1 /*
2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "memory/allocation.inline.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/connode.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/loopnode.hpp"
40 #include "opto/machnode.hpp"
41 #include "opto/matcher.hpp"
42 #include "opto/memnode.hpp"
43 #include "opto/mulnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/phaseX.hpp"
46 #include "opto/regmask.hpp"
47 #include "opto/valuetypenode.hpp"
48 #include "utilities/align.hpp"
49 #include "utilities/copy.hpp"
50 #include "utilities/macros.hpp"
51 #include "utilities/vmError.hpp"
52 #if INCLUDE_ZGC
53 #include "gc/z/c2/zBarrierSetC2.hpp"
54 #endif
55
56 // Portions of code courtesy of Clifford Click
57
58 // Optimization - Graph Style
59
60 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
61
62 //=============================================================================
63 uint MemNode::size_of() const { return sizeof(*this); }
64
65 const TypePtr *MemNode::adr_type() const {
66 Node* adr = in(Address);
67 if (adr == NULL) return NULL; // node is dead
68 const TypePtr* cross_check = NULL;
69 DEBUG_ONLY(cross_check = _adr_type);
70 return calculate_adr_type(adr->bottom_type(), cross_check);
71 }
72
73 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
74 if (adr != NULL) {
75 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
76 return true;
77 }
78 }
79 return false;
80 }
81
82 #ifndef PRODUCT
83 void MemNode::dump_spec(outputStream *st) const {
84 if (in(Address) == NULL) return; // node is dead
85 #ifndef ASSERT
86 // fake the missing field
87 const TypePtr* _adr_type = NULL;
88 if (in(Address) != NULL)
89 _adr_type = in(Address)->bottom_type()->isa_ptr();
90 #endif
91 dump_adr_type(this, _adr_type, st);
92
93 Compile* C = Compile::current();
94 if (C->alias_type(_adr_type)->is_volatile()) {
95 st->print(" Volatile!");
96 }
97 if (_unaligned_access) {
98 st->print(" unaligned");
99 }
100 if (_mismatched_access) {
101 st->print(" mismatched");
102 }
103 }
104
105 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
106 st->print(" @");
107 if (adr_type == NULL) {
108 st->print("NULL");
109 } else {
110 adr_type->dump_on(st);
111 Compile* C = Compile::current();
112 Compile::AliasType* atp = NULL;
113 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
114 if (atp == NULL)
115 st->print(", idx=?\?;");
116 else if (atp->index() == Compile::AliasIdxBot)
117 st->print(", idx=Bot;");
118 else if (atp->index() == Compile::AliasIdxTop)
119 st->print(", idx=Top;");
120 else if (atp->index() == Compile::AliasIdxRaw)
121 st->print(", idx=Raw;");
122 else {
123 ciField* field = atp->field();
124 if (field) {
125 st->print(", name=");
126 field->print_name_on(st);
127 }
128 st->print(", idx=%d;", atp->index());
129 }
130 }
131 }
132
133 extern void print_alias_types();
134
135 #endif
136
137 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
138 assert((t_oop != NULL), "sanity");
139 bool is_instance = t_oop->is_known_instance_field();
140 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
141 (load != NULL) && load->is_Load() &&
142 (phase->is_IterGVN() != NULL);
143 if (!(is_instance || is_boxed_value_load))
144 return mchain; // don't try to optimize non-instance types
145 uint instance_id = t_oop->instance_id();
146 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
147 Node *prev = NULL;
148 Node *result = mchain;
149 while (prev != result) {
150 prev = result;
151 if (result == start_mem)
152 break; // hit one of our sentinels
153 // skip over a call which does not affect this memory slice
154 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
155 Node *proj_in = result->in(0);
156 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
157 break; // hit one of our sentinels
158 } else if (proj_in->is_Call()) {
159 // ArrayCopyNodes processed here as well
160 CallNode *call = proj_in->as_Call();
161 if (!call->may_modify(t_oop, phase)) { // returns false for instances
162 result = call->in(TypeFunc::Memory);
163 }
164 } else if (proj_in->is_Initialize()) {
165 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
166 // Stop if this is the initialization for the object instance which
167 // contains this memory slice, otherwise skip over it.
168 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
169 break;
170 }
171 if (is_instance) {
172 result = proj_in->in(TypeFunc::Memory);
173 } else if (is_boxed_value_load) {
174 Node* klass = alloc->in(AllocateNode::KlassNode);
175 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
176 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
177 result = proj_in->in(TypeFunc::Memory); // not related allocation
178 }
179 }
180 } else if (proj_in->is_MemBar()) {
181 ArrayCopyNode* ac = NULL;
182 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
183 break;
184 }
185 result = proj_in->in(TypeFunc::Memory);
186 } else {
187 assert(false, "unexpected projection");
188 }
189 } else if (result->is_ClearArray()) {
190 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
191 // Can not bypass initialization of the instance
192 // we are looking for.
193 break;
194 }
195 // Otherwise skip it (the call updated 'result' value).
196 } else if (result->is_MergeMem()) {
197 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
198 }
199 }
200 return result;
201 }
202
203 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
204 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
205 if (t_oop == NULL)
206 return mchain; // don't try to optimize non-oop types
207 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
208 bool is_instance = t_oop->is_known_instance_field();
209 PhaseIterGVN *igvn = phase->is_IterGVN();
210 if (is_instance && igvn != NULL && result->is_Phi()) {
211 PhiNode *mphi = result->as_Phi();
212 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
213 const TypePtr *t = mphi->adr_type();
214 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
215 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
216 t->is_oopptr()->cast_to_exactness(true)
217 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
218 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) {
219 // clone the Phi with our address type
220 result = mphi->split_out_instance(t_adr, igvn);
221 } else {
222 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
223 }
224 }
225 return result;
226 }
227
228 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
229 uint alias_idx = phase->C->get_alias_index(tp);
230 Node *mem = mmem;
231 #ifdef ASSERT
232 {
233 // Check that current type is consistent with the alias index used during graph construction
234 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
235 bool consistent = adr_check == NULL || adr_check->empty() ||
236 phase->C->must_alias(adr_check, alias_idx );
237 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
238 if( !consistent && adr_check != NULL && !adr_check->empty() &&
239 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
240 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
241 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
242 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
243 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
244 // don't assert if it is dead code.
245 consistent = true;
246 }
247 if( !consistent ) {
248 st->print("alias_idx==%d, adr_check==", alias_idx);
249 if( adr_check == NULL ) {
250 st->print("NULL");
251 } else {
252 adr_check->dump();
253 }
254 st->cr();
255 print_alias_types();
256 assert(consistent, "adr_check must match alias idx");
257 }
258 }
259 #endif
260 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
261 // means an array I have not precisely typed yet. Do not do any
262 // alias stuff with it any time soon.
263 const TypeOopPtr *toop = tp->isa_oopptr();
264 if( tp->base() != Type::AnyPtr &&
265 !(toop &&
266 toop->klass() != NULL &&
267 toop->klass()->is_java_lang_Object() &&
268 toop->offset() == Type::OffsetBot) ) {
269 // compress paths and change unreachable cycles to TOP
270 // If not, we can update the input infinitely along a MergeMem cycle
271 // Equivalent code in PhiNode::Ideal
272 Node* m = phase->transform(mmem);
273 // If transformed to a MergeMem, get the desired slice
274 // Otherwise the returned node represents memory for every slice
275 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
276 // Update input if it is progress over what we have now
277 }
278 return mem;
279 }
280
281 //--------------------------Ideal_common---------------------------------------
282 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
283 // Unhook non-raw memories from complete (macro-expanded) initializations.
284 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
285 // If our control input is a dead region, kill all below the region
286 Node *ctl = in(MemNode::Control);
287 if (ctl && remove_dead_region(phase, can_reshape))
288 return this;
289 ctl = in(MemNode::Control);
290 // Don't bother trying to transform a dead node
291 if (ctl && ctl->is_top()) return NodeSentinel;
292
293 PhaseIterGVN *igvn = phase->is_IterGVN();
294 // Wait if control on the worklist.
295 if (ctl && can_reshape && igvn != NULL) {
296 Node* bol = NULL;
297 Node* cmp = NULL;
298 if (ctl->in(0)->is_If()) {
299 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
300 bol = ctl->in(0)->in(1);
301 if (bol->is_Bool())
302 cmp = ctl->in(0)->in(1)->in(1);
303 }
304 if (igvn->_worklist.member(ctl) ||
305 (bol != NULL && igvn->_worklist.member(bol)) ||
306 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
307 // This control path may be dead.
308 // Delay this memory node transformation until the control is processed.
309 phase->is_IterGVN()->_worklist.push(this);
310 return NodeSentinel; // caller will return NULL
311 }
312 }
313 // Ignore if memory is dead, or self-loop
314 Node *mem = in(MemNode::Memory);
315 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
316 assert(mem != this, "dead loop in MemNode::Ideal");
317
318 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
319 // This memory slice may be dead.
320 // Delay this mem node transformation until the memory is processed.
321 phase->is_IterGVN()->_worklist.push(this);
322 return NodeSentinel; // caller will return NULL
323 }
324
325 Node *address = in(MemNode::Address);
326 const Type *t_adr = phase->type(address);
327 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
328
329 if (can_reshape && igvn != NULL &&
330 (igvn->_worklist.member(address) ||
331 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
332 // The address's base and type may change when the address is processed.
333 // Delay this mem node transformation until the address is processed.
334 phase->is_IterGVN()->_worklist.push(this);
335 return NodeSentinel; // caller will return NULL
336 }
337
338 // Do NOT remove or optimize the next lines: ensure a new alias index
339 // is allocated for an oop pointer type before Escape Analysis.
340 // Note: C++ will not remove it since the call has side effect.
341 if (t_adr->isa_oopptr()) {
342 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
343 }
344
345 Node* base = NULL;
346 if (address->is_AddP()) {
347 base = address->in(AddPNode::Base);
348 }
349 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
350 !t_adr->isa_rawptr()) {
351 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
352 // Skip this node optimization if its address has TOP base.
353 return NodeSentinel; // caller will return NULL
354 }
355
356 // Avoid independent memory operations
357 Node* old_mem = mem;
358
359 // The code which unhooks non-raw memories from complete (macro-expanded)
360 // initializations was removed. After macro-expansion all stores catched
361 // by Initialize node became raw stores and there is no information
362 // which memory slices they modify. So it is unsafe to move any memory
363 // operation above these stores. Also in most cases hooked non-raw memories
364 // were already unhooked by using information from detect_ptr_independence()
365 // and find_previous_store().
366
367 if (mem->is_MergeMem()) {
368 MergeMemNode* mmem = mem->as_MergeMem();
369 const TypePtr *tp = t_adr->is_ptr();
370
371 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
372 }
373
374 if (mem != old_mem) {
375 set_req(MemNode::Memory, mem);
376 if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) {
377 igvn->_worklist.push(old_mem);
378 }
379 if (phase->type(mem) == Type::TOP) return NodeSentinel;
380 return this;
381 }
382
383 // let the subclass continue analyzing...
384 return NULL;
385 }
386
387 // Helper function for proving some simple control dominations.
388 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
389 // Already assumes that 'dom' is available at 'sub', and that 'sub'
390 // is not a constant (dominated by the method's StartNode).
391 // Used by MemNode::find_previous_store to prove that the
392 // control input of a memory operation predates (dominates)
393 // an allocation it wants to look past.
394 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
395 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
396 return false; // Conservative answer for dead code
397
398 // Check 'dom'. Skip Proj and CatchProj nodes.
399 dom = dom->find_exact_control(dom);
400 if (dom == NULL || dom->is_top())
401 return false; // Conservative answer for dead code
402
403 if (dom == sub) {
404 // For the case when, for example, 'sub' is Initialize and the original
405 // 'dom' is Proj node of the 'sub'.
406 return false;
407 }
408
409 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
410 return true;
411
412 // 'dom' dominates 'sub' if its control edge and control edges
413 // of all its inputs dominate or equal to sub's control edge.
414
415 // Currently 'sub' is either Allocate, Initialize or Start nodes.
416 // Or Region for the check in LoadNode::Ideal();
417 // 'sub' should have sub->in(0) != NULL.
418 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
419 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
420
421 // Get control edge of 'sub'.
422 Node* orig_sub = sub;
423 sub = sub->find_exact_control(sub->in(0));
424 if (sub == NULL || sub->is_top())
425 return false; // Conservative answer for dead code
426
427 assert(sub->is_CFG(), "expecting control");
428
429 if (sub == dom)
430 return true;
431
432 if (sub->is_Start() || sub->is_Root())
433 return false;
434
435 {
436 // Check all control edges of 'dom'.
437
438 ResourceMark rm;
439 Arena* arena = Thread::current()->resource_area();
440 Node_List nlist(arena);
441 Unique_Node_List dom_list(arena);
442
443 dom_list.push(dom);
444 bool only_dominating_controls = false;
445
446 for (uint next = 0; next < dom_list.size(); next++) {
447 Node* n = dom_list.at(next);
448 if (n == orig_sub)
449 return false; // One of dom's inputs dominated by sub.
450 if (!n->is_CFG() && n->pinned()) {
451 // Check only own control edge for pinned non-control nodes.
452 n = n->find_exact_control(n->in(0));
453 if (n == NULL || n->is_top())
454 return false; // Conservative answer for dead code
455 assert(n->is_CFG(), "expecting control");
456 dom_list.push(n);
457 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
458 only_dominating_controls = true;
459 } else if (n->is_CFG()) {
460 if (n->dominates(sub, nlist))
461 only_dominating_controls = true;
462 else
463 return false;
464 } else {
465 // First, own control edge.
466 Node* m = n->find_exact_control(n->in(0));
467 if (m != NULL) {
468 if (m->is_top())
469 return false; // Conservative answer for dead code
470 dom_list.push(m);
471 }
472 // Now, the rest of edges.
473 uint cnt = n->req();
474 for (uint i = 1; i < cnt; i++) {
475 m = n->find_exact_control(n->in(i));
476 if (m == NULL || m->is_top())
477 continue;
478 dom_list.push(m);
479 }
480 }
481 }
482 return only_dominating_controls;
483 }
484 }
485
486 //---------------------detect_ptr_independence---------------------------------
487 // Used by MemNode::find_previous_store to prove that two base
488 // pointers are never equal.
489 // The pointers are accompanied by their associated allocations,
490 // if any, which have been previously discovered by the caller.
491 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
492 Node* p2, AllocateNode* a2,
493 PhaseTransform* phase) {
494 // Attempt to prove that these two pointers cannot be aliased.
495 // They may both manifestly be allocations, and they should differ.
496 // Or, if they are not both allocations, they can be distinct constants.
497 // Otherwise, one is an allocation and the other a pre-existing value.
498 if (a1 == NULL && a2 == NULL) { // neither an allocation
499 return (p1 != p2) && p1->is_Con() && p2->is_Con();
500 } else if (a1 != NULL && a2 != NULL) { // both allocations
501 return (a1 != a2);
502 } else if (a1 != NULL) { // one allocation a1
503 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
504 return all_controls_dominate(p2, a1);
505 } else { //(a2 != NULL) // one allocation a2
506 return all_controls_dominate(p1, a2);
507 }
508 return false;
509 }
510
511
512 // Find an arraycopy that must have set (can_see_stored_value=true) or
513 // could have set (can_see_stored_value=false) the value for this load
514 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
515 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
516 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
517 Node* mb = mem->in(0);
518 if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
519 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
520 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy();
521 if (ac->is_clonebasic()) {
522 intptr_t offset;
523 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset);
524 if (alloc != NULL && alloc == ld_alloc) {
525 return ac;
526 }
527 }
528 }
529 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
530 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
531
532 if (ac->is_arraycopy_validated() ||
533 ac->is_copyof_validated() ||
534 ac->is_copyofrange_validated()) {
535 Node* ld_addp = in(MemNode::Address);
536 if (ld_addp->is_AddP()) {
537 Node* ld_base = ld_addp->in(AddPNode::Address);
538 Node* ld_offs = ld_addp->in(AddPNode::Offset);
539
540 Node* dest = ac->in(ArrayCopyNode::Dest);
541
542 if (dest == ld_base) {
543 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
544 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
545 return ac;
546 }
547 if (!can_see_stored_value) {
548 mem = ac->in(TypeFunc::Memory);
549 }
550 }
551 }
552 }
553 }
554 return NULL;
555 }
556
557 // The logic for reordering loads and stores uses four steps:
558 // (a) Walk carefully past stores and initializations which we
559 // can prove are independent of this load.
560 // (b) Observe that the next memory state makes an exact match
561 // with self (load or store), and locate the relevant store.
562 // (c) Ensure that, if we were to wire self directly to the store,
563 // the optimizer would fold it up somehow.
564 // (d) Do the rewiring, and return, depending on some other part of
565 // the optimizer to fold up the load.
566 // This routine handles steps (a) and (b). Steps (c) and (d) are
567 // specific to loads and stores, so they are handled by the callers.
568 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
569 //
570 Node* MemNode::find_previous_store(PhaseTransform* phase) {
571 Node* ctrl = in(MemNode::Control);
572 Node* adr = in(MemNode::Address);
573 intptr_t offset = 0;
574 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
575 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
576
577 if (offset == Type::OffsetBot)
578 return NULL; // cannot unalias unless there are precise offsets
579
580 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
581 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
582
583 intptr_t size_in_bytes = memory_size();
584
585 Node* mem = in(MemNode::Memory); // start searching here...
586
587 int cnt = 50; // Cycle limiter
588 for (;;) { // While we can dance past unrelated stores...
589 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
590
591 Node* prev = mem;
592 if (mem->is_Store()) {
593 Node* st_adr = mem->in(MemNode::Address);
594 intptr_t st_offset = 0;
595 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
596 if (st_base == NULL)
597 break; // inscrutable pointer
598
599 // For raw accesses it's not enough to prove that constant offsets don't intersect.
600 // We need the bases to be the equal in order for the offset check to make sense.
601 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
602 break;
603 }
604
605 if (st_offset != offset && st_offset != Type::OffsetBot) {
606 const int MAX_STORE = BytesPerLong;
607 if (st_offset >= offset + size_in_bytes ||
608 st_offset <= offset - MAX_STORE ||
609 st_offset <= offset - mem->as_Store()->memory_size()) {
610 // Success: The offsets are provably independent.
611 // (You may ask, why not just test st_offset != offset and be done?
612 // The answer is that stores of different sizes can co-exist
613 // in the same sequence of RawMem effects. We sometimes initialize
614 // a whole 'tile' of array elements with a single jint or jlong.)
615 mem = mem->in(MemNode::Memory);
616 continue; // (a) advance through independent store memory
617 }
618 }
619 if (st_base != base &&
620 detect_ptr_independence(base, alloc,
621 st_base,
622 AllocateNode::Ideal_allocation(st_base, phase),
623 phase)) {
624 // Success: The bases are provably independent.
625 mem = mem->in(MemNode::Memory);
626 continue; // (a) advance through independent store memory
627 }
628
629 // (b) At this point, if the bases or offsets do not agree, we lose,
630 // since we have not managed to prove 'this' and 'mem' independent.
631 if (st_base == base && st_offset == offset) {
632 return mem; // let caller handle steps (c), (d)
633 }
634
635 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
636 InitializeNode* st_init = mem->in(0)->as_Initialize();
637 AllocateNode* st_alloc = st_init->allocation();
638 if (st_alloc == NULL)
639 break; // something degenerated
640 bool known_identical = false;
641 bool known_independent = false;
642 if (alloc == st_alloc)
643 known_identical = true;
644 else if (alloc != NULL)
645 known_independent = true;
646 else if (all_controls_dominate(this, st_alloc))
647 known_independent = true;
648
649 if (known_independent) {
650 // The bases are provably independent: Either they are
651 // manifestly distinct allocations, or else the control
652 // of this load dominates the store's allocation.
653 int alias_idx = phase->C->get_alias_index(adr_type());
654 if (alias_idx == Compile::AliasIdxRaw) {
655 mem = st_alloc->in(TypeFunc::Memory);
656 } else {
657 mem = st_init->memory(alias_idx);
658 }
659 continue; // (a) advance through independent store memory
660 }
661
662 // (b) at this point, if we are not looking at a store initializing
663 // the same allocation we are loading from, we lose.
664 if (known_identical) {
665 // From caller, can_see_stored_value will consult find_captured_store.
666 return mem; // let caller handle steps (c), (d)
667 }
668
669 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
670 if (prev != mem) {
671 // Found an arraycopy but it doesn't affect that load
672 continue;
673 }
674 // Found an arraycopy that may affect that load
675 return mem;
676 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
677 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
678 if (mem->is_Proj() && mem->in(0)->is_Call()) {
679 // ArrayCopyNodes processed here as well.
