(new_base);
4121 }
4122 }
4123
4124 // Make a new, untransformed MergeMem with the same base as 'mem'.
4125 // If mem is itself a MergeMem, populate the result with the same edges.
4126 MergeMemNode* MergeMemNode::make(Node* mem) {
4127 return new MergeMemNode(mem);
4128 }
4129
4130 //------------------------------cmp--------------------------------------------
4131 uint MergeMemNode::hash() const { return NO_HASH; }
4132 uint MergeMemNode::cmp( const Node &n ) const {
4133 return (&n == this); // Always fail except on self
4134 }
4135
4136 //------------------------------Identity---------------------------------------
4137 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4138 // Identity if this merge point does not record any interesting memory
4139 // disambiguations.
4140 Node* base_mem = base_memory();
4141 Node* empty_mem = empty_memory();
4142 if (base_mem != empty_mem) { // Memory path is not dead?
4143 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4144 Node* mem = in(i);
4145 if (mem != empty_mem && mem != base_mem) {
4146 return this; // Many memory splits; no change
4147 }
4148 }
4149 }
4150 return base_mem; // No memory splits; ID on the one true input
4151 }
4152
4153 //------------------------------Ideal------------------------------------------
4154 // This method is invoked recursively on chains of MergeMem nodes
4155 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4156 // Remove chain'd MergeMems
4157 //
4158 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4159 // relative to the "in(Bot)". Since we are patching both at the same time,
4160 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4161 // but rewrite each "in(i)" relative to the new "in(Bot)".
4162 Node *progress = NULL;
4163
4164
4165 Node* old_base = base_memory();
4166 Node* empty_mem = empty_memory();
4167 if (old_base == empty_mem)
4168 return NULL; // Dead memory path.
4169
4170 MergeMemNode* old_mbase;
4171 if (old_base != NULL && old_base->is_MergeMem())
4172 old_mbase = old_base->as_MergeMem();
4173 else
4174 old_mbase = NULL;
4175 Node* new_base = old_base;
4176
4177 // simplify stacked MergeMems in base memory
4178 if (old_mbase) new_base = old_mbase->base_memory();
4179
4180 // the base memory might contribute new slices beyond my req()
4181 if (old_mbase) grow_to_match(old_mbase);
4182
4183 // Look carefully at the base node if it is a phi.
4184 PhiNode* phi_base;
4185 if (new_base != NULL && new_base->is_Phi())
4186 phi_base = new_base->as_Phi();
4187 else
4188 phi_base = NULL;
4189
4190 Node* phi_reg = NULL;
4191 uint phi_len = (uint)-1;
4192 if (phi_base != NULL && !phi_base->is_copy()) {
4193 // do not examine phi if degraded to a copy
4194 phi_reg = phi_base->region();
4195 phi_len = phi_base->req();
4196 // see if the phi is unfinished
4197 for (uint i = 1; i < phi_len; i++) {
4198 if (phi_base->in(i) == NULL) {
4199 // incomplete phi; do not look at it yet!
4200 phi_reg = NULL;
4201 phi_len = (uint)-1;
4202 break;
4203 }
4204 }
4205 }
4206
4207 // Note: We do not call verify_sparse on entry, because inputs
4208 // can normalize to the base_memory via subsume_node or similar
4209 // mechanisms. This method repairs that damage.
4210
4211 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4212
4213 // Look at each slice.
4214 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4215 Node* old_in = in(i);
4216 // calculate the old memory value
4217 Node* old_mem = old_in;
4218 if (old_mem == empty_mem) old_mem = old_base;
4219 assert(old_mem == memory_at(i), "");
4220
4221 // maybe update (reslice) the old memory value
4222
4223 // simplify stacked MergeMems
4224 Node* new_mem = old_mem;
4225 MergeMemNode* old_mmem;
4226 if (old_mem != NULL && old_mem->is_MergeMem())
4227 old_mmem = old_mem->as_MergeMem();
4228 else
4229 old_mmem = NULL;
4230 if (old_mmem == this) {
4231 // This can happen if loops break up and safepoints disappear.
4232 // A merge of BotPtr (default) with a RawPtr memory derived from a
4233 // safepoint can be rewritten to a merge of the same BotPtr with
4234 // the BotPtr phi coming into the loop. If that phi disappears
4235 // also, we can end up with a self-loop of the mergemem.
4236 // In general, if loops degenerate and memory effects disappear,
4237 // a mergemem can be left looking at itself. This simply means
4238 // that the mergemem's default should be used, since there is
4239 // no longer any apparent effect on this slice.
4240 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4241 // from start. Update the input to TOP.
