1 /*
2 * Copyright (c) 1997, 2020, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "memory/allocation.inline.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/connode.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/loopnode.hpp"
40 #include "opto/machnode.hpp"
41 #include "opto/matcher.hpp"
42 #include "opto/memnode.hpp"
43 #include "opto/mulnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/phaseX.hpp"
46 #include "opto/regmask.hpp"
47 #include "opto/rootnode.hpp"
48 #include "utilities/align.hpp"
49 #include "utilities/copy.hpp"
50 #include "utilities/macros.hpp"
51 #include "utilities/powerOfTwo.hpp"
52 #include "utilities/vmError.hpp"
53
54 // Portions of code courtesy of Clifford Click
55
56 // Optimization - Graph Style
57
58 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
59
60 //=============================================================================
61 uint MemNode::size_of() const { return sizeof(*this); }
62
63 const TypePtr *MemNode::adr_type() const {
64 Node* adr = in(Address);
65 if (adr == NULL) return NULL; // node is dead
66 const TypePtr* cross_check = NULL;
67 DEBUG_ONLY(cross_check = _adr_type);
68 return calculate_adr_type(adr->bottom_type(), cross_check);
69 }
70
71 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
72 if (adr != NULL) {
73 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
74 return true;
75 }
76 }
77 return false;
78 }
79
80 #ifndef PRODUCT
81 void MemNode::dump_spec(outputStream *st) const {
82 if (in(Address) == NULL) return; // node is dead
83 #ifndef ASSERT
84 // fake the missing field
85 const TypePtr* _adr_type = NULL;
86 if (in(Address) != NULL)
87 _adr_type = in(Address)->bottom_type()->isa_ptr();
88 #endif
89 dump_adr_type(this, _adr_type, st);
90
91 Compile* C = Compile::current();
92 if (C->alias_type(_adr_type)->is_volatile()) {
93 st->print(" Volatile!");
94 }
95 if (_unaligned_access) {
96 st->print(" unaligned");
97 }
98 if (_mismatched_access) {
99 st->print(" mismatched");
100 }
101 if (_unsafe_access) {
102 st->print(" unsafe");
103 }
104 }
105
106 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
107 st->print(" @");
108 if (adr_type == NULL) {
109 st->print("NULL");
110 } else {
111 adr_type->dump_on(st);
112 Compile* C = Compile::current();
113 Compile::AliasType* atp = NULL;
114 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
115 if (atp == NULL)
116 st->print(", idx=?\?;");
117 else if (atp->index() == Compile::AliasIdxBot)
118 st->print(", idx=Bot;");
119 else if (atp->index() == Compile::AliasIdxTop)
120 st->print(", idx=Top;");
121 else if (atp->index() == Compile::AliasIdxRaw)
122 st->print(", idx=Raw;");
123 else {
124 ciField* field = atp->field();
125 if (field) {
126 st->print(", name=");
127 field->print_name_on(st);
128 }
129 st->print(", idx=%d;", atp->index());
130 }
131 }
132 }
133
134 extern void print_alias_types();
135
136 #endif
137
138 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
139 assert((t_oop != NULL), "sanity");
140 bool is_instance = t_oop->is_known_instance_field();
141 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
142 (load != NULL) && load->is_Load() &&
143 (phase->is_IterGVN() != NULL);
144 if (!(is_instance || is_boxed_value_load))
145 return mchain; // don't try to optimize non-instance types
146 uint instance_id = t_oop->instance_id();
147 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
148 Node *prev = NULL;
149 Node *result = mchain;
150 while (prev != result) {
151 prev = result;
152 if (result == start_mem)
153 break; // hit one of our sentinels
154 // skip over a call which does not affect this memory slice
155 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
156 Node *proj_in = result->in(0);
157 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
158 break; // hit one of our sentinels
159 } else if (proj_in->is_Call()) {
160 // ArrayCopyNodes processed here as well
161 CallNode *call = proj_in->as_Call();
162 if (!call->may_modify(t_oop, phase)) { // returns false for instances
163 result = call->in(TypeFunc::Memory);
164 }
165 } else if (proj_in->is_Initialize()) {
166 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
167 // Stop if this is the initialization for the object instance which
168 // contains this memory slice, otherwise skip over it.
169 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
170 break;
171 }
172 if (is_instance) {
173 result = proj_in->in(TypeFunc::Memory);
174 } else if (is_boxed_value_load) {
175 Node* klass = alloc->in(AllocateNode::KlassNode);
176 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
177 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
178 result = proj_in->in(TypeFunc::Memory); // not related allocation
179 }
180 }
181 } else if (proj_in->is_MemBar()) {
182 ArrayCopyNode* ac = NULL;
183 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
184 break;
185 }
186 result = proj_in->in(TypeFunc::Memory);
187 } else {
188 assert(false, "unexpected projection");
189 }
190 } else if (result->is_ClearArray()) {
191 intptr_t offset;
192 AllocateNode* alloc = AllocateNode::Ideal_allocation(result->in(3), phase, offset);
193
194 if (!is_instance || (alloc == NULL) || !ClearArrayNode::step_through(&result, instance_id, phase)) {
195 // Can not bypass initialization of the instance
196 // we are looking for.
197 break;
198 }
199 // Otherwise skip it (the call updated 'result' value).
200 } else if (result->is_MergeMem()) {
201 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
202 }
203 }
204 return result;
205 }
206
207 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
208 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
209 if (t_oop == NULL)
210 return mchain; // don't try to optimize non-oop types
211 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
212 bool is_instance = t_oop->is_known_instance_field();
213 PhaseIterGVN *igvn = phase->is_IterGVN();
214 if (is_instance && igvn != NULL && result->is_Phi()) {
215 PhiNode *mphi = result->as_Phi();
216 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
217 const TypePtr *t = mphi->adr_type();
218 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
219 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
220 t->is_oopptr()->cast_to_exactness(true)
221 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
222 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) {
223 // clone the Phi with our address type
224 result = mphi->split_out_instance(t_adr, igvn);
225 } else {
226 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
227 }
228 }
229 return result;
230 }
231
232 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
233 uint alias_idx = phase->C->get_alias_index(tp);
234 Node *mem = mmem;
235 #ifdef ASSERT
236 {
237 // Check that current type is consistent with the alias index used during graph construction
238 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
239 bool consistent = adr_check == NULL || adr_check->empty() ||
240 phase->C->must_alias(adr_check, alias_idx );
241 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
242 if( !consistent && adr_check != NULL && !adr_check->empty() &&
243 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
244 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
245 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
246 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
247 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
248 // don't assert if it is dead code.
249 consistent = true;
250 }
251 if( !consistent ) {
252 st->print("alias_idx==%d, adr_check==", alias_idx);
253 if( adr_check == NULL ) {
254 st->print("NULL");
255 } else {
256 adr_check->dump();
257 }
258 st->cr();
259 print_alias_types();
260 assert(consistent, "adr_check must match alias idx");
261 }
262 }
263 #endif
264 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
265 // means an array I have not precisely typed yet. Do not do any
266 // alias stuff with it any time soon.
267 const TypeOopPtr *toop = tp->isa_oopptr();
268 if( tp->base() != Type::AnyPtr &&
269 !(toop &&
270 toop->klass() != NULL &&
271 toop->klass()->is_java_lang_Object() &&
272 toop->offset() == Type::OffsetBot) ) {
273 // compress paths and change unreachable cycles to TOP
274 // If not, we can update the input infinitely along a MergeMem cycle
275 // Equivalent code in PhiNode::Ideal
276 Node* m = phase->transform(mmem);
277 // If transformed to a MergeMem, get the desired slice
278 // Otherwise the returned node represents memory for every slice
279 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
280 // Update input if it is progress over what we have now
281 }
282 return mem;
283 }
284
285 //--------------------------Ideal_common---------------------------------------
286 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
287 // Unhook non-raw memories from complete (macro-expanded) initializations.
288 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
289 // If our control input is a dead region, kill all below the region
290 Node *ctl = in(MemNode::Control);
291 if (ctl && remove_dead_region(phase, can_reshape))
292 return this;
293 ctl = in(MemNode::Control);
294 // Don't bother trying to transform a dead node
295 if (ctl && ctl->is_top()) return NodeSentinel;
296
297 PhaseIterGVN *igvn = phase->is_IterGVN();
298 // Wait if control on the worklist.
299 if (ctl && can_reshape && igvn != NULL) {
300 Node* bol = NULL;
301 Node* cmp = NULL;
302 if (ctl->in(0)->is_If()) {
303 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
304 bol = ctl->in(0)->in(1);
305 if (bol->is_Bool())
306 cmp = ctl->in(0)->in(1)->in(1);
307 }
308 if (igvn->_worklist.member(ctl) ||
309 (bol != NULL && igvn->_worklist.member(bol)) ||
310 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
311 // This control path may be dead.
312 // Delay this memory node transformation until the control is processed.
313 phase->is_IterGVN()->_worklist.push(this);
314 return NodeSentinel; // caller will return NULL
315 }
316 }
317 // Ignore if memory is dead, or self-loop
318 Node *mem = in(MemNode::Memory);
319 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
320 assert(mem != this, "dead loop in MemNode::Ideal");
321
322 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
323 // This memory slice may be dead.
324 // Delay this mem node transformation until the memory is processed.
325 phase->is_IterGVN()->_worklist.push(this);
326 return NodeSentinel; // caller will return NULL
327 }
328
329 Node *address = in(MemNode::Address);
330 const Type *t_adr = phase->type(address);
331 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
332
333 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
334 // Unsafe off-heap access with zero address. Remove access and other control users
335 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
336 assert(ctl != NULL, "unsafe accesses should be control dependent");
337 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
338 Node* u = ctl->fast_out(i);
339 if (u != ctl) {
340 igvn->rehash_node_delayed(u);
341 int nb = u->replace_edge(ctl, phase->C->top());
342 --i, imax -= nb;
343 }
344 }
345 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
346 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
347 phase->C->root()->add_req(halt);
348 return this;
349 }
350
351 if (can_reshape && igvn != NULL &&
352 (igvn->_worklist.member(address) ||
353 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
354 // The address's base and type may change when the address is processed.
355 // Delay this mem node transformation until the address is processed.
356 phase->is_IterGVN()->_worklist.push(this);
357 return NodeSentinel; // caller will return NULL
358 }
359
360 // Do NOT remove or optimize the next lines: ensure a new alias index
361 // is allocated for an oop pointer type before Escape Analysis.
362 // Note: C++ will not remove it since the call has side effect.
363 if (t_adr->isa_oopptr()) {
364 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
365 }
366
367 Node* base = NULL;
368 if (address->is_AddP()) {
369 base = address->in(AddPNode::Base);
370 }
371 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
372 !t_adr->isa_rawptr()) {
373 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
374 // Skip this node optimization if its address has TOP base.
375 return NodeSentinel; // caller will return NULL
376 }
377
378 // Avoid independent memory operations
379 Node* old_mem = mem;
380
381 // The code which unhooks non-raw memories from complete (macro-expanded)
382 // initializations was removed. After macro-expansion all stores catched
383 // by Initialize node became raw stores and there is no information
384 // which memory slices they modify. So it is unsafe to move any memory
385 // operation above these stores. Also in most cases hooked non-raw memories
386 // were already unhooked by using information from detect_ptr_independence()
387 // and find_previous_store().
388
389 if (mem->is_MergeMem()) {
390 MergeMemNode* mmem = mem->as_MergeMem();
391 const TypePtr *tp = t_adr->is_ptr();
392
393 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
394 }
395
396 if (mem != old_mem) {
397 set_req(MemNode::Memory, mem);
398 if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) {
399 igvn->_worklist.push(old_mem);
400 }
401 if (phase->type(mem) == Type::TOP) return NodeSentinel;
402 return this;
403 }
404
405 // let the subclass continue analyzing...
406 return NULL;
407 }
408
409 // Helper function for proving some simple control dominations.
410 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
411 // Already assumes that 'dom' is available at 'sub', and that 'sub'
412 // is not a constant (dominated by the method's StartNode).
413 // Used by MemNode::find_previous_store to prove that the
414 // control input of a memory operation predates (dominates)
415 // an allocation it wants to look past.
416 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
417 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
418 return false; // Conservative answer for dead code
419
420 // Check 'dom'. Skip Proj and CatchProj nodes.
421 dom = dom->find_exact_control(dom);
422 if (dom == NULL || dom->is_top())
423 return false; // Conservative answer for dead code
424
425 if (dom == sub) {
426 // For the case when, for example, 'sub' is Initialize and the original
427 // 'dom' is Proj node of the 'sub'.
428 return false;
429 }
430
431 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
432 return true;
433
434 // 'dom' dominates 'sub' if its control edge and control edges
435 // of all its inputs dominate or equal to sub's control edge.
436
437 // Currently 'sub' is either Allocate, Initialize or Start nodes.
438 // Or Region for the check in LoadNode::Ideal();
439 // 'sub' should have sub->in(0) != NULL.
440 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
441 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
442
443 // Get control edge of 'sub'.
444 Node* orig_sub = sub;
445 sub = sub->find_exact_control(sub->in(0));
446 if (sub == NULL || sub->is_top())
447 return false; // Conservative answer for dead code
448
449 assert(sub->is_CFG(), "expecting control");
450
451 if (sub == dom)
452 return true;
453
454 if (sub->is_Start() || sub->is_Root())
455 return false;
456
457 {
458 // Check all control edges of 'dom'.
459
460 ResourceMark rm;
461 Arena* arena = Thread::current()->resource_area();
462 Node_List nlist(arena);
463 Unique_Node_List dom_list(arena);
464
465 dom_list.push(dom);
466 bool only_dominating_controls = false;
467
468 for (uint next = 0; next < dom_list.size(); next++) {
469 Node* n = dom_list.at(next);
470 if (n == orig_sub)
471 return false; // One of dom's inputs dominated by sub.
472 if (!n->is_CFG() && n->pinned()) {
473 // Check only own control edge for pinned non-control nodes.
474 n = n->find_exact_control(n->in(0));
475 if (n == NULL || n->is_top())
476 return false; // Conservative answer for dead code
477 assert(n->is_CFG(), "expecting control");
478 dom_list.push(n);
479 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
480 only_dominating_controls = true;
481 } else if (n->is_CFG()) {
482 if (n->dominates(sub, nlist))
483 only_dominating_controls = true;
484 else
485 return false;
486 } else {
487 // First, own control edge.
488 Node* m = n->find_exact_control(n->in(0));
489 if (m != NULL) {
490 if (m->is_top())
491 return false; // Conservative answer for dead code
492 dom_list.push(m);
493 }
494 // Now, the rest of edges.
495 uint cnt = n->req();
496 for (uint i = 1; i < cnt; i++) {
497 m = n->find_exact_control(n->in(i));
498 if (m == NULL || m->is_top())
499 continue;
500 dom_list.push(m);
501 }
502 }
503 }
504 return only_dominating_controls;
505 }
506 }
507
508 //---------------------detect_ptr_independence---------------------------------
509 // Used by MemNode::find_previous_store to prove that two base
510 // pointers are never equal.
511 // The pointers are accompanied by their associated allocations,
512 // if any, which have been previously discovered by the caller.
513 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
514 Node* p2, AllocateNode* a2,
515 PhaseTransform* phase) {
516 // Attempt to prove that these two pointers cannot be aliased.
517 // They may both manifestly be allocations, and they should differ.
518 // Or, if they are not both allocations, they can be distinct constants.
519 // Otherwise, one is an allocation and the other a pre-existing value.
520 if (a1 == NULL && a2 == NULL) { // neither an allocation
521 return (p1 != p2) && p1->is_Con() && p2->is_Con();
522 } else if (a1 != NULL && a2 != NULL) { // both allocations
523 return (a1 != a2);
524 } else if (a1 != NULL) { // one allocation a1
525 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
526 return all_controls_dominate(p2, a1);
527 } else { //(a2 != NULL) // one allocation a2
528 return all_controls_dominate(p1, a2);
529 }
530 return false;
531 }
532
533
534 // Find an arraycopy that must have set (can_see_stored_value=true) or
535 // could have set (can_see_stored_value=false) the value for this load
536 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
537 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
538 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
539 Node* mb = mem->in(0);
540 if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
541 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
542 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy();
543 if (ac->is_clonebasic()) {
544 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase);
545 if (alloc != NULL && alloc == ld_alloc) {
546 return ac;
547 }
548 }
549 }
550 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
551 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
552
553 if (ac->is_arraycopy_validated() ||
554 ac->is_copyof_validated() ||
555 ac->is_copyofrange_validated()) {
556 Node* ld_addp = in(MemNode::Address);
557 if (ld_addp->is_AddP()) {
558 Node* ld_base = ld_addp->in(AddPNode::Address);
559 Node* ld_offs = ld_addp->in(AddPNode::Offset);
560
561 Node* dest = ac->in(ArrayCopyNode::Dest);
562
563 if (dest == ld_base) {
564 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
565 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
566 return ac;
567 }
568 if (!can_see_stored_value) {
569 mem = ac->in(TypeFunc::Memory);
570 }
571 }
572 }
573 }
574 }
575 return NULL;
576 }
577
578 // The logic for reordering loads and stores uses four steps:
579 // (a) Walk carefully past stores and initializations which we
580 // can prove are independent of this load.
581 // (b) Observe that the next memory state makes an exact match
582 // with self (load or store), and locate the relevant store.
583 // (c) Ensure that, if we were to wire self directly to the store,
584 // the optimizer would fold it up somehow.
585 // (d) Do the rewiring, and return, depending on some other part of
586 // the optimizer to fold up the load.
587 // This routine handles steps (a) and (b). Steps (c) and (d) are
588 // specific to loads and stores, so they are handled by the callers.
589 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
590 //
591 Node* MemNode::find_previous_store(PhaseTransform* phase) {
592 Node* ctrl = in(MemNode::Control);
593 Node* adr = in(MemNode::Address);
594 intptr_t offset = 0;
595 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
596 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
597
598 if (offset == Type::OffsetBot)
599 return NULL; // cannot unalias unless there are precise offsets
600
601 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
602 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
603
604 intptr_t size_in_bytes = memory_size();
605
606 Node* mem = in(MemNode::Memory); // start searching here...
607
608 int cnt = 50; // Cycle limiter
609 for (;;) { // While we can dance past unrelated stores...
610 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
611
612 Node* prev = mem;
613 if (mem->is_Store()) {
614 Node* st_adr = mem->in(MemNode::Address);
615 intptr_t st_offset = 0;
616 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
617 if (st_base == NULL)
618 break; // inscrutable pointer
619
620 // For raw accesses it's not enough to prove that constant offsets don't intersect.
621 // We need the bases to be the equal in order for the offset check to make sense.
622 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
623 break;
624 }
625
626 if (st_offset != offset && st_offset != Type::OffsetBot) {
627 const int MAX_STORE = BytesPerLong;
628 if (st_offset >= offset + size_in_bytes ||
629 st_offset <= offset - MAX_STORE ||
630 st_offset <= offset - mem->as_Store()->memory_size()) {
631 // Success: The offsets are provably independent.
632 // (You may ask, why not just test st_offset != offset and be done?
633 // The answer is that stores of different sizes can co-exist
634 // in the same sequence of RawMem effects. We sometimes initialize
635 // a whole 'tile' of array elements with a single jint or jlong.)
636 mem = mem->in(MemNode::Memory);
637 continue; // (a) advance through independent store memory
638 }
639 }
640 if (st_base != base &&
641 detect_ptr_independence(base, alloc,
642 st_base,
643 AllocateNode::Ideal_allocation(st_base, phase),
644 phase)) {
645 // Success: The bases are provably independent.
646 mem = mem->in(MemNode::Memory);
647 continue; // (a) advance through independent store memory
648 }
649
650 // (b) At this point, if the bases or offsets do not agree, we lose,
651 // since we have not managed to prove 'this' and 'mem' independent.
652 if (st_base == base && st_offset == offset) {
653 return mem; // let caller handle steps (c), (d)
654 }
655
656 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
657 InitializeNode* st_init = mem->in(0)->as_Initialize();
658 AllocateNode* st_alloc = st_init->allocation();
659 if (st_alloc == NULL)
660 break; // something degenerated
661 bool known_identical = false;
662 bool known_independent = false;
663 if (alloc == st_alloc)
664 known_identical = true;
665 else if (alloc != NULL)
666 known_independent = true;
667 else if (all_controls_dominate(this, st_alloc))
668 known_independent = true;
669
670 if (known_independent) {
671 // The bases are provably independent: Either they are
672 // manifestly distinct allocations, or else the control
673 // of this load dominates the store's allocation.
