1 /*
2 * Copyright (c) 1997, 2020, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/javaClasses.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/barrierSet.hpp"
29 #include "gc/shared/c2/barrierSetC2.hpp"
30 #include "memory/allocation.inline.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/connode.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/loopnode.hpp"
40 #include "opto/machnode.hpp"
41 #include "opto/matcher.hpp"
42 #include "opto/memnode.hpp"
43 #include "opto/mulnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/phaseX.hpp"
46 #include "opto/regmask.hpp"
47 #include "opto/rootnode.hpp"
48 #include "utilities/align.hpp"
49 #include "utilities/copy.hpp"
50 #include "utilities/macros.hpp"
51 #include "utilities/powerOfTwo.hpp"
52 #include "utilities/vmError.hpp"
53
54 // Portions of code courtesy of Clifford Click
55
56 // Optimization - Graph Style
57
58 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st);
59
60 //=============================================================================
size_of() const61 uint MemNode::size_of() const { return sizeof(*this); }
62
adr_type() const63 const TypePtr *MemNode::adr_type() const {
64 Node* adr = in(Address);
65 if (adr == NULL) return NULL; // node is dead
66 const TypePtr* cross_check = NULL;
67 DEBUG_ONLY(cross_check = _adr_type);
68 return calculate_adr_type(adr->bottom_type(), cross_check);
69 }
70
check_if_adr_maybe_raw(Node * adr)71 bool MemNode::check_if_adr_maybe_raw(Node* adr) {
72 if (adr != NULL) {
73 if (adr->bottom_type()->base() == Type::RawPtr || adr->bottom_type()->base() == Type::AnyPtr) {
74 return true;
75 }
76 }
77 return false;
78 }
79
80 #ifndef PRODUCT
dump_spec(outputStream * st) const81 void MemNode::dump_spec(outputStream *st) const {
82 if (in(Address) == NULL) return; // node is dead
83 #ifndef ASSERT
84 // fake the missing field
85 const TypePtr* _adr_type = NULL;
86 if (in(Address) != NULL)
87 _adr_type = in(Address)->bottom_type()->isa_ptr();
88 #endif
89 dump_adr_type(this, _adr_type, st);
90
91 Compile* C = Compile::current();
92 if (C->alias_type(_adr_type)->is_volatile()) {
93 st->print(" Volatile!");
94 }
95 if (_unaligned_access) {
96 st->print(" unaligned");
97 }
98 if (_mismatched_access) {
99 st->print(" mismatched");
100 }
101 if (_unsafe_access) {
102 st->print(" unsafe");
103 }
104 }
105
dump_adr_type(const Node * mem,const TypePtr * adr_type,outputStream * st)106 void MemNode::dump_adr_type(const Node* mem, const TypePtr* adr_type, outputStream *st) {
107 st->print(" @");
108 if (adr_type == NULL) {
109 st->print("NULL");
110 } else {
111 adr_type->dump_on(st);
112 Compile* C = Compile::current();
113 Compile::AliasType* atp = NULL;
114 if (C->have_alias_type(adr_type)) atp = C->alias_type(adr_type);
115 if (atp == NULL)
116 st->print(", idx=?\?;");
117 else if (atp->index() == Compile::AliasIdxBot)
118 st->print(", idx=Bot;");
119 else if (atp->index() == Compile::AliasIdxTop)
120 st->print(", idx=Top;");
121 else if (atp->index() == Compile::AliasIdxRaw)
122 st->print(", idx=Raw;");
123 else {
124 ciField* field = atp->field();
125 if (field) {
126 st->print(", name=");
127 field->print_name_on(st);
128 }
129 st->print(", idx=%d;", atp->index());
130 }
131 }
132 }
133
134 extern void print_alias_types();
135
136 #endif
137
optimize_simple_memory_chain(Node * mchain,const TypeOopPtr * t_oop,Node * load,PhaseGVN * phase)138 Node *MemNode::optimize_simple_memory_chain(Node *mchain, const TypeOopPtr *t_oop, Node *load, PhaseGVN *phase) {
139 assert((t_oop != NULL), "sanity");
140 bool is_instance = t_oop->is_known_instance_field();
141 bool is_boxed_value_load = t_oop->is_ptr_to_boxed_value() &&
142 (load != NULL) && load->is_Load() &&
143 (phase->is_IterGVN() != NULL);
144 if (!(is_instance || is_boxed_value_load))
145 return mchain; // don't try to optimize non-instance types
146 uint instance_id = t_oop->instance_id();
147 Node *start_mem = phase->C->start()->proj_out_or_null(TypeFunc::Memory);
148 Node *prev = NULL;
149 Node *result = mchain;
150 while (prev != result) {
151 prev = result;
152 if (result == start_mem)
153 break; // hit one of our sentinels
154 // skip over a call which does not affect this memory slice
155 if (result->is_Proj() && result->as_Proj()->_con == TypeFunc::Memory) {
156 Node *proj_in = result->in(0);
157 if (proj_in->is_Allocate() && proj_in->_idx == instance_id) {
158 break; // hit one of our sentinels
159 } else if (proj_in->is_Call()) {
160 // ArrayCopyNodes processed here as well
161 CallNode *call = proj_in->as_Call();
162 if (!call->may_modify(t_oop, phase)) { // returns false for instances
163 result = call->in(TypeFunc::Memory);
164 }
165 } else if (proj_in->is_Initialize()) {
166 AllocateNode* alloc = proj_in->as_Initialize()->allocation();
167 // Stop if this is the initialization for the object instance which
168 // contains this memory slice, otherwise skip over it.
169 if ((alloc == NULL) || (alloc->_idx == instance_id)) {
170 break;
171 }
172 if (is_instance) {
173 result = proj_in->in(TypeFunc::Memory);
174 } else if (is_boxed_value_load) {
175 Node* klass = alloc->in(AllocateNode::KlassNode);
176 const TypeKlassPtr* tklass = phase->type(klass)->is_klassptr();
177 if (tklass->klass_is_exact() && !tklass->klass()->equals(t_oop->klass())) {
178 result = proj_in->in(TypeFunc::Memory); // not related allocation
179 }
180 }
181 } else if (proj_in->is_MemBar()) {
182 ArrayCopyNode* ac = NULL;
183 if (ArrayCopyNode::may_modify(t_oop, proj_in->as_MemBar(), phase, ac)) {
184 break;
185 }
186 result = proj_in->in(TypeFunc::Memory);
187 } else {
188 assert(false, "unexpected projection");
189 }
190 } else if (result->is_ClearArray()) {
191 if (!is_instance || !ClearArrayNode::step_through(&result, instance_id, phase)) {
192 // Can not bypass initialization of the instance
193 // we are looking for.
194 break;
195 }
196 // Otherwise skip it (the call updated 'result' value).
197 } else if (result->is_MergeMem()) {
198 result = step_through_mergemem(phase, result->as_MergeMem(), t_oop, NULL, tty);
199 }
200 }
201 return result;
202 }
203
optimize_memory_chain(Node * mchain,const TypePtr * t_adr,Node * load,PhaseGVN * phase)204 Node *MemNode::optimize_memory_chain(Node *mchain, const TypePtr *t_adr, Node *load, PhaseGVN *phase) {
205 const TypeOopPtr* t_oop = t_adr->isa_oopptr();
206 if (t_oop == NULL)
207 return mchain; // don't try to optimize non-oop types
208 Node* result = optimize_simple_memory_chain(mchain, t_oop, load, phase);
209 bool is_instance = t_oop->is_known_instance_field();
210 PhaseIterGVN *igvn = phase->is_IterGVN();
211 if (is_instance && igvn != NULL && result->is_Phi()) {
212 PhiNode *mphi = result->as_Phi();
213 assert(mphi->bottom_type() == Type::MEMORY, "memory phi required");
214 const TypePtr *t = mphi->adr_type();
215 if (t == TypePtr::BOTTOM || t == TypeRawPtr::BOTTOM ||
216 (t->isa_oopptr() && !t->is_oopptr()->is_known_instance() &&
217 t->is_oopptr()->cast_to_exactness(true)
218 ->is_oopptr()->cast_to_ptr_type(t_oop->ptr())
219 ->is_oopptr()->cast_to_instance_id(t_oop->instance_id()) == t_oop)) {
220 // clone the Phi with our address type
221 result = mphi->split_out_instance(t_adr, igvn);
222 } else {
223 assert(phase->C->get_alias_index(t) == phase->C->get_alias_index(t_adr), "correct memory chain");
224 }
225 }
226 return result;
227 }
228
step_through_mergemem(PhaseGVN * phase,MergeMemNode * mmem,const TypePtr * tp,const TypePtr * adr_check,outputStream * st)229 static Node *step_through_mergemem(PhaseGVN *phase, MergeMemNode *mmem, const TypePtr *tp, const TypePtr *adr_check, outputStream *st) {
230 uint alias_idx = phase->C->get_alias_index(tp);
231 Node *mem = mmem;
232 #ifdef ASSERT
233 {
234 // Check that current type is consistent with the alias index used during graph construction
235 assert(alias_idx >= Compile::AliasIdxRaw, "must not be a bad alias_idx");
236 bool consistent = adr_check == NULL || adr_check->empty() ||
237 phase->C->must_alias(adr_check, alias_idx );
238 // Sometimes dead array references collapse to a[-1], a[-2], or a[-3]
239 if( !consistent && adr_check != NULL && !adr_check->empty() &&
240 tp->isa_aryptr() && tp->offset() == Type::OffsetBot &&
241 adr_check->isa_aryptr() && adr_check->offset() != Type::OffsetBot &&
242 ( adr_check->offset() == arrayOopDesc::length_offset_in_bytes() ||
243 adr_check->offset() == oopDesc::klass_offset_in_bytes() ||
244 adr_check->offset() == oopDesc::mark_offset_in_bytes() ) ) {
245 // don't assert if it is dead code.
246 consistent = true;
247 }
248 if( !consistent ) {
249 st->print("alias_idx==%d, adr_check==", alias_idx);
250 if( adr_check == NULL ) {
251 st->print("NULL");
252 } else {
253 adr_check->dump();
254 }
255 st->cr();
256 print_alias_types();
257 assert(consistent, "adr_check must match alias idx");
258 }
259 }
260 #endif
261 // TypeOopPtr::NOTNULL+any is an OOP with unknown offset - generally
262 // means an array I have not precisely typed yet. Do not do any
263 // alias stuff with it any time soon.
264 const TypeOopPtr *toop = tp->isa_oopptr();
265 if( tp->base() != Type::AnyPtr &&
266 !(toop &&
267 toop->klass() != NULL &&
268 toop->klass()->is_java_lang_Object() &&
269 toop->offset() == Type::OffsetBot) ) {
270 // compress paths and change unreachable cycles to TOP
271 // If not, we can update the input infinitely along a MergeMem cycle
272 // Equivalent code in PhiNode::Ideal
273 Node* m = phase->transform(mmem);
274 // If transformed to a MergeMem, get the desired slice
275 // Otherwise the returned node represents memory for every slice
276 mem = (m->is_MergeMem())? m->as_MergeMem()->memory_at(alias_idx) : m;
277 // Update input if it is progress over what we have now
278 }
279 return mem;
280 }
281
282 //--------------------------Ideal_common---------------------------------------
283 // Look for degenerate control and memory inputs. Bypass MergeMem inputs.
284 // Unhook non-raw memories from complete (macro-expanded) initializations.
Ideal_common(PhaseGVN * phase,bool can_reshape)285 Node *MemNode::Ideal_common(PhaseGVN *phase, bool can_reshape) {
286 // If our control input is a dead region, kill all below the region
287 Node *ctl = in(MemNode::Control);
288 if (ctl && remove_dead_region(phase, can_reshape))
289 return this;
290 ctl = in(MemNode::Control);
291 // Don't bother trying to transform a dead node
292 if (ctl && ctl->is_top()) return NodeSentinel;
293
294 PhaseIterGVN *igvn = phase->is_IterGVN();
295 // Wait if control on the worklist.
296 if (ctl && can_reshape && igvn != NULL) {
297 Node* bol = NULL;
298 Node* cmp = NULL;
299 if (ctl->in(0)->is_If()) {
300 assert(ctl->is_IfTrue() || ctl->is_IfFalse(), "sanity");
301 bol = ctl->in(0)->in(1);
302 if (bol->is_Bool())
303 cmp = ctl->in(0)->in(1)->in(1);
304 }
305 if (igvn->_worklist.member(ctl) ||
306 (bol != NULL && igvn->_worklist.member(bol)) ||
307 (cmp != NULL && igvn->_worklist.member(cmp)) ) {
308 // This control path may be dead.
309 // Delay this memory node transformation until the control is processed.
310 igvn->_worklist.push(this);
311 return NodeSentinel; // caller will return NULL
312 }
313 }
314 // Ignore if memory is dead, or self-loop
315 Node *mem = in(MemNode::Memory);
316 if (phase->type( mem ) == Type::TOP) return NodeSentinel; // caller will return NULL
317 assert(mem != this, "dead loop in MemNode::Ideal");
318
319 if (can_reshape && igvn != NULL && igvn->_worklist.member(mem)) {
320 // This memory slice may be dead.
321 // Delay this mem node transformation until the memory is processed.
322 igvn->_worklist.push(this);
323 return NodeSentinel; // caller will return NULL
324 }
325
326 Node *address = in(MemNode::Address);
327 const Type *t_adr = phase->type(address);
328 if (t_adr == Type::TOP) return NodeSentinel; // caller will return NULL
329
330 if (can_reshape && is_unsafe_access() && (t_adr == TypePtr::NULL_PTR)) {
331 // Unsafe off-heap access with zero address. Remove access and other control users
332 // to not confuse optimizations and add a HaltNode to fail if this is ever executed.
333 assert(ctl != NULL, "unsafe accesses should be control dependent");
334 for (DUIterator_Fast imax, i = ctl->fast_outs(imax); i < imax; i++) {
335 Node* u = ctl->fast_out(i);
336 if (u != ctl) {
337 igvn->rehash_node_delayed(u);
338 int nb = u->replace_edge(ctl, phase->C->top());
339 --i, imax -= nb;
340 }
341 }
342 Node* frame = igvn->transform(new ParmNode(phase->C->start(), TypeFunc::FramePtr));
343 Node* halt = igvn->transform(new HaltNode(ctl, frame, "unsafe off-heap access with zero address"));
344 phase->C->root()->add_req(halt);
345 return this;
346 }
347
348 if (can_reshape && igvn != NULL &&
349 (igvn->_worklist.member(address) ||
350 (igvn->_worklist.size() > 0 && t_adr != adr_type())) ) {
351 // The address's base and type may change when the address is processed.
352 // Delay this mem node transformation until the address is processed.
353 igvn->_worklist.push(this);
354 return NodeSentinel; // caller will return NULL
355 }
356
357 // Do NOT remove or optimize the next lines: ensure a new alias index
358 // is allocated for an oop pointer type before Escape Analysis.
359 // Note: C++ will not remove it since the call has side effect.
360 if (t_adr->isa_oopptr()) {
361 int alias_idx = phase->C->get_alias_index(t_adr->is_ptr());
362 }
363
364 Node* base = NULL;
365 if (address->is_AddP()) {
366 base = address->in(AddPNode::Base);
367 }
368 if (base != NULL && phase->type(base)->higher_equal(TypePtr::NULL_PTR) &&
369 !t_adr->isa_rawptr()) {
370 // Note: raw address has TOP base and top->higher_equal(TypePtr::NULL_PTR) is true.
371 // Skip this node optimization if its address has TOP base.
372 return NodeSentinel; // caller will return NULL
373 }
374
375 // Avoid independent memory operations
376 Node* old_mem = mem;
377
378 // The code which unhooks non-raw memories from complete (macro-expanded)
379 // initializations was removed. After macro-expansion all stores catched
380 // by Initialize node became raw stores and there is no information
381 // which memory slices they modify. So it is unsafe to move any memory
382 // operation above these stores. Also in most cases hooked non-raw memories
383 // were already unhooked by using information from detect_ptr_independence()
384 // and find_previous_store().
385
386 if (mem->is_MergeMem()) {
387 MergeMemNode* mmem = mem->as_MergeMem();
388 const TypePtr *tp = t_adr->is_ptr();
389
390 mem = step_through_mergemem(phase, mmem, tp, adr_type(), tty);
391 }
392
393 if (mem != old_mem) {
394 set_req(MemNode::Memory, mem);
395 if (can_reshape && old_mem->outcnt() == 0 && igvn != NULL) {
396 igvn->_worklist.push(old_mem);
397 }
398 if (phase->type(mem) == Type::TOP) return NodeSentinel;
399 return this;
400 }
401
402 // let the subclass continue analyzing...
