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