674 int alias_idx = phase->C->get_alias_index(adr_type());
675 if (alias_idx == Compile::AliasIdxRaw) {
676 mem = st_alloc->in(TypeFunc::Memory);
677 } else {
678 mem = st_init->memory(alias_idx);
679 }
680 continue; // (a) advance through independent store memory
681 }
682
683 // (b) at this point, if we are not looking at a store initializing
684 // the same allocation we are loading from, we lose.
685 if (known_identical) {
686 // From caller, can_see_stored_value will consult find_captured_store.
687 return mem; // let caller handle steps (c), (d)
688 }
689
690 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
691 if (prev != mem) {
692 // Found an arraycopy but it doesn't affect that load
693 continue;
694 }
695 // Found an arraycopy that may affect that load
696 return mem;
697 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
698 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
699 if (mem->is_Proj() && mem->in(0)->is_Call()) {
700 // ArrayCopyNodes processed here as well.
701 CallNode *call = mem->in(0)->as_Call();
702 if (!call->may_modify(addr_t, phase)) {
703 mem = call->in(TypeFunc::Memory);
704 continue; // (a) advance through independent call memory
705 }
706 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
707 ArrayCopyNode* ac = NULL;
708 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
709 break;
710 }
711 mem = mem->in(0)->in(TypeFunc::Memory);
712 continue; // (a) advance through independent MemBar memory
713 } else if (mem->is_ClearArray()) {
714 intptr_t offset;
715 AllocateNode* alloc = AllocateNode::Ideal_allocation(mem->in(3), phase, offset);
716
717 if ((alloc != NULL) && ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
718 // (the call updated 'mem' value)
719 continue; // (a) advance through independent allocation memory
720 } else {
721 // Can not bypass initialization of the instance
722 // we are looking for.
723 return mem;
724 }
725 } else if (mem->is_MergeMem()) {
726 int alias_idx = phase->C->get_alias_index(adr_type());
727 mem = mem->as_MergeMem()->memory_at(alias_idx);
728 continue; // (a) advance through independent MergeMem memory
729 }
730 }
731
732 // Unless there is an explicit 'continue', we must bail out here,
733 // because 'mem' is an inscrutable memory state (e.g., a call).
734 break;
735 }
736
737 return NULL; // bail out
738 }
739
740 //----------------------calculate_adr_type-------------------------------------
741 // Helper function. Notices when the given type of address hits top or bottom.
742 // Also, asserts a cross-check of the type against the expected address type.
743 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
744 if (t == Type::TOP) return NULL; // does not touch memory any more?
745 #ifdef ASSERT
746 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL;
747 #endif
748 const TypePtr* tp = t->isa_ptr();
749 if (tp == NULL) {
750 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
751 return TypePtr::BOTTOM; // touches lots of memory
752 } else {
753 #ifdef ASSERT
754 // %%%% [phh] We don't check the alias index if cross_check is
755 // TypeRawPtr::BOTTOM. Needs to be investigated.
756 if (cross_check != NULL &&
757 cross_check != TypePtr::BOTTOM &&
758 cross_check != TypeRawPtr::BOTTOM) {
759 // Recheck the alias index, to see if it has changed (due to a bug).
760 Compile* C = Compile::current();
761 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
762 "must stay in the original alias category");
763 // The type of the address must be contained in the adr_type,
764 // disregarding "null"-ness.
765 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
766 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
767 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
768 "real address must not escape from expected memory type");
769 }
770 #endif
771 return tp;
772 }
773 }
774
775 //=============================================================================
776 // Should LoadNode::Ideal() attempt to remove control edges?
777 bool LoadNode::can_remove_control() const {
778 return true;
779 }
780 uint LoadNode::size_of() const { return sizeof(*this); }
781 bool LoadNode::cmp( const Node &n ) const
782 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
783 const Type *LoadNode::bottom_type() const { return _type; }
784 uint LoadNode::ideal_reg() const {
785 return _type->ideal_reg();
786 }
787
788 #ifndef PRODUCT
789 void LoadNode::dump_spec(outputStream *st) const {
790 MemNode::dump_spec(st);
791 if( !Verbose && !WizardMode ) {
792 // standard dump does this in Verbose and WizardMode
793 st->print(" #"); _type->dump_on(st);
794 }
795 if (!depends_only_on_test()) {
796 st->print(" (does not depend only on test)");
797 }
798 }
799 #endif
800
801 #ifdef ASSERT
802 //----------------------------is_immutable_value-------------------------------
803 // Helper function to allow a raw load without control edge for some cases
804 bool LoadNode::is_immutable_value(Node* adr) {
805 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
806 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
807 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
808 in_bytes(JavaThread::osthread_offset())));
809 }
810 #endif
811
812 //----------------------------LoadNode::make-----------------------------------
813 // Polymorphic factory method:
814 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo,
815 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
816 Compile* C = gvn.C;
817
818 // sanity check the alias category against the created node type
819 assert(!(adr_type->isa_oopptr() &&
820 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
821 "use LoadKlassNode instead");
822 assert(!(adr_type->isa_aryptr() &&
823 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
824 "use LoadRangeNode instead");
825 // Check control edge of raw loads
826 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
827 // oop will be recorded in oop map if load crosses safepoint
828 rt->isa_oopptr() || is_immutable_value(adr),
829 "raw memory operations should have control edge");
830 LoadNode* load = NULL;
831 switch (bt) {
832 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
833 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
834 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
835 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
836 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
837 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break;
838 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
839 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
840 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
841 case T_OBJECT:
842 #ifdef _LP64
843 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
844 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
845 } else
846 #endif
847 {
848 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
849 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
850 }
851 break;
852 default:
853 ShouldNotReachHere();
854 break;
855 }
856 assert(load != NULL, "LoadNode should have been created");
857 if (unaligned) {
858 load->set_unaligned_access();
859 }
860 if (mismatched) {
861 load->set_mismatched_access();
862 }
863 if (unsafe) {
864 load->set_unsafe_access();
865 }
866 load->set_barrier_data(barrier_data);
867 if (load->Opcode() == Op_LoadN) {
868 Node* ld = gvn.transform(load);
869 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
870 }
871
872 return load;
873 }
874
875 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
876 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
877 bool require_atomic = true;
878 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic);
879 if (unaligned) {
880 load->set_unaligned_access();
881 }
882 if (mismatched) {
883 load->set_mismatched_access();
884 }
885 if (unsafe) {
886 load->set_unsafe_access();
887 }
888 load->set_barrier_data(barrier_data);
889 return load;
890 }
891
892 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
893 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
894 bool require_atomic = true;
895 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic);
896 if (unaligned) {
897 load->set_unaligned_access();
898 }
899 if (mismatched) {
900 load->set_mismatched_access();
901 }
902 if (unsafe) {
903 load->set_unsafe_access();
904 }
905 load->set_barrier_data(barrier_data);
906 return load;
907 }
908
909
910
911 //------------------------------hash-------------------------------------------
912 uint LoadNode::hash() const {
913 // unroll addition of interesting fields
914 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
915 }
916
917 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
918 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
919 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
920 bool is_stable_ary = FoldStableValues &&
921 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
922 tp->isa_aryptr()->is_stable();
923
924 return (eliminate_boxing && non_volatile) || is_stable_ary;
925 }
926
927 return false;
928 }
929
930 // Is the value loaded previously stored by an arraycopy? If so return
931 // a load node that reads from the source array so we may be able to
932 // optimize out the ArrayCopy node later.
933 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
934 Node* ld_adr = in(MemNode::Address);
935 intptr_t ld_off = 0;
936 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
937 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
938 if (ac != NULL) {
939 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
940
941 Node* mem = ac->in(TypeFunc::Memory);
942 Node* ctl = ac->in(0);
943 Node* src = ac->in(ArrayCopyNode::Src);
944
945 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
946 return NULL;
947 }
948
949 LoadNode* ld = clone()->as_Load();
950 Node* addp = in(MemNode::Address)->clone();
951 if (ac->as_ArrayCopy()->is_clonebasic()) {
952 assert(ld_alloc != NULL, "need an alloc");
953 assert(addp->is_AddP(), "address must be addp");
954 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
955 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
956 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
957 addp->set_req(AddPNode::Base, src);
958 addp->set_req(AddPNode::Address, src);
959 } else {
960 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
961 ac->as_ArrayCopy()->is_copyof_validated() ||
962 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
963 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
964 addp->set_req(AddPNode::Base, src);
965 addp->set_req(AddPNode::Address, src);
966
967 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
968 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type();
969 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
970 uint shift = exact_log2(type2aelembytes(ary_elem));
971
972 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
973 #ifdef _LP64
974 diff = phase->transform(new ConvI2LNode(diff));
975 #endif
976 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
977
978 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
979 addp->set_req(AddPNode::Offset, offset);
980 }
981 addp = phase->transform(addp);
982 #ifdef ASSERT
983 const TypePtr* adr_type = phase->type(addp)->is_ptr();
984 ld->_adr_type = adr_type;
985 #endif
986 ld->set_req(MemNode::Address, addp);
987 ld->set_req(0, ctl);
988 ld->set_req(MemNode::Memory, mem);
989 // load depends on the tests that validate the arraycopy
990 ld->_control_dependency = UnknownControl;
991 return ld;
992 }
993 return NULL;
994 }
995
996
997 //---------------------------can_see_stored_value------------------------------
998 // This routine exists to make sure this set of tests is done the same
999 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
1000 // will change the graph shape in a way which makes memory alive twice at the
1001 // same time (uses the Oracle model of aliasing), then some
1002 // LoadXNode::Identity will fold things back to the equivalence-class model
1003 // of aliasing.
1004 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
1005 Node* ld_adr = in(MemNode::Address);
1006 intptr_t ld_off = 0;
1007 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1008 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base, phase);
1009 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1010 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
1011 // This is more general than load from boxing objects.
1012 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1013 uint alias_idx = atp->index();
1014 bool final = !atp->is_rewritable();
1015 Node* result = NULL;
1016 Node* current = st;
1017 // Skip through chains of MemBarNodes checking the MergeMems for
1018 // new states for the slice of this load. Stop once any other
1019 // kind of node is encountered. Loads from final memory can skip
1020 // through any kind of MemBar but normal loads shouldn't skip
1021 // through MemBarAcquire since the could allow them to move out of
1022 // a synchronized region.
1023 while (current->is_Proj()) {
1024 int opc = current->in(0)->Opcode();
1025 if ((final && (opc == Op_MemBarAcquire ||
1026 opc == Op_MemBarAcquireLock ||
1027 opc == Op_LoadFence)) ||
1028 opc == Op_MemBarRelease ||
1029 opc == Op_StoreFence ||
1030 opc == Op_MemBarReleaseLock ||
1031 opc == Op_MemBarStoreStore ||
1032 opc == Op_MemBarCPUOrder) {
1033 Node* mem = current->in(0)->in(TypeFunc::Memory);
1034 if (mem->is_MergeMem()) {
1035 MergeMemNode* merge = mem->as_MergeMem();
1036 Node* new_st = merge->memory_at(alias_idx);
1037 if (new_st == merge->base_memory()) {
1038 // Keep searching
1039 current = new_st;
1040 continue;
1041 }
1042 // Save the new memory state for the slice and fall through
1043 // to exit.
1044 result = new_st;
1045 }
1046 }
1047 break;
1048 }
1049 if (result != NULL) {
1050 st = result;
1051 }
1052 }
1053
1054 // Loop around twice in the case Load -> Initialize -> Store.
1055 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1056 for (int trip = 0; trip <= 1; trip++) {
1057
1058 if (st->is_Store()) {
1059 Node* st_adr = st->in(MemNode::Address);
1060 if (!phase->eqv(st_adr, ld_adr)) {
1061 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1062 intptr_t st_off = 0;
1063 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1064 if (ld_base == NULL) return NULL;
1065 if (st_base == NULL) return NULL;
1066 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return NULL;
1067 if (ld_off != st_off) return NULL;
1068 if (ld_off == Type::OffsetBot) return NULL;
1069 // Same base, same offset.
1070 // Possible improvement for arrays: check index value instead of absolute offset.
1071
1072 // At this point we have proven something like this setup:
1073 // B = << base >>
1074 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1075 // S = StoreQ(AddP( B , #Off), V)
1076 // (Actually, we haven't yet proven the Q's are the same.)
1077 // In other words, we are loading from a casted version of
1078 // the same pointer-and-offset that we stored to.
1079 // Casted version may carry a dependency and it is respected.
1080 // Thus, we are able to replace L by V.
1081 }
1082 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1083 if (store_Opcode() != st->Opcode())
1084 return NULL;
1085 return st->in(MemNode::ValueIn);
1086 }
1087
1088 // A load from a freshly-created object always returns zero.
1089 // (This can happen after LoadNode::Ideal resets the load's memory input
1090 // to find_captured_store, which returned InitializeNode::zero_memory.)
1091 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1092 (st->in(0) == ld_alloc) &&
1093 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1094 // return a zero value for the load's basic type
1095 // (This is one of the few places where a generic PhaseTransform
1096 // can create new nodes. Think of it as lazily manifesting
1097 // virtually pre-existing constants.)
1098 return phase->zerocon(memory_type());
1099 }
1100
1101 // A load from an initialization barrier can match a captured store.
1102 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1103 InitializeNode* init = st->in(0)->as_Initialize();
1104 AllocateNode* alloc = init->allocation();
1105 if ((alloc != NULL) && (alloc == ld_alloc)) {
1106 // examine a captured store value
1107 st = init->find_captured_store(ld_off, memory_size(), phase);
1108 if (st != NULL) {
1109 continue; // take one more trip around
1110 }
1111 }
1112 }
1113
1114 // Load boxed value from result of valueOf() call is input parameter.
1115 if (this->is_Load() && ld_adr->is_AddP() &&
1116 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1117 intptr_t ignore = 0;
1118 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1119 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1120 base = bs->step_over_gc_barrier(base);
1121 if (base != NULL && base->is_Proj() &&
1122 base->as_Proj()->_con == TypeFunc::Parms &&
1123 base->in(0)->is_CallStaticJava() &&
1124 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1125 return base->in(0)->in(TypeFunc::Parms);
1126 }
1127 }
1128
1129 break;
1130 }
1131
1132 return NULL;
1133 }
1134
1135 //----------------------is_instance_field_load_with_local_phi------------------
1136 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1137 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1138 in(Address)->is_AddP() ) {
1139 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1140 // Only instances and boxed values.
1141 if( t_oop != NULL &&
1142 (t_oop->is_ptr_to_boxed_value() ||
1143 t_oop->is_known_instance_field()) &&
1144 t_oop->offset() != Type::OffsetBot &&
1145 t_oop->offset() != Type::OffsetTop) {
1146 return true;
1147 }
1148 }
1149 return false;
1150 }
1151
1152 //------------------------------Identity---------------------------------------
1153 // Loads are identity if previous store is to same address
1154 Node* LoadNode::Identity(PhaseGVN* phase) {
1155 // If the previous store-maker is the right kind of Store, and the store is
1156 // to the same address, then we are equal to the value stored.
1157 Node* mem = in(Memory);
1158 Node* value = can_see_stored_value(mem, phase);
1159 if( value ) {
1160 // byte, short & char stores truncate naturally.
1161 // A load has to load the truncated value which requires
1162 // some sort of masking operation and that requires an
1163 // Ideal call instead of an Identity call.
1164 if (memory_size() < BytesPerInt) {
1165 // If the input to the store does not fit with the load's result type,
1166 // it must be truncated via an Ideal call.
1167 if (!phase->type(value)->higher_equal(phase->type(this)))
1168 return this;
1169 }
1170 // (This works even when value is a Con, but LoadNode::Value
1171 // usually runs first, producing the singleton type of the Con.)
1172 return value;
1173 }
1174
1175 // Search for an existing data phi which was generated before for the same
1176 // instance's field to avoid infinite generation of phis in a loop.
1177 Node *region = mem->in(0);
1178 if (is_instance_field_load_with_local_phi(region)) {
1179 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1180 int this_index = phase->C->get_alias_index(addr_t);
1181 int this_offset = addr_t->offset();
1182 int this_iid = addr_t->instance_id();
1183 if (!addr_t->is_known_instance() &&
1184 addr_t->is_ptr_to_boxed_value()) {
1185 // Use _idx of address base (could be Phi node) for boxed values.
1186 intptr_t ignore = 0;
1187 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1188 if (base == NULL) {
1189 return this;
1190 }
1191 this_iid = base->_idx;
1192 }
1193 const Type* this_type = bottom_type();
1194 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1195 Node* phi = region->fast_out(i);
1196 if (phi->is_Phi() && phi != mem &&
1197 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1198 return phi;
1199 }
1200 }
1201 }
1202
1203 return this;
1204 }
1205
1206 // Construct an equivalent unsigned load.
1207 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1208 BasicType bt = T_ILLEGAL;
1209 const Type* rt = NULL;
1210 switch (Opcode()) {
1211 case Op_LoadUB: return this;
1212 case Op_LoadUS: return this;
1213 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1214 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1215 default:
1216 assert(false, "no unsigned variant: %s", Name());
1217 return NULL;
1218 }
1219 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1220 raw_adr_type(), rt, bt, _mo, _control_dependency,
1221 is_unaligned_access(), is_mismatched_access());
1222 }
1223
1224 // Construct an equivalent signed load.
1225 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1226 BasicType bt = T_ILLEGAL;
1227 const Type* rt = NULL;
1228 switch (Opcode()) {
1229 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1230 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1231 case Op_LoadB: // fall through
1232 case Op_LoadS: // fall through
1233 case Op_LoadI: // fall through
1234 case Op_LoadL: return this;
1235 default:
1236 assert(false, "no signed variant: %s", Name());
1237 return NULL;
1238 }
1239 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1240 raw_adr_type(), rt, bt, _mo, _control_dependency,
1241 is_unaligned_access(), is_mismatched_access());
1242 }
1243
1244 // We're loading from an object which has autobox behaviour.
1245 // If this object is result of a valueOf call we'll have a phi
1246 // merging a newly allocated object and a load from the cache.
1247 // We want to replace this load with the original incoming
1248 // argument to the valueOf call.
1249 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1250 assert(phase->C->eliminate_boxing(), "sanity");
1251 intptr_t ignore = 0;
1252 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1253 if ((base == NULL) || base->is_Phi()) {
1254 // Push the loads from the phi that comes from valueOf up
1255 // through it to allow elimination of the loads and the recovery
1256 // of the original value. It is done in split_through_phi().
1257 return NULL;
1258 } else if (base->is_Load() ||
1259 (base->is_DecodeN() && base->in(1)->is_Load())) {
1260 // Eliminate the load of boxed value for integer types from the cache
1261 // array by deriving the value from the index into the array.
1262 // Capture the offset of the load and then reverse the computation.
1263
1264 // Get LoadN node which loads a boxing object from 'cache' array.
1265 if (base->is_DecodeN()) {
1266 base = base->in(1);
1267 }
1268 if (!base->in(Address)->is_AddP()) {
1269 return NULL; // Complex address
1270 }
1271 AddPNode* address = base->in(Address)->as_AddP();
1272 Node* cache_base = address->in(AddPNode::Base);
1273 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1274 // Get ConP node which is static 'cache' field.
1275 cache_base = cache_base->in(1);
1276 }
1277 if ((cache_base != NULL) && cache_base->is_Con()) {
1278 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1279 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1280 Node* elements[4];
1281 int shift = exact_log2(type2aelembytes(T_OBJECT));
1282 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1283 if (count > 0 && elements[0]->is_Con() &&
1284 (count == 1 ||
1285 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1286 elements[1]->in(2) == phase->intcon(shift)))) {
1287 ciObjArray* array = base_type->const_oop()->as_obj_array();
1288 // Fetch the box object cache[0] at the base of the array and get its value
1289 ciInstance* box = array->obj_at(0)->as_instance();
1290 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1291 assert(ik->is_box_klass(), "sanity");
1292 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1293 if (ik->nof_nonstatic_fields() == 1) {
1294 // This should be true nonstatic_field_at requires calling
1295 // nof_nonstatic_fields so check it anyway
1296 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1297 BasicType bt = c.basic_type();
1298 // Only integer types have boxing cache.