680 CallNode *call = mem->in(0)->as_Call();
681 if (!call->may_modify(addr_t, phase)) {
682 mem = call->in(TypeFunc::Memory);
683 continue; // (a) advance through independent call memory
684 }
685 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
686 ArrayCopyNode* ac = NULL;
687 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
688 break;
689 }
690 mem = mem->in(0)->in(TypeFunc::Memory);
691 continue; // (a) advance through independent MemBar memory
692 } else if (mem->is_ClearArray()) {
693 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
694 // (the call updated 'mem' value)
695 continue; // (a) advance through independent allocation memory
696 } else {
697 // Can not bypass initialization of the instance
698 // we are looking for.
699 return mem;
700 }
701 } else if (mem->is_MergeMem()) {
702 int alias_idx = phase->C->get_alias_index(adr_type());
703 mem = mem->as_MergeMem()->memory_at(alias_idx);
704 continue; // (a) advance through independent MergeMem memory
705 }
706 }
707
708 // Unless there is an explicit 'continue', we must bail out here,
709 // because 'mem' is an inscrutable memory state (e.g., a call).
710 break;
711 }
712
713 return NULL; // bail out
714 }
715
716 //----------------------calculate_adr_type-------------------------------------
717 // Helper function. Notices when the given type of address hits top or bottom.
718 // Also, asserts a cross-check of the type against the expected address type.
719 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
720 if (t == Type::TOP) return NULL; // does not touch memory any more?
721 #ifdef PRODUCT
722 cross_check = NULL;
723 #else
724 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL;
725 #endif
726 const TypePtr* tp = t->isa_ptr();
727 if (tp == NULL) {
728 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
729 return TypePtr::BOTTOM; // touches lots of memory
730 } else {
731 #ifdef ASSERT
732 // %%%% [phh] We don't check the alias index if cross_check is
733 // TypeRawPtr::BOTTOM. Needs to be investigated.
734 if (cross_check != NULL &&
735 cross_check != TypePtr::BOTTOM &&
736 cross_check != TypeRawPtr::BOTTOM) {
737 // Recheck the alias index, to see if it has changed (due to a bug).
738 Compile* C = Compile::current();
739 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
740 "must stay in the original alias category");
741 // The type of the address must be contained in the adr_type,
742 // disregarding "null"-ness.
743 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
744 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
745 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
746 "real address must not escape from expected memory type");
747 }
748 #endif
749 return tp;
750 }
751 }
752
753 //=============================================================================
754 // Should LoadNode::Ideal() attempt to remove control edges?
755 bool LoadNode::can_remove_control() const {
756 return true;
757 }
758 uint LoadNode::size_of() const { return sizeof(*this); }
759 uint LoadNode::cmp( const Node &n ) const
760 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
761 const Type *LoadNode::bottom_type() const { return _type; }
762 uint LoadNode::ideal_reg() const {
763 return _type->ideal_reg();
764 }
765
766 #ifndef PRODUCT
767 void LoadNode::dump_spec(outputStream *st) const {
768 MemNode::dump_spec(st);
769 if( !Verbose && !WizardMode ) {
770 // standard dump does this in Verbose and WizardMode
771 st->print(" #"); _type->dump_on(st);
772 }
773 if (!depends_only_on_test()) {
774 st->print(" (does not depend only on test)");
775 }
776 }
777 #endif
778
779 #ifdef ASSERT
780 //----------------------------is_immutable_value-------------------------------
781 // Helper function to allow a raw load without control edge for some cases
782 bool LoadNode::is_immutable_value(Node* adr) {
783 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
784 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
785 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
786 in_bytes(JavaThread::osthread_offset())));
787 }
788 #endif
789
790 //----------------------------LoadNode::make-----------------------------------
791 // Polymorphic factory method:
792 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo,
793 ControlDependency control_dependency, bool unaligned, bool mismatched) {
794 Compile* C = gvn.C;
795
796 // sanity check the alias category against the created node type
797 assert(!(adr_type->isa_oopptr() &&
798 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
799 "use LoadKlassNode instead");
800 assert(!(adr_type->isa_aryptr() &&
801 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
802 "use LoadRangeNode instead");
803 // Check control edge of raw loads
804 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
805 // oop will be recorded in oop map if load crosses safepoint
806 rt->isa_oopptr() || is_immutable_value(adr),
807 "raw memory operations should have control edge");
808 LoadNode* load = NULL;
809 switch (bt) {
810 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
811 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
812 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
813 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
814 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
815 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break;
816 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
817 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
818 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
819 case T_VALUETYPE:
820 case T_OBJECT:
821 #ifdef _LP64
822 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
823 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
824 } else
825 #endif
826 {
827 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
828 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
829 }
830 break;
831 default:
832 ShouldNotReachHere();
833 break;
834 }
835 assert(load != NULL, "LoadNode should have been created");
836 if (unaligned) {
837 load->set_unaligned_access();
838 }
839 if (mismatched) {
840 load->set_mismatched_access();
841 }
842 if (load->Opcode() == Op_LoadN) {
843 Node* ld = gvn.transform(load);
844 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
845 }
846
847 return load;
848 }
849
850 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
851 ControlDependency control_dependency, bool unaligned, bool mismatched) {
852 bool require_atomic = true;
853 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic);
854 if (unaligned) {
855 load->set_unaligned_access();
856 }
857 if (mismatched) {
858 load->set_mismatched_access();
859 }
860 return load;
861 }
862
863 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
864 ControlDependency control_dependency, bool unaligned, bool mismatched) {
865 bool require_atomic = true;
866 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic);
867 if (unaligned) {
868 load->set_unaligned_access();
869 }
870 if (mismatched) {
871 load->set_mismatched_access();
872 }
873 return load;
874 }
875
876
877
878 //------------------------------hash-------------------------------------------
879 uint LoadNode::hash() const {
880 // unroll addition of interesting fields
881 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
882 }
883
884 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
885 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
886 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
887 bool is_stable_ary = FoldStableValues &&
888 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
889 tp->isa_aryptr()->is_stable();
890
891 return (eliminate_boxing && non_volatile) || is_stable_ary;
892 }
893
894 return false;
895 }
896
897 // Is the value loaded previously stored by an arraycopy? If so return
898 // a load node that reads from the source array so we may be able to
899 // optimize out the ArrayCopy node later.
900 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
901 #if INCLUDE_ZGC
902 if (UseZGC) {
903 if (bottom_type()->make_oopptr() != NULL) {
904 return NULL;
905 }
906 }
907 #endif
908
909 Node* ld_adr = in(MemNode::Address);
910 intptr_t ld_off = 0;
911 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
912 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
913 if (ac != NULL) {
914 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
915
916 Node* mem = ac->in(TypeFunc::Memory);
917 Node* ctl = ac->in(0);
918 Node* src = ac->in(ArrayCopyNode::Src);
919
920 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
921 return NULL;
922 }
923
924 LoadNode* ld = clone()->as_Load();
925 Node* addp = in(MemNode::Address)->clone();
926 if (ac->as_ArrayCopy()->is_clonebasic()) {
927 assert(ld_alloc != NULL, "need an alloc");
928 assert(addp->is_AddP(), "address must be addp");
929 assert(addp->in(AddPNode::Base) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base), "strange pattern");
930 assert(addp->in(AddPNode::Address) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address), "strange pattern");
931 addp->set_req(AddPNode::Base, src->in(AddPNode::Base));
932 addp->set_req(AddPNode::Address, src->in(AddPNode::Address));
933 } else {
934 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
935 ac->as_ArrayCopy()->is_copyof_validated() ||
936 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
937 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
938 addp->set_req(AddPNode::Base, src);
939 addp->set_req(AddPNode::Address, src);
940
941 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
942 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type();
943 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
944 uint shift = exact_log2(type2aelembytes(ary_elem));
945
946 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
947 #ifdef _LP64
948 diff = phase->transform(new ConvI2LNode(diff));
949 #endif
950 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
951
952 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
953 addp->set_req(AddPNode::Offset, offset);
954 }
955 addp = phase->transform(addp);
956 #ifdef ASSERT
957 const TypePtr* adr_type = phase->type(addp)->is_ptr();
958 ld->_adr_type = adr_type;
959 #endif
960 ld->set_req(MemNode::Address, addp);
961 ld->set_req(0, ctl);
962 ld->set_req(MemNode::Memory, mem);
963 // load depends on the tests that validate the arraycopy
964 ld->_control_dependency = Pinned;
965 return ld;
966 }
967 return NULL;
968 }
969
970
971 //---------------------------can_see_stored_value------------------------------
972 // This routine exists to make sure this set of tests is done the same
973 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
974 // will change the graph shape in a way which makes memory alive twice at the
975 // same time (uses the Oracle model of aliasing), then some
976 // LoadXNode::Identity will fold things back to the equivalence-class model
977 // of aliasing.
978 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
979 Node* ld_adr = in(MemNode::Address);
980 intptr_t ld_off = 0;
981 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
982 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
983 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
984 // This is more general than load from boxing objects.
985 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
986 uint alias_idx = atp->index();
987 bool final = !atp->is_rewritable();
988 Node* result = NULL;
989 Node* current = st;
990 // Skip through chains of MemBarNodes checking the MergeMems for
991 // new states for the slice of this load. Stop once any other
992 // kind of node is encountered. Loads from final memory can skip
993 // through any kind of MemBar but normal loads shouldn't skip
994 // through MemBarAcquire since the could allow them to move out of
995 // a synchronized region.
996 while (current->is_Proj()) {
997 int opc = current->in(0)->Opcode();
998 if ((final && (opc == Op_MemBarAcquire ||
999 opc == Op_MemBarAcquireLock ||
1000 opc == Op_LoadFence)) ||
1001 opc == Op_MemBarRelease ||
1002 opc == Op_StoreFence ||
1003 opc == Op_MemBarReleaseLock ||
1004 opc == Op_MemBarStoreStore ||
1005 opc == Op_MemBarCPUOrder) {
1006 Node* mem = current->in(0)->in(TypeFunc::Memory);
1007 if (mem->is_MergeMem()) {
1008 MergeMemNode* merge = mem->as_MergeMem();
1009 Node* new_st = merge->memory_at(alias_idx);
1010 if (new_st == merge->base_memory()) {
1011 // Keep searching
1012 current = new_st;
1013 continue;
1014 }
1015 // Save the new memory state for the slice and fall through
1016 // to exit.
1017 result = new_st;
1018 }
1019 }
1020 break;
1021 }
1022 if (result != NULL) {
1023 st = result;
1024 }
1025 }
1026
1027 // Loop around twice in the case Load -> Initialize -> Store.
1028 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1029 for (int trip = 0; trip <= 1; trip++) {
1030
1031 if (st->is_Store()) {
1032 Node* st_adr = st->in(MemNode::Address);
1033 if (!phase->eqv(st_adr, ld_adr)) {
1034 // Try harder before giving up... Match raw and non-raw pointers.
1035 intptr_t st_off = 0;
1036 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
1037 if (alloc == NULL) return NULL;
1038 if (alloc != ld_alloc) return NULL;
1039 if (ld_off != st_off) return NULL;
1040 // At this point we have proven something like this setup:
1041 // A = Allocate(...)
1042 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
1043 // S = StoreQ(, AddP(, A.Parm , #Off), V)
1044 // (Actually, we haven't yet proven the Q's are the same.)
1045 // In other words, we are loading from a casted version of
1046 // the same pointer-and-offset that we stored to.
1047 // Thus, we are able to replace L by V.
1048 }
1049 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1050 if (store_Opcode() != st->Opcode())
1051 return NULL;
1052 return st->in(MemNode::ValueIn);
1053 }
1054
1055 // A load from a freshly-created object always returns zero.
1056 // (This can happen after LoadNode::Ideal resets the load's memory input
1057 // to find_captured_store, which returned InitializeNode::zero_memory.)
1058 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1059 (st->in(0) == ld_alloc) &&
1060 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1061 // return a zero value for the load's basic type
1062 // (This is one of the few places where a generic PhaseTransform
1063 // can create new nodes. Think of it as lazily manifesting
1064 // virtually pre-existing constants.)
1065 assert(memory_type() != T_VALUETYPE, "should not be used for value types");
1066 Node* default_value = ld_alloc->in(AllocateNode::DefaultValue);
1067 if (default_value != NULL) {
1068 return default_value;
1069 }
1070 assert(ld_alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
1071 return phase->zerocon(memory_type());
1072 }
1073
1074 // A load from an initialization barrier can match a captured store.
1075 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1076 InitializeNode* init = st->in(0)->as_Initialize();
1077 AllocateNode* alloc = init->allocation();
1078 if ((alloc != NULL) && (alloc == ld_alloc)) {
1079 // examine a captured store value
1080 st = init->find_captured_store(ld_off, memory_size(), phase);
1081 if (st != NULL) {
1082 continue; // take one more trip around
1083 }
1084 }
1085 }
1086
1087 // Load boxed value from result of valueOf() call is input parameter.
1088 if (this->is_Load() && ld_adr->is_AddP() &&
1089 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1090 intptr_t ignore = 0;
1091 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1092 if (base != NULL && base->is_Proj() &&
1093 base->as_Proj()->_con == TypeFunc::Parms &&
1094 base->in(0)->is_CallStaticJava() &&
1095 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1096 return base->in(0)->in(TypeFunc::Parms);
1097 }
1098 }
1099
1100 break;
1101 }
1102
1103 return NULL;
1104 }
1105
1106 //----------------------is_instance_field_load_with_local_phi------------------
1107 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1108 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1109 in(Address)->is_AddP() ) {
1110 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1111 // Only instances and boxed values.
1112 if( t_oop != NULL &&
1113 (t_oop->is_ptr_to_boxed_value() ||
1114 t_oop->is_known_instance_field()) &&
1115 t_oop->offset() != Type::OffsetBot &&
1116 t_oop->offset() != Type::OffsetTop) {
1117 return true;
1118 }
1119 }
1120 return false;
1121 }
1122
1123 //------------------------------Identity---------------------------------------
1124 // Loads are identity if previous store is to same address
1125 Node* LoadNode::Identity(PhaseGVN* phase) {
1126 // Loading from a ValueTypePtr? The ValueTypePtr has the values of
1127 // all fields as input. Look for the field with matching offset.
1128 Node* addr = in(Address);
1129 intptr_t offset;
1130 Node* base = AddPNode::Ideal_base_and_offset(addr, phase, offset);
1131 if (base != NULL && base->is_ValueTypePtr() && offset > oopDesc::klass_offset_in_bytes()) {
1132 Node* value = base->as_ValueTypePtr()->field_value_by_offset((int)offset, true);
1133 if (value->is_ValueType()) {
1134 // Non-flattened value type field
1135 ValueTypeNode* vt = value->as_ValueType();
1136 if (vt->is_allocated(phase)) {
1137 value = vt->get_oop();
1138 } else {
1139 // Not yet allocated, bail out
1140 value = NULL;
1141 }
1142 }
1143 if (value != NULL) {
1144 if (Opcode() == Op_LoadN) {
1145 // Encode oop value if we are loading a narrow oop
1146 assert(!phase->type(value)->isa_narrowoop(), "should already be decoded");
1147 value = phase->transform(new EncodePNode(value, bottom_type()));
1148 }
1149 return value;
1150 }
1151 }
1152
1153 // If the previous store-maker is the right kind of Store, and the store is
1154 // to the same address, then we are equal to the value stored.
1155 Node* mem = in(Memory);
1156 Node* value = can_see_stored_value(mem, phase);
1157 if( value ) {
1158 // byte, short & char stores truncate naturally.
1159 // A load has to load the truncated value which requires
1160 // some sort of masking operation and that requires an
1161 // Ideal call instead of an Identity call.
1162 if (memory_size() < BytesPerInt) {
1163 // If the input to the store does not fit with the load's result type,
1164 // it must be truncated via an Ideal call.
1165 if (!phase->type(value)->higher_equal(phase->type(this)))
1166 return this;
1167 }
1168 // (This works even when value is a Con, but LoadNode::Value
1169 // usually runs first, producing the singleton type of the Con.)
1170 return value;
1171 }
1172
1173 // Search for an existing data phi which was generated before for the same
1174 // instance's field to avoid infinite generation of phis in a loop.
1175 Node *region = mem->in(0);
1176 if (is_instance_field_load_with_local_phi(region)) {
1177 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1178 int this_index = phase->C->get_alias_index(addr_t);
1179 int this_offset = addr_t->offset();
1180 int this_iid = addr_t->instance_id();
1181 if (!addr_t->is_known_instance() &&
1182 addr_t->is_ptr_to_boxed_value()) {
1183 // Use _idx of address base (could be Phi node) for boxed values.
1184 intptr_t ignore = 0;
1185 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1186 if (base == NULL) {
1187 return this;
1188 }
1189 this_iid = base->_idx;
1190 }
1191 const Type* this_type = bottom_type();
1192 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1193 Node* phi = region->fast_out(i);
1194 if (phi->is_Phi() && phi != mem &&
1195 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1196 return phi;
1197 }
1198 }
1199 }
1200
1201 return this;
1202 }
1203
1204 // Construct an equivalent unsigned load.
1205 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1206 BasicType bt = T_ILLEGAL;
1207 const Type* rt = NULL;
1208 switch (Opcode()) {
1209 case Op_LoadUB: return this;
1210 case Op_LoadUS: return this;
1211 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1212 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1213 default:
1214 assert(false, "no unsigned variant: %s", Name());
1215 return NULL;
1216 }
1217 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1218 raw_adr_type(), rt, bt, _mo, _control_dependency,
1219 is_unaligned_access(), is_mismatched_access());
1220 }
1221
1222 // Construct an equivalent signed load.
1223 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1224 BasicType bt = T_ILLEGAL;
1225 const Type* rt = NULL;
1226 switch (Opcode()) {
1227 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1228 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1229 case Op_LoadB: // fall through
1230 case Op_LoadS: // fall through
1231 case Op_LoadI: // fall through
1232 case Op_LoadL: return this;
1233 default:
1234 assert(false, "no signed variant: %s", Name());
1235 return NULL;
1236 }
1237 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1238 raw_adr_type(), rt, bt, _mo, _control_dependency,
1239 is_unaligned_access(), is_mismatched_access());
1240 }
1241
1242 // We're loading from an object which has autobox behaviour.
1243 // If this object is result of a valueOf call we'll have a phi
1244 // merging a newly allocated object and a load from the cache.
1245 // We want to replace this load with the original incoming
1246 // argument to the valueOf call.
1247 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1248 assert(phase->C->eliminate_boxing(), "sanity");
1249 intptr_t ignore = 0;
1250 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1251 if ((base == NULL) || base->is_Phi()) {
1252 // Push the loads from the phi that comes from valueOf up
1253 // through it to allow elimination of the loads and the recovery
1254 // of the original value. It is done in split_through_phi().
1255 return NULL;
1256 } else if (base->is_Load() ||
1257 (base->is_DecodeN() && base->in(1)->is_Load())) {
1258 // Eliminate the load of boxed value for integer types from the cache
1259 // array by deriving the value from the index into the array.
1260 // Capture the offset of the load and then reverse the computation.
1261
1262 // Get LoadN node which loads a boxing object from 'cache' array.
1263 if (base->is_DecodeN()) {
1264 base = base->in(1);
1265 }
1266 if (!base->in(Address)->is_AddP()) {
1267 return NULL; // Complex address
1268 }
1269 AddPNode* address = base->in(Address)->as_AddP();
1270 Node* cache_base = address->in(AddPNode::Base);
1271 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1272 // Get ConP node which is static 'cache' field.
1273 cache_base = cache_base->in(1);
1274 }
1275 if ((cache_base != NULL) && cache_base->is_Con()) {
1276 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1277 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1278 Node* elements[4];
1279 int shift = exact_log2(type2aelembytes(T_OBJECT));
1280 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1281 if (count > 0 && elements[0]->is_Con() &&
1282 (count == 1 ||
1283 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1284 elements[1]->in(2) == phase->intcon(shift)))) {
1285 ciObjArray* array = base_type->const_oop()->as_obj_array();
1286 // Fetch the box object cache[0] at the base of the array and get its value
1287 ciInstance* box = array->obj_at(0)->as_instance();
1288 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1289 assert(ik->is_box_klass(), "sanity");
1290 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1291 if (ik->nof_nonstatic_fields() == 1) {
1292 // This should be true nonstatic_field_at requires calling
1293 // nof_nonstatic_fields so check it anyway
1294 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1295 BasicType bt = c.basic_type();
1296 // Only integer types have boxing cache.