4242 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4243 }
4244 else if (old_mmem != NULL) {
4245 new_mem = old_mmem->memory_at(i);
4246 }
4247 // else preceding memory was not a MergeMem
4248
4249 // replace equivalent phis (unfortunately, they do not GVN together)
4250 if (new_mem != NULL && new_mem != new_base &&
4251 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4252 if (new_mem->is_Phi()) {
4253 PhiNode* phi_mem = new_mem->as_Phi();
4254 for (uint i = 1; i < phi_len; i++) {
4255 if (phi_base->in(i) != phi_mem->in(i)) {
4256 phi_mem = NULL;
4257 break;
4258 }
4259 }
4260 if (phi_mem != NULL) {
4261 // equivalent phi nodes; revert to the def
4262 new_mem = new_base;
4263 }
4264 }
4265 }
4266
4267 // maybe store down a new value
4268 Node* new_in = new_mem;
4269 if (new_in == new_base) new_in = empty_mem;
4270
4271 if (new_in != old_in) {
4272 // Warning: Do not combine this "if" with the previous "if"
4273 // A memory slice might have be be rewritten even if it is semantically
4274 // unchanged, if the base_memory value has changed.
4275 set_req(i, new_in);
4276 progress = this; // Report progress
4277 }
4278 }
4279
4280 if (new_base != old_base) {
4281 set_req(Compile::AliasIdxBot, new_base);
4282 // Don't use set_base_memory(new_base), because we need to update du.
4283 assert(base_memory() == new_base, "");
4284 progress = this;
4285 }
4286
4287 if( base_memory() == this ) {
4288 // a self cycle indicates this memory path is dead
4289 set_req(Compile::AliasIdxBot, empty_mem);
4290 }
4291
4292 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4293 // Recursion must occur after the self cycle check above
4294 if( base_memory()->is_MergeMem() ) {
4295 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4296 Node *m = phase->transform(new_mbase); // Rollup any cycles
4297 if( m != NULL && (m->is_top() ||
4298 m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem) ) {
4299 // propagate rollup of dead cycle to self
4300 set_req(Compile::AliasIdxBot, empty_mem);
4301 }
4302 }
4303
4304 if( base_memory() == empty_mem ) {
4305 progress = this;
4306 // Cut inputs during Parse phase only.
4307 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4308 if( !can_reshape ) {
4309 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4310 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4311 }
4312 }
4313 }
4314
4315 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4316 // Check if PhiNode::Ideal's "Split phis through memory merges"
4317 // transform should be attempted. Look for this->phi->this cycle.
4318 uint merge_width = req();
4319 if (merge_width > Compile::AliasIdxRaw) {
4320 PhiNode* phi = base_memory()->as_Phi();
4321 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4322 if (phi->in(i) == this) {
4323 phase->is_IterGVN()->_worklist.push(phi);
4324 break;
4325 }
4326 }
4327 }
4328 }
4329
4330 assert(progress || verify_sparse(), "please, no dups of base");
4331 return progress;
4332 }
4333
4334 //-------------------------set_base_memory-------------------------------------
4335 void MergeMemNode::set_base_memory(Node *new_base) {
4336 Node* empty_mem = empty_memory();
4337 set_req(Compile::AliasIdxBot, new_base);
4338 assert(memory_at(req()) == new_base, "must set default memory");
4339 // Clear out other occurrences of new_base:
4340 if (new_base != empty_mem) {
4341 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4342 if (in(i) == new_base) set_req(i, empty_mem);
4343 }
4344 }
4345 }
4346
4347 //------------------------------out_RegMask------------------------------------
4348 const RegMask &MergeMemNode::out_RegMask() const {
4349 return RegMask::Empty;
4350 }
4351
4352 //------------------------------dump_spec--------------------------------------
4353 #ifndef PRODUCT
4354 void MergeMemNode::dump_spec(outputStream *st) const {
4355 st->print(" {");
4356 Node* base_mem = base_memory();
4357 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4358 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4359 if (mem == base_mem) { st->print(" -"); continue; }
4360 st->print( " N%d:", mem->_idx );
4361 Compile::current()->get_adr_type(i)->dump_on(st);
4362 }
4363 st->print(" }");
4364 }
4365 #endif // !PRODUCT
4366
4367
4368 #ifdef ASSERT
4369 static bool might_be_same(Node* a, Node* b) {
4370 if (a == b) return true;
4371 if (!(a->is_Phi() || b->is_Phi())) return false;
4372 // phis shift around during optimization
4373 return true; // pretty stupid...
