1 /*
2 * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "memory/allocation.inline.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/connode.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/loopnode.hpp"
40 #include "opto/machnode.hpp"
41 #include "opto/matcher.hpp"
42 #include "opto/memnode.hpp"
43 #include "opto/mulnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/phaseX.hpp"
46 #include "opto/regmask.hpp"
47 #include "opto/rootnode.hpp"
48 #include "utilities/align.hpp"
49 #include "utilities/copy.hpp"
50 #include "utilities/macros.hpp"
51 #include "utilities/vmError.hpp"
52 #if INCLUDE_ZGC
53 #include "gc/z/c2/zBarrierSetC2.hpp"
54 #endif
55 #if INCLUDE_SHENANDOAHGC
56 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
57 #endif
58
59 // Portions of code courtesy of Clifford Click
60
61 // Optimization - Graph Style
62
63 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
64
65 //=============================================================================
66 uint MemNode::size_of() const { return sizeof(*this); }
67
68 const TypePtr *MemNode::adr_type() const {
69 Node* adr = in(Address);
70 if (adr == NULL) return NULL; // node is dead
71 const TypePtr* cross_check = NULL;
72 DEBUG_ONLY(cross_check = _adr_type);
73 return calculate_adr_type(adr->bottom_type(), cross_check);
74 }
75
76 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
77 if (adr != NULL) {
78 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
79 return true;
80 }
81 }
82 return false;
83 }
84
85 #ifndef PRODUCT
86 void MemNode::dump_spec(outputStream *st) const {
87 if (in(Address) == NULL) return; // node is dead
88 #ifndef ASSERT
89 // fake the missing field
90 const TypePtr* _adr_type = NULL;
91 if (in(Address) != NULL)
92 _adr_type = in(Address)->bottom_type()->isa_ptr();
93 #endif
94 dump_adr_type(this, _adr_type, st);
95
96 Compile* C = Compile::current();
97 if (C->alias_type(_adr_type)->is_volatile()) {
98 st->print(" Volatile!");
99 }
100 if (_unaligned_access) {
101 st->print(" unaligned");
102 }
103 if (_mismatched_access) {
104 st->print(" mismatched");
105 }
106 if (_unsafe_access) {
107 st->print(" unsafe");
108 }
109 }
110
111 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
112 st->print(" @");
113 if (adr_type == NULL) {
114 st->print("NULL");
115 } else {
116 adr_type->dump_on(st);
117 Compile* C = Compile::current();
118 Compile::AliasType* atp = NULL;
119 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
120 if (atp == NULL)
121 st->print(", idx=?\?;");
122 else if (atp->index() == Compile::AliasIdxBot)
123 st->print(", idx=Bot;");
124 else if (atp->index() == Compile::AliasIdxTop)
125 st->print(", idx=Top;");
126 else if (atp->index() == Compile::AliasIdxRaw)
127 st->print(", idx=Raw;");
128 else {
129 ciField* field = atp->field();
130 if (field) {
131 st->print(", name=");
132 field->print_name_on(st);
133 }
134 st->print(", idx=%d;", atp->index());
135 }
136 }
137 }
138
139 extern void print_alias_types();
140
141 #endif
142
143 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
144 assert((t_oop != NULL), "sanity");
145 bool is_instance = t_oop->is_known_instance_field();
146 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
147 (load != NULL) && load->is_Load() &&
148 (phase->is_IterGVN() != NULL);
149 if (!(is_instance || is_boxed_value_load))
150 return mchain; // don't try to optimize non-instance types
151 uint instance_id = t_oop->instance_id();
152 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
153 Node *prev = NULL;
154 Node *result = mchain;
155 while (prev != result) {
156 prev = result;
157 if (result == start_mem)
158 break; // hit one of our sentinels
159 // skip over a call which does not affect this memory slice
160 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
161 Node *proj_in = result->in(0);
162 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
163 break; // hit one of our sentinels
164 } else if (proj_in->is_Call()) {
165 // ArrayCopyNodes processed here as well
166 CallNode *call = proj_in->as_Call();
167 if (!call->may_modify(t_oop, phase)) { // returns false for instances
168 result = call->in(TypeFunc::Memory);
169 }
170 } else if (proj_in->is_Initialize()) {
171 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
172 // Stop if this is the initialization for the object instance which
173 // contains this memory slice, otherwise skip over it.
174 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
175 break;
176 }
177 if (is_instance) {
178 result = proj_in->in(TypeFunc::Memory);
179 } else if (is_boxed_value_load) {
180 Node* klass = alloc->in(AllocateNode::KlassNode);
181 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
182 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
183 result = proj_in->in(TypeFunc::Memory); // not related allocation
184 }
185 }
186 } else if (proj_in->is_MemBar()) {
187 ArrayCopyNode* ac = NULL;
188 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
189 break;
190 }
191 result = proj_in->in(TypeFunc::Memory);
192 } else {
193 assert(false, "unexpected projection");
194 }
195 } else if (result->is_ClearArray()) {
196 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
197 // Can not bypass initialization of the instance
198 // we are looking for.
199 break;
200 }
201 // Otherwise skip it (the call updated 'result' value).
202 } else if (result->is_MergeMem()) {
203 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
204 }
205 }
206 return result;
207 }
208
209 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
210 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
211 if (t_oop == NULL)
212 return mchain; // don't try to optimize non-oop types
213 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
214 bool is_instance = t_oop->is_known_instance_field();
215 PhaseIterGVN *igvn = phase->is_IterGVN();
216 if (is_instance && igvn != NULL && result->is_Phi()) {
217 PhiNode *mphi = result->as_Phi();
218 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
219 const TypePtr *t = mphi->adr_type();
220 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
221 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
222 t->is_oopptr()->cast_to_exactness(true)
223 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
224 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) {
225 // clone the Phi with our address type
226 result = mphi->split_out_instance(t_adr, igvn);
227 } else {
228 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
229 }
230 }
231 return result;
232 }
233
234 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
235 uint alias_idx = phase->C->get_alias_index(tp);
236 Node *mem = mmem;
237 #ifdef ASSERT
238 {
239 // Check that current type is consistent with the alias index used during graph construction
240 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
241 bool consistent = adr_check == NULL || adr_check->empty() ||
242 phase->C->must_alias(adr_check, alias_idx );
243 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
244 if( !consistent && adr_check != NULL && !adr_check->empty() &&
245 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
246 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
247 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
248 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
249 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
250 // don't assert if it is dead code.
251 consistent = true;
252 }
253 if( !consistent ) {
254 st->print("alias_idx==%d, adr_check==", alias_idx);
255 if( adr_check == NULL ) {
256 st->print("NULL");
257 } else {
258 adr_check->dump();
259 }
260 st->cr();
261 print_alias_types();
262 assert(consistent, "adr_check must match alias idx");
263 }
264 }
265 #endif
266 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
267 // means an array I have not precisely typed yet. Do not do any
268 // alias stuff with it any time soon.
269 const TypeOopPtr *toop = tp->isa_oopptr();
270 if( tp->base() != Type::AnyPtr &&
271 !(toop &&
272 toop->klass() != NULL &&
273 toop->klass()->is_java_lang_Object() &&
274 toop->offset() == Type::OffsetBot) ) {
275 // compress paths and change unreachable cycles to TOP
276 // If not, we can update the input infinitely along a MergeMem cycle
277 // Equivalent code in PhiNode::Ideal
278 Node* m = phase->transform(mmem);
279 // If transformed to a MergeMem, get the desired slice
280 // Otherwise the returned node represents memory for every slice
281 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
282 // Update input if it is progress over what we have now
283 }
284 return mem;
285 }
286
287 //--------------------------Ideal_common---------------------------------------
288 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
289 // Unhook non-raw memories from complete (macro-expanded) initializations.
290 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
291 // If our control input is a dead region, kill all below the region
292 Node *ctl = in(MemNode::Control);
293 if (ctl && remove_dead_region(phase, can_reshape))
294 return this;
295 ctl = in(MemNode::Control);
296 // Don't bother trying to transform a dead node
297 if (ctl && ctl->is_top()) return NodeSentinel;
298
299 PhaseIterGVN *igvn = phase->is_IterGVN();
300 // Wait if control on the worklist.
301 if (ctl && can_reshape && igvn != NULL) {
302 Node* bol = NULL;
303 Node* cmp = NULL;
304 if (ctl->in(0)->is_If()) {
305 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
306 bol = ctl->in(0)->in(1);
307 if (bol->is_Bool())
308 cmp = ctl->in(0)->in(1)->in(1);
309 }
310 if (igvn->_worklist.member(ctl) ||
311 (bol != NULL && igvn->_worklist.member(bol)) ||
312 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
313 // This control path may be dead.
314 // Delay this memory node transformation until the control is processed.
315 phase->is_IterGVN()->_worklist.push(this);
316 return NodeSentinel; // caller will return NULL
317 }
318 }
319 // Ignore if memory is dead, or self-loop
320 Node *mem = in(MemNode::Memory);
321 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
322 assert(mem != this, "dead loop in MemNode::Ideal");
323
324 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
325 // This memory slice may be dead.
326 // Delay this mem node transformation until the memory is processed.
327 phase->is_IterGVN()->_worklist.push(this);
328 return NodeSentinel; // caller will return NULL
329 }
330
331 Node *address = in(MemNode::Address);
332 const Type *t_adr = phase->type(address);
333 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
334
335 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
336 // Unsafe off-heap access with zero address. Remove access and other control users
337 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
338 assert(ctl != NULL, "unsafe accesses should be control dependent");
339 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
340 Node* u = ctl->fast_out(i);
341 if (u != ctl) {
342 igvn->rehash_node_delayed(u);
343 int nb = u->replace_edge(ctl, phase->C->top());
344 --i, imax -= nb;
345 }
346 }
347 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
348 Node* halt = igvn->transform(new HaltNode(ctl, frame));
349 phase->C->root()->add_req(halt);
350 return this;
351 }
352
353 if (can_reshape && igvn != NULL &&
354 (igvn->_worklist.member(address) ||
355 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
356 // The address's base and type may change when the address is processed.
357 // Delay this mem node transformation until the address is processed.
358 phase->is_IterGVN()->_worklist.push(this);
359 return NodeSentinel; // caller will return NULL
360 }
361
362 // Do NOT remove or optimize the next lines: ensure a new alias index
363 // is allocated for an oop pointer type before Escape Analysis.
364 // Note: C++ will not remove it since the call has side effect.
365 if (t_adr->isa_oopptr()) {
366 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
367 }
368
369 Node* base = NULL;
370 if (address->is_AddP()) {
371 base = address->in(AddPNode::Base);
372 }
373 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
374 !t_adr->isa_rawptr()) {
375 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
376 // Skip this node optimization if its address has TOP base.
377 return NodeSentinel; // caller will return NULL
378 }
379
380 // Avoid independent memory operations
381 Node* old_mem = mem;
382
383 // The code which unhooks non-raw memories from complete (macro-expanded)
384 // initializations was removed. After macro-expansion all stores catched
385 // by Initialize node became raw stores and there is no information
386 // which memory slices they modify. So it is unsafe to move any memory
387 // operation above these stores. Also in most cases hooked non-raw memories
388 // were already unhooked by using information from detect_ptr_independence()
389 // and find_previous_store().
390
391 if (mem->is_MergeMem()) {
392 MergeMemNode* mmem = mem->as_MergeMem();
393 const TypePtr *tp = t_adr->is_ptr();
394
395 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
396 }
397
398 if (mem != old_mem) {
399 set_req(MemNode::Memory, mem);
400 if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) {
401 igvn->_worklist.push(old_mem);
402 }
403 if (phase->type(mem) == Type::TOP) return NodeSentinel;
404 return this;
405 }
406
407 // let the subclass continue analyzing...
408 return NULL;
409 }
410
411 // Helper function for proving some simple control dominations.
412 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
413 // Already assumes that 'dom' is available at 'sub', and that 'sub'
414 // is not a constant (dominated by the method's StartNode).
415 // Used by MemNode::find_previous_store to prove that the
416 // control input of a memory operation predates (dominates)
417 // an allocation it wants to look past.
418 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
419 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
420 return false; // Conservative answer for dead code
421
422 // Check 'dom'. Skip Proj and CatchProj nodes.
423 dom = dom->find_exact_control(dom);
424 if (dom == NULL || dom->is_top())
425 return false; // Conservative answer for dead code
426
427 if (dom == sub) {
428 // For the case when, for example, 'sub' is Initialize and the original
429 // 'dom' is Proj node of the 'sub'.
430 return false;
431 }
432
433 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
434 return true;
435
436 // 'dom' dominates 'sub' if its control edge and control edges
437 // of all its inputs dominate or equal to sub's control edge.
438
439 // Currently 'sub' is either Allocate, Initialize or Start nodes.
440 // Or Region for the check in LoadNode::Ideal();
441 // 'sub' should have sub->in(0) != NULL.
442 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
443 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
444
445 // Get control edge of 'sub'.
446 Node* orig_sub = sub;
447 sub = sub->find_exact_control(sub->in(0));
448 if (sub == NULL || sub->is_top())
449 return false; // Conservative answer for dead code
450
451 assert(sub->is_CFG(), "expecting control");
452
453 if (sub == dom)
454 return true;
455
456 if (sub->is_Start() || sub->is_Root())
457 return false;
458
459 {
460 // Check all control edges of 'dom'.
461
462 ResourceMark rm;
463 Arena* arena = Thread::current()->resource_area();
464 Node_List nlist(arena);
465 Unique_Node_List dom_list(arena);
466
467 dom_list.push(dom);
468 bool only_dominating_controls = false;
469
470 for (uint next = 0; next < dom_list.size(); next++) {
471 Node* n = dom_list.at(next);
472 if (n == orig_sub)
473 return false; // One of dom's inputs dominated by sub.
474 if (!n->is_CFG() && n->pinned()) {
475 // Check only own control edge for pinned non-control nodes.
476 n = n->find_exact_control(n->in(0));
477 if (n == NULL || n->is_top())
478 return false; // Conservative answer for dead code
479 assert(n->is_CFG(), "expecting control");
480 dom_list.push(n);
481 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
482 only_dominating_controls = true;
483 } else if (n->is_CFG()) {
484 if (n->dominates(sub, nlist))
485 only_dominating_controls = true;
486 else
487 return false;
488 } else {
489 // First, own control edge.
490 Node* m = n->find_exact_control(n->in(0));
491 if (m != NULL) {
492 if (m->is_top())
493 return false; // Conservative answer for dead code
494 dom_list.push(m);
495 }
496 // Now, the rest of edges.
497 uint cnt = n->req();
498 for (uint i = 1; i < cnt; i++) {
499 m = n->find_exact_control(n->in(i));
500 if (m == NULL || m->is_top())
501 continue;
502 dom_list.push(m);
503 }
504 }
505 }
506 return only_dominating_controls;
507 }
508 }
509
510 //---------------------detect_ptr_independence---------------------------------
511 // Used by MemNode::find_previous_store to prove that two base
512 // pointers are never equal.
513 // The pointers are accompanied by their associated allocations,
514 // if any, which have been previously discovered by the caller.
515 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
516 Node* p2, AllocateNode* a2,
517 PhaseTransform* phase) {
518 // Attempt to prove that these two pointers cannot be aliased.
519 // They may both manifestly be allocations, and they should differ.
520 // Or, if they are not both allocations, they can be distinct constants.
521 // Otherwise, one is an allocation and the other a pre-existing value.
522 if (a1 == NULL && a2 == NULL) { // neither an allocation
523 return (p1 != p2) && p1->is_Con() && p2->is_Con();
524 } else if (a1 != NULL && a2 != NULL) { // both allocations
525 return (a1 != a2);
526 } else if (a1 != NULL) { // one allocation a1
527 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
528 return all_controls_dominate(p2, a1);
529 } else { //(a2 != NULL) // one allocation a2
530 return all_controls_dominate(p1, a2);
531 }
532 return false;
533 }
534
535
536 // Find an arraycopy that must have set (can_see_stored_value=true) or
537 // could have set (can_see_stored_value=false) the value for this load
538 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
539 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
540 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
541 Node* mb = mem->in(0);
542 if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
543 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
544 ArrayCopyNode* ac = mb->in(0)->in(0)->as_ArrayCopy();
545 if (ac->is_clonebasic()) {
546 intptr_t offset;
547 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase, offset);
548 if (alloc != NULL && alloc == ld_alloc) {
549 return ac;
550 }
551 }
552 }
553 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
554 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
555
556 if (ac->is_arraycopy_validated() ||
557 ac->is_copyof_validated() ||
558 ac->is_copyofrange_validated()) {
559 Node* ld_addp = in(MemNode::Address);
560 if (ld_addp->is_AddP()) {
561 Node* ld_base = ld_addp->in(AddPNode::Address);
562 Node* ld_offs = ld_addp->in(AddPNode::Offset);
563
564 Node* dest = ac->in(ArrayCopyNode::Dest);
565
566 if (dest == ld_base) {
567 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
568 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
569 return ac;
570 }
571 if (!can_see_stored_value) {
572 mem = ac->in(TypeFunc::Memory);
573 }
574 }
575 }
576 }
577 }
578 return NULL;
579 }
580
581 // The logic for reordering loads and stores uses four steps:
582 // (a) Walk carefully past stores and initializations which we
583 // can prove are independent of this load.
584 // (b) Observe that the next memory state makes an exact match
585 // with self (load or store), and locate the relevant store.
586 // (c) Ensure that, if we were to wire self directly to the store,
587 // the optimizer would fold it up somehow.
588 // (d) Do the rewiring, and return, depending on some other part of
589 // the optimizer to fold up the load.
590 // This routine handles steps (a) and (b). Steps (c) and (d) are
591 // specific to loads and stores, so they are handled by the callers.
592 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
593 //
594 Node* MemNode::find_previous_store(PhaseTransform* phase) {
595 Node* ctrl = in(MemNode::Control);
596 Node* adr = in(MemNode::Address);
597 intptr_t offset = 0;
598 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
599 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
600
601 if (offset == Type::OffsetBot)
602 return NULL; // cannot unalias unless there are precise offsets
603
604 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
605 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
606
607 intptr_t size_in_bytes = memory_size();
608
609 Node* mem = in(MemNode::Memory); // start searching here...
610
611 int cnt = 50; // Cycle limiter
612 for (;;) { // While we can dance past unrelated stores...
613 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
614
615 Node* prev = mem;
616 if (mem->is_Store()) {
617 Node* st_adr = mem->in(MemNode::Address);
618 intptr_t st_offset = 0;
619 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
620 if (st_base == NULL)
621 break; // inscrutable pointer
622
623 // For raw accesses it's not enough to prove that constant offsets don't intersect.
624 // We need the bases to be the equal in order for the offset check to make sense.
625 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
626 break;
627 }
628
629 if (st_offset != offset && st_offset != Type::OffsetBot) {
630 const int MAX_STORE = BytesPerLong;
631 if (st_offset >= offset + size_in_bytes ||
632 st_offset <= offset - MAX_STORE ||
633 st_offset <= offset - mem->as_Store()->memory_size()) {
634 // Success: The offsets are provably independent.
635 // (You may ask, why not just test st_offset != offset and be done?
636 // The answer is that stores of different sizes can co-exist
637 // in the same sequence of RawMem effects. We sometimes initialize
638 // a whole 'tile' of array elements with a single jint or jlong.)
639 mem = mem->in(MemNode::Memory);
640 continue; // (a) advance through independent store memory
641 }
642 }
643 if (st_base != base &&
644 detect_ptr_independence(base, alloc,
645 st_base,
646 AllocateNode::Ideal_allocation(st_base, phase),
647 phase)) {
648 // Success: The bases are provably independent.
649 mem = mem->in(MemNode::Memory);
650 continue; // (a) advance through independent store memory
651 }
652
653 // (b) At this point, if the bases or offsets do not agree, we lose,
654 // since we have not managed to prove 'this' and 'mem' independent.
655 if (st_base == base && st_offset == offset) {
656 return mem; // let caller handle steps (c), (d)
657 }
658
659 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
660 InitializeNode* st_init = mem->in(0)->as_Initialize();
661 AllocateNode* st_alloc = st_init->allocation();
662 if (st_alloc == NULL)
663 break; // something degenerated
664 bool known_identical = false;
665 bool known_independent = false;
666 if (alloc == st_alloc)
667 known_identical = true;
668 else if (alloc != NULL)
669 known_independent = true;
670 else if (all_controls_dominate(this, st_alloc))
671 known_independent = true;
672
673 if (known_independent) {
674 // The bases are provably independent: Either they are
675 // manifestly distinct allocations, or else the control
676 // of this load dominates the store's allocation.