403 return NULL;
404 }
405
406 // Helper function for proving some simple control dominations.
407 // Attempt to prove that all control inputs of 'dom' dominate 'sub'.
408 // Already assumes that 'dom' is available at 'sub', and that 'sub'
409 // is not a constant (dominated by the method's StartNode).
410 // Used by MemNode::find_previous_store to prove that the
411 // control input of a memory operation predates (dominates)
412 // an allocation it wants to look past.
all_controls_dominate(Node * dom,Node * sub)413 bool MemNode::all_controls_dominate(Node* dom, Node* sub) {
414 if (dom == NULL || dom->is_top() || sub == NULL || sub->is_top())
415 return false; // Conservative answer for dead code
416
417 // Check 'dom'. Skip Proj and CatchProj nodes.
418 dom = dom->find_exact_control(dom);
419 if (dom == NULL || dom->is_top())
420 return false; // Conservative answer for dead code
421
422 if (dom == sub) {
423 // For the case when, for example, 'sub' is Initialize and the original
424 // 'dom' is Proj node of the 'sub'.
425 return false;
426 }
427
428 if (dom->is_Con() || dom->is_Start() || dom->is_Root() || dom == sub)
429 return true;
430
431 // 'dom' dominates 'sub' if its control edge and control edges
432 // of all its inputs dominate or equal to sub's control edge.
433
434 // Currently 'sub' is either Allocate, Initialize or Start nodes.
435 // Or Region for the check in LoadNode::Ideal();
436 // 'sub' should have sub->in(0) != NULL.
437 assert(sub->is_Allocate() || sub->is_Initialize() || sub->is_Start() ||
438 sub->is_Region() || sub->is_Call(), "expecting only these nodes");
439
440 // Get control edge of 'sub'.
441 Node* orig_sub = sub;
442 sub = sub->find_exact_control(sub->in(0));
443 if (sub == NULL || sub->is_top())
444 return false; // Conservative answer for dead code
445
446 assert(sub->is_CFG(), "expecting control");
447
448 if (sub == dom)
449 return true;
450
451 if (sub->is_Start() || sub->is_Root())
452 return false;
453
454 {
455 // Check all control edges of 'dom'.
456
457 ResourceMark rm;
458 Node_List nlist;
459 Unique_Node_List dom_list;
460
461 dom_list.push(dom);
462 bool only_dominating_controls = false;
463
464 for (uint next = 0; next < dom_list.size(); next++) {
465 Node* n = dom_list.at(next);
466 if (n == orig_sub)
467 return false; // One of dom's inputs dominated by sub.
468 if (!n->is_CFG() && n->pinned()) {
469 // Check only own control edge for pinned non-control nodes.
470 n = n->find_exact_control(n->in(0));
471 if (n == NULL || n->is_top())
472 return false; // Conservative answer for dead code
473 assert(n->is_CFG(), "expecting control");
474 dom_list.push(n);
475 } else if (n->is_Con() || n->is_Start() || n->is_Root()) {
476 only_dominating_controls = true;
477 } else if (n->is_CFG()) {
478 if (n->dominates(sub, nlist))
479 only_dominating_controls = true;
480 else
481 return false;
482 } else {
483 // First, own control edge.
484 Node* m = n->find_exact_control(n->in(0));
485 if (m != NULL) {
486 if (m->is_top())
487 return false; // Conservative answer for dead code
488 dom_list.push(m);
489 }
490 // Now, the rest of edges.
491 uint cnt = n->req();
492 for (uint i = 1; i < cnt; i++) {
493 m = n->find_exact_control(n->in(i));
494 if (m == NULL || m->is_top())
495 continue;
496 dom_list.push(m);
497 }
498 }
499 }
500 return only_dominating_controls;
501 }
502 }
503
504 //---------------------detect_ptr_independence---------------------------------
505 // Used by MemNode::find_previous_store to prove that two base
506 // pointers are never equal.
507 // The pointers are accompanied by their associated allocations,
508 // if any, which have been previously discovered by the caller.
detect_ptr_independence(Node * p1,AllocateNode * a1,Node * p2,AllocateNode * a2,PhaseTransform * phase)509 bool MemNode::detect_ptr_independence(Node* p1, AllocateNode* a1,
510 Node* p2, AllocateNode* a2,
511 PhaseTransform* phase) {
512 // Attempt to prove that these two pointers cannot be aliased.
513 // They may both manifestly be allocations, and they should differ.
514 // Or, if they are not both allocations, they can be distinct constants.
515 // Otherwise, one is an allocation and the other a pre-existing value.
516 if (a1 == NULL && a2 == NULL) { // neither an allocation
517 return (p1 != p2) && p1->is_Con() && p2->is_Con();
518 } else if (a1 != NULL && a2 != NULL) { // both allocations
519 return (a1 != a2);
520 } else if (a1 != NULL) { // one allocation a1
521 // (Note: p2->is_Con implies p2->in(0)->is_Root, which dominates.)
522 return all_controls_dominate(p2, a1);
523 } else { //(a2 != NULL) // one allocation a2
524 return all_controls_dominate(p1, a2);
525 }
526 return false;
527 }
528
529
530 // Find an arraycopy that must have set (can_see_stored_value=true) or
531 // could have set (can_see_stored_value=false) the value for this load
find_previous_arraycopy(PhaseTransform * phase,Node * ld_alloc,Node * & mem,bool can_see_stored_value) const532 Node* LoadNode::find_previous_arraycopy(PhaseTransform* phase, Node* ld_alloc, Node*& mem, bool can_see_stored_value) const {
533 ArrayCopyNode* ac = find_array_copy_clone(phase, ld_alloc, mem);
534 if (ac != NULL) {
535 Node* ld_addp = in(MemNode::Address);
536 Node* src = ac->in(ArrayCopyNode::Src);
537 const TypeAryPtr* ary_t = phase->type(src)->isa_aryptr();
538
539 // This is a load from a cloned array. The corresponding arraycopy ac must
540 // have set the value for the load and we can return ac but only if the load
541 // is known to be within bounds. This is checked below.
542 if (ary_t != NULL && ld_addp->is_AddP()) {
543 Node* ld_offs = ld_addp->in(AddPNode::Offset);
544 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type();
545 jlong header = arrayOopDesc::base_offset_in_bytes(ary_elem);
546 jlong elemsize = type2aelembytes(ary_elem);
547
548 const TypeX* ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
549 const TypeInt* sizetype = ary_t->size();
550
551 if (ld_offs_t->_lo >= header && ld_offs_t->_hi < (sizetype->_lo * elemsize + header)) {
552 // The load is known to be within bounds. It receives its value from ac.
553 return ac;
554 }
555 // The load is known to be out-of-bounds.
556 }
557 // The load could be out-of-bounds. It must not be hoisted but must remain
558 // dependent on the runtime range check. This is achieved by returning NULL.
559 } else if (mem->is_Proj() && mem->in(0) != NULL && mem->in(0)->is_ArrayCopy()) {
560 ArrayCopyNode* ac = mem->in(0)->as_ArrayCopy();
561
562 if (ac->is_arraycopy_validated() ||
563 ac->is_copyof_validated() ||
564 ac->is_copyofrange_validated()) {
565 Node* ld_addp = in(MemNode::Address);
566 if (ld_addp->is_AddP()) {
567 Node* ld_base = ld_addp->in(AddPNode::Address);
568 Node* ld_offs = ld_addp->in(AddPNode::Offset);
569
570 Node* dest = ac->in(ArrayCopyNode::Dest);
571
572 if (dest == ld_base) {
573 const TypeX *ld_offs_t = phase->type(ld_offs)->isa_intptr_t();
574 if (ac->modifies(ld_offs_t->_lo, ld_offs_t->_hi, phase, can_see_stored_value)) {
575 return ac;
576 }
577 if (!can_see_stored_value) {
578 mem = ac->in(TypeFunc::Memory);
579 }
580 }
581 }
582 }
583 }
584 return NULL;
585 }
586
find_array_copy_clone(PhaseTransform * phase,Node * ld_alloc,Node * mem) const587 ArrayCopyNode* MemNode::find_array_copy_clone(PhaseTransform* phase, Node* ld_alloc, Node* mem) const {
588 if (mem->is_Proj() && mem->in(0) != NULL && (mem->in(0)->Opcode() == Op_MemBarStoreStore ||
589 mem->in(0)->Opcode() == Op_MemBarCPUOrder)) {
590 if (ld_alloc != NULL) {
591 // Check if there is an array copy for a clone
592 Node* mb = mem->in(0);
593 ArrayCopyNode* ac = NULL;
594 if (mb->in(0) != NULL && mb->in(0)->is_Proj() &&
595 mb->in(0)->in(0) != NULL && mb->in(0)->in(0)->is_ArrayCopy()) {
596 ac = mb->in(0)->in(0)->as_ArrayCopy();
597 } else {
598 // Step over GC barrier when ReduceInitialCardMarks is disabled
599 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
600 Node* control_proj_ac = bs->step_over_gc_barrier(mb->in(0));
601
602 if (control_proj_ac->is_Proj() && control_proj_ac->in(0)->is_ArrayCopy()) {
603 ac = control_proj_ac->in(0)->as_ArrayCopy();
604 }
605 }
606
607 if (ac != NULL && ac->is_clonebasic()) {
608 AllocateNode* alloc = AllocateNode::Ideal_allocation(ac->in(ArrayCopyNode::Dest), phase);
609 if (alloc != NULL && alloc == ld_alloc) {
610 return ac;
611 }
612 }
613 }
614 }
615 return NULL;
616 }
617
618 // The logic for reordering loads and stores uses four steps:
619 // (a) Walk carefully past stores and initializations which we
620 // can prove are independent of this load.
621 // (b) Observe that the next memory state makes an exact match
622 // with self (load or store), and locate the relevant store.
623 // (c) Ensure that, if we were to wire self directly to the store,
624 // the optimizer would fold it up somehow.
625 // (d) Do the rewiring, and return, depending on some other part of
626 // the optimizer to fold up the load.
627 // This routine handles steps (a) and (b). Steps (c) and (d) are
628 // specific to loads and stores, so they are handled by the callers.
629 // (Currently, only LoadNode::Ideal has steps (c), (d). More later.)
630 //
find_previous_store(PhaseTransform * phase)631 Node* MemNode::find_previous_store(PhaseTransform* phase) {
632 Node* ctrl = in(MemNode::Control);
633 Node* adr = in(MemNode::Address);
634 intptr_t offset = 0;
635 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
636 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
637
638 if (offset == Type::OffsetBot)
639 return NULL; // cannot unalias unless there are precise offsets
640
641 const bool adr_maybe_raw = check_if_adr_maybe_raw(adr);
642 const TypeOopPtr *addr_t = adr->bottom_type()->isa_oopptr();
643
644 intptr_t size_in_bytes = memory_size();
645
646 Node* mem = in(MemNode::Memory); // start searching here...
647
648 int cnt = 50; // Cycle limiter
649 for (;;) { // While we can dance past unrelated stores...
650 if (--cnt < 0) break; // Caught in cycle or a complicated dance?
651
652 Node* prev = mem;
653 if (mem->is_Store()) {
654 Node* st_adr = mem->in(MemNode::Address);
655 intptr_t st_offset = 0;
656 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
657 if (st_base == NULL)
658 break; // inscrutable pointer
659
660 // For raw accesses it's not enough to prove that constant offsets don't intersect.
661 // We need the bases to be the equal in order for the offset check to make sense.
662 if ((adr_maybe_raw || check_if_adr_maybe_raw(st_adr)) && st_base != base) {
663 break;
664 }
665
666 if (st_offset != offset && st_offset != Type::OffsetBot) {
667 const int MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
668 assert(mem->as_Store()->memory_size() <= MAX_STORE, "");
669 if (st_offset >= offset + size_in_bytes ||
670 st_offset <= offset - MAX_STORE ||
671 st_offset <= offset - mem->as_Store()->memory_size()) {
672 // Success: The offsets are provably independent.
673 // (You may ask, why not just test st_offset != offset and be done?
674 // The answer is that stores of different sizes can co-exist
675 // in the same sequence of RawMem effects. We sometimes initialize
676 // a whole 'tile' of array elements with a single jint or jlong.)
677 mem = mem->in(MemNode::Memory);
678 continue; // (a) advance through independent store memory
679 }
680 }
681 if (st_base != base &&
682 detect_ptr_independence(base, alloc,
683 st_base,
684 AllocateNode::Ideal_allocation(st_base, phase),
685 phase)) {
686 // Success: The bases are provably independent.
687 mem = mem->in(MemNode::Memory);
688 continue; // (a) advance through independent store memory
689 }
690
691 // (b) At this point, if the bases or offsets do not agree, we lose,
692 // since we have not managed to prove 'this' and 'mem' independent.
693 if (st_base == base && st_offset == offset) {
694 return mem; // let caller handle steps (c), (d)
695 }
696
697 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
698 InitializeNode* st_init = mem->in(0)->as_Initialize();
699 AllocateNode* st_alloc = st_init->allocation();
700 if (st_alloc == NULL)
701 break; // something degenerated
702 bool known_identical = false;
703 bool known_independent = false;
704 if (alloc == st_alloc)
705 known_identical = true;
706 else if (alloc != NULL)
707 known_independent = true;
708 else if (all_controls_dominate(this, st_alloc))
709 known_independent = true;
710
711 if (known_independent) {
712 // The bases are provably independent: Either they are
713 // manifestly distinct allocations, or else the control
714 // of this load dominates the store's allocation.
715 int alias_idx = phase->C->get_alias_index(adr_type());
716 if (alias_idx == Compile::AliasIdxRaw) {
717 mem = st_alloc->in(TypeFunc::Memory);
718 } else {
719 mem = st_init->memory(alias_idx);
720 }
721 continue; // (a) advance through independent store memory
722 }
723
724 // (b) at this point, if we are not looking at a store initializing
725 // the same allocation we are loading from, we lose.
726 if (known_identical) {
727 // From caller, can_see_stored_value will consult find_captured_store.
728 return mem; // let caller handle steps (c), (d)
729 }
730
731 } else if (find_previous_arraycopy(phase, alloc, mem, false) != NULL) {
732 if (prev != mem) {
733 // Found an arraycopy but it doesn't affect that load
734 continue;
735 }
736 // Found an arraycopy that may affect that load
737 return mem;
738 } else if (addr_t != NULL && addr_t->is_known_instance_field()) {
739 // Can't use optimize_simple_memory_chain() since it needs PhaseGVN.
740 if (mem->is_Proj() && mem->in(0)->is_Call()) {
741 // ArrayCopyNodes processed here as well.
742 CallNode *call = mem->in(0)->as_Call();
743 if (!call->may_modify(addr_t, phase)) {
744 mem = call->in(TypeFunc::Memory);
745 continue; // (a) advance through independent call memory
746 }
747 } else if (mem->is_Proj() && mem->in(0)->is_MemBar()) {
748 ArrayCopyNode* ac = NULL;
749 if (ArrayCopyNode::may_modify(addr_t, mem->in(0)->as_MemBar(), phase, ac)) {
750 break;
751 }
752 mem = mem->in(0)->in(TypeFunc::Memory);
753 continue; // (a) advance through independent MemBar memory
754 } else if (mem->is_ClearArray()) {
755 if (ClearArrayNode::step_through(&mem, (uint)addr_t->instance_id(), phase)) {
756 // (the call updated 'mem' value)
757 continue; // (a) advance through independent allocation memory
758 } else {
759 // Can not bypass initialization of the instance
760 // we are looking for.
761 return mem;
762 }
763 } else if (mem->is_MergeMem()) {
764 int alias_idx = phase->C->get_alias_index(adr_type());
765 mem = mem->as_MergeMem()->memory_at(alias_idx);
766 continue; // (a) advance through independent MergeMem memory
767 }
768 }
769
770 // Unless there is an explicit 'continue', we must bail out here,
771 // because 'mem' is an inscrutable memory state (e.g., a call).
772 break;
773 }
774
775 return NULL; // bail out
776 }
777
778 //----------------------calculate_adr_type-------------------------------------
779 // Helper function. Notices when the given type of address hits top or bottom.
780 // Also, asserts a cross-check of the type against the expected address type.
calculate_adr_type(const Type * t,const TypePtr * cross_check)781 const TypePtr* MemNode::calculate_adr_type(const Type* t, const TypePtr* cross_check) {
782 if (t == Type::TOP) return NULL; // does not touch memory any more?