1297 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1298 bt == T_BYTE || bt == T_SHORT ||
1299 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1300 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1301 if (cache_low != (int)cache_low) {
1302 return NULL; // should not happen since cache is array indexed by value
1303 }
1304 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1305 if (offset != (int)offset) {
1306 return NULL; // should not happen since cache is array indexed by value
1307 }
1308 // Add up all the offsets making of the address of the load
1309 Node* result = elements[0];
1310 for (int i = 1; i < count; i++) {
1311 result = phase->transform(new AddXNode(result, elements[i]));
1312 }
1313 // Remove the constant offset from the address and then
1314 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset)));
1315 // remove the scaling of the offset to recover the original index.
1316 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1317 // Peel the shift off directly but wrap it in a dummy node
1318 // since Ideal can't return existing nodes
1319 result = new RShiftXNode(result->in(1), phase->intcon(0));
1320 } else if (result->is_Add() && result->in(2)->is_Con() &&
1321 result->in(1)->Opcode() == Op_LShiftX &&
1322 result->in(1)->in(2) == phase->intcon(shift)) {
1323 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1324 // but for boxing cache access we know that X<<Z will not overflow
1325 // (there is range check) so we do this optimizatrion by hand here.
1326 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift));
1327 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con));
1328 } else {
1329 result = new RShiftXNode(result, phase->intcon(shift));
1330 }
1331 #ifdef _LP64
1332 if (bt != T_LONG) {
1333 result = new ConvL2INode(phase->transform(result));
1334 }
1335 #else
1336 if (bt == T_LONG) {
1337 result = new ConvI2LNode(phase->transform(result));
1338 }
1339 #endif
1340 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1341 // Need to preserve unboxing load type if it is unsigned.
1342 switch(this->Opcode()) {
1343 case Op_LoadUB:
1344 result = new AndINode(phase->transform(result), phase->intcon(0xFF));
1345 break;
1346 case Op_LoadUS:
1347 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF));
1348 break;
1349 }
1350 return result;
1351 }
1352 }
1353 }
1354 }
1355 }
1356 return NULL;
1357 }
1358
1359 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1360 Node* region = phi->in(0);
1361 if (region == NULL) {
1362 return false; // Wait stable graph
1363 }
1364 uint cnt = phi->req();
1365 for (uint i = 1; i < cnt; i++) {
1366 Node* rc = region->in(i);
1367 if (rc == NULL || phase->type(rc) == Type::TOP)
1368 return false; // Wait stable graph
1369 Node* in = phi->in(i);
1370 if (in == NULL || phase->type(in) == Type::TOP)
1371 return false; // Wait stable graph
1372 }
1373 return true;
1374 }
1375 //------------------------------split_through_phi------------------------------
1376 // Split instance or boxed field load through Phi.
1377 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1378 Node* mem = in(Memory);
1379 Node* address = in(Address);
1380 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1381
1382 assert((t_oop != NULL) &&
1383 (t_oop->is_known_instance_field() ||
1384 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1385
1386 Compile* C = phase->C;
1387 intptr_t ignore = 0;
1388 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1389 bool base_is_phi = (base != NULL) && base->is_Phi();
1390 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1391 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1392 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1393
1394 if (!((mem->is_Phi() || base_is_phi) &&
1395 (load_boxed_values || t_oop->is_known_instance_field()))) {
1396 return NULL; // memory is not Phi
1397 }
1398
1399 if (mem->is_Phi()) {
1400 if (!stable_phi(mem->as_Phi(), phase)) {
1401 return NULL; // Wait stable graph
1402 }
1403 uint cnt = mem->req();
1404 // Check for loop invariant memory.
1405 if (cnt == 3) {
1406 for (uint i = 1; i < cnt; i++) {
1407 Node* in = mem->in(i);
1408 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1409 if (m == mem) {
1410 set_req(Memory, mem->in(cnt - i));
1411 return this; // made change
1412 }
1413 }
1414 }
1415 }
1416 if (base_is_phi) {
1417 if (!stable_phi(base->as_Phi(), phase)) {
1418 return NULL; // Wait stable graph
1419 }
1420 uint cnt = base->req();
1421 // Check for loop invariant memory.
1422 if (cnt == 3) {
1423 for (uint i = 1; i < cnt; i++) {
1424 if (base->in(i) == base) {
1425 return NULL; // Wait stable graph
1426 }
1427 }
1428 }
1429 }
1430
1431 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1432
1433 // Split through Phi (see original code in loopopts.cpp).
1434 assert(C->have_alias_type(t_oop), "instance should have alias type");
1435
1436 // Do nothing here if Identity will find a value
1437 // (to avoid infinite chain of value phis generation).
1438 if (!phase->eqv(this, this->Identity(phase)))
1439 return NULL;
1440
1441 // Select Region to split through.
1442 Node* region;
1443 if (!base_is_phi) {
1444 assert(mem->is_Phi(), "sanity");
1445 region = mem->in(0);
1446 // Skip if the region dominates some control edge of the address.
1447 if (!MemNode::all_controls_dominate(address, region))
1448 return NULL;
1449 } else if (!mem->is_Phi()) {
1450 assert(base_is_phi, "sanity");
1451 region = base->in(0);
1452 // Skip if the region dominates some control edge of the memory.
1453 if (!MemNode::all_controls_dominate(mem, region))
1454 return NULL;
1455 } else if (base->in(0) != mem->in(0)) {
1456 assert(base_is_phi && mem->is_Phi(), "sanity");
1457 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1458 region = base->in(0);
1459 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1460 region = mem->in(0);
1461 } else {
1462 return NULL; // complex graph
1463 }
1464 } else {
1465 assert(base->in(0) == mem->in(0), "sanity");
1466 region = mem->in(0);
1467 }
1468
1469 const Type* this_type = this->bottom_type();
1470 int this_index = C->get_alias_index(t_oop);
1471 int this_offset = t_oop->offset();
1472 int this_iid = t_oop->instance_id();
1473 if (!t_oop->is_known_instance() && load_boxed_values) {
1474 // Use _idx of address base for boxed values.
1475 this_iid = base->_idx;
1476 }
1477 PhaseIterGVN* igvn = phase->is_IterGVN();
1478 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1479 for (uint i = 1; i < region->req(); i++) {
1480 Node* x;
1481 Node* the_clone = NULL;
1482 if (region->in(i) == C->top()) {
1483 x = C->top(); // Dead path? Use a dead data op
1484 } else {
1485 x = this->clone(); // Else clone up the data op
1486 the_clone = x; // Remember for possible deletion.
1487 // Alter data node to use pre-phi inputs
1488 if (this->in(0) == region) {
1489 x->set_req(0, region->in(i));
1490 } else {
1491 x->set_req(0, NULL);
1492 }
1493 if (mem->is_Phi() && (mem->in(0) == region)) {
1494 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1495 }
1496 if (address->is_Phi() && address->in(0) == region) {
1497 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1498 }
1499 if (base_is_phi && (base->in(0) == region)) {
1500 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1501 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1502 x->set_req(Address, adr_x);
1503 }
1504 }
1505 // Check for a 'win' on some paths
1506 const Type *t = x->Value(igvn);
1507
1508 bool singleton = t->singleton();
1509
1510 // See comments in PhaseIdealLoop::split_thru_phi().
1511 if (singleton && t == Type::TOP) {
1512 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1513 }
1514
1515 if (singleton) {
1516 x = igvn->makecon(t);
1517 } else {
1518 // We now call Identity to try to simplify the cloned node.
1519 // Note that some Identity methods call phase->type(this).
1520 // Make sure that the type array is big enough for
1521 // our new node, even though we may throw the node away.
1522 // (This tweaking with igvn only works because x is a new node.)
1523 igvn->set_type(x, t);
1524 // If x is a TypeNode, capture any more-precise type permanently into Node
1525 // otherwise it will be not updated during igvn->transform since
1526 // igvn->type(x) is set to x->Value() already.
1527 x->raise_bottom_type(t);
1528 Node *y = x->Identity(igvn);
1529 if (y != x) {
1530 x = y;
1531 } else {
1532 y = igvn->hash_find_insert(x);
1533 if (y) {
1534 x = y;
1535 } else {
1536 // Else x is a new node we are keeping
1537 // We do not need register_new_node_with_optimizer
1538 // because set_type has already been called.
1539 igvn->_worklist.push(x);
1540 }
1541 }
1542 }
1543 if (x != the_clone && the_clone != NULL) {
1544 igvn->remove_dead_node(the_clone);
1545 }
1546 phi->set_req(i, x);
1547 }
1548 // Record Phi
1549 igvn->register_new_node_with_optimizer(phi);
1550 return phi;
1551 }
1552
1553 //------------------------------Ideal------------------------------------------
1554 // If the load is from Field memory and the pointer is non-null, it might be possible to
1555 // zero out the control input.
1556 // If the offset is constant and the base is an object allocation,
1557 // try to hook me up to the exact initializing store.
1558 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1559 Node* p = MemNode::Ideal_common(phase, can_reshape);
1560 if (p) return (p == NodeSentinel) ? NULL : p;
1561
1562 Node* ctrl = in(MemNode::Control);
1563 Node* address = in(MemNode::Address);
1564 bool progress = false;
1565
1566 // Skip up past a SafePoint control. Cannot do this for Stores because
1567 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1568 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1569 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw ) {
1570 ctrl = ctrl->in(0);
1571 set_req(MemNode::Control,ctrl);
1572 progress = true;
1573 }
1574
1575 intptr_t ignore = 0;
1576 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1577 if (base != NULL
1578 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1579 // Check for useless control edge in some common special cases
1580 if (in(MemNode::Control) != NULL
1581 && can_remove_control()
1582 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1583 && all_controls_dominate(base, phase->C->start())) {
1584 // A method-invariant, non-null address (constant or 'this' argument).
1585 set_req(MemNode::Control, NULL);
1586 progress = true;
1587 }
1588 }
1589
1590 Node* mem = in(MemNode::Memory);
1591 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1592
1593 if (can_reshape && (addr_t != NULL)) {
1594 // try to optimize our memory input
1595 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1596 if (opt_mem != mem) {
1597 set_req(MemNode::Memory, opt_mem);
1598 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1599 return this;
1600 }
1601 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1602 if ((t_oop != NULL) &&
1603 (t_oop->is_known_instance_field() ||
1604 t_oop->is_ptr_to_boxed_value())) {
1605 PhaseIterGVN *igvn = phase->is_IterGVN();
1606 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1607 // Delay this transformation until memory Phi is processed.
1608 phase->is_IterGVN()->_worklist.push(this);
1609 return NULL;
1610 }
1611 // Split instance field load through Phi.
1612 Node* result = split_through_phi(phase);
1613 if (result != NULL) return result;
1614
1615 if (t_oop->is_ptr_to_boxed_value()) {
1616 Node* result = eliminate_autobox(phase);
1617 if (result != NULL) return result;
1618 }
1619 }
1620 }
1621
1622 // Is there a dominating load that loads the same value? Leave
1623 // anything that is not a load of a field/array element (like
1624 // barriers etc.) alone
1625 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) {
1626 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1627 Node *use = mem->fast_out(i);
1628 if (use != this &&
1629 use->Opcode() == Opcode() &&
1630 use->in(0) != NULL &&
1631 use->in(0) != in(0) &&
1632 use->in(Address) == in(Address)) {
1633 Node* ctl = in(0);
1634 for (int i = 0; i < 10 && ctl != NULL; i++) {
1635 ctl = IfNode::up_one_dom(ctl);
1636 if (ctl == use->in(0)) {
1637 set_req(0, use->in(0));
1638 return this;
1639 }
1640 }
1641 }
1642 }
1643 }
1644
1645 // Check for prior store with a different base or offset; make Load
1646 // independent. Skip through any number of them. Bail out if the stores
1647 // are in an endless dead cycle and report no progress. This is a key
1648 // transform for Reflection. However, if after skipping through the Stores
1649 // we can't then fold up against a prior store do NOT do the transform as
1650 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1651 // array memory alive twice: once for the hoisted Load and again after the
1652 // bypassed Store. This situation only works if EVERYBODY who does
1653 // anti-dependence work knows how to bypass. I.e. we need all
1654 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1655 // the alias index stuff. So instead, peek through Stores and IFF we can
1656 // fold up, do so.
1657 Node* prev_mem = find_previous_store(phase);
1658 if (prev_mem != NULL) {
1659 Node* value = can_see_arraycopy_value(prev_mem, phase);
1660 if (value != NULL) {
1661 return value;
1662 }
1663 }
1664 // Steps (a), (b): Walk past independent stores to find an exact match.
1665 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1666 // (c) See if we can fold up on the spot, but don't fold up here.
1667 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1668 // just return a prior value, which is done by Identity calls.
1669 if (can_see_stored_value(prev_mem, phase)) {
1670 // Make ready for step (d):
1671 set_req(MemNode::Memory, prev_mem);
1672 return this;
1673 }
1674 }
1675
1676 return progress ? this : NULL;
1677 }
1678
1679 // Helper to recognize certain Klass fields which are invariant across
1680 // some group of array types (e.g., int[] or all T[] where T < Object).
1681 const Type*
1682 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1683 ciKlass* klass) const {
1684 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1685 // The field is Klass::_modifier_flags. Return its (constant) value.
1686 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1687 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1688 return TypeInt::make(klass->modifier_flags());
1689 }
1690 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1691 // The field is Klass::_access_flags. Return its (constant) value.
1692 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1693 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1694 return TypeInt::make(klass->access_flags());
1695 }
1696 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1697 // The field is Klass::_layout_helper. Return its constant value if known.
1698 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1699 return TypeInt::make(klass->layout_helper());
1700 }
1701
1702 // No match.
1703 return NULL;
1704 }
1705
1706 //------------------------------Value-----------------------------------------
1707 const Type* LoadNode::Value(PhaseGVN* phase) const {
1708 // Either input is TOP ==> the result is TOP
1709 Node* mem = in(MemNode::Memory);
1710 const Type *t1 = phase->type(mem);
1711 if (t1 == Type::TOP) return Type::TOP;
1712 Node* adr = in(MemNode::Address);
1713 const TypePtr* tp = phase->type(adr)->isa_ptr();
1714 if (tp == NULL || tp->empty()) return Type::TOP;
1715 int off = tp->offset();
1716 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1717 Compile* C = phase->C;
1718
1719 // Try to guess loaded type from pointer type
1720 if (tp->isa_aryptr()) {
1721 const TypeAryPtr* ary = tp->is_aryptr();
1722 const Type* t = ary->elem();
1723
1724 // Determine whether the reference is beyond the header or not, by comparing
1725 // the offset against the offset of the start of the array's data.
1726 // Different array types begin at slightly different offsets (12 vs. 16).
1727 // We choose T_BYTE as an example base type that is least restrictive
1728 // as to alignment, which will therefore produce the smallest
1729 // possible base offset.
1730 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1731 const bool off_beyond_header = ((uint)off >= (uint)min_base_off);
1732
1733 // Try to constant-fold a stable array element.
1734 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1735 // Make sure the reference is not into the header and the offset is constant
1736 ciObject* aobj = ary->const_oop();
1737 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1738 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1739 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1740 stable_dimension,
1741 memory_type(), is_unsigned());
1742 if (con_type != NULL) {
1743 return con_type;
1744 }
1745 }
1746 }
1747
1748 // Don't do this for integer types. There is only potential profit if
1749 // the element type t is lower than _type; that is, for int types, if _type is
1750 // more restrictive than t. This only happens here if one is short and the other
1751 // char (both 16 bits), and in those cases we've made an intentional decision
1752 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1753 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1754 //
1755 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1756 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1757 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1758 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1759 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1760 // In fact, that could have been the original type of p1, and p1 could have
1761 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1762 // expression (LShiftL quux 3) independently optimized to the constant 8.
1763 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1764 && (_type->isa_vect() == NULL)
1765 && t->isa_valuetype() == NULL
1766 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1767 // t might actually be lower than _type, if _type is a unique
1768 // concrete subclass of abstract class t.
1769 if (off_beyond_header) { // is the offset beyond the header?
1770 const Type* jt = t->join_speculative(_type);
1771 // In any case, do not allow the join, per se, to empty out the type.
1772 if (jt->empty() && !t->empty()) {
1773 // This can happen if a interface-typed array narrows to a class type.
1774 jt = _type;
1775 }
1776 #ifdef ASSERT
1777 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1778 // The pointers in the autobox arrays are always non-null
1779 Node* base = adr->in(AddPNode::Base);
1780 if ((base != NULL) && base->is_DecodeN()) {
1781 // Get LoadN node which loads IntegerCache.cache field
1782 base = base->in(1);
1783 }
1784 if ((base != NULL) && base->is_Con()) {
1785 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1786 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1787 // It could be narrow oop
1788 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1789 }
1790 }
1791 }
1792 #endif
1793 return jt;
1794 }
1795 }
1796 } else if (tp->base() == Type::InstPtr) {
1797 assert( off != Type::OffsetBot ||
1798 // arrays can be cast to Objects
1799 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1800 tp->is_oopptr()->klass() == ciEnv::current()->Class_klass() ||
1801 // unsafe field access may not have a constant offset
1802 C->has_unsafe_access(),
1803 "Field accesses must be precise" );
1804 // For oop loads, we expect the _type to be precise.
1805
1806 // Optimize loads from constant fields.
1807 const TypeInstPtr* tinst = tp->is_instptr();
1808 ciObject* const_oop = tinst->const_oop();
1809 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1810 BasicType bt = memory_type();
1811 ciType* mirror_type = const_oop->as_instance()->java_mirror_type();
1812 if (mirror_type != NULL && mirror_type->is_valuetype()) {
1813 ciValueKlass* vk = mirror_type->as_value_klass();
1814 if (off == vk->default_value_offset()) {
1815 // Loading a special hidden field that contains the oop of the default value type
1816 const Type* const_oop = TypeInstPtr::make(vk->default_value_instance());
1817 return (bt == T_NARROWOOP) ? const_oop->make_narrowoop() : const_oop;
1818 }
1819 }
1820 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), bt);
1821 if (con_type != NULL) {
1822 return con_type;
1823 }
1824 }
1825 } else if (tp->base() == Type::KlassPtr) {
1826 assert( off != Type::OffsetBot ||
1827 // arrays can be cast to Objects
1828 tp->is_klassptr()->klass() == NULL ||
1829 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1830 // also allow array-loading from the primary supertype
1831 // array during subtype checks
1832 Opcode() == Op_LoadKlass,
1833 "Field accesses must be precise" );
1834 // For klass/static loads, we expect the _type to be precise
1835 } else if (tp->base() == Type::RawPtr && !StressReflectiveCode) {
1836 if (adr->is_Load() && off == 0) {
1837 /* With mirrors being an indirect in the Klass*
1838 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
1839 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
1840 *
1841 * So check the type and klass of the node before the LoadP.
1842 */
1843 Node* adr2 = adr->in(MemNode::Address);
1844 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1845 if (tkls != NULL) {
1846 ciKlass* klass = tkls->klass();
1847 if (klass != NULL && klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1848 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1849 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1850 return TypeInstPtr::make(klass->java_mirror());
1851 }
1852 }
1853 } else {
1854 // Check for a load of the default value offset from the ValueKlassFixedBlock:
1855 // LoadI(LoadP(value_klass, adr_valueklass_fixed_block_offset), default_value_offset_offset)
1856 intptr_t offset = 0;
1857 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
1858 if (base != NULL && base->is_Load() && offset == in_bytes(ValueKlass::default_value_offset_offset())) {
1859 const TypeKlassPtr* tkls = phase->type(base->in(MemNode::Address))->isa_klassptr();
1860 if (tkls != NULL && tkls->is_loaded() && tkls->klass_is_exact() && tkls->isa_valuetype() &&
1861 tkls->offset() == in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset())) {
1862 assert(base->Opcode() == Op_LoadP, "must load an oop from klass");
1863 assert(Opcode() == Op_LoadI, "must load an int from fixed block");
1864 return TypeInt::make(tkls->klass()->as_value_klass()->default_value_offset());
1865 }
1866 }
1867 }
1868 }
1869
1870 const TypeKlassPtr *tkls = tp->isa_klassptr();
1871 if (tkls != NULL && !StressReflectiveCode) {
1872 ciKlass* klass = tkls->klass();
1873 if (tkls->is_loaded() && tkls->klass_is_exact()) {
1874 // We are loading a field from a Klass metaobject whose identity
1875 // is known at compile time (the type is "exact" or "precise").
1876 // Check for fields we know are maintained as constants by the VM.
1877 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1878 // The field is Klass::_super_check_offset. Return its (constant) value.
1879 // (Folds up type checking code.)
1880 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1881 return TypeInt::make(klass->super_check_offset());
1882 }
1883 // Compute index into primary_supers array
1884 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1885 // Check for overflowing; use unsigned compare to handle the negative case.
1886 if( depth < ciKlass::primary_super_limit() ) {
1887 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1888 // (Folds up type checking code.)