677 int alias_idx = phase->C->get_alias_index(adr_type());
678 if (alias_idx == Compile::AliasIdxRaw) {
679 mem = st_alloc->in(TypeFunc::Memory);
680 } else {
681 mem = st_init->memory(alias_idx);
682 }
683 continue; // (a) advance through independent store memory
684 }
685
686 // (b) at this point, if we are not looking at a store initializing
687 // the same allocation we are loading from, we lose.
688 if (known_identical) {
689 // From caller, can_see_stored_value will consult find_captured_store.
690 return mem; // let caller handle steps (c), (d)
691 }
692
693 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
694 if (prev != mem) {
695 // Found an arraycopy but it doesn't affect that load
696 continue;
697 }
698 // Found an arraycopy that may affect that load
699 return mem;
700 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
701 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
702 if (mem->is_Proj() && mem->in(0)->is_Call()) {
703 // ArrayCopyNodes processed here as well.
704 CallNode *call = mem->in(0)->as_Call();
705 if (!call->may_modify(addr_t, phase)) {
706 mem = call->in(TypeFunc::Memory);
707 continue; // (a) advance through independent call memory
708 }
709 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
710 ArrayCopyNode* ac = NULL;
711 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
712 break;
713 }
714 mem = mem->in(0)->in(TypeFunc::Memory);
715 continue; // (a) advance through independent MemBar memory
716 } else if (mem->is_ClearArray()) {
717 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
718 // (the call updated 'mem' value)
719 continue; // (a) advance through independent allocation memory
720 } else {
721 // Can not bypass initialization of the instance
722 // we are looking for.
723 return mem;
724 }
725 } else if (mem->is_MergeMem()) {
726 int alias_idx = phase->C->get_alias_index(adr_type());
727 mem = mem->as_MergeMem()->memory_at(alias_idx);
728 continue; // (a) advance through independent MergeMem memory
729 }
730 }
731
732 // Unless there is an explicit 'continue', we must bail out here,
733 // because 'mem' is an inscrutable memory state (e.g., a call).
734 break;
735 }
736
737 return NULL; // bail out
738 }
739
740 //----------------------calculate_adr_type-------------------------------------
741 // Helper function. Notices when the given type of address hits top or bottom.
742 // Also, asserts a cross-check of the type against the expected address type.
743 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
744 if (t == Type::TOP) return NULL; // does not touch memory any more?
745 #ifdef PRODUCT
746 cross_check = NULL;
747 #else
748 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL;
749 #endif
750 const TypePtr* tp = t->isa_ptr();
751 if (tp == NULL) {
752 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
753 return TypePtr::BOTTOM; // touches lots of memory
754 } else {
755 #ifdef ASSERT
756 // %%%% [phh] We don't check the alias index if cross_check is
757 // TypeRawPtr::BOTTOM. Needs to be investigated.
758 if (cross_check != NULL &&
759 cross_check != TypePtr::BOTTOM &&
760 cross_check != TypeRawPtr::BOTTOM) {
761 // Recheck the alias index, to see if it has changed (due to a bug).
762 Compile* C = Compile::current();
763 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
764 "must stay in the original alias category");
765 // The type of the address must be contained in the adr_type,
766 // disregarding "null"-ness.
767 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
768 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
769 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
770 "real address must not escape from expected memory type");
771 }
772 #endif
773 return tp;
774 }
775 }
776
777 //=============================================================================
778 // Should LoadNode::Ideal() attempt to remove control edges?
779 bool LoadNode::can_remove_control() const {
780 return true;
781 }
782 uint LoadNode::size_of() const { return sizeof(*this); }
783 uint LoadNode::cmp( const Node &n ) const
784 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
785 const Type *LoadNode::bottom_type() const { return _type; }
786 uint LoadNode::ideal_reg() const {
787 return _type->ideal_reg();
788 }
789
790 #ifndef PRODUCT
791 void LoadNode::dump_spec(outputStream *st) const {
792 MemNode::dump_spec(st);
793 if( !Verbose && !WizardMode ) {
794 // standard dump does this in Verbose and WizardMode
795 st->print(" #"); _type->dump_on(st);
796 }
797 if (!depends_only_on_test()) {
798 st->print(" (does not depend only on test)");
799 }
800 }
801 #endif
802
803 #ifdef ASSERT
804 //----------------------------is_immutable_value-------------------------------
805 // Helper function to allow a raw load without control edge for some cases
806 bool LoadNode::is_immutable_value(Node* adr) {
807 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
808 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
809 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
810 in_bytes(JavaThread::osthread_offset())));
811 }
812 #endif
813
814 //----------------------------LoadNode::make-----------------------------------
815 // Polymorphic factory method:
816 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo,
817 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) {
818 Compile* C = gvn.C;
819
820 // sanity check the alias category against the created node type
821 assert(!(adr_type->isa_oopptr() &&
822 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
823 "use LoadKlassNode instead");
824 assert(!(adr_type->isa_aryptr() &&
825 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
826 "use LoadRangeNode instead");
827 // Check control edge of raw loads
828 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
829 // oop will be recorded in oop map if load crosses safepoint
830 rt->isa_oopptr() || is_immutable_value(adr),
831 "raw memory operations should have control edge");
832 LoadNode* load = NULL;
833 switch (bt) {
834 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
835 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
836 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
837 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
838 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
839 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break;
840 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
841 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
842 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
843 case T_OBJECT:
844 #ifdef _LP64
845 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
846 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
847 } else
848 #endif
849 {
850 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
851 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
852 }
853 break;
854 default:
855 ShouldNotReachHere();
856 break;
857 }
858 assert(load != NULL, "LoadNode should have been created");
859 if (unaligned) {
860 load->set_unaligned_access();
861 }
862 if (mismatched) {
863 load->set_mismatched_access();
864 }
865 if (unsafe) {
866 load->set_unsafe_access();
867 }
868 if (load->Opcode() == Op_LoadN) {
869 Node* ld = gvn.transform(load);
870 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
871 }
872
873 return load;
874 }
875
876 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
877 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) {
878 bool require_atomic = true;
879 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic);
880 if (unaligned) {
881 load->set_unaligned_access();
882 }
883 if (mismatched) {
884 load->set_mismatched_access();
885 }
886 if (unsafe) {
887 load->set_unsafe_access();
888 }
889 return load;
890 }
891
892 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
893 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe) {
894 bool require_atomic = true;
895 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic);
896 if (unaligned) {
897 load->set_unaligned_access();
898 }
899 if (mismatched) {
900 load->set_mismatched_access();
901 }
902 if (unsafe) {
903 load->set_unsafe_access();
904 }
905 return load;
906 }
907
908
909
910 //------------------------------hash-------------------------------------------
911 uint LoadNode::hash() const {
912 // unroll addition of interesting fields
913 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
914 }
915
916 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
917 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
918 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
919 bool is_stable_ary = FoldStableValues &&
920 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
921 tp->isa_aryptr()->is_stable();
922
923 return (eliminate_boxing && non_volatile) || is_stable_ary;
924 }
925
926 return false;
927 }
928
929 // Is the value loaded previously stored by an arraycopy? If so return
930 // a load node that reads from the source array so we may be able to
931 // optimize out the ArrayCopy node later.
932 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
933 #if INCLUDE_ZGC
934 if (UseZGC) {
935 if (bottom_type()->make_oopptr() != NULL) {
936 return NULL;
937 }
938 }
939 #endif
940
941 Node* ld_adr = in(MemNode::Address);
942 intptr_t ld_off = 0;
943 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
944 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
945 if (ac != NULL) {
946 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
947
948 Node* mem = ac->in(TypeFunc::Memory);
949 Node* ctl = ac->in(0);
950 Node* src = ac->in(ArrayCopyNode::Src);
951
952 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
953 return NULL;
954 }
955
956 LoadNode* ld = clone()->as_Load();
957 Node* addp = in(MemNode::Address)->clone();
958 if (ac->as_ArrayCopy()->is_clonebasic()) {
959 assert(ld_alloc != NULL, "need an alloc");
960 assert(addp->is_AddP(), "address must be addp");
961 assert(addp->in(AddPNode::Base) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Base), "strange pattern");
962 assert(addp->in(AddPNode::Address) == ac->in(ArrayCopyNode::Dest)->in(AddPNode::Address), "strange pattern");
963 addp->set_req(AddPNode::Base, src->in(AddPNode::Base));
964 addp->set_req(AddPNode::Address, src->in(AddPNode::Address));
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 = Pinned;
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 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
1014 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1015 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
1016 // This is more general than load from boxing objects.
1017 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1018 uint alias_idx = atp->index();
1019 bool final = !atp->is_rewritable();
1020 Node* result = NULL;
1021 Node* current = st;
1022 // Skip through chains of MemBarNodes checking the MergeMems for
1023 // new states for the slice of this load. Stop once any other
1024 // kind of node is encountered. Loads from final memory can skip
1025 // through any kind of MemBar but normal loads shouldn't skip
1026 // through MemBarAcquire since the could allow them to move out of
1027 // a synchronized region.
1028 while (current->is_Proj()) {
1029 int opc = current->in(0)->Opcode();
1030 if ((final && (opc == Op_MemBarAcquire ||
1031 opc == Op_MemBarAcquireLock ||
1032 opc == Op_LoadFence)) ||
1033 opc == Op_MemBarRelease ||
1034 opc == Op_StoreFence ||
1035 opc == Op_MemBarReleaseLock ||
1036 opc == Op_MemBarStoreStore ||
1037 opc == Op_MemBarCPUOrder) {
1038 Node* mem = current->in(0)->in(TypeFunc::Memory);
1039 if (mem->is_MergeMem()) {
1040 MergeMemNode* merge = mem->as_MergeMem();
1041 Node* new_st = merge->memory_at(alias_idx);
1042 if (new_st == merge->base_memory()) {
1043 // Keep searching
1044 current = new_st;
1045 continue;
1046 }
1047 // Save the new memory state for the slice and fall through
1048 // to exit.
1049 result = new_st;
1050 }
1051 }
1052 break;
1053 }
1054 if (result != NULL) {
1055 st = result;
1056 }
1057 }
1058
1059 // Loop around twice in the case Load -> Initialize -> Store.
1060 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1061 for (int trip = 0; trip <= 1; trip++) {
1062
1063 if (st->is_Store()) {
1064 Node* st_adr = st->in(MemNode::Address);
1065 if (!phase->eqv(st_adr, ld_adr)) {
1066 // Try harder before giving up... Match raw and non-raw pointers.
1067 intptr_t st_off = 0;
1068 AllocateNode* alloc = AllocateNode::Ideal_allocation(st_adr, phase, st_off);
1069 if (alloc == NULL) return NULL;
1070 if (alloc != ld_alloc) return NULL;
1071 if (ld_off != st_off) return NULL;
1072 // At this point we have proven something like this setup:
1073 // A = Allocate(...)
1074 // L = LoadQ(, AddP(CastPP(, A.Parm),, #Off))
1075 // S = StoreQ(, AddP(, A.Parm , #Off), V)
1076 // (Actually, we haven't yet proven the Q's are the same.)
1077 // In other words, we are loading from a casted version of
1078 // the same pointer-and-offset that we stored to.
1079 // Thus, we are able to replace L by V.
1080 }
1081 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1082 if (store_Opcode() != st->Opcode())
1083 return NULL;
1084 return st->in(MemNode::ValueIn);
1085 }
1086
1087 // A load from a freshly-created object always returns zero.
1088 // (This can happen after LoadNode::Ideal resets the load's memory input
1089 // to find_captured_store, which returned InitializeNode::zero_memory.)
1090 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1091 (st->in(0) == ld_alloc) &&
1092 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1093 // return a zero value for the load's basic type
1094 // (This is one of the few places where a generic PhaseTransform
1095 // can create new nodes. Think of it as lazily manifesting
1096 // virtually pre-existing constants.)
1097 return phase->zerocon(memory_type());
1098 }
1099
1100 // A load from an initialization barrier can match a captured store.
1101 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1102 InitializeNode* init = st->in(0)->as_Initialize();
1103 AllocateNode* alloc = init->allocation();
1104 if ((alloc != NULL) && (alloc == ld_alloc)) {
1105 // examine a captured store value
1106 st = init->find_captured_store(ld_off, memory_size(), phase);
1107 if (st != NULL) {
1108 continue; // take one more trip around
1109 }
1110 }
1111 }
1112
1113 // Load boxed value from result of valueOf() call is input parameter.
1114 if (this->is_Load() && ld_adr->is_AddP() &&
1115 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1116 intptr_t ignore = 0;
1117 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1118 #if INCLUDE_SHENANDOAHGC
1119 if (UseShenandoahGC) {
1120 base = ((ShenandoahBarrierSetC2*) BarrierSet::barrier_set()->barrier_set_c2())->step_over_gc_barrier(base);
1121 }
1122 #endif
1123 if (base != NULL && base->is_Proj() &&
1124 base->as_Proj()->_con == TypeFunc::Parms &&
1125 base->in(0)->is_CallStaticJava() &&
1126 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1127 return base->in(0)->in(TypeFunc::Parms);
1128 }
1129 }
1130
1131 break;
1132 }
1133
1134 return NULL;
1135 }
1136
1137 //----------------------is_instance_field_load_with_local_phi------------------
1138 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1139 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1140 in(Address)->is_AddP() ) {
1141 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1142 // Only instances and boxed values.
1143 if( t_oop != NULL &&
1144 (t_oop->is_ptr_to_boxed_value() ||
1145 t_oop->is_known_instance_field()) &&
1146 t_oop->offset() != Type::OffsetBot &&
1147 t_oop->offset() != Type::OffsetTop) {
1148 return true;
1149 }
1150 }
1151 return false;
1152 }
1153
1154 //------------------------------Identity---------------------------------------
1155 // Loads are identity if previous store is to same address
1156 Node* LoadNode::Identity(PhaseGVN* phase) {
1157 // If the previous store-maker is the right kind of Store, and the store is
1158 // to the same address, then we are equal to the value stored.
1159 Node* mem = in(Memory);
1160 Node* value = can_see_stored_value(mem, phase);
1161 if( value ) {
1162 // byte, short & char stores truncate naturally.
1163 // A load has to load the truncated value which requires
1164 // some sort of masking operation and that requires an
1165 // Ideal call instead of an Identity call.
1166 if (memory_size() < BytesPerInt) {
1167 // If the input to the store does not fit with the load's result type,
1168 // it must be truncated via an Ideal call.
1169 if (!phase->type(value)->higher_equal(phase->type(this)))
1170 return this;
1171 }
1172 // (This works even when value is a Con, but LoadNode::Value
1173 // usually runs first, producing the singleton type of the Con.)
1174 return value;
1175 }
1176
1177 // Search for an existing data phi which was generated before for the same
1178 // instance's field to avoid infinite generation of phis in a loop.
1179 Node *region = mem->in(0);
1180 if (is_instance_field_load_with_local_phi(region)) {
1181 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1182 int this_index = phase->C->get_alias_index(addr_t);
1183 int this_offset = addr_t->offset();
1184 int this_iid = addr_t->instance_id();
1185 if (!addr_t->is_known_instance() &&
1186 addr_t->is_ptr_to_boxed_value()) {
1187 // Use _idx of address base (could be Phi node) for boxed values.
1188 intptr_t ignore = 0;
1189 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1190 if (base == NULL) {
1191 return this;
1192 }
1193 this_iid = base->_idx;
1194 }
1195 const Type* this_type = bottom_type();
1196 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1197 Node* phi = region->fast_out(i);
1198 if (phi->is_Phi() && phi != mem &&
1199 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1200 return phi;
1201 }
1202 }
1203 }
1204
1205 return this;
1206 }
1207
1208 // Construct an equivalent unsigned load.
1209 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1210 BasicType bt = T_ILLEGAL;
1211 const Type* rt = NULL;
1212 switch (Opcode()) {
1213 case Op_LoadUB: return this;
1214 case Op_LoadUS: return this;
1215 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1216 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1217 default:
1218 assert(false, "no unsigned variant: %s", Name());
1219 return NULL;
1220 }
1221 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1222 raw_adr_type(), rt, bt, _mo, _control_dependency,
1223 is_unaligned_access(), is_mismatched_access());
1224 }
1225
1226 // Construct an equivalent signed load.
1227 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1228 BasicType bt = T_ILLEGAL;
1229 const Type* rt = NULL;
1230 switch (Opcode()) {
1231 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1232 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1233 case Op_LoadB: // fall through
1234 case Op_LoadS: // fall through
1235 case Op_LoadI: // fall through
1236 case Op_LoadL: return this;
1237 default:
1238 assert(false, "no signed variant: %s", Name());
1239 return NULL;
1240 }
1241 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1242 raw_adr_type(), rt, bt, _mo, _control_dependency,
1243 is_unaligned_access(), is_mismatched_access());
1244 }
1245
1246 // We're loading from an object which has autobox behaviour.
1247 // If this object is result of a valueOf call we'll have a phi
1248 // merging a newly allocated object and a load from the cache.
1249 // We want to replace this load with the original incoming
1250 // argument to the valueOf call.
1251 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1252 assert(phase->C->eliminate_boxing(), "sanity");
1253 intptr_t ignore = 0;
1254 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1255 if ((base == NULL) || base->is_Phi()) {
1256 // Push the loads from the phi that comes from valueOf up
1257 // through it to allow elimination of the loads and the recovery
1258 // of the original value. It is done in split_through_phi().
1259 return NULL;
1260 } else if (base->is_Load() ||
1261 (base->is_DecodeN() && base->in(1)->is_Load())) {
1262 // Eliminate the load of boxed value for integer types from the cache
1263 // array by deriving the value from the index into the array.
1264 // Capture the offset of the load and then reverse the computation.
1265
1266 // Get LoadN node which loads a boxing object from 'cache' array.
1267 if (base->is_DecodeN()) {
1268 base = base->in(1);
1269 }
1270 if (!base->in(Address)->is_AddP()) {
1271 return NULL; // Complex address
1272 }
1273 AddPNode* address = base->in(Address)->as_AddP();
1274 Node* cache_base = address->in(AddPNode::Base);
1275 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1276 // Get ConP node which is static 'cache' field.
1277 cache_base = cache_base->in(1);
1278 }
1279 if ((cache_base != NULL) && cache_base->is_Con()) {
1280 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1281 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1282 Node* elements[4];
1283 int shift = exact_log2(type2aelembytes(T_OBJECT));
1284 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1285 if (count > 0 && elements[0]->is_Con() &&
1286 (count == 1 ||
1287 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1288 elements[1]->in(2) == phase->intcon(shift)))) {
1289 ciObjArray* array = base_type->const_oop()->as_obj_array();
1290 // Fetch the box object cache[0] at the base of the array and get its value
1291 ciInstance* box = array->obj_at(0)->as_instance();
1292 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1293 assert(ik->is_box_klass(), "sanity");
1294 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1295 if (ik->nof_nonstatic_fields() == 1) {
1296 // This should be true nonstatic_field_at requires calling
1297 // nof_nonstatic_fields so check it anyway
1298 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1299 BasicType bt = c.basic_type();
1300 // Only integer types have boxing cache.
1301 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1302 bt == T_BYTE || bt == T_SHORT ||
1303 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1304 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1305 if (cache_low != (int)cache_low) {
1306 return NULL; // should not happen since cache is array indexed by value
1307 }
1308 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1309 if (offset != (int)offset) {
1310 return NULL; // should not happen since cache is array indexed by value
1311 }
1312 // Add up all the offsets making of the address of the load
1313 Node* result = elements[0];
1314 for (int i = 1; i < count; i++) {
1315 result = phase->transform(new AddXNode(result, elements[i]));
1316 }
1317 // Remove the constant offset from the address and then
1318 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset)));
1319 // remove the scaling of the offset to recover the original index.
1320 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1321 // Peel the shift off directly but wrap it in a dummy node
1322 // since Ideal can't return existing nodes
1323 result = new RShiftXNode(result->in(1), phase->intcon(0));
1324 } else if (result->is_Add() && result->in(2)->is_Con() &&
1325 result->in(1)->Opcode() == Op_LShiftX &&
1326 result->in(1)->in(2) == phase->intcon(shift)) {
1327 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1328 // but for boxing cache access we know that X<<Z will not overflow
1329 // (there is range check) so we do this optimizatrion by hand here.