783 #ifdef ASSERT
784 if (!VerifyAliases || VMError::is_error_reported() || Node::in_dump()) cross_check = NULL;
785 #endif
786 const TypePtr* tp = t->isa_ptr();
787 if (tp == NULL) {
788 assert(cross_check == NULL || cross_check == TypePtr::BOTTOM, "expected memory type must be wide");
789 return TypePtr::BOTTOM; // touches lots of memory
790 } else {
791 #ifdef ASSERT
792 // %%%% [phh] We don't check the alias index if cross_check is
793 // TypeRawPtr::BOTTOM. Needs to be investigated.
794 if (cross_check != NULL &&
795 cross_check != TypePtr::BOTTOM &&
796 cross_check != TypeRawPtr::BOTTOM) {
797 // Recheck the alias index, to see if it has changed (due to a bug).
798 Compile* C = Compile::current();
799 assert(C->get_alias_index(cross_check) == C->get_alias_index(tp),
800 "must stay in the original alias category");
801 // The type of the address must be contained in the adr_type,
802 // disregarding "null"-ness.
803 // (We make an exception for TypeRawPtr::BOTTOM, which is a bit bucket.)
804 const TypePtr* tp_notnull = tp->join(TypePtr::NOTNULL)->is_ptr();
805 assert(cross_check->meet(tp_notnull) == cross_check->remove_speculative(),
806 "real address must not escape from expected memory type");
807 }
808 #endif
809 return tp;
810 }
811 }
812
813 //=============================================================================
814 // Should LoadNode::Ideal() attempt to remove control edges?
can_remove_control() const815 bool LoadNode::can_remove_control() const {
816 return true;
817 }
size_of() const818 uint LoadNode::size_of() const { return sizeof(*this); }
cmp(const Node & n) const819 bool LoadNode::cmp( const Node &n ) const
820 { return !Type::cmp( _type, ((LoadNode&)n)._type ); }
bottom_type() const821 const Type *LoadNode::bottom_type() const { return _type; }
ideal_reg() const822 uint LoadNode::ideal_reg() const {
823 return _type->ideal_reg();
824 }
825
826 #ifndef PRODUCT
dump_spec(outputStream * st) const827 void LoadNode::dump_spec(outputStream *st) const {
828 MemNode::dump_spec(st);
829 if( !Verbose && !WizardMode ) {
830 // standard dump does this in Verbose and WizardMode
831 st->print(" #"); _type->dump_on(st);
832 }
833 if (!depends_only_on_test()) {
834 st->print(" (does not depend only on test)");
835 }
836 }
837 #endif
838
839 #ifdef ASSERT
840 //----------------------------is_immutable_value-------------------------------
841 // Helper function to allow a raw load without control edge for some cases
is_immutable_value(Node * adr)842 bool LoadNode::is_immutable_value(Node* adr) {
843 return (adr->is_AddP() && adr->in(AddPNode::Base)->is_top() &&
844 adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
845 (adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
846 in_bytes(JavaThread::osthread_offset()) ||
847 adr->in(AddPNode::Offset)->find_intptr_t_con(-1) ==
848 in_bytes(JavaThread::threadObj_offset())));
849 }
850 #endif
851
852 //----------------------------LoadNode::make-----------------------------------
853 // Polymorphic factory method:
make(PhaseGVN & gvn,Node * ctl,Node * mem,Node * adr,const TypePtr * adr_type,const Type * rt,BasicType bt,MemOrd mo,ControlDependency control_dependency,bool unaligned,bool mismatched,bool unsafe,uint8_t barrier_data)854 Node *LoadNode::make(PhaseGVN& gvn, Node *ctl, Node *mem, Node *adr, const TypePtr* adr_type, const Type *rt, BasicType bt, MemOrd mo,
855 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
856 Compile* C = gvn.C;
857
858 // sanity check the alias category against the created node type
859 assert(!(adr_type->isa_oopptr() &&
860 adr_type->offset() == oopDesc::klass_offset_in_bytes()),
861 "use LoadKlassNode instead");
862 assert(!(adr_type->isa_aryptr() &&
863 adr_type->offset() == arrayOopDesc::length_offset_in_bytes()),
864 "use LoadRangeNode instead");
865 // Check control edge of raw loads
866 assert( ctl != NULL || C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
867 // oop will be recorded in oop map if load crosses safepoint
868 rt->isa_oopptr() || is_immutable_value(adr),
869 "raw memory operations should have control edge");
870 LoadNode* load = NULL;
871 switch (bt) {
872 case T_BOOLEAN: load = new LoadUBNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
873 case T_BYTE: load = new LoadBNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
874 case T_INT: load = new LoadINode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
875 case T_CHAR: load = new LoadUSNode(ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
876 case T_SHORT: load = new LoadSNode (ctl, mem, adr, adr_type, rt->is_int(), mo, control_dependency); break;
877 case T_LONG: load = new LoadLNode (ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency); break;
878 case T_FLOAT: load = new LoadFNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
879 case T_DOUBLE: load = new LoadDNode (ctl, mem, adr, adr_type, rt, mo, control_dependency); break;
880 case T_ADDRESS: load = new LoadPNode (ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency); break;
881 case T_OBJECT:
882 #ifdef _LP64
883 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
884 load = new LoadNNode(ctl, mem, adr, adr_type, rt->make_narrowoop(), mo, control_dependency);
885 } else
886 #endif
887 {
888 assert(!adr->bottom_type()->is_ptr_to_narrowoop() && !adr->bottom_type()->is_ptr_to_narrowklass(), "should have got back a narrow oop");
889 load = new LoadPNode(ctl, mem, adr, adr_type, rt->is_ptr(), mo, control_dependency);
890 }
891 break;
892 default:
893 ShouldNotReachHere();
894 break;
895 }
896 assert(load != NULL, "LoadNode should have been created");
897 if (unaligned) {
898 load->set_unaligned_access();
899 }
900 if (mismatched) {
901 load->set_mismatched_access();
902 }
903 if (unsafe) {
904 load->set_unsafe_access();
905 }
906 load->set_barrier_data(barrier_data);
907 if (load->Opcode() == Op_LoadN) {
908 Node* ld = gvn.transform(load);
909 return new DecodeNNode(ld, ld->bottom_type()->make_ptr());
910 }
911
912 return load;
913 }
914
make_atomic(Node * ctl,Node * mem,Node * adr,const TypePtr * adr_type,const Type * rt,MemOrd mo,ControlDependency control_dependency,bool unaligned,bool mismatched,bool unsafe,uint8_t barrier_data)915 LoadLNode* LoadLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
916 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
917 bool require_atomic = true;
918 LoadLNode* load = new LoadLNode(ctl, mem, adr, adr_type, rt->is_long(), mo, control_dependency, require_atomic);
919 if (unaligned) {
920 load->set_unaligned_access();
921 }
922 if (mismatched) {
923 load->set_mismatched_access();
924 }
925 if (unsafe) {
926 load->set_unsafe_access();
927 }
928 load->set_barrier_data(barrier_data);
929 return load;
930 }
931
make_atomic(Node * ctl,Node * mem,Node * adr,const TypePtr * adr_type,const Type * rt,MemOrd mo,ControlDependency control_dependency,bool unaligned,bool mismatched,bool unsafe,uint8_t barrier_data)932 LoadDNode* LoadDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, const Type* rt, MemOrd mo,
933 ControlDependency control_dependency, bool unaligned, bool mismatched, bool unsafe, uint8_t barrier_data) {
934 bool require_atomic = true;
935 LoadDNode* load = new LoadDNode(ctl, mem, adr, adr_type, rt, mo, control_dependency, require_atomic);
936 if (unaligned) {
937 load->set_unaligned_access();
938 }
939 if (mismatched) {
940 load->set_mismatched_access();
941 }
942 if (unsafe) {
943 load->set_unsafe_access();
944 }
945 load->set_barrier_data(barrier_data);
946 return load;
947 }
948
949
950
951 //------------------------------hash-------------------------------------------
hash() const952 uint LoadNode::hash() const {
953 // unroll addition of interesting fields
954 return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address);
955 }
956
skip_through_membars(Compile::AliasType * atp,const TypeInstPtr * tp,bool eliminate_boxing)957 static bool skip_through_membars(Compile::AliasType* atp, const TypeInstPtr* tp, bool eliminate_boxing) {
958 if ((atp != NULL) && (atp->index() >= Compile::AliasIdxRaw)) {
959 bool non_volatile = (atp->field() != NULL) && !atp->field()->is_volatile();
960 bool is_stable_ary = FoldStableValues &&
961 (tp != NULL) && (tp->isa_aryptr() != NULL) &&
962 tp->isa_aryptr()->is_stable();
963
964 return (eliminate_boxing && non_volatile) || is_stable_ary;
965 }
966
967 return false;
968 }
969
970 // Is the value loaded previously stored by an arraycopy? If so return
971 // a load node that reads from the source array so we may be able to
972 // optimize out the ArrayCopy node later.
can_see_arraycopy_value(Node * st,PhaseGVN * phase) const973 Node* LoadNode::can_see_arraycopy_value(Node* st, PhaseGVN* phase) const {
974 Node* ld_adr = in(MemNode::Address);
975 intptr_t ld_off = 0;
976 AllocateNode* ld_alloc = AllocateNode::Ideal_allocation(ld_adr, phase, ld_off);
977 Node* ac = find_previous_arraycopy(phase, ld_alloc, st, true);
978 if (ac != NULL) {
979 assert(ac->is_ArrayCopy(), "what kind of node can this be?");
980
981 Node* mem = ac->in(TypeFunc::Memory);
982 Node* ctl = ac->in(0);
983 Node* src = ac->in(ArrayCopyNode::Src);
984
985 if (!ac->as_ArrayCopy()->is_clonebasic() && !phase->type(src)->isa_aryptr()) {
986 return NULL;
987 }
988
989 LoadNode* ld = clone()->as_Load();
990 Node* addp = in(MemNode::Address)->clone();
991 if (ac->as_ArrayCopy()->is_clonebasic()) {
992 assert(ld_alloc != NULL, "need an alloc");
993 assert(addp->is_AddP(), "address must be addp");
994 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
995 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Base)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
996 assert(bs->step_over_gc_barrier(addp->in(AddPNode::Address)) == bs->step_over_gc_barrier(ac->in(ArrayCopyNode::Dest)), "strange pattern");
997 addp->set_req(AddPNode::Base, src);
998 addp->set_req(AddPNode::Address, src);
999 } else {
1000 assert(ac->as_ArrayCopy()->is_arraycopy_validated() ||
1001 ac->as_ArrayCopy()->is_copyof_validated() ||
1002 ac->as_ArrayCopy()->is_copyofrange_validated(), "only supported cases");
1003 assert(addp->in(AddPNode::Base) == addp->in(AddPNode::Address), "should be");
1004 addp->set_req(AddPNode::Base, src);
1005 addp->set_req(AddPNode::Address, src);
1006
1007 const TypeAryPtr* ary_t = phase->type(in(MemNode::Address))->isa_aryptr();
1008 BasicType ary_elem = ary_t->klass()->as_array_klass()->element_type()->basic_type();
1009 uint header = arrayOopDesc::base_offset_in_bytes(ary_elem);
1010 uint shift = exact_log2(type2aelembytes(ary_elem));
1011
1012 Node* diff = phase->transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
1013 #ifdef _LP64
1014 diff = phase->transform(new ConvI2LNode(diff));
1015 #endif
1016 diff = phase->transform(new LShiftXNode(diff, phase->intcon(shift)));
1017
1018 Node* offset = phase->transform(new AddXNode(addp->in(AddPNode::Offset), diff));
1019 addp->set_req(AddPNode::Offset, offset);
1020 }
1021 addp = phase->transform(addp);
1022 #ifdef ASSERT
1023 const TypePtr* adr_type = phase->type(addp)->is_ptr();
1024 ld->_adr_type = adr_type;
1025 #endif
1026 ld->set_req(MemNode::Address, addp);
1027 ld->set_req(0, ctl);
1028 ld->set_req(MemNode::Memory, mem);
1029 // load depends on the tests that validate the arraycopy
1030 ld->_control_dependency = UnknownControl;
1031 return ld;
1032 }
1033 return NULL;
1034 }
1035
1036
1037 //---------------------------can_see_stored_value------------------------------
1038 // This routine exists to make sure this set of tests is done the same
1039 // everywhere. We need to make a coordinated change: first LoadNode::Ideal
1040 // will change the graph shape in a way which makes memory alive twice at the
1041 // same time (uses the Oracle model of aliasing), then some
1042 // LoadXNode::Identity will fold things back to the equivalence-class model
1043 // of aliasing.
can_see_stored_value(Node * st,PhaseTransform * phase) const1044 Node* MemNode::can_see_stored_value(Node* st, PhaseTransform* phase) const {
1045 Node* ld_adr = in(MemNode::Address);
1046 intptr_t ld_off = 0;
1047 Node* ld_base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ld_off);
1048 Node* ld_alloc = AllocateNode::Ideal_allocation(ld_base, phase);
1049 const TypeInstPtr* tp = phase->type(ld_adr)->isa_instptr();
1050 Compile::AliasType* atp = (tp != NULL) ? phase->C->alias_type(tp) : NULL;
1051 // This is more general than load from boxing objects.
1052 if (skip_through_membars(atp, tp, phase->C->eliminate_boxing())) {
1053 uint alias_idx = atp->index();
1054 bool final = !atp->is_rewritable();
1055 Node* result = NULL;
1056 Node* current = st;
1057 // Skip through chains of MemBarNodes checking the MergeMems for
1058 // new states for the slice of this load. Stop once any other
1059 // kind of node is encountered. Loads from final memory can skip
1060 // through any kind of MemBar but normal loads shouldn't skip
1061 // through MemBarAcquire since the could allow them to move out of
1062 // a synchronized region.
1063 while (current->is_Proj()) {
1064 int opc = current->in(0)->Opcode();
1065 if ((final && (opc == Op_MemBarAcquire ||
1066 opc == Op_MemBarAcquireLock ||
1067 opc == Op_LoadFence)) ||
1068 opc == Op_MemBarRelease ||
1069 opc == Op_StoreFence ||
1070 opc == Op_MemBarReleaseLock ||
1071 opc == Op_MemBarStoreStore ||
1072 opc == Op_MemBarCPUOrder) {
1073 Node* mem = current->in(0)->in(TypeFunc::Memory);
1074 if (mem->is_MergeMem()) {
1075 MergeMemNode* merge = mem->as_MergeMem();
1076 Node* new_st = merge->memory_at(alias_idx);
1077 if (new_st == merge->base_memory()) {
1078 // Keep searching
1079 current = new_st;
1080 continue;
1081 }
1082 // Save the new memory state for the slice and fall through
1083 // to exit.
1084 result = new_st;
1085 }
1086 }
1087 break;
1088 }
1089 if (result != NULL) {
1090 st = result;
1091 }
1092 }
1093
1094 // Loop around twice in the case Load -> Initialize -> Store.
1095 // (See PhaseIterGVN::add_users_to_worklist, which knows about this case.)
1096 for (int trip = 0; trip <= 1; trip++) {
1097
1098 if (st->is_Store()) {
1099 Node* st_adr = st->in(MemNode::Address);
1100 if (st_adr != ld_adr) {
1101 // Try harder before giving up. Unify base pointers with casts (e.g., raw/non-raw pointers).
1102 intptr_t st_off = 0;
1103 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_off);
1104 if (ld_base == NULL) return NULL;
1105 if (st_base == NULL) return NULL;
1106 if (!ld_base->eqv_uncast(st_base, /*keep_deps=*/true)) return NULL;
1107 if (ld_off != st_off) return NULL;
1108 if (ld_off == Type::OffsetBot) return NULL;
1109 // Same base, same offset.
1110 // Possible improvement for arrays: check index value instead of absolute offset.
1111
1112 // At this point we have proven something like this setup:
1113 // B = << base >>
1114 // L = LoadQ(AddP(Check/CastPP(B), #Off))
1115 // S = StoreQ(AddP( B , #Off), V)
1116 // (Actually, we haven't yet proven the Q's are the same.)
1117 // In other words, we are loading from a casted version of
1118 // the same pointer-and-offset that we stored to.
1119 // Casted version may carry a dependency and it is respected.
1120 // Thus, we are able to replace L by V.
1121 }
1122 // Now prove that we have a LoadQ matched to a StoreQ, for some Q.
1123 if (store_Opcode() != st->Opcode())
1124 return NULL;
1125 return st->in(MemNode::ValueIn);
1126 }
1127
1128 // A load from a freshly-created object always returns zero.
1129 // (This can happen after LoadNode::Ideal resets the load's memory input
1130 // to find_captured_store, which returned InitializeNode::zero_memory.)