1889 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1890 ciKlass *ss = klass->super_of_depth(depth);
1891 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1892 }
1893 const Type* aift = load_array_final_field(tkls, klass);
1894 if (aift != NULL) return aift;
1895 }
1896
1897 // We can still check if we are loading from the primary_supers array at a
1898 // shallow enough depth. Even though the klass is not exact, entries less
1899 // than or equal to its super depth are correct.
1900 if (tkls->is_loaded()) {
1901 ciType *inner = klass;
1902 while( inner->is_obj_array_klass() )
1903 inner = inner->as_obj_array_klass()->base_element_type();
1904 if( inner->is_instance_klass() &&
1905 !inner->as_instance_klass()->flags().is_interface() ) {
1906 // Compute index into primary_supers array
1907 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1908 // Check for overflowing; use unsigned compare to handle the negative case.
1909 if( depth < ciKlass::primary_super_limit() &&
1910 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1911 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1912 // (Folds up type checking code.)
1913 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1914 ciKlass *ss = klass->super_of_depth(depth);
1915 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1916 }
1917 }
1918 }
1919
1920 // If the type is enough to determine that the thing is not an array,
1921 // we can give the layout_helper a positive interval type.
1922 // This will help short-circuit some reflective code.
1923 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1924 && !klass->is_array_klass() // not directly typed as an array
1925 && !klass->is_interface() // specifically not Serializable & Cloneable
1926 && !klass->is_java_lang_Object() // not the supertype of all T[]
1927 ) {
1928 // Note: When interfaces are reliable, we can narrow the interface
1929 // test to (klass != Serializable && klass != Cloneable).
1930 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1931 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1932 // The key property of this type is that it folds up tests
1933 // for array-ness, since it proves that the layout_helper is positive.
1934 // Thus, a generic value like the basic object layout helper works fine.
1935 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1936 }
1937 }
1938
1939 // If we are loading from a freshly-allocated object, produce a zero,
1940 // if the load is provably beyond the header of the object.
1941 // (Also allow a variable load from a fresh array to produce zero.)
1942 const TypeOopPtr *tinst = tp->isa_oopptr();
1943 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1944 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1945 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1946 Node* value = can_see_stored_value(mem,phase);
1947 if (value != NULL && value->is_Con()) {
1948 assert(value->bottom_type()->higher_equal(_type),"sanity");
1949 return value->bottom_type();
1950 }
1951 }
1952
1953 if (is_instance) {
1954 // If we have an instance type and our memory input is the
1955 // programs's initial memory state, there is no matching store,
1956 // so just return a zero of the appropriate type
1957 Node *mem = in(MemNode::Memory);
1958 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1959 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1960 return Type::get_zero_type(_type->basic_type());
1961 }
1962 }
1963 return _type;
1964 }
1965
1966 //------------------------------match_edge-------------------------------------
1967 // Do we Match on this edge index or not? Match only the address.
1968 uint LoadNode::match_edge(uint idx) const {
1969 return idx == MemNode::Address;
1970 }
1971
1972 //--------------------------LoadBNode::Ideal--------------------------------------
1973 //
1974 // If the previous store is to the same address as this load,
1975 // and the value stored was larger than a byte, replace this load
1976 // with the value stored truncated to a byte. If no truncation is
1977 // needed, the replacement is done in LoadNode::Identity().
1978 //
1979 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1980 Node* mem = in(MemNode::Memory);
1981 Node* value = can_see_stored_value(mem,phase);
1982 if( value && !phase->type(value)->higher_equal( _type ) ) {
1983 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
1984 return new RShiftINode(result, phase->intcon(24));
1985 }
1986 // Identity call will handle the case where truncation is not needed.
1987 return LoadNode::Ideal(phase, can_reshape);
1988 }
1989
1990 const Type* LoadBNode::Value(PhaseGVN* phase) const {
1991 Node* mem = in(MemNode::Memory);
1992 Node* value = can_see_stored_value(mem,phase);
1993 if (value != NULL && value->is_Con() &&
1994 !value->bottom_type()->higher_equal(_type)) {
1995 // If the input to the store does not fit with the load's result type,
1996 // it must be truncated. We can't delay until Ideal call since
1997 // a singleton Value is needed for split_thru_phi optimization.
1998 int con = value->get_int();
1999 return TypeInt::make((con << 24) >> 24);
2000 }
2001 return LoadNode::Value(phase);
2002 }
2003
2004 //--------------------------LoadUBNode::Ideal-------------------------------------
2005 //
2006 // If the previous store is to the same address as this load,
2007 // and the value stored was larger than a byte, replace this load
2008 // with the value stored truncated to a byte. If no truncation is
2009 // needed, the replacement is done in LoadNode::Identity().
2010 //
2011 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2012 Node* mem = in(MemNode::Memory);
2013 Node* value = can_see_stored_value(mem, phase);
2014 if (value && !phase->type(value)->higher_equal(_type))
2015 return new AndINode(value, phase->intcon(0xFF));
2016 // Identity call will handle the case where truncation is not needed.
2017 return LoadNode::Ideal(phase, can_reshape);
2018 }
2019
2020 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2021 Node* mem = in(MemNode::Memory);
2022 Node* value = can_see_stored_value(mem,phase);
2023 if (value != NULL && value->is_Con() &&
2024 !value->bottom_type()->higher_equal(_type)) {
2025 // If the input to the store does not fit with the load's result type,
2026 // it must be truncated. We can't delay until Ideal call since
2027 // a singleton Value is needed for split_thru_phi optimization.
2028 int con = value->get_int();
2029 return TypeInt::make(con & 0xFF);
2030 }
2031 return LoadNode::Value(phase);
2032 }
2033
2034 //--------------------------LoadUSNode::Ideal-------------------------------------
2035 //
2036 // If the previous store is to the same address as this load,
2037 // and the value stored was larger than a char, replace this load
2038 // with the value stored truncated to a char. If no truncation is
2039 // needed, the replacement is done in LoadNode::Identity().
2040 //
2041 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2042 Node* mem = in(MemNode::Memory);
2043 Node* value = can_see_stored_value(mem,phase);
2044 if( value && !phase->type(value)->higher_equal( _type ) )
2045 return new AndINode(value,phase->intcon(0xFFFF));
2046 // Identity call will handle the case where truncation is not needed.
2047 return LoadNode::Ideal(phase, can_reshape);
2048 }
2049
2050 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2051 Node* mem = in(MemNode::Memory);
2052 Node* value = can_see_stored_value(mem,phase);
2053 if (value != NULL && value->is_Con() &&
2054 !value->bottom_type()->higher_equal(_type)) {
2055 // If the input to the store does not fit with the load's result type,
2056 // it must be truncated. We can't delay until Ideal call since
2057 // a singleton Value is needed for split_thru_phi optimization.
2058 int con = value->get_int();
2059 return TypeInt::make(con & 0xFFFF);
2060 }
2061 return LoadNode::Value(phase);
2062 }
2063
2064 //--------------------------LoadSNode::Ideal--------------------------------------
2065 //
2066 // If the previous store is to the same address as this load,
2067 // and the value stored was larger than a short, replace this load
2068 // with the value stored truncated to a short. If no truncation is
2069 // needed, the replacement is done in LoadNode::Identity().
2070 //
2071 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2072 Node* mem = in(MemNode::Memory);
2073 Node* value = can_see_stored_value(mem,phase);
2074 if( value && !phase->type(value)->higher_equal( _type ) ) {
2075 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
2076 return new RShiftINode(result, phase->intcon(16));
2077 }
2078 // Identity call will handle the case where truncation is not needed.
2079 return LoadNode::Ideal(phase, can_reshape);
2080 }
2081
2082 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2083 Node* mem = in(MemNode::Memory);
2084 Node* value = can_see_stored_value(mem,phase);
2085 if (value != NULL && value->is_Con() &&
2086 !value->bottom_type()->higher_equal(_type)) {
2087 // If the input to the store does not fit with the load's result type,
2088 // it must be truncated. We can't delay until Ideal call since
2089 // a singleton Value is needed for split_thru_phi optimization.
2090 int con = value->get_int();
2091 return TypeInt::make((con << 16) >> 16);
2092 }
2093 return LoadNode::Value(phase);
2094 }
2095
2096 //=============================================================================
2097 //----------------------------LoadKlassNode::make------------------------------
2098 // Polymorphic factory method:
2099 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2100 // sanity check the alias category against the created node type
2101 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2102 assert(adr_type != NULL, "expecting TypeKlassPtr");
2103 #ifdef _LP64
2104 if (adr_type->is_ptr_to_narrowklass()) {
2105 assert(UseCompressedClassPointers, "no compressed klasses");
2106 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2107 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2108 }
2109 #endif
2110 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2111 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2112 }
2113
2114 //------------------------------Value------------------------------------------
2115 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2116 return klass_value_common(phase);
2117 }
2118
2119 // In most cases, LoadKlassNode does not have the control input set. If the control
2120 // input is set, it must not be removed (by LoadNode::Ideal()).
2121 bool LoadKlassNode::can_remove_control() const {
2122 return false;
2123 }
2124
2125 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2126 // Either input is TOP ==> the result is TOP
2127 const Type *t1 = phase->type( in(MemNode::Memory) );
2128 if (t1 == Type::TOP) return Type::TOP;
2129 Node *adr = in(MemNode::Address);
2130 const Type *t2 = phase->type( adr );
2131 if (t2 == Type::TOP) return Type::TOP;
2132 const TypePtr *tp = t2->is_ptr();
2133 if (TypePtr::above_centerline(tp->ptr()) ||
2134 tp->ptr() == TypePtr::Null) return Type::TOP;
2135
2136 // Return a more precise klass, if possible
2137 const TypeInstPtr *tinst = tp->isa_instptr();
2138 if (tinst != NULL) {
2139 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2140 int offset = tinst->offset();
2141 if (ik == phase->C->env()->Class_klass()
2142 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2143 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2144 // We are loading a special hidden field from a Class mirror object,
2145 // the field which points to the VM's Klass metaobject.
2146 ciType* t = tinst->java_mirror_type();
2147 // java_mirror_type returns non-null for compile-time Class constants.
2148 if (t != NULL) {
2149 // constant oop => constant klass
2150 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2151 if (t->is_void()) {
2152 // We cannot create a void array. Since void is a primitive type return null
2153 // klass. Users of this result need to do a null check on the returned klass.
2154 return TypePtr::NULL_PTR;
2155 }
2156 return TypeKlassPtr::make(ciArrayKlass::make(t));
2157 }
2158 if (!t->is_klass()) {
2159 // a primitive Class (e.g., int.class) has NULL for a klass field
2160 return TypePtr::NULL_PTR;
2161 }
2162 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2163 return TypeKlassPtr::make(t->as_klass());
2164 }
2165 // non-constant mirror, so we can't tell what's going on
2166 }
2167 if( !ik->is_loaded() )
2168 return _type; // Bail out if not loaded
2169 if (offset == oopDesc::klass_offset_in_bytes()) {
2170 if (tinst->klass_is_exact()) {
2171 return TypeKlassPtr::make(ik);
2172 }
2173 // See if we can become precise: no subklasses and no interface
2174 // (Note: We need to support verified interfaces.)
2175 if (!ik->is_interface() && !ik->has_subklass()) {
2176 //assert(!UseExactTypes, "this code should be useless with exact types");
2177 // Add a dependence; if any subclass added we need to recompile
2178 if (!ik->is_final()) {
2179 // %%% should use stronger assert_unique_concrete_subtype instead
2180 phase->C->dependencies()->assert_leaf_type(ik);
2181 }
2182 // Return precise klass
2183 return TypeKlassPtr::make(ik);
2184 }
2185
2186 // Return root of possible klass
2187 return TypeKlassPtr::make(TypePtr::NotNull, ik, Type::Offset(0));
2188 }
2189 }
2190
2191 // Check for loading klass from an array
2192 const TypeAryPtr *tary = tp->isa_aryptr();
2193 if( tary != NULL ) {
2194 ciKlass *tary_klass = tary->klass();
2195 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2196 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2197 if (tary->klass_is_exact()) {
2198 return TypeKlassPtr::make(tary_klass);
2199 }
2200 ciArrayKlass *ak = tary->klass()->as_array_klass();
2201 // If the klass is an object array, we defer the question to the
2202 // array component klass.
2203 if( ak->is_obj_array_klass() ) {
2204 assert( ak->is_loaded(), "" );
2205 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2206 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2207 ciInstanceKlass* ik = base_k->as_instance_klass();
2208 // See if we can become precise: no subklasses and no interface
2209 if (!ik->is_interface() && !ik->has_subklass()) {
2210 //assert(!UseExactTypes, "this code should be useless with exact types");
2211 // Add a dependence; if any subclass added we need to recompile
2212 if (!ik->is_final()) {
2213 phase->C->dependencies()->assert_leaf_type(ik);
2214 }
2215 // Return precise array klass
2216 return TypeKlassPtr::make(ak);
2217 }
2218 }
2219 return TypeKlassPtr::make(TypePtr::NotNull, ak, Type::Offset(0));
2220 } else { // Found a type-array?
2221 //assert(!UseExactTypes, "this code should be useless with exact types");
2222 assert( ak->is_type_array_klass(), "" );
2223 return TypeKlassPtr::make(ak); // These are always precise
2224 }
2225 }
2226 }
2227
2228 // Check for loading klass from an array klass
2229 const TypeKlassPtr *tkls = tp->isa_klassptr();
2230 if (tkls != NULL && !StressReflectiveCode) {
2231 if (!tkls->is_loaded()) {
2232 return _type; // Bail out if not loaded
2233 }
2234 ciKlass* klass = tkls->klass();
2235 if( klass->is_obj_array_klass() &&
2236 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2237 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2238 // // Always returning precise element type is incorrect,
2239 // // e.g., element type could be object and array may contain strings
2240 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2241
2242 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2243 // according to the element type's subclassing.
2244 return TypeKlassPtr::make(tkls->ptr(), elem, Type::Offset(0));
2245 }
2246 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2247 tkls->offset() == in_bytes(Klass::super_offset())) {
2248 ciKlass* sup = klass->as_instance_klass()->super();
2249 // The field is Klass::_super. Return its (constant) value.
2250 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2251 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2252 }
2253 }
2254
2255 // Bailout case
2256 return LoadNode::Value(phase);
2257 }
2258
2259 //------------------------------Identity---------------------------------------
2260 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2261 // Also feed through the klass in Allocate(...klass...)._klass.
2262 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2263 return klass_identity_common(phase);
2264 }
2265
2266 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2267 Node* x = LoadNode::Identity(phase);
2268 if (x != this) return x;
2269
2270 // Take apart the address into an oop and and offset.
2271 // Return 'this' if we cannot.
2272 Node* adr = in(MemNode::Address);
2273 intptr_t offset = 0;
2274 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2275 if (base == NULL) return this;
2276 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2277 if (toop == NULL) return this;
2278
2279 // Step over potential GC barrier for OopHandle resolve
2280 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2281 if (bs->is_gc_barrier_node(base)) {
2282 base = bs->step_over_gc_barrier(base);
2283 }
2284
2285 // We can fetch the klass directly through an AllocateNode.
2286 // This works even if the klass is not constant (clone or newArray).
2287 if (offset == oopDesc::klass_offset_in_bytes()) {
2288 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2289 if (allocated_klass != NULL) {
2290 return allocated_klass;
2291 }
2292 }
2293
2294 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2295 // See inline_native_Class_query for occurrences of these patterns.
2296 // Java Example: x.getClass().isAssignableFrom(y)
2297 //
2298 // This improves reflective code, often making the Class
2299 // mirror go completely dead. (Current exception: Class
2300 // mirrors may appear in debug info, but we could clean them out by
2301 // introducing a new debug info operator for Klass.java_mirror).
2302
2303 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2304 && offset == java_lang_Class::klass_offset_in_bytes()) {
2305 if (base->is_Load()) {
2306 Node* base2 = base->in(MemNode::Address);
2307 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2308 Node* adr2 = base2->in(MemNode::Address);
2309 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2310 if (tkls != NULL && !tkls->empty()
2311 && (tkls->klass()->is_instance_klass() ||
2312 tkls->klass()->is_array_klass())
2313 && adr2->is_AddP()
2314 ) {
2315 int mirror_field = in_bytes(Klass::java_mirror_offset());
2316 if (tkls->offset() == mirror_field) {
2317 return adr2->in(AddPNode::Base);
2318 }
2319 }
2320 }
2321 }
2322 }
2323
2324 return this;
2325 }
2326
2327
2328 //------------------------------Value------------------------------------------
2329 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2330 const Type *t = klass_value_common(phase);
2331 if (t == Type::TOP)
2332 return t;
2333
2334 return t->make_narrowklass();
2335 }
2336
2337 //------------------------------Identity---------------------------------------
2338 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2339 // Also feed through the klass in Allocate(...klass...)._klass.
2340 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2341 Node *x = klass_identity_common(phase);
2342
2343 const Type *t = phase->type( x );
2344 if( t == Type::TOP ) return x;
2345 if( t->isa_narrowklass()) return x;
2346 assert (!t->isa_narrowoop(), "no narrow oop here");
2347
2348 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2349 }
2350
2351 //------------------------------Value-----------------------------------------
2352 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2353 // Either input is TOP ==> the result is TOP
2354 const Type *t1 = phase->type( in(MemNode::Memory) );
2355 if( t1 == Type::TOP ) return Type::TOP;
2356 Node *adr = in(MemNode::Address);
2357 const Type *t2 = phase->type( adr );
2358 if( t2 == Type::TOP ) return Type::TOP;
2359 const TypePtr *tp = t2->is_ptr();
2360 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2361 const TypeAryPtr *tap = tp->isa_aryptr();
2362 if( !tap ) return _type;
2363 return tap->size();
2364 }
2365
2366 //-------------------------------Ideal---------------------------------------
2367 // Feed through the length in AllocateArray(...length...)._length.
2368 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2369 Node* p = MemNode::Ideal_common(phase, can_reshape);
2370 if (p) return (p == NodeSentinel) ? NULL : p;
2371
2372 // Take apart the address into an oop and and offset.
2373 // Return 'this' if we cannot.
2374 Node* adr = in(MemNode::Address);
2375 intptr_t offset = 0;
2376 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2377 if (base == NULL) return NULL;
2378 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2379 if (tary == NULL) return NULL;
2380
2381 // We can fetch the length directly through an AllocateArrayNode.
2382 // This works even if the length is not constant (clone or newArray).
2383 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2384 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2385 if (alloc != NULL) {
2386 Node* allocated_length = alloc->Ideal_length();
2387 Node* len = alloc->make_ideal_length(tary, phase);
2388 if (allocated_length != len) {
2389 // New CastII improves on this.
2390 return len;
2391 }
2392 }
2393 }
2394
2395 return NULL;
2396 }
2397
2398 //------------------------------Identity---------------------------------------
2399 // Feed through the length in AllocateArray(...length...)._length.
2400 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2401 Node* x = LoadINode::Identity(phase);
2402 if (x != this) return x;
2403
2404 // Take apart the address into an oop and and offset.
2405 // Return 'this' if we cannot.
2406 Node* adr = in(MemNode::Address);
2407 intptr_t offset = 0;
2408 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2409 if (base == NULL) return this;
2410 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2411 if (tary == NULL) return this;
2412
2413 // We can fetch the length directly through an AllocateArrayNode.
2414 // This works even if the length is not constant (clone or newArray).
2415 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2416 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2417 if (alloc != NULL) {
2418 Node* allocated_length = alloc->Ideal_length();
2419 // Do not allow make_ideal_length to allocate a CastII node.
2420 Node* len = alloc->make_ideal_length(tary, phase, false);
2421 if (allocated_length == len) {
2422 // Return allocated_length only if it would not be improved by a CastII.
2423 return allocated_length;
2424 }
2425 }
2426 }
2427
2428 return this;
2429
2430 }
2431
2432 //=============================================================================
2433 //---------------------------StoreNode::make-----------------------------------
2434 // Polymorphic factory method:
2435 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2436 assert((mo == unordered || mo == release), "unexpected");
2437 Compile* C = gvn.C;
2438 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2439 ctl != NULL, "raw memory operations should have control edge");
2440
2441 switch (bt) {
2442 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2443 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2444 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2445 case T_CHAR:
2446 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2447 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2448 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2449 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2450 case T_METADATA:
2451 case T_ADDRESS:
2452 case T_VALUETYPE:
2453 case T_OBJECT:
2454 #ifdef _LP64
2455 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2456 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2457 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2458 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2459 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2460 adr->bottom_type()->isa_rawptr())) {
2461 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2462 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2463 }
2464 #endif
2465 {
2466 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2467 }
2468 default:
2469 ShouldNotReachHere();
2470 return (StoreNode*)NULL;
2471 }
2472 }
2473
2474 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2475 bool require_atomic = true;
2476 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2477 }
2478
2479 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2480 bool require_atomic = true;
2481 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2482 }
2483
2484
2485 //--------------------------bottom_type----------------------------------------
2486 const Type *StoreNode::bottom_type() const {
2487 return Type::MEMORY;
2488 }
2489
2490 //------------------------------hash-------------------------------------------
2491 uint StoreNode::hash() const {
2492 // unroll addition of interesting fields
2493 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2494
2495 // Since they are not commoned, do not hash them:
2496 return NO_HASH;
2497 }
2498
2499 //------------------------------Ideal------------------------------------------
2500 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2501 // When a store immediately follows a relevant allocation/initialization,
2502 // try to capture it into the initialization, or hoist it above.