1330 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift));
1331 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con));
1332 } else {
1333 result = new RShiftXNode(result, phase->intcon(shift));
1334 }
1335 #ifdef _LP64
1336 if (bt != T_LONG) {
1337 result = new ConvL2INode(phase->transform(result));
1338 }
1339 #else
1340 if (bt == T_LONG) {
1341 result = new ConvI2LNode(phase->transform(result));
1342 }
1343 #endif
1344 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1345 // Need to preserve unboxing load type if it is unsigned.
1346 switch(this->Opcode()) {
1347 case Op_LoadUB:
1348 result = new AndINode(phase->transform(result), phase->intcon(0xFF));
1349 break;
1350 case Op_LoadUS:
1351 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF));
1352 break;
1353 }
1354 return result;
1355 }
1356 }
1357 }
1358 }
1359 }
1360 return NULL;
1361 }
1362
1363 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1364 Node* region = phi->in(0);
1365 if (region == NULL) {
1366 return false; // Wait stable graph
1367 }
1368 uint cnt = phi->req();
1369 for (uint i = 1; i < cnt; i++) {
1370 Node* rc = region->in(i);
1371 if (rc == NULL || phase->type(rc) == Type::TOP)
1372 return false; // Wait stable graph
1373 Node* in = phi->in(i);
1374 if (in == NULL || phase->type(in) == Type::TOP)
1375 return false; // Wait stable graph
1376 }
1377 return true;
1378 }
1379 //------------------------------split_through_phi------------------------------
1380 // Split instance or boxed field load through Phi.
1381 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1382 Node* mem = in(Memory);
1383 Node* address = in(Address);
1384 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1385
1386 assert((t_oop != NULL) &&
1387 (t_oop->is_known_instance_field() ||
1388 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1389
1390 Compile* C = phase->C;
1391 intptr_t ignore = 0;
1392 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1393 bool base_is_phi = (base != NULL) && base->is_Phi();
1394 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1395 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1396 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1397
1398 if (!((mem->is_Phi() || base_is_phi) &&
1399 (load_boxed_values || t_oop->is_known_instance_field()))) {
1400 return NULL; // memory is not Phi
1401 }
1402
1403 if (mem->is_Phi()) {
1404 if (!stable_phi(mem->as_Phi(), phase)) {
1405 return NULL; // Wait stable graph
1406 }
1407 uint cnt = mem->req();
1408 // Check for loop invariant memory.
1409 if (cnt == 3) {
1410 for (uint i = 1; i < cnt; i++) {
1411 Node* in = mem->in(i);
1412 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1413 if (m == mem) {
1414 if (i == 1) {
1415 // if the first edge was a loop, check second edge too.
1416 // If both are replaceable - we are in an infinite loop
1417 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1418 if (n == mem) {
1419 break;
1420 }
1421 }
1422 set_req(Memory, mem->in(cnt - i));
1423 return this; // made change
1424 }
1425 }
1426 }
1427 }
1428 if (base_is_phi) {
1429 if (!stable_phi(base->as_Phi(), phase)) {
1430 return NULL; // Wait stable graph
1431 }
1432 uint cnt = base->req();
1433 // Check for loop invariant memory.
1434 if (cnt == 3) {
1435 for (uint i = 1; i < cnt; i++) {
1436 if (base->in(i) == base) {
1437 return NULL; // Wait stable graph
1438 }
1439 }
1440 }
1441 }
1442
1443 bool load_boxed_phi = load_boxed_values && base_is_phi && (base->in(0) == mem->in(0));
1444
1445 // Split through Phi (see original code in loopopts.cpp).
1446 assert(C->have_alias_type(t_oop), "instance should have alias type");
1447
1448 // Do nothing here if Identity will find a value
1449 // (to avoid infinite chain of value phis generation).
1450 if (!phase->eqv(this, phase->apply_identity(this)))
1451 return NULL;
1452
1453 // Select Region to split through.
1454 Node* region;
1455 if (!base_is_phi) {
1456 assert(mem->is_Phi(), "sanity");
1457 region = mem->in(0);
1458 // Skip if the region dominates some control edge of the address.
1459 if (!MemNode::all_controls_dominate(address, region))
1460 return NULL;
1461 } else if (!mem->is_Phi()) {
1462 assert(base_is_phi, "sanity");
1463 region = base->in(0);
1464 // Skip if the region dominates some control edge of the memory.
1465 if (!MemNode::all_controls_dominate(mem, region))
1466 return NULL;
1467 } else if (base->in(0) != mem->in(0)) {
1468 assert(base_is_phi && mem->is_Phi(), "sanity");
1469 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1470 region = base->in(0);
1471 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1472 region = mem->in(0);
1473 } else {
1474 return NULL; // complex graph
1475 }
1476 } else {
1477 assert(base->in(0) == mem->in(0), "sanity");
1478 region = mem->in(0);
1479 }
1480
1481 const Type* this_type = this->bottom_type();
1482 int this_index = C->get_alias_index(t_oop);
1483 int this_offset = t_oop->offset();
1484 int this_iid = t_oop->instance_id();
1485 if (!t_oop->is_known_instance() && load_boxed_values) {
1486 // Use _idx of address base for boxed values.
1487 this_iid = base->_idx;
1488 }
1489 PhaseIterGVN* igvn = phase->is_IterGVN();
1490 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1491 for (uint i = 1; i < region->req(); i++) {
1492 Node* x;
1493 Node* the_clone = NULL;
1494 if (region->in(i) == C->top()) {
1495 x = C->top(); // Dead path? Use a dead data op
1496 } else {
1497 x = this->clone(); // Else clone up the data op
1498 the_clone = x; // Remember for possible deletion.
1499 // Alter data node to use pre-phi inputs
1500 if (this->in(0) == region) {
1501 x->set_req(0, region->in(i));
1502 } else {
1503 x->set_req(0, NULL);
1504 }
1505 if (mem->is_Phi() && (mem->in(0) == region)) {
1506 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1507 }
1508 if (address->is_Phi() && address->in(0) == region) {
1509 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1510 }
1511 if (base_is_phi && (base->in(0) == region)) {
1512 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1513 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1514 x->set_req(Address, adr_x);
1515 }
1516 }
1517 // Check for a 'win' on some paths
1518 const Type *t = x->Value(igvn);
1519
1520 bool singleton = t->singleton();
1521
1522 // See comments in PhaseIdealLoop::split_thru_phi().
1523 if (singleton && t == Type::TOP) {
1524 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1525 }
1526
1527 if (singleton) {
1528 x = igvn->makecon(t);
1529 } else {
1530 // We now call Identity to try to simplify the cloned node.
1531 // Note that some Identity methods call phase->type(this).
1532 // Make sure that the type array is big enough for
1533 // our new node, even though we may throw the node away.
1534 // (This tweaking with igvn only works because x is a new node.)
1535 igvn->set_type(x, t);
1536 // If x is a TypeNode, capture any more-precise type permanently into Node
1537 // otherwise it will be not updated during igvn->transform since
1538 // igvn->type(x) is set to x->Value() already.
1539 x->raise_bottom_type(t);
1540 Node *y = igvn->apply_identity(x);
1541 if (y != x) {
1542 x = y;
1543 } else {
1544 y = igvn->hash_find_insert(x);
1545 if (y) {
1546 x = y;
1547 } else {
1548 // Else x is a new node we are keeping
1549 // We do not need register_new_node_with_optimizer
1550 // because set_type has already been called.
1551 igvn->_worklist.push(x);
1552 }
1553 }
1554 }
1555 if (x != the_clone && the_clone != NULL) {
1556 igvn->remove_dead_node(the_clone);
1557 }
1558 phi->set_req(i, x);
1559 }
1560 // Record Phi
1561 igvn->register_new_node_with_optimizer(phi);
1562 return phi;
1563 }
1564
1565 //------------------------------Ideal------------------------------------------
1566 // If the load is from Field memory and the pointer is non-null, it might be possible to
1567 // zero out the control input.
1568 // If the offset is constant and the base is an object allocation,
1569 // try to hook me up to the exact initializing store.
1570 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1571 Node* p = MemNode::Ideal_common(phase, can_reshape);
1572 if (p) return (p == NodeSentinel) ? NULL : p;
1573
1574 Node* ctrl = in(MemNode::Control);
1575 Node* address = in(MemNode::Address);
1576 bool progress = false;
1577
1578 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1579 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1580
1581 // Skip up past a SafePoint control. Cannot do this for Stores because
1582 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1583 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1584 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1585 !addr_mark ) {
1586 ctrl = ctrl->in(0);
1587 set_req(MemNode::Control,ctrl);
1588 progress = true;
1589 }
1590
1591 intptr_t ignore = 0;
1592 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1593 if (base != NULL
1594 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1595 // Check for useless control edge in some common special cases
1596 if (in(MemNode::Control) != NULL
1597 && can_remove_control()
1598 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1599 && all_controls_dominate(base, phase->C->start())) {
1600 // A method-invariant, non-null address (constant or 'this' argument).
1601 set_req(MemNode::Control, NULL);
1602 progress = true;
1603 }
1604 }
1605
1606 Node* mem = in(MemNode::Memory);
1607 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1608
1609 if (can_reshape && (addr_t != NULL)) {
1610 // try to optimize our memory input
1611 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1612 if (opt_mem != mem) {
1613 set_req(MemNode::Memory, opt_mem);
1614 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1615 return this;
1616 }
1617 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1618 if ((t_oop != NULL) &&
1619 (t_oop->is_known_instance_field() ||
1620 t_oop->is_ptr_to_boxed_value())) {
1621 PhaseIterGVN *igvn = phase->is_IterGVN();
1622 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1623 // Delay this transformation until memory Phi is processed.
1624 phase->is_IterGVN()->_worklist.push(this);
1625 return NULL;
1626 }
1627 // Split instance field load through Phi.
1628 Node* result = split_through_phi(phase);
1629 if (result != NULL) return result;
1630
1631 if (t_oop->is_ptr_to_boxed_value()) {
1632 Node* result = eliminate_autobox(phase);
1633 if (result != NULL) return result;
1634 }
1635 }
1636 }
1637
1638 // Is there a dominating load that loads the same value? Leave
1639 // anything that is not a load of a field/array element (like
1640 // barriers etc.) alone
1641 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) {
1642 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1643 Node *use = mem->fast_out(i);
1644 if (use != this &&
1645 use->Opcode() == Opcode() &&
1646 use->in(0) != NULL &&
1647 use->in(0) != in(0) &&
1648 use->in(Address) == in(Address)) {
1649 Node* ctl = in(0);
1650 for (int i = 0; i < 10 && ctl != NULL; i++) {
1651 ctl = IfNode::up_one_dom(ctl);
1652 if (ctl == use->in(0)) {
1653 set_req(0, use->in(0));
1654 return this;
1655 }
1656 }
1657 }
1658 }
1659 }
1660
1661 // Check for prior store with a different base or offset; make Load
1662 // independent. Skip through any number of them. Bail out if the stores
1663 // are in an endless dead cycle and report no progress. This is a key
1664 // transform for Reflection. However, if after skipping through the Stores
1665 // we can't then fold up against a prior store do NOT do the transform as
1666 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1667 // array memory alive twice: once for the hoisted Load and again after the
1668 // bypassed Store. This situation only works if EVERYBODY who does
1669 // anti-dependence work knows how to bypass. I.e. we need all
1670 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1671 // the alias index stuff. So instead, peek through Stores and IFF we can
1672 // fold up, do so.
1673 Node* prev_mem = find_previous_store(phase);
1674 if (prev_mem != NULL) {
1675 Node* value = can_see_arraycopy_value(prev_mem, phase);
1676 if (value != NULL) {
1677 return value;
1678 }
1679 }
1680 // Steps (a), (b): Walk past independent stores to find an exact match.
1681 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1682 // (c) See if we can fold up on the spot, but don't fold up here.
1683 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1684 // just return a prior value, which is done by Identity calls.
1685 if (can_see_stored_value(prev_mem, phase)) {
1686 // Make ready for step (d):
1687 set_req(MemNode::Memory, prev_mem);
1688 return this;
1689 }
1690 }
1691
1692 return progress ? this : NULL;
1693 }
1694
1695 // Helper to recognize certain Klass fields which are invariant across
1696 // some group of array types (e.g., int[] or all T[] where T < Object).
1697 const Type*
1698 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1699 ciKlass* klass) const {
1700 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1701 // The field is Klass::_modifier_flags. Return its (constant) value.
1702 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1703 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1704 return TypeInt::make(klass->modifier_flags());
1705 }
1706 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1707 // The field is Klass::_access_flags. Return its (constant) value.
1708 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1709 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1710 return TypeInt::make(klass->access_flags());
1711 }
1712 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1713 // The field is Klass::_layout_helper. Return its constant value if known.
1714 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1715 return TypeInt::make(klass->layout_helper());
1716 }
1717
1718 // No match.
1719 return NULL;
1720 }
1721
1722 //------------------------------Value-----------------------------------------
1723 const Type* LoadNode::Value(PhaseGVN* phase) const {
1724 // Either input is TOP ==> the result is TOP
1725 Node* mem = in(MemNode::Memory);
1726 const Type *t1 = phase->type(mem);
1727 if (t1 == Type::TOP) return Type::TOP;
1728 Node* adr = in(MemNode::Address);
1729 const TypePtr* tp = phase->type(adr)->isa_ptr();
1730 if (tp == NULL || tp->empty()) return Type::TOP;
1731 int off = tp->offset();
1732 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1733 Compile* C = phase->C;
1734
1735 // Try to guess loaded type from pointer type
1736 if (tp->isa_aryptr()) {
1737 const TypeAryPtr* ary = tp->is_aryptr();
1738 const Type* t = ary->elem();
1739
1740 // Determine whether the reference is beyond the header or not, by comparing
1741 // the offset against the offset of the start of the array's data.
1742 // Different array types begin at slightly different offsets (12 vs. 16).
1743 // We choose T_BYTE as an example base type that is least restrictive
1744 // as to alignment, which will therefore produce the smallest
1745 // possible base offset.
1746 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1747 const bool off_beyond_header = ((uint)off >= (uint)min_base_off);
1748
1749 // Try to constant-fold a stable array element.
1750 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1751 // Make sure the reference is not into the header and the offset is constant
1752 ciObject* aobj = ary->const_oop();
1753 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1754 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1755 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1756 stable_dimension,
1757 memory_type(), is_unsigned());
1758 if (con_type != NULL) {
1759 return con_type;
1760 }
1761 }
1762 }
1763
1764 // Don't do this for integer types. There is only potential profit if
1765 // the element type t is lower than _type; that is, for int types, if _type is
1766 // more restrictive than t. This only happens here if one is short and the other
1767 // char (both 16 bits), and in those cases we've made an intentional decision
1768 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1769 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1770 //
1771 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1772 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1773 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1774 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1775 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1776 // In fact, that could have been the original type of p1, and p1 could have
1777 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1778 // expression (LShiftL quux 3) independently optimized to the constant 8.
1779 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1780 && (_type->isa_vect() == NULL)
1781 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1782 // t might actually be lower than _type, if _type is a unique
1783 // concrete subclass of abstract class t.
1784 if (off_beyond_header) { // is the offset beyond the header?
1785 const Type* jt = t->join_speculative(_type);
1786 // In any case, do not allow the join, per se, to empty out the type.
1787 if (jt->empty() && !t->empty()) {
1788 // This can happen if a interface-typed array narrows to a class type.
1789 jt = _type;
1790 }
1791 #ifdef ASSERT
1792 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1793 // The pointers in the autobox arrays are always non-null
1794 Node* base = adr->in(AddPNode::Base);
1795 if ((base != NULL) && base->is_DecodeN()) {
1796 // Get LoadN node which loads IntegerCache.cache field
1797 base = base->in(1);
1798 }
1799 if ((base != NULL) && base->is_Con()) {
1800 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1801 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1802 // It could be narrow oop
1803 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1804 }
1805 }
1806 }
1807 #endif
1808 return jt;
1809 }
1810 }
1811 } else if (tp->base() == Type::InstPtr) {
1812 assert( off != Type::OffsetBot ||
1813 // arrays can be cast to Objects
1814 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1815 // unsafe field access may not have a constant offset
1816 C->has_unsafe_access(),
1817 "Field accesses must be precise" );
1818 // For oop loads, we expect the _type to be precise.
1819
1820 // Optimize loads from constant fields.
1821 const TypeInstPtr* tinst = tp->is_instptr();
1822 ciObject* const_oop = tinst->const_oop();
1823 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1824 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
1825 if (con_type != NULL) {
1826 return con_type;
1827 }
1828 }
1829 } else if (tp->base() == Type::KlassPtr) {
1830 assert( off != Type::OffsetBot ||
1831 // arrays can be cast to Objects
1832 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1833 // also allow array-loading from the primary supertype
1834 // array during subtype checks
1835 Opcode() == Op_LoadKlass,
1836 "Field accesses must be precise" );
1837 // For klass/static loads, we expect the _type to be precise
1838 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
1839 /* With mirrors being an indirect in the Klass*
1840 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
1841 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
1842 *
1843 * So check the type and klass of the node before the LoadP.
1844 */
1845 Node* adr2 = adr->in(MemNode::Address);
1846 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1847 if (tkls != NULL && !StressReflectiveCode) {
1848 ciKlass* klass = tkls->klass();
1849 if (klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1850 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1851 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1852 return TypeInstPtr::make(klass->java_mirror());
1853 }
1854 }
1855 }
1856
1857 const TypeKlassPtr *tkls = tp->isa_klassptr();
1858 if (tkls != NULL && !StressReflectiveCode) {
1859 ciKlass* klass = tkls->klass();
1860 if (klass->is_loaded() && tkls->klass_is_exact()) {
1861 // We are loading a field from a Klass metaobject whose identity
1862 // is known at compile time (the type is "exact" or "precise").
1863 // Check for fields we know are maintained as constants by the VM.
1864 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1865 // The field is Klass::_super_check_offset. Return its (constant) value.
1866 // (Folds up type checking code.)
1867 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1868 return TypeInt::make(klass->super_check_offset());
1869 }
1870 // Compute index into primary_supers array
1871 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1872 // Check for overflowing; use unsigned compare to handle the negative case.
1873 if( depth < ciKlass::primary_super_limit() ) {
1874 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1875 // (Folds up type checking code.)
1876 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1877 ciKlass *ss = klass->super_of_depth(depth);
1878 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1879 }
1880 const Type* aift = load_array_final_field(tkls, klass);
1881 if (aift != NULL) return aift;
1882 }
1883
1884 // We can still check if we are loading from the primary_supers array at a
1885 // shallow enough depth. Even though the klass is not exact, entries less
1886 // than or equal to its super depth are correct.
1887 if (klass->is_loaded() ) {
1888 ciType *inner = klass;
1889 while( inner->is_obj_array_klass() )
1890 inner = inner->as_obj_array_klass()->base_element_type();
1891 if( inner->is_instance_klass() &&
1892 !inner->as_instance_klass()->flags().is_interface() ) {
1893 // Compute index into primary_supers array
1894 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
1895 // Check for overflowing; use unsigned compare to handle the negative case.
1896 if( depth < ciKlass::primary_super_limit() &&
1897 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
1898 // The field is an element of Klass::_primary_supers. Return its (constant) value.
1899 // (Folds up type checking code.)
1900 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
1901 ciKlass *ss = klass->super_of_depth(depth);
1902 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
1903 }
1904 }
1905 }
1906
1907 // If the type is enough to determine that the thing is not an array,
1908 // we can give the layout_helper a positive interval type.
1909 // This will help short-circuit some reflective code.
1910 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
1911 && !klass->is_array_klass() // not directly typed as an array
1912 && !klass->is_interface() // specifically not Serializable & Cloneable
1913 && !klass->is_java_lang_Object() // not the supertype of all T[]
1914 ) {
1915 // Note: When interfaces are reliable, we can narrow the interface
1916 // test to (klass != Serializable && klass != Cloneable).
1917 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
1918 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
1919 // The key property of this type is that it folds up tests
1920 // for array-ness, since it proves that the layout_helper is positive.
1921 // Thus, a generic value like the basic object layout helper works fine.