1131 if (st->is_Proj() && st->in(0)->is_Allocate() &&
1132 (st->in(0) == ld_alloc) &&
1133 (ld_off >= st->in(0)->as_Allocate()->minimum_header_size())) {
1134 // return a zero value for the load's basic type
1135 // (This is one of the few places where a generic PhaseTransform
1136 // can create new nodes. Think of it as lazily manifesting
1137 // virtually pre-existing constants.)
1138 if (memory_type() != T_VOID) {
1139 if (ReduceBulkZeroing || find_array_copy_clone(phase, ld_alloc, in(MemNode::Memory)) == NULL) {
1140 // If ReduceBulkZeroing is disabled, we need to check if the allocation does not belong to an
1141 // ArrayCopyNode clone. If it does, then we cannot assume zero since the initialization is done
1142 // by the ArrayCopyNode.
1143 return phase->zerocon(memory_type());
1144 }
1145 } else {
1146 // TODO: materialize all-zero vector constant
1147 assert(!isa_Load() || as_Load()->type()->isa_vect(), "");
1148 }
1149 }
1150
1151 // A load from an initialization barrier can match a captured store.
1152 if (st->is_Proj() && st->in(0)->is_Initialize()) {
1153 InitializeNode* init = st->in(0)->as_Initialize();
1154 AllocateNode* alloc = init->allocation();
1155 if ((alloc != NULL) && (alloc == ld_alloc)) {
1156 // examine a captured store value
1157 st = init->find_captured_store(ld_off, memory_size(), phase);
1158 if (st != NULL) {
1159 continue; // take one more trip around
1160 }
1161 }
1162 }
1163
1164 // Load boxed value from result of valueOf() call is input parameter.
1165 if (this->is_Load() && ld_adr->is_AddP() &&
1166 (tp != NULL) && tp->is_ptr_to_boxed_value()) {
1167 intptr_t ignore = 0;
1168 Node* base = AddPNode::Ideal_base_and_offset(ld_adr, phase, ignore);
1169 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1170 base = bs->step_over_gc_barrier(base);
1171 if (base != NULL && base->is_Proj() &&
1172 base->as_Proj()->_con == TypeFunc::Parms &&
1173 base->in(0)->is_CallStaticJava() &&
1174 base->in(0)->as_CallStaticJava()->is_boxing_method()) {
1175 return base->in(0)->in(TypeFunc::Parms);
1176 }
1177 }
1178
1179 break;
1180 }
1181
1182 return NULL;
1183 }
1184
1185 //----------------------is_instance_field_load_with_local_phi------------------
is_instance_field_load_with_local_phi(Node * ctrl)1186 bool LoadNode::is_instance_field_load_with_local_phi(Node* ctrl) {
1187 if( in(Memory)->is_Phi() && in(Memory)->in(0) == ctrl &&
1188 in(Address)->is_AddP() ) {
1189 const TypeOopPtr* t_oop = in(Address)->bottom_type()->isa_oopptr();
1190 // Only instances and boxed values.
1191 if( t_oop != NULL &&
1192 (t_oop->is_ptr_to_boxed_value() ||
1193 t_oop->is_known_instance_field()) &&
1194 t_oop->offset() != Type::OffsetBot &&
1195 t_oop->offset() != Type::OffsetTop) {
1196 return true;
1197 }
1198 }
1199 return false;
1200 }
1201
1202 //------------------------------Identity---------------------------------------
1203 // Loads are identity if previous store is to same address
Identity(PhaseGVN * phase)1204 Node* LoadNode::Identity(PhaseGVN* phase) {
1205 // If the previous store-maker is the right kind of Store, and the store is
1206 // to the same address, then we are equal to the value stored.
1207 Node* mem = in(Memory);
1208 Node* value = can_see_stored_value(mem, phase);
1209 if( value ) {
1210 // byte, short & char stores truncate naturally.
1211 // A load has to load the truncated value which requires
1212 // some sort of masking operation and that requires an
1213 // Ideal call instead of an Identity call.
1214 if (memory_size() < BytesPerInt) {
1215 // If the input to the store does not fit with the load's result type,
1216 // it must be truncated via an Ideal call.
1217 if (!phase->type(value)->higher_equal(phase->type(this)))
1218 return this;
1219 }
1220 // (This works even when value is a Con, but LoadNode::Value
1221 // usually runs first, producing the singleton type of the Con.)
1222 return value;
1223 }
1224
1225 // Search for an existing data phi which was generated before for the same
1226 // instance's field to avoid infinite generation of phis in a loop.
1227 Node *region = mem->in(0);
1228 if (is_instance_field_load_with_local_phi(region)) {
1229 const TypeOopPtr *addr_t = in(Address)->bottom_type()->isa_oopptr();
1230 int this_index = phase->C->get_alias_index(addr_t);
1231 int this_offset = addr_t->offset();
1232 int this_iid = addr_t->instance_id();
1233 if (!addr_t->is_known_instance() &&
1234 addr_t->is_ptr_to_boxed_value()) {
1235 // Use _idx of address base (could be Phi node) for boxed values.
1236 intptr_t ignore = 0;
1237 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1238 if (base == NULL) {
1239 return this;
1240 }
1241 this_iid = base->_idx;
1242 }
1243 const Type* this_type = bottom_type();
1244 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1245 Node* phi = region->fast_out(i);
1246 if (phi->is_Phi() && phi != mem &&
1247 phi->as_Phi()->is_same_inst_field(this_type, (int)mem->_idx, this_iid, this_index, this_offset)) {
1248 return phi;
1249 }
1250 }
1251 }
1252
1253 return this;
1254 }
1255
1256 // Construct an equivalent unsigned load.
convert_to_unsigned_load(PhaseGVN & gvn)1257 Node* LoadNode::convert_to_unsigned_load(PhaseGVN& gvn) {
1258 BasicType bt = T_ILLEGAL;
1259 const Type* rt = NULL;
1260 switch (Opcode()) {
1261 case Op_LoadUB: return this;
1262 case Op_LoadUS: return this;
1263 case Op_LoadB: bt = T_BOOLEAN; rt = TypeInt::UBYTE; break;
1264 case Op_LoadS: bt = T_CHAR; rt = TypeInt::CHAR; break;
1265 default:
1266 assert(false, "no unsigned variant: %s", Name());
1267 return NULL;
1268 }
1269 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1270 raw_adr_type(), rt, bt, _mo, _control_dependency,
1271 is_unaligned_access(), is_mismatched_access());
1272 }
1273
1274 // Construct an equivalent signed load.
convert_to_signed_load(PhaseGVN & gvn)1275 Node* LoadNode::convert_to_signed_load(PhaseGVN& gvn) {
1276 BasicType bt = T_ILLEGAL;
1277 const Type* rt = NULL;
1278 switch (Opcode()) {
1279 case Op_LoadUB: bt = T_BYTE; rt = TypeInt::BYTE; break;
1280 case Op_LoadUS: bt = T_SHORT; rt = TypeInt::SHORT; break;
1281 case Op_LoadB: // fall through
1282 case Op_LoadS: // fall through
1283 case Op_LoadI: // fall through
1284 case Op_LoadL: return this;
1285 default:
1286 assert(false, "no signed variant: %s", Name());
1287 return NULL;
1288 }
1289 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1290 raw_adr_type(), rt, bt, _mo, _control_dependency,
1291 is_unaligned_access(), is_mismatched_access());
1292 }
1293
has_reinterpret_variant(const Type * rt)1294 bool LoadNode::has_reinterpret_variant(const Type* rt) {
1295 BasicType bt = rt->basic_type();
1296 switch (Opcode()) {
1297 case Op_LoadI: return (bt == T_FLOAT);
1298 case Op_LoadL: return (bt == T_DOUBLE);
1299 case Op_LoadF: return (bt == T_INT);
1300 case Op_LoadD: return (bt == T_LONG);
1301
1302 default: return false;
1303 }
1304 }
1305
convert_to_reinterpret_load(PhaseGVN & gvn,const Type * rt)1306 Node* LoadNode::convert_to_reinterpret_load(PhaseGVN& gvn, const Type* rt) {
1307 BasicType bt = rt->basic_type();
1308 assert(has_reinterpret_variant(rt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1309 bool is_mismatched = is_mismatched_access();
1310 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1311 if (raw_type == NULL) {
1312 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1313 }
1314 return LoadNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address),
1315 raw_adr_type(), rt, bt, _mo, _control_dependency,
1316 is_unaligned_access(), is_mismatched);
1317 }
1318
has_reinterpret_variant(const Type * vt)1319 bool StoreNode::has_reinterpret_variant(const Type* vt) {
1320 BasicType bt = vt->basic_type();
1321 switch (Opcode()) {
1322 case Op_StoreI: return (bt == T_FLOAT);
1323 case Op_StoreL: return (bt == T_DOUBLE);
1324 case Op_StoreF: return (bt == T_INT);
1325 case Op_StoreD: return (bt == T_LONG);
1326
1327 default: return false;
1328 }
1329 }
1330
convert_to_reinterpret_store(PhaseGVN & gvn,Node * val,const Type * vt)1331 Node* StoreNode::convert_to_reinterpret_store(PhaseGVN& gvn, Node* val, const Type* vt) {
1332 BasicType bt = vt->basic_type();
1333 assert(has_reinterpret_variant(vt), "no reinterpret variant: %s %s", Name(), type2name(bt));
1334 StoreNode* st = StoreNode::make(gvn, in(MemNode::Control), in(MemNode::Memory), in(MemNode::Address), raw_adr_type(), val, bt, _mo);
1335
1336 bool is_mismatched = is_mismatched_access();
1337 const TypeRawPtr* raw_type = gvn.type(in(MemNode::Memory))->isa_rawptr();
1338 if (raw_type == NULL) {
1339 is_mismatched = true; // conservatively match all non-raw accesses as mismatched
1340 }
1341 if (is_mismatched) {
1342 st->set_mismatched_access();
1343 }
1344 return st;
1345 }
1346
1347 // We're loading from an object which has autobox behaviour.
1348 // If this object is result of a valueOf call we'll have a phi
1349 // merging a newly allocated object and a load from the cache.
1350 // We want to replace this load with the original incoming
1351 // argument to the valueOf call.
eliminate_autobox(PhaseGVN * phase)1352 Node* LoadNode::eliminate_autobox(PhaseGVN* phase) {
1353 assert(phase->C->eliminate_boxing(), "sanity");
1354 intptr_t ignore = 0;
1355 Node* base = AddPNode::Ideal_base_and_offset(in(Address), phase, ignore);
1356 if ((base == NULL) || base->is_Phi()) {
1357 // Push the loads from the phi that comes from valueOf up
1358 // through it to allow elimination of the loads and the recovery
1359 // of the original value. It is done in split_through_phi().
1360 return NULL;
1361 } else if (base->is_Load() ||
1362 (base->is_DecodeN() && base->in(1)->is_Load())) {
1363 // Eliminate the load of boxed value for integer types from the cache
1364 // array by deriving the value from the index into the array.
1365 // Capture the offset of the load and then reverse the computation.
1366
1367 // Get LoadN node which loads a boxing object from 'cache' array.
1368 if (base->is_DecodeN()) {
1369 base = base->in(1);
1370 }
1371 if (!base->in(Address)->is_AddP()) {
1372 return NULL; // Complex address
1373 }
1374 AddPNode* address = base->in(Address)->as_AddP();
1375 Node* cache_base = address->in(AddPNode::Base);
1376 if ((cache_base != NULL) && cache_base->is_DecodeN()) {
1377 // Get ConP node which is static 'cache' field.
1378 cache_base = cache_base->in(1);
1379 }
1380 if ((cache_base != NULL) && cache_base->is_Con()) {
1381 const TypeAryPtr* base_type = cache_base->bottom_type()->isa_aryptr();
1382 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1383 Node* elements[4];
1384 int shift = exact_log2(type2aelembytes(T_OBJECT));
1385 int count = address->unpack_offsets(elements, ARRAY_SIZE(elements));
1386 if (count > 0 && elements[0]->is_Con() &&
1387 (count == 1 ||
1388 (count == 2 && elements[1]->Opcode() == Op_LShiftX &&
1389 elements[1]->in(2) == phase->intcon(shift)))) {
1390 ciObjArray* array = base_type->const_oop()->as_obj_array();
1391 // Fetch the box object cache[0] at the base of the array and get its value
1392 ciInstance* box = array->obj_at(0)->as_instance();
1393 ciInstanceKlass* ik = box->klass()->as_instance_klass();
1394 assert(ik->is_box_klass(), "sanity");
1395 assert(ik->nof_nonstatic_fields() == 1, "change following code");
1396 if (ik->nof_nonstatic_fields() == 1) {
1397 // This should be true nonstatic_field_at requires calling
1398 // nof_nonstatic_fields so check it anyway
1399 ciConstant c = box->field_value(ik->nonstatic_field_at(0));
1400 BasicType bt = c.basic_type();
1401 // Only integer types have boxing cache.
1402 assert(bt == T_BOOLEAN || bt == T_CHAR ||
1403 bt == T_BYTE || bt == T_SHORT ||
1404 bt == T_INT || bt == T_LONG, "wrong type = %s", type2name(bt));
1405 jlong cache_low = (bt == T_LONG) ? c.as_long() : c.as_int();
1406 if (cache_low != (int)cache_low) {
1407 return NULL; // should not happen since cache is array indexed by value
1408 }
1409 jlong offset = arrayOopDesc::base_offset_in_bytes(T_OBJECT) - (cache_low << shift);
1410 if (offset != (int)offset) {
1411 return NULL; // should not happen since cache is array indexed by value
1412 }
1413 // Add up all the offsets making of the address of the load
1414 Node* result = elements[0];
1415 for (int i = 1; i < count; i++) {
1416 result = phase->transform(new AddXNode(result, elements[i]));
1417 }
1418 // Remove the constant offset from the address and then
1419 result = phase->transform(new AddXNode(result, phase->MakeConX(-(int)offset)));
1420 // remove the scaling of the offset to recover the original index.
1421 if (result->Opcode() == Op_LShiftX && result->in(2) == phase->intcon(shift)) {
1422 // Peel the shift off directly but wrap it in a dummy node
1423 // since Ideal can't return existing nodes
1424 result = new RShiftXNode(result->in(1), phase->intcon(0));
1425 } else if (result->is_Add() && result->in(2)->is_Con() &&
1426 result->in(1)->Opcode() == Op_LShiftX &&
1427 result->in(1)->in(2) == phase->intcon(shift)) {
1428 // We can't do general optimization: ((X<<Z) + Y) >> Z ==> X + (Y>>Z)
1429 // but for boxing cache access we know that X<<Z will not overflow
1430 // (there is range check) so we do this optimizatrion by hand here.
1431 Node* add_con = new RShiftXNode(result->in(2), phase->intcon(shift));
1432 result = new AddXNode(result->in(1)->in(1), phase->transform(add_con));
1433 } else {
1434 result = new RShiftXNode(result, phase->intcon(shift));
1435 }
1436 #ifdef _LP64
1437 if (bt != T_LONG) {
1438 result = new ConvL2INode(phase->transform(result));
1439 }
1440 #else
1441 if (bt == T_LONG) {
1442 result = new ConvI2LNode(phase->transform(result));
1443 }
1444 #endif
1445 // Boxing/unboxing can be done from signed & unsigned loads (e.g. LoadUB -> ... -> LoadB pair).
1446 // Need to preserve unboxing load type if it is unsigned.
1447 switch(this->Opcode()) {
1448 case Op_LoadUB:
1449 result = new AndINode(phase->transform(result), phase->intcon(0xFF));
1450 break;
1451 case Op_LoadUS:
1452 result = new AndINode(phase->transform(result), phase->intcon(0xFFFF));
1453 break;
1454 }
1455 return result;
1456 }
1457 }
1458 }
1459 }
1460 }
1461 return NULL;
1462 }
1463
stable_phi(PhiNode * phi,PhaseGVN * phase)1464 static bool stable_phi(PhiNode* phi, PhaseGVN *phase) {
1465 Node* region = phi->in(0);
1466 if (region == NULL) {
1467 return false; // Wait stable graph
1468 }
1469 uint cnt = phi->req();
1470 for (uint i = 1; i < cnt; i++) {
1471 Node* rc = region->in(i);
1472 if (rc == NULL || phase->type(rc) == Type::TOP)
1473 return false; // Wait stable graph
1474 Node* in = phi->in(i);
1475 if (in == NULL || phase->type(in) == Type::TOP)
1476 return false; // Wait stable graph
1477 }
1478 return true;
1479 }
1480 //------------------------------split_through_phi------------------------------
1481 // Split instance or boxed field load through Phi.
split_through_phi(PhaseGVN * phase)1482 Node *LoadNode::split_through_phi(PhaseGVN *phase) {
1483 Node* mem = in(Memory);
1484 Node* address = in(Address);
1485 const TypeOopPtr *t_oop = phase->type(address)->isa_oopptr();
1486
1487 assert((t_oop != NULL) &&
1488 (t_oop->is_known_instance_field() ||
1489 t_oop->is_ptr_to_boxed_value()), "invalide conditions");
1490
1491 Compile* C = phase->C;
1492 intptr_t ignore = 0;
1493 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1494 bool base_is_phi = (base != NULL) && base->is_Phi();
1495 bool load_boxed_values = t_oop->is_ptr_to_boxed_value() && C->aggressive_unboxing() &&
1496 (base != NULL) && (base == address->in(AddPNode::Base)) &&
1497 phase->type(base)->higher_equal(TypePtr::NOTNULL);
1498
1499 if (!((mem->is_Phi() || base_is_phi) &&
1500 (load_boxed_values || t_oop->is_known_instance_field()))) {
1501 return NULL; // memory is not Phi
1502 }
1503
1504 if (mem->is_Phi()) {
1505 if (!stable_phi(mem->as_Phi(), phase)) {
1506 return NULL; // Wait stable graph
1507 }
1508 uint cnt = mem->req();
1509 // Check for loop invariant memory.