1299 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1300 bt == T_BYTE || bt == T_SHORT ||
1301 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1302 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1303 if (cache_low != (int)cache_low) {
1304 return NULL; // should not happen since cache is array indexed by value
1305 }
1306 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1307 if (offset != (int)offset) {
1308 return NULL; // should not happen since cache is array indexed by value
1309 }
1310 // Add up all the offsets making of the address of the load
1311 Node* result = elements[0];
1312 for (int i = 1; i < count; i++) {
1313 result = phase->transform(new AddXNode(result, elements[i]));
1314 }
1315 // Remove the constant offset from the address and then
1316 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset)));
1317 // remove the scaling of the offset to recover the original index.
1318 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1319 // Peel the shift off directly but wrap it in a dummy node
1320 // since Ideal can't return existing nodes
1321 result = new RShiftXNode(result->in(1), phase->intcon(0));
1322 } else if (result->is_Add() && result->in(2)->is_Con() &&
1323 result->in(1)->Opcode() == Op_LShiftX &&
1324 result->in(1)->in(2) == phase->intcon(shift)) {
1325 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1326 // but for boxing cache access we know that X<<Z will not overflow
1327 // (there is range check) so we do this optimizatrion by hand here.
1328 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift));
1329 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con));
1330 } else {
1331 result = new RShiftXNode(result, phase->intcon(shift));
1332 }
1333 #ifdef _LP64
1334 if (bt != T_LONG) {
1335 result = new ConvL2INode(phase->transform(result));
1336 }
1337 #else
1338 if (bt == T_LONG) {
1339 result = new ConvI2LNode(phase->transform(result));
1340 }
1341 #endif
1342 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1343 // Need to preserve unboxing load type if it is unsigned.
1344 switch(this->Opcode()) {
1345 case Op_LoadUB:
1346 result = new AndINode(phase->transform(result), phase->intcon(0xFF));
1347 break;
1348 case Op_LoadUS:
1349 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF));
1350 break;
1351 }
1352 return result;
1353 }
1354 }
1355 }
1356 }
1357 }
1358 return NULL;
1359 }
1360
1361 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1362 Node* region = phi->in(0);
1363 if (region == NULL) {
1364 return false; // Wait stable graph
1365 }
1366 uint cnt = phi->req();
1367 for (uint i = 1; i < cnt; i++) {
1368 Node* rc = region->in(i);
1369 if (rc == NULL || phase->type(rc) == Type::TOP)
1370 return false; // Wait stable graph
1371 Node* in = phi->in(i);
1372 if (in == NULL || phase->type(in) == Type::TOP)
1373 return false; // Wait stable graph
1374 }
1375 return true;
1376 }
1377 //------------------------------split_through_phi------------------------------
1378 // Split instance or boxed field load through Phi.
1379 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1380 Node* mem = in(Memory);
1381 Node* address = in(Address);
1382 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1383
1384 assert((t_oop != NULL) &&
1385 (t_oop->is_known_instance_field() ||
1386 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1387
1388 Compile* C = phase->C;
1389 intptr_t ignore = 0;
1390 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1391 bool base_is_phi = (base != NULL) && base->is_Phi();
1392 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1393 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1394 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1395
1396 if (!((mem->is_Phi() || base_is_phi) &&
1397 (load_boxed_values || t_oop->is_known_instance_field()))) {
1398 return NULL; // memory is not Phi
1399 }
1400
1401 if (mem->is_Phi()) {
1402 if (!stable_phi(mem->as_Phi(), phase)) {
1403 return NULL; // Wait stable graph
1404 }
1405 uint cnt = mem->req();
1406 // Check for loop invariant memory.
1407 if (cnt == 3) {
1408 for (uint i = 1; i < cnt; i++) {
1409 Node* in = mem->in(i);
1410 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1411 if (m == mem) {
1412 if (i == 1) {
1413 // if the first edge was a loop, check second edge too.
1414 // If both are replaceable - we are in an infinite loop
1415 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1416 if (n == mem) {
1417 break;
1418 }
1419 }
1420 set_req(Memory, mem->in(cnt - i));
1421 return this; // made change
1422 }
1423 }
1424 }
1425 }
1426 if (base_is_phi) {
1427 if (!stable_phi(base->as_Phi(), phase)) {
1428 return NULL; // Wait stable graph
1429 }
1430 uint cnt = base->req();
1431 // Check for loop invariant memory.
1432 if (cnt == 3) {
1433 for (uint i = 1; i < cnt; i++) {
1434 if (base->in(i) == base) {
1435 return NULL; // Wait stable graph
1436 }
1437 }
1438 }
1439 }
1440
1441 // Split through Phi (see original code in loopopts.cpp).
1442 assert(C->have_alias_type(t_oop), "instance should have alias type");
1443
1444 // Do nothing here if Identity will find a value
1445 // (to avoid infinite chain of value phis generation).
1446 if (!phase->eqv(this, this->Identity(phase))) {
1447 return NULL;
1448 }
1449
1450 // Select Region to split through.
1451 Node* region;
1452 if (!base_is_phi) {
1453 assert(mem->is_Phi(), "sanity");
1454 region = mem->in(0);
1455 // Skip if the region dominates some control edge of the address.
1456 if (!MemNode::all_controls_dominate(address, region))
1457 return NULL;
1458 } else if (!mem->is_Phi()) {
1459 assert(base_is_phi, "sanity");
1460 region = base->in(0);
1461 // Skip if the region dominates some control edge of the memory.
1462 if (!MemNode::all_controls_dominate(mem, region))
1463 return NULL;
1464 } else if (base->in(0) != mem->in(0)) {
1465 assert(base_is_phi && mem->is_Phi(), "sanity");
1466 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1467 region = base->in(0);
1468 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1469 region = mem->in(0);
1470 } else {
1471 return NULL; // complex graph
1472 }
1473 } else {
1474 assert(base->in(0) == mem->in(0), "sanity");
1475 region = mem->in(0);
1476 }
1477
1478 const Type* this_type = this->bottom_type();
1479 int this_index = C->get_alias_index(t_oop);
1480 int this_offset = t_oop->offset();
1481 int this_iid = t_oop->instance_id();
1482 if (!t_oop->is_known_instance() && load_boxed_values) {
1483 // Use _idx of address base for boxed values.
1484 this_iid = base->_idx;
1485 }
1486 PhaseIterGVN* igvn = phase->is_IterGVN();
1487 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1488 for (uint i = 1; i < region->req(); i++) {
1489 Node* x;
1490 Node* the_clone = NULL;
1491 Node* in = region->in(i);
1492 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1493 in != NULL && in->is_OuterStripMinedLoop()) {
1494 // No node should go in the outer strip mined loop
1495 in = in->in(LoopNode::EntryControl);
1496 }
1497 if (in == NULL || in == C->top()) {
1498 x = C->top(); // Dead path? Use a dead data op
1499 } else {
1500 x = this->clone(); // Else clone up the data op
1501 the_clone = x; // Remember for possible deletion.
1502 // Alter data node to use pre-phi inputs
1503 if (this->in(0) == region) {
1504 x->set_req(0, in);
1505 } else {
1506 x->set_req(0, NULL);
1507 }
1508 if (mem->is_Phi() && (mem->in(0) == region)) {
1509 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1510 }
1511 if (address->is_Phi() && address->in(0) == region) {
1512 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1513 }
1514 if (base_is_phi && (base->in(0) == region)) {
1515 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1516 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1517 x->set_req(Address, adr_x);
1518 }
1519 }
1520 // Check for a 'win' on some paths
1521 const Type *t = x->Value(igvn);
1522
1523 bool singleton = t->singleton();
1524
1525 // See comments in PhaseIdealLoop::split_thru_phi().
1526 if (singleton && t == Type::TOP) {
1527 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1528 }
1529
1530 if (singleton) {
1531 x = igvn->makecon(t);
1532 } else {
1533 // We now call Identity to try to simplify the cloned node.
1534 // Note that some Identity methods call phase->type(this).
1535 // Make sure that the type array is big enough for
1536 // our new node, even though we may throw the node away.
1537 // (This tweaking with igvn only works because x is a new node.)
1538 igvn->set_type(x, t);
1539 // If x is a TypeNode, capture any more-precise type permanently into Node
1540 // otherwise it will be not updated during igvn->transform since
1541 // igvn->type(x) is set to x->Value() already.
1542 x->raise_bottom_type(t);
1543 Node* y = x->Identity(igvn);
1544 if (y != x) {
1545 x = y;
1546 } else {
1547 y = igvn->hash_find_insert(x);
1548 if (y) {
1549 x = y;
1550 } else {
1551 // Else x is a new node we are keeping
1552 // We do not need register_new_node_with_optimizer
1553 // because set_type has already been called.
1554 igvn->_worklist.push(x);
1555 }
1556 }
1557 }
1558 if (x != the_clone && the_clone != NULL) {
1559 igvn->remove_dead_node(the_clone);
1560 }
1561 phi->set_req(i, x);
1562 }
1563 // Record Phi
1564 igvn->register_new_node_with_optimizer(phi);
1565 return phi;
1566 }
1567
1568 AllocateNode* LoadNode::is_new_object_mark_load(PhaseGVN *phase) const {
1569 if (Opcode() == Op_LoadX) {
1570 Node* address = in(MemNode::Address);
1571 AllocateNode* alloc = AllocateNode::Ideal_allocation(address, phase);
1572 Node* mem = in(MemNode::Memory);
1573 if (alloc != NULL && mem->is_Proj() &&
1574 mem->in(0) != NULL &&
1575 mem->in(0) == alloc->initialization() &&
1576 alloc->initialization()->proj_out_or_null(0) != NULL) {
1577 return alloc;
1578 }
1579 }
1580 return NULL;
1581 }
1582
1583
1584 //------------------------------Ideal------------------------------------------
1585 // If the load is from Field memory and the pointer is non-null, it might be possible to
1586 // zero out the control input.
1587 // If the offset is constant and the base is an object allocation,
1588 // try to hook me up to the exact initializing store.
1589 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1590 Node* p = MemNode::Ideal_common(phase, can_reshape);
1591 if (p) return (p == NodeSentinel) ? NULL : p;
1592
1593 Node* ctrl = in(MemNode::Control);
1594 Node* address = in(MemNode::Address);
1595 bool progress = false;
1596
1597 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1598 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1599
1600 // Skip up past a SafePoint control. Cannot do this for Stores because
1601 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1602 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1603 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1604 !addr_mark &&
1605 (depends_only_on_test() || has_unknown_control_dependency())) {
1606 ctrl = ctrl->in(0);
1607 set_req(MemNode::Control,ctrl);
1608 progress = true;
1609 }
1610
1611 intptr_t ignore = 0;
1612 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1613 if (base != NULL
1614 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1615 // Check for useless control edge in some common special cases
1616 if (in(MemNode::Control) != NULL
1617 && can_remove_control()
1618 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1619 && all_controls_dominate(base, phase->C->start())) {
1620 // A method-invariant, non-null address (constant or 'this' argument).
1621 set_req(MemNode::Control, NULL);
1622 progress = true;
1623 }
1624 }
1625
1626 Node* mem = in(MemNode::Memory);
1627 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1628
1629 if (can_reshape && (addr_t != NULL)) {
1630 // try to optimize our memory input
1631 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1632 if (opt_mem != mem) {
1633 set_req(MemNode::Memory, opt_mem);
1634 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1635 return this;
1636 }
1637 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1638 if ((t_oop != NULL) &&
1639 (t_oop->is_known_instance_field() ||
1640 t_oop->is_ptr_to_boxed_value())) {
1641 PhaseIterGVN *igvn = phase->is_IterGVN();
1642 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1643 // Delay this transformation until memory Phi is processed.
1644 phase->is_IterGVN()->_worklist.push(this);
1645 return NULL;
1646 }
1647 // Split instance field load through Phi.
1648 Node* result = split_through_phi(phase);
1649 if (result != NULL) return result;
1650
1651 if (t_oop->is_ptr_to_boxed_value()) {
1652 Node* result = eliminate_autobox(phase);
1653 if (result != NULL) return result;
1654 }
1655 }
1656 }
1657
1658 // Is there a dominating load that loads the same value? Leave
1659 // anything that is not a load of a field/array element (like
1660 // barriers etc.) alone
1661 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) {
1662 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1663 Node *use = mem->fast_out(i);
1664 if (use != this &&
1665 use->Opcode() == Opcode() &&
1666 use->in(0) != NULL &&
1667 use->in(0) != in(0) &&
1668 use->in(Address) == in(Address)) {
1669 Node* ctl = in(0);
1670 for (int i = 0; i < 10 && ctl != NULL; i++) {
1671 ctl = IfNode::up_one_dom(ctl);
1672 if (ctl == use->in(0)) {
1673 set_req(0, use->in(0));
1674 return this;
1675 }
1676 }
1677 }
1678 }
1679 }
1680
1681 // Check for prior store with a different base or offset; make Load
1682 // independent. Skip through any number of them. Bail out if the stores
1683 // are in an endless dead cycle and report no progress. This is a key
1684 // transform for Reflection. However, if after skipping through the Stores
1685 // we can't then fold up against a prior store do NOT do the transform as
1686 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1687 // array memory alive twice: once for the hoisted Load and again after the
1688 // bypassed Store. This situation only works if EVERYBODY who does
1689 // anti-dependence work knows how to bypass. I.e. we need all
1690 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1691 // the alias index stuff. So instead, peek through Stores and IFF we can
1692 // fold up, do so.
1693 Node* prev_mem = find_previous_store(phase);
1694 if (prev_mem != NULL) {
1695 Node* value = can_see_arraycopy_value(prev_mem, phase);
1696 if (value != NULL) {
1697 return value;
1698 }
1699 }
1700 // Steps (a), (b): Walk past independent stores to find an exact match.
1701 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1702 // (c) See if we can fold up on the spot, but don't fold up here.
1703 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1704 // just return a prior value, which is done by Identity calls.
1705 if (can_see_stored_value(prev_mem, phase)) {
1706 // Make ready for step (d):
1707 set_req(MemNode::Memory, prev_mem);
1708 return this;
1709 }
1710 }
1711
1712 AllocateNode* alloc = is_new_object_mark_load(phase);
1713 if (alloc != NULL && alloc->Opcode() == Op_Allocate && UseBiasedLocking) {
1714 InitializeNode* init = alloc->initialization();
1715 Node* control = init->proj_out(0);
1716 return alloc->make_ideal_mark(phase, address, control, mem);
1717 }
1718
1719 return progress ? this : NULL;
1720 }
1721
1722 // Helper to recognize certain Klass fields which are invariant across
1723 // some group of array types (e.g., int[] or all T[] where T < Object).
1724 const Type*
1725 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1726 ciKlass* klass) const {
1727 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1728 // The field is Klass::_modifier_flags. Return its (constant) value.
1729 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1730 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1731 return TypeInt::make(klass->modifier_flags());
1732 }
1733 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1734 // The field is Klass::_access_flags. Return its (constant) value.
1735 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1736 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1737 return TypeInt::make(klass->access_flags());
1738 }
1739 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1740 // The field is Klass::_layout_helper. Return its constant value if known.
1741 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1742 return TypeInt::make(klass->layout_helper());
1743 }
1744
1745 // No match.
1746 return NULL;
1747 }
1748
1749 //------------------------------Value-----------------------------------------
1750 const Type* LoadNode::Value(PhaseGVN* phase) const {
1751 // Either input is TOP ==> the result is TOP
1752 Node* mem = in(MemNode::Memory);
1753 const Type *t1 = phase->type(mem);
1754 if (t1 == Type::TOP) return Type::TOP;
1755 Node* adr = in(MemNode::Address);
1756 const TypePtr* tp = phase->type(adr)->isa_ptr();
1757 if (tp == NULL || tp->empty()) return Type::TOP;
1758 int off = tp->offset();
1759 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1760 Compile* C = phase->C;
1761
1762 // Try to guess loaded type from pointer type
1763 if (tp->isa_aryptr()) {
1764 const TypeAryPtr* ary = tp->is_aryptr();
1765 const Type* t = ary->elem();
1766
1767 // Determine whether the reference is beyond the header or not, by comparing
1768 // the offset against the offset of the start of the array's data.
1769 // Different array types begin at slightly different offsets (12 vs. 16).
1770 // We choose T_BYTE as an example base type that is least restrictive
1771 // as to alignment, which will therefore produce the smallest
1772 // possible base offset.
1773 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1774 const bool off_beyond_header = (off >= min_base_off);
1775
1776 // Try to constant-fold a stable array element.
1777 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1778 // Make sure the reference is not into the header and the offset is constant
1779 ciObject* aobj = ary->const_oop();
1780 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1781 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1782 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1783 stable_dimension,
1784 memory_type(), is_unsigned());
1785 if (con_type != NULL) {
1786 return con_type;
1787 }
1788 }
1789 }
1790
1791 // Don't do this for integer types. There is only potential profit if
1792 // the element type t is lower than _type; that is, for int types, if _type is
1793 // more restrictive than t. This only happens here if one is short and the other
1794 // char (both 16 bits), and in those cases we've made an intentional decision
1795 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1796 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1797 //
1798 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1799 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1800 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1801 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1802 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1803 // In fact, that could have been the original type of p1, and p1 could have
1804 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1805 // expression (LShiftL quux 3) independently optimized to the constant 8.
1806 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1807 && (_type->isa_vect() == NULL)
1808 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1809 // t might actually be lower than _type, if _type is a unique
1810 // concrete subclass of abstract class t.
1811 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
1812 const Type* jt = t->join_speculative(_type);
1813 // In any case, do not allow the join, per se, to empty out the type.
1814 if (jt->empty() && !t->empty()) {
1815 // This can happen if a interface-typed array narrows to a class type.
1816 jt = _type;
1817 }
1818 #ifdef ASSERT
1819 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1820 // The pointers in the autobox arrays are always non-null
1821 Node* base = adr->in(AddPNode::Base);
1822 if ((base != NULL) && base->is_DecodeN()) {
1823 // Get LoadN node which loads IntegerCache.cache field
1824 base = base->in(1);
1825 }
1826 if ((base != NULL) && base->is_Con()) {
1827 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1828 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1829 // It could be narrow oop
1830 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1831 }
1832 }
1833 }
1834 #endif
1835 return jt;
1836 }
1837 }
1838 } else if (tp->base() == Type::InstPtr) {
1839 assert( off != Type::OffsetBot ||
1840 // arrays can be cast to Objects
1841 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1842 // unsafe field access may not have a constant offset
1843 C->has_unsafe_access(),
1844 "Field accesses must be precise" );
1845 // For oop loads, we expect the _type to be precise.
1846
1847 // Optimize loads from constant fields.
1848 const TypeInstPtr* tinst = tp->is_instptr();
1849 ciObject* const_oop = tinst->const_oop();
1850 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1851 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
1852 if (con_type != NULL) {
1853 return con_type;
1854 }
1855 }
1856 } else if (tp->base() == Type::KlassPtr) {
1857 assert( off != Type::OffsetBot ||
1858 // arrays can be cast to Objects
1859 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1860 // also allow array-loading from the primary supertype
1861 // array during subtype checks
1862 Opcode() == Op_LoadKlass,
1863 "Field accesses must be precise" );
1864 // For klass/static loads, we expect the _type to be precise
1865 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
1866 /* With mirrors being an indirect in the Klass*
1867 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
1868 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
1869 *
1870 * So check the type and klass of the node before the LoadP.
1871 */
1872 Node* adr2 = adr->in(MemNode::Address);
1873 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1874 if (tkls != NULL && !StressReflectiveCode) {
1875 ciKlass* klass = tkls->klass();
1876 if (klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1877 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1878 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1879 return TypeInstPtr::make(klass->java_mirror());
1880 }
1881 }
1882 }
1883
1884 const TypeKlassPtr *tkls = tp->isa_klassptr();
1885 if (tkls != NULL && !StressReflectiveCode) {
1886 ciKlass* klass = tkls->klass();
1887 if (klass->is_loaded() && tkls->klass_is_exact()) {
1888 // We are loading a field from a Klass metaobject whose identity
1889 // is known at compile time (the type is "exact" or "precise").
1890 // Check for fields we know are maintained as constants by the VM.
1891 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1892 // The field is Klass::_super_check_offset. Return its (constant) value.
1893 // (Folds up type checking code.)
1894 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1895 return TypeInt::make(klass->super_check_offset());
1896 }
1897 // Compute index into primary_supers array
1898 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1899 // Check for overflowing; use unsigned compare to handle the negative case.
1900 if( depth < ciKlass::primary_super_limit() ) {
1901 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1902 // (Folds up type checking code.)