1922 return TypeInt::make(min_size, max_jint, Type::WidenMin);
1923 }
1924 }
1925
1926 // If we are loading from a freshly-allocated object, produce a zero,
1927 // if the load is provably beyond the header of the object.
1928 // (Also allow a variable load from a fresh array to produce zero.)
1929 const TypeOopPtr *tinst = tp->isa_oopptr();
1930 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
1931 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
1932 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
1933 Node* value = can_see_stored_value(mem,phase);
1934 if (value != NULL && value->is_Con()) {
1935 assert(value->bottom_type()->higher_equal(_type),"sanity");
1936 return value->bottom_type();
1937 }
1938 }
1939
1940 if (is_instance) {
1941 // If we have an instance type and our memory input is the
1942 // programs's initial memory state, there is no matching store,
1943 // so just return a zero of the appropriate type
1944 Node *mem = in(MemNode::Memory);
1945 if (mem->is_Parm() && mem->in(0)->is_Start()) {
1946 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
1947 return Type::get_zero_type(_type->basic_type());
1948 }
1949 }
1950 return _type;
1951 }
1952
1953 //------------------------------match_edge-------------------------------------
1954 // Do we Match on this edge index or not? Match only the address.
1955 uint LoadNode::match_edge(uint idx) const {
1956 return idx == MemNode::Address;
1957 }
1958
1959 //--------------------------LoadBNode::Ideal--------------------------------------
1960 //
1961 // If the previous store is to the same address as this load,
1962 // and the value stored was larger than a byte, replace this load
1963 // with the value stored truncated to a byte. If no truncation is
1964 // needed, the replacement is done in LoadNode::Identity().
1965 //
1966 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1967 Node* mem = in(MemNode::Memory);
1968 Node* value = can_see_stored_value(mem,phase);
1969 if( value && !phase->type(value)->higher_equal( _type ) ) {
1970 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
1971 return new RShiftINode(result, phase->intcon(24));
1972 }
1973 // Identity call will handle the case where truncation is not needed.
1974 return LoadNode::Ideal(phase, can_reshape);
1975 }
1976
1977 const Type* LoadBNode::Value(PhaseGVN* phase) const {
1978 Node* mem = in(MemNode::Memory);
1979 Node* value = can_see_stored_value(mem,phase);
1980 if (value != NULL && value->is_Con() &&
1981 !value->bottom_type()->higher_equal(_type)) {
1982 // If the input to the store does not fit with the load's result type,
1983 // it must be truncated. We can't delay until Ideal call since
1984 // a singleton Value is needed for split_thru_phi optimization.
1985 int con = value->get_int();
1986 return TypeInt::make((con << 24) >> 24);
1987 }
1988 return LoadNode::Value(phase);
1989 }
1990
1991 //--------------------------LoadUBNode::Ideal-------------------------------------
1992 //
1993 // If the previous store is to the same address as this load,
1994 // and the value stored was larger than a byte, replace this load
1995 // with the value stored truncated to a byte. If no truncation is
1996 // needed, the replacement is done in LoadNode::Identity().
1997 //
1998 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
1999 Node* mem = in(MemNode::Memory);
2000 Node* value = can_see_stored_value(mem, phase);
2001 if (value && !phase->type(value)->higher_equal(_type))
2002 return new AndINode(value, phase->intcon(0xFF));
2003 // Identity call will handle the case where truncation is not needed.
2004 return LoadNode::Ideal(phase, can_reshape);
2005 }
2006
2007 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2008 Node* mem = in(MemNode::Memory);
2009 Node* value = can_see_stored_value(mem,phase);
2010 if (value != NULL && value->is_Con() &&
2011 !value->bottom_type()->higher_equal(_type)) {
2012 // If the input to the store does not fit with the load's result type,
2013 // it must be truncated. We can't delay until Ideal call since
2014 // a singleton Value is needed for split_thru_phi optimization.
2015 int con = value->get_int();
2016 return TypeInt::make(con & 0xFF);
2017 }
2018 return LoadNode::Value(phase);
2019 }
2020
2021 //--------------------------LoadUSNode::Ideal-------------------------------------
2022 //
2023 // If the previous store is to the same address as this load,
2024 // and the value stored was larger than a char, replace this load
2025 // with the value stored truncated to a char. If no truncation is
2026 // needed, the replacement is done in LoadNode::Identity().
2027 //
2028 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2029 Node* mem = in(MemNode::Memory);
2030 Node* value = can_see_stored_value(mem,phase);
2031 if( value && !phase->type(value)->higher_equal( _type ) )
2032 return new AndINode(value,phase->intcon(0xFFFF));
2033 // Identity call will handle the case where truncation is not needed.
2034 return LoadNode::Ideal(phase, can_reshape);
2035 }
2036
2037 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2038 Node* mem = in(MemNode::Memory);
2039 Node* value = can_see_stored_value(mem,phase);
2040 if (value != NULL && value->is_Con() &&
2041 !value->bottom_type()->higher_equal(_type)) {
2042 // If the input to the store does not fit with the load's result type,
2043 // it must be truncated. We can't delay until Ideal call since
2044 // a singleton Value is needed for split_thru_phi optimization.
2045 int con = value->get_int();
2046 return TypeInt::make(con & 0xFFFF);
2047 }
2048 return LoadNode::Value(phase);
2049 }
2050
2051 //--------------------------LoadSNode::Ideal--------------------------------------
2052 //
2053 // If the previous store is to the same address as this load,
2054 // and the value stored was larger than a short, replace this load
2055 // with the value stored truncated to a short. If no truncation is
2056 // needed, the replacement is done in LoadNode::Identity().
2057 //
2058 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2059 Node* mem = in(MemNode::Memory);
2060 Node* value = can_see_stored_value(mem,phase);
2061 if( value && !phase->type(value)->higher_equal( _type ) ) {
2062 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
2063 return new RShiftINode(result, phase->intcon(16));
2064 }
2065 // Identity call will handle the case where truncation is not needed.
2066 return LoadNode::Ideal(phase, can_reshape);
2067 }
2068
2069 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2070 Node* mem = in(MemNode::Memory);
2071 Node* value = can_see_stored_value(mem,phase);
2072 if (value != NULL && value->is_Con() &&
2073 !value->bottom_type()->higher_equal(_type)) {
2074 // If the input to the store does not fit with the load's result type,
2075 // it must be truncated. We can't delay until Ideal call since
2076 // a singleton Value is needed for split_thru_phi optimization.
2077 int con = value->get_int();
2078 return TypeInt::make((con << 16) >> 16);
2079 }
2080 return LoadNode::Value(phase);
2081 }
2082
2083 //=============================================================================
2084 //----------------------------LoadKlassNode::make------------------------------
2085 // Polymorphic factory method:
2086 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2087 // sanity check the alias category against the created node type
2088 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2089 assert(adr_type != NULL, "expecting TypeKlassPtr");
2090 #ifdef _LP64
2091 if (adr_type->is_ptr_to_narrowklass()) {
2092 assert(UseCompressedClassPointers, "no compressed klasses");
2093 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2094 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2095 }
2096 #endif
2097 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2098 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2099 }
2100
2101 //------------------------------Value------------------------------------------
2102 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2103 return klass_value_common(phase);
2104 }
2105
2106 // In most cases, LoadKlassNode does not have the control input set. If the control
2107 // input is set, it must not be removed (by LoadNode::Ideal()).
2108 bool LoadKlassNode::can_remove_control() const {
2109 return false;
2110 }
2111
2112 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2113 // Either input is TOP ==> the result is TOP
2114 const Type *t1 = phase->type( in(MemNode::Memory) );
2115 if (t1 == Type::TOP) return Type::TOP;
2116 Node *adr = in(MemNode::Address);
2117 const Type *t2 = phase->type( adr );
2118 if (t2 == Type::TOP) return Type::TOP;
2119 const TypePtr *tp = t2->is_ptr();
2120 if (TypePtr::above_centerline(tp->ptr()) ||
2121 tp->ptr() == TypePtr::Null) return Type::TOP;
2122
2123 // Return a more precise klass, if possible
2124 const TypeInstPtr *tinst = tp->isa_instptr();
2125 if (tinst != NULL) {
2126 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2127 int offset = tinst->offset();
2128 if (ik == phase->C->env()->Class_klass()
2129 && (offset == java_lang_Class::klass_offset_in_bytes() ||
2130 offset == java_lang_Class::array_klass_offset_in_bytes())) {
2131 // We are loading a special hidden field from a Class mirror object,
2132 // the field which points to the VM's Klass metaobject.
2133 ciType* t = tinst->java_mirror_type();
2134 // java_mirror_type returns non-null for compile-time Class constants.
2135 if (t != NULL) {
2136 // constant oop => constant klass
2137 if (offset == java_lang_Class::array_klass_offset_in_bytes()) {
2138 if (t->is_void()) {
2139 // We cannot create a void array. Since void is a primitive type return null
2140 // klass. Users of this result need to do a null check on the returned klass.
2141 return TypePtr::NULL_PTR;
2142 }
2143 return TypeKlassPtr::make(ciArrayKlass::make(t));
2144 }
2145 if (!t->is_klass()) {
2146 // a primitive Class (e.g., int.class) has NULL for a klass field
2147 return TypePtr::NULL_PTR;
2148 }
2149 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2150 return TypeKlassPtr::make(t->as_klass());
2151 }
2152 // non-constant mirror, so we can't tell what's going on
2153 }
2154 if( !ik->is_loaded() )
2155 return _type; // Bail out if not loaded
2156 if (offset == oopDesc::klass_offset_in_bytes()) {
2157 if (tinst->klass_is_exact()) {
2158 return TypeKlassPtr::make(ik);
2159 }
2160 // See if we can become precise: no subklasses and no interface
2161 // (Note: We need to support verified interfaces.)
2162 if (!ik->is_interface() && !ik->has_subklass()) {
2163 //assert(!UseExactTypes, "this code should be useless with exact types");
2164 // Add a dependence; if any subclass added we need to recompile
2165 if (!ik->is_final()) {
2166 // %%% should use stronger assert_unique_concrete_subtype instead
2167 phase->C->dependencies()->assert_leaf_type(ik);
2168 }
2169 // Return precise klass
2170 return TypeKlassPtr::make(ik);
2171 }
2172
2173 // Return root of possible klass
2174 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2175 }
2176 }
2177
2178 // Check for loading klass from an array
2179 const TypeAryPtr *tary = tp->isa_aryptr();
2180 if( tary != NULL ) {
2181 ciKlass *tary_klass = tary->klass();
2182 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2183 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2184 if (tary->klass_is_exact()) {
2185 return TypeKlassPtr::make(tary_klass);
2186 }
2187 ciArrayKlass *ak = tary->klass()->as_array_klass();
2188 // If the klass is an object array, we defer the question to the
2189 // array component klass.
2190 if( ak->is_obj_array_klass() ) {
2191 assert( ak->is_loaded(), "" );
2192 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2193 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2194 ciInstanceKlass* ik = base_k->as_instance_klass();
2195 // See if we can become precise: no subklasses and no interface
2196 if (!ik->is_interface() && !ik->has_subklass()) {
2197 //assert(!UseExactTypes, "this code should be useless with exact types");
2198 // Add a dependence; if any subclass added we need to recompile
2199 if (!ik->is_final()) {
2200 phase->C->dependencies()->assert_leaf_type(ik);
2201 }
2202 // Return precise array klass
2203 return TypeKlassPtr::make(ak);
2204 }
2205 }
2206 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2207 } else { // Found a type-array?
2208 //assert(!UseExactTypes, "this code should be useless with exact types");
2209 assert( ak->is_type_array_klass(), "" );
2210 return TypeKlassPtr::make(ak); // These are always precise
2211 }
2212 }
2213 }
2214
2215 // Check for loading klass from an array klass
2216 const TypeKlassPtr *tkls = tp->isa_klassptr();
2217 if (tkls != NULL && !StressReflectiveCode) {
2218 ciKlass* klass = tkls->klass();
2219 if( !klass->is_loaded() )
2220 return _type; // Bail out if not loaded
2221 if( klass->is_obj_array_klass() &&
2222 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2223 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2224 // // Always returning precise element type is incorrect,
2225 // // e.g., element type could be object and array may contain strings
2226 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2227
2228 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2229 // according to the element type's subclassing.
2230 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2231 }
2232 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2233 tkls->offset() == in_bytes(Klass::super_offset())) {
2234 ciKlass* sup = klass->as_instance_klass()->super();
2235 // The field is Klass::_super. Return its (constant) value.
2236 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2237 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2238 }
2239 }
2240
2241 // Bailout case
2242 return LoadNode::Value(phase);
2243 }
2244
2245 //------------------------------Identity---------------------------------------
2246 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2247 // Also feed through the klass in Allocate(...klass...)._klass.
2248 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2249 return klass_identity_common(phase);
2250 }
2251
2252 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2253 Node* x = LoadNode::Identity(phase);
2254 if (x != this) return x;
2255
2256 // Take apart the address into an oop and and offset.
2257 // Return 'this' if we cannot.
2258 Node* adr = in(MemNode::Address);
2259 intptr_t offset = 0;
2260 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2261 if (base == NULL) return this;
2262 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2263 if (toop == NULL) return this;
2264
2265 // Step over potential GC barrier for OopHandle resolve
2266 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2267 if (bs->is_gc_barrier_node(base)) {
2268 base = bs->step_over_gc_barrier(base);
2269 }
2270
2271 // We can fetch the klass directly through an AllocateNode.
2272 // This works even if the klass is not constant (clone or newArray).
2273 if (offset == oopDesc::klass_offset_in_bytes()) {
2274 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2275 if (allocated_klass != NULL) {
2276 return allocated_klass;
2277 }
2278 }
2279
2280 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2281 // See inline_native_Class_query for occurrences of these patterns.
2282 // Java Example: x.getClass().isAssignableFrom(y)
2283 //
2284 // This improves reflective code, often making the Class
2285 // mirror go completely dead. (Current exception: Class
2286 // mirrors may appear in debug info, but we could clean them out by
2287 // introducing a new debug info operator for Klass.java_mirror).
2288
2289 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2290 && offset == java_lang_Class::klass_offset_in_bytes()) {
2291 if (base->is_Load()) {
2292 Node* base2 = base->in(MemNode::Address);
2293 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2294 Node* adr2 = base2->in(MemNode::Address);
2295 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2296 if (tkls != NULL && !tkls->empty()
2297 && (tkls->klass()->is_instance_klass() ||
2298 tkls->klass()->is_array_klass())
2299 && adr2->is_AddP()
2300 ) {
2301 int mirror_field = in_bytes(Klass::java_mirror_offset());
2302 if (tkls->offset() == mirror_field) {
2303 return adr2->in(AddPNode::Base);
2304 }
2305 }
2306 }
2307 }
2308 }
2309
2310 return this;
2311 }
2312
2313
2314 //------------------------------Value------------------------------------------
2315 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2316 const Type *t = klass_value_common(phase);
2317 if (t == Type::TOP)
2318 return t;
2319
2320 return t->make_narrowklass();
2321 }
2322
2323 //------------------------------Identity---------------------------------------
2324 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2325 // Also feed through the klass in Allocate(...klass...)._klass.
2326 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2327 Node *x = klass_identity_common(phase);
2328
2329 const Type *t = phase->type( x );
2330 if( t == Type::TOP ) return x;
2331 if( t->isa_narrowklass()) return x;
2332 assert (!t->isa_narrowoop(), "no narrow oop here");
2333
2334 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2335 }
2336
2337 //------------------------------Value-----------------------------------------
2338 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2339 // Either input is TOP ==> the result is TOP
2340 const Type *t1 = phase->type( in(MemNode::Memory) );
2341 if( t1 == Type::TOP ) return Type::TOP;
2342 Node *adr = in(MemNode::Address);
2343 const Type *t2 = phase->type( adr );
2344 if( t2 == Type::TOP ) return Type::TOP;
2345 const TypePtr *tp = t2->is_ptr();
2346 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2347 const TypeAryPtr *tap = tp->isa_aryptr();
2348 if( !tap ) return _type;
2349 return tap->size();
2350 }
2351
2352 //-------------------------------Ideal---------------------------------------
2353 // Feed through the length in AllocateArray(...length...)._length.
2354 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2355 Node* p = MemNode::Ideal_common(phase, can_reshape);
2356 if (p) return (p == NodeSentinel) ? NULL : p;
2357
2358 // Take apart the address into an oop and and offset.
2359 // Return 'this' if we cannot.
2360 Node* adr = in(MemNode::Address);
2361 intptr_t offset = 0;
2362 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2363 if (base == NULL) return NULL;
2364 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2365 if (tary == NULL) return NULL;
2366
2367 // We can fetch the length directly through an AllocateArrayNode.
2368 // This works even if the length is not constant (clone or newArray).
2369 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2370 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2371 if (alloc != NULL) {
2372 Node* allocated_length = alloc->Ideal_length();
2373 Node* len = alloc->make_ideal_length(tary, phase);
2374 if (allocated_length != len) {
2375 // New CastII improves on this.
2376 return len;
2377 }
2378 }
2379 }
2380
2381 return NULL;
2382 }
2383
2384 //------------------------------Identity---------------------------------------
2385 // Feed through the length in AllocateArray(...length...)._length.
2386 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2387 Node* x = LoadINode::Identity(phase);
2388 if (x != this) return x;
2389
2390 // Take apart the address into an oop and and offset.
2391 // Return 'this' if we cannot.
2392 Node* adr = in(MemNode::Address);
2393 intptr_t offset = 0;
2394 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2395 if (base == NULL) return this;
2396 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2397 if (tary == NULL) return this;
2398
2399 // We can fetch the length directly through an AllocateArrayNode.
2400 // This works even if the length is not constant (clone or newArray).
2401 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2402 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2403 if (alloc != NULL) {
2404 Node* allocated_length = alloc->Ideal_length();
2405 // Do not allow make_ideal_length to allocate a CastII node.
2406 Node* len = alloc->make_ideal_length(tary, phase, false);
2407 if (allocated_length == len) {
2408 // Return allocated_length only if it would not be improved by a CastII.
2409 return allocated_length;
2410 }
2411 }
2412 }
2413
2414 return this;
2415
2416 }
2417
2418 //=============================================================================
2419 //---------------------------StoreNode::make-----------------------------------
2420 // Polymorphic factory method:
2421 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2422 assert((mo == unordered || mo == release), "unexpected");
2423 Compile* C = gvn.C;
2424 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2425 ctl != NULL, "raw memory operations should have control edge");
2426
2427 switch (bt) {
2428 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2429 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2430 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2431 case T_CHAR:
2432 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2433 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2434 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2435 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2436 case T_METADATA:
2437 case T_ADDRESS:
2438 case T_OBJECT:
2439 #ifdef _LP64
2440 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2441 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2442 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2443 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2444 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2445 adr->bottom_type()->isa_rawptr())) {
2446 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2447 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2448 }
2449 #endif
2450 {
2451 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2452 }
2453 default:
2454 ShouldNotReachHere();
2455 return (StoreNode*)NULL;
2456 }
2457 }
2458
2459 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2460 bool require_atomic = true;
2461 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2462 }
2463
2464 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2465 bool require_atomic = true;
2466 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2467 }
2468
2469
2470 //--------------------------bottom_type----------------------------------------
2471 const Type *StoreNode::bottom_type() const {
2472 return Type::MEMORY;
2473 }
2474
2475 //------------------------------hash-------------------------------------------
2476 uint StoreNode::hash() const {
2477 // unroll addition of interesting fields
2478 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2479
2480 // Since they are not commoned, do not hash them:
2481 return NO_HASH;
2482 }
2483
2484 //------------------------------Ideal------------------------------------------
2485 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2486 // When a store immediately follows a relevant allocation/initialization,
2487 // try to capture it into the initialization, or hoist it above.
2488 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2489 Node* p = MemNode::Ideal_common(phase, can_reshape);
2490 if (p) return (p == NodeSentinel) ? NULL : p;
2491
2492 Node* mem = in(MemNode::Memory);
2493 Node* address = in(MemNode::Address);
2494 // Back-to-back stores to same address? Fold em up. Generally
2495 // unsafe if I have intervening uses... Also disallowed for StoreCM
2496 // since they must follow each StoreP operation. Redundant StoreCMs
2497 // are eliminated just before matching in final_graph_reshape.
2498 {
2499 Node* st = mem;
2500 // If Store 'st' has more than one use, we cannot fold 'st' away.