2503 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2504 Node* p = MemNode::Ideal_common(phase, can_reshape);
2505 if (p) return (p == NodeSentinel) ? NULL : p;
2506
2507 Node* mem = in(MemNode::Memory);
2508 Node* address = in(MemNode::Address);
2509 // Back-to-back stores to same address? Fold em up. Generally
2510 // unsafe if I have intervening uses... Also disallowed for StoreCM
2511 // since they must follow each StoreP operation. Redundant StoreCMs
2512 // are eliminated just before matching in final_graph_reshape.
2513 {
2514 Node* st = mem;
2515 // If Store 'st' has more than one use, we cannot fold 'st' away.
2516 // For example, 'st' might be the final state at a conditional
2517 // return. Or, 'st' might be used by some node which is live at
2518 // the same time 'st' is live, which might be unschedulable. So,
2519 // require exactly ONE user until such time as we clone 'mem' for
2520 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2521 // true).
2522 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2523 // Looking at a dead closed cycle of memory?
2524 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2525 assert(Opcode() == st->Opcode() ||
2526 st->Opcode() == Op_StoreVector ||
2527 Opcode() == Op_StoreVector ||
2528 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2529 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2530 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2531 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreN) ||
2532 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2533 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2534
2535 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2536 st->as_Store()->memory_size() <= this->memory_size()) {
2537 Node* use = st->raw_out(0);
2538 phase->igvn_rehash_node_delayed(use);
2539 if (can_reshape) {
2540 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2541 } else {
2542 // It's OK to do this in the parser, since DU info is always accurate,
2543 // and the parser always refers to nodes via SafePointNode maps.
2544 use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2545 }
2546 return this;
2547 }
2548 st = st->in(MemNode::Memory);
2549 }
2550 }
2551
2552
2553 // Capture an unaliased, unconditional, simple store into an initializer.
2554 // Or, if it is independent of the allocation, hoist it above the allocation.
2555 if (ReduceFieldZeroing && /*can_reshape &&*/
2556 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2557 InitializeNode* init = mem->in(0)->as_Initialize();
2558 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2559 if (offset > 0) {
2560 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2561 // If the InitializeNode captured me, it made a raw copy of me,
2562 // and I need to disappear.
2563 if (moved != NULL) {
2564 // %%% hack to ensure that Ideal returns a new node:
2565 mem = MergeMemNode::make(mem);
2566 return mem; // fold me away
2567 }
2568 }
2569 }
2570
2571 return NULL; // No further progress
2572 }
2573
2574 //------------------------------Value-----------------------------------------
2575 const Type* StoreNode::Value(PhaseGVN* phase) const {
2576 // Either input is TOP ==> the result is TOP
2577 const Type *t1 = phase->type( in(MemNode::Memory) );
2578 if( t1 == Type::TOP ) return Type::TOP;
2579 const Type *t2 = phase->type( in(MemNode::Address) );
2580 if( t2 == Type::TOP ) return Type::TOP;
2581 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2582 if( t3 == Type::TOP ) return Type::TOP;
2583 return Type::MEMORY;
2584 }
2585
2586 //------------------------------Identity---------------------------------------
2587 // Remove redundant stores:
2588 // Store(m, p, Load(m, p)) changes to m.
2589 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2590 Node* StoreNode::Identity(PhaseGVN* phase) {
2591 Node* mem = in(MemNode::Memory);
2592 Node* adr = in(MemNode::Address);
2593 Node* val = in(MemNode::ValueIn);
2594
2595 Node* result = this;
2596
2597 // Load then Store? Then the Store is useless
2598 if (val->is_Load() &&
2599 val->in(MemNode::Address)->eqv_uncast(adr) &&
2600 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2601 val->as_Load()->store_Opcode() == Opcode()) {
2602 result = mem;
2603 }
2604
2605 // Two stores in a row of the same value?
2606 if (result == this &&
2607 mem->is_Store() &&
2608 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2609 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2610 mem->Opcode() == Opcode()) {
2611 result = mem;
2612 }
2613
2614 // Store of zero anywhere into a freshly-allocated object?
2615 // Then the store is useless.
2616 // (It must already have been captured by the InitializeNode.)
2617 if (result == this && ReduceFieldZeroing) {
2618 // a newly allocated object is already all-zeroes everywhere
2619 if (mem->is_Proj() && mem->in(0)->is_Allocate() &&
2620 (phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == val)) {
2621 assert(!phase->type(val)->is_zero_type() || mem->in(0)->in(AllocateNode::DefaultValue) == NULL, "storing null to value array is forbidden");
2622 result = mem;
2623 }
2624
2625 if (result == this) {
2626 // the store may also apply to zero-bits in an earlier object
2627 Node* prev_mem = find_previous_store(phase);
2628 // Steps (a), (b): Walk past independent stores to find an exact match.
2629 if (prev_mem != NULL) {
2630 Node* prev_val = can_see_stored_value(prev_mem, phase);
2631 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2632 // prev_val and val might differ by a cast; it would be good
2633 // to keep the more informative of the two.
2634 if (phase->type(val)->is_zero_type()) {
2635 result = mem;
2636 } else if (prev_mem->is_Proj() && prev_mem->in(0)->is_Initialize()) {
2637 InitializeNode* init = prev_mem->in(0)->as_Initialize();
2638 AllocateNode* alloc = init->allocation();
2639 if (alloc != NULL && alloc->in(AllocateNode::DefaultValue) == val) {
2640 result = mem;
2641 }
2642 }
2643 }
2644 }
2645 }
2646 }
2647
2648 if (result != this && phase->is_IterGVN() != NULL) {
2649 MemBarNode* trailing = trailing_membar();
2650 if (trailing != NULL) {
2651 #ifdef ASSERT
2652 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2653 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects");
2654 #endif
2655 PhaseIterGVN* igvn = phase->is_IterGVN();
2656 trailing->remove(igvn);
2657 }
2658 }
2659
2660 return result;
2661 }
2662
2663 //------------------------------match_edge-------------------------------------
2664 // Do we Match on this edge index or not? Match only memory & value
2665 uint StoreNode::match_edge(uint idx) const {
2666 return idx == MemNode::Address || idx == MemNode::ValueIn;
2667 }
2668
2669 //------------------------------cmp--------------------------------------------
2670 // Do not common stores up together. They generally have to be split
2671 // back up anyways, so do not bother.
2672 uint StoreNode::cmp( const Node &n ) const {
2673 return (&n == this); // Always fail except on self
2674 }
2675
2676 //------------------------------Ideal_masked_input-----------------------------
2677 // Check for a useless mask before a partial-word store
2678 // (StoreB ... (AndI valIn conIa) )
2679 // If (conIa & mask == mask) this simplifies to
2680 // (StoreB ... (valIn) )
2681 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2682 Node *val = in(MemNode::ValueIn);
2683 if( val->Opcode() == Op_AndI ) {
2684 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2685 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2686 set_req(MemNode::ValueIn, val->in(1));
2687 return this;
2688 }
2689 }
2690 return NULL;
2691 }
2692
2693
2694 //------------------------------Ideal_sign_extended_input----------------------
2695 // Check for useless sign-extension before a partial-word store
2696 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2697 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2698 // (StoreB ... (valIn) )
2699 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2700 Node *val = in(MemNode::ValueIn);
2701 if( val->Opcode() == Op_RShiftI ) {
2702 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2703 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2704 Node *shl = val->in(1);
2705 if( shl->Opcode() == Op_LShiftI ) {
2706 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2707 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2708 set_req(MemNode::ValueIn, shl->in(1));
2709 return this;
2710 }
2711 }
2712 }
2713 }
2714 return NULL;
2715 }
2716
2717 //------------------------------value_never_loaded-----------------------------------
2718 // Determine whether there are any possible loads of the value stored.
2719 // For simplicity, we actually check if there are any loads from the
2720 // address stored to, not just for loads of the value stored by this node.
2721 //
2722 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2723 Node *adr = in(Address);
2724 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2725 if (adr_oop == NULL)
2726 return false;
2727 if (!adr_oop->is_known_instance_field())
2728 return false; // if not a distinct instance, there may be aliases of the address
2729 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2730 Node *use = adr->fast_out(i);
2731 if (use->is_Load() || use->is_LoadStore()) {
2732 return false;
2733 }
2734 }
2735 return true;
2736 }
2737
2738 MemBarNode* StoreNode::trailing_membar() const {
2739 if (is_release()) {
2740 MemBarNode* trailing_mb = NULL;
2741 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2742 Node* u = fast_out(i);
2743 if (u->is_MemBar()) {
2744 if (u->as_MemBar()->trailing_store()) {
2745 assert(u->Opcode() == Op_MemBarVolatile, "");
2746 assert(trailing_mb == NULL, "only one");
2747 trailing_mb = u->as_MemBar();
2748 #ifdef ASSERT
2749 Node* leading = u->as_MemBar()->leading_membar();
2750 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2751 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2752 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2753 #endif
2754 } else {
2755 assert(u->as_MemBar()->standalone(), "");
2756 }
2757 }
2758 }
2759 return trailing_mb;
2760 }
2761 return NULL;
2762 }
2763
2764
2765 //=============================================================================
2766 //------------------------------Ideal------------------------------------------
2767 // If the store is from an AND mask that leaves the low bits untouched, then
2768 // we can skip the AND operation. If the store is from a sign-extension
2769 // (a left shift, then right shift) we can skip both.
2770 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2771 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2772 if( progress != NULL ) return progress;
2773
2774 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2775 if( progress != NULL ) return progress;
2776
2777 // Finally check the default case
2778 return StoreNode::Ideal(phase, can_reshape);
2779 }
2780
2781 //=============================================================================
2782 //------------------------------Ideal------------------------------------------
2783 // If the store is from an AND mask that leaves the low bits untouched, then
2784 // we can skip the AND operation
2785 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2786 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2787 if( progress != NULL ) return progress;
2788
2789 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2790 if( progress != NULL ) return progress;
2791
2792 // Finally check the default case
2793 return StoreNode::Ideal(phase, can_reshape);
2794 }
2795
2796 //=============================================================================
2797 //------------------------------Identity---------------------------------------
2798 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2799 // No need to card mark when storing a null ptr
2800 Node* my_store = in(MemNode::OopStore);
2801 if (my_store->is_Store()) {
2802 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2803 if( t1 == TypePtr::NULL_PTR ) {
2804 return in(MemNode::Memory);
2805 }
2806 }
2807 return this;
2808 }
2809
2810 //=============================================================================
2811 //------------------------------Ideal---------------------------------------
2812 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2813 Node* progress = StoreNode::Ideal(phase, can_reshape);
2814 if (progress != NULL) return progress;
2815
2816 Node* my_store = in(MemNode::OopStore);
2817 if (my_store->is_MergeMem()) {
2818 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2819 set_req(MemNode::OopStore, mem);
2820 return this;
2821 }
2822
2823 return NULL;
2824 }
2825
2826 //------------------------------Value-----------------------------------------
2827 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2828 // Either input is TOP ==> the result is TOP
2829 const Type *t = phase->type( in(MemNode::Memory) );
2830 if( t == Type::TOP ) return Type::TOP;
2831 t = phase->type( in(MemNode::Address) );
2832 if( t == Type::TOP ) return Type::TOP;
2833 t = phase->type( in(MemNode::ValueIn) );
2834 if( t == Type::TOP ) return Type::TOP;
2835 // If extra input is TOP ==> the result is TOP
2836 t = phase->type( in(MemNode::OopStore) );
2837 if( t == Type::TOP ) return Type::TOP;
2838
2839 return StoreNode::Value( phase );
2840 }
2841
2842
2843 //=============================================================================
2844 //----------------------------------SCMemProjNode------------------------------
2845 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2846 {
2847 return bottom_type();
2848 }
2849
2850 //=============================================================================
2851 //----------------------------------LoadStoreNode------------------------------
2852 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2853 : Node(required),
2854 _type(rt),
2855 _adr_type(at)
2856 {
2857 init_req(MemNode::Control, c );
2858 init_req(MemNode::Memory , mem);
2859 init_req(MemNode::Address, adr);
2860 init_req(MemNode::ValueIn, val);
2861 init_class_id(Class_LoadStore);
2862 }
2863
2864 uint LoadStoreNode::ideal_reg() const {
2865 return _type->ideal_reg();
2866 }
2867
2868 bool LoadStoreNode::result_not_used() const {
2869 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2870 Node *x = fast_out(i);
2871 if (x->Opcode() == Op_SCMemProj) continue;
2872 return false;
2873 }
2874 return true;
2875 }
2876
2877 MemBarNode* LoadStoreNode::trailing_membar() const {
2878 MemBarNode* trailing = NULL;
2879 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2880 Node* u = fast_out(i);
2881 if (u->is_MemBar()) {
2882 if (u->as_MemBar()->trailing_load_store()) {
2883 assert(u->Opcode() == Op_MemBarAcquire, "");
2884 assert(trailing == NULL, "only one");
2885 trailing = u->as_MemBar();
2886 #ifdef ASSERT
2887 Node* leading = trailing->leading_membar();
2888 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2889 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
2890 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
2891 #endif
2892 } else {
2893 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
2894 }
2895 }
2896 }
2897
2898 return trailing;
2899 }
2900
2901 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2902
2903 //=============================================================================
2904 //----------------------------------LoadStoreConditionalNode--------------------
2905 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2906 init_req(ExpectedIn, ex );
2907 }
2908
2909 //=============================================================================
2910 //-------------------------------adr_type--------------------------------------
2911 const TypePtr* ClearArrayNode::adr_type() const {
2912 Node *adr = in(3);
2913 if (adr == NULL) return NULL; // node is dead
2914 return MemNode::calculate_adr_type(adr->bottom_type());
2915 }
2916
2917 //------------------------------match_edge-------------------------------------
2918 // Do we Match on this edge index or not? Do not match memory
2919 uint ClearArrayNode::match_edge(uint idx) const {
2920 return idx > 1;
2921 }
2922
2923 //------------------------------Identity---------------------------------------
2924 // Clearing a zero length array does nothing
2925 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
2926 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2927 }
2928
2929 //------------------------------Idealize---------------------------------------
2930 // Clearing a short array is faster with stores
2931 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2932 // Already know this is a large node, do not try to ideal it
2933 if (!IdealizeClearArrayNode || _is_large) return NULL;
2934
2935 const int unit = BytesPerLong;
2936 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2937 if (!t) return NULL;
2938 if (!t->is_con()) return NULL;
2939 intptr_t raw_count = t->get_con();
2940 intptr_t size = raw_count;
2941 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2942 // Clearing nothing uses the Identity call.
2943 // Negative clears are possible on dead ClearArrays
2944 // (see jck test stmt114.stmt11402.val).
2945 if (size <= 0 || size % unit != 0) return NULL;
2946 intptr_t count = size / unit;
2947 // Length too long; communicate this to matchers and assemblers.
2948 // Assemblers are responsible to produce fast hardware clears for it.
2949 if (size > InitArrayShortSize) {
2950 return new ClearArrayNode(in(0), in(1), in(2), in(3), in(4), true);
2951 }
2952 Node *mem = in(1);
2953 if( phase->type(mem)==Type::TOP ) return NULL;
2954 Node *adr = in(3);
2955 const Type* at = phase->type(adr);
2956 if( at==Type::TOP ) return NULL;
2957 const TypePtr* atp = at->isa_ptr();
2958 // adjust atp to be the correct array element address type
2959 if (atp == NULL) atp = TypePtr::BOTTOM;
2960 else atp = atp->add_offset(Type::OffsetBot);
2961 // Get base for derived pointer purposes
2962 if( adr->Opcode() != Op_AddP ) Unimplemented();
2963 Node *base = adr->in(1);
2964
2965 Node *val = in(4);
2966 Node *off = phase->MakeConX(BytesPerLong);
2967 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
2968 count--;
2969 while( count-- ) {
2970 mem = phase->transform(mem);
2971 adr = phase->transform(new AddPNode(base,adr,off));
2972 mem = new StoreLNode(in(0), mem, adr, atp, val, MemNode::unordered, false);
2973 }
2974 return mem;
2975 }
2976
2977 //----------------------------step_through----------------------------------
2978 // Return allocation input memory edge if it is different instance
2979 // or itself if it is the one we are looking for.
2980 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2981 Node* n = *np;
2982 assert(n->is_ClearArray(), "sanity");
2983 intptr_t offset;
2984 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2985 // This method is called only before Allocate nodes are expanded
2986 // during macro nodes expansion. Before that ClearArray nodes are
2987 // only generated in PhaseMacroExpand::generate_arraycopy() (before
2988 // Allocate nodes are expanded) which follows allocations.
2989 assert(alloc != NULL, "should have allocation");
2990 if (alloc->_idx == instance_id) {
2991 // Can not bypass initialization of the instance we are looking for.
2992 return false;
2993 }
2994 // Otherwise skip it.
2995 InitializeNode* init = alloc->initialization();
2996 if (init != NULL)
2997 *np = init->in(TypeFunc::Memory);
2998 else
2999 *np = alloc->in(TypeFunc::Memory);
3000 return true;
3001 }
3002
3003 //----------------------------clear_memory-------------------------------------
3004 // Generate code to initialize object storage to zero.
3005 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3006 Node* val,
3007 Node* raw_val,
3008 intptr_t start_offset,
3009 Node* end_offset,
3010 PhaseGVN* phase) {
3011 intptr_t offset = start_offset;
3012
3013 int unit = BytesPerLong;
3014 if ((offset % unit) != 0) {
3015 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
3016 adr = phase->transform(adr);
3017 const TypePtr* atp = TypeRawPtr::BOTTOM;
3018 if (val != NULL) {
3019 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
3020 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
3021 } else {
3022 assert(raw_val == NULL, "val may not be null");
3023 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3024 }
3025 mem = phase->transform(mem);
3026 offset += BytesPerInt;
3027 }
3028 assert((offset % unit) == 0, "");
3029
3030 // Initialize the remaining stuff, if any, with a ClearArray.