1903 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1904 ciKlass *ss = klass->super_of_depth(depth);
1905 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1906 }
1907 const Type* aift = load_array_final_field(tkls, klass);
1908 if (aift != NULL) return aift;
1909 }
1910
1911 // We can still check if we are loading from the primary_supers array at a
1912 // shallow enough depth. Even though the klass is not exact, entries less
1913 // than or equal to its super depth are correct.
1914 if (klass->is_loaded() ) {
1915 ciType *inner = klass;
1916 while( inner->is_obj_array_klass() )
1917 inner = inner->as_obj_array_klass()->base_element_type();
1918 if( inner->is_instance_klass() &&
1919 !inner->as_instance_klass()->flags().is_interface() ) {
1920 // Compute index into primary_supers array
1921 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1922 // Check for overflowing; use unsigned compare to handle the negative case.
1923 if( depth < ciKlass::primary_super_limit() &&
1924 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1925 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1926 // (Folds up type checking code.)
1927 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1928 ciKlass *ss = klass->super_of_depth(depth);
1929 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1930 }
1931 }
1932 }
1933
1934 // If the type is enough to determine that the thing is not an array,
1935 // we can give the layout_helper a positive interval type.
1936 // This will help short-circuit some reflective code.
1937 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1938 && !klass->is_array_klass() // not directly typed as an array
1939 && !klass->is_interface() // specifically not Serializable & Cloneable
1940 && !klass->is_java_lang_Object() // not the supertype of all T[]
1941 ) {
1942 // Note: When interfaces are reliable, we can narrow the interface
1943 // test to (klass != Serializable && klass != Cloneable).
1944 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1945 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1946 // The key property of this type is that it folds up tests
1947 // for array-ness, since it proves that the layout_helper is positive.
1948 // Thus, a generic value like the basic object layout helper works fine.
1949 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1950 }
1951 }
1952
1953 // If we are loading from a freshly-allocated object, produce a zero,
1954 // if the load is provably beyond the header of the object.
1955 // (Also allow a variable load from a fresh array to produce zero.)
1956 const TypeOopPtr *tinst = tp->isa_oopptr();
1957 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1958 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1959 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1960 Node* value = can_see_stored_value(mem,phase);
1961 if (value != NULL && value->is_Con()) {
1962 assert(value->bottom_type()->higher_equal(_type),"sanity");
1963 return value->bottom_type();
1964 }
1965 }
1966
1967 if (is_instance) {
1968 // If we have an instance type and our memory input is the
1969 // programs's initial memory state, there is no matching store,
1970 // so just return a zero of the appropriate type
1971 Node *mem = in(MemNode::Memory);
1972 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1973 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1974 return Type::get_zero_type(_type->basic_type());
1975 }
1976 }
1977
1978 Node* alloc = is_new_object_mark_load(phase);
1979 if (alloc != NULL && !(alloc->Opcode() == Op_Allocate && UseBiasedLocking)) {
1980 return TypeX::make(markWord::prototype().value());
1981 }
1982
1983 return _type;
1984 }
1985
1986 //------------------------------match_edge-------------------------------------
1987 // Do we Match on this edge index or not? Match only the address.
1988 uint LoadNode::match_edge(uint idx) const {
1989 return idx == MemNode::Address;
1990 }
1991
1992 //--------------------------LoadBNode::Ideal--------------------------------------
1993 //
1994 // If the previous store is to the same address as this load,
1995 // and the value stored was larger than a byte, replace this load
1996 // with the value stored truncated to a byte. If no truncation is
1997 // needed, the replacement is done in LoadNode::Identity().
1998 //
1999 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2000 Node* mem = in(MemNode::Memory);
2001 Node* value = can_see_stored_value(mem,phase);
2002 if( value && !phase->type(value)->higher_equal( _type ) ) {
2003 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
2004 return new RShiftINode(result, phase->intcon(24));
2005 }
2006 // Identity call will handle the case where truncation is not needed.
2007 return LoadNode::Ideal(phase, can_reshape);
2008 }
2009
2010 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2011 Node* mem = in(MemNode::Memory);
2012 Node* value = can_see_stored_value(mem,phase);
2013 if (value != NULL && value->is_Con() &&
2014 !value->bottom_type()->higher_equal(_type)) {
2015 // If the input to the store does not fit with the load's result type,
2016 // it must be truncated. We can't delay until Ideal call since
2017 // a singleton Value is needed for split_thru_phi optimization.
2018 int con = value->get_int();
2019 return TypeInt::make((con << 24) >> 24);
2020 }
2021 return LoadNode::Value(phase);
2022 }
2023
2024 //--------------------------LoadUBNode::Ideal-------------------------------------
2025 //
2026 // If the previous store is to the same address as this load,
2027 // and the value stored was larger than a byte, replace this load
2028 // with the value stored truncated to a byte. If no truncation is
2029 // needed, the replacement is done in LoadNode::Identity().
2030 //
2031 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2032 Node* mem = in(MemNode::Memory);
2033 Node* value = can_see_stored_value(mem, phase);
2034 if (value && !phase->type(value)->higher_equal(_type))
2035 return new AndINode(value, phase->intcon(0xFF));
2036 // Identity call will handle the case where truncation is not needed.
2037 return LoadNode::Ideal(phase, can_reshape);
2038 }
2039
2040 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2041 Node* mem = in(MemNode::Memory);
2042 Node* value = can_see_stored_value(mem,phase);
2043 if (value != NULL && value->is_Con() &&
2044 !value->bottom_type()->higher_equal(_type)) {
2045 // If the input to the store does not fit with the load's result type,
2046 // it must be truncated. We can't delay until Ideal call since
2047 // a singleton Value is needed for split_thru_phi optimization.
2048 int con = value->get_int();
2049 return TypeInt::make(con & 0xFF);
2050 }
2051 return LoadNode::Value(phase);
2052 }
2053
2054 //--------------------------LoadUSNode::Ideal-------------------------------------
2055 //
2056 // If the previous store is to the same address as this load,
2057 // and the value stored was larger than a char, replace this load
2058 // with the value stored truncated to a char. If no truncation is
2059 // needed, the replacement is done in LoadNode::Identity().
2060 //
2061 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2062 Node* mem = in(MemNode::Memory);
2063 Node* value = can_see_stored_value(mem,phase);
2064 if( value && !phase->type(value)->higher_equal( _type ) )
2065 return new AndINode(value,phase->intcon(0xFFFF));
2066 // Identity call will handle the case where truncation is not needed.
2067 return LoadNode::Ideal(phase, can_reshape);
2068 }
2069
2070 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2071 Node* mem = in(MemNode::Memory);
2072 Node* value = can_see_stored_value(mem,phase);
2073 if (value != NULL && value->is_Con() &&
2074 !value->bottom_type()->higher_equal(_type)) {
2075 // If the input to the store does not fit with the load's result type,
2076 // it must be truncated. We can't delay until Ideal call since
2077 // a singleton Value is needed for split_thru_phi optimization.
2078 int con = value->get_int();
2079 return TypeInt::make(con & 0xFFFF);
2080 }
2081 return LoadNode::Value(phase);
2082 }
2083
2084 //--------------------------LoadSNode::Ideal--------------------------------------
2085 //
2086 // If the previous store is to the same address as this load,
2087 // and the value stored was larger than a short, replace this load
2088 // with the value stored truncated to a short. If no truncation is
2089 // needed, the replacement is done in LoadNode::Identity().
2090 //
2091 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2092 Node* mem = in(MemNode::Memory);
2093 Node* value = can_see_stored_value(mem,phase);
2094 if( value && !phase->type(value)->higher_equal( _type ) ) {
2095 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
2096 return new RShiftINode(result, phase->intcon(16));
2097 }
2098 // Identity call will handle the case where truncation is not needed.
2099 return LoadNode::Ideal(phase, can_reshape);
2100 }
2101
2102 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2103 Node* mem = in(MemNode::Memory);
2104 Node* value = can_see_stored_value(mem,phase);
2105 if (value != NULL && value->is_Con() &&
2106 !value->bottom_type()->higher_equal(_type)) {
2107 // If the input to the store does not fit with the load's result type,
2108 // it must be truncated. We can't delay until Ideal call since
2109 // a singleton Value is needed for split_thru_phi optimization.
2110 int con = value->get_int();
2111 return TypeInt::make((con << 16) >> 16);
2112 }
2113 return LoadNode::Value(phase);
2114 }
2115
2116 //=============================================================================
2117 //----------------------------LoadKlassNode::make------------------------------
2118 // Polymorphic factory method:
2119 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2120 // sanity check the alias category against the created node type
2121 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2122 assert(adr_type != NULL, "expecting TypeKlassPtr");
2123 #ifdef _LP64
2124 if (adr_type->is_ptr_to_narrowklass()) {
2125 assert(UseCompressedClassPointers, "no compressed klasses");
2126 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2127 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2128 }
2129 #endif
2130 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2131 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2132 }
2133
2134 //------------------------------Value------------------------------------------
2135 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2136 return klass_value_common(phase);
2137 }
2138
2139 // In most cases, LoadKlassNode does not have the control input set. If the control
2140 // input is set, it must not be removed (by LoadNode::Ideal()).
2141 bool LoadKlassNode::can_remove_control() const {
2142 return false;
2143 }
2144
2145 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2146 // Either input is TOP ==> the result is TOP
2147 const Type *t1 = phase->type( in(MemNode::Memory) );
2148 if (t1 == Type::TOP) return Type::TOP;
2149 Node *adr = in(MemNode::Address);
2150 const Type *t2 = phase->type( adr );
2151 if (t2 == Type::TOP) return Type::TOP;
2152 const TypePtr *tp = t2->is_ptr();
2153 if (TypePtr::above_centerline(tp->ptr()) ||
2154 tp->ptr() == TypePtr::Null) return Type::TOP;
2155
2156 // Return a more precise klass, if possible
2157 const TypeInstPtr *tinst = tp->isa_instptr();
2158 if (tinst != NULL) {
2159 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2160 int offset = tinst->offset();
2161 if (ik == phase->C->env()->Class_klass()
2162 && (offset == java_lang_Class::klass_offset() ||
2163 offset == java_lang_Class::array_klass_offset())) {
2164 // We are loading a special hidden field from a Class mirror object,
2165 // the field which points to the VM's Klass metaobject.
2166 ciType* t = tinst->java_mirror_type();
2167 // java_mirror_type returns non-null for compile-time Class constants.
2168 if (t != NULL) {
2169 // constant oop => constant klass
2170 if (offset == java_lang_Class::array_klass_offset()) {
2171 if (t->is_void()) {
2172 // We cannot create a void array. Since void is a primitive type return null
2173 // klass. Users of this result need to do a null check on the returned klass.
2174 return TypePtr::NULL_PTR;
2175 }
2176 return TypeKlassPtr::make(ciArrayKlass::make(t));
2177 }
2178 if (!t->is_klass()) {
2179 // a primitive Class (e.g., int.class) has NULL for a klass field
2180 return TypePtr::NULL_PTR;
2181 }
2182 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2183 return TypeKlassPtr::make(t->as_klass());
2184 }
2185 // non-constant mirror, so we can't tell what's going on
2186 }
2187 if( !ik->is_loaded() )
2188 return _type; // Bail out if not loaded
2189 if (offset == oopDesc::klass_offset_in_bytes()) {
2190 if (tinst->klass_is_exact()) {
2191 return TypeKlassPtr::make(ik);
2192 }
2193 // See if we can become precise: no subklasses and no interface
2194 // (Note: We need to support verified interfaces.)
2195 if (!ik->is_interface() && !ik->has_subklass()) {
2196 // Add a dependence; if any subclass added we need to recompile
2197 if (!ik->is_final()) {
2198 // %%% should use stronger assert_unique_concrete_subtype instead
2199 phase->C->dependencies()->assert_leaf_type(ik);
2200 }
2201 // Return precise klass
2202 return TypeKlassPtr::make(ik);
2203 }
2204
2205 // Return root of possible klass
2206 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2207 }
2208 }
2209
2210 // Check for loading klass from an array
2211 const TypeAryPtr *tary = tp->isa_aryptr();
2212 if( tary != NULL ) {
2213 ciKlass *tary_klass = tary->klass();
2214 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2215 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2216 if (tary->klass_is_exact()) {
2217 return TypeKlassPtr::make(tary_klass);
2218 }
2219 ciArrayKlass *ak = tary->klass()->as_array_klass();
2220 // If the klass is an object array, we defer the question to the
2221 // array component klass.
2222 if( ak->is_obj_array_klass() ) {
2223 assert( ak->is_loaded(), "" );
2224 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2225 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2226 ciInstanceKlass* ik = base_k->as_instance_klass();
2227 // See if we can become precise: no subklasses and no interface
2228 if (!ik->is_interface() && !ik->has_subklass()) {
2229 // Add a dependence; if any subclass added we need to recompile
2230 if (!ik->is_final()) {
2231 phase->C->dependencies()->assert_leaf_type(ik);
2232 }
2233 // Return precise array klass
2234 return TypeKlassPtr::make(ak);
2235 }
2236 }
2237 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2238 } else { // Found a type-array?
2239 assert( ak->is_type_array_klass(), "" );
2240 return TypeKlassPtr::make(ak); // These are always precise
2241 }
2242 }
2243 }
2244
2245 // Check for loading klass from an array klass
2246 const TypeKlassPtr *tkls = tp->isa_klassptr();
2247 if (tkls != NULL && !StressReflectiveCode) {
2248 ciKlass* klass = tkls->klass();
2249 if( !klass->is_loaded() )
2250 return _type; // Bail out if not loaded
2251 if( klass->is_obj_array_klass() &&
2252 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2253 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2254 // // Always returning precise element type is incorrect,
2255 // // e.g., element type could be object and array may contain strings
2256 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2257
2258 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2259 // according to the element type's subclassing.
2260 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2261 }
2262 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2263 tkls->offset() == in_bytes(Klass::super_offset())) {
2264 ciKlass* sup = klass->as_instance_klass()->super();
2265 // The field is Klass::_super. Return its (constant) value.
2266 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2267 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2268 }
2269 }
2270
2271 // Bailout case
2272 return LoadNode::Value(phase);
2273 }
2274
2275 //------------------------------Identity---------------------------------------
2276 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2277 // Also feed through the klass in Allocate(...klass...)._klass.
2278 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2279 return klass_identity_common(phase);
2280 }
2281
2282 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2283 Node* x = LoadNode::Identity(phase);
2284 if (x != this) return x;
2285
2286 // Take apart the address into an oop and and offset.
2287 // Return 'this' if we cannot.
2288 Node* adr = in(MemNode::Address);
2289 intptr_t offset = 0;
2290 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2291 if (base == NULL) return this;
2292 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2293 if (toop == NULL) return this;
2294
2295 // Step over potential GC barrier for OopHandle resolve
2296 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2297 if (bs->is_gc_barrier_node(base)) {
2298 base = bs->step_over_gc_barrier(base);
2299 }
2300
2301 // We can fetch the klass directly through an AllocateNode.
2302 // This works even if the klass is not constant (clone or newArray).
2303 if (offset == oopDesc::klass_offset_in_bytes()) {
2304 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2305 if (allocated_klass != NULL) {
2306 return allocated_klass;
2307 }
2308 }
2309
2310 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2311 // See inline_native_Class_query for occurrences of these patterns.
2312 // Java Example: x.getClass().isAssignableFrom(y)
2313 //
2314 // This improves reflective code, often making the Class
2315 // mirror go completely dead. (Current exception: Class
2316 // mirrors may appear in debug info, but we could clean them out by
2317 // introducing a new debug info operator for Klass.java_mirror).
2318
2319 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2320 && offset == java_lang_Class::klass_offset()) {
2321 if (base->is_Load()) {
2322 Node* base2 = base->in(MemNode::Address);
2323 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2324 Node* adr2 = base2->in(MemNode::Address);
2325 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2326 if (tkls != NULL && !tkls->empty()
2327 && (tkls->klass()->is_instance_klass() ||
2328 tkls->klass()->is_array_klass())
2329 && adr2->is_AddP()
2330 ) {
2331 int mirror_field = in_bytes(Klass::java_mirror_offset());
2332 if (tkls->offset() == mirror_field) {
2333 return adr2->in(AddPNode::Base);
2334 }
2335 }
2336 }
2337 }
2338 }
2339
2340 return this;
2341 }
2342
2343
2344 //------------------------------Value------------------------------------------
2345 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2346 const Type *t = klass_value_common(phase);
2347 if (t == Type::TOP)
2348 return t;
2349
2350 return t->make_narrowklass();
2351 }
2352
2353 //------------------------------Identity---------------------------------------
2354 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2355 // Also feed through the klass in Allocate(...klass...)._klass.
2356 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2357 Node *x = klass_identity_common(phase);
2358
2359 const Type *t = phase->type( x );
2360 if( t == Type::TOP ) return x;
2361 if( t->isa_narrowklass()) return x;
2362 assert (!t->isa_narrowoop(), "no narrow oop here");
2363
2364 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2365 }
2366
2367 //------------------------------Value-----------------------------------------
2368 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2369 // Either input is TOP ==> the result is TOP
2370 const Type *t1 = phase->type( in(MemNode::Memory) );
2371 if( t1 == Type::TOP ) return Type::TOP;
2372 Node *adr = in(MemNode::Address);
2373 const Type *t2 = phase->type( adr );
2374 if( t2 == Type::TOP ) return Type::TOP;
2375 const TypePtr *tp = t2->is_ptr();
2376 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2377 const TypeAryPtr *tap = tp->isa_aryptr();
2378 if( !tap ) return _type;
2379 return tap->size();
2380 }
2381
2382 //-------------------------------Ideal---------------------------------------
2383 // Feed through the length in AllocateArray(...length...)._length.
2384 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2385 Node* p = MemNode::Ideal_common(phase, can_reshape);
2386 if (p) return (p == NodeSentinel) ? NULL : p;
2387
2388 // Take apart the address into an oop and and offset.
2389 // Return 'this' if we cannot.
2390 Node* adr = in(MemNode::Address);
2391 intptr_t offset = 0;
2392 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2393 if (base == NULL) return NULL;
2394 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2395 if (tary == NULL) return NULL;
2396
2397 // We can fetch the length directly through an AllocateArrayNode.
2398 // This works even if the length is not constant (clone or newArray).
2399 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2400 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2401 if (alloc != NULL) {
2402 Node* allocated_length = alloc->Ideal_length();
2403 Node* len = alloc->make_ideal_length(tary, phase);
2404 if (allocated_length != len) {
2405 // New CastII improves on this.
2406 return len;
2407 }
2408 }
2409 }
2410
2411 return NULL;
2412 }
2413
2414 //------------------------------Identity---------------------------------------
2415 // Feed through the length in AllocateArray(...length...)._length.
2416 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2417 Node* x = LoadINode::Identity(phase);
2418 if (x != this) return x;
2419
2420 // Take apart the address into an oop and and offset.
2421 // Return 'this' if we cannot.
2422 Node* adr = in(MemNode::Address);
2423 intptr_t offset = 0;
2424 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2425 if (base == NULL) return this;
2426 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2427 if (tary == NULL) return this;
2428
2429 // We can fetch the length directly through an AllocateArrayNode.
2430 // This works even if the length is not constant (clone or newArray).
2431 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2432 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2433 if (alloc != NULL) {
2434 Node* allocated_length = alloc->Ideal_length();
2435 // Do not allow make_ideal_length to allocate a CastII node.
2436 Node* len = alloc->make_ideal_length(tary, phase, false);
2437 if (allocated_length == len) {
2438 // Return allocated_length only if it would not be improved by a CastII.
2439 return allocated_length;
2440 }
2441 }
2442 }
2443
2444 return this;
2445
2446 }
2447
2448 //=============================================================================
2449 //---------------------------StoreNode::make-----------------------------------
2450 // Polymorphic factory method:
2451 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2452 assert((mo == unordered || mo == release), "unexpected");
2453 Compile* C = gvn.C;
2454 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2455 ctl != NULL, "raw memory operations should have control edge");
2456
2457 switch (bt) {
2458 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2459 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2460 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2461 case T_CHAR:
2462 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2463 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2464 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2465 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2466 case T_METADATA:
2467 case T_ADDRESS:
2468 case T_OBJECT:
2469 #ifdef _LP64
2470 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2471 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2472 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2473 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2474 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2475 adr->bottom_type()->isa_rawptr())) {
2476 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2477 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2478 }
2479 #endif
2480 {
2481 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2482 }
2483 default:
2484 ShouldNotReachHere();
2485 return (StoreNode*)NULL;
2486 }
2487 }
2488
2489 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2490 bool require_atomic = true;
2491 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2492 }
2493
2494 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2495 bool require_atomic = true;
2496 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2497 }
2498
2499
2500 //--------------------------bottom_type----------------------------------------
2501 const Type *StoreNode::bottom_type() const {
2502 return Type::MEMORY;
2503 }
2504
2505 //------------------------------hash-------------------------------------------
2506 uint StoreNode::hash() const {
2507 // unroll addition of interesting fields
2508 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2509
2510 // Since they are not commoned, do not hash them:
2511 return NO_HASH;
2512 }
2513
2514 //------------------------------Ideal------------------------------------------
2515 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2516 // When a store immediately follows a relevant allocation/initialization,
2517 // try to capture it into the initialization, or hoist it above.
