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