1510 if (cnt == 3) {
1511 for (uint i = 1; i < cnt; i++) {
1512 Node* in = mem->in(i);
1513 Node* m = optimize_memory_chain(in, t_oop, this, phase);
1514 if (m == mem) {
1515 if (i == 1) {
1516 // if the first edge was a loop, check second edge too.
1517 // If both are replaceable - we are in an infinite loop
1518 Node *n = optimize_memory_chain(mem->in(2), t_oop, this, phase);
1519 if (n == mem) {
1520 break;
1521 }
1522 }
1523 set_req(Memory, mem->in(cnt - i));
1524 return this; // made change
1525 }
1526 }
1527 }
1528 }
1529 if (base_is_phi) {
1530 if (!stable_phi(base->as_Phi(), phase)) {
1531 return NULL; // Wait stable graph
1532 }
1533 uint cnt = base->req();
1534 // Check for loop invariant memory.
1535 if (cnt == 3) {
1536 for (uint i = 1; i < cnt; i++) {
1537 if (base->in(i) == base) {
1538 return NULL; // Wait stable graph
1539 }
1540 }
1541 }
1542 }
1543
1544 // Split through Phi (see original code in loopopts.cpp).
1545 assert(C->have_alias_type(t_oop), "instance should have alias type");
1546
1547 // Do nothing here if Identity will find a value
1548 // (to avoid infinite chain of value phis generation).
1549 if (this != Identity(phase)) {
1550 return NULL;
1551 }
1552
1553 // Select Region to split through.
1554 Node* region;
1555 if (!base_is_phi) {
1556 assert(mem->is_Phi(), "sanity");
1557 region = mem->in(0);
1558 // Skip if the region dominates some control edge of the address.
1559 if (!MemNode::all_controls_dominate(address, region))
1560 return NULL;
1561 } else if (!mem->is_Phi()) {
1562 assert(base_is_phi, "sanity");
1563 region = base->in(0);
1564 // Skip if the region dominates some control edge of the memory.
1565 if (!MemNode::all_controls_dominate(mem, region))
1566 return NULL;
1567 } else if (base->in(0) != mem->in(0)) {
1568 assert(base_is_phi && mem->is_Phi(), "sanity");
1569 if (MemNode::all_controls_dominate(mem, base->in(0))) {
1570 region = base->in(0);
1571 } else if (MemNode::all_controls_dominate(address, mem->in(0))) {
1572 region = mem->in(0);
1573 } else {
1574 return NULL; // complex graph
1575 }
1576 } else {
1577 assert(base->in(0) == mem->in(0), "sanity");
1578 region = mem->in(0);
1579 }
1580
1581 const Type* this_type = this->bottom_type();
1582 int this_index = C->get_alias_index(t_oop);
1583 int this_offset = t_oop->offset();
1584 int this_iid = t_oop->instance_id();
1585 if (!t_oop->is_known_instance() && load_boxed_values) {
1586 // Use _idx of address base for boxed values.
1587 this_iid = base->_idx;
1588 }
1589 PhaseIterGVN* igvn = phase->is_IterGVN();
1590 Node* phi = new PhiNode(region, this_type, NULL, mem->_idx, this_iid, this_index, this_offset);
1591 for (uint i = 1; i < region->req(); i++) {
1592 Node* x;
1593 Node* the_clone = NULL;
1594 Node* in = region->in(i);
1595 if (region->is_CountedLoop() && region->as_Loop()->is_strip_mined() && i == LoopNode::EntryControl &&
1596 in != NULL && in->is_OuterStripMinedLoop()) {
1597 // No node should go in the outer strip mined loop
1598 in = in->in(LoopNode::EntryControl);
1599 }
1600 if (in == NULL || in == C->top()) {
1601 x = C->top(); // Dead path? Use a dead data op
1602 } else {
1603 x = this->clone(); // Else clone up the data op
1604 the_clone = x; // Remember for possible deletion.
1605 // Alter data node to use pre-phi inputs
1606 if (this->in(0) == region) {
1607 x->set_req(0, in);
1608 } else {
1609 x->set_req(0, NULL);
1610 }
1611 if (mem->is_Phi() && (mem->in(0) == region)) {
1612 x->set_req(Memory, mem->in(i)); // Use pre-Phi input for the clone.
1613 }
1614 if (address->is_Phi() && address->in(0) == region) {
1615 x->set_req(Address, address->in(i)); // Use pre-Phi input for the clone
1616 }
1617 if (base_is_phi && (base->in(0) == region)) {
1618 Node* base_x = base->in(i); // Clone address for loads from boxed objects.
1619 Node* adr_x = phase->transform(new AddPNode(base_x,base_x,address->in(AddPNode::Offset)));
1620 x->set_req(Address, adr_x);
1621 }
1622 }
1623 // Check for a 'win' on some paths
1624 const Type *t = x->Value(igvn);
1625
1626 bool singleton = t->singleton();
1627
1628 // See comments in PhaseIdealLoop::split_thru_phi().
1629 if (singleton && t == Type::TOP) {
1630 singleton &= region->is_Loop() && (i != LoopNode::EntryControl);
1631 }
1632
1633 if (singleton) {
1634 x = igvn->makecon(t);
1635 } else {
1636 // We now call Identity to try to simplify the cloned node.
1637 // Note that some Identity methods call phase->type(this).
1638 // Make sure that the type array is big enough for
1639 // our new node, even though we may throw the node away.
1640 // (This tweaking with igvn only works because x is a new node.)
1641 igvn->set_type(x, t);
1642 // If x is a TypeNode, capture any more-precise type permanently into Node
1643 // otherwise it will be not updated during igvn->transform since
1644 // igvn->type(x) is set to x->Value() already.
1645 x->raise_bottom_type(t);
1646 Node* y = x->Identity(igvn);
1647 if (y != x) {
1648 x = y;
1649 } else {
1650 y = igvn->hash_find_insert(x);
1651 if (y) {
1652 x = y;
1653 } else {
1654 // Else x is a new node we are keeping
1655 // We do not need register_new_node_with_optimizer
1656 // because set_type has already been called.
1657 igvn->_worklist.push(x);
1658 }
1659 }
1660 }
1661 if (x != the_clone && the_clone != NULL) {
1662 igvn->remove_dead_node(the_clone);
1663 }
1664 phi->set_req(i, x);
1665 }
1666 // Record Phi
1667 igvn->register_new_node_with_optimizer(phi);
1668 return phi;
1669 }
1670
is_new_object_mark_load(PhaseGVN * phase) const1671 AllocateNode* LoadNode::is_new_object_mark_load(PhaseGVN *phase) const {
1672 if (Opcode() == Op_LoadX) {
1673 Node* address = in(MemNode::Address);
1674 AllocateNode* alloc = AllocateNode::Ideal_allocation(address, phase);
1675 Node* mem = in(MemNode::Memory);
1676 if (alloc != NULL && mem->is_Proj() &&
1677 mem->in(0) != NULL &&
1678 mem->in(0) == alloc->initialization() &&
1679 alloc->initialization()->proj_out_or_null(0) != NULL) {
1680 return alloc;
1681 }
1682 }
1683 return NULL;
1684 }
1685
1686
1687 //------------------------------Ideal------------------------------------------
1688 // If the load is from Field memory and the pointer is non-null, it might be possible to
1689 // zero out the control input.
1690 // If the offset is constant and the base is an object allocation,
1691 // try to hook me up to the exact initializing store.
Ideal(PhaseGVN * phase,bool can_reshape)1692 Node *LoadNode::Ideal(PhaseGVN *phase, bool can_reshape) {
1693 Node* p = MemNode::Ideal_common(phase, can_reshape);
1694 if (p) return (p == NodeSentinel) ? NULL : p;
1695
1696 Node* ctrl = in(MemNode::Control);
1697 Node* address = in(MemNode::Address);
1698 bool progress = false;
1699
1700 bool addr_mark = ((phase->type(address)->isa_oopptr() || phase->type(address)->isa_narrowoop()) &&
1701 phase->type(address)->is_ptr()->offset() == oopDesc::mark_offset_in_bytes());
1702
1703 // Skip up past a SafePoint control. Cannot do this for Stores because
1704 // pointer stores & cardmarks must stay on the same side of a SafePoint.
1705 if( ctrl != NULL && ctrl->Opcode() == Op_SafePoint &&
1706 phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw &&
1707 !addr_mark &&
1708 (depends_only_on_test() || has_unknown_control_dependency())) {
1709 ctrl = ctrl->in(0);
1710 set_req(MemNode::Control,ctrl);
1711 progress = true;
1712 }
1713
1714 intptr_t ignore = 0;
1715 Node* base = AddPNode::Ideal_base_and_offset(address, phase, ignore);
1716 if (base != NULL
1717 && phase->C->get_alias_index(phase->type(address)->is_ptr()) != Compile::AliasIdxRaw) {
1718 // Check for useless control edge in some common special cases
1719 if (in(MemNode::Control) != NULL
1720 && can_remove_control()
1721 && phase->type(base)->higher_equal(TypePtr::NOTNULL)
1722 && all_controls_dominate(base, phase->C->start())) {
1723 // A method-invariant, non-null address (constant or 'this' argument).
1724 set_req(MemNode::Control, NULL);
1725 progress = true;
1726 }
1727 }
1728
1729 Node* mem = in(MemNode::Memory);
1730 const TypePtr *addr_t = phase->type(address)->isa_ptr();
1731
1732 if (can_reshape && (addr_t != NULL)) {
1733 // try to optimize our memory input
1734 Node* opt_mem = MemNode::optimize_memory_chain(mem, addr_t, this, phase);
1735 if (opt_mem != mem) {
1736 set_req(MemNode::Memory, opt_mem);
1737 if (phase->type( opt_mem ) == Type::TOP) return NULL;
1738 return this;
1739 }
1740 const TypeOopPtr *t_oop = addr_t->isa_oopptr();
1741 if ((t_oop != NULL) &&
1742 (t_oop->is_known_instance_field() ||
1743 t_oop->is_ptr_to_boxed_value())) {
1744 PhaseIterGVN *igvn = phase->is_IterGVN();
1745 if (igvn != NULL && igvn->_worklist.member(opt_mem)) {
1746 // Delay this transformation until memory Phi is processed.
1747 igvn->_worklist.push(this);
1748 return NULL;
1749 }
1750 // Split instance field load through Phi.
1751 Node* result = split_through_phi(phase);
1752 if (result != NULL) return result;
1753
1754 if (t_oop->is_ptr_to_boxed_value()) {
1755 Node* result = eliminate_autobox(phase);
1756 if (result != NULL) return result;
1757 }
1758 }
1759 }
1760
1761 // Is there a dominating load that loads the same value? Leave
1762 // anything that is not a load of a field/array element (like
1763 // barriers etc.) alone
1764 if (in(0) != NULL && !adr_type()->isa_rawptr() && can_reshape) {
1765 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
1766 Node *use = mem->fast_out(i);
1767 if (use != this &&
1768 use->Opcode() == Opcode() &&
1769 use->in(0) != NULL &&
1770 use->in(0) != in(0) &&
1771 use->in(Address) == in(Address)) {
1772 Node* ctl = in(0);
1773 for (int i = 0; i < 10 && ctl != NULL; i++) {
1774 ctl = IfNode::up_one_dom(ctl);
1775 if (ctl == use->in(0)) {
1776 set_req(0, use->in(0));
1777 return this;
1778 }
1779 }
1780 }
1781 }
1782 }
1783
1784 // Check for prior store with a different base or offset; make Load
1785 // independent. Skip through any number of them. Bail out if the stores
1786 // are in an endless dead cycle and report no progress. This is a key
1787 // transform for Reflection. However, if after skipping through the Stores
1788 // we can't then fold up against a prior store do NOT do the transform as
1789 // this amounts to using the 'Oracle' model of aliasing. It leaves the same
1790 // array memory alive twice: once for the hoisted Load and again after the
1791 // bypassed Store. This situation only works if EVERYBODY who does
1792 // anti-dependence work knows how to bypass. I.e. we need all
1793 // anti-dependence checks to ask the same Oracle. Right now, that Oracle is
1794 // the alias index stuff. So instead, peek through Stores and IFF we can
1795 // fold up, do so.
1796 Node* prev_mem = find_previous_store(phase);
1797 if (prev_mem != NULL) {
1798 Node* value = can_see_arraycopy_value(prev_mem, phase);
1799 if (value != NULL) {
1800 return value;
1801 }
1802 }
1803 // Steps (a), (b): Walk past independent stores to find an exact match.
1804 if (prev_mem != NULL && prev_mem != in(MemNode::Memory)) {
1805 // (c) See if we can fold up on the spot, but don't fold up here.
1806 // Fold-up might require truncation (for LoadB/LoadS/LoadUS) or
1807 // just return a prior value, which is done by Identity calls.
1808 if (can_see_stored_value(prev_mem, phase)) {
1809 // Make ready for step (d):
1810 set_req(MemNode::Memory, prev_mem);
1811 return this;
1812 }
1813 }
1814
1815 AllocateNode* alloc = is_new_object_mark_load(phase);
1816 if (alloc != NULL && alloc->Opcode() == Op_Allocate && UseBiasedLocking) {
1817 InitializeNode* init = alloc->initialization();
1818 Node* control = init->proj_out(0);
1819 return alloc->make_ideal_mark(phase, address, control, mem);
1820 }
1821
1822 return progress ? this : NULL;
1823 }
1824
1825 // Helper to recognize certain Klass fields which are invariant across
1826 // some group of array types (e.g., int[] or all T[] where T < Object).
1827 const Type*
load_array_final_field(const TypeKlassPtr * tkls,ciKlass * klass) const1828 LoadNode::load_array_final_field(const TypeKlassPtr *tkls,
1829 ciKlass* klass) const {
1830 if (tkls->offset() == in_bytes(Klass::modifier_flags_offset())) {
1831 // The field is Klass::_modifier_flags. Return its (constant) value.
1832 // (Folds up the 2nd indirection in aClassConstant.getModifiers().)
1833 assert(this->Opcode() == Op_LoadI, "must load an int from _modifier_flags");
1834 return TypeInt::make(klass->modifier_flags());
1835 }
1836 if (tkls->offset() == in_bytes(Klass::access_flags_offset())) {
1837 // The field is Klass::_access_flags. Return its (constant) value.
1838 // (Folds up the 2nd indirection in Reflection.getClassAccessFlags(aClassConstant).)
1839 assert(this->Opcode() == Op_LoadI, "must load an int from _access_flags");
1840 return TypeInt::make(klass->access_flags());
1841 }
1842 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())) {
1843 // The field is Klass::_layout_helper. Return its constant value if known.
1844 assert(this->Opcode() == Op_LoadI, "must load an int from _layout_helper");
1845 return TypeInt::make(klass->layout_helper());
1846 }
1847
1848 // No match.
1849 return NULL;
1850 }
1851
1852 //------------------------------Value-----------------------------------------
Value(PhaseGVN * phase) const1853 const Type* LoadNode::Value(PhaseGVN* phase) const {
1854 // Either input is TOP ==> the result is TOP
1855 Node* mem = in(MemNode::Memory);
1856 const Type *t1 = phase->type(mem);
1857 if (t1 == Type::TOP) return Type::TOP;
1858 Node* adr = in(MemNode::Address);
1859 const TypePtr* tp = phase->type(adr)->isa_ptr();
1860 if (tp == NULL || tp->empty()) return Type::TOP;
1861 int off = tp->offset();
1862 assert(off != Type::OffsetTop, "case covered by TypePtr::empty");
1863 Compile* C = phase->C;
1864
1865 // Try to guess loaded type from pointer type
1866 if (tp->isa_aryptr()) {
1867 const TypeAryPtr* ary = tp->is_aryptr();
1868 const Type* t = ary->elem();
1869
1870 // Determine whether the reference is beyond the header or not, by comparing
1871 // the offset against the offset of the start of the array's data.