2518 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2519 Node* p = MemNode::Ideal_common(phase, can_reshape);
2520 if (p) return (p == NodeSentinel) ? NULL : p;
2521
2522 Node* mem = in(MemNode::Memory);
2523 Node* address = in(MemNode::Address);
2524 // Back-to-back stores to same address? Fold em up. Generally
2525 // unsafe if I have intervening uses... Also disallowed for StoreCM
2526 // since they must follow each StoreP operation. Redundant StoreCMs
2527 // are eliminated just before matching in final_graph_reshape.
2528 {
2529 Node* st = mem;
2530 // If Store 'st' has more than one use, we cannot fold 'st' away.
2531 // For example, 'st' might be the final state at a conditional
2532 // return. Or, 'st' might be used by some node which is live at
2533 // the same time 'st' is live, which might be unschedulable. So,
2534 // require exactly ONE user until such time as we clone 'mem' for
2535 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2536 // true).
2537 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2538 // Looking at a dead closed cycle of memory?
2539 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2540 assert(Opcode() == st->Opcode() ||
2541 st->Opcode() == Op_StoreVector ||
2542 Opcode() == Op_StoreVector ||
2543 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2544 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2545 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2546 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2547 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2548
2549 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2550 st->as_Store()->memory_size() <= this->memory_size()) {
2551 Node* use = st->raw_out(0);
2552 phase->igvn_rehash_node_delayed(use);
2553 if (can_reshape) {
2554 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2555 } else {
2556 // It's OK to do this in the parser, since DU info is always accurate,
2557 // and the parser always refers to nodes via SafePointNode maps.
2558 use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2559 }
2560 return this;
2561 }
2562 st = st->in(MemNode::Memory);
2563 }
2564 }
2565
2566
2567 // Capture an unaliased, unconditional, simple store into an initializer.
2568 // Or, if it is independent of the allocation, hoist it above the allocation.
2569 if (ReduceFieldZeroing && /*can_reshape &&*/
2570 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2571 InitializeNode* init = mem->in(0)->as_Initialize();
2572 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2573 if (offset > 0) {
2574 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2575 // If the InitializeNode captured me, it made a raw copy of me,
2576 // and I need to disappear.
2577 if (moved != NULL) {
2578 // %%% hack to ensure that Ideal returns a new node:
2579 mem = MergeMemNode::make(mem);
2580 return mem; // fold me away
2581 }
2582 }
2583 }
2584
2585 return NULL; // No further progress
2586 }
2587
2588 //------------------------------Value-----------------------------------------
2589 const Type* StoreNode::Value(PhaseGVN* phase) const {
2590 // Either input is TOP ==> the result is TOP
2591 const Type *t1 = phase->type( in(MemNode::Memory) );
2592 if( t1 == Type::TOP ) return Type::TOP;
2593 const Type *t2 = phase->type( in(MemNode::Address) );
2594 if( t2 == Type::TOP ) return Type::TOP;
2595 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2596 if( t3 == Type::TOP ) return Type::TOP;
2597 return Type::MEMORY;
2598 }
2599
2600 //------------------------------Identity---------------------------------------
2601 // Remove redundant stores:
2602 // Store(m, p, Load(m, p)) changes to m.
2603 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2604 Node* StoreNode::Identity(PhaseGVN* phase) {
2605 Node* mem = in(MemNode::Memory);
2606 Node* adr = in(MemNode::Address);
2607 Node* val = in(MemNode::ValueIn);
2608
2609 Node* result = this;
2610
2611 // Load then Store? Then the Store is useless
2612 if (val->is_Load() &&
2613 val->in(MemNode::Address)->eqv_uncast(adr) &&
2614 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2615 val->as_Load()->store_Opcode() == Opcode()) {
2616 result = mem;
2617 }
2618
2619 // Two stores in a row of the same value?
2620 if (result == this &&
2621 mem->is_Store() &&
2622 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2623 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2624 mem->Opcode() == Opcode()) {
2625 result = mem;
2626 }
2627
2628 // Store of zero anywhere into a freshly-allocated object?
2629 // Then the store is useless.
2630 // (It must already have been captured by the InitializeNode.)
2631 if (result == this &&
2632 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2633 // a newly allocated object is already all-zeroes everywhere
2634 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2635 result = mem;
2636 }
2637
2638 if (result == this) {
2639 // the store may also apply to zero-bits in an earlier object
2640 Node* prev_mem = find_previous_store(phase);
2641 // Steps (a), (b): Walk past independent stores to find an exact match.
2642 if (prev_mem != NULL) {
2643 Node* prev_val = can_see_stored_value(prev_mem, phase);
2644 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2645 // prev_val and val might differ by a cast; it would be good
2646 // to keep the more informative of the two.
2647 result = mem;
2648 }
2649 }
2650 }
2651 }
2652
2653 if (result != this && phase->is_IterGVN() != NULL) {
2654 MemBarNode* trailing = trailing_membar();
2655 if (trailing != NULL) {
2656 #ifdef ASSERT
2657 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2658 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects");
2659 #endif
2660 PhaseIterGVN* igvn = phase->is_IterGVN();
2661 trailing->remove(igvn);
2662 }
2663 }
2664
2665 return result;
2666 }
2667
2668 //------------------------------match_edge-------------------------------------
2669 // Do we Match on this edge index or not? Match only memory & value
2670 uint StoreNode::match_edge(uint idx) const {
2671 return idx == MemNode::Address || idx == MemNode::ValueIn;
2672 }
2673
2674 //------------------------------cmp--------------------------------------------
2675 // Do not common stores up together. They generally have to be split
2676 // back up anyways, so do not bother.
2677 bool StoreNode::cmp( const Node &n ) const {
2678 return (&n == this); // Always fail except on self
2679 }
2680
2681 //------------------------------Ideal_masked_input-----------------------------
2682 // Check for a useless mask before a partial-word store
2683 // (StoreB ... (AndI valIn conIa) )
2684 // If (conIa & mask == mask) this simplifies to
2685 // (StoreB ... (valIn) )
2686 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2687 Node *val = in(MemNode::ValueIn);
2688 if( val->Opcode() == Op_AndI ) {
2689 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2690 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2691 set_req(MemNode::ValueIn, val->in(1));
2692 return this;
2693 }
2694 }
2695 return NULL;
2696 }
2697
2698
2699 //------------------------------Ideal_sign_extended_input----------------------
2700 // Check for useless sign-extension before a partial-word store
2701 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2702 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2703 // (StoreB ... (valIn) )
2704 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2705 Node *val = in(MemNode::ValueIn);
2706 if( val->Opcode() == Op_RShiftI ) {
2707 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2708 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2709 Node *shl = val->in(1);
2710 if( shl->Opcode() == Op_LShiftI ) {
2711 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2712 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2713 set_req(MemNode::ValueIn, shl->in(1));
2714 return this;
2715 }
2716 }
2717 }
2718 }
2719 return NULL;
2720 }
2721
2722 //------------------------------value_never_loaded-----------------------------------
2723 // Determine whether there are any possible loads of the value stored.
2724 // For simplicity, we actually check if there are any loads from the
2725 // address stored to, not just for loads of the value stored by this node.
2726 //
2727 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2728 Node *adr = in(Address);
2729 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2730 if (adr_oop == NULL)
2731 return false;
2732 if (!adr_oop->is_known_instance_field())
2733 return false; // if not a distinct instance, there may be aliases of the address
2734 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2735 Node *use = adr->fast_out(i);
2736 if (use->is_Load() || use->is_LoadStore()) {
2737 return false;
2738 }
2739 }
2740 return true;
2741 }
2742
2743 MemBarNode* StoreNode::trailing_membar() const {
2744 if (is_release()) {
2745 MemBarNode* trailing_mb = NULL;
2746 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2747 Node* u = fast_out(i);
2748 if (u->is_MemBar()) {
2749 if (u->as_MemBar()->trailing_store()) {
2750 assert(u->Opcode() == Op_MemBarVolatile, "");
2751 assert(trailing_mb == NULL, "only one");
2752 trailing_mb = u->as_MemBar();
2753 #ifdef ASSERT
2754 Node* leading = u->as_MemBar()->leading_membar();
2755 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2756 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2757 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2758 #endif
2759 } else {
2760 assert(u->as_MemBar()->standalone(), "");
2761 }
2762 }
2763 }
2764 return trailing_mb;
2765 }
2766 return NULL;
2767 }
2768
2769
2770 //=============================================================================
2771 //------------------------------Ideal------------------------------------------
2772 // If the store is from an AND mask that leaves the low bits untouched, then
2773 // we can skip the AND operation. If the store is from a sign-extension
2774 // (a left shift, then right shift) we can skip both.
2775 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2776 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2777 if( progress != NULL ) return progress;
2778
2779 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2780 if( progress != NULL ) return progress;
2781
2782 // Finally check the default case
2783 return StoreNode::Ideal(phase, can_reshape);
2784 }
2785
2786 //=============================================================================
2787 //------------------------------Ideal------------------------------------------
2788 // If the store is from an AND mask that leaves the low bits untouched, then
2789 // we can skip the AND operation
2790 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2791 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2792 if( progress != NULL ) return progress;
2793
2794 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2795 if( progress != NULL ) return progress;
2796
2797 // Finally check the default case
2798 return StoreNode::Ideal(phase, can_reshape);
2799 }
2800
2801 //=============================================================================
2802 //------------------------------Identity---------------------------------------
2803 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2804 // No need to card mark when storing a null ptr
2805 Node* my_store = in(MemNode::OopStore);
2806 if (my_store->is_Store()) {
2807 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2808 if( t1 == TypePtr::NULL_PTR ) {
2809 return in(MemNode::Memory);
2810 }
2811 }
2812 return this;
2813 }
2814
2815 //=============================================================================
2816 //------------------------------Ideal---------------------------------------
2817 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2818 Node* progress = StoreNode::Ideal(phase, can_reshape);
2819 if (progress != NULL) return progress;
2820
2821 Node* my_store = in(MemNode::OopStore);
2822 if (my_store->is_MergeMem()) {
2823 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2824 set_req(MemNode::OopStore, mem);
2825 return this;
2826 }
2827
2828 return NULL;
2829 }
2830
2831 //------------------------------Value-----------------------------------------
2832 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2833 // Either input is TOP ==> the result is TOP
2834 const Type *t = phase->type( in(MemNode::Memory) );
2835 if( t == Type::TOP ) return Type::TOP;
2836 t = phase->type( in(MemNode::Address) );
2837 if( t == Type::TOP ) return Type::TOP;
2838 t = phase->type( in(MemNode::ValueIn) );
2839 if( t == Type::TOP ) return Type::TOP;
2840 // If extra input is TOP ==> the result is TOP
2841 t = phase->type( in(MemNode::OopStore) );
2842 if( t == Type::TOP ) return Type::TOP;
2843
2844 return StoreNode::Value( phase );
2845 }
2846
2847
2848 //=============================================================================
2849 //----------------------------------SCMemProjNode------------------------------
2850 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2851 {
2852 return bottom_type();
2853 }
2854
2855 //=============================================================================
2856 //----------------------------------LoadStoreNode------------------------------
2857 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2858 : Node(required),
2859 _type(rt),
2860 _adr_type(at),
2861 _barrier(0)
2862 {
2863 init_req(MemNode::Control, c );
2864 init_req(MemNode::Memory , mem);
2865 init_req(MemNode::Address, adr);
2866 init_req(MemNode::ValueIn, val);
2867 init_class_id(Class_LoadStore);
2868 }
2869
2870 uint LoadStoreNode::ideal_reg() const {
2871 return _type->ideal_reg();
2872 }
2873
2874 bool LoadStoreNode::result_not_used() const {
2875 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2876 Node *x = fast_out(i);
2877 if (x->Opcode() == Op_SCMemProj) continue;
2878 return false;
2879 }
2880 return true;
2881 }
2882
2883 MemBarNode* LoadStoreNode::trailing_membar() const {
2884 MemBarNode* trailing = NULL;
2885 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2886 Node* u = fast_out(i);
2887 if (u->is_MemBar()) {
2888 if (u->as_MemBar()->trailing_load_store()) {
2889 assert(u->Opcode() == Op_MemBarAcquire, "");
2890 assert(trailing == NULL, "only one");
2891 trailing = u->as_MemBar();
2892 #ifdef ASSERT
2893 Node* leading = trailing->leading_membar();
2894 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2895 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
2896 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
2897 #endif
2898 } else {
2899 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
2900 }
2901 }
2902 }
2903
2904 return trailing;
2905 }
2906
2907 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2908
2909 //=============================================================================
2910 //----------------------------------LoadStoreConditionalNode--------------------
2911 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2912 init_req(ExpectedIn, ex );
2913 }
2914
2915 //=============================================================================
2916 //-------------------------------adr_type--------------------------------------
2917 const TypePtr* ClearArrayNode::adr_type() const {
2918 Node *adr = in(3);
2919 if (adr == NULL) return NULL; // node is dead
2920 return MemNode::calculate_adr_type(adr->bottom_type());
2921 }
2922
2923 //------------------------------match_edge-------------------------------------
2924 // Do we Match on this edge index or not? Do not match memory
2925 uint ClearArrayNode::match_edge(uint idx) const {
2926 return idx > 1;
2927 }
2928
2929 //------------------------------Identity---------------------------------------
2930 // Clearing a zero length array does nothing
2931 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
2932 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2933 }
2934
2935 //------------------------------Idealize---------------------------------------
2936 // Clearing a short array is faster with stores
2937 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2938 // Already know this is a large node, do not try to ideal it
2939 if (!IdealizeClearArrayNode || _is_large) return NULL;
2940
2941 const int unit = BytesPerLong;
2942 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2943 if (!t) return NULL;
2944 if (!t->is_con()) return NULL;
2945 intptr_t raw_count = t->get_con();
2946 intptr_t size = raw_count;
2947 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2948 // Clearing nothing uses the Identity call.
2949 // Negative clears are possible on dead ClearArrays
2950 // (see jck test stmt114.stmt11402.val).
2951 if (size <= 0 || size % unit != 0) return NULL;
2952 intptr_t count = size / unit;
2953 // Length too long; communicate this to matchers and assemblers.
2954 // Assemblers are responsible to produce fast hardware clears for it.
2955 if (size > InitArrayShortSize) {
2956 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
2957 }
2958 Node *mem = in(1);
2959 if( phase->type(mem)==Type::TOP ) return NULL;
2960 Node *adr = in(3);
2961 const Type* at = phase->type(adr);
2962 if( at==Type::TOP ) return NULL;
2963 const TypePtr* atp = at->isa_ptr();
2964 // adjust atp to be the correct array element address type
2965 if (atp == NULL) atp = TypePtr::BOTTOM;
2966 else atp = atp->add_offset(Type::OffsetBot);
2967 // Get base for derived pointer purposes
2968 if( adr->Opcode() != Op_AddP ) Unimplemented();
2969 Node *base = adr->in(1);
2970
2971 Node *zero = phase->makecon(TypeLong::ZERO);
2972 Node *off = phase->MakeConX(BytesPerLong);
2973 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2974 count--;
2975 while( count-- ) {
2976 mem = phase->transform(mem);
2977 adr = phase->transform(new AddPNode(base,adr,off));
2978 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2979 }
2980 return mem;
2981 }
2982
2983 //----------------------------step_through----------------------------------
2984 // Return allocation input memory edge if it is different instance
2985 // or itself if it is the one we are looking for.
2986 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2987 Node* n = *np;
2988 assert(n->is_ClearArray(), "sanity");
2989 intptr_t offset;
2990 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2991 // This method is called only before Allocate nodes are expanded
2992 // during macro nodes expansion. Before that ClearArray nodes are
2993 // only generated in PhaseMacroExpand::generate_arraycopy() (before
2994 // Allocate nodes are expanded) which follows allocations.
2995 assert(alloc != NULL, "should have allocation");
2996 if (alloc->_idx == instance_id) {
2997 // Can not bypass initialization of the instance we are looking for.
2998 return false;
2999 }
3000 // Otherwise skip it.
3001 InitializeNode* init = alloc->initialization();
3002 if (init != NULL)
3003 *np = init->in(TypeFunc::Memory);
3004 else
3005 *np = alloc->in(TypeFunc::Memory);
3006 return true;
3007 }
3008
3009 //----------------------------clear_memory-------------------------------------
3010 // Generate code to initialize object storage to zero.
3011 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3012 intptr_t start_offset,
3013 Node* end_offset,
3014 PhaseGVN* phase) {
3015 intptr_t offset = start_offset;
3016
3017 int unit = BytesPerLong;
3018 if ((offset % unit) != 0) {
3019 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
3020 adr = phase->transform(adr);
3021 const TypePtr* atp = TypeRawPtr::BOTTOM;
3022 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3023 mem = phase->transform(mem);
3024 offset += BytesPerInt;
3025 }
3026 assert((offset % unit) == 0, "");
3027
3028 // Initialize the remaining stuff, if any, with a ClearArray.
3029 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
3030 }
3031
3032 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3033 Node* start_offset,
3034 Node* end_offset,
3035 PhaseGVN* phase) {
3036 if (start_offset == end_offset) {
3037 // nothing to do
3038 return mem;
3039 }
3040
3041 int unit = BytesPerLong;
3042 Node* zbase = start_offset;
3043 Node* zend = end_offset;
3044
3045 // Scale to the unit required by the CPU:
3046 if (!Matcher::init_array_count_is_in_bytes) {
3047 Node* shift = phase->intcon(exact_log2(unit));
3048 zbase = phase->transform(new URShiftXNode(zbase, shift) );
3049 zend = phase->transform(new URShiftXNode(zend, shift) );
3050 }
3051
3052 // Bulk clear double-words
3053 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
3054 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
3055 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
3056 return phase->transform(mem);
3057 }
3058
3059 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3060 intptr_t start_offset,
3061 intptr_t end_offset,
3062 PhaseGVN* phase) {
3063 if (start_offset == end_offset) {
3064 // nothing to do
3065 return mem;
3066 }
3067
3068 assert((end_offset % BytesPerInt) == 0, "odd end offset");
3069 intptr_t done_offset = end_offset;
3070 if ((done_offset % BytesPerLong) != 0) {
3071 done_offset -= BytesPerInt;
3072 }
3073 if (done_offset > start_offset) {
3074 mem = clear_memory(ctl, mem, dest,
3075 start_offset, phase->MakeConX(done_offset), phase);
3076 }
3077 if (done_offset < end_offset) { // emit the final 32-bit store
3078 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3079 adr = phase->transform(adr);
3080 const TypePtr* atp = TypeRawPtr::BOTTOM;
3081 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3082 mem = phase->transform(mem);
3083 done_offset += BytesPerInt;
3084 }
3085 assert(done_offset == end_offset, "");
3086 return mem;
3087 }
3088
3089 //=============================================================================
3090 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3091 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
3092 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3093 #ifdef ASSERT
3094 , _pair_idx(0)
3095 #endif
3096 {
3097 init_class_id(Class_MemBar);
3098 Node* top = C->top();
3099 init_req(TypeFunc::I_O,top);
3100 init_req(TypeFunc::FramePtr,top);
3101 init_req(TypeFunc::ReturnAdr,top);
3102 if (precedent != NULL)
3103 init_req(TypeFunc::Parms, precedent);
3104 }
3105
3106 //------------------------------cmp--------------------------------------------
3107 uint MemBarNode::hash() const { return NO_HASH; }
3108 bool MemBarNode::cmp( const Node &n ) const {
3109 return (&n == this); // Always fail except on self
3110 }
3111
3112 //------------------------------make-------------------------------------------
3113 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3114 switch (opcode) {
3115 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
3116 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
3117 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
3118 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
3119 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3120 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3121 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
3122 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
3123 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
3124 case Op_Initialize: return new InitializeNode(C, atp, pn);
3125 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
3126 default: ShouldNotReachHere(); return NULL;
3127 }
3128 }
3129
3130 void MemBarNode::remove(PhaseIterGVN *igvn) {
3131 if (outcnt() != 2) {
3132 return;
3133 }
3134 if (trailing_store() || trailing_load_store()) {
3135 MemBarNode* leading = leading_membar();
3136 if (leading != NULL) {
3137 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3138 leading->remove(igvn);
3139 }
3140 }
3141 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3142 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3143 }
3144
3145 //------------------------------Ideal------------------------------------------
3146 // Return a node which is more "ideal" than the current node. Strip out
3147 // control copies
3148 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3149 if (remove_dead_region(phase, can_reshape)) return this;
3150 // Don't bother trying to transform a dead node
3151 if (in(0) && in(0)->is_top()) {
3152 return NULL;
3153 }
3154
3155 bool progress = false;
3156 // Eliminate volatile MemBars for scalar replaced objects.