2501 // For example, 'st' might be the final state at a conditional
2502 // return. Or, 'st' might be used by some node which is live at
2503 // the same time 'st' is live, which might be unschedulable. So,
2504 // require exactly ONE user until such time as we clone 'mem' for
2505 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2506 // true).
2507 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2508 // Looking at a dead closed cycle of memory?
2509 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2510 assert(Opcode() == st->Opcode() ||
2511 st->Opcode() == Op_StoreVector ||
2512 Opcode() == Op_StoreVector ||
2513 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2514 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2515 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2516 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2517 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2518
2519 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2520 st->as_Store()->memory_size() <= this->memory_size()) {
2521 Node* use = st->raw_out(0);
2522 phase->igvn_rehash_node_delayed(use);
2523 if (can_reshape) {
2524 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2525 } else {
2526 // It's OK to do this in the parser, since DU info is always accurate,
2527 // and the parser always refers to nodes via SafePointNode maps.
2528 use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2529 }
2530 return this;
2531 }
2532 st = st->in(MemNode::Memory);
2533 }
2534 }
2535
2536
2537 // Capture an unaliased, unconditional, simple store into an initializer.
2538 // Or, if it is independent of the allocation, hoist it above the allocation.
2539 if (ReduceFieldZeroing && /*can_reshape &&*/
2540 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2541 InitializeNode* init = mem->in(0)->as_Initialize();
2542 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2543 if (offset > 0) {
2544 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2545 // If the InitializeNode captured me, it made a raw copy of me,
2546 // and I need to disappear.
2547 if (moved != NULL) {
2548 // %%% hack to ensure that Ideal returns a new node:
2549 mem = MergeMemNode::make(mem);
2550 return mem; // fold me away
2551 }
2552 }
2553 }
2554
2555 return NULL; // No further progress
2556 }
2557
2558 //------------------------------Value-----------------------------------------
2559 const Type* StoreNode::Value(PhaseGVN* phase) const {
2560 // Either input is TOP ==> the result is TOP
2561 const Type *t1 = phase->type( in(MemNode::Memory) );
2562 if( t1 == Type::TOP ) return Type::TOP;
2563 const Type *t2 = phase->type( in(MemNode::Address) );
2564 if( t2 == Type::TOP ) return Type::TOP;
2565 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2566 if( t3 == Type::TOP ) return Type::TOP;
2567 return Type::MEMORY;
2568 }
2569
2570 //------------------------------Identity---------------------------------------
2571 // Remove redundant stores:
2572 // Store(m, p, Load(m, p)) changes to m.
2573 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
2574 Node* StoreNode::Identity(PhaseGVN* phase) {
2575 Node* mem = in(MemNode::Memory);
2576 Node* adr = in(MemNode::Address);
2577 Node* val = in(MemNode::ValueIn);
2578
2579 Node* result = this;
2580
2581 // Load then Store? Then the Store is useless
2582 if (val->is_Load() &&
2583 val->in(MemNode::Address)->eqv_uncast(adr) &&
2584 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2585 val->as_Load()->store_Opcode() == Opcode()) {
2586 result = mem;
2587 }
2588
2589 // Two stores in a row of the same value?
2590 if (result == this &&
2591 mem->is_Store() &&
2592 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2593 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2594 mem->Opcode() == Opcode()) {
2595 result = mem;
2596 }
2597
2598 // Store of zero anywhere into a freshly-allocated object?
2599 // Then the store is useless.
2600 // (It must already have been captured by the InitializeNode.)
2601 if (result == this &&
2602 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2603 // a newly allocated object is already all-zeroes everywhere
2604 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2605 result = mem;
2606 }
2607
2608 if (result == this) {
2609 // the store may also apply to zero-bits in an earlier object
2610 Node* prev_mem = find_previous_store(phase);
2611 // Steps (a), (b): Walk past independent stores to find an exact match.
2612 if (prev_mem != NULL) {
2613 Node* prev_val = can_see_stored_value(prev_mem, phase);
2614 if (prev_val != NULL && phase->eqv(prev_val, val)) {
2615 // prev_val and val might differ by a cast; it would be good
2616 // to keep the more informative of the two.
2617 result = mem;
2618 }
2619 }
2620 }
2621 }
2622
2623 if (result != this && phase->is_IterGVN() != NULL) {
2624 MemBarNode* trailing = trailing_membar();
2625 if (trailing != NULL) {
2626 #ifdef ASSERT
2627 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2628 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects");
2629 #endif
2630 PhaseIterGVN* igvn = phase->is_IterGVN();
2631 trailing->remove(igvn);
2632 }
2633 }
2634
2635 return result;
2636 }
2637
2638 //------------------------------match_edge-------------------------------------
2639 // Do we Match on this edge index or not? Match only memory & value
2640 uint StoreNode::match_edge(uint idx) const {
2641 return idx == MemNode::Address || idx == MemNode::ValueIn;
2642 }
2643
2644 //------------------------------cmp--------------------------------------------
2645 // Do not common stores up together. They generally have to be split
2646 // back up anyways, so do not bother.
2647 uint StoreNode::cmp( const Node &n ) const {
2648 return (&n == this); // Always fail except on self
2649 }
2650
2651 //------------------------------Ideal_masked_input-----------------------------
2652 // Check for a useless mask before a partial-word store
2653 // (StoreB ... (AndI valIn conIa) )
2654 // If (conIa & mask == mask) this simplifies to
2655 // (StoreB ... (valIn) )
2656 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2657 Node *val = in(MemNode::ValueIn);
2658 if( val->Opcode() == Op_AndI ) {
2659 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2660 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2661 set_req(MemNode::ValueIn, val->in(1));
2662 return this;
2663 }
2664 }
2665 return NULL;
2666 }
2667
2668
2669 //------------------------------Ideal_sign_extended_input----------------------
2670 // Check for useless sign-extension before a partial-word store
2671 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2672 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2673 // (StoreB ... (valIn) )
2674 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2675 Node *val = in(MemNode::ValueIn);
2676 if( val->Opcode() == Op_RShiftI ) {
2677 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2678 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2679 Node *shl = val->in(1);
2680 if( shl->Opcode() == Op_LShiftI ) {
2681 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2682 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2683 set_req(MemNode::ValueIn, shl->in(1));
2684 return this;
2685 }
2686 }
2687 }
2688 }
2689 return NULL;
2690 }
2691
2692 //------------------------------value_never_loaded-----------------------------------
2693 // Determine whether there are any possible loads of the value stored.
2694 // For simplicity, we actually check if there are any loads from the
2695 // address stored to, not just for loads of the value stored by this node.
2696 //
2697 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2698 Node *adr = in(Address);
2699 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2700 if (adr_oop == NULL)
2701 return false;
2702 if (!adr_oop->is_known_instance_field())
2703 return false; // if not a distinct instance, there may be aliases of the address
2704 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2705 Node *use = adr->fast_out(i);
2706 if (use->is_Load() || use->is_LoadStore()) {
2707 return false;
2708 }
2709 }
2710 return true;
2711 }
2712
2713 MemBarNode* StoreNode::trailing_membar() const {
2714 if (is_release()) {
2715 MemBarNode* trailing_mb = NULL;
2716 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2717 Node* u = fast_out(i);
2718 if (u->is_MemBar()) {
2719 if (u->as_MemBar()->trailing_store()) {
2720 assert(u->Opcode() == Op_MemBarVolatile, "");
2721 assert(trailing_mb == NULL, "only one");
2722 trailing_mb = u->as_MemBar();
2723 #ifdef ASSERT
2724 Node* leading = u->as_MemBar()->leading_membar();
2725 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2726 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2727 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2728 #endif
2729 } else {
2730 assert(u->as_MemBar()->standalone(), "");
2731 }
2732 }
2733 }
2734 return trailing_mb;
2735 }
2736 return NULL;
2737 }
2738
2739
2740 //=============================================================================
2741 //------------------------------Ideal------------------------------------------
2742 // If the store is from an AND mask that leaves the low bits untouched, then
2743 // we can skip the AND operation. If the store is from a sign-extension
2744 // (a left shift, then right shift) we can skip both.
2745 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2746 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2747 if( progress != NULL ) return progress;
2748
2749 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2750 if( progress != NULL ) return progress;
2751
2752 // Finally check the default case
2753 return StoreNode::Ideal(phase, can_reshape);
2754 }
2755
2756 //=============================================================================
2757 //------------------------------Ideal------------------------------------------
2758 // If the store is from an AND mask that leaves the low bits untouched, then
2759 // we can skip the AND operation
2760 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2761 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2762 if( progress != NULL ) return progress;
2763
2764 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2765 if( progress != NULL ) return progress;
2766
2767 // Finally check the default case
2768 return StoreNode::Ideal(phase, can_reshape);
2769 }
2770
2771 //=============================================================================
2772 //------------------------------Identity---------------------------------------
2773 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2774 // No need to card mark when storing a null ptr
2775 Node* my_store = in(MemNode::OopStore);
2776 if (my_store->is_Store()) {
2777 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2778 if( t1 == TypePtr::NULL_PTR ) {
2779 return in(MemNode::Memory);
2780 }
2781 }
2782 return this;
2783 }
2784
2785 //=============================================================================
2786 //------------------------------Ideal---------------------------------------
2787 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2788 Node* progress = StoreNode::Ideal(phase, can_reshape);
2789 if (progress != NULL) return progress;
2790
2791 Node* my_store = in(MemNode::OopStore);
2792 if (my_store->is_MergeMem()) {
2793 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2794 set_req(MemNode::OopStore, mem);
2795 return this;
2796 }
2797
2798 return NULL;
2799 }
2800
2801 //------------------------------Value-----------------------------------------
2802 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2803 // Either input is TOP ==> the result is TOP
2804 const Type *t = phase->type( in(MemNode::Memory) );
2805 if( t == Type::TOP ) return Type::TOP;
2806 t = phase->type( in(MemNode::Address) );
2807 if( t == Type::TOP ) return Type::TOP;
2808 t = phase->type( in(MemNode::ValueIn) );
2809 if( t == Type::TOP ) return Type::TOP;
2810 // If extra input is TOP ==> the result is TOP
2811 t = phase->type( in(MemNode::OopStore) );
2812 if( t == Type::TOP ) return Type::TOP;
2813
2814 return StoreNode::Value( phase );
2815 }
2816
2817
2818 //=============================================================================
2819 //----------------------------------SCMemProjNode------------------------------
2820 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2821 {
2822 return bottom_type();
2823 }
2824
2825 //=============================================================================
2826 //----------------------------------LoadStoreNode------------------------------
2827 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2828 : Node(required),
2829 _type(rt),
2830 _adr_type(at)
2831 {
2832 init_req(MemNode::Control, c );
2833 init_req(MemNode::Memory , mem);
2834 init_req(MemNode::Address, adr);
2835 init_req(MemNode::ValueIn, val);
2836 init_class_id(Class_LoadStore);
2837 }
2838
2839 uint LoadStoreNode::ideal_reg() const {
2840 return _type->ideal_reg();
2841 }
2842
2843 bool LoadStoreNode::result_not_used() const {
2844 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2845 Node *x = fast_out(i);
2846 if (x->Opcode() == Op_SCMemProj) continue;
2847 return false;
2848 }
2849 return true;
2850 }
2851
2852 MemBarNode* LoadStoreNode::trailing_membar() const {
2853 MemBarNode* trailing = NULL;
2854 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2855 Node* u = fast_out(i);
2856 if (u->is_MemBar()) {
2857 if (u->as_MemBar()->trailing_load_store()) {
2858 assert(u->Opcode() == Op_MemBarAcquire, "");
2859 assert(trailing == NULL, "only one");
2860 trailing = u->as_MemBar();
2861 #ifdef ASSERT
2862 Node* leading = trailing->leading_membar();
2863 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2864 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
2865 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
2866 #endif
2867 } else {
2868 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
2869 }
2870 }
2871 }
2872
2873 return trailing;
2874 }
2875
2876 uint LoadStoreNode::size_of() const { return sizeof(*this); }
2877
2878 //=============================================================================
2879 //----------------------------------LoadStoreConditionalNode--------------------
2880 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
2881 init_req(ExpectedIn, ex );
2882 }
2883
2884 //=============================================================================
2885 //-------------------------------adr_type--------------------------------------
2886 const TypePtr* ClearArrayNode::adr_type() const {
2887 Node *adr = in(3);
2888 if (adr == NULL) return NULL; // node is dead
2889 return MemNode::calculate_adr_type(adr->bottom_type());
2890 }
2891
2892 //------------------------------match_edge-------------------------------------
2893 // Do we Match on this edge index or not? Do not match memory
2894 uint ClearArrayNode::match_edge(uint idx) const {
2895 return idx > 1;
2896 }
2897
2898 //------------------------------Identity---------------------------------------
2899 // Clearing a zero length array does nothing
2900 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
2901 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
2902 }
2903
2904 //------------------------------Idealize---------------------------------------
2905 // Clearing a short array is faster with stores
2906 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2907 // Already know this is a large node, do not try to ideal it
2908 if (!IdealizeClearArrayNode || _is_large) return NULL;
2909
2910 const int unit = BytesPerLong;
2911 const TypeX* t = phase->type(in(2))->isa_intptr_t();
2912 if (!t) return NULL;
2913 if (!t->is_con()) return NULL;
2914 intptr_t raw_count = t->get_con();
2915 intptr_t size = raw_count;
2916 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
2917 // Clearing nothing uses the Identity call.
2918 // Negative clears are possible on dead ClearArrays
2919 // (see jck test stmt114.stmt11402.val).
2920 if (size <= 0 || size % unit != 0) return NULL;
2921 intptr_t count = size / unit;
2922 // Length too long; communicate this to matchers and assemblers.
2923 // Assemblers are responsible to produce fast hardware clears for it.
2924 if (size > InitArrayShortSize) {
2925 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
2926 }
2927 Node *mem = in(1);
2928 if( phase->type(mem)==Type::TOP ) return NULL;
2929 Node *adr = in(3);
2930 const Type* at = phase->type(adr);
2931 if( at==Type::TOP ) return NULL;
2932 const TypePtr* atp = at->isa_ptr();
2933 // adjust atp to be the correct array element address type
2934 if (atp == NULL) atp = TypePtr::BOTTOM;
2935 else atp = atp->add_offset(Type::OffsetBot);
2936 // Get base for derived pointer purposes
2937 if( adr->Opcode() != Op_AddP ) Unimplemented();
2938 Node *base = adr->in(1);
2939
2940 Node *zero = phase->makecon(TypeLong::ZERO);
2941 Node *off = phase->MakeConX(BytesPerLong);
2942 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2943 count--;
2944 while( count-- ) {
2945 mem = phase->transform(mem);
2946 adr = phase->transform(new AddPNode(base,adr,off));
2947 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
2948 }
2949 return mem;
2950 }
2951
2952 //----------------------------step_through----------------------------------
2953 // Return allocation input memory edge if it is different instance
2954 // or itself if it is the one we are looking for.
2955 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
2956 Node* n = *np;
2957 assert(n->is_ClearArray(), "sanity");
2958 intptr_t offset;
2959 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
2960 // This method is called only before Allocate nodes are expanded
2961 // during macro nodes expansion. Before that ClearArray nodes are
2962 // only generated in PhaseMacroExpand::generate_arraycopy() (before
2963 // Allocate nodes are expanded) which follows allocations.
2964 assert(alloc != NULL, "should have allocation");
2965 if (alloc->_idx == instance_id) {
2966 // Can not bypass initialization of the instance we are looking for.
2967 return false;
2968 }
2969 // Otherwise skip it.
2970 InitializeNode* init = alloc->initialization();
2971 if (init != NULL)
2972 *np = init->in(TypeFunc::Memory);
2973 else
2974 *np = alloc->in(TypeFunc::Memory);
2975 return true;
2976 }
2977
2978 //----------------------------clear_memory-------------------------------------
2979 // Generate code to initialize object storage to zero.
2980 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
2981 intptr_t start_offset,
2982 Node* end_offset,
2983 PhaseGVN* phase) {
2984 intptr_t offset = start_offset;
2985
2986 int unit = BytesPerLong;
2987 if ((offset % unit) != 0) {
2988 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
2989 adr = phase->transform(adr);
2990 const TypePtr* atp = TypeRawPtr::BOTTOM;
2991 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
2992 mem = phase->transform(mem);
2993 offset += BytesPerInt;
2994 }
2995 assert((offset % unit) == 0, "");
2996
2997 // Initialize the remaining stuff, if any, with a ClearArray.
2998 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
2999 }
3000
3001 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3002 Node* start_offset,
3003 Node* end_offset,
3004 PhaseGVN* phase) {
3005 if (start_offset == end_offset) {
3006 // nothing to do
3007 return mem;
3008 }
3009
3010 int unit = BytesPerLong;
3011 Node* zbase = start_offset;
3012 Node* zend = end_offset;
3013
3014 // Scale to the unit required by the CPU:
3015 if (!Matcher::init_array_count_is_in_bytes) {
3016 Node* shift = phase->intcon(exact_log2(unit));
3017 zbase = phase->transform(new URShiftXNode(zbase, shift) );
3018 zend = phase->transform(new URShiftXNode(zend, shift) );
3019 }
3020
3021 // Bulk clear double-words
3022 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
3023 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
3024 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
3025 return phase->transform(mem);
3026 }
3027
3028 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3029 intptr_t start_offset,
3030 intptr_t end_offset,
3031 PhaseGVN* phase) {
3032 if (start_offset == end_offset) {
3033 // nothing to do
3034 return mem;
3035 }
3036
3037 assert((end_offset % BytesPerInt) == 0, "odd end offset");
3038 intptr_t done_offset = end_offset;
3039 if ((done_offset % BytesPerLong) != 0) {
3040 done_offset -= BytesPerInt;
3041 }
3042 if (done_offset > start_offset) {
3043 mem = clear_memory(ctl, mem, dest,
3044 start_offset, phase->MakeConX(done_offset), phase);
3045 }
3046 if (done_offset < end_offset) { // emit the final 32-bit store
3047 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3048 adr = phase->transform(adr);
3049 const TypePtr* atp = TypeRawPtr::BOTTOM;
3050 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3051 mem = phase->transform(mem);
3052 done_offset += BytesPerInt;
3053 }
3054 assert(done_offset == end_offset, "");
3055 return mem;
3056 }
3057
3058 //=============================================================================
3059 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3060 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
3061 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3062 #ifdef ASSERT
3063 , _pair_idx(0)
3064 #endif
3065 {
3066 init_class_id(Class_MemBar);
3067 Node* top = C->top();
3068 init_req(TypeFunc::I_O,top);
3069 init_req(TypeFunc::FramePtr,top);
3070 init_req(TypeFunc::ReturnAdr,top);
3071 if (precedent != NULL)
3072 init_req(TypeFunc::Parms, precedent);
3073 }
3074
3075 //------------------------------cmp--------------------------------------------
3076 uint MemBarNode::hash() const { return NO_HASH; }
3077 uint MemBarNode::cmp( const Node &n ) const {
3078 return (&n == this); // Always fail except on self
3079 }
3080
3081 //------------------------------make-------------------------------------------
3082 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3083 switch (opcode) {
3084 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
3085 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
3086 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
3087 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
3088 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3089 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3090 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
3091 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
3092 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
3093 case Op_Initialize: return new InitializeNode(C, atp, pn);
3094 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
3095 default: ShouldNotReachHere(); return NULL;
3096 }
3097 }
3098
3099 void MemBarNode::remove(PhaseIterGVN *igvn) {
3100 if (outcnt() != 2) {
3101 return;
3102 }
3103 if (trailing_store() || trailing_load_store()) {
3104 MemBarNode* leading = leading_membar();
3105 if (leading != NULL) {
3106 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3107 leading->remove(igvn);
3108 }
3109 }
3110 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3111 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3112 }
3113
3114 //------------------------------Ideal------------------------------------------
3115 // Return a node which is more "ideal" than the current node. Strip out
3116 // control copies
3117 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3118 if (remove_dead_region(phase, can_reshape)) return this;
3119 // Don't bother trying to transform a dead node
3120 if (in(0) && in(0)->is_top()) {
3121 return NULL;
3122 }
3123
3124 #if INCLUDE_ZGC
3125 if (UseZGC) {
3126 if (req() == (Precedent+1) && in(MemBarNode::Precedent)->in(0) != NULL && in(MemBarNode::Precedent)->in(0)->is_LoadBarrier()) {
3127 Node* load_node = in(MemBarNode::Precedent)->in(0)->in(LoadBarrierNode::Oop);
3128 set_req(MemBarNode::Precedent, load_node);
3129 return this;
3130 }
3131 }
3132 #endif
3133
3134 bool progress = false;
3135 // Eliminate volatile MemBars for scalar replaced objects.
3136 if (can_reshape && req() == (Precedent+1)) {
3137 bool eliminate = false;
3138 int opc = Opcode();
3139 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3140 // Volatile field loads and stores.
3141 Node* my_mem = in(MemBarNode::Precedent);
3142 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3143 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3144 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3145 // replace this Precedent (decodeN) with the Load instead.
3146 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3147 Node* load_node = my_mem->in(1);
3148 set_req(MemBarNode::Precedent, load_node);
3149 phase->is_IterGVN()->_worklist.push(my_mem);
3150 my_mem = load_node;
3151 } else {
3152 assert(my_mem->unique_out() == this, "sanity");
3153 del_req(Precedent);
3154 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3155 my_mem = NULL;
3156 }
3157 progress = true;
3158 }
3159 if (my_mem != NULL && my_mem->is_Mem()) {
3160 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3161 // Check for scalar replaced object reference.
3162 if( t_oop != NULL && t_oop->is_known_instance_field() &&
3163 t_oop->offset() != Type::OffsetBot &&
3164 t_oop->offset() != Type::OffsetTop) {
3165 eliminate = true;
3166 }
3167 }
3168 } else if (opc == Op_MemBarRelease) {
3169 // Final field stores.
3170 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3171 if ((alloc != NULL) && alloc->is_Allocate() &&
3172 alloc->as_Allocate()->does_not_escape_thread()) {
3173 // The allocated object does not escape.
3174 eliminate = true;
3175 }
3176 }
3177 if (eliminate) {
3178 // Replace MemBar projections by its inputs.
3179 PhaseIterGVN* igvn = phase->is_IterGVN();
3180 remove(igvn);
3181 // Must return either the original node (now dead) or a new node
3182 // (Do not return a top here, since that would break the uniqueness of top.)
3183 return new ConINode(TypeInt::ZERO);
3184 }
3185 }
3186 return progress ? this : NULL;
3187 }
3188
3189 //------------------------------Value------------------------------------------
3190 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3191 if( !in(0) ) return Type::TOP;
3192 if( phase->type(in(0)) == Type::TOP )
3193 return Type::TOP;
3194 return TypeTuple::MEMBAR;
3195 }
3196
3197 //------------------------------match------------------------------------------
3198 // Construct projections for memory.
3199 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3200 switch (proj->_con) {
3201 case TypeFunc::Control:
3202 case TypeFunc::Memory:
3203 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3204 }
3205 ShouldNotReachHere();
3206 return NULL;
3207 }
3208
3209 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3210 trailing->_kind = TrailingStore;
3211 leading->_kind = LeadingStore;
3212 #ifdef ASSERT
3213 trailing->_pair_idx = leading->_idx;
3214 leading->_pair_idx = leading->_idx;
3215 #endif
3216 }
3217
3218 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3219 trailing->_kind = TrailingLoadStore;
3220 leading->_kind = LeadingLoadStore;
3221 #ifdef ASSERT
3222 trailing->_pair_idx = leading->_idx;
3223 leading->_pair_idx = leading->_idx;
3224 #endif
3225 }
3226
3227 MemBarNode* MemBarNode::trailing_membar() const {
3228 ResourceMark rm;
3229 Node* trailing = (Node*)this;
3230 VectorSet seen(Thread::current()->resource_area());
3231 Node_Stack multis(0);
3232 do {
3233 Node* c = trailing;
3234 uint i = 0;
3235 do {
3236 trailing = NULL;
3237 for (; i < c->outcnt(); i++) {
3238 Node* next = c->raw_out(i);
3239 if (next != c && next->is_CFG()) {
3240 if (c->is_MultiBranch()) {
3241 if (multis.node() == c) {
3242 multis.set_index(i+1);
3243 } else {
3244 multis.push(c, i+1);
3245 }
3246 }
3247 trailing = next;
3248 break;
3249 }
3250 }
3251 if (trailing != NULL && !seen.test_set(trailing->_idx)) {
3252 break;
3253 }
3254 while (multis.size() > 0) {
3255 c = multis.node();
3256 i = multis.index();
3257 if (i < c->req()) {
3258 break;
3259 }
3260 multis.pop();
3261 }
3262 } while (multis.size() > 0);
3263 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
3264
3265 MemBarNode* mb = trailing->as_MemBar();
3266 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3267 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3268 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3269 return mb;
3270 }
3271
3272 MemBarNode* MemBarNode::leading_membar() const {
3273 ResourceMark rm;
3274 VectorSet seen(Thread::current()->resource_area());
3275 Node_Stack regions(0);
3276 Node* leading = in(0);
3277 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3278 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) {
3279 leading = NULL;
3280 while (regions.size() > 0 && leading == NULL) {
3281 Node* r = regions.node();
3282 uint i = regions.index();
3283 if (i < r->req()) {
3284 leading = r->in(i);
3285 regions.set_index(i+1);
3286 } else {
3287 regions.pop();
3288 }
3289 }
3290 if (leading == NULL) {
3291 assert(regions.size() == 0, "all paths should have been tried");
3292 return NULL;
3293 }
3294 }
3295 if (leading->is_Region()) {
3296 regions.push(leading, 2);
3297 leading = leading->in(1);
3298 } else {
3299 leading = leading->in(0);
3300 }
3301 }
3302 #ifdef ASSERT
3303 Unique_Node_List wq;
3304 wq.push((Node*)this);
3305 uint found = 0;
3306 for (uint i = 0; i < wq.size(); i++) {
3307 Node* n = wq.at(i);
3308 if (n->is_Region()) {
3309 for (uint j = 1; j < n->req(); j++) {
3310 Node* in = n->in(j);
3311 if (in != NULL && !in->is_top()) {
3312 wq.push(in);
3313 }
3314 }
3315 } else {
3316 if (n->is_MemBar() && n->as_MemBar()->leading()) {
3317 assert(n == leading, "consistency check failed");
3318 found++;
3319 } else {
3320 Node* in = n->in(0);
3321 if (in != NULL && !in->is_top()) {
3322 wq.push(in);
3323 }
3324 }
3325 }
3326 }
3327 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed");
3328 #endif
3329 if (leading == NULL) {
3330 return NULL;
3331 }
3332 MemBarNode* mb = leading->as_MemBar();
3333 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3334 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3335 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3336 return mb;
3337 }
3338
3339 //===========================InitializeNode====================================
3340 // SUMMARY:
3341 // This node acts as a memory barrier on raw memory, after some raw stores.
3342 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3343 // The Initialize can 'capture' suitably constrained stores as raw inits.
3344 // It can coalesce related raw stores into larger units (called 'tiles').
3345 // It can avoid zeroing new storage for memory units which have raw inits.
3346 // At macro-expansion, it is marked 'complete', and does not optimize further.
3347 //
3348 // EXAMPLE:
3349 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3350 // ctl = incoming control; mem* = incoming memory
3351 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3352 // First allocate uninitialized memory and fill in the header:
3353 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3354 // ctl := alloc.Control; mem* := alloc.Memory*
3355 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3356 // Then initialize to zero the non-header parts of the raw memory block:
3357 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3358 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3359 // After the initialize node executes, the object is ready for service:
3360 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3361 // Suppose its body is immediately initialized as {1,2}:
3362 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3363 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3364 // mem.SLICE(#short[*]) := store2
3365 //
3366 // DETAILS:
3367 // An InitializeNode collects and isolates object initialization after
3368 // an AllocateNode and before the next possible safepoint. As a
3369 // memory barrier (MemBarNode), it keeps critical stores from drifting
3370 // down past any safepoint or any publication of the allocation.
3371 // Before this barrier, a newly-allocated object may have uninitialized bits.
3372 // After this barrier, it may be treated as a real oop, and GC is allowed.
3373 //
3374 // The semantics of the InitializeNode include an implicit zeroing of
3375 // the new object from object header to the end of the object.
3376 // (The object header and end are determined by the AllocateNode.)
3377 //
3378 // Certain stores may be added as direct inputs to the InitializeNode.
3379 // These stores must update raw memory, and they must be to addresses
3380 // derived from the raw address produced by AllocateNode, and with
3381 // a constant offset. They must be ordered by increasing offset.
3382 // The first one is at in(RawStores), the last at in(req()-1).
3383 // Unlike most memory operations, they are not linked in a chain,
3384 // but are displayed in parallel as users of the rawmem output of
3385 // the allocation.
3386 //
3387 // (See comments in InitializeNode::capture_store, which continue
3388 // the example given above.)
3389 //
3390 // When the associated Allocate is macro-expanded, the InitializeNode
3391 // may be rewritten to optimize collected stores. A ClearArrayNode
3392 // may also be created at that point to represent any required zeroing.
3393 // The InitializeNode is then marked 'complete', prohibiting further
3394 // capturing of nearby memory operations.
3395 //
3396 // During macro-expansion, all captured initializations which store
3397 // constant values of 32 bits or smaller are coalesced (if advantageous)
3398 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3399 // initialized in fewer memory operations. Memory words which are
3400 // covered by neither tiles nor non-constant stores are pre-zeroed
3401 // by explicit stores of zero. (The code shape happens to do all
3402 // zeroing first, then all other stores, with both sequences occurring
3403 // in order of ascending offsets.)
3404 //
3405 // Alternatively, code may be inserted between an AllocateNode and its
3406 // InitializeNode, to perform arbitrary initialization of the new object.
3407 // E.g., the object copying intrinsics insert complex data transfers here.
3408 // The initialization must then be marked as 'complete' disable the
3409 // built-in zeroing semantics and the collection of initializing stores.
3410 //
3411 // While an InitializeNode is incomplete, reads from the memory state
3412 // produced by it are optimizable if they match the control edge and
3413 // new oop address associated with the allocation/initialization.
3414 // They return a stored value (if the offset matches) or else zero.
3415 // A write to the memory state, if it matches control and address,
3416 // and if it is to a constant offset, may be 'captured' by the
3417 // InitializeNode. It is cloned as a raw memory operation and rewired
3418 // inside the initialization, to the raw oop produced by the allocation.
3419 // Operations on addresses which are provably distinct (e.g., to
3420 // other AllocateNodes) are allowed to bypass the initialization.
3421 //
3422 // The effect of all this is to consolidate object initialization
3423 // (both arrays and non-arrays, both piecewise and bulk) into a
3424 // single location, where it can be optimized as a unit.
3425 //
3426 // Only stores with an offset less than TrackedInitializationLimit words
3427 // will be considered for capture by an InitializeNode. This puts a
3428 // reasonable limit on the complexity of optimized initializations.
3429
3430 //---------------------------InitializeNode------------------------------------
3431 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3432 : _is_complete(Incomplete), _does_not_escape(false),
3433 MemBarNode(C, adr_type, rawoop)
3434 {
3435 init_class_id(Class_Initialize);
3436
3437 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3438 assert(in(RawAddress) == rawoop, "proper init");
3439 // Note: allocation() can be NULL, for secondary initialization barriers
3440 }
3441
3442 // Since this node is not matched, it will be processed by the
3443 // register allocator. Declare that there are no constraints
3444 // on the allocation of the RawAddress edge.
3445 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3446 // This edge should be set to top, by the set_complete. But be conservative.
3447 if (idx == InitializeNode::RawAddress)
3448 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3449 return RegMask::Empty;
3450 }
3451
3452 Node* InitializeNode::memory(uint alias_idx) {
3453 Node* mem = in(Memory);
3454 if (mem->is_MergeMem()) {
3455 return mem->as_MergeMem()->memory_at(alias_idx);
3456 } else {
3457 // incoming raw memory is not split
3458 return mem;
3459 }
3460 }
3461
3462 bool InitializeNode::is_non_zero() {
3463 if (is_complete()) return false;
3464 remove_extra_zeroes();
3465 return (req() > RawStores);
3466 }
3467
3468 void InitializeNode::set_complete(PhaseGVN* phase) {
3469 assert(!is_complete(), "caller responsibility");
3470 _is_complete = Complete;
3471
3472 // After this node is complete, it contains a bunch of
3473 // raw-memory initializations. There is no need for
3474 // it to have anything to do with non-raw memory effects.
3475 // Therefore, tell all non-raw users to re-optimize themselves,
3476 // after skipping the memory effects of this initialization.
3477 PhaseIterGVN* igvn = phase->is_IterGVN();
3478 if (igvn) igvn->add_users_to_worklist(this);
3479 }
3480
3481 // convenience function
3482 // return false if the init contains any stores already
3483 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3484 InitializeNode* init = initialization();
3485 if (init == NULL || init->is_complete()) return false;
3486 init->remove_extra_zeroes();
3487 // for now, if this allocation has already collected any inits, bail:
3488 if (init->is_non_zero()) return false;
3489 init->set_complete(phase);
3490 return true;
3491 }
3492
3493 void InitializeNode::remove_extra_zeroes() {
3494 if (req() == RawStores) return;
3495 Node* zmem = zero_memory();
3496 uint fill = RawStores;
3497 for (uint i = fill; i < req(); i++) {
3498 Node* n = in(i);
3499 if (n->is_top() || n == zmem) continue; // skip
3500 if (fill < i) set_req(fill, n); // compact
3501 ++fill;
3502 }
3503 // delete any empty spaces created:
3504 while (fill < req()) {
3505 del_req(fill);
3506 }
3507 }
3508
3509 // Helper for remembering which stores go with which offsets.
3510 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3511 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3512 intptr_t offset = -1;
3513 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3514 phase, offset);
3515 if (base == NULL) return -1; // something is dead,
3516 if (offset < 0) return -1; // dead, dead
3517 return offset;
3518 }
3519
3520 // Helper for proving that an initialization expression is
3521 // "simple enough" to be folded into an object initialization.
3522 // Attempts to prove that a store's initial value 'n' can be captured
3523 // within the initialization without creating a vicious cycle, such as:
3524 // { Foo p = new Foo(); p.next = p; }
3525 // True for constants and parameters and small combinations thereof.
3526 bool InitializeNode::detect_init_independence(Node* n, int& count) {
3527 if (n == NULL) return true; // (can this really happen?)
3528 if (n->is_Proj()) n = n->in(0);
3529 if (n == this) return false; // found a cycle
3530 if (n->is_Con()) return true;
3531 if (n->is_Start()) return true; // params, etc., are OK
3532 if (n->is_Root()) return true; // even better
3533
3534 Node* ctl = n->in(0);
3535 if (ctl != NULL && !ctl->is_top()) {
3536 if (ctl->is_Proj()) ctl = ctl->in(0);
3537 if (ctl == this) return false;
3538
3539 // If we already know that the enclosing memory op is pinned right after
3540 // the init, then any control flow that the store has picked up
3541 // must have preceded the init, or else be equal to the init.
3542 // Even after loop optimizations (which might change control edges)
3543 // a store is never pinned *before* the availability of its inputs.
3544 if (!MemNode::all_controls_dominate(n, this))
3545 return false; // failed to prove a good control
3546 }
3547
3548 // Check data edges for possible dependencies on 'this'.
3549 if ((count += 1) > 20) return false; // complexity limit
3550 for (uint i = 1; i < n->req(); i++) {
3551 Node* m = n->in(i);
3552 if (m == NULL || m == n || m->is_top()) continue;
3553 uint first_i = n->find_edge(m);
3554 if (i != first_i) continue; // process duplicate edge just once
3555 if (!detect_init_independence(m, count)) {
3556 return false;
3557 }
3558 }
3559
3560 return true;
3561 }
3562
3563 // Here are all the checks a Store must pass before it can be moved into
3564 // an initialization. Returns zero if a check fails.
3565 // On success, returns the (constant) offset to which the store applies,
3566 // within the initialized memory.
3567 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseTransform* phase, bool can_reshape) {
3568 const int FAIL = 0;
3569 if (st->req() != MemNode::ValueIn + 1)
3570 return FAIL; // an inscrutable StoreNode (card mark?)
3571 Node* ctl = st->in(MemNode::Control);
3572 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3573 return FAIL; // must be unconditional after the initialization
3574 Node* mem = st->in(MemNode::Memory);
3575 if (!(mem->is_Proj() && mem->in(0) == this))
3576 return FAIL; // must not be preceded by other stores
3577 Node* adr = st->in(MemNode::Address);
3578 intptr_t offset;
3579 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3580 if (alloc == NULL)
3581 return FAIL; // inscrutable address
3582 if (alloc != allocation())
3583 return FAIL; // wrong allocation! (store needs to float up)
3584 int size_in_bytes = st->memory_size();
3585 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
3586 return FAIL; // mismatched access
3587 }
3588 Node* val = st->in(MemNode::ValueIn);
3589 int complexity_count = 0;
3590 if (!detect_init_independence(val, complexity_count))
3591 return FAIL; // stored value must be 'simple enough'
3592
3593 // The Store can be captured only if nothing after the allocation
3594 // and before the Store is using the memory location that the store
3595 // overwrites.
3596 bool failed = false;
3597 // If is_complete_with_arraycopy() is true the shape of the graph is
3598 // well defined and is safe so no need for extra checks.
3599 if (!is_complete_with_arraycopy()) {
3600 // We are going to look at each use of the memory state following
3601 // the allocation to make sure nothing reads the memory that the
3602 // Store writes.
3603 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3604 int alias_idx = phase->C->get_alias_index(t_adr);
3605 ResourceMark rm;
3606 Unique_Node_List mems;
3607 mems.push(mem);
3608 Node* unique_merge = NULL;
3609 for (uint next = 0; next < mems.size(); ++next) {
3610 Node *m = mems.at(next);
3611 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3612 Node *n = m->fast_out(j);
3613 if (n->outcnt() == 0) {
3614 continue;
3615 }
3616 if (n == st) {
3617 continue;
3618 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3619 // If the control of this use is different from the control
3620 // of the Store which is right after the InitializeNode then
3621 // this node cannot be between the InitializeNode and the
3622 // Store.
3623 continue;
3624 } else if (n->is_MergeMem()) {
3625 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3626 // We can hit a MergeMemNode (that will likely go away
3627 // later) that is a direct use of the memory state
3628 // following the InitializeNode on the same slice as the
3629 // store node that we'd like to capture. We need to check
3630 // the uses of the MergeMemNode.
3631 mems.push(n);
3632 }
3633 } else if (n->is_Mem()) {
3634 Node* other_adr = n->in(MemNode::Address);
3635 if (other_adr == adr) {
3636 failed = true;
3637 break;
3638 } else {
3639 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3640 if (other_t_adr != NULL) {
3641 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3642 if (other_alias_idx == alias_idx) {
3643 // A load from the same memory slice as the store right
3644 // after the InitializeNode. We check the control of the
3645 // object/array that is loaded from. If it's the same as
3646 // the store control then we cannot capture the store.
3647 assert(!n->is_Store(), "2 stores to same slice on same control?");
3648 Node* base = other_adr;
3649 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3650 base = base->in(AddPNode::Base);
3651 if (base != NULL) {
3652 base = base->uncast();
3653 if (base->is_Proj() && base->in(0) == alloc) {
3654 failed = true;
3655 break;
3656 }
3657 }
3658 }
3659 }
3660 }
3661 } else {
3662 failed = true;
3663 break;
3664 }
3665 }
3666 }
3667 }
3668 if (failed) {
3669 if (!can_reshape) {
3670 // We decided we couldn't capture the store during parsing. We
3671 // should try again during the next IGVN once the graph is
3672 // cleaner.
3673 phase->C->record_for_igvn(st);
3674 }
3675 return FAIL;
3676 }
3677
3678 return offset; // success
3679 }
3680
3681 // Find the captured store in(i) which corresponds to the range
3682 // [start..start+size) in the initialized object.
3683 // If there is one, return its index i. If there isn't, return the
3684 // negative of the index where it should be inserted.
3685 // Return 0 if the queried range overlaps an initialization boundary
3686 // or if dead code is encountered.
3687 // If size_in_bytes is zero, do not bother with overlap checks.
3688 int InitializeNode::captured_store_insertion_point(intptr_t start,
3689 int size_in_bytes,
3690 PhaseTransform* phase) {
3691 const int FAIL = 0, MAX_STORE = BytesPerLong;
3692
3693 if (is_complete())
3694 return FAIL; // arraycopy got here first; punt
3695
3696 assert(allocation() != NULL, "must be present");
3697
3698 // no negatives, no header fields:
3699 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3700
3701 // after a certain size, we bail out on tracking all the stores:
3702 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3703 if (start >= ti_limit) return FAIL;
3704
3705 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3706 if (i >= limit) return -(int)i; // not found; here is where to put it
3707
3708 Node* st = in(i);
3709 intptr_t st_off = get_store_offset(st, phase);
3710 if (st_off < 0) {
3711 if (st != zero_memory()) {
3712 return FAIL; // bail out if there is dead garbage
3713 }
3714 } else if (st_off > start) {
3715 // ...we are done, since stores are ordered
3716 if (st_off < start + size_in_bytes) {
3717 return FAIL; // the next store overlaps
3718 }
3719 return -(int)i; // not found; here is where to put it
3720 } else if (st_off < start) {
3721 if (size_in_bytes != 0 &&
3722 start < st_off + MAX_STORE &&
3723 start < st_off + st->as_Store()->memory_size()) {
3724 return FAIL; // the previous store overlaps
3725 }
3726 } else {
3727 if (size_in_bytes != 0 &&
3728 st->as_Store()->memory_size() != size_in_bytes) {
3729 return FAIL; // mismatched store size
3730 }
3731 return i;
3732 }
3733
3734 ++i;
3735 }
3736 }
3737
3738 // Look for a captured store which initializes at the offset 'start'
3739 // with the given size. If there is no such store, and no other
3740 // initialization interferes, then return zero_memory (the memory
3741 // projection of the AllocateNode).
3742 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3743 PhaseTransform* phase) {
3744 assert(stores_are_sane(phase), "");
3745 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3746 if (i == 0) {
3747 return NULL; // something is dead
3748 } else if (i < 0) {
3749 return zero_memory(); // just primordial zero bits here
3750 } else {
3751 Node* st = in(i); // here is the store at this position
3752 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3753 return st;
3754 }
3755 }
3756
3757 // Create, as a raw pointer, an address within my new object at 'offset'.
3758 Node* InitializeNode::make_raw_address(intptr_t offset,
3759 PhaseTransform* phase) {
3760 Node* addr = in(RawAddress);
3761 if (offset != 0) {
3762 Compile* C = phase->C;
3763 addr = phase->transform( new AddPNode(C->top(), addr,
3764 phase->MakeConX(offset)) );
3765 }
3766 return addr;
3767 }
3768
3769 // Clone the given store, converting it into a raw store
3770 // initializing a field or element of my new object.
3771 // Caller is responsible for retiring the original store,
3772 // with subsume_node or the like.
3773 //
3774 // From the example above InitializeNode::InitializeNode,
3775 // here are the old stores to be captured:
3776 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3777 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3778 //
3779 // Here is the changed code; note the extra edges on init:
3780 // alloc = (Allocate ...)
3781 // rawoop = alloc.RawAddress
3782 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3783 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3784 // init = (Initialize alloc.Control alloc.Memory rawoop
3785 // rawstore1 rawstore2)
3786 //
3787 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3788 PhaseTransform* phase, bool can_reshape) {
3789 assert(stores_are_sane(phase), "");
3790
3791 if (start < 0) return NULL;
3792 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3793
3794 Compile* C = phase->C;
3795 int size_in_bytes = st->memory_size();
3796 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3797 if (i == 0) return NULL; // bail out
3798 Node* prev_mem = NULL; // raw memory for the captured store
3799 if (i > 0) {
3800 prev_mem = in(i); // there is a pre-existing store under this one
3801 set_req(i, C->top()); // temporarily disconnect it
3802 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3803 } else {
3804 i = -i; // no pre-existing store
3805 prev_mem = zero_memory(); // a slice of the newly allocated object
3806 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3807 set_req(--i, C->top()); // reuse this edge; it has been folded away
3808 else
3809 ins_req(i, C->top()); // build a new edge
3810 }
3811 Node* new_st = st->clone();
3812 new_st->set_req(MemNode::Control, in(Control));
3813 new_st->set_req(MemNode::Memory, prev_mem);
3814 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3815 new_st = phase->transform(new_st);
3816
3817 // At this point, new_st might have swallowed a pre-existing store
3818 // at the same offset, or perhaps new_st might have disappeared,
3819 // if it redundantly stored the same value (or zero to fresh memory).
3820
3821 // In any case, wire it in:
3822 phase->igvn_rehash_node_delayed(this);
3823 set_req(i, new_st);
3824
3825 // The caller may now kill the old guy.
3826 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3827 assert(check_st == new_st || check_st == NULL, "must be findable");
3828 assert(!is_complete(), "");
3829 return new_st;
3830 }
3831
3832 static bool store_constant(jlong* tiles, int num_tiles,
3833 intptr_t st_off, int st_size,
3834 jlong con) {
3835 if ((st_off & (st_size-1)) != 0)
3836 return false; // strange store offset (assume size==2**N)
3837 address addr = (address)tiles + st_off;
3838 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
3839 switch (st_size) {
3840 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
3841 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
3842 case sizeof(jint): *(jint*) addr = (jint) con; break;
3843 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
3844 default: return false; // strange store size (detect size!=2**N here)
3845 }
3846 return true; // return success to caller
3847 }
3848
3849 // Coalesce subword constants into int constants and possibly
3850 // into long constants. The goal, if the CPU permits,
3851 // is to initialize the object with a small number of 64-bit tiles.
3852 // Also, convert floating-point constants to bit patterns.
3853 // Non-constants are not relevant to this pass.
3854 //
3855 // In terms of the running example on InitializeNode::InitializeNode
3856 // and InitializeNode::capture_store, here is the transformation
3857 // of rawstore1 and rawstore2 into rawstore12:
3858 // alloc = (Allocate ...)
3859 // rawoop = alloc.RawAddress
3860 // tile12 = 0x00010002
3861 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
3862 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
3863 //
3864 void
3865 InitializeNode::coalesce_subword_stores(intptr_t header_size,
3866 Node* size_in_bytes,
3867 PhaseGVN* phase) {
3868 Compile* C = phase->C;
3869
3870 assert(stores_are_sane(phase), "");
3871 // Note: After this pass, they are not completely sane,
3872 // since there may be some overlaps.
3873
3874 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
3875
3876 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3877 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
3878 size_limit = MIN2(size_limit, ti_limit);
3879 size_limit = align_up(size_limit, BytesPerLong);
3880 int num_tiles = size_limit / BytesPerLong;
3881
3882 // allocate space for the tile map:
3883 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
3884 jlong tiles_buf[small_len];
3885 Node* nodes_buf[small_len];
3886 jlong inits_buf[small_len];
3887 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
3888 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3889 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
3890 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
3891 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
3892 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
3893 // tiles: exact bitwise model of all primitive constants
3894 // nodes: last constant-storing node subsumed into the tiles model
3895 // inits: which bytes (in each tile) are touched by any initializations
3896
3897 //// Pass A: Fill in the tile model with any relevant stores.
3898
3899 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
3900 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
3901 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
3902 Node* zmem = zero_memory(); // initially zero memory state
3903 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
3904 Node* st = in(i);
3905 intptr_t st_off = get_store_offset(st, phase);
3906
3907 // Figure out the store's offset and constant value:
3908 if (st_off < header_size) continue; //skip (ignore header)
3909 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
3910 int st_size = st->as_Store()->memory_size();
3911 if (st_off + st_size > size_limit) break;
3912
3913 // Record which bytes are touched, whether by constant or not.
3914 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
3915 continue; // skip (strange store size)
3916
3917 const Type* val = phase->type(st->in(MemNode::ValueIn));
3918 if (!val->singleton()) continue; //skip (non-con store)
3919 BasicType type = val->basic_type();
3920
3921 jlong con = 0;
3922 switch (type) {
3923 case T_INT: con = val->is_int()->get_con(); break;
3924 case T_LONG: con = val->is_long()->get_con(); break;
3925 case T_FLOAT: con = jint_cast(val->getf()); break;
3926 case T_DOUBLE: con = jlong_cast(val->getd()); break;
3927 default: continue; //skip (odd store type)
3928 }
3929
3930 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
3931 st->Opcode() == Op_StoreL) {
3932 continue; // This StoreL is already optimal.
3933 }
3934
3935 // Store down the constant.
3936 store_constant(tiles, num_tiles, st_off, st_size, con);
3937
3938 intptr_t j = st_off >> LogBytesPerLong;
3939
3940 if (type == T_INT && st_size == BytesPerInt
3941 && (st_off & BytesPerInt) == BytesPerInt) {
3942 jlong lcon = tiles[j];
3943 if (!Matcher::isSimpleConstant64(lcon) &&
3944 st->Opcode() == Op_StoreI) {
3945 // This StoreI is already optimal by itself.
3946 jint* intcon = (jint*) &tiles[j];
3947 intcon[1] = 0; // undo the store_constant()
3948
3949 // If the previous store is also optimal by itself, back up and
3950 // undo the action of the previous loop iteration... if we can.
3951 // But if we can't, just let the previous half take care of itself.
3952 st = nodes[j];
3953 st_off -= BytesPerInt;
3954 con = intcon[0];
3955 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
3956 assert(st_off >= header_size, "still ignoring header");
3957 assert(get_store_offset(st, phase) == st_off, "must be");
3958 assert(in(i-1) == zmem, "must be");
3959 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
3960 assert(con == tcon->is_int()->get_con(), "must be");
3961 // Undo the effects of the previous loop trip, which swallowed st:
3962 intcon[0] = 0; // undo store_constant()
3963 set_req(i-1, st); // undo set_req(i, zmem)
3964 nodes[j] = NULL; // undo nodes[j] = st
3965 --old_subword; // undo ++old_subword
3966 }
3967 continue; // This StoreI is already optimal.
3968 }
3969 }
3970
3971 // This store is not needed.
3972 set_req(i, zmem);
3973 nodes[j] = st; // record for the moment
3974 if (st_size < BytesPerLong) // something has changed
3975 ++old_subword; // includes int/float, but who's counting...
3976 else ++old_long;
3977 }
3978
3979 if ((old_subword + old_long) == 0)
3980 return; // nothing more to do
3981
3982 //// Pass B: Convert any non-zero tiles into optimal constant stores.
3983 // Be sure to insert them before overlapping non-constant stores.
3984 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
3985 for (int j = 0; j < num_tiles; j++) {
3986 jlong con = tiles[j];
3987 jlong init = inits[j];
3988 if (con == 0) continue;
3989 jint con0, con1; // split the constant, address-wise
3990 jint init0, init1; // split the init map, address-wise
3991 { union { jlong con; jint intcon[2]; } u;
3992 u.con = con;
3993 con0 = u.intcon[0];
3994 con1 = u.intcon[1];
3995 u.con = init;
3996 init0 = u.intcon[0];
3997 init1 = u.intcon[1];
3998 }
3999
4000 Node* old = nodes[j];
4001 assert(old != NULL, "need the prior store");
4002 intptr_t offset = (j * BytesPerLong);
4003
4004 bool split = !Matcher::isSimpleConstant64(con);
4005
4006 if (offset < header_size) {
4007 assert(offset + BytesPerInt >= header_size, "second int counts");
4008 assert(*(jint*)&tiles[j] == 0, "junk in header");
4009 split = true; // only the second word counts
4010 // Example: int a[] = { 42 ... }
4011 } else if (con0 == 0 && init0 == -1) {
4012 split = true; // first word is covered by full inits
4013 // Example: int a[] = { ... foo(), 42 ... }
4014 } else if (con1 == 0 && init1 == -1) {
4015 split = true; // second word is covered by full inits
4016 // Example: int a[] = { ... 42, foo() ... }
4017 }
4018
4019 // Here's a case where init0 is neither 0 nor -1:
4020 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
4021 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
4022 // In this case the tile is not split; it is (jlong)42.
4023 // The big tile is stored down, and then the foo() value is inserted.
4024 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
4025
4026 Node* ctl = old->in(MemNode::Control);
4027 Node* adr = make_raw_address(offset, phase);
4028 const TypePtr* atp = TypeRawPtr::BOTTOM;
4029
4030 // One or two coalesced stores to plop down.
4031 Node* st[2];
4032 intptr_t off[2];
4033 int nst = 0;
4034 if (!split) {
4035 ++new_long;
4036 off[nst] = offset;
4037 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4038 phase->longcon(con), T_LONG, MemNode::unordered);
4039 } else {
4040 // Omit either if it is a zero.
4041 if (con0 != 0) {
4042 ++new_int;
4043 off[nst] = offset;
4044 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4045 phase->intcon(con0), T_INT, MemNode::unordered);
4046 }
4047 if (con1 != 0) {
4048 ++new_int;
4049 offset += BytesPerInt;
4050 adr = make_raw_address(offset, phase);
4051 off[nst] = offset;
4052 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4053 phase->intcon(con1), T_INT, MemNode::unordered);
4054 }
4055 }
4056
4057 // Insert second store first, then the first before the second.
4058 // Insert each one just before any overlapping non-constant stores.
4059 while (nst > 0) {
4060 Node* st1 = st[--nst];
4061 C->copy_node_notes_to(st1, old);
4062 st1 = phase->transform(st1);
4063 offset = off[nst];
4064 assert(offset >= header_size, "do not smash header");
4065 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4066 guarantee(ins_idx != 0, "must re-insert constant store");
4067 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
4068 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4069 set_req(--ins_idx, st1);
4070 else
4071 ins_req(ins_idx, st1);
4072 }
4073 }
4074
4075 if (PrintCompilation && WizardMode)
4076 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4077 old_subword, old_long, new_int, new_long);
4078 if (C->log() != NULL)
4079 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4080 old_subword, old_long, new_int, new_long);
4081
4082 // Clean up any remaining occurrences of zmem:
4083 remove_extra_zeroes();
4084 }
4085
4086 // Explore forward from in(start) to find the first fully initialized
4087 // word, and return its offset. Skip groups of subword stores which
4088 // together initialize full words. If in(start) is itself part of a
4089 // fully initialized word, return the offset of in(start). If there
4090 // are no following full-word stores, or if something is fishy, return
4091 // a negative value.
4092 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4093 int int_map = 0;
4094 intptr_t int_map_off = 0;
4095 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
4096
4097 for (uint i = start, limit = req(); i < limit; i++) {
4098 Node* st = in(i);
4099
4100 intptr_t st_off = get_store_offset(st, phase);
4101 if (st_off < 0) break; // return conservative answer
4102
4103 int st_size = st->as_Store()->memory_size();
4104 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4105 return st_off; // we found a complete word init
4106 }
4107
4108 // update the map:
4109
4110 intptr_t this_int_off = align_down(st_off, BytesPerInt);
4111 if (this_int_off != int_map_off) {
4112 // reset the map:
4113 int_map = 0;
4114 int_map_off = this_int_off;
4115 }
4116
4117 int subword_off = st_off - this_int_off;
4118 int_map |= right_n_bits(st_size) << subword_off;
4119 if ((int_map & FULL_MAP) == FULL_MAP) {
4120 return this_int_off; // we found a complete word init
4121 }
4122
4123 // Did this store hit or cross the word boundary?
4124 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4125 if (next_int_off == this_int_off + BytesPerInt) {
4126 // We passed the current int, without fully initializing it.
4127 int_map_off = next_int_off;
4128 int_map >>= BytesPerInt;
4129 } else if (next_int_off > this_int_off + BytesPerInt) {
4130 // We passed the current and next int.
4131 return this_int_off + BytesPerInt;
4132 }
4133 }
4134
4135 return -1;
4136 }
4137
4138
4139 // Called when the associated AllocateNode is expanded into CFG.
4140 // At this point, we may perform additional optimizations.
4141 // Linearize the stores by ascending offset, to make memory
4142 // activity as coherent as possible.
4143 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4144 intptr_t header_size,
4145 Node* size_in_bytes,
4146 PhaseGVN* phase) {
4147 assert(!is_complete(), "not already complete");
4148 assert(stores_are_sane(phase), "");
4149 assert(allocation() != NULL, "must be present");
4150
4151 remove_extra_zeroes();
4152
4153 if (ReduceFieldZeroing || ReduceBulkZeroing)
4154 // reduce instruction count for common initialization patterns
4155 coalesce_subword_stores(header_size, size_in_bytes, phase);
4156
4157 Node* zmem = zero_memory(); // initially zero memory state
4158 Node* inits = zmem; // accumulating a linearized chain of inits
4159 #ifdef ASSERT
4160 intptr_t first_offset = allocation()->minimum_header_size();
4161 intptr_t last_init_off = first_offset; // previous init offset
4162 intptr_t last_init_end = first_offset; // previous init offset+size
4163 intptr_t last_tile_end = first_offset; // previous tile offset+size
4164 #endif
4165 intptr_t zeroes_done = header_size;
4166
4167 bool do_zeroing = true; // we might give up if inits are very sparse
4168 int big_init_gaps = 0; // how many large gaps have we seen?
4169
4170 if (UseTLAB && ZeroTLAB) do_zeroing = false;
4171 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
4172
4173 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4174 Node* st = in(i);
4175 intptr_t st_off = get_store_offset(st, phase);
4176 if (st_off < 0)
4177 break; // unknown junk in the inits
4178 if (st->in(MemNode::Memory) != zmem)
4179 break; // complicated store chains somehow in list
4180
4181 int st_size = st->as_Store()->memory_size();
4182 intptr_t next_init_off = st_off + st_size;
4183
4184 if (do_zeroing && zeroes_done < next_init_off) {
4185 // See if this store needs a zero before it or under it.
4186 intptr_t zeroes_needed = st_off;
4187
4188 if (st_size < BytesPerInt) {
4189 // Look for subword stores which only partially initialize words.
4190 // If we find some, we must lay down some word-level zeroes first,
4191 // underneath the subword stores.
4192 //
4193 // Examples:
4194 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
4195 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
4196 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
4197 //
4198 // Note: coalesce_subword_stores may have already done this,
4199 // if it was prompted by constant non-zero subword initializers.
4200 // But this case can still arise with non-constant stores.
4201
4202 intptr_t next_full_store = find_next_fullword_store(i, phase);
4203
4204 // In the examples above:
4205 // in(i) p q r s x y z
4206 // st_off 12 13 14 15 12 13 14
4207 // st_size 1 1 1 1 1 1 1
4208 // next_full_s. 12 16 16 16 16 16 16
4209 // z's_done 12 16 16 16 12 16 12
4210 // z's_needed 12 16 16 16 16 16 16
4211 // zsize 0 0 0 0 4 0 4
4212 if (next_full_store < 0) {
4213 // Conservative tack: Zero to end of current word.
4214 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4215 } else {
4216 // Zero to beginning of next fully initialized word.
4217 // Or, don't zero at all, if we are already in that word.
4218 assert(next_full_store >= zeroes_needed, "must go forward");
4219 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4220 zeroes_needed = next_full_store;
4221 }
4222 }
4223
4224 if (zeroes_needed > zeroes_done) {
4225 intptr_t zsize = zeroes_needed - zeroes_done;
4226 // Do some incremental zeroing on rawmem, in parallel with inits.
4227 zeroes_done = align_down(zeroes_done, BytesPerInt);
4228 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4229 zeroes_done, zeroes_needed,
4230 phase);
4231 zeroes_done = zeroes_needed;
4232 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4233 do_zeroing = false; // leave the hole, next time
4234 }
4235 }
4236
4237 // Collect the store and move on:
4238 st->set_req(MemNode::Memory, inits);
4239 inits = st; // put it on the linearized chain
4240 set_req(i, zmem); // unhook from previous position
4241
4242 if (zeroes_done == st_off)
4243 zeroes_done = next_init_off;
4244
4245 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4246
4247 #ifdef ASSERT
4248 // Various order invariants. Weaker than stores_are_sane because
4249 // a large constant tile can be filled in by smaller non-constant stores.
4250 assert(st_off >= last_init_off, "inits do not reverse");
4251 last_init_off = st_off;
4252 const Type* val = NULL;
4253 if (st_size >= BytesPerInt &&
4254 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4255 (int)val->basic_type() < (int)T_OBJECT) {
4256 assert(st_off >= last_tile_end, "tiles do not overlap");
4257 assert(st_off >= last_init_end, "tiles do not overwrite inits");
4258 last_tile_end = MAX2(last_tile_end, next_init_off);
4259 } else {
4260 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4261 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4262 assert(st_off >= last_init_end, "inits do not overlap");
4263 last_init_end = next_init_off; // it's a non-tile
4264 }
4265 #endif //ASSERT
4266 }
4267
4268 remove_extra_zeroes(); // clear out all the zmems left over
4269 add_req(inits);
4270
4271 if (!(UseTLAB && ZeroTLAB)) {
4272 // If anything remains to be zeroed, zero it all now.
4273 zeroes_done = align_down(zeroes_done, BytesPerInt);
4274 // if it is the last unused 4 bytes of an instance, forget about it
4275 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4276 if (zeroes_done + BytesPerLong >= size_limit) {
4277 AllocateNode* alloc = allocation();
4278 assert(alloc != NULL, "must be present");
4279 if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
4280 Node* klass_node = alloc->in(AllocateNode::KlassNode);
4281 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4282 if (zeroes_done == k->layout_helper())
4283 zeroes_done = size_limit;
4284 }
4285 }
4286 if (zeroes_done < size_limit) {
4287 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4288 zeroes_done, size_in_bytes, phase);
4289 }
4290 }
4291
4292 set_complete(phase);
4293 return rawmem;
4294 }
4295
4296
4297 #ifdef ASSERT
4298 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4299 if (is_complete())
4300 return true; // stores could be anything at this point
4301 assert(allocation() != NULL, "must be present");
4302 intptr_t last_off = allocation()->minimum_header_size();
4303 for (uint i = InitializeNode::RawStores; i < req(); i++) {
4304 Node* st = in(i);
4305 intptr_t st_off = get_store_offset(st, phase);
4306 if (st_off < 0) continue; // ignore dead garbage
4307 if (last_off > st_off) {
4308 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4309 this->dump(2);
4310 assert(false, "ascending store offsets");
4311 return false;
4312 }
4313 last_off = st_off + st->as_Store()->memory_size();
4314 }
4315 return true;
4316 }
4317 #endif //ASSERT
4318
4319
4320
4321
4322 //============================MergeMemNode=====================================
4323 //
4324 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4325 // contributing store or call operations. Each contributor provides the memory
4326 // state for a particular "alias type" (see Compile::alias_type). For example,
4327 // if a MergeMem has an input X for alias category #6, then any memory reference
4328 // to alias category #6 may use X as its memory state input, as an exact equivalent
4329 // to using the MergeMem as a whole.
4330 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4331 //
4332 // (Here, the <N> notation gives the index of the relevant adr_type.)
4333 //
4334 // In one special case (and more cases in the future), alias categories overlap.
4335 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4336 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4337 // it is exactly equivalent to that state W:
4338 // MergeMem(<Bot>: W) <==> W
4339 //
4340 // Usually, the merge has more than one input. In that case, where inputs
4341 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4342 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4343 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4344 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4345 //
4346 // A merge can take a "wide" memory state as one of its narrow inputs.
4347 // This simply means that the merge observes out only the relevant parts of
4348 // the wide input. That is, wide memory states arriving at narrow merge inputs
4349 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4350 //
4351 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4352 // and that memory slices "leak through":
4353 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4354 //
4355 // But, in such a cascade, repeated memory slices can "block the leak":
4356 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4357 //
4358 // In the last example, Y is not part of the combined memory state of the
4359 // outermost MergeMem. The system must, of course, prevent unschedulable
4360 // memory states from arising, so you can be sure that the state Y is somehow
4361 // a precursor to state Y'.
4362 //
4363 //
4364 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4365 // of each MergeMemNode array are exactly the numerical alias indexes, including
4366 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4367 // Compile::alias_type (and kin) produce and manage these indexes.
4368 //
4369 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4370 // (Note that this provides quick access to the top node inside MergeMem methods,
4371 // without the need to reach out via TLS to Compile::current.)
4372 //
4373 // As a consequence of what was just described, a MergeMem that represents a full
4374 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4375 // containing all alias categories.
4376 //
4377 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4378 //
4379 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4380 // a memory state for the alias type <N>, or else the top node, meaning that
4381 // there is no particular input for that alias type. Note that the length of
4382 // a MergeMem is variable, and may be extended at any time to accommodate new
4383 // memory states at larger alias indexes. When merges grow, they are of course
4384 // filled with "top" in the unused in() positions.
4385 //
4386 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4387 // (Top was chosen because it works smoothly with passes like GCM.)
4388 //
4389 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4390 // the type of random VM bits like TLS references.) Since it is always the
4391 // first non-Bot memory slice, some low-level loops use it to initialize an
4392 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4393 //
4394 //
4395 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4396 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4397 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4398 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4399 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4400 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4401 //
4402 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4403 // really that different from the other memory inputs. An abbreviation called
4404 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4405 //
4406 //
4407 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4408 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4409 // that "emerges though" the base memory will be marked as excluding the alias types
4410 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4411 //
4412 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4413 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4414 //
4415 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4416 // (It is currently unimplemented.) As you can see, the resulting merge is
4417 // actually a disjoint union of memory states, rather than an overlay.
4418 //
4419
4420 //------------------------------MergeMemNode-----------------------------------
4421 Node* MergeMemNode::make_empty_memory() {
4422 Node* empty_memory = (Node*) Compile::current()->top();
4423 assert(empty_memory->is_top(), "correct sentinel identity");
4424 return empty_memory;
4425 }
4426
4427 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4428 init_class_id(Class_MergeMem);
4429 // all inputs are nullified in Node::Node(int)
4430 // set_input(0, NULL); // no control input
4431
4432 // Initialize the edges uniformly to top, for starters.
4433 Node* empty_mem = make_empty_memory();
4434 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4435 init_req(i,empty_mem);
4436 }
4437 assert(empty_memory() == empty_mem, "");
4438
4439 if( new_base != NULL && new_base->is_MergeMem() ) {
4440 MergeMemNode* mdef = new_base->as_MergeMem();
4441 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4442 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4443 mms.set_memory(mms.memory2());
4444 }
4445 assert(base_memory() == mdef->base_memory(), "");
4446 } else {
4447 set_base_memory(new_base);
4448 }
4449 }
4450
4451 // Make a new, untransformed MergeMem with the same base as 'mem'.
4452 // If mem is itself a MergeMem, populate the result with the same edges.
4453 MergeMemNode* MergeMemNode::make(Node* mem) {
4454 return new MergeMemNode(mem);
4455 }
4456
4457 //------------------------------cmp--------------------------------------------
4458 uint MergeMemNode::hash() const { return NO_HASH; }
4459 uint MergeMemNode::cmp( const Node &n ) const {
4460 return (&n == this); // Always fail except on self
4461 }
4462
4463 //------------------------------Identity---------------------------------------
4464 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4465 // Identity if this merge point does not record any interesting memory
4466 // disambiguations.
4467 Node* base_mem = base_memory();
4468 Node* empty_mem = empty_memory();
4469 if (base_mem != empty_mem) { // Memory path is not dead?
4470 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4471 Node* mem = in(i);
4472 if (mem != empty_mem && mem != base_mem) {
4473 return this; // Many memory splits; no change
4474 }
4475 }
4476 }
4477 return base_mem; // No memory splits; ID on the one true input
4478 }
4479
4480 //------------------------------Ideal------------------------------------------
4481 // This method is invoked recursively on chains of MergeMem nodes
4482 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4483 // Remove chain'd MergeMems
4484 //
4485 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4486 // relative to the "in(Bot)". Since we are patching both at the same time,
4487 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4488 // but rewrite each "in(i)" relative to the new "in(Bot)".
4489 Node *progress = NULL;
4490
4491
4492 Node* old_base = base_memory();
4493 Node* empty_mem = empty_memory();
4494 if (old_base == empty_mem)
4495 return NULL; // Dead memory path.
4496
4497 MergeMemNode* old_mbase;
4498 if (old_base != NULL && old_base->is_MergeMem())
4499 old_mbase = old_base->as_MergeMem();
4500 else
4501 old_mbase = NULL;
4502 Node* new_base = old_base;
4503
4504 // simplify stacked MergeMems in base memory
4505 if (old_mbase) new_base = old_mbase->base_memory();
4506
4507 // the base memory might contribute new slices beyond my req()
4508 if (old_mbase) grow_to_match(old_mbase);
4509
4510 // Look carefully at the base node if it is a phi.
4511 PhiNode* phi_base;
4512 if (new_base != NULL && new_base->is_Phi())
4513 phi_base = new_base->as_Phi();
4514 else
4515 phi_base = NULL;
4516
4517 Node* phi_reg = NULL;
4518 uint phi_len = (uint)-1;
4519 if (phi_base != NULL && !phi_base->is_copy()) {
4520 // do not examine phi if degraded to a copy
4521 phi_reg = phi_base->region();
4522 phi_len = phi_base->req();
4523 // see if the phi is unfinished
4524 for (uint i = 1; i < phi_len; i++) {
4525 if (phi_base->in(i) == NULL) {
4526 // incomplete phi; do not look at it yet!
4527 phi_reg = NULL;
4528 phi_len = (uint)-1;
4529 break;
4530 }
4531 }
4532 }
4533
4534 // Note: We do not call verify_sparse on entry, because inputs
4535 // can normalize to the base_memory via subsume_node or similar
4536 // mechanisms. This method repairs that damage.
4537
4538 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4539
4540 // Look at each slice.
4541 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4542 Node* old_in = in(i);
4543 // calculate the old memory value
4544 Node* old_mem = old_in;
4545 if (old_mem == empty_mem) old_mem = old_base;
4546 assert(old_mem == memory_at(i), "");
4547
4548 // maybe update (reslice) the old memory value
4549
4550 // simplify stacked MergeMems
4551 Node* new_mem = old_mem;
4552 MergeMemNode* old_mmem;
4553 if (old_mem != NULL && old_mem->is_MergeMem())
4554 old_mmem = old_mem->as_MergeMem();
4555 else
4556 old_mmem = NULL;
4557 if (old_mmem == this) {
4558 // This can happen if loops break up and safepoints disappear.
4559 // A merge of BotPtr (default) with a RawPtr memory derived from a
4560 // safepoint can be rewritten to a merge of the same BotPtr with
4561 // the BotPtr phi coming into the loop. If that phi disappears
4562 // also, we can end up with a self-loop of the mergemem.
4563 // In general, if loops degenerate and memory effects disappear,
4564 // a mergemem can be left looking at itself. This simply means
4565 // that the mergemem's default should be used, since there is
4566 // no longer any apparent effect on this slice.
4567 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4568 // from start. Update the input to TOP.
4569 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4570 }
4571 else if (old_mmem != NULL) {
4572 new_mem = old_mmem->memory_at(i);
4573 }
4574 // else preceding memory was not a MergeMem
4575
4576 // replace equivalent phis (unfortunately, they do not GVN together)
4577 if (new_mem != NULL && new_mem != new_base &&
4578 new_mem->req() == phi_len && new_mem->in(0) == phi_reg) {
4579 if (new_mem->is_Phi()) {
4580 PhiNode* phi_mem = new_mem->as_Phi();
4581 for (uint i = 1; i < phi_len; i++) {
4582 if (phi_base->in(i) != phi_mem->in(i)) {
4583 phi_mem = NULL;
4584 break;
4585 }
4586 }
4587 if (phi_mem != NULL) {
4588 // equivalent phi nodes; revert to the def
4589 new_mem = new_base;
4590 }
4591 }
4592 }
4593
4594 // maybe store down a new value
4595 Node* new_in = new_mem;
4596 if (new_in == new_base) new_in = empty_mem;
4597
4598 if (new_in != old_in) {
4599 // Warning: Do not combine this "if" with the previous "if"
4600 // A memory slice might have be be rewritten even if it is semantically
4601 // unchanged, if the base_memory value has changed.
4602 set_req(i, new_in);
4603 progress = this; // Report progress
4604 }
4605 }
4606
4607 if (new_base != old_base) {
4608 set_req(Compile::AliasIdxBot, new_base);
4609 // Don't use set_base_memory(new_base), because we need to update du.
4610 assert(base_memory() == new_base, "");
4611 progress = this;
4612 }
4613
4614 if( base_memory() == this ) {
4615 // a self cycle indicates this memory path is dead
4616 set_req(Compile::AliasIdxBot, empty_mem);
4617 }
4618
4619 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4620 // Recursion must occur after the self cycle check above
4621 if( base_memory()->is_MergeMem() ) {
4622 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4623 Node *m = phase->transform(new_mbase); // Rollup any cycles
4624 if( m != NULL &&
4625 (m->is_top() ||
4626 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4627 // propagate rollup of dead cycle to self
4628 set_req(Compile::AliasIdxBot, empty_mem);
4629 }
4630 }
4631
4632 if( base_memory() == empty_mem ) {
4633 progress = this;
4634 // Cut inputs during Parse phase only.
4635 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4636 if( !can_reshape ) {
4637 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4638 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4639 }
4640 }
4641 }
4642
4643 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4644 // Check if PhiNode::Ideal's "Split phis through memory merges"
4645 // transform should be attempted. Look for this->phi->this cycle.
4646 uint merge_width = req();
4647 if (merge_width > Compile::AliasIdxRaw) {
4648 PhiNode* phi = base_memory()->as_Phi();
4649 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4650 if (phi->in(i) == this) {
4651 phase->is_IterGVN()->_worklist.push(phi);
4652 break;
4653 }
4654 }
4655 }
4656 }
4657
4658 assert(progress || verify_sparse(), "please, no dups of base");
4659 return progress;
4660 }
4661
4662 //-------------------------set_base_memory-------------------------------------
4663 void MergeMemNode::set_base_memory(Node *new_base) {
4664 Node* empty_mem = empty_memory();
4665 set_req(Compile::AliasIdxBot, new_base);
4666 assert(memory_at(req()) == new_base, "must set default memory");
4667 // Clear out other occurrences of new_base:
4668 if (new_base != empty_mem) {
4669 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4670 if (in(i) == new_base) set_req(i, empty_mem);
4671 }
4672 }
4673 }
4674
4675 //------------------------------out_RegMask------------------------------------
4676 const RegMask &MergeMemNode::out_RegMask() const {
4677 return RegMask::Empty;
4678 }
4679
4680 //------------------------------dump_spec--------------------------------------
4681 #ifndef PRODUCT
4682 void MergeMemNode::dump_spec(outputStream *st) const {
4683 st->print(" {");
4684 Node* base_mem = base_memory();
4685 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4686 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4687 if (mem == base_mem) { st->print(" -"); continue; }
4688 st->print( " N%d:", mem->_idx );
4689 Compile::current()->get_adr_type(i)->dump_on(st);
4690 }
4691 st->print(" }");
4692 }
4693 #endif // !PRODUCT
4694
4695
4696 #ifdef ASSERT
4697 static bool might_be_same(Node* a, Node* b) {
4698 if (a == b) return true;
4699 if (!(a->is_Phi() || b->is_Phi())) return false;
4700 // phis shift around during optimization
4701 return true; // pretty stupid...
4702 }
4703
4704 // verify a narrow slice (either incoming or outgoing)
4705 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4706 if (!VerifyAliases) return; // don't bother to verify unless requested
4707 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
4708 if (Node::in_dump()) return; // muzzle asserts when printing
4709 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4710 assert(n != NULL, "");
4711 // Elide intervening MergeMem's
4712 while (n->is_MergeMem()) {
4713 n = n->as_MergeMem()->memory_at(alias_idx);
4714 }
4715 Compile* C = Compile::current();
4716 const TypePtr* n_adr_type = n->adr_type();
4717 if (n == m->empty_memory()) {
4718 // Implicit copy of base_memory()
4719 } else if (n_adr_type != TypePtr::BOTTOM) {
4720 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4721 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4722 } else {
4723 // A few places like make_runtime_call "know" that VM calls are narrow,
4724 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4725 bool expected_wide_mem = false;
4726 if (n == m->base_memory()) {
4727 expected_wide_mem = true;
4728 } else if (alias_idx == Compile::AliasIdxRaw ||
4729 n == m->memory_at(Compile::AliasIdxRaw)) {
4730 expected_wide_mem = true;
4731 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4732 // memory can "leak through" calls on channels that
4733 // are write-once. Allow this also.
4734 expected_wide_mem = true;
4735 }
4736 assert(expected_wide_mem, "expected narrow slice replacement");
4737 }
4738 }
4739 #else // !ASSERT
4740 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4741 #endif
4742
4743
4744 //-----------------------------memory_at---------------------------------------
4745 Node* MergeMemNode::memory_at(uint alias_idx) const {
4746 assert(alias_idx >= Compile::AliasIdxRaw ||
4747 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4748 "must avoid base_memory and AliasIdxTop");
4749
4750 // Otherwise, it is a narrow slice.
4751 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4752 Compile *C = Compile::current();
4753 if (is_empty_memory(n)) {
4754 // the array is sparse; empty slots are the "top" node
4755 n = base_memory();
4756 assert(Node::in_dump()
4757 || n == NULL || n->bottom_type() == Type::TOP
4758 || n->adr_type() == NULL // address is TOP
4759 || n->adr_type() == TypePtr::BOTTOM
4760 || n->adr_type() == TypeRawPtr::BOTTOM
4761 || Compile::current()->AliasLevel() == 0,
4762 "must be a wide memory");
4763 // AliasLevel == 0 if we are organizing the memory states manually.
4764 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4765 } else {
4766 // make sure the stored slice is sane
4767 #ifdef ASSERT
4768 if (VMError::is_error_reported() || Node::in_dump()) {
4769 } else if (might_be_same(n, base_memory())) {
4770 // Give it a pass: It is a mostly harmless repetition of the base.
4771 // This can arise normally from node subsumption during optimization.
4772 } else {
4773 verify_memory_slice(this, alias_idx, n);
4774 }
4775 #endif
4776 }
4777 return n;
4778 }
4779
4780 //---------------------------set_memory_at-------------------------------------
4781 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4782 verify_memory_slice(this, alias_idx, n);
4783 Node* empty_mem = empty_memory();
4784 if (n == base_memory()) n = empty_mem; // collapse default
4785 uint need_req = alias_idx+1;
4786 if (req() < need_req) {
4787 if (n == empty_mem) return; // already the default, so do not grow me
4788 // grow the sparse array
4789 do {
4790 add_req(empty_mem);
4791 } while (req() < need_req);
4792 }
4793 set_req( alias_idx, n );
4794 }
4795
4796
4797
4798 //--------------------------iteration_setup------------------------------------
4799 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4800 if (other != NULL) {
4801 grow_to_match(other);
4802 // invariant: the finite support of mm2 is within mm->req()
4803 #ifdef ASSERT
4804 for (uint i = req(); i < other->req(); i++) {
4805 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4806 }
4807 #endif
4808 }
4809 // Replace spurious copies of base_memory by top.
4810 Node* base_mem = base_memory();
4811 if (base_mem != NULL && !base_mem->is_top()) {
4812 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4813 if (in(i) == base_mem)
4814 set_req(i, empty_memory());
4815 }
4816 }
4817 }
4818
4819 //---------------------------grow_to_match-------------------------------------
4820 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4821 Node* empty_mem = empty_memory();
4822 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4823 // look for the finite support of the other memory
4824 for (uint i = other->req(); --i >= req(); ) {
4825 if (other->in(i) != empty_mem) {
4826 uint new_len = i+1;
4827 while (req() < new_len) add_req(empty_mem);
4828 break;
4829 }
4830 }
4831 }
4832
4833 //---------------------------verify_sparse-------------------------------------
4834 #ifndef PRODUCT
4835 bool MergeMemNode::verify_sparse() const {
4836 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4837 Node* base_mem = base_memory();
4838 // The following can happen in degenerate cases, since empty==top.
4839 if (is_empty_memory(base_mem)) return true;
4840 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4841 assert(in(i) != NULL, "sane slice");
4842 if (in(i) == base_mem) return false; // should have been the sentinel value!
4843 }
4844 return true;
4845 }
4846
4847 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4848 Node* n;
4849 n = mm->in(idx);
4850 if (mem == n) return true; // might be empty_memory()
4851 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4852 if (mem == n) return true;
4853 while (n->is_Phi() && (n = n->as_Phi()->is_copy()) != NULL) {
4854 if (mem == n) return true;
4855 if (n == NULL) break;
4856 }
4857 return false;
4858 }
4859 #endif // !PRODUCT