4374 }
4375
4376 // verify a narrow slice (either incoming or outgoing)
4377 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4378 if (!VerifyAliases) return; // don't bother to verify unless requested
4379 if (is_error_reported()) return; // muzzle asserts when debugging an error
4380 if (Node::in_dump()) return; // muzzle asserts when printing
4381 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4382 assert(n != NULL, "");
4383 // Elide intervening MergeMem's
4384 while (n->is_MergeMem()) {
4385 n = n->as_MergeMem()->memory_at(alias_idx);
4386 }
4387 Compile* C = Compile::current();
4388 const TypePtr* n_adr_type = n->adr_type();
4389 if (n == m->empty_memory()) {
4390 // Implicit copy of base_memory()
4391 } else if (n_adr_type != TypePtr::BOTTOM) {
4392 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4393 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4394 } else {
4395 // A few places like make_runtime_call "know" that VM calls are narrow,
4396 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4397 bool expected_wide_mem = false;
4398 if (n == m->base_memory()) {
4399 expected_wide_mem = true;
4400 } else if (alias_idx == Compile::AliasIdxRaw ||
4401 n == m->memory_at(Compile::AliasIdxRaw)) {
4402 expected_wide_mem = true;
4403 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4404 // memory can "leak through" calls on channels that
4405 // are write-once. Allow this also.
4406 expected_wide_mem = true;
4407 }
4408 assert(expected_wide_mem, "expected narrow slice replacement");
4409 }
4410 }
4411 #else // !ASSERT
4412 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4413 #endif
4414
4415
4416 //-----------------------------memory_at---------------------------------------
4417 Node* MergeMemNode::memory_at(uint alias_idx) const {
4418 assert(alias_idx >= Compile::AliasIdxRaw ||
4419 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4420 "must avoid base_memory and AliasIdxTop");
4421
4422 // Otherwise, it is a narrow slice.
4423 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4424 Compile *C = Compile::current();
4425 if (is_empty_memory(n)) {
4426 // the array is sparse; empty slots are the "top" node
4427 n = base_memory();
4428 assert(Node::in_dump()
4429 || n == NULL || n->bottom_type() == Type::TOP
4430 || n->adr_type() == NULL // address is TOP
4431 || n->adr_type() == TypePtr::BOTTOM
4432 || n->adr_type() == TypeRawPtr::BOTTOM
4433 || Compile::current()->AliasLevel() == 0,
4434 "must be a wide memory");
4435 // AliasLevel == 0 if we are organizing the memory states manually.
4436 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4437 } else {
4438 // make sure the stored slice is sane
4439 #ifdef ASSERT
4440 if (is_error_reported() || Node::in_dump()) {
4441 } else if (might_be_same(n, base_memory())) {
4442 // Give it a pass: It is a mostly harmless repetition of the base.
4443 // This can arise normally from node subsumption during optimization.
4444 } else {
4445 verify_memory_slice(this, alias_idx, n);
4446 }
4447 #endif
4448 }
4449 return n;
4450 }
4451
4452 //---------------------------set_memory_at-------------------------------------
4453 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4454 verify_memory_slice(this, alias_idx, n);
4455 Node* empty_mem = empty_memory();
4456 if (n == base_memory()) n = empty_mem; // collapse default
4457 uint need_req = alias_idx+1;
4458 if (req() < need_req) {
4459 if (n == empty_mem) return; // already the default, so do not grow me
4460 // grow the sparse array
4461 do {
4462 add_req(empty_mem);
4463 } while (req() < need_req);
4464 }
4465 set_req( alias_idx, n );
4466 }
4467
4468
4469
4470 //--------------------------iteration_setup------------------------------------
4471 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4472 if (other != NULL) {
4473 grow_to_match(other);
4474 // invariant: the finite support of mm2 is within mm->req()
4475 #ifdef ASSERT
4476 for (uint i = req(); i < other->req(); i++) {
4477 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4478 }
4479 #endif
4480 }
4481 // Replace spurious copies of base_memory by top.
4482 Node* base_mem = base_memory();
4483 if (base_mem != NULL && !base_mem->is_top()) {
4484 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4485 if (in(i) == base_mem)
4486 set_req(i, empty_memory());
4487 }
4488 }
4489 }
4490
4491 //---------------------------grow_to_match-------------------------------------
4492 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4493 Node* empty_mem = empty_memory();
4494 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4495 // look for the finite support of the other memory
4496 for (uint i = other->req(); --i >= req(); ) {
4497 if (other->in(i) != empty_mem) {
4498 uint new_len = i+1;
4499 while (req() < new_len) add_req(empty_mem);
4500 break;
4501 }
4502 }
4503 }
4504
4505 //---------------------------verify_sparse-------------------------------------
4506 #ifndef PRODUCT
4507 bool MergeMemNode::verify_sparse() const {
4508 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4509 Node* base_mem = base_memory();
4510 // The following can happen in degenerate cases, since empty==top.
4511 if (is_empty_memory(base_mem)) return true;
4512 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4513 assert(in(i) != NULL, "sane slice");
4514 if (in(i) == base_mem) return false; // should have been the sentinel value!
4515 }
4516 return true;
4517 }
4518
4519 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4520 Node* n;
4521 n = mm->in(idx);
4522 if (mem == n) return true; // might be empty_memory()
4523 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4524 if (mem == n) return true;
4525 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4526 if (mem == n) return true;
4527 if (n == NULL) break;
4528 }
4529 return false;
4530 }
4531 #endif // !PRODUCT