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