3031 return clear_memory(ctl, mem, dest, raw_val, phase->MakeConX(offset), end_offset, phase);
3032 }
3033
3034 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3035 Node* raw_val,
3036 Node* start_offset,
3037 Node* end_offset,
3038 PhaseGVN* phase) {
3039 if (start_offset == end_offset) {
3040 // nothing to do
3041 return mem;
3042 }
3043
3044 int unit = BytesPerLong;
3045 Node* zbase = start_offset;
3046 Node* zend = end_offset;
3047
3048 // Scale to the unit required by the CPU:
3049 if (!Matcher::init_array_count_is_in_bytes) {
3050 Node* shift = phase->intcon(exact_log2(unit));
3051 zbase = phase->transform(new URShiftXNode(zbase, shift) );
3052 zend = phase->transform(new URShiftXNode(zend, shift) );
3053 }
3054
3055 // Bulk clear double-words
3056 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
3057 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
3058 if (raw_val == NULL) {
3059 raw_val = phase->MakeConX(0);
3060 }
3061 mem = new ClearArrayNode(ctl, mem, zsize, adr, raw_val, false);
3062 return phase->transform(mem);
3063 }
3064
3065 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3066 Node* val,
3067 Node* raw_val,
3068 intptr_t start_offset,
3069 intptr_t end_offset,
3070 PhaseGVN* phase) {
3071 if (start_offset == end_offset) {
3072 // nothing to do
3073 return mem;
3074 }
3075
3076 assert((end_offset % BytesPerInt) == 0, "odd end offset");
3077 intptr_t done_offset = end_offset;
3078 if ((done_offset % BytesPerLong) != 0) {
3079 done_offset -= BytesPerInt;
3080 }
3081 if (done_offset > start_offset) {
3082 mem = clear_memory(ctl, mem, dest, val, raw_val,
3083 start_offset, phase->MakeConX(done_offset), phase);
3084 }
3085 if (done_offset < end_offset) { // emit the final 32-bit store
3086 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3087 adr = phase->transform(adr);
3088 const TypePtr* atp = TypeRawPtr::BOTTOM;
3089 if (val != NULL) {
3090 assert(phase->type(val)->isa_narrowoop(), "should be narrow oop");
3091 mem = new StoreNNode(ctl, mem, adr, atp, val, MemNode::unordered);
3092 } else {
3093 assert(raw_val == NULL, "val may not be null");
3094 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3095 }
3096 mem = phase->transform(mem);
3097 done_offset += BytesPerInt;
3098 }
3099 assert(done_offset == end_offset, "");
3100 return mem;
3101 }
3102
3103 //=============================================================================
3104 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3105 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
3106 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3107 #ifdef ASSERT
3108 , _pair_idx(0)
3109 #endif
3110 {
3111 init_class_id(Class_MemBar);
3112 Node* top = C->top();
3113 init_req(TypeFunc::I_O,top);
3114 init_req(TypeFunc::FramePtr,top);
3115 init_req(TypeFunc::ReturnAdr,top);
3116 if (precedent != NULL)
3117 init_req(TypeFunc::Parms, precedent);
3118 }
3119
3120 //------------------------------cmp--------------------------------------------
3121 uint MemBarNode::hash() const { return NO_HASH; }
3122 uint MemBarNode::cmp( const Node &n ) const {
3123 return (&n == this); // Always fail except on self
3124 }
3125
3126 //------------------------------make-------------------------------------------
3127 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3128 switch (opcode) {
3129 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
3130 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
3131 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
3132 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
3133 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3134 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3135 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
3136 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
3137 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
3138 case Op_Initialize: return new InitializeNode(C, atp, pn);
3139 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
3140 default: ShouldNotReachHere(); return NULL;
3141 }
3142 }
3143
3144 void MemBarNode::remove(PhaseIterGVN *igvn) {
3145 if (outcnt() != 2) {
3146 return;
3147 }
3148 if (trailing_store() || trailing_load_store()) {
3149 MemBarNode* leading = leading_membar();
3150 if (leading != NULL) {
3151 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3152 leading->remove(igvn);
3153 }
3154 }
3155 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3156 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3157 }
3158
3159 //------------------------------Ideal------------------------------------------
3160 // Return a node which is more "ideal" than the current node. Strip out
3161 // control copies
3162 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3163 if (remove_dead_region(phase, can_reshape)) return this;
3164 // Don't bother trying to transform a dead node
3165 if (in(0) && in(0)->is_top()) {
3166 return NULL;
3167 }
3168
3169 #if INCLUDE_ZGC
3170 if (UseZGC) {
3171 if (req() == (Precedent+1) && in(MemBarNode::Precedent)->in(0) != NULL && in(MemBarNode::Precedent)->in(0)->is_LoadBarrier()) {
3172 Node* load_node = in(MemBarNode::Precedent)->in(0)->in(LoadBarrierNode::Oop);
3173 set_req(MemBarNode::Precedent, load_node);
3174 return this;
3175 }
3176 }
3177 #endif
3178
3179 bool progress = false;
3180 // Eliminate volatile MemBars for scalar replaced objects.
3181 if (can_reshape && req() == (Precedent+1)) {
3182 bool eliminate = false;
3183 int opc = Opcode();
3184 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3185 // Volatile field loads and stores.
3186 Node* my_mem = in(MemBarNode::Precedent);
3187 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3188 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3189 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3190 // replace this Precedent (decodeN) with the Load instead.
3191 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3192 Node* load_node = my_mem->in(1);
3193 set_req(MemBarNode::Precedent, load_node);
3194 phase->is_IterGVN()->_worklist.push(my_mem);
3195 my_mem = load_node;
3196 } else {
3197 assert(my_mem->unique_out() == this, "sanity");
3198 del_req(Precedent);
3199 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3200 my_mem = NULL;
3201 }
3202 progress = true;
3203 }
3204 if (my_mem != NULL && my_mem->is_Mem()) {
3205 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3206 // Check for scalar replaced object reference.
3207 if( t_oop != NULL && t_oop->is_known_instance_field() &&
3208 t_oop->offset() != Type::OffsetBot &&
3209 t_oop->offset() != Type::OffsetTop) {
3210 eliminate = true;
3211 }
3212 }
3213 } else if (opc == Op_MemBarRelease) {
3214 // Final field stores.
3215 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3216 if ((alloc != NULL) && alloc->is_Allocate() &&
3217 alloc->as_Allocate()->does_not_escape_thread()) {
3218 // The allocated object does not escape.
3219 eliminate = true;
3220 }
3221 }
3222 if (eliminate) {
3223 // Replace MemBar projections by its inputs.
3224 PhaseIterGVN* igvn = phase->is_IterGVN();
3225 remove(igvn);
3226 // Must return either the original node (now dead) or a new node
3227 // (Do not return a top here, since that would break the uniqueness of top.)
3228 return new ConINode(TypeInt::ZERO);
3229 }
3230 }
3231 return progress ? this : NULL;
3232 }
3233
3234 //------------------------------Value------------------------------------------
3235 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3236 if( !in(0) ) return Type::TOP;
3237 if( phase->type(in(0)) == Type::TOP )
3238 return Type::TOP;
3239 return TypeTuple::MEMBAR;
3240 }
3241
3242 //------------------------------match------------------------------------------
3243 // Construct projections for memory.
3244 Node *MemBarNode::match(const ProjNode *proj, const Matcher *m, const RegMask* mask) {
3245 switch (proj->_con) {
3246 case TypeFunc::Control:
3247 case TypeFunc::Memory:
3248 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3249 }
3250 ShouldNotReachHere();
3251 return NULL;
3252 }
3253
3254 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3255 trailing->_kind = TrailingStore;
3256 leading->_kind = LeadingStore;
3257 #ifdef ASSERT
3258 trailing->_pair_idx = leading->_idx;
3259 leading->_pair_idx = leading->_idx;
3260 #endif
3261 }
3262
3263 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3264 trailing->_kind = TrailingLoadStore;
3265 leading->_kind = LeadingLoadStore;
3266 #ifdef ASSERT
3267 trailing->_pair_idx = leading->_idx;
3268 leading->_pair_idx = leading->_idx;
3269 #endif
3270 }
3271
3272 MemBarNode* MemBarNode::trailing_membar() const {
3273 Node* trailing = (Node*)this;
3274 VectorSet seen(Thread::current()->resource_area());
3275 while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing()) {
3276 if (seen.test_set(trailing->_idx)) {
3277 // Dying subgraph?
3278 return NULL;
3279 }
3280 for (DUIterator_Fast jmax, j = trailing->fast_outs(jmax); j < jmax; j++) {
3281 Node* next = trailing->fast_out(j);
3282 if (next != trailing && next->is_CFG()) {
3283 trailing = next;
3284 break;
3285 }
3286 }
3287 }
3288 MemBarNode* mb = trailing->as_MemBar();
3289 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3290 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3291 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3292 return mb;
3293 }
3294
3295 MemBarNode* MemBarNode::leading_membar() const {
3296 VectorSet seen(Thread::current()->resource_area());
3297 Node* leading = in(0);
3298 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3299 if (seen.test_set(leading->_idx)) {
3300 // Dying subgraph?
3301 return NULL;
3302 }
3303 if (leading->is_Region()) {
3304 leading = leading->in(1);
3305 } else {
3306 leading = leading->in(0);
3307 }
3308 }
3309 #ifdef ASSERT
3310 Unique_Node_List wq;
3311 wq.push((Node*)this);
3312 uint found = 0;
3313 for (uint i = 0; i < wq.size(); i++) {
3314 Node* n = wq.at(i);
3315 if (n->is_Region()) {
3316 for (uint j = 1; j < n->req(); j++) {
3317 Node* in = n->in(j);
3318 if (in != NULL && !in->is_top()) {
3319 wq.push(in);
3320 }
3321 }
3322 } else {
3323 if (n->is_MemBar() && n->as_MemBar()->leading()) {
3324 assert(n == leading, "consistency check failed");
3325 found++;
3326 } else {
3327 Node* in = n->in(0);
3328 if (in != NULL && !in->is_top()) {
3329 wq.push(in);
3330 }
3331 }
3332 }
3333 }
3334 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed");
3335 #endif
3336 if (leading == NULL) {
3337 return NULL;
3338 }
3339 MemBarNode* mb = leading->as_MemBar();
3340 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3341 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3342 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3343 return mb;
3344 }
3345
3346 //===========================InitializeNode====================================
3347 // SUMMARY:
3348 // This node acts as a memory barrier on raw memory, after some raw stores.
3349 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3350 // The Initialize can 'capture' suitably constrained stores as raw inits.
3351 // It can coalesce related raw stores into larger units (called 'tiles').
3352 // It can avoid zeroing new storage for memory units which have raw inits.
3353 // At macro-expansion, it is marked 'complete', and does not optimize further.
3354 //
3355 // EXAMPLE:
3356 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3357 // ctl = incoming control; mem* = incoming memory
3358 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3359 // First allocate uninitialized memory and fill in the header:
3360 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3361 // ctl := alloc.Control; mem* := alloc.Memory*
3362 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3363 // Then initialize to zero the non-header parts of the raw memory block:
3364 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3365 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3366 // After the initialize node executes, the object is ready for service:
3367 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3368 // Suppose its body is immediately initialized as {1,2}:
3369 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3370 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3371 // mem.SLICE(#short[*]) := store2
3372 //
3373 // DETAILS:
3374 // An InitializeNode collects and isolates object initialization after
3375 // an AllocateNode and before the next possible safepoint. As a
3376 // memory barrier (MemBarNode), it keeps critical stores from drifting
3377 // down past any safepoint or any publication of the allocation.
3378 // Before this barrier, a newly-allocated object may have uninitialized bits.
3379 // After this barrier, it may be treated as a real oop, and GC is allowed.
3380 //
3381 // The semantics of the InitializeNode include an implicit zeroing of
3382 // the new object from object header to the end of the object.
3383 // (The object header and end are determined by the AllocateNode.)
3384 //
3385 // Certain stores may be added as direct inputs to the InitializeNode.
3386 // These stores must update raw memory, and they must be to addresses
3387 // derived from the raw address produced by AllocateNode, and with
3388 // a constant offset. They must be ordered by increasing offset.
3389 // The first one is at in(RawStores), the last at in(req()-1).
3390 // Unlike most memory operations, they are not linked in a chain,
3391 // but are displayed in parallel as users of the rawmem output of
3392 // the allocation.
3393 //
3394 // (See comments in InitializeNode::capture_store, which continue
3395 // the example given above.)
3396 //
3397 // When the associated Allocate is macro-expanded, the InitializeNode
3398 // may be rewritten to optimize collected stores. A ClearArrayNode
3399 // may also be created at that point to represent any required zeroing.
3400 // The InitializeNode is then marked 'complete', prohibiting further
3401 // capturing of nearby memory operations.
3402 //
3403 // During macro-expansion, all captured initializations which store
3404 // constant values of 32 bits or smaller are coalesced (if advantageous)
3405 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3406 // initialized in fewer memory operations. Memory words which are
3407 // covered by neither tiles nor non-constant stores are pre-zeroed
3408 // by explicit stores of zero. (The code shape happens to do all
3409 // zeroing first, then all other stores, with both sequences occurring
3410 // in order of ascending offsets.)
3411 //
3412 // Alternatively, code may be inserted between an AllocateNode and its
3413 // InitializeNode, to perform arbitrary initialization of the new object.
3414 // E.g., the object copying intrinsics insert complex data transfers here.
3415 // The initialization must then be marked as 'complete' disable the
3416 // built-in zeroing semantics and the collection of initializing stores.
3417 //
3418 // While an InitializeNode is incomplete, reads from the memory state
3419 // produced by it are optimizable if they match the control edge and
3420 // new oop address associated with the allocation/initialization.
3421 // They return a stored value (if the offset matches) or else zero.
3422 // A write to the memory state, if it matches control and address,
3423 // and if it is to a constant offset, may be 'captured' by the
3424 // InitializeNode. It is cloned as a raw memory operation and rewired
3425 // inside the initialization, to the raw oop produced by the allocation.
3426 // Operations on addresses which are provably distinct (e.g., to
3427 // other AllocateNodes) are allowed to bypass the initialization.
3428 //
3429 // The effect of all this is to consolidate object initialization
3430 // (both arrays and non-arrays, both piecewise and bulk) into a
3431 // single location, where it can be optimized as a unit.
3432 //
3433 // Only stores with an offset less than TrackedInitializationLimit words
3434 // will be considered for capture by an InitializeNode. This puts a
3435 // reasonable limit on the complexity of optimized initializations.
3436
3437 //---------------------------InitializeNode------------------------------------
3438 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3439 : MemBarNode(C, adr_type, rawoop),
3440 _is_complete(Incomplete), _does_not_escape(false)
3441 {
3442 init_class_id(Class_Initialize);
3443
3444 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3445 assert(in(RawAddress) == rawoop, "proper init");
3446 // Note: allocation() can be NULL, for secondary initialization barriers
3447 }
3448
3449 // Since this node is not matched, it will be processed by the
3450 // register allocator. Declare that there are no constraints
3451 // on the allocation of the RawAddress edge.
3452 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3453 // This edge should be set to top, by the set_complete. But be conservative.
3454 if (idx == InitializeNode::RawAddress)
3455 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3456 return RegMask::Empty;
3457 }
3458
3459 Node* InitializeNode::memory(uint alias_idx) {
3460 Node* mem = in(Memory);
3461 if (mem->is_MergeMem()) {
3462 return mem->as_MergeMem()->memory_at(alias_idx);
3463 } else {
3464 // incoming raw memory is not split
3465 return mem;
3466 }
3467 }
3468
3469 bool InitializeNode::is_non_zero() {
3470 if (is_complete()) return false;
3471 remove_extra_zeroes();
3472 return (req() > RawStores);
3473 }
3474
3475 void InitializeNode::set_complete(PhaseGVN* phase) {
3476 assert(!is_complete(), "caller responsibility");
3477 _is_complete = Complete;
3478
3479 // After this node is complete, it contains a bunch of
3480 // raw-memory initializations. There is no need for
3481 // it to have anything to do with non-raw memory effects.
3482 // Therefore, tell all non-raw users to re-optimize themselves,
3483 // after skipping the memory effects of this initialization.
3484 PhaseIterGVN* igvn = phase->is_IterGVN();
3485 if (igvn) igvn->add_users_to_worklist(this);
3486 }
3487
3488 // convenience function
3489 // return false if the init contains any stores already
3490 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3491 InitializeNode* init = initialization();
3492 if (init == NULL || init->is_complete()) {
3493 return false;
3494 }
3495 init->remove_extra_zeroes();
3496 // for now, if this allocation has already collected any inits, bail:
3497 if (init->is_non_zero()) return false;
3498 init->set_complete(phase);
3499 return true;
3500 }
3501
3502 void InitializeNode::remove_extra_zeroes() {
3503 if (req() == RawStores) return;
3504 Node* zmem = zero_memory();
3505 uint fill = RawStores;
3506 for (uint i = fill; i < req(); i++) {
3507 Node* n = in(i);
3508 if (n->is_top() || n == zmem) continue; // skip
3509 if (fill < i) set_req(fill, n); // compact
3510 ++fill;
3511 }
3512 // delete any empty spaces created:
3513 while (fill < req()) {
3514 del_req(fill);
3515 }
3516 }
3517
3518 // Helper for remembering which stores go with which offsets.
3519 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3520 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3521 intptr_t offset = -1;
3522 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3523 phase, offset);
3524 if (base == NULL) return -1; // something is dead,
3525 if (offset < 0) return -1; // dead, dead
3526 return offset;
3527 }
3528
3529 // Helper for proving that an initialization expression is
3530 // "simple enough" to be folded into an object initialization.
3531 // Attempts to prove that a store's initial value 'n' can be captured
3532 // within the initialization without creating a vicious cycle, such as:
3533 // { Foo p = new Foo(); p.next = p; }
3534 // True for constants and parameters and small combinations thereof.
3535 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3536 if (n == NULL) return true; // (can this really happen?)
3537 if (n->is_Proj()) n = n->in(0);
3538 if (n == this) return false; // found a cycle
3539 if (n->is_Con()) return true;
3540 if (n->is_Start()) return true; // params, etc., are OK
3541 if (n->is_Root()) return true; // even better
3542
3543 Node* ctl = n->in(0);
3544 if (ctl != NULL && !ctl->is_top()) {
3545 if (ctl->is_Proj()) ctl = ctl->in(0);
3546 if (ctl == this) return false;
3547
3548 // If we already know that the enclosing memory op is pinned right after
3549 // the init, then any control flow that the store has picked up
3550 // must have preceded the init, or else be equal to the init.
3551 // Even after loop optimizations (which might change control edges)
3552 // a store is never pinned *before* the availability of its inputs.
3553 if (!MemNode::all_controls_dominate(n, this))
3554 return false; // failed to prove a good control
3555 }
3556
3557 // Check data edges for possible dependencies on 'this'.
3558 if ((count += 1) > 20) return false; // complexity limit
3559 for (uint i = 1; i < n->req(); i++) {
3560 Node* m = n->in(i);
3561 if (m == NULL || m == n || m->is_top()) continue;
3562 uint first_i = n->find_edge(m);
3563 if (i != first_i) continue; // process duplicate edge just once
3564 if (!detect_init_independence(m, count)) {
3565 return false;
3566 }
3567 }
3568
3569 return true;
3570 }
3571
3572 // Here are all the checks a Store must pass before it can be moved into
3573 // an initialization. Returns zero if a check fails.
3574 // On success, returns the (constant) offset to which the store applies,
3575 // within the initialized memory.
3576 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3577 const int FAIL = 0;
3578 if (st->is_unaligned_access()) {
3579 return FAIL;
3580 }
3581 if (st->req() != MemNode::ValueIn + 1)
3582 return FAIL; // an inscrutable StoreNode (card mark?)
3583 Node* ctl = st->in(MemNode::Control);
3584 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3585 return FAIL; // must be unconditional after the initialization
3586 Node* mem = st->in(MemNode::Memory);
3587 if (!(mem->is_Proj() && mem->in(0) == this))
3588 return FAIL; // must not be preceded by other stores
3589 Node* adr = st->in(MemNode::Address);
3590 intptr_t offset;
3591 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3592 if (alloc == NULL)
3593 return FAIL; // inscrutable address
3594 if (alloc != allocation())
3595 return FAIL; // wrong allocation! (store needs to float up)
3596 Node* val = st->in(MemNode::ValueIn);
3597 int complexity_count = 0;
3598 if (!detect_init_independence(val, complexity_count))
3599 return FAIL; // stored value must be 'simple enough'
3600
3601 // The Store can be captured only if nothing after the allocation
3602 // and before the Store is using the memory location that the store
3603 // overwrites.
3604 bool failed = false;
3605 // If is_complete_with_arraycopy() is true the shape of the graph is
3606 // well defined and is safe so no need for extra checks.
3607 if (!is_complete_with_arraycopy()) {
3608 // We are going to look at each use of the memory state following
3609 // the allocation to make sure nothing reads the memory that the
3610 // Store writes.
3611 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3612 int alias_idx = phase->C->get_alias_index(t_adr);
3613 ResourceMark rm;
3614 Unique_Node_List mems;
3615 mems.push(mem);
3616 Node* unique_merge = NULL;
3617 for (uint next = 0; next < mems.size(); ++next) {
3618 Node *m = mems.at(next);
3619 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3620 Node *n = m->fast_out(j);
3621 if (n->outcnt() == 0) {
3622 continue;
3623 }
3624 if (n == st) {
3625 continue;
3626 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3627 // If the control of this use is different from the control
3628 // of the Store which is right after the InitializeNode then
3629 // this node cannot be between the InitializeNode and the
3630 // Store.
3631 continue;
3632 } else if (n->is_MergeMem()) {
3633 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3634 // We can hit a MergeMemNode (that will likely go away
3635 // later) that is a direct use of the memory state
3636 // following the InitializeNode on the same slice as the
3637 // store node that we'd like to capture. We need to check
3638 // the uses of the MergeMemNode.
3639 mems.push(n);
3640 }
3641 } else if (n->is_Mem()) {
3642 Node* other_adr = n->in(MemNode::Address);
3643 if (other_adr == adr) {
3644 failed = true;
3645 break;
3646 } else {
3647 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3648 if (other_t_adr != NULL) {
3649 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3650 if (other_alias_idx == alias_idx) {
3651 // A load from the same memory slice as the store right
3652 // after the InitializeNode. We check the control of the
3653 // object/array that is loaded from. If it's the same as
3654 // the store control then we cannot capture the store.
3655 assert(!n->is_Store(), "2 stores to same slice on same control?");
3656 Node* base = other_adr;
3657 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3658 base = base->in(AddPNode::Base);
3659 if (base != NULL) {
3660 base = base->uncast();
3661 if (base->is_Proj() && base->in(0) == alloc) {
3662 failed = true;
3663 break;
3664 }
3665 }
3666 }
3667 }
3668 }
3669 } else {
3670 failed = true;
3671 break;
3672 }
3673 }
3674 }
3675 }
3676 if (failed) {
3677 if (!can_reshape) {
3678 // We decided we couldn't capture the store during parsing. We
3679 // should try again during the next IGVN once the graph is
3680 // cleaner.