3157 if (can_reshape && req() == (Precedent+1)) {
3158 bool eliminate = false;
3159 int opc = Opcode();
3160 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3161 // Volatile field loads and stores.
3162 Node* my_mem = in(MemBarNode::Precedent);
3163 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3164 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3165 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3166 // replace this Precedent (decodeN) with the Load instead.
3167 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3168 Node* load_node = my_mem->in(1);
3169 set_req(MemBarNode::Precedent, load_node);
3170 phase->is_IterGVN()->_worklist.push(my_mem);
3171 my_mem = load_node;
3172 } else {
3173 assert(my_mem->unique_out() == this, "sanity");
3174 del_req(Precedent);
3175 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3176 my_mem = NULL;
3177 }
3178 progress = true;
3179 }
3180 if (my_mem != NULL && my_mem->is_Mem()) {
3181 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3182 // Check for scalar replaced object reference.
3183 if( t_oop != NULL && t_oop->is_known_instance_field() &&
3184 t_oop->offset() != Type::OffsetBot &&
3185 t_oop->offset() != Type::OffsetTop) {
3186 eliminate = true;
3187 }
3188 }
3189 } else if (opc == Op_MemBarRelease) {
3190 // Final field stores.
3191 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3192 if ((alloc != NULL) && alloc->is_Allocate() &&
3193 alloc->as_Allocate()->does_not_escape_thread()) {
3194 // The allocated object does not escape.
3195 eliminate = true;
3196 }
3197 }
3198 if (eliminate) {
3199 // Replace MemBar projections by its inputs.
3200 PhaseIterGVN* igvn = phase->is_IterGVN();
3201 remove(igvn);
3202 // Must return either the original node (now dead) or a new node
3203 // (Do not return a top here, since that would break the uniqueness of top.)
3204 return new ConINode(TypeInt::ZERO);
3205 }
3206 }
3207 return progress ? this : NULL;
3208 }
3209
3210 //------------------------------Value------------------------------------------
3211 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3212 if( !in(0) ) return Type::TOP;
3213 if( phase->type(in(0)) == Type::TOP )
3214 return Type::TOP;
3215 return TypeTuple::MEMBAR;
3216 }
3217
3218 //------------------------------match------------------------------------------
3219 // Construct projections for memory.
3220 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3221 switch (proj->_con) {
3222 case TypeFunc::Control:
3223 case TypeFunc::Memory:
3224 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3225 }
3226 ShouldNotReachHere();
3227 return NULL;
3228 }
3229
3230 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3231 trailing->_kind = TrailingStore;
3232 leading->_kind = LeadingStore;
3233 #ifdef ASSERT
3234 trailing->_pair_idx = leading->_idx;
3235 leading->_pair_idx = leading->_idx;
3236 #endif
3237 }
3238
3239 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3240 trailing->_kind = TrailingLoadStore;
3241 leading->_kind = LeadingLoadStore;
3242 #ifdef ASSERT
3243 trailing->_pair_idx = leading->_idx;
3244 leading->_pair_idx = leading->_idx;
3245 #endif
3246 }
3247
3248 MemBarNode* MemBarNode::trailing_membar() const {
3249 ResourceMark rm;
3250 Node* trailing = (Node*)this;
3251 VectorSet seen(Thread::current()->resource_area());
3252 Node_Stack multis(0);
3253 do {
3254 Node* c = trailing;
3255 uint i = 0;
3256 do {
3257 trailing = NULL;
3258 for (; i < c->outcnt(); i++) {
3259 Node* next = c->raw_out(i);
3260 if (next != c && next->is_CFG()) {
3261 if (c->is_MultiBranch()) {
3262 if (multis.node() == c) {
3263 multis.set_index(i+1);
3264 } else {
3265 multis.push(c, i+1);
3266 }
3267 }
3268 trailing = next;
3269 break;
3270 }
3271 }
3272 if (trailing != NULL && !seen.test_set(trailing->_idx)) {
3273 break;
3274 }
3275 while (multis.size() > 0) {
3276 c = multis.node();
3277 i = multis.index();
3278 if (i < c->req()) {
3279 break;
3280 }
3281 multis.pop();
3282 }
3283 } while (multis.size() > 0);
3284 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
3285
3286 MemBarNode* mb = trailing->as_MemBar();
3287 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3288 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3289 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3290 return mb;
3291 }
3292
3293 MemBarNode* MemBarNode::leading_membar() const {
3294 ResourceMark rm;
3295 VectorSet seen(Thread::current()->resource_area());
3296 Node_Stack regions(0);
3297 Node* leading = in(0);
3298 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3299 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) {
3300 leading = NULL;
3301 while (regions.size() > 0 && leading == NULL) {
3302 Node* r = regions.node();
3303 uint i = regions.index();
3304 if (i < r->req()) {
3305 leading = r->in(i);
3306 regions.set_index(i+1);
3307 } else {
3308 regions.pop();
3309 }
3310 }
3311 if (leading == NULL) {
3312 assert(regions.size() == 0, "all paths should have been tried");
3313 return NULL;
3314 }
3315 }
3316 if (leading->is_Region()) {
3317 regions.push(leading, 2);
3318 leading = leading->in(1);
3319 } else {
3320 leading = leading->in(0);
3321 }
3322 }
3323 #ifdef ASSERT
3324 Unique_Node_List wq;
3325 wq.push((Node*)this);
3326 uint found = 0;
3327 for (uint i = 0; i < wq.size(); i++) {
3328 Node* n = wq.at(i);
3329 if (n->is_Region()) {
3330 for (uint j = 1; j < n->req(); j++) {
3331 Node* in = n->in(j);
3332 if (in != NULL && !in->is_top()) {
3333 wq.push(in);
3334 }
3335 }
3336 } else {
3337 if (n->is_MemBar() && n->as_MemBar()->leading()) {
3338 assert(n == leading, "consistency check failed");
3339 found++;
3340 } else {
3341 Node* in = n->in(0);
3342 if (in != NULL && !in->is_top()) {
3343 wq.push(in);
3344 }
3345 }
3346 }
3347 }
3348 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed");
3349 #endif
3350 if (leading == NULL) {
3351 return NULL;
3352 }
3353 MemBarNode* mb = leading->as_MemBar();
3354 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3355 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3356 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3357 return mb;
3358 }
3359
3360 //===========================InitializeNode====================================
3361 // SUMMARY:
3362 // This node acts as a memory barrier on raw memory, after some raw stores.
3363 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3364 // The Initialize can 'capture' suitably constrained stores as raw inits.
3365 // It can coalesce related raw stores into larger units (called 'tiles').
3366 // It can avoid zeroing new storage for memory units which have raw inits.
3367 // At macro-expansion, it is marked 'complete', and does not optimize further.
3368 //
3369 // EXAMPLE:
3370 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3371 // ctl = incoming control; mem* = incoming memory
3372 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3373 // First allocate uninitialized memory and fill in the header:
3374 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3375 // ctl := alloc.Control; mem* := alloc.Memory*
3376 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3377 // Then initialize to zero the non-header parts of the raw memory block:
3378 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3379 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3380 // After the initialize node executes, the object is ready for service:
3381 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3382 // Suppose its body is immediately initialized as {1,2}:
3383 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3384 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3385 // mem.SLICE(#short[*]) := store2
3386 //
3387 // DETAILS:
3388 // An InitializeNode collects and isolates object initialization after
3389 // an AllocateNode and before the next possible safepoint. As a
3390 // memory barrier (MemBarNode), it keeps critical stores from drifting
3391 // down past any safepoint or any publication of the allocation.
3392 // Before this barrier, a newly-allocated object may have uninitialized bits.
3393 // After this barrier, it may be treated as a real oop, and GC is allowed.
3394 //
3395 // The semantics of the InitializeNode include an implicit zeroing of
3396 // the new object from object header to the end of the object.
3397 // (The object header and end are determined by the AllocateNode.)
3398 //
3399 // Certain stores may be added as direct inputs to the InitializeNode.
3400 // These stores must update raw memory, and they must be to addresses
3401 // derived from the raw address produced by AllocateNode, and with
3402 // a constant offset. They must be ordered by increasing offset.
3403 // The first one is at in(RawStores), the last at in(req()-1).
3404 // Unlike most memory operations, they are not linked in a chain,
3405 // but are displayed in parallel as users of the rawmem output of
3406 // the allocation.
3407 //
3408 // (See comments in InitializeNode::capture_store, which continue
3409 // the example given above.)
3410 //
3411 // When the associated Allocate is macro-expanded, the InitializeNode
3412 // may be rewritten to optimize collected stores. A ClearArrayNode
3413 // may also be created at that point to represent any required zeroing.
3414 // The InitializeNode is then marked 'complete', prohibiting further
3415 // capturing of nearby memory operations.
3416 //
3417 // During macro-expansion, all captured initializations which store
3418 // constant values of 32 bits or smaller are coalesced (if advantageous)
3419 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3420 // initialized in fewer memory operations. Memory words which are
3421 // covered by neither tiles nor non-constant stores are pre-zeroed
3422 // by explicit stores of zero. (The code shape happens to do all
3423 // zeroing first, then all other stores, with both sequences occurring
3424 // in order of ascending offsets.)
3425 //
3426 // Alternatively, code may be inserted between an AllocateNode and its
3427 // InitializeNode, to perform arbitrary initialization of the new object.
3428 // E.g., the object copying intrinsics insert complex data transfers here.
3429 // The initialization must then be marked as 'complete' disable the
3430 // built-in zeroing semantics and the collection of initializing stores.
3431 //
3432 // While an InitializeNode is incomplete, reads from the memory state
3433 // produced by it are optimizable if they match the control edge and
3434 // new oop address associated with the allocation/initialization.
3435 // They return a stored value (if the offset matches) or else zero.
3436 // A write to the memory state, if it matches control and address,
3437 // and if it is to a constant offset, may be 'captured' by the
3438 // InitializeNode. It is cloned as a raw memory operation and rewired
3439 // inside the initialization, to the raw oop produced by the allocation.
3440 // Operations on addresses which are provably distinct (e.g., to
3441 // other AllocateNodes) are allowed to bypass the initialization.
3442 //
3443 // The effect of all this is to consolidate object initialization
3444 // (both arrays and non-arrays, both piecewise and bulk) into a
3445 // single location, where it can be optimized as a unit.
3446 //
3447 // Only stores with an offset less than TrackedInitializationLimit words
3448 // will be considered for capture by an InitializeNode. This puts a
3449 // reasonable limit on the complexity of optimized initializations.
3450
3451 //---------------------------InitializeNode------------------------------------
3452 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3453 : MemBarNode(C, adr_type, rawoop),
3454 _is_complete(Incomplete), _does_not_escape(false)
3455 {
3456 init_class_id(Class_Initialize);
3457
3458 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3459 assert(in(RawAddress) == rawoop, "proper init");
3460 // Note: allocation() can be NULL, for secondary initialization barriers
3461 }
3462
3463 // Since this node is not matched, it will be processed by the
3464 // register allocator. Declare that there are no constraints
3465 // on the allocation of the RawAddress edge.
3466 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3467 // This edge should be set to top, by the set_complete. But be conservative.
3468 if (idx == InitializeNode::RawAddress)
3469 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3470 return RegMask::Empty;
3471 }
3472
3473 Node* InitializeNode::memory(uint alias_idx) {
3474 Node* mem = in(Memory);
3475 if (mem->is_MergeMem()) {
3476 return mem->as_MergeMem()->memory_at(alias_idx);
3477 } else {
3478 // incoming raw memory is not split
3479 return mem;
3480 }
3481 }
3482
3483 bool InitializeNode::is_non_zero() {
3484 if (is_complete()) return false;
3485 remove_extra_zeroes();
3486 return (req() > RawStores);
3487 }
3488
3489 void InitializeNode::set_complete(PhaseGVN* phase) {
3490 assert(!is_complete(), "caller responsibility");
3491 _is_complete = Complete;
3492
3493 // After this node is complete, it contains a bunch of
3494 // raw-memory initializations. There is no need for
3495 // it to have anything to do with non-raw memory effects.
3496 // Therefore, tell all non-raw users to re-optimize themselves,
3497 // after skipping the memory effects of this initialization.
3498 PhaseIterGVN* igvn = phase->is_IterGVN();
3499 if (igvn) igvn->add_users_to_worklist(this);
3500 }
3501
3502 // convenience function
3503 // return false if the init contains any stores already
3504 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3505 InitializeNode* init = initialization();
3506 if (init == NULL || init->is_complete()) return false;
3507 init->remove_extra_zeroes();
3508 // for now, if this allocation has already collected any inits, bail:
3509 if (init->is_non_zero()) return false;
3510 init->set_complete(phase);
3511 return true;
3512 }
3513
3514 void InitializeNode::remove_extra_zeroes() {
3515 if (req() == RawStores) return;
3516 Node* zmem = zero_memory();
3517 uint fill = RawStores;
3518 for (uint i = fill; i < req(); i++) {
3519 Node* n = in(i);
3520 if (n->is_top() || n == zmem) continue; // skip
3521 if (fill < i) set_req(fill, n); // compact
3522 ++fill;
3523 }
3524 // delete any empty spaces created:
3525 while (fill < req()) {
3526 del_req(fill);
3527 }
3528 }
3529
3530 // Helper for remembering which stores go with which offsets.
3531 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3532 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3533 intptr_t offset = -1;
3534 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3535 phase, offset);
3536 if (base == NULL) return -1; // something is dead,
3537 if (offset < 0) return -1; // dead, dead
3538 return offset;
3539 }
3540
3541 // Helper for proving that an initialization expression is
3542 // "simple enough" to be folded into an object initialization.
3543 // Attempts to prove that a store's initial value 'n' can be captured
3544 // within the initialization without creating a vicious cycle, such as:
3545 // { Foo p = new Foo(); p.next = p; }
3546 // True for constants and parameters and small combinations thereof.
3547 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
3548 ResourceMark rm;
3549 Unique_Node_List worklist;
3550 worklist.push(value);
3551
3552 uint complexity_limit = 20;
3553 for (uint j = 0; j < worklist.size(); j++) {
3554 if (j >= complexity_limit) {
3555 return false; // Bail out if processed too many nodes
3556 }
3557
3558 Node* n = worklist.at(j);
3559 if (n == NULL) continue; // (can this really happen?)
3560 if (n->is_Proj()) n = n->in(0);
3561 if (n == this) return false; // found a cycle
3562 if (n->is_Con()) continue;
3563 if (n->is_Start()) continue; // params, etc., are OK
3564 if (n->is_Root()) continue; // even better
3565
3566 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
3567 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
3568 continue;
3569 }
3570
3571 Node* ctl = n->in(0);
3572 if (ctl != NULL && !ctl->is_top()) {
3573 if (ctl->is_Proj()) ctl = ctl->in(0);
3574 if (ctl == this) return false;
3575
3576 // If we already know that the enclosing memory op is pinned right after
3577 // the init, then any control flow that the store has picked up
3578 // must have preceded the init, or else be equal to the init.
3579 // Even after loop optimizations (which might change control edges)
3580 // a store is never pinned *before* the availability of its inputs.
3581 if (!MemNode::all_controls_dominate(n, this))
3582 return false; // failed to prove a good control
3583 }
3584
3585 // Check data edges for possible dependencies on 'this'.
3586 for (uint i = 1; i < n->req(); i++) {
3587 Node* m = n->in(i);
3588 if (m == NULL || m == n || m->is_top()) continue;
3589
3590 // Only process data inputs once
3591 worklist.push(m);
3592 }
3593 }
3594
3595 return true;
3596 }
3597
3598 // Here are all the checks a Store must pass before it can be moved into
3599 // an initialization. Returns zero if a check fails.
3600 // On success, returns the (constant) offset to which the store applies,
3601 // within the initialized memory.
3602 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
3603 const int FAIL = 0;
3604 if (st->req() != MemNode::ValueIn + 1)
3605 return FAIL; // an inscrutable StoreNode (card mark?)
3606 Node* ctl = st->in(MemNode::Control);
3607 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3608 return FAIL; // must be unconditional after the initialization
3609 Node* mem = st->in(MemNode::Memory);
3610 if (!(mem->is_Proj() && mem->in(0) == this))
3611 return FAIL; // must not be preceded by other stores
3612 Node* adr = st->in(MemNode::Address);
3613 intptr_t offset;
3614 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3615 if (alloc == NULL)
3616 return FAIL; // inscrutable address
3617 if (alloc != allocation())
3618 return FAIL; // wrong allocation! (store needs to float up)
3619 int size_in_bytes = st->memory_size();
3620 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
3621 return FAIL; // mismatched access
3622 }
3623 Node* val = st->in(MemNode::ValueIn);
3624
3625 if (!detect_init_independence(val, phase))
3626 return FAIL; // stored value must be 'simple enough'
3627
3628 // The Store can be captured only if nothing after the allocation
3629 // and before the Store is using the memory location that the store
3630 // overwrites.
3631 bool failed = false;
3632 // If is_complete_with_arraycopy() is true the shape of the graph is
3633 // well defined and is safe so no need for extra checks.
3634 if (!is_complete_with_arraycopy()) {
3635 // We are going to look at each use of the memory state following
3636 // the allocation to make sure nothing reads the memory that the
3637 // Store writes.
3638 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3639 int alias_idx = phase->C->get_alias_index(t_adr);
3640 ResourceMark rm;
3641 Unique_Node_List mems;
3642 mems.push(mem);
3643 Node* unique_merge = NULL;
3644 for (uint next = 0; next < mems.size(); ++next) {
3645 Node *m = mems.at(next);
3646 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3647 Node *n = m->fast_out(j);
3648 if (n->outcnt() == 0) {
3649 continue;
3650 }
3651 if (n == st) {
3652 continue;
3653 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3654 // If the control of this use is different from the control
3655 // of the Store which is right after the InitializeNode then
3656 // this node cannot be between the InitializeNode and the
3657 // Store.
3658 continue;
3659 } else if (n->is_MergeMem()) {
3660 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3661 // We can hit a MergeMemNode (that will likely go away
3662 // later) that is a direct use of the memory state
3663 // following the InitializeNode on the same slice as the
3664 // store node that we'd like to capture. We need to check
3665 // the uses of the MergeMemNode.
3666 mems.push(n);
3667 }
3668 } else if (n->is_Mem()) {
3669 Node* other_adr = n->in(MemNode::Address);
3670 if (other_adr == adr) {
3671 failed = true;
3672 break;
3673 } else {
3674 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3675 if (other_t_adr != NULL) {
3676 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3677 if (other_alias_idx == alias_idx) {
3678 // A load from the same memory slice as the store right
3679 // after the InitializeNode. We check the control of the
3680 // object/array that is loaded from. If it's the same as
3681 // the store control then we cannot capture the store.
3682 assert(!n->is_Store(), "2 stores to same slice on same control?");
3683 Node* base = other_adr;
3684 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3685 base = base->in(AddPNode::Base);
3686 if (base != NULL) {
3687 base = base->uncast();
3688 if (base->is_Proj() && base->in(0) == alloc) {
3689 failed = true;
3690 break;
3691 }
3692 }
3693 }
3694 }
3695 }
3696 } else {
3697 failed = true;
3698 break;
3699 }
3700 }
3701 }
3702 }
3703 if (failed) {
3704 if (!can_reshape) {
3705 // We decided we couldn't capture the store during parsing. We
3706 // should try again during the next IGVN once the graph is
3707 // cleaner.