1872 // Different array types begin at slightly different offsets (12 vs. 16).
1873 // We choose T_BYTE as an example base type that is least restrictive
1874 // as to alignment, which will therefore produce the smallest
1875 // possible base offset.
1876 const int min_base_off = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1877 const bool off_beyond_header = (off >= min_base_off);
1878
1879 // Try to constant-fold a stable array element.
1880 if (FoldStableValues && !is_mismatched_access() && ary->is_stable()) {
1881 // Make sure the reference is not into the header and the offset is constant
1882 ciObject* aobj = ary->const_oop();
1883 if (aobj != NULL && off_beyond_header && adr->is_AddP() && off != Type::OffsetBot) {
1884 int stable_dimension = (ary->stable_dimension() > 0 ? ary->stable_dimension() - 1 : 0);
1885 const Type* con_type = Type::make_constant_from_array_element(aobj->as_array(), off,
1886 stable_dimension,
1887 memory_type(), is_unsigned());
1888 if (con_type != NULL) {
1889 return con_type;
1890 }
1891 }
1892 }
1893
1894 // Don't do this for integer types. There is only potential profit if
1895 // the element type t is lower than _type; that is, for int types, if _type is
1896 // more restrictive than t. This only happens here if one is short and the other
1897 // char (both 16 bits), and in those cases we've made an intentional decision
1898 // to use one kind of load over the other. See AndINode::Ideal and 4965907.
1899 // Also, do not try to narrow the type for a LoadKlass, regardless of offset.
1900 //
1901 // Yes, it is possible to encounter an expression like (LoadKlass p1:(AddP x x 8))
1902 // where the _gvn.type of the AddP is wider than 8. This occurs when an earlier
1903 // copy p0 of (AddP x x 8) has been proven equal to p1, and the p0 has been
1904 // subsumed by p1. If p1 is on the worklist but has not yet been re-transformed,
1905 // it is possible that p1 will have a type like Foo*[int+]:NotNull*+any.
1906 // In fact, that could have been the original type of p1, and p1 could have
1907 // had an original form like p1:(AddP x x (LShiftL quux 3)), where the
1908 // expression (LShiftL quux 3) independently optimized to the constant 8.
1909 if ((t->isa_int() == NULL) && (t->isa_long() == NULL)
1910 && (_type->isa_vect() == NULL)
1911 && Opcode() != Op_LoadKlass && Opcode() != Op_LoadNKlass) {
1912 // t might actually be lower than _type, if _type is a unique
1913 // concrete subclass of abstract class t.
1914 if (off_beyond_header || off == Type::OffsetBot) { // is the offset beyond the header?
1915 const Type* jt = t->join_speculative(_type);
1916 // In any case, do not allow the join, per se, to empty out the type.
1917 if (jt->empty() && !t->empty()) {
1918 // This can happen if a interface-typed array narrows to a class type.
1919 jt = _type;
1920 }
1921 #ifdef ASSERT
1922 if (phase->C->eliminate_boxing() && adr->is_AddP()) {
1923 // The pointers in the autobox arrays are always non-null
1924 Node* base = adr->in(AddPNode::Base);
1925 if ((base != NULL) && base->is_DecodeN()) {
1926 // Get LoadN node which loads IntegerCache.cache field
1927 base = base->in(1);
1928 }
1929 if ((base != NULL) && base->is_Con()) {
1930 const TypeAryPtr* base_type = base->bottom_type()->isa_aryptr();
1931 if ((base_type != NULL) && base_type->is_autobox_cache()) {
1932 // It could be narrow oop
1933 assert(jt->make_ptr()->ptr() == TypePtr::NotNull,"sanity");
1934 }
1935 }
1936 }
1937 #endif
1938 return jt;
1939 }
1940 }
1941 } else if (tp->base() == Type::InstPtr) {
1942 assert( off != Type::OffsetBot ||
1943 // arrays can be cast to Objects
1944 tp->is_oopptr()->klass()->is_java_lang_Object() ||
1945 // unsafe field access may not have a constant offset
1946 C->has_unsafe_access(),
1947 "Field accesses must be precise" );
1948 // For oop loads, we expect the _type to be precise.
1949
1950 // Optimize loads from constant fields.
1951 const TypeInstPtr* tinst = tp->is_instptr();
1952 ciObject* const_oop = tinst->const_oop();
1953 if (!is_mismatched_access() && off != Type::OffsetBot && const_oop != NULL && const_oop->is_instance()) {
1954 const Type* con_type = Type::make_constant_from_field(const_oop->as_instance(), off, is_unsigned(), memory_type());
1955 if (con_type != NULL) {
1956 return con_type;
1957 }
1958 }
1959 } else if (tp->base() == Type::KlassPtr) {
1960 assert( off != Type::OffsetBot ||
1961 // arrays can be cast to Objects
1962 tp->is_klassptr()->klass()->is_java_lang_Object() ||
1963 // also allow array-loading from the primary supertype
1964 // array during subtype checks
1965 Opcode() == Op_LoadKlass,
1966 "Field accesses must be precise" );
1967 // For klass/static loads, we expect the _type to be precise
1968 } else if (tp->base() == Type::RawPtr && adr->is_Load() && off == 0) {
1969 /* With mirrors being an indirect in the Klass*
1970 * the VM is now using two loads. LoadKlass(LoadP(LoadP(Klass, mirror_offset), zero_offset))
1971 * The LoadP from the Klass has a RawPtr type (see LibraryCallKit::load_mirror_from_klass).
1972 *
1973 * So check the type and klass of the node before the LoadP.
1974 */
1975 Node* adr2 = adr->in(MemNode::Address);
1976 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
1977 if (tkls != NULL && !StressReflectiveCode) {
1978 ciKlass* klass = tkls->klass();
1979 if (klass->is_loaded() && tkls->klass_is_exact() && tkls->offset() == in_bytes(Klass::java_mirror_offset())) {
1980 assert(adr->Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1981 assert(Opcode() == Op_LoadP, "must load an oop from _java_mirror");
1982 return TypeInstPtr::make(klass->java_mirror());
1983 }
1984 }
1985 }
1986
1987 const TypeKlassPtr *tkls = tp->isa_klassptr();
1988 if (tkls != NULL && !StressReflectiveCode) {
1989 ciKlass* klass = tkls->klass();
1990 if (klass->is_loaded() && tkls->klass_is_exact()) {
1991 // We are loading a field from a Klass metaobject whose identity
1992 // is known at compile time (the type is "exact" or "precise").
1993 // Check for fields we know are maintained as constants by the VM.
1994 if (tkls->offset() == in_bytes(Klass::super_check_offset_offset())) {
1995 // The field is Klass::_super_check_offset. Return its (constant) value.
1996 // (Folds up type checking code.)
1997 assert(Opcode() == Op_LoadI, "must load an int from _super_check_offset");
1998 return TypeInt::make(klass->super_check_offset());
1999 }
2000 // Compute index into primary_supers array
2001 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2002 // Check for overflowing; use unsigned compare to handle the negative case.
2003 if( depth < ciKlass::primary_super_limit() ) {
2004 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2005 // (Folds up type checking code.)
2006 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2007 ciKlass *ss = klass->super_of_depth(depth);
2008 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
2009 }
2010 const Type* aift = load_array_final_field(tkls, klass);
2011 if (aift != NULL) return aift;
2012 }
2013
2014 // We can still check if we are loading from the primary_supers array at a
2015 // shallow enough depth. Even though the klass is not exact, entries less
2016 // than or equal to its super depth are correct.
2017 if (klass->is_loaded() ) {
2018 ciType *inner = klass;
2019 while( inner->is_obj_array_klass() )
2020 inner = inner->as_obj_array_klass()->base_element_type();
2021 if( inner->is_instance_klass() &&
2022 !inner->as_instance_klass()->flags().is_interface() ) {
2023 // Compute index into primary_supers array
2024 juint depth = (tkls->offset() - in_bytes(Klass::primary_supers_offset())) / sizeof(Klass*);
2025 // Check for overflowing; use unsigned compare to handle the negative case.
2026 if( depth < ciKlass::primary_super_limit() &&
2027 depth <= klass->super_depth() ) { // allow self-depth checks to handle self-check case
2028 // The field is an element of Klass::_primary_supers. Return its (constant) value.
2029 // (Folds up type checking code.)
2030 assert(Opcode() == Op_LoadKlass, "must load a klass from _primary_supers");
2031 ciKlass *ss = klass->super_of_depth(depth);
2032 return ss ? TypeKlassPtr::make(ss) : TypePtr::NULL_PTR;
2033 }
2034 }
2035 }
2036
2037 // If the type is enough to determine that the thing is not an array,
2038 // we can give the layout_helper a positive interval type.
2039 // This will help short-circuit some reflective code.
2040 if (tkls->offset() == in_bytes(Klass::layout_helper_offset())
2041 && !klass->is_array_klass() // not directly typed as an array
2042 && !klass->is_interface() // specifically not Serializable & Cloneable
2043 && !klass->is_java_lang_Object() // not the supertype of all T[]
2044 ) {
2045 // Note: When interfaces are reliable, we can narrow the interface
2046 // test to (klass != Serializable && klass != Cloneable).
2047 assert(Opcode() == Op_LoadI, "must load an int from _layout_helper");
2048 jint min_size = Klass::instance_layout_helper(oopDesc::header_size(), false);
2049 // The key property of this type is that it folds up tests
2050 // for array-ness, since it proves that the layout_helper is positive.
2051 // Thus, a generic value like the basic object layout helper works fine.
2052 return TypeInt::make(min_size, max_jint, Type::WidenMin);
2053 }
2054 }
2055
2056 // If we are loading from a freshly-allocated object, produce a zero,
2057 // if the load is provably beyond the header of the object.
2058 // (Also allow a variable load from a fresh array to produce zero.)
2059 const TypeOopPtr *tinst = tp->isa_oopptr();
2060 bool is_instance = (tinst != NULL) && tinst->is_known_instance_field();
2061 bool is_boxed_value = (tinst != NULL) && tinst->is_ptr_to_boxed_value();
2062 if (ReduceFieldZeroing || is_instance || is_boxed_value) {
2063 Node* value = can_see_stored_value(mem,phase);
2064 if (value != NULL && value->is_Con()) {
2065 assert(value->bottom_type()->higher_equal(_type),"sanity");
2066 return value->bottom_type();
2067 }
2068 }
2069
2070 bool is_vect = (_type->isa_vect() != NULL);
2071 if (is_instance && !is_vect) {
2072 // If we have an instance type and our memory input is the
2073 // programs's initial memory state, there is no matching store,
2074 // so just return a zero of the appropriate type -
2075 // except if it is vectorized - then we have no zero constant.
2076 Node *mem = in(MemNode::Memory);
2077 if (mem->is_Parm() && mem->in(0)->is_Start()) {
2078 assert(mem->as_Parm()->_con == TypeFunc::Memory, "must be memory Parm");
2079 return Type::get_zero_type(_type->basic_type());
2080 }
2081 }
2082
2083 Node* alloc = is_new_object_mark_load(phase);
2084 if (alloc != NULL && !(alloc->Opcode() == Op_Allocate && UseBiasedLocking)) {
2085 return TypeX::make(markWord::prototype().value());
2086 }
2087
2088 return _type;
2089 }
2090
2091 //------------------------------match_edge-------------------------------------
2092 // Do we Match on this edge index or not? Match only the address.
match_edge(uint idx) const2093 uint LoadNode::match_edge(uint idx) const {
2094 return idx == MemNode::Address;
2095 }
2096
2097 //--------------------------LoadBNode::Ideal--------------------------------------
2098 //
2099 // If the previous store is to the same address as this load,
2100 // and the value stored was larger than a byte, replace this load
2101 // with the value stored truncated to a byte. If no truncation is
2102 // needed, the replacement is done in LoadNode::Identity().
2103 //
Ideal(PhaseGVN * phase,bool can_reshape)2104 Node *LoadBNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2105 Node* mem = in(MemNode::Memory);
2106 Node* value = can_see_stored_value(mem,phase);
2107 if( value && !phase->type(value)->higher_equal( _type ) ) {
2108 Node *result = phase->transform( new LShiftINode(value, phase->intcon(24)) );
2109 return new RShiftINode(result, phase->intcon(24));
2110 }
2111 // Identity call will handle the case where truncation is not needed.
2112 return LoadNode::Ideal(phase, can_reshape);
2113 }
2114
Value(PhaseGVN * phase) const2115 const Type* LoadBNode::Value(PhaseGVN* phase) const {
2116 Node* mem = in(MemNode::Memory);
2117 Node* value = can_see_stored_value(mem,phase);
2118 if (value != NULL && value->is_Con() &&
2119 !value->bottom_type()->higher_equal(_type)) {
2120 // If the input to the store does not fit with the load's result type,
2121 // it must be truncated. We can't delay until Ideal call since
2122 // a singleton Value is needed for split_thru_phi optimization.
2123 int con = value->get_int();
2124 return TypeInt::make((con << 24) >> 24);
2125 }
2126 return LoadNode::Value(phase);
2127 }
2128
2129 //--------------------------LoadUBNode::Ideal-------------------------------------
2130 //
2131 // If the previous store is to the same address as this load,
2132 // and the value stored was larger than a byte, replace this load
2133 // with the value stored truncated to a byte. If no truncation is
2134 // needed, the replacement is done in LoadNode::Identity().
2135 //
Ideal(PhaseGVN * phase,bool can_reshape)2136 Node* LoadUBNode::Ideal(PhaseGVN* phase, bool can_reshape) {
2137 Node* mem = in(MemNode::Memory);
2138 Node* value = can_see_stored_value(mem, phase);
2139 if (value && !phase->type(value)->higher_equal(_type))
2140 return new AndINode(value, phase->intcon(0xFF));
2141 // Identity call will handle the case where truncation is not needed.
2142 return LoadNode::Ideal(phase, can_reshape);
2143 }
2144
Value(PhaseGVN * phase) const2145 const Type* LoadUBNode::Value(PhaseGVN* phase) const {
2146 Node* mem = in(MemNode::Memory);
2147 Node* value = can_see_stored_value(mem,phase);
2148 if (value != NULL && value->is_Con() &&
2149 !value->bottom_type()->higher_equal(_type)) {
2150 // If the input to the store does not fit with the load's result type,
2151 // it must be truncated. We can't delay until Ideal call since
2152 // a singleton Value is needed for split_thru_phi optimization.
2153 int con = value->get_int();
2154 return TypeInt::make(con & 0xFF);
2155 }
2156 return LoadNode::Value(phase);
2157 }
2158
2159 //--------------------------LoadUSNode::Ideal-------------------------------------
2160 //
2161 // If the previous store is to the same address as this load,
2162 // and the value stored was larger than a char, replace this load
2163 // with the value stored truncated to a char. If no truncation is
2164 // needed, the replacement is done in LoadNode::Identity().
2165 //
Ideal(PhaseGVN * phase,bool can_reshape)2166 Node *LoadUSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2167 Node* mem = in(MemNode::Memory);
2168 Node* value = can_see_stored_value(mem,phase);
2169 if( value && !phase->type(value)->higher_equal( _type ) )
2170 return new AndINode(value,phase->intcon(0xFFFF));
2171 // Identity call will handle the case where truncation is not needed.
2172 return LoadNode::Ideal(phase, can_reshape);
2173 }
2174
Value(PhaseGVN * phase) const2175 const Type* LoadUSNode::Value(PhaseGVN* phase) const {
2176 Node* mem = in(MemNode::Memory);
2177 Node* value = can_see_stored_value(mem,phase);
2178 if (value != NULL && value->is_Con() &&
2179 !value->bottom_type()->higher_equal(_type)) {
2180 // If the input to the store does not fit with the load's result type,
2181 // it must be truncated. We can't delay until Ideal call since
2182 // a singleton Value is needed for split_thru_phi optimization.
2183 int con = value->get_int();
2184 return TypeInt::make(con & 0xFFFF);
2185 }
2186 return LoadNode::Value(phase);
2187 }
2188
2189 //--------------------------LoadSNode::Ideal--------------------------------------
2190 //
2191 // If the previous store is to the same address as this load,
2192 // and the value stored was larger than a short, replace this load
2193 // with the value stored truncated to a short. If no truncation is
2194 // needed, the replacement is done in LoadNode::Identity().
2195 //
Ideal(PhaseGVN * phase,bool can_reshape)2196 Node *LoadSNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2197 Node* mem = in(MemNode::Memory);
2198 Node* value = can_see_stored_value(mem,phase);
2199 if( value && !phase->type(value)->higher_equal( _type ) ) {
2200 Node *result = phase->transform( new LShiftINode(value, phase->intcon(16)) );
2201 return new RShiftINode(result, phase->intcon(16));
2202 }
2203 // Identity call will handle the case where truncation is not needed.
2204 return LoadNode::Ideal(phase, can_reshape);
2205 }
2206
Value(PhaseGVN * phase) const2207 const Type* LoadSNode::Value(PhaseGVN* phase) const {
2208 Node* mem = in(MemNode::Memory);
2209 Node* value = can_see_stored_value(mem,phase);
2210 if (value != NULL && value->is_Con() &&
2211 !value->bottom_type()->higher_equal(_type)) {
2212 // If the input to the store does not fit with the load's result type,
2213 // it must be truncated. We can't delay until Ideal call since
2214 // a singleton Value is needed for split_thru_phi optimization.
2215 int con = value->get_int();
2216 return TypeInt::make((con << 16) >> 16);
2217 }
2218 return LoadNode::Value(phase);
2219 }
2220
2221 //=============================================================================
2222 //----------------------------LoadKlassNode::make------------------------------
2223 // Polymorphic factory method:
make(PhaseGVN & gvn,Node * ctl,Node * mem,Node * adr,const TypePtr * at,const TypeKlassPtr * tk)2224 Node* LoadKlassNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* at, const TypeKlassPtr* tk) {
2225 // sanity check the alias category against the created node type
2226 const TypePtr *adr_type = adr->bottom_type()->isa_ptr();
2227 assert(adr_type != NULL, "expecting TypeKlassPtr");
2228 #ifdef _LP64
2229 if (adr_type->is_ptr_to_narrowklass()) {
2230 assert(UseCompressedClassPointers, "no compressed klasses");
2231 Node* load_klass = gvn.transform(new LoadNKlassNode(ctl, mem, adr, at, tk->make_narrowklass(), MemNode::unordered));
2232 return new DecodeNKlassNode(load_klass, load_klass->bottom_type()->make_ptr());
2233 }
2234 #endif
2235 assert(!adr_type->is_ptr_to_narrowklass() && !adr_type->is_ptr_to_narrowoop(), "should have got back a narrow oop");
2236 return new LoadKlassNode(ctl, mem, adr, at, tk, MemNode::unordered);
2237 }
2238
2239 //------------------------------Value------------------------------------------
Value(PhaseGVN * phase) const2240 const Type* LoadKlassNode::Value(PhaseGVN* phase) const {
2241 return klass_value_common(phase);
2242 }
2243
2244 // In most cases, LoadKlassNode does not have the control input set. If the control
2245 // input is set, it must not be removed (by LoadNode::Ideal()).
can_remove_control() const2246 bool LoadKlassNode::can_remove_control() const {
2247 return false;
2248 }
2249
klass_value_common(PhaseGVN * phase) const2250 const Type* LoadNode::klass_value_common(PhaseGVN* phase) const {
2251 // Either input is TOP ==> the result is TOP
2252 const Type *t1 = phase->type( in(MemNode::Memory) );
2253 if (t1 == Type::TOP) return Type::TOP;
2254 Node *adr = in(MemNode::Address);
2255 const Type *t2 = phase->type( adr );
2256 if (t2 == Type::TOP) return Type::TOP;
2257 const TypePtr *tp = t2->is_ptr();
2258 if (TypePtr::above_centerline(tp->ptr()) ||
2259 tp->ptr() == TypePtr::Null) return Type::TOP;
2260
2261 // Return a more precise klass, if possible
2262 const TypeInstPtr *tinst = tp->isa_instptr();
2263 if (tinst != NULL) {
2264 ciInstanceKlass* ik = tinst->klass()->as_instance_klass();
2265 int offset = tinst->offset();
2266 if (ik == phase->C->env()->Class_klass()
2267 && (offset == java_lang_Class::klass_offset() ||
2268 offset == java_lang_Class::array_klass_offset())) {
2269 // We are loading a special hidden field from a Class mirror object,
2270 // the field which points to the VM's Klass metaobject.
2271 ciType* t = tinst->java_mirror_type();
2272 // java_mirror_type returns non-null for compile-time Class constants.
2273 if (t != NULL) {
2274 // constant oop => constant klass
2275 if (offset == java_lang_Class::array_klass_offset()) {
2276 if (t->is_void()) {
2277 // We cannot create a void array. Since void is a primitive type return null
2278 // klass. Users of this result need to do a null check on the returned klass.
2279 return TypePtr::NULL_PTR;
2280 }
2281 return TypeKlassPtr::make(ciArrayKlass::make(t));
2282 }
2283 if (!t->is_klass()) {
2284 // a primitive Class (e.g., int.class) has NULL for a klass field
2285 return TypePtr::NULL_PTR;
2286 }
2287 // (Folds up the 1st indirection in aClassConstant.getModifiers().)
2288 return TypeKlassPtr::make(t->as_klass());
2289 }
2290 // non-constant mirror, so we can't tell what's going on
2291 }
2292 if( !ik->is_loaded() )
2293 return _type; // Bail out if not loaded
2294 if (offset == oopDesc::klass_offset_in_bytes()) {
2295 if (tinst->klass_is_exact()) {
2296 return TypeKlassPtr::make(ik);
2297 }
2298 // See if we can become precise: no subklasses and no interface
2299 // (Note: We need to support verified interfaces.)
2300 if (!ik->is_interface() && !ik->has_subklass()) {
2301 // Add a dependence; if any subclass added we need to recompile
2302 if (!ik->is_final()) {
2303 // %%% should use stronger assert_unique_concrete_subtype instead
2304 phase->C->dependencies()->assert_leaf_type(ik);
2305 }
2306 // Return precise klass
2307 return TypeKlassPtr::make(ik);
2308 }
2309
2310 // Return root of possible klass
2311 return TypeKlassPtr::make(TypePtr::NotNull, ik, 0/*offset*/);
2312 }
2313 }
2314
2315 // Check for loading klass from an array
2316 const TypeAryPtr *tary = tp->isa_aryptr();
2317 if( tary != NULL ) {
2318 ciKlass *tary_klass = tary->klass();
2319 if (tary_klass != NULL // can be NULL when at BOTTOM or TOP
2320 && tary->offset() == oopDesc::klass_offset_in_bytes()) {
2321 if (tary->klass_is_exact()) {
2322 return TypeKlassPtr::make(tary_klass);
2323 }
2324 ciArrayKlass *ak = tary->klass()->as_array_klass();
2325 // If the klass is an object array, we defer the question to the
2326 // array component klass.
2327 if( ak->is_obj_array_klass() ) {
2328 assert( ak->is_loaded(), "" );
2329 ciKlass *base_k = ak->as_obj_array_klass()->base_element_klass();
2330 if( base_k->is_loaded() && base_k->is_instance_klass() ) {
2331 ciInstanceKlass* ik = base_k->as_instance_klass();
2332 // See if we can become precise: no subklasses and no interface
2333 if (!ik->is_interface() && !ik->has_subklass()) {
2334 // Add a dependence; if any subclass added we need to recompile
2335 if (!ik->is_final()) {
2336 phase->C->dependencies()->assert_leaf_type(ik);
2337 }
2338 // Return precise array klass
2339 return TypeKlassPtr::make(ak);
2340 }
2341 }
2342 return TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
2343 } else { // Found a type-array?
2344 assert( ak->is_type_array_klass(), "" );
2345 return TypeKlassPtr::make(ak); // These are always precise
2346 }
2347 }
2348 }
2349
2350 // Check for loading klass from an array klass
2351 const TypeKlassPtr *tkls = tp->isa_klassptr();
2352 if (tkls != NULL && !StressReflectiveCode) {
2353 ciKlass* klass = tkls->klass();
2354 if( !klass->is_loaded() )
2355 return _type; // Bail out if not loaded
2356 if( klass->is_obj_array_klass() &&
2357 tkls->offset() == in_bytes(ObjArrayKlass::element_klass_offset())) {
2358 ciKlass* elem = klass->as_obj_array_klass()->element_klass();
2359 // // Always returning precise element type is incorrect,
2360 // // e.g., element type could be object and array may contain strings
2361 // return TypeKlassPtr::make(TypePtr::Constant, elem, 0);
2362
2363 // The array's TypeKlassPtr was declared 'precise' or 'not precise'
2364 // according to the element type's subclassing.
2365 return TypeKlassPtr::make(tkls->ptr(), elem, 0/*offset*/);
2366 }
2367 if( klass->is_instance_klass() && tkls->klass_is_exact() &&
2368 tkls->offset() == in_bytes(Klass::super_offset())) {
2369 ciKlass* sup = klass->as_instance_klass()->super();
2370 // The field is Klass::_super. Return its (constant) value.
2371 // (Folds up the 2nd indirection in aClassConstant.getSuperClass().)
2372 return sup ? TypeKlassPtr::make(sup) : TypePtr::NULL_PTR;
2373 }
2374 }
2375
2376 // Bailout case
2377 return LoadNode::Value(phase);
2378 }
2379
2380 //------------------------------Identity---------------------------------------
2381 // To clean up reflective code, simplify k.java_mirror.as_klass to plain k.
2382 // Also feed through the klass in Allocate(...klass...)._klass.
Identity(PhaseGVN * phase)2383 Node* LoadKlassNode::Identity(PhaseGVN* phase) {
2384 return klass_identity_common(phase);
2385 }
2386
klass_identity_common(PhaseGVN * phase)2387 Node* LoadNode::klass_identity_common(PhaseGVN* phase) {
2388 Node* x = LoadNode::Identity(phase);
2389 if (x != this) return x;
2390
2391 // Take apart the address into an oop and and offset.
2392 // Return 'this' if we cannot.
2393 Node* adr = in(MemNode::Address);
2394 intptr_t offset = 0;
2395 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2396 if (base == NULL) return this;
2397 const TypeOopPtr* toop = phase->type(adr)->isa_oopptr();
2398 if (toop == NULL) return this;
2399
2400 // Step over potential GC barrier for OopHandle resolve
2401 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2402 if (bs->is_gc_barrier_node(base)) {
2403 base = bs->step_over_gc_barrier(base);
2404 }
2405
2406 // We can fetch the klass directly through an AllocateNode.
2407 // This works even if the klass is not constant (clone or newArray).
2408 if (offset == oopDesc::klass_offset_in_bytes()) {
2409 Node* allocated_klass = AllocateNode::Ideal_klass(base, phase);
2410 if (allocated_klass != NULL) {
2411 return allocated_klass;
2412 }
2413 }
2414
2415 // Simplify k.java_mirror.as_klass to plain k, where k is a Klass*.
2416 // See inline_native_Class_query for occurrences of these patterns.
2417 // Java Example: x.getClass().isAssignableFrom(y)
2418 //
2419 // This improves reflective code, often making the Class
2420 // mirror go completely dead. (Current exception: Class
2421 // mirrors may appear in debug info, but we could clean them out by
2422 // introducing a new debug info operator for Klass.java_mirror).
2423
2424 if (toop->isa_instptr() && toop->klass() == phase->C->env()->Class_klass()
2425 && offset == java_lang_Class::klass_offset()) {
2426 if (base->is_Load()) {
2427 Node* base2 = base->in(MemNode::Address);
2428 if (base2->is_Load()) { /* direct load of a load which is the OopHandle */
2429 Node* adr2 = base2->in(MemNode::Address);
2430 const TypeKlassPtr* tkls = phase->type(adr2)->isa_klassptr();
2431 if (tkls != NULL && !tkls->empty()
2432 && (tkls->klass()->is_instance_klass() ||
2433 tkls->klass()->is_array_klass())
2434 && adr2->is_AddP()
2435 ) {
2436 int mirror_field = in_bytes(Klass::java_mirror_offset());
2437 if (tkls->offset() == mirror_field) {
2438 return adr2->in(AddPNode::Base);
2439 }
2440 }
2441 }
2442 }
2443 }
2444
2445 return this;
2446 }
2447
2448
2449 //------------------------------Value------------------------------------------
Value(PhaseGVN * phase) const2450 const Type* LoadNKlassNode::Value(PhaseGVN* phase) const {
2451 const Type *t = klass_value_common(phase);
2452 if (t == Type::TOP)
2453 return t;
2454
2455 return t->make_narrowklass();
2456 }
2457
2458 //------------------------------Identity---------------------------------------
2459 // To clean up reflective code, simplify k.java_mirror.as_klass to narrow k.
2460 // Also feed through the klass in Allocate(...klass...)._klass.
Identity(PhaseGVN * phase)2461 Node* LoadNKlassNode::Identity(PhaseGVN* phase) {
2462 Node *x = klass_identity_common(phase);
2463
2464 const Type *t = phase->type( x );
2465 if( t == Type::TOP ) return x;
2466 if( t->isa_narrowklass()) return x;
2467 assert (!t->isa_narrowoop(), "no narrow oop here");
2468
2469 return phase->transform(new EncodePKlassNode(x, t->make_narrowklass()));
2470 }
2471
2472 //------------------------------Value-----------------------------------------
Value(PhaseGVN * phase) const2473 const Type* LoadRangeNode::Value(PhaseGVN* phase) const {
2474 // Either input is TOP ==> the result is TOP
2475 const Type *t1 = phase->type( in(MemNode::Memory) );
2476 if( t1 == Type::TOP ) return Type::TOP;
2477 Node *adr = in(MemNode::Address);
2478 const Type *t2 = phase->type( adr );
2479 if( t2 == Type::TOP ) return Type::TOP;
2480 const TypePtr *tp = t2->is_ptr();
2481 if (TypePtr::above_centerline(tp->ptr())) return Type::TOP;
2482 const TypeAryPtr *tap = tp->isa_aryptr();
2483 if( !tap ) return _type;
2484 return tap->size();
2485 }
2486
2487 //-------------------------------Ideal---------------------------------------
2488 // Feed through the length in AllocateArray(...length...)._length.
Ideal(PhaseGVN * phase,bool can_reshape)2489 Node *LoadRangeNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2490 Node* p = MemNode::Ideal_common(phase, can_reshape);
2491 if (p) return (p == NodeSentinel) ? NULL : p;
2492
2493 // Take apart the address into an oop and and offset.
2494 // Return 'this' if we cannot.
2495 Node* adr = in(MemNode::Address);
2496 intptr_t offset = 0;
2497 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2498 if (base == NULL) return NULL;
2499 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2500 if (tary == NULL) return NULL;
2501
2502 // We can fetch the length directly through an AllocateArrayNode.
2503 // This works even if the length is not constant (clone or newArray).
2504 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2505 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2506 if (alloc != NULL) {
2507 Node* allocated_length = alloc->Ideal_length();
2508 Node* len = alloc->make_ideal_length(tary, phase);
2509 if (allocated_length != len) {
2510 // New CastII improves on this.
2511 return len;
2512 }
2513 }
2514 }
2515
2516 return NULL;
2517 }
2518
2519 //------------------------------Identity---------------------------------------
2520 // Feed through the length in AllocateArray(...length...)._length.
Identity(PhaseGVN * phase)2521 Node* LoadRangeNode::Identity(PhaseGVN* phase) {
2522 Node* x = LoadINode::Identity(phase);
2523 if (x != this) return x;
2524
2525 // Take apart the address into an oop and and offset.
2526 // Return 'this' if we cannot.
2527 Node* adr = in(MemNode::Address);
2528 intptr_t offset = 0;
2529 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
2530 if (base == NULL) return this;
2531 const TypeAryPtr* tary = phase->type(adr)->isa_aryptr();
2532 if (tary == NULL) return this;
2533
2534 // We can fetch the length directly through an AllocateArrayNode.
2535 // This works even if the length is not constant (clone or newArray).
2536 if (offset == arrayOopDesc::length_offset_in_bytes()) {
2537 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(base, phase);
2538 if (alloc != NULL) {
2539 Node* allocated_length = alloc->Ideal_length();
2540 // Do not allow make_ideal_length to allocate a CastII node.
2541 Node* len = alloc->make_ideal_length(tary, phase, false);
2542 if (allocated_length == len) {
2543 // Return allocated_length only if it would not be improved by a CastII.
2544 return allocated_length;
2545 }
2546 }
2547 }
2548
2549 return this;
2550
2551 }
2552
2553 //=============================================================================
2554 //---------------------------StoreNode::make-----------------------------------
2555 // Polymorphic factory method:
make(PhaseGVN & gvn,Node * ctl,Node * mem,Node * adr,const TypePtr * adr_type,Node * val,BasicType bt,MemOrd mo)2556 StoreNode* StoreNode::make(PhaseGVN& gvn, Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, BasicType bt, MemOrd mo) {
2557 assert((mo == unordered || mo == release), "unexpected");
2558 Compile* C = gvn.C;
2559 assert(C->get_alias_index(adr_type) != Compile::AliasIdxRaw ||
2560 ctl != NULL, "raw memory operations should have control edge");
2561
2562 switch (bt) {
2563 case T_BOOLEAN: val = gvn.transform(new AndINode(val, gvn.intcon(0x1))); // Fall through to T_BYTE case
2564 case T_BYTE: return new StoreBNode(ctl, mem, adr, adr_type, val, mo);
2565 case T_INT: return new StoreINode(ctl, mem, adr, adr_type, val, mo);
2566 case T_CHAR:
2567 case T_SHORT: return new StoreCNode(ctl, mem, adr, adr_type, val, mo);
2568 case T_LONG: return new StoreLNode(ctl, mem, adr, adr_type, val, mo);
2569 case T_FLOAT: return new StoreFNode(ctl, mem, adr, adr_type, val, mo);
2570 case T_DOUBLE: return new StoreDNode(ctl, mem, adr, adr_type, val, mo);
2571 case T_METADATA:
2572 case T_ADDRESS:
2573 case T_OBJECT:
2574 #ifdef _LP64
2575 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2576 val = gvn.transform(new EncodePNode(val, val->bottom_type()->make_narrowoop()));
2577 return new StoreNNode(ctl, mem, adr, adr_type, val, mo);
2578 } else if (adr->bottom_type()->is_ptr_to_narrowklass() ||
2579 (UseCompressedClassPointers && val->bottom_type()->isa_klassptr() &&
2580 adr->bottom_type()->isa_rawptr())) {
2581 val = gvn.transform(new EncodePKlassNode(val, val->bottom_type()->make_narrowklass()));
2582 return new StoreNKlassNode(ctl, mem, adr, adr_type, val, mo);
2583 }
2584 #endif
2585 {
2586 return new StorePNode(ctl, mem, adr, adr_type, val, mo);
2587 }
2588 default:
2589 ShouldNotReachHere();
2590 return (StoreNode*)NULL;
2591 }
2592 }
2593
make_atomic(Node * ctl,Node * mem,Node * adr,const TypePtr * adr_type,Node * val,MemOrd mo)2594 StoreLNode* StoreLNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2595 bool require_atomic = true;
2596 return new StoreLNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2597 }
2598
make_atomic(Node * ctl,Node * mem,Node * adr,const TypePtr * adr_type,Node * val,MemOrd mo)2599 StoreDNode* StoreDNode::make_atomic(Node* ctl, Node* mem, Node* adr, const TypePtr* adr_type, Node* val, MemOrd mo) {
2600 bool require_atomic = true;
2601 return new StoreDNode(ctl, mem, adr, adr_type, val, mo, require_atomic);
2602 }
2603
2604
2605 //--------------------------bottom_type----------------------------------------
bottom_type() const2606 const Type *StoreNode::bottom_type() const {
2607 return Type::MEMORY;
2608 }
2609
2610 //------------------------------hash-------------------------------------------
hash() const2611 uint StoreNode::hash() const {
2612 // unroll addition of interesting fields
2613 //return (uintptr_t)in(Control) + (uintptr_t)in(Memory) + (uintptr_t)in(Address) + (uintptr_t)in(ValueIn);
2614
2615 // Since they are not commoned, do not hash them:
2616 return NO_HASH;
2617 }
2618
2619 //------------------------------Ideal------------------------------------------
2620 // Change back-to-back Store(, p, x) -> Store(m, p, y) to Store(m, p, x).
2621 // When a store immediately follows a relevant allocation/initialization,
2622 // try to capture it into the initialization, or hoist it above.
Ideal(PhaseGVN * phase,bool can_reshape)2623 Node *StoreNode::Ideal(PhaseGVN *phase, bool can_reshape) {
2624 Node* p = MemNode::Ideal_common(phase, can_reshape);
2625 if (p) return (p == NodeSentinel) ? NULL : p;
2626
2627 Node* mem = in(MemNode::Memory);
2628 Node* address = in(MemNode::Address);
2629 Node* value = in(MemNode::ValueIn);
2630 // Back-to-back stores to same address? Fold em up. Generally
2631 // unsafe if I have intervening uses... Also disallowed for StoreCM
2632 // since they must follow each StoreP operation. Redundant StoreCMs
2633 // are eliminated just before matching in final_graph_reshape.
2634 {
2635 Node* st = mem;
2636 // If Store 'st' has more than one use, we cannot fold 'st' away.
2637 // For example, 'st' might be the final state at a conditional
2638 // return. Or, 'st' might be used by some node which is live at
2639 // the same time 'st' is live, which might be unschedulable. So,
2640 // require exactly ONE user until such time as we clone 'mem' for
2641 // each of 'mem's uses (thus making the exactly-1-user-rule hold
2642 // true).
2643 while (st->is_Store() && st->outcnt() == 1 && st->Opcode() != Op_StoreCM) {
2644 // Looking at a dead closed cycle of memory?
2645 assert(st != st->in(MemNode::Memory), "dead loop in StoreNode::Ideal");
2646 assert(Opcode() == st->Opcode() ||
2647 st->Opcode() == Op_StoreVector ||
2648 Opcode() == Op_StoreVector ||
2649 st->Opcode() == Op_StoreVectorScatter ||
2650 Opcode() == Op_StoreVectorScatter ||
2651 phase->C->get_alias_index(adr_type()) == Compile::AliasIdxRaw ||
2652 (Opcode() == Op_StoreL && st->Opcode() == Op_StoreI) || // expanded ClearArrayNode
2653 (Opcode() == Op_StoreI && st->Opcode() == Op_StoreL) || // initialization by arraycopy
2654 (is_mismatched_access() || st->as_Store()->is_mismatched_access()),
2655 "no mismatched stores, except on raw memory: %s %s", NodeClassNames[Opcode()], NodeClassNames[st->Opcode()]);
2656
2657 if (st->in(MemNode::Address)->eqv_uncast(address) &&
2658 st->as_Store()->memory_size() <= this->memory_size()) {
2659 Node* use = st->raw_out(0);
2660 if (phase->is_IterGVN()) {
2661 phase->is_IterGVN()->rehash_node_delayed(use);
2662 }
2663 if (can_reshape) {
2664 use->set_req_X(MemNode::Memory, st->in(MemNode::Memory), phase->is_IterGVN());
2665 } else {
2666 // It's OK to do this in the parser, since DU info is always accurate,
2667 // and the parser always refers to nodes via SafePointNode maps.
2668 use->set_req(MemNode::Memory, st->in(MemNode::Memory));
2669 }
2670 return this;
2671 }
2672 st = st->in(MemNode::Memory);
2673 }
2674 }
2675
2676
2677 // Capture an unaliased, unconditional, simple store into an initializer.
2678 // Or, if it is independent of the allocation, hoist it above the allocation.
2679 if (ReduceFieldZeroing && /*can_reshape &&*/
2680 mem->is_Proj() && mem->in(0)->is_Initialize()) {
2681 InitializeNode* init = mem->in(0)->as_Initialize();
2682 intptr_t offset = init->can_capture_store(this, phase, can_reshape);
2683 if (offset > 0) {
2684 Node* moved = init->capture_store(this, offset, phase, can_reshape);
2685 // If the InitializeNode captured me, it made a raw copy of me,
2686 // and I need to disappear.
2687 if (moved != NULL) {
2688 // %%% hack to ensure that Ideal returns a new node:
2689 mem = MergeMemNode::make(mem);
2690 return mem; // fold me away
2691 }
2692 }
2693 }
2694
2695 // Fold reinterpret cast into memory operation:
2696 // StoreX mem (MoveY2X v) => StoreY mem v
2697 if (value->is_Move()) {
2698 const Type* vt = value->in(1)->bottom_type();
2699 if (has_reinterpret_variant(vt)) {
2700 if (phase->C->post_loop_opts_phase()) {
2701 return convert_to_reinterpret_store(*phase, value->in(1), vt);
2702 } else {
2703 phase->C->record_for_post_loop_opts_igvn(this); // attempt the transformation once loop opts are over
2704 }
2705 }
2706 }
2707
2708 return NULL; // No further progress
2709 }
2710
2711 //------------------------------Value-----------------------------------------
Value(PhaseGVN * phase) const2712 const Type* StoreNode::Value(PhaseGVN* phase) const {
2713 // Either input is TOP ==> the result is TOP
2714 const Type *t1 = phase->type( in(MemNode::Memory) );
2715 if( t1 == Type::TOP ) return Type::TOP;
2716 const Type *t2 = phase->type( in(MemNode::Address) );
2717 if( t2 == Type::TOP ) return Type::TOP;
2718 const Type *t3 = phase->type( in(MemNode::ValueIn) );
2719 if( t3 == Type::TOP ) return Type::TOP;
2720 return Type::MEMORY;
2721 }
2722
2723 //------------------------------Identity---------------------------------------
2724 // Remove redundant stores:
2725 // Store(m, p, Load(m, p)) changes to m.
2726 // Store(, p, x) -> Store(m, p, x) changes to Store(m, p, x).
Identity(PhaseGVN * phase)2727 Node* StoreNode::Identity(PhaseGVN* phase) {
2728 Node* mem = in(MemNode::Memory);
2729 Node* adr = in(MemNode::Address);
2730 Node* val = in(MemNode::ValueIn);
2731
2732 Node* result = this;
2733
2734 // Load then Store? Then the Store is useless
2735 if (val->is_Load() &&
2736 val->in(MemNode::Address)->eqv_uncast(adr) &&
2737 val->in(MemNode::Memory )->eqv_uncast(mem) &&
2738 val->as_Load()->store_Opcode() == Opcode()) {
2739 result = mem;
2740 }
2741
2742 // Two stores in a row of the same value?
2743 if (result == this &&
2744 mem->is_Store() &&
2745 mem->in(MemNode::Address)->eqv_uncast(adr) &&
2746 mem->in(MemNode::ValueIn)->eqv_uncast(val) &&
2747 mem->Opcode() == Opcode()) {
2748 result = mem;
2749 }
2750
2751 // Store of zero anywhere into a freshly-allocated object?
2752 // Then the store is useless.
2753 // (It must already have been captured by the InitializeNode.)
2754 if (result == this &&
2755 ReduceFieldZeroing && phase->type(val)->is_zero_type()) {
2756 // a newly allocated object is already all-zeroes everywhere
2757 if (mem->is_Proj() && mem->in(0)->is_Allocate()) {
2758 result = mem;
2759 }
2760
2761 if (result == this) {
2762 // the store may also apply to zero-bits in an earlier object
2763 Node* prev_mem = find_previous_store(phase);
2764 // Steps (a), (b): Walk past independent stores to find an exact match.
2765 if (prev_mem != NULL) {
2766 Node* prev_val = can_see_stored_value(prev_mem, phase);
2767 if (prev_val != NULL && prev_val == val) {
2768 // prev_val and val might differ by a cast; it would be good
2769 // to keep the more informative of the two.
2770 result = mem;
2771 }
2772 }
2773 }
2774 }
2775
2776 PhaseIterGVN* igvn = phase->is_IterGVN();
2777 if (result != this && igvn != NULL) {
2778 MemBarNode* trailing = trailing_membar();
2779 if (trailing != NULL) {
2780 #ifdef ASSERT
2781 const TypeOopPtr* t_oop = phase->type(in(Address))->isa_oopptr();
2782 assert(t_oop == NULL || t_oop->is_known_instance_field(), "only for non escaping objects");
2783 #endif
2784 trailing->remove(igvn);
2785 }
2786 }
2787
2788 return result;
2789 }
2790
2791 //------------------------------match_edge-------------------------------------
2792 // Do we Match on this edge index or not? Match only memory & value
match_edge(uint idx) const2793 uint StoreNode::match_edge(uint idx) const {
2794 return idx == MemNode::Address || idx == MemNode::ValueIn;
2795 }
2796
2797 //------------------------------cmp--------------------------------------------
2798 // Do not common stores up together. They generally have to be split
2799 // back up anyways, so do not bother.
cmp(const Node & n) const2800 bool StoreNode::cmp( const Node &n ) const {
2801 return (&n == this); // Always fail except on self
2802 }
2803
2804 //------------------------------Ideal_masked_input-----------------------------
2805 // Check for a useless mask before a partial-word store
2806 // (StoreB ... (AndI valIn conIa) )
2807 // If (conIa & mask == mask) this simplifies to
2808 // (StoreB ... (valIn) )
Ideal_masked_input(PhaseGVN * phase,uint mask)2809 Node *StoreNode::Ideal_masked_input(PhaseGVN *phase, uint mask) {
2810 Node *val = in(MemNode::ValueIn);
2811 if( val->Opcode() == Op_AndI ) {
2812 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2813 if( t && t->is_con() && (t->get_con() & mask) == mask ) {
2814 set_req(MemNode::ValueIn, val->in(1));
2815 return this;
2816 }
2817 }
2818 return NULL;
2819 }
2820
2821
2822 //------------------------------Ideal_sign_extended_input----------------------
2823 // Check for useless sign-extension before a partial-word store
2824 // (StoreB ... (RShiftI _ (LShiftI _ valIn conIL ) conIR) )
2825 // If (conIL == conIR && conIR <= num_bits) this simplifies to
2826 // (StoreB ... (valIn) )
Ideal_sign_extended_input(PhaseGVN * phase,int num_bits)2827 Node *StoreNode::Ideal_sign_extended_input(PhaseGVN *phase, int num_bits) {
2828 Node *val = in(MemNode::ValueIn);
2829 if( val->Opcode() == Op_RShiftI ) {
2830 const TypeInt *t = phase->type( val->in(2) )->isa_int();
2831 if( t && t->is_con() && (t->get_con() <= num_bits) ) {
2832 Node *shl = val->in(1);
2833 if( shl->Opcode() == Op_LShiftI ) {
2834 const TypeInt *t2 = phase->type( shl->in(2) )->isa_int();
2835 if( t2 && t2->is_con() && (t2->get_con() == t->get_con()) ) {
2836 set_req(MemNode::ValueIn, shl->in(1));
2837 return this;
2838 }
2839 }
2840 }
2841 }
2842 return NULL;
2843 }
2844
2845 //------------------------------value_never_loaded-----------------------------------
2846 // Determine whether there are any possible loads of the value stored.
2847 // For simplicity, we actually check if there are any loads from the
2848 // address stored to, not just for loads of the value stored by this node.
2849 //
value_never_loaded(PhaseTransform * phase) const2850 bool StoreNode::value_never_loaded( PhaseTransform *phase) const {
2851 Node *adr = in(Address);
2852 const TypeOopPtr *adr_oop = phase->type(adr)->isa_oopptr();
2853 if (adr_oop == NULL)
2854 return false;
2855 if (!adr_oop->is_known_instance_field())
2856 return false; // if not a distinct instance, there may be aliases of the address
2857 for (DUIterator_Fast imax, i = adr->fast_outs(imax); i < imax; i++) {
2858 Node *use = adr->fast_out(i);
2859 if (use->is_Load() || use->is_LoadStore()) {
2860 return false;
2861 }
2862 }
2863 return true;
2864 }
2865
trailing_membar() const2866 MemBarNode* StoreNode::trailing_membar() const {
2867 if (is_release()) {
2868 MemBarNode* trailing_mb = NULL;
2869 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2870 Node* u = fast_out(i);
2871 if (u->is_MemBar()) {
2872 if (u->as_MemBar()->trailing_store()) {
2873 assert(u->Opcode() == Op_MemBarVolatile, "");
2874 assert(trailing_mb == NULL, "only one");
2875 trailing_mb = u->as_MemBar();
2876 #ifdef ASSERT
2877 Node* leading = u->as_MemBar()->leading_membar();
2878 assert(leading->Opcode() == Op_MemBarRelease, "incorrect membar");
2879 assert(leading->as_MemBar()->leading_store(), "incorrect membar pair");
2880 assert(leading->as_MemBar()->trailing_membar() == u, "incorrect membar pair");
2881 #endif
2882 } else {
2883 assert(u->as_MemBar()->standalone(), "");
2884 }
2885 }
2886 }
2887 return trailing_mb;
2888 }
2889 return NULL;
2890 }
2891
2892
2893 //=============================================================================
2894 //------------------------------Ideal------------------------------------------
2895 // If the store is from an AND mask that leaves the low bits untouched, then
2896 // we can skip the AND operation. If the store is from a sign-extension
2897 // (a left shift, then right shift) we can skip both.
Ideal(PhaseGVN * phase