3681 phase->C->record_for_igvn(st);
3682 }
3683 return FAIL;
3684 }
3685
3686 return offset; // success
3687 }
3688
3689 // Find the captured store in(i) which corresponds to the range
3690 // [start..start+size) in the initialized object.
3691 // If there is one, return its index i. If there isn't, return the
3692 // negative of the index where it should be inserted.
3693 // Return 0 if the queried range overlaps an initialization boundary
3694 // or if dead code is encountered.
3695 // If size_in_bytes is zero, do not bother with overlap checks.
3696 int InitializeNode::captured_store_insertion_point(intptr_t start,
3697 int size_in_bytes,
3698 PhaseTransform* phase) {
3699 const int FAIL = 0, MAX_STORE = BytesPerLong;
3700
3701 if (is_complete())
3702 return FAIL; // arraycopy got here first; punt
3703
3704 assert(allocation() != NULL, "must be present");
3705
3706 // no negatives, no header fields:
3707 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3708
3709 // after a certain size, we bail out on tracking all the stores:
3710 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3711 if (start >= ti_limit) return FAIL;
3712
3713 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3714 if (i >= limit) return -(int)i; // not found; here is where to put it
3715
3716 Node* st = in(i);
3717 intptr_t st_off = get_store_offset(st, phase);
3718 if (st_off < 0) {
3719 if (st != zero_memory()) {
3720 return FAIL; // bail out if there is dead garbage
3721 }
3722 } else if (st_off > start) {
3723 // ...we are done, since stores are ordered
3724 if (st_off < start + size_in_bytes) {
3725 return FAIL; // the next store overlaps
3726 }
3727 return -(int)i; // not found; here is where to put it
3728 } else if (st_off < start) {
3729 if (size_in_bytes != 0 &&
3730 start < st_off + MAX_STORE &&
3731 start < st_off + st->as_Store()->memory_size()) {
3732 return FAIL; // the previous store overlaps
3733 }
3734 } else {
3735 if (size_in_bytes != 0 &&
3736 st->as_Store()->memory_size() != size_in_bytes) {
3737 return FAIL; // mismatched store size
3738 }
3739 return i;
3740 }
3741
3742 ++i;
3743 }
3744 }
3745
3746 // Look for a captured store which initializes at the offset 'start'
3747 // with the given size. If there is no such store, and no other
3748 // initialization interferes, then return zero_memory (the memory
3749 // projection of the AllocateNode).
3750 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3751 PhaseTransform* phase) {
3752 assert(stores_are_sane(phase), "");
3753 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3754 if (i == 0) {
3755 return NULL; // something is dead
3756 } else if (i < 0) {
3757 return zero_memory(); // just primordial zero bits here
3758 } else {
3759 Node* st = in(i); // here is the store at this position
3760 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3761 return st;
3762 }
3763 }
3764
3765 // Create, as a raw pointer, an address within my new object at 'offset'.
3766 Node* InitializeNode::make_raw_address(intptr_t offset,
3767 PhaseTransform* phase) {
3768 Node* addr = in(RawAddress);
3769 if (offset != 0) {
3770 Compile* C = phase->C;
3771 addr = phase->transform( new AddPNode(C->top(), addr,
3772 phase->MakeConX(offset)) );
3773 }
3774 return addr;
3775 }
3776
3777 // Clone the given store, converting it into a raw store
3778 // initializing a field or element of my new object.
3779 // Caller is responsible for retiring the original store,
3780 // with subsume_node or the like.
3781 //
3782 // From the example above InitializeNode::InitializeNode,
3783 // here are the old stores to be captured:
3784 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3785 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3786 //
3787 // Here is the changed code; note the extra edges on init:
3788 // alloc = (Allocate ...)
3789 // rawoop = alloc.RawAddress
3790 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3791 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3792 // init = (Initialize alloc.Control alloc.Memory rawoop
3793 // rawstore1 rawstore2)
3794 //
3795 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3796 PhaseTransform* phase, bool can_reshape) {
3797 assert(stores_are_sane(phase), "");
3798
3799 if (start < 0) return NULL;
3800 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3801
3802 Compile* C = phase->C;
3803 int size_in_bytes = st->memory_size();
3804 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3805 if (i == 0) return NULL; // bail out
3806 Node* prev_mem = NULL; // raw memory for the captured store
3807 if (i > 0) {
3808 prev_mem = in(i); // there is a pre-existing store under this one
3809 set_req(i, C->top()); // temporarily disconnect it
3810 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3811 } else {
3812 i = -i; // no pre-existing store
3813 prev_mem = zero_memory(); // a slice of the newly allocated object
3814 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3815 set_req(--i, C->top()); // reuse this edge; it has been folded away
3816 else
3817 ins_req(i, C->top()); // build a new edge
3818 }
3819 Node* new_st = st->clone();
3820 new_st->set_req(MemNode::Control, in(Control));
3821 new_st->set_req(MemNode::Memory, prev_mem);
3822 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3823 new_st = phase->transform(new_st);
3824
3825 // At this point, new_st might have swallowed a pre-existing store
3826 // at the same offset, or perhaps new_st might have disappeared,
3827 // if it redundantly stored the same value (or zero to fresh memory).
3828
3829 // In any case, wire it in:
3830 phase->igvn_rehash_node_delayed(this);
3831 set_req(i, new_st);
3832
3833 // The caller may now kill the old guy.
3834 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3835 assert(check_st == new_st || check_st == NULL, "must be findable");
3836 assert(!is_complete(), "");
3837 return new_st;
3838 }
3839
3840 static bool store_constant(jlong* tiles, int num_tiles,
3841 intptr_t st_off, int st_size,
3842 jlong con) {
3843 if ((st_off & (st_size-1)) != 0)
3844 return false; // strange store offset (assume size==2**N)
3845 address addr = (address)tiles + st_off;
3846 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3847 switch (st_size) {
3848 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3849 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3850 case sizeof(jint): *(jint*) addr = (jint) con; break;
3851 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3852 default: return false; // strange store size (detect size!=2**N here)
3853 }
3854 return true; // return success to caller
3855 }
3856
3857 // Coalesce subword constants into int constants and possibly
3858 // into long constants. The goal, if the CPU permits,
3859 // is to initialize the object with a small number of 64-bit tiles.
3860 // Also, convert floating-point constants to bit patterns.
3861 // Non-constants are not relevant to this pass.
3862 //
3863 // In terms of the running example on InitializeNode::InitializeNode
3864 // and InitializeNode::capture_store, here is the transformation
3865 // of rawstore1 and rawstore2 into rawstore12:
3866 // alloc = (Allocate ...)
3867 // rawoop = alloc.RawAddress
3868 // tile12 = 0x00010002
3869 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3870 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3871 //
3872 void
3873 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3874 Node* size_in_bytes,
3875 PhaseGVN* phase) {
3876 Compile* C = phase->C;
3877
3878 assert(stores_are_sane(phase), "");
3879 // Note: After this pass, they are not completely sane,
3880 // since there may be some overlaps.
3881
3882 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3883
3884 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3885 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3886 size_limit = MIN2(size_limit, ti_limit);
3887 size_limit = align_up(size_limit, BytesPerLong);
3888 int num_tiles = size_limit / BytesPerLong;
3889
3890 // allocate space for the tile map:
3891 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3892 jlong tiles_buf[small_len];
3893 Node* nodes_buf[small_len];
3894 jlong inits_buf[small_len];
3895 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3896 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3897 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3898 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3899 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3900 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3901 // tiles: exact bitwise model of all primitive constants
3902 // nodes: last constant-storing node subsumed into the tiles model
3903 // inits: which bytes (in each tile) are touched by any initializations
3904
3905 //// Pass A: Fill in the tile model with any relevant stores.
3906
3907 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3908 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3909 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3910 Node* zmem = zero_memory(); // initially zero memory state
3911 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3912 Node* st = in(i);
3913 intptr_t st_off = get_store_offset(st, phase);
3914
3915 // Figure out the store's offset and constant value:
3916 if (st_off < header_size) continue; //skip (ignore header)
3917 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3918 int st_size = st->as_Store()->memory_size();
3919 if (st_off + st_size > size_limit) break;
3920
3921 // Record which bytes are touched, whether by constant or not.
3922 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3923 continue; // skip (strange store size)
3924
3925 const Type* val = phase->type(st->in(MemNode::ValueIn));
3926 if (!val->singleton()) continue; //skip (non-con store)
3927 BasicType type = val->basic_type();
3928
3929 jlong con = 0;
3930 switch (type) {
3931 case T_INT: con = val->is_int()->get_con(); break;
3932 case T_LONG: con = val->is_long()->get_con(); break;
3933 case T_FLOAT: con = jint_cast(val->getf()); break;
3934 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3935 default: continue; //skip (odd store type)
3936 }
3937
3938 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3939 st->Opcode() == Op_StoreL) {
3940 continue; // This StoreL is already optimal.
3941 }
3942
3943 // Store down the constant.
3944 store_constant(tiles, num_tiles, st_off, st_size, con);
3945
3946 intptr_t j = st_off >> LogBytesPerLong;
3947
3948 if (type == T_INT && st_size == BytesPerInt
3949 && (st_off & BytesPerInt) == BytesPerInt) {
3950 jlong lcon = tiles[j];
3951 if (!Matcher::isSimpleConstant64(lcon) &&
3952 st->Opcode() == Op_StoreI) {
3953 // This StoreI is already optimal by itself.
3954 jint* intcon = (jint*) &tiles[j];
3955 intcon[1] = 0; // undo the store_constant()
3956
3957 // If the previous store is also optimal by itself, back up and
3958 // undo the action of the previous loop iteration... if we can.
3959 // But if we can't, just let the previous half take care of itself.
3960 st = nodes[j];
3961 st_off -= BytesPerInt;
3962 con = intcon[0];
3963 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3964 assert(st_off >= header_size, "still ignoring header");
3965 assert(get_store_offset(st, phase) == st_off, "must be");
3966 assert(in(i-1) == zmem, "must be");
3967 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3968 assert(con == tcon->is_int()->get_con(), "must be");
3969 // Undo the effects of the previous loop trip, which swallowed st:
3970 intcon[0] = 0; // undo store_constant()
3971 set_req(i-1, st); // undo set_req(i, zmem)
3972 nodes[j] = NULL; // undo nodes[j] = st
3973 --old_subword; // undo ++old_subword
3974 }
3975 continue; // This StoreI is already optimal.
3976 }
3977 }
3978
3979 // This store is not needed.
3980 set_req(i, zmem);
3981 nodes[j] = st; // record for the moment
3982 if (st_size < BytesPerLong) // something has changed
3983 ++old_subword; // includes int/float, but who's counting...
3984 else ++old_long;
3985 }
3986
3987 if ((old_subword + old_long) == 0)
3988 return; // nothing more to do
3989
3990 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3991 // Be sure to insert them before overlapping non-constant stores.
3992 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3993 for (int j = 0; j < num_tiles; j++) {
3994 jlong con = tiles[j];
3995 jlong init = inits[j];
3996 if (con == 0) continue;
3997 jint con0, con1; // split the constant, address-wise
3998 jint init0, init1; // split the init map, address-wise
3999 { union { jlong con; jint intcon[2]; } u;
4000 u.con = con;
4001 con0 = u.intcon[0];
4002 con1 = u.intcon[1];
4003 u.con = init;
4004 init0 = u.intcon[0];
4005 init1 = u.intcon[1];
4006 }
4007
4008 Node* old = nodes[j];
4009 assert(old != NULL, "need the prior store");
4010 intptr_t offset = (j * BytesPerLong);
4011
4012 bool split = !Matcher::isSimpleConstant64(con);
4013
4014 if (offset < header_size) {
4015 assert(offset + BytesPerInt >= header_size, "second int counts");
4016 assert(*(jint*)&tiles[j] == 0, "junk in header");
4017 split = true; // only the second word counts
4018 // Example: int a[] = { 42 ... }
4019 } else if (con0 == 0 && init0 == -1) {
4020 split = true; // first word is covered by full inits
4021 // Example: int a[] = { ... foo(), 42 ... }
4022 } else if (con1 == 0 && init1 == -1) {
4023 split = true; // second word is covered by full inits
4024 // Example: int a[] = { ... 42, foo() ... }
4025 }
4026
4027 // Here's a case where init0 is neither 0 nor -1:
4028 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
4029 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
4030 // In this case the tile is not split; it is (jlong)42.
4031 // The big tile is stored down, and then the foo() value is inserted.
4032 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
4033
4034 Node* ctl = old->in(MemNode::Control);
4035 Node* adr = make_raw_address(offset, phase);
4036 const TypePtr* atp = TypeRawPtr::BOTTOM;
4037
4038 // One or two coalesced stores to plop down.
4039 Node* st[2];
4040 intptr_t off[2];
4041 int nst = 0;
4042 if (!split) {
4043 ++new_long;
4044 off[nst] = offset;
4045 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4046 phase->longcon(con), T_LONG, MemNode::unordered);
4047 } else {
4048 // Omit either if it is a zero.
4049 if (con0 != 0) {
4050 ++new_int;
4051 off[nst] = offset;
4052 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4053 phase->intcon(con0), T_INT, MemNode::unordered);
4054 }
4055 if (con1 != 0) {
4056 ++new_int;
4057 offset += BytesPerInt;
4058 adr = make_raw_address(offset, phase);
4059 off[nst] = offset;
4060 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4061 phase->intcon(con1), T_INT, MemNode::unordered);
4062 }
4063 }
4064
4065 // Insert second store first, then the first before the second.
4066 // Insert each one just before any overlapping non-constant stores.
4067 while (nst > 0) {
4068 Node* st1 = st[--nst];
4069 C->copy_node_notes_to(st1, old);
4070 st1 = phase->transform(st1);
4071 offset = off[nst];
4072 assert(offset >= header_size, "do not smash header");
4073 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4074 guarantee(ins_idx != 0, "must re-insert constant store");
4075 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
4076 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4077 set_req(--ins_idx, st1);
4078 else
4079 ins_req(ins_idx, st1);
4080 }
4081 }
4082
4083 if (PrintCompilation && WizardMode)
4084 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4085 old_subword, old_long, new_int, new_long);
4086 if (C->log() != NULL)
4087 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4088 old_subword, old_long, new_int, new_long);
4089
4090 // Clean up any remaining occurrences of zmem:
4091 remove_extra_zeroes();
4092 }
4093
4094 // Explore forward from in(start) to find the first fully initialized
4095 // word, and return its offset. Skip groups of subword stores which
4096 // together initialize full words. If in(start) is itself part of a
4097 // fully initialized word, return the offset of in(start). If there
4098 // are no following full-word stores, or if something is fishy, return
4099 // a negative value.
4100 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4101 int int_map = 0;
4102 intptr_t int_map_off = 0;
4103 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
4104
4105 for (uint i = start, limit = req(); i < limit; i++) {
4106 Node* st = in(i);
4107
4108 intptr_t st_off = get_store_offset(st, phase);
4109 if (st_off < 0) break; // return conservative answer
4110
4111 int st_size = st->as_Store()->memory_size();
4112 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4113 return st_off; // we found a complete word init
4114 }
4115
4116 // update the map:
4117
4118 intptr_t this_int_off = align_down(st_off, BytesPerInt);
4119 if (this_int_off != int_map_off) {
4120 // reset the map:
4121 int_map = 0;
4122 int_map_off = this_int_off;
4123 }
4124
4125 int subword_off = st_off - this_int_off;
4126 int_map |= right_n_bits(st_size) << subword_off;
4127 if ((int_map & FULL_MAP) == FULL_MAP) {
4128 return this_int_off; // we found a complete word init
4129 }
4130
4131 // Did this store hit or cross the word boundary?
4132 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4133 if (next_int_off == this_int_off + BytesPerInt) {
4134 // We passed the current int, without fully initializing it.
4135 int_map_off = next_int_off;
4136 int_map >>= BytesPerInt;
4137 } else if (next_int_off > this_int_off + BytesPerInt) {
4138 // We passed the current and next int.
4139 return this_int_off + BytesPerInt;
4140 }
4141 }
4142
4143 return -1;
4144 }
4145
4146
4147 // Called when the associated AllocateNode is expanded into CFG.
4148 // At this point, we may perform additional optimizations.
4149 // Linearize the stores by ascending offset, to make memory
4150 // activity as coherent as possible.
4151 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4152 intptr_t header_size,
4153 Node* size_in_bytes,
4154 PhaseGVN* phase) {
4155 assert(!is_complete(), "not already complete");
4156 assert(stores_are_sane(phase), "");
4157 assert(allocation() != NULL, "must be present");
4158
4159 remove_extra_zeroes();
4160
4161 if (ReduceFieldZeroing || ReduceBulkZeroing)
4162 // reduce instruction count for common initialization patterns
4163 coalesce_subword_stores(header_size, size_in_bytes, phase);
4164
4165 Node* zmem = zero_memory(); // initially zero memory state
4166 Node* inits = zmem; // accumulating a linearized chain of inits
4167 #ifdef ASSERT
4168 intptr_t first_offset = allocation()->minimum_header_size();
4169 intptr_t last_init_off = first_offset; // previous init offset
4170 intptr_t last_init_end = first_offset; // previous init offset+size
4171 intptr_t last_tile_end = first_offset; // previous tile offset+size
4172 #endif
4173 intptr_t zeroes_done = header_size;
4174
4175 bool do_zeroing = true; // we might give up if inits are very sparse
4176 int big_init_gaps = 0; // how many large gaps have we seen?
4177
4178 if (UseTLAB && ZeroTLAB) do_zeroing = false;
4179 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
4180
4181 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4182 Node* st = in(i);
4183 intptr_t st_off = get_store_offset(st, phase);
4184 if (st_off < 0)
4185 break; // unknown junk in the inits
4186 if (st->in(MemNode::Memory) != zmem)
4187 break; // complicated store chains somehow in list
4188
4189 int st_size = st->as_Store()->memory_size();
4190 intptr_t next_init_off = st_off + st_size;
4191
4192 if (do_zeroing && zeroes_done < next_init_off) {
4193 // See if this store needs a zero before it or under it.
4194 intptr_t zeroes_needed = st_off;
4195
4196 if (st_size < BytesPerInt) {
4197 // Look for subword stores which only partially initialize words.
4198 // If we find some, we must lay down some word-level zeroes first,
4199 // underneath the subword stores.
4200 //
4201 // Examples:
4202 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
4203 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
4204 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
4205 //
4206 // Note: coalesce_subword_stores may have already done this,
4207 // if it was prompted by constant non-zero subword initializers.
4208 // But this case can still arise with non-constant stores.
4209
4210 intptr_t next_full_store = find_next_fullword_store(i, phase);
4211
4212 // In the examples above:
4213 // in(i) p q r s x y z
4214 // st_off 12 13 14 15 12 13 14
4215 // st_size 1 1 1 1 1 1 1
4216 // next_full_s. 12 16 16 16 16 16 16
4217 // z's_done 12 16 16 16 12 16 12
4218 // z's_needed 12 16 16 16 16 16 16
4219 // zsize 0 0 0 0 4 0 4
4220 if (next_full_store < 0) {
4221 // Conservative tack: Zero to end of current word.
4222 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4223 } else {
4224 // Zero to beginning of next fully initialized word.
4225 // Or, don't zero at all, if we are already in that word.
4226 assert(next_full_store >= zeroes_needed, "must go forward");
4227 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4228 zeroes_needed = next_full_store;
4229 }
4230 }
4231
4232 if (zeroes_needed > zeroes_done) {
4233 intptr_t zsize = zeroes_needed - zeroes_done;
4234 // Do some incremental zeroing on rawmem, in parallel with inits.
4235 zeroes_done = align_down(zeroes_done, BytesPerInt);
4236 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4237 allocation()->in(AllocateNode::DefaultValue),
4238 allocation()->in(AllocateNode::RawDefaultValue),
4239 zeroes_done, zeroes_needed,
4240 phase);
4241 zeroes_done = zeroes_needed;
4242 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4243 do_zeroing = false; // leave the hole, next time
4244 }
4245 }
4246
4247 // Collect the store and move on:
4248 st->set_req(MemNode::Memory, inits);
4249 inits = st; // put it on the linearized chain
4250 set_req(i, zmem); // unhook from previous position
4251
4252 if (zeroes_done == st_off)
4253 zeroes_done = next_init_off;
4254
4255 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4256
4257 #ifdef ASSERT
4258 // Various order invariants. Weaker than stores_are_sane because
4259 // a large constant tile can be filled in by smaller non-constant stores.
4260 assert(st_off >= last_init_off, "inits do not reverse");
4261 last_init_off = st_off;
4262 const Type* val = NULL;
4263 if (st_size >= BytesPerInt &&
4264 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4265 (int)val->basic_type() < (int)T_OBJECT) {
4266 assert(st_off >= last_tile_end, "tiles do not overlap");
4267 assert(st_off >= last_init_end, "tiles do not overwrite inits");
4268 last_tile_end = MAX2(last_tile_end, next_init_off);
4269 } else {
4270 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4271 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4272 assert(st_off >= last_init_end, "inits do not overlap");
4273 last_init_end = next_init_off; // it's a non-tile
4274 }
4275 #endif //ASSERT
4276 }
4277
4278 remove_extra_zeroes(); // clear out all the zmems left over
4279 add_req(inits);
4280
4281 if (!(UseTLAB && ZeroTLAB)) {
4282 // If anything remains to be zeroed, zero it all now.
4283 zeroes_done = align_down(zeroes_done, BytesPerInt);
4284 // if it is the last unused 4 bytes of an instance, forget about it
4285 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4286 if (zeroes_done + BytesPerLong >= size_limit) {
4287 AllocateNode* alloc = allocation();
4288 assert(alloc != NULL, "must be present");
4289 if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
4290 Node* klass_node = alloc->in(AllocateNode::KlassNode);
4291 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4292 if (zeroes_done == k->layout_helper())
4293 zeroes_done = size_limit;
4294 }
4295 }
4296 if (zeroes_done < size_limit) {
4297 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4298 allocation()->in(AllocateNode::DefaultValue),
4299 allocation()->in(AllocateNode::RawDefaultValue),
4300 zeroes_done, size_in_bytes, phase);
4301 }
4302 }
4303
4304 set_complete(phase);
4305 return rawmem;
4306 }
4307
4308
4309 #ifdef ASSERT
4310 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4311 if (is_complete())
4312 return true; // stores could be anything at this point
4313 assert(allocation() != NULL, "must be present");
4314 intptr_t last_off = allocation()->minimum_header_size();
4315 for (uint i = InitializeNode::RawStores; i < req(); i++) {
4316 Node* st = in(i);
4317 intptr_t st_off = get_store_offset(st, phase);
4318 if (st_off < 0) continue; // ignore dead garbage
4319 if (last_off > st_off) {
4320 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4321 this->dump(2);
4322 assert(false, "ascending store offsets");
4323 return false;
4324 }
4325 last_off = st_off + st->as_Store()->memory_size();
4326 }
4327 return true;
4328 }
4329 #endif //ASSERT
4330
4331
4332
4333
4334 //============================MergeMemNode=====================================
4335 //
4336 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4337 // contributing store or call operations. Each contributor provides the memory
4338 // state for a particular "alias type" (see Compile::alias_type). For example,
4339 // if a MergeMem has an input X for alias category #6, then any memory reference
4340 // to alias category #6 may use X as its memory state input, as an exact equivalent
4341 // to using the MergeMem as a whole.
4342 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4343 //
4344 // (Here, the <N> notation gives the index of the relevant adr_type.)
4345 //
4346 // In one special case (and more cases in the future), alias categories overlap.
4347 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4348 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4349 // it is exactly equivalent to that state W:
4350 // MergeMem(<Bot>: W) <==> W
4351 //
4352 // Usually, the merge has more than one input. In that case, where inputs
4353 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4354 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4355 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4356 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4357 //
4358 // A merge can take a "wide" memory state as one of its narrow inputs.
4359 // This simply means that the merge observes out only the relevant parts of
4360 // the wide input. That is, wide memory states arriving at narrow merge inputs
4361 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4362 //
4363 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4364 // and that memory slices "leak through":
4365 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4366 //
4367 // But, in such a cascade, repeated memory slices can "block the leak":
4368 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4369 //
4370 // In the last example, Y is not part of the combined memory state of the
4371 // outermost MergeMem. The system must, of course, prevent unschedulable
4372 // memory states from arising, so you can be sure that the state Y is somehow
4373 // a precursor to state Y'.
4374 //
4375 //
4376 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4377 // of each MergeMemNode array are exactly the numerical alias indexes, including
4378 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4379 // Compile::alias_type (and kin) produce and manage these indexes.
4380 //
4381 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4382 // (Note that this provides quick access to the top node inside MergeMem methods,
4383 // without the need to reach out via TLS to Compile::current.)
4384 //
4385 // As a consequence of what was just described, a MergeMem that represents a full
4386 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4387 // containing all alias categories.
4388 //
4389 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4390 //
4391 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4392 // a memory state for the alias type <N>, or else the top node, meaning that
4393 // there is no particular input for that alias type. Note that the length of
4394 // a MergeMem is variable, and may be extended at any time to accommodate new
4395 // memory states at larger alias indexes. When merges grow, they are of course
4396 // filled with "top" in the unused in() positions.
4397 //
4398 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4399 // (Top was chosen because it works smoothly with passes like GCM.)
4400 //
4401 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4402 // the type of random VM bits like TLS references.) Since it is always the
4403 // first non-Bot memory slice, some low-level loops use it to initialize an
4404 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4405 //
4406 //
4407 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4408 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4409 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4410 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4411 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4412 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4413 //
4414 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4415 // really that different from the other memory inputs. An abbreviation called
4416 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4417 //
4418 //
4419 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4420 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4421 // that "emerges though" the base memory will be marked as excluding the alias types
4422 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4423 //
4424 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4425 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4426 //
4427 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4428 // (It is currently unimplemented.) As you can see, the resulting merge is
4429 // actually a disjoint union of memory states, rather than an overlay.
4430 //
4431
4432 //------------------------------MergeMemNode-----------------------------------
4433 Node* MergeMemNode::make_empty_memory() {
4434 Node* empty_memory = (Node*) Compile::current()->top();
4435 assert(empty_memory->is_top(), "correct sentinel identity");
4436 return empty_memory;
4437 }
4438
4439 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4440 init_class_id(Class_MergeMem);
4441 // all inputs are nullified in Node::Node(int)
4442 // set_input(0, NULL); // no control input
4443
4444 // Initialize the edges uniformly to top, for starters.
4445 Node* empty_mem = make_empty_memory();
4446 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4447 init_req(i,empty_mem);
4448 }
4449 assert(empty_memory() == empty_mem, "");
4450
4451 if( new_base != NULL && new_base->is_MergeMem() ) {
4452 MergeMemNode* mdef = new_base->as_MergeMem();
4453 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4454 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4455 mms.set_memory(mms.memory2());
4456 }
4457 assert(base_memory() == mdef->base_memory(), "");
4458 } else {
4459 set_base_memory(new_base);
4460 }
4461 }
4462
4463 // Make a new, untransformed MergeMem with the same base as 'mem'.
4464 // If mem is itself a MergeMem, populate the result with the same edges.
4465 MergeMemNode* MergeMemNode::make(Node* mem) {
4466 return new MergeMemNode(mem);
4467 }
4468
4469 //------------------------------cmp--------------------------------------------
4470 uint MergeMemNode::hash() const { return NO_HASH; }
4471 uint MergeMemNode::cmp( const Node &n ) const {
4472 return (&n == this); // Always fail except on self
4473 }
4474
4475 //------------------------------Identity---------------------------------------
4476 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4477 // Identity if this merge point does not record any interesting memory
4478 // disambiguations.
4479 Node* base_mem = base_memory();
4480 Node* empty_mem = empty_memory();
4481 if (base_mem != empty_mem) { // Memory path is not dead?
4482 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4483 Node* mem = in(i);
4484 if (mem != empty_mem && mem != base_mem) {
4485 return this; // Many memory splits; no change
4486 }
4487 }
4488 }
4489 return base_mem; // No memory splits; ID on the one true input
4490 }
4491
4492 //------------------------------Ideal------------------------------------------
4493 // This method is invoked recursively on chains of MergeMem nodes
4494 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4495 // Remove chain'd MergeMems
4496 //
4497 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4498 // relative to the "in(Bot)". Since we are patching both at the same time,
4499 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4500 // but rewrite each "in(i)" relative to the new "in(Bot)".
4501 Node *progress = NULL;
4502
4503
4504 Node* old_base = base_memory();
4505 Node* empty_mem = empty_memory();
4506 if (old_base == empty_mem)
4507 return NULL; // Dead memory path.
4508
4509 MergeMemNode* old_mbase;
4510 if (old_base != NULL && old_base->is_MergeMem())
4511 old_mbase = old_base->as_MergeMem();
4512 else
4513 old_mbase = NULL;
4514 Node* new_base = old_base;
4515
4516 // simplify stacked MergeMems in base memory
4517 if (old_mbase) new_base = old_mbase->base_memory();
4518
4519 // the base memory might contribute new slices beyond my req()
4520 if (old_mbase) grow_to_match(old_mbase);
4521
4522 // Look carefully at the base node if it is a phi.
4523 PhiNode* phi_base;
4524 if (new_base != NULL && new_base->is_Phi())
4525 phi_base = new_base->as_Phi();
4526 else
4527 phi_base = NULL;
4528
4529 Node* phi_reg = NULL;
4530 uint phi_len = (uint)-1;
4531 if (phi_base != NULL && !phi_base->is_copy()) {
4532 // do not examine phi if degraded to a copy
4533 phi_reg = phi_base->region();
4534 phi_len = phi_base->req();
4535 // see if the phi is unfinished
4536 for (uint i = 1; i < phi_len; i++) {
4537 if (phi_base->in(i) == NULL) {
4538 // incomplete phi; do not look at it yet!
4539 phi_reg = NULL;
4540 phi_len = (uint)-1;
4541 break;
4542 }
4543 }
4544 }
4545
4546 // Note: We do not call verify_sparse on entry, because inputs
4547 // can normalize to the base_memory via subsume_node or similar
4548 // mechanisms. This method repairs that damage.
4549
4550 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4551
4552 // Look at each slice.
4553 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4554 Node* old_in = in(i);
4555 // calculate the old memory value
4556 Node* old_mem = old_in;
4557 if (old_mem == empty_mem) old_mem = old_base;
4558 assert(old_mem == memory_at(i), "");
4559
4560 // maybe update (reslice) the old memory value
4561
4562 // simplify stacked MergeMems
4563 Node* new_mem = old_mem;
4564 MergeMemNode* old_mmem;
4565 if (old_mem != NULL && old_mem->is_MergeMem())
4566 old_mmem = old_mem->as_MergeMem();
4567 else
4568 old_mmem = NULL;
4569 if (old_mmem == this) {
4570 // This can happen if loops break up and safepoints disappear.
4571 // A merge of BotPtr (default) with a RawPtr memory derived from a
4572 // safepoint can be rewritten to a merge of the same BotPtr with
4573 // the BotPtr phi coming into the loop. If that phi disappears
4574 // also, we can end up with a self-loop of the mergemem.
4575 // In general, if loops degenerate and memory effects disappear,
4576 // a mergemem can be left looking at itself. This simply means
4577 // that the mergemem's default should be used, since there is
4578 // no longer any apparent effect on this slice.
4579 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4580 // from start. Update the input to TOP.
4581 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4582 }
4583 else if (old_mmem != NULL) {
4584 new_mem = old_mmem->memory_at(i);
4585 }
4586 // else preceding memory was not a MergeMem
4587
4588 // replace equivalent phis (unfortunately, they do not GVN together)
4589 if (new_mem != NULL && new_mem != new_base &&
4590 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4591 if (new_mem->is_Phi()) {
4592 PhiNode* phi_mem = new_mem->as_Phi();
4593 for (uint i = 1; i < phi_len; i++) {
4594 if (phi_base->in(i) != phi_mem->in(i)) {
4595 phi_mem = NULL;
4596 break;
4597 }
4598 }
4599 if (phi_mem != NULL) {
4600 // equivalent phi nodes; revert to the def
4601 new_mem = new_base;
4602 }
4603 }
4604 }
4605
4606 // maybe store down a new value
4607 Node* new_in = new_mem;
4608 if (new_in == new_base) new_in = empty_mem;
4609
4610 if (new_in != old_in) {
4611 // Warning: Do not combine this "if" with the previous "if"
4612 // A memory slice might have be be rewritten even if it is semantically
4613 // unchanged, if the base_memory value has changed.
4614 set_req(i, new_in);
4615 progress = this; // Report progress
4616 }
4617 }
4618
4619 if (new_base != old_base) {
4620 set_req(Compile::AliasIdxBot, new_base);
4621 // Don't use set_base_memory(new_base), because we need to update du.
4622 assert(base_memory() == new_base, "");
4623 progress = this;
4624 }
4625
4626 if( base_memory() == this ) {
4627 // a self cycle indicates this memory path is dead
4628 set_req(Compile::AliasIdxBot, empty_mem);
4629 }
4630
4631 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4632 // Recursion must occur after the self cycle check above
4633 if( base_memory()->is_MergeMem() ) {
4634 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4635 Node *m = phase->transform(new_mbase); // Rollup any cycles
4636 if( m != NULL &&
4637 (m->is_top() ||
4638 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4639 // propagate rollup of dead cycle to self
4640 set_req(Compile::AliasIdxBot, empty_mem);
4641 }
4642 }
4643
4644 if( base_memory() == empty_mem ) {
4645 progress = this;
4646 // Cut inputs during Parse phase only.
4647 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4648 if( !can_reshape ) {
4649 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4650 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4651 }
4652 }
4653 }
4654
4655 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4656 // Check if PhiNode::Ideal's "Split phis through memory merges"
4657 // transform should be attempted. Look for this->phi->this cycle.
4658 uint merge_width = req();
4659 if (merge_width > Compile::AliasIdxRaw) {
4660 PhiNode* phi = base_memory()->as_Phi();
4661 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4662 if (phi->in(i) == this) {
4663 phase->is_IterGVN()->_worklist.push(phi);
4664 break;
4665 }
4666 }
4667 }
4668 }
4669
4670 assert(progress || verify_sparse(), "please, no dups of base");
4671 return progress;
4672 }
4673
4674 //-------------------------set_base_memory-------------------------------------
4675 void MergeMemNode::set_base_memory(Node *new_base) {
4676 Node* empty_mem = empty_memory();
4677 set_req(Compile::AliasIdxBot, new_base);
4678 assert(memory_at(req()) == new_base, "must set default memory");
4679 // Clear out other occurrences of new_base:
4680 if (new_base != empty_mem) {
4681 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4682 if (in(i) == new_base) set_req(i, empty_mem);
4683 }
4684 }
4685 }
4686
4687 //------------------------------out_RegMask------------------------------------
4688 const RegMask &MergeMemNode::out_RegMask() const {
4689 return RegMask::Empty;
4690 }
4691
4692 //------------------------------dump_spec--------------------------------------
4693 #ifndef PRODUCT
4694 void MergeMemNode::dump_spec(outputStream *st) const {
4695 st->print(" {");
4696 Node* base_mem = base_memory();
4697 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4698 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4699 if (mem == base_mem) { st->print(" -"); continue; }
4700 st->print( " N%d:", mem->_idx );
4701 Compile::current()->get_adr_type(i)->dump_on(st);
4702 }
4703 st->print(" }");
4704 }
4705 #endif // !PRODUCT
4706
4707
4708 #ifdef ASSERT
4709 static bool might_be_same(Node* a, Node* b) {
4710 if (a == b) return true;
4711 if (!(a->is_Phi() || b->is_Phi())) return false;
4712 // phis shift around during optimization
4713 return true; // pretty stupid...
4714 }
4715
4716 // verify a narrow slice (either incoming or outgoing)
4717 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4718 if (!VerifyAliases) return; // don't bother to verify unless requested
4719 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
4720 if (Node::in_dump()) return; // muzzle asserts when printing
4721 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4722 assert(n != NULL, "");
4723 // Elide intervening MergeMem's
4724 while (n->is_MergeMem()) {
4725 n = n->as_MergeMem()->memory_at(alias_idx);
4726 }
4727 Compile* C = Compile::current();
4728 const TypePtr* n_adr_type = n->adr_type();
4729 if (n == m->empty_memory()) {
4730 // Implicit copy of base_memory()
4731 } else if (n_adr_type != TypePtr::BOTTOM) {
4732 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4733 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4734 } else {
4735 // A few places like make_runtime_call "know" that VM calls are narrow,
4736 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4737 bool expected_wide_mem = false;
4738 if (n == m->base_memory()) {
4739 expected_wide_mem = true;
4740 } else if (alias_idx == Compile::AliasIdxRaw ||
4741 n == m->memory_at(Compile::AliasIdxRaw)) {
4742 expected_wide_mem = true;
4743 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4744 // memory can "leak through" calls on channels that
4745 // are write-once. Allow this also.
4746 expected_wide_mem = true;
4747 }
4748 assert(expected_wide_mem, "expected narrow slice replacement");
4749 }
4750 }
4751 #else // !ASSERT
4752 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4753 #endif
4754
4755
4756 //-----------------------------memory_at---------------------------------------
4757 Node* MergeMemNode::memory_at(uint alias_idx) const {
4758 assert(alias_idx >= Compile::AliasIdxRaw ||
4759 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4760 "must avoid base_memory and AliasIdxTop");
4761
4762 // Otherwise, it is a narrow slice.
4763 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4764 Compile *C = Compile::current();
4765 if (is_empty_memory(n)) {
4766 // the array is sparse; empty slots are the "top" node
4767 n = base_memory();
4768 assert(Node::in_dump()
4769 || n == NULL || n->bottom_type() == Type::TOP
4770 || n->adr_type() == NULL // address is TOP
4771 || n->adr_type() == TypePtr::BOTTOM
4772 || n->adr_type() == TypeRawPtr::BOTTOM
4773 || Compile::current()->AliasLevel() == 0,
4774 "must be a wide memory");
4775 // AliasLevel == 0 if we are organizing the memory states manually.
4776 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4777 } else {
4778 // make sure the stored slice is sane
4779 #ifdef ASSERT
4780 if (VMError::is_error_reported() || Node::in_dump()) {
4781 } else if (might_be_same(n, base_memory())) {
4782 // Give it a pass: It is a mostly harmless repetition of the base.
4783 // This can arise normally from node subsumption during optimization.
4784 } else {
4785 verify_memory_slice(this, alias_idx, n);
4786 }
4787 #endif
4788 }
4789 return n;
4790 }
4791
4792 //---------------------------set_memory_at-------------------------------------
4793 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4794 verify_memory_slice(this, alias_idx, n);
4795 Node* empty_mem = empty_memory();
4796 if (n == base_memory()) n = empty_mem; // collapse default
4797 uint need_req = alias_idx+1;
4798 if (req() < need_req) {
4799 if (n == empty_mem) return; // already the default, so do not grow me
4800 // grow the sparse array
4801 do {
4802 add_req(empty_mem);
4803 } while (req() < need_req);
4804 }
4805 set_req( alias_idx, n );
4806 }
4807
4808
4809
4810 //--------------------------iteration_setup------------------------------------
4811 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4812 if (other != NULL) {
4813 grow_to_match(other);
4814 // invariant: the finite support of mm2 is within mm->req()
4815 #ifdef ASSERT
4816 for (uint i = req(); i < other->req(); i++) {
4817 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4818 }
4819 #endif
4820 }
4821 // Replace spurious copies of base_memory by top.
4822 Node* base_mem = base_memory();
4823 if (base_mem != NULL && !base_mem->is_top()) {
4824 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4825 if (in(i) == base_mem)
4826 set_req(i, empty_memory());
4827 }
4828 }
4829 }
4830
4831 //---------------------------grow_to_match-------------------------------------
4832 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4833 Node* empty_mem = empty_memory();
4834 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4835 // look for the finite support of the other memory
4836 for (uint i = other->req(); --i >= req(); ) {
4837 if (other->in(i) != empty_mem) {
4838 uint new_len = i+1;
4839 while (req() < new_len) add_req(empty_mem);
4840 break;
4841 }
4842 }
4843 }
4844
4845 //---------------------------verify_sparse-------------------------------------
4846 #ifndef PRODUCT
4847 bool MergeMemNode::verify_sparse() const {
4848 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4849 Node* base_mem = base_memory();
4850 // The following can happen in degenerate cases, since empty==top.
4851 if (is_empty_memory(base_mem)) return true;
4852 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4853 assert(in(i) != NULL, "sane slice");
4854 if (in(i) == base_mem) return false; // should have been the sentinel value!
4855 }
4856 return true;
4857 }
4858
4859 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4860 Node* n;
4861 n = mm->in(idx);
4862 if (mem == n) return true; // might be empty_memory()
4863 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4864 if (mem == n) return true;
4865 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4866 if (mem == n) return true;
4867 if (n == NULL) break;
4868 }
4869 return false;
4870 }
4871 #endif // !PRODUCT