3708 phase->C->record_for_igvn(st);
3709 }
3710 return FAIL;
3711 }
3712
3713 return offset; // success
3714 }
3715
3716 // Find the captured store in(i) which corresponds to the range
3717 // [start..start+size) in the initialized object.
3718 // If there is one, return its index i. If there isn't, return the
3719 // negative of the index where it should be inserted.
3720 // Return 0 if the queried range overlaps an initialization boundary
3721 // or if dead code is encountered.
3722 // If size_in_bytes is zero, do not bother with overlap checks.
3723 int InitializeNode::captured_store_insertion_point(intptr_t start,
3724 int size_in_bytes,
3725 PhaseTransform* phase) {
3726 const int FAIL = 0, MAX_STORE = BytesPerLong;
3727
3728 if (is_complete())
3729 return FAIL; // arraycopy got here first; punt
3730
3731 assert(allocation() != NULL, "must be present");
3732
3733 // no negatives, no header fields:
3734 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3735
3736 // after a certain size, we bail out on tracking all the stores:
3737 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3738 if (start >= ti_limit) return FAIL;
3739
3740 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3741 if (i >= limit) return -(int)i; // not found; here is where to put it
3742
3743 Node* st = in(i);
3744 intptr_t st_off = get_store_offset(st, phase);
3745 if (st_off < 0) {
3746 if (st != zero_memory()) {
3747 return FAIL; // bail out if there is dead garbage
3748 }
3749 } else if (st_off > start) {
3750 // ...we are done, since stores are ordered
3751 if (st_off < start + size_in_bytes) {
3752 return FAIL; // the next store overlaps
3753 }
3754 return -(int)i; // not found; here is where to put it
3755 } else if (st_off < start) {
3756 if (size_in_bytes != 0 &&
3757 start < st_off + MAX_STORE &&
3758 start < st_off + st->as_Store()->memory_size()) {
3759 return FAIL; // the previous store overlaps
3760 }
3761 } else {
3762 if (size_in_bytes != 0 &&
3763 st->as_Store()->memory_size() != size_in_bytes) {
3764 return FAIL; // mismatched store size
3765 }
3766 return i;
3767 }
3768
3769 ++i;
3770 }
3771 }
3772
3773 // Look for a captured store which initializes at the offset 'start'
3774 // with the given size. If there is no such store, and no other
3775 // initialization interferes, then return zero_memory (the memory
3776 // projection of the AllocateNode).
3777 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3778 PhaseTransform* phase) {
3779 assert(stores_are_sane(phase), "");
3780 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3781 if (i == 0) {
3782 return NULL; // something is dead
3783 } else if (i < 0) {
3784 return zero_memory(); // just primordial zero bits here
3785 } else {
3786 Node* st = in(i); // here is the store at this position
3787 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3788 return st;
3789 }
3790 }
3791
3792 // Create, as a raw pointer, an address within my new object at 'offset'.
3793 Node* InitializeNode::make_raw_address(intptr_t offset,
3794 PhaseTransform* phase) {
3795 Node* addr = in(RawAddress);
3796 if (offset != 0) {
3797 Compile* C = phase->C;
3798 addr = phase->transform( new AddPNode(C->top(), addr,
3799 phase->MakeConX(offset)) );
3800 }
3801 return addr;
3802 }
3803
3804 // Clone the given store, converting it into a raw store
3805 // initializing a field or element of my new object.
3806 // Caller is responsible for retiring the original store,
3807 // with subsume_node or the like.
3808 //
3809 // From the example above InitializeNode::InitializeNode,
3810 // here are the old stores to be captured:
3811 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3812 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3813 //
3814 // Here is the changed code; note the extra edges on init:
3815 // alloc = (Allocate ...)
3816 // rawoop = alloc.RawAddress
3817 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3818 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3819 // init = (Initialize alloc.Control alloc.Memory rawoop
3820 // rawstore1 rawstore2)
3821 //
3822 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3823 PhaseGVN* phase, bool can_reshape) {
3824 assert(stores_are_sane(phase), "");
3825
3826 if (start < 0) return NULL;
3827 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3828
3829 Compile* C = phase->C;
3830 int size_in_bytes = st->memory_size();
3831 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3832 if (i == 0) return NULL; // bail out
3833 Node* prev_mem = NULL; // raw memory for the captured store
3834 if (i > 0) {
3835 prev_mem = in(i); // there is a pre-existing store under this one
3836 set_req(i, C->top()); // temporarily disconnect it
3837 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3838 } else {
3839 i = -i; // no pre-existing store
3840 prev_mem = zero_memory(); // a slice of the newly allocated object
3841 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3842 set_req(--i, C->top()); // reuse this edge; it has been folded away
3843 else
3844 ins_req(i, C->top()); // build a new edge
3845 }
3846 Node* new_st = st->clone();
3847 new_st->set_req(MemNode::Control, in(Control));
3848 new_st->set_req(MemNode::Memory, prev_mem);
3849 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3850 new_st = phase->transform(new_st);
3851
3852 // At this point, new_st might have swallowed a pre-existing store
3853 // at the same offset, or perhaps new_st might have disappeared,
3854 // if it redundantly stored the same value (or zero to fresh memory).
3855
3856 // In any case, wire it in:
3857 phase->igvn_rehash_node_delayed(this);
3858 set_req(i, new_st);
3859
3860 // The caller may now kill the old guy.
3861 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3862 assert(check_st == new_st || check_st == NULL, "must be findable");
3863 assert(!is_complete(), "");
3864 return new_st;
3865 }
3866
3867 static bool store_constant(jlong* tiles, int num_tiles,
3868 intptr_t st_off, int st_size,
3869 jlong con) {
3870 if ((st_off & (st_size-1)) != 0)
3871 return false; // strange store offset (assume size==2**N)
3872 address addr = (address)tiles + st_off;
3873 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3874 switch (st_size) {
3875 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3876 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3877 case sizeof(jint): *(jint*) addr = (jint) con; break;
3878 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3879 default: return false; // strange store size (detect size!=2**N here)
3880 }
3881 return true; // return success to caller
3882 }
3883
3884 // Coalesce subword constants into int constants and possibly
3885 // into long constants. The goal, if the CPU permits,
3886 // is to initialize the object with a small number of 64-bit tiles.
3887 // Also, convert floating-point constants to bit patterns.
3888 // Non-constants are not relevant to this pass.
3889 //
3890 // In terms of the running example on InitializeNode::InitializeNode
3891 // and InitializeNode::capture_store, here is the transformation
3892 // of rawstore1 and rawstore2 into rawstore12:
3893 // alloc = (Allocate ...)
3894 // rawoop = alloc.RawAddress
3895 // tile12 = 0x00010002
3896 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3897 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3898 //
3899 void
3900 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3901 Node* size_in_bytes,
3902 PhaseGVN* phase) {
3903 Compile* C = phase->C;
3904
3905 assert(stores_are_sane(phase), "");
3906 // Note: After this pass, they are not completely sane,
3907 // since there may be some overlaps.
3908
3909 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3910
3911 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3912 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3913 size_limit = MIN2(size_limit, ti_limit);
3914 size_limit = align_up(size_limit, BytesPerLong);
3915 int num_tiles = size_limit / BytesPerLong;
3916
3917 // allocate space for the tile map:
3918 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3919 jlong tiles_buf[small_len];
3920 Node* nodes_buf[small_len];
3921 jlong inits_buf[small_len];
3922 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3923 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3924 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3925 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3926 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3927 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3928 // tiles: exact bitwise model of all primitive constants
3929 // nodes: last constant-storing node subsumed into the tiles model
3930 // inits: which bytes (in each tile) are touched by any initializations
3931
3932 //// Pass A: Fill in the tile model with any relevant stores.
3933
3934 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3935 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3936 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3937 Node* zmem = zero_memory(); // initially zero memory state
3938 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3939 Node* st = in(i);
3940 intptr_t st_off = get_store_offset(st, phase);
3941
3942 // Figure out the store's offset and constant value:
3943 if (st_off < header_size) continue; //skip (ignore header)
3944 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3945 int st_size = st->as_Store()->memory_size();
3946 if (st_off + st_size > size_limit) break;
3947
3948 // Record which bytes are touched, whether by constant or not.
3949 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3950 continue; // skip (strange store size)
3951
3952 const Type* val = phase->type(st->in(MemNode::ValueIn));
3953 if (!val->singleton()) continue; //skip (non-con store)
3954 BasicType type = val->basic_type();
3955
3956 jlong con = 0;
3957 switch (type) {
3958 case T_INT: con = val->is_int()->get_con(); break;
3959 case T_LONG: con = val->is_long()->get_con(); break;
3960 case T_FLOAT: con = jint_cast(val->getf()); break;
3961 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3962 default: continue; //skip (odd store type)
3963 }
3964
3965 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3966 st->Opcode() == Op_StoreL) {
3967 continue; // This StoreL is already optimal.
3968 }
3969
3970 // Store down the constant.
3971 store_constant(tiles, num_tiles, st_off, st_size, con);
3972
3973 intptr_t j = st_off >> LogBytesPerLong;
3974
3975 if (type == T_INT && st_size == BytesPerInt
3976 && (st_off & BytesPerInt) == BytesPerInt) {
3977 jlong lcon = tiles[j];
3978 if (!Matcher::isSimpleConstant64(lcon) &&
3979 st->Opcode() == Op_StoreI) {
3980 // This StoreI is already optimal by itself.
3981 jint* intcon = (jint*) &tiles[j];
3982 intcon[1] = 0; // undo the store_constant()
3983
3984 // If the previous store is also optimal by itself, back up and
3985 // undo the action of the previous loop iteration... if we can.
3986 // But if we can't, just let the previous half take care of itself.
3987 st = nodes[j];
3988 st_off -= BytesPerInt;
3989 con = intcon[0];
3990 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3991 assert(st_off >= header_size, "still ignoring header");
3992 assert(get_store_offset(st, phase) == st_off, "must be");
3993 assert(in(i-1) == zmem, "must be");
3994 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3995 assert(con == tcon->is_int()->get_con(), "must be");
3996 // Undo the effects of the previous loop trip, which swallowed st:
3997 intcon[0] = 0; // undo store_constant()
3998 set_req(i-1, st); // undo set_req(i, zmem)
3999 nodes[j] = NULL; // undo nodes[j] = st
4000 --old_subword; // undo ++old_subword
4001 }
4002 continue; // This StoreI is already optimal.
4003 }
4004 }
4005
4006 // This store is not needed.
4007 set_req(i, zmem);
4008 nodes[j] = st; // record for the moment
4009 if (st_size < BytesPerLong) // something has changed
4010 ++old_subword; // includes int/float, but who's counting...
4011 else ++old_long;
4012 }
4013
4014 if ((old_subword + old_long) == 0)
4015 return; // nothing more to do
4016
4017 //// Pass B: Convert any non-zero tiles into optimal constant stores.
4018 // Be sure to insert them before overlapping non-constant stores.
4019 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
4020 for (int j = 0; j < num_tiles; j++) {
4021 jlong con = tiles[j];
4022 jlong init = inits[j];
4023 if (con == 0) continue;
4024 jint con0, con1; // split the constant, address-wise
4025 jint init0, init1; // split the init map, address-wise
4026 { union { jlong con; jint intcon[2]; } u;
4027 u.con = con;
4028 con0 = u.intcon[0];
4029 con1 = u.intcon[1];
4030 u.con = init;
4031 init0 = u.intcon[0];
4032 init1 = u.intcon[1];
4033 }
4034
4035 Node* old = nodes[j];
4036 assert(old != NULL, "need the prior store");
4037 intptr_t offset = (j * BytesPerLong);
4038
4039 bool split = !Matcher::isSimpleConstant64(con);
4040
4041 if (offset < header_size) {
4042 assert(offset + BytesPerInt >= header_size, "second int counts");
4043 assert(*(jint*)&tiles[j] == 0, "junk in header");
4044 split = true; // only the second word counts
4045 // Example: int a[] = { 42 ... }
4046 } else if (con0 == 0 && init0 == -1) {
4047 split = true; // first word is covered by full inits
4048 // Example: int a[] = { ... foo(), 42 ... }
4049 } else if (con1 == 0 && init1 == -1) {
4050 split = true; // second word is covered by full inits
4051 // Example: int a[] = { ... 42, foo() ... }
4052 }
4053
4054 // Here's a case where init0 is neither 0 nor -1:
4055 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
4056 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
4057 // In this case the tile is not split; it is (jlong)42.
4058 // The big tile is stored down, and then the foo() value is inserted.
4059 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
4060
4061 Node* ctl = old->in(MemNode::Control);
4062 Node* adr = make_raw_address(offset, phase);
4063 const TypePtr* atp = TypeRawPtr::BOTTOM;
4064
4065 // One or two coalesced stores to plop down.
4066 Node* st[2];
4067 intptr_t off[2];
4068 int nst = 0;
4069 if (!split) {
4070 ++new_long;
4071 off[nst] = offset;
4072 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4073 phase->longcon(con), T_LONG, MemNode::unordered);
4074 } else {
4075 // Omit either if it is a zero.
4076 if (con0 != 0) {
4077 ++new_int;
4078 off[nst] = offset;
4079 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4080 phase->intcon(con0), T_INT, MemNode::unordered);
4081 }
4082 if (con1 != 0) {
4083 ++new_int;
4084 offset += BytesPerInt;
4085 adr = make_raw_address(offset, phase);
4086 off[nst] = offset;
4087 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4088 phase->intcon(con1), T_INT, MemNode::unordered);
4089 }
4090 }
4091
4092 // Insert second store first, then the first before the second.
4093 // Insert each one just before any overlapping non-constant stores.
4094 while (nst > 0) {
4095 Node* st1 = st[--nst];
4096 C->copy_node_notes_to(st1, old);
4097 st1 = phase->transform(st1);
4098 offset = off[nst];
4099 assert(offset >= header_size, "do not smash header");
4100 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4101 guarantee(ins_idx != 0, "must re-insert constant store");
4102 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
4103 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4104 set_req(--ins_idx, st1);
4105 else
4106 ins_req(ins_idx, st1);
4107 }
4108 }
4109
4110 if (PrintCompilation && WizardMode)
4111 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4112 old_subword, old_long, new_int, new_long);
4113 if (C->log() != NULL)
4114 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4115 old_subword, old_long, new_int, new_long);
4116
4117 // Clean up any remaining occurrences of zmem:
4118 remove_extra_zeroes();
4119 }
4120
4121 // Explore forward from in(start) to find the first fully initialized
4122 // word, and return its offset. Skip groups of subword stores which
4123 // together initialize full words. If in(start) is itself part of a
4124 // fully initialized word, return the offset of in(start). If there
4125 // are no following full-word stores, or if something is fishy, return
4126 // a negative value.
4127 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4128 int int_map = 0;
4129 intptr_t int_map_off = 0;
4130 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
4131
4132 for (uint i = start, limit = req(); i < limit; i++) {
4133 Node* st = in(i);
4134
4135 intptr_t st_off = get_store_offset(st, phase);
4136 if (st_off < 0) break; // return conservative answer
4137
4138 int st_size = st->as_Store()->memory_size();
4139 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4140 return st_off; // we found a complete word init
4141 }
4142
4143 // update the map:
4144
4145 intptr_t this_int_off = align_down(st_off, BytesPerInt);
4146 if (this_int_off != int_map_off) {
4147 // reset the map:
4148 int_map = 0;
4149 int_map_off = this_int_off;
4150 }
4151
4152 int subword_off = st_off - this_int_off;
4153 int_map |= right_n_bits(st_size) << subword_off;
4154 if ((int_map & FULL_MAP) == FULL_MAP) {
4155 return this_int_off; // we found a complete word init
4156 }
4157
4158 // Did this store hit or cross the word boundary?
4159 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4160 if (next_int_off == this_int_off + BytesPerInt) {
4161 // We passed the current int, without fully initializing it.
4162 int_map_off = next_int_off;
4163 int_map >>= BytesPerInt;
4164 } else if (next_int_off > this_int_off + BytesPerInt) {
4165 // We passed the current and next int.
4166 return this_int_off + BytesPerInt;
4167 }
4168 }
4169
4170 return -1;
4171 }
4172
4173
4174 // Called when the associated AllocateNode is expanded into CFG.
4175 // At this point, we may perform additional optimizations.
4176 // Linearize the stores by ascending offset, to make memory
4177 // activity as coherent as possible.
4178 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4179 intptr_t header_size,
4180 Node* size_in_bytes,
4181 PhaseIterGVN* phase) {
4182 assert(!is_complete(), "not already complete");
4183 assert(stores_are_sane(phase), "");
4184 assert(allocation() != NULL, "must be present");
4185
4186 remove_extra_zeroes();
4187
4188 if (ReduceFieldZeroing || ReduceBulkZeroing)
4189 // reduce instruction count for common initialization patterns
4190 coalesce_subword_stores(header_size, size_in_bytes, phase);
4191
4192 Node* zmem = zero_memory(); // initially zero memory state
4193 Node* inits = zmem; // accumulating a linearized chain of inits
4194 #ifdef ASSERT
4195 intptr_t first_offset = allocation()->minimum_header_size();
4196 intptr_t last_init_off = first_offset; // previous init offset
4197 intptr_t last_init_end = first_offset; // previous init offset+size
4198 intptr_t last_tile_end = first_offset; // previous tile offset+size
4199 #endif
4200 intptr_t zeroes_done = header_size;
4201
4202 bool do_zeroing = true; // we might give up if inits are very sparse
4203 int big_init_gaps = 0; // how many large gaps have we seen?
4204
4205 if (UseTLAB && ZeroTLAB) do_zeroing = false;
4206 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
4207
4208 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4209 Node* st = in(i);
4210 intptr_t st_off = get_store_offset(st, phase);
4211 if (st_off < 0)
4212 break; // unknown junk in the inits
4213 if (st->in(MemNode::Memory) != zmem)
4214 break; // complicated store chains somehow in list
4215
4216 int st_size = st->as_Store()->memory_size();
4217 intptr_t next_init_off = st_off + st_size;
4218
4219 if (do_zeroing && zeroes_done < next_init_off) {
4220 // See if this store needs a zero before it or under it.
4221 intptr_t zeroes_needed = st_off;
4222
4223 if (st_size < BytesPerInt) {
4224 // Look for subword stores which only partially initialize words.
4225 // If we find some, we must lay down some word-level zeroes first,
4226 // underneath the subword stores.
4227 //
4228 // Examples:
4229 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
4230 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
4231 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
4232 //
4233 // Note: coalesce_subword_stores may have already done this,
4234 // if it was prompted by constant non-zero subword initializers.
4235 // But this case can still arise with non-constant stores.
4236
4237 intptr_t next_full_store = find_next_fullword_store(i, phase);
4238
4239 // In the examples above:
4240 // in(i) p q r s x y z
4241 // st_off 12 13 14 15 12 13 14
4242 // st_size 1 1 1 1 1 1 1
4243 // next_full_s. 12 16 16 16 16 16 16
4244 // z's_done 12 16 16 16 12 16 12
4245 // z's_needed 12 16 16 16 16 16 16
4246 // zsize 0 0 0 0 4 0 4
4247 if (next_full_store < 0) {
4248 // Conservative tack: Zero to end of current word.
4249 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4250 } else {
4251 // Zero to beginning of next fully initialized word.
4252 // Or, don't zero at all, if we are already in that word.
4253 assert(next_full_store >= zeroes_needed, "must go forward");
4254 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4255 zeroes_needed = next_full_store;
4256 }
4257 }
4258
4259 if (zeroes_needed > zeroes_done) {
4260 intptr_t zsize = zeroes_needed - zeroes_done;
4261 // Do some incremental zeroing on rawmem, in parallel with inits.
4262 zeroes_done = align_down(zeroes_done, BytesPerInt);
4263 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4264 zeroes_done, zeroes_needed,
4265 phase);
4266 zeroes_done = zeroes_needed;
4267 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4268 do_zeroing = false; // leave the hole, next time
4269 }
4270 }
4271
4272 // Collect the store and move on:
4273 phase->replace_input_of(st, MemNode::Memory, inits);
4274 inits = st; // put it on the linearized chain
4275 set_req(i, zmem); // unhook from previous position
4276
4277 if (zeroes_done == st_off)
4278 zeroes_done = next_init_off;
4279
4280 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4281
4282 #ifdef ASSERT
4283 // Various order invariants. Weaker than stores_are_sane because
4284 // a large constant tile can be filled in by smaller non-constant stores.
4285 assert(st_off >= last_init_off, "inits do not reverse");
4286 last_init_off = st_off;
4287 const Type* val = NULL;
4288 if (st_size >= BytesPerInt &&
4289 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4290 (int)val->basic_type() < (int)T_OBJECT) {
4291 assert(st_off >= last_tile_end, "tiles do not overlap");
4292 assert(st_off >= last_init_end, "tiles do not overwrite inits");
4293 last_tile_end = MAX2(last_tile_end, next_init_off);
4294 } else {
4295 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4296 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4297 assert(st_off >= last_init_end, "inits do not overlap");
4298 last_init_end = next_init_off; // it's a non-tile
4299 }
4300 #endif //ASSERT
4301 }
4302
4303 remove_extra_zeroes(); // clear out all the zmems left over
4304 add_req(inits);
4305
4306 if (!(UseTLAB && ZeroTLAB)) {
4307 // If anything remains to be zeroed, zero it all now.
4308 zeroes_done = align_down(zeroes_done, BytesPerInt);
4309 // if it is the last unused 4 bytes of an instance, forget about it
4310 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4311 if (zeroes_done + BytesPerLong >= size_limit) {
4312 AllocateNode* alloc = allocation();
4313 assert(alloc != NULL, "must be present");
4314 if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
4315 Node* klass_node = alloc->in(AllocateNode::KlassNode);
4316 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4317 if (zeroes_done == k->layout_helper())
4318 zeroes_done = size_limit;
4319 }
4320 }
4321 if (zeroes_done < size_limit) {
4322 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4323 zeroes_done, size_in_bytes, phase);
4324 }
4325 }
4326
4327 set_complete(phase);
4328 return rawmem;
4329 }
4330
4331
4332 #ifdef ASSERT
4333 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4334 if (is_complete())
4335 return true; // stores could be anything at this point
4336 assert(allocation() != NULL, "must be present");
4337 intptr_t last_off = allocation()->minimum_header_size();
4338 for (uint i = InitializeNode::RawStores; i < req(); i++) {
4339 Node* st = in(i);
4340 intptr_t st_off = get_store_offset(st, phase);
4341 if (st_off < 0) continue; // ignore dead garbage
4342 if (last_off > st_off) {
4343 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4344 this->dump(2);
4345 assert(false, "ascending store offsets");
4346 return false;
4347 }
4348 last_off = st_off + st->as_Store()->memory_size();
4349 }
4350 return true;
4351 }
4352 #endif //ASSERT
4353
4354
4355
4356
4357 //============================MergeMemNode=====================================
4358 //
4359 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4360 // contributing store or call operations. Each contributor provides the memory
4361 // state for a particular "alias type" (see Compile::alias_type). For example,
4362 // if a MergeMem has an input X for alias category #6, then any memory reference
4363 // to alias category #6 may use X as its memory state input, as an exact equivalent
4364 // to using the MergeMem as a whole.
4365 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4366 //
4367 // (Here, the <N> notation gives the index of the relevant adr_type.)
4368 //
4369 // In one special case (and more cases in the future), alias categories overlap.
4370 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4371 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4372 // it is exactly equivalent to that state W:
4373 // MergeMem(<Bot>: W) <==> W
4374 //
4375 // Usually, the merge has more than one input. In that case, where inputs
4376 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4377 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4378 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4379 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4380 //
4381 // A merge can take a "wide" memory state as one of its narrow inputs.
4382 // This simply means that the merge observes out only the relevant parts of
4383 // the wide input. That is, wide memory states arriving at narrow merge inputs
4384 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4385 //
4386 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4387 // and that memory slices "leak through":
4388 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4389 //
4390 // But, in such a cascade, repeated memory slices can "block the leak":
4391 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4392 //
4393 // In the last example, Y is not part of the combined memory state of the
4394 // outermost MergeMem. The system must, of course, prevent unschedulable
4395 // memory states from arising, so you can be sure that the state Y is somehow
4396 // a precursor to state Y'.
4397 //
4398 //
4399 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4400 // of each MergeMemNode array are exactly the numerical alias indexes, including
4401 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4402 // Compile::alias_type (and kin) produce and manage these indexes.
4403 //
4404 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4405 // (Note that this provides quick access to the top node inside MergeMem methods,
4406 // without the need to reach out via TLS to Compile::current.)
4407 //
4408 // As a consequence of what was just described, a MergeMem that represents a full
4409 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4410 // containing all alias categories.
4411 //
4412 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4413 //
4414 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4415 // a memory state for the alias type <N>, or else the top node, meaning that
4416 // there is no particular input for that alias type. Note that the length of
4417 // a MergeMem is variable, and may be extended at any time to accommodate new
4418 // memory states at larger alias indexes. When merges grow, they are of course
4419 // filled with "top" in the unused in() positions.
4420 //
4421 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4422 // (Top was chosen because it works smoothly with passes like GCM.)
4423 //
4424 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4425 // the type of random VM bits like TLS references.) Since it is always the
4426 // first non-Bot memory slice, some low-level loops use it to initialize an
4427 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4428 //
4429 //
4430 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4431 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4432 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4433 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4434 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4435 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4436 //
4437 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4438 // really that different from the other memory inputs. An abbreviation called
4439 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4440 //
4441 //
4442 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4443 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4444 // that "emerges though" the base memory will be marked as excluding the alias types
4445 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4446 //
4447 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4448 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4449 //
4450 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4451 // (It is currently unimplemented.) As you can see, the resulting merge is
4452 // actually a disjoint union of memory states, rather than an overlay.
4453 //
4454
4455 //------------------------------MergeMemNode-----------------------------------
4456 Node* MergeMemNode::make_empty_memory() {
4457 Node* empty_memory = (Node*) Compile::current()->top();
4458 assert(empty_memory->is_top(), "correct sentinel identity");
4459 return empty_memory;
4460 }
4461
4462 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4463 init_class_id(Class_MergeMem);
4464 // all inputs are nullified in Node::Node(int)
4465 // set_input(0, NULL); // no control input
4466
4467 // Initialize the edges uniformly to top, for starters.
4468 Node* empty_mem = make_empty_memory();
4469 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4470 init_req(i,empty_mem);
4471 }
4472 assert(empty_memory() == empty_mem, "");
4473
4474 if( new_base != NULL && new_base->is_MergeMem() ) {
4475 MergeMemNode* mdef = new_base->as_MergeMem();
4476 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4477 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4478 mms.set_memory(mms.memory2());
4479 }
4480 assert(base_memory() == mdef->base_memory(), "");
4481 } else {
4482 set_base_memory(new_base);
4483 }
4484 }
4485
4486 // Make a new, untransformed MergeMem with the same base as 'mem'.
4487 // If mem is itself a MergeMem, populate the result with the same edges.
4488 MergeMemNode* MergeMemNode::make(Node* mem) {
4489 return new MergeMemNode(mem);
4490 }
4491
4492 //------------------------------cmp--------------------------------------------
4493 uint MergeMemNode::hash() const { return NO_HASH; }
4494 bool MergeMemNode::cmp( const Node &n ) const {
4495 return (&n == this); // Always fail except on self
4496 }
4497
4498 //------------------------------Identity---------------------------------------
4499 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4500 // Identity if this merge point does not record any interesting memory
4501 // disambiguations.
4502 Node* base_mem = base_memory();
4503 Node* empty_mem = empty_memory();
4504 if (base_mem != empty_mem) { // Memory path is not dead?
4505 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4506 Node* mem = in(i);
4507 if (mem != empty_mem && mem != base_mem) {
4508 return this; // Many memory splits; no change
4509 }
4510 }
4511 }
4512 return base_mem; // No memory splits; ID on the one true input
4513 }
4514
4515 //------------------------------Ideal------------------------------------------
4516 // This method is invoked recursively on chains of MergeMem nodes
4517 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4518 // Remove chain'd MergeMems
4519 //
4520 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4521 // relative to the "in(Bot)". Since we are patching both at the same time,
4522 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4523 // but rewrite each "in(i)" relative to the new "in(Bot)".
4524 Node *progress = NULL;
4525
4526
4527 Node* old_base = base_memory();
4528 Node* empty_mem = empty_memory();
4529 if (old_base == empty_mem)
4530 return NULL; // Dead memory path.
4531
4532 MergeMemNode* old_mbase;
4533 if (old_base != NULL && old_base->is_MergeMem())
4534 old_mbase = old_base->as_MergeMem();
4535 else
4536 old_mbase = NULL;
4537 Node* new_base = old_base;
4538
4539 // simplify stacked MergeMems in base memory
4540 if (old_mbase) new_base = old_mbase->base_memory();
4541
4542 // the base memory might contribute new slices beyond my req()
4543 if (old_mbase) grow_to_match(old_mbase);
4544
4545 // Look carefully at the base node if it is a phi.
4546 PhiNode* phi_base;
4547 if (new_base != NULL && new_base->is_Phi())
4548 phi_base = new_base->as_Phi();
4549 else
4550 phi_base = NULL;
4551
4552 Node* phi_reg = NULL;
4553 uint phi_len = (uint)-1;
4554 if (phi_base != NULL && !phi_base->is_copy()) {
4555 // do not examine phi if degraded to a copy
4556 phi_reg = phi_base->region();
4557 phi_len = phi_base->req();
4558 // see if the phi is unfinished
4559 for (uint i = 1; i < phi_len; i++) {
4560 if (phi_base->in(i) == NULL) {
4561 // incomplete phi; do not look at it yet!
4562 phi_reg = NULL;
4563 phi_len = (uint)-1;
4564 break;
4565 }
4566 }
4567 }
4568
4569 // Note: We do not call verify_sparse on entry, because inputs
4570 // can normalize to the base_memory via subsume_node or similar
4571 // mechanisms. This method repairs that damage.
4572
4573 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4574
4575 // Look at each slice.
4576 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4577 Node* old_in = in(i);
4578 // calculate the old memory value
4579 Node* old_mem = old_in;
4580 if (old_mem == empty_mem) old_mem = old_base;
4581 assert(old_mem == memory_at(i), "");
4582
4583 // maybe update (reslice) the old memory value
4584
4585 // simplify stacked MergeMems
4586 Node* new_mem = old_mem;
4587 MergeMemNode* old_mmem;
4588 if (old_mem != NULL && old_mem->is_MergeMem())
4589 old_mmem = old_mem->as_MergeMem();
4590 else
4591 old_mmem = NULL;
4592 if (old_mmem == this) {
4593 // This can happen if loops break up and safepoints disappear.
4594 // A merge of BotPtr (default) with a RawPtr memory derived from a
4595 // safepoint can be rewritten to a merge of the same BotPtr with
4596 // the BotPtr phi coming into the loop. If that phi disappears
4597 // also, we can end up with a self-loop of the mergemem.
4598 // In general, if loops degenerate and memory effects disappear,
4599 // a mergemem can be left looking at itself. This simply means
4600 // that the mergemem's default should be used, since there is
4601 // no longer any apparent effect on this slice.
4602 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4603 // from start. Update the input to TOP.
4604 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4605 }
4606 else if (old_mmem != NULL) {
4607 new_mem = old_mmem->memory_at(i);
4608 }
4609 // else preceding memory was not a MergeMem
4610
4611 // replace equivalent phis (unfortunately, they do not GVN together)
4612 if (new_mem != NULL && new_mem != new_base &&
4613 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4614 if (new_mem->is_Phi()) {
4615 PhiNode* phi_mem = new_mem->as_Phi();
4616 for (uint i = 1; i < phi_len; i++) {
4617 if (phi_base->in(i) != phi_mem->in(i)) {
4618 phi_mem = NULL;
4619 break;
4620 }
4621 }
4622 if (phi_mem != NULL) {
4623 // equivalent phi nodes; revert to the def
4624 new_mem = new_base;
4625 }
4626 }
4627 }
4628
4629 // maybe store down a new value
4630 Node* new_in = new_mem;
4631 if (new_in == new_base) new_in = empty_mem;
4632
4633 if (new_in != old_in) {
4634 // Warning: Do not combine this "if" with the previous "if"
4635 // A memory slice might have be be rewritten even if it is semantically
4636 // unchanged, if the base_memory value has changed.
4637 set_req(i, new_in);
4638 progress = this; // Report progress
4639 }
4640 }
4641
4642 if (new_base != old_base) {
4643 set_req(Compile::AliasIdxBot, new_base);
4644 // Don't use set_base_memory(new_base), because we need to update du.
4645 assert(base_memory() == new_base, "");
4646 progress = this;
4647 }
4648
4649 if( base_memory() == this ) {
4650 // a self cycle indicates this memory path is dead
4651 set_req(Compile::AliasIdxBot, empty_mem);
4652 }
4653
4654 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4655 // Recursion must occur after the self cycle check above
4656 if( base_memory()->is_MergeMem() ) {
4657 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4658 Node *m = phase->transform(new_mbase); // Rollup any cycles
4659 if( m != NULL &&
4660 (m->is_top() ||
4661 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4662 // propagate rollup of dead cycle to self
4663 set_req(Compile::AliasIdxBot, empty_mem);
4664 }
4665 }
4666
4667 if( base_memory() == empty_mem ) {
4668 progress = this;
4669 // Cut inputs during Parse phase only.
4670 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4671 if( !can_reshape ) {
4672 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4673 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4674 }
4675 }
4676 }
4677
4678 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4679 // Check if PhiNode::Ideal's "Split phis through memory merges"
4680 // transform should be attempted. Look for this->phi->this cycle.
4681 uint merge_width = req();
4682 if (merge_width > Compile::AliasIdxRaw) {
4683 PhiNode* phi = base_memory()->as_Phi();
4684 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4685 if (phi->in(i) == this) {
4686 phase->is_IterGVN()->_worklist.push(phi);
4687 break;
4688 }
4689 }
4690 }
4691 }
4692
4693 assert(progress || verify_sparse(), "please, no dups of base");
4694 return progress;
4695 }
4696
4697 //-------------------------set_base_memory-------------------------------------
4698 void MergeMemNode::set_base_memory(Node *new_base) {
4699 Node* empty_mem = empty_memory();
4700 set_req(Compile::AliasIdxBot, new_base);
4701 assert(memory_at(req()) == new_base, "must set default memory");
4702 // Clear out other occurrences of new_base:
4703 if (new_base != empty_mem) {
4704 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4705 if (in(i) == new_base) set_req(i, empty_mem);
4706 }
4707 }
4708 }
4709
4710 //------------------------------out_RegMask------------------------------------
4711 const RegMask &MergeMemNode::out_RegMask() const {
4712 return RegMask::Empty;
4713 }
4714
4715 //------------------------------dump_spec--------------------------------------
4716 #ifndef PRODUCT
4717 void MergeMemNode::dump_spec(outputStream *st) const {
4718 st->print(" {");
4719 Node* base_mem = base_memory();
4720 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4721 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4722 if (mem == base_mem) { st->print(" -"); continue; }
4723 st->print( " N%d:", mem->_idx );
4724 Compile::current()->get_adr_type(i)->dump_on(st);
4725 }
4726 st->print(" }");
4727 }
4728 #endif // !PRODUCT
4729
4730
4731 #ifdef ASSERT
4732 static bool might_be_same(Node* a, Node* b) {
4733 if (a == b) return true;
4734 if (!(a->is_Phi() || b->is_Phi())) return false;
4735 // phis shift around during optimization
4736 return true; // pretty stupid...
4737 }
4738
4739 // verify a narrow slice (either incoming or outgoing)
4740 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4741 if (!VerifyAliases) return; // don't bother to verify unless requested
4742 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
4743 if (Node::in_dump()) return; // muzzle asserts when printing
4744 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4745 assert(n != NULL, "");
4746 // Elide intervening MergeMem's
4747 while (n->is_MergeMem()) {
4748 n = n->as_MergeMem()->memory_at(alias_idx);
4749 }
4750 Compile* C = Compile::current();
4751 const TypePtr* n_adr_type = n->adr_type();
4752 if (n == m->empty_memory()) {
4753 // Implicit copy of base_memory()
4754 } else if (n_adr_type != TypePtr::BOTTOM) {
4755 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4756 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4757 } else {
4758 // A few places like make_runtime_call "know" that VM calls are narrow,
4759 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4760 bool expected_wide_mem = false;
4761 if (n == m->base_memory()) {
4762 expected_wide_mem = true;
4763 } else if (alias_idx == Compile::AliasIdxRaw ||
4764 n == m->memory_at(Compile::AliasIdxRaw)) {
4765 expected_wide_mem = true;
4766 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4767 // memory can "leak through" calls on channels that
4768 // are write-once. Allow this also.
4769 expected_wide_mem = true;
4770 }
4771 assert(expected_wide_mem, "expected narrow slice replacement");
4772 }
4773 }
4774 #else // !ASSERT
4775 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4776 #endif
4777
4778
4779 //-----------------------------memory_at---------------------------------------
4780 Node* MergeMemNode::memory_at(uint alias_idx) const {
4781 assert(alias_idx >= Compile::AliasIdxRaw ||
4782 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4783 "must avoid base_memory and AliasIdxTop");
4784
4785 // Otherwise, it is a narrow slice.
4786 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4787 Compile *C = Compile::current();
4788 if (is_empty_memory(n)) {
4789 // the array is sparse; empty slots are the "top" node
4790 n = base_memory();
4791 assert(Node::in_dump()
4792 || n == NULL || n->bottom_type() == Type::TOP
4793 || n->adr_type() == NULL // address is TOP
4794 || n->adr_type() == TypePtr::BOTTOM
4795 || n->adr_type() == TypeRawPtr::BOTTOM
4796 || Compile::current()->AliasLevel() == 0,
4797 "must be a wide memory");
4798 // AliasLevel == 0 if we are organizing the memory states manually.
4799 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4800 } else {
4801 // make sure the stored slice is sane
4802 #ifdef ASSERT
4803 if (VMError::is_error_reported() || Node::in_dump()) {
4804 } else if (might_be_same(n, base_memory())) {
4805 // Give it a pass: It is a mostly harmless repetition of the base.
4806 // This can arise normally from node subsumption during optimization.
4807 } else {
4808 verify_memory_slice(this, alias_idx, n);
4809 }
4810 #endif
4811 }
4812 return n;
4813 }
4814
4815 //---------------------------set_memory_at-------------------------------------
4816 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4817 verify_memory_slice(this, alias_idx, n);
4818 Node* empty_mem = empty_memory();
4819 if (n == base_memory()) n = empty_mem; // collapse default
4820 uint need_req = alias_idx+1;
4821 if (req() < need_req) {
4822 if (n == empty_mem) return; // already the default, so do not grow me
4823 // grow the sparse array
4824 do {
4825 add_req(empty_mem);
4826 } while (req() < need_req);
4827 }
4828 set_req( alias_idx, n );
4829 }
4830
4831
4832
4833 //--------------------------iteration_setup------------------------------------
4834 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4835 if (other != NULL) {
4836 grow_to_match(other);
4837 // invariant: the finite support of mm2 is within mm->req()
4838 #ifdef ASSERT
4839 for (uint i = req(); i < other->req(); i++) {
4840 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4841 }
4842 #endif
4843 }
4844 // Replace spurious copies of base_memory by top.
4845 Node* base_mem = base_memory();
4846 if (base_mem != NULL && !base_mem->is_top()) {
4847 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4848 if (in(i) == base_mem)
4849 set_req(i, empty_memory());
4850 }
4851 }
4852 }
4853
4854 //---------------------------grow_to_match-------------------------------------
4855 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4856 Node* empty_mem = empty_memory();
4857 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4858 // look for the finite support of the other memory
4859 for (uint i = other->req(); --i >= req(); ) {
4860 if (other->in(i) != empty_mem) {
4861 uint new_len = i+1;
4862 while (req() < new_len) add_req(empty_mem);
4863 break;
4864 }
4865 }
4866 }
4867
4868 //---------------------------verify_sparse-------------------------------------
4869 #ifndef PRODUCT
4870 bool MergeMemNode::verify_sparse() const {
4871 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4872 Node* base_mem = base_memory();
4873 // The following can happen in degenerate cases, since empty==top.
4874 if (is_empty_memory(base_mem)) return true;
4875 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4876 assert(in(i) != NULL, "sane slice");
4877 if (in(i) == base_mem) return false; // should have been the sentinel value!
4878 }
4879 return true;
4880 }
4881
4882 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4883 Node* n;
4884 n = mm->in(idx);
4885 if (mem == n) return true; // might be empty_memory()
4886 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4887 if (mem == n) return true;
4888 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4889 if (mem == n) return true;
4890 if (n == NULL) break;
4891 }
4892 return false;
4893 }
4894 #endif // !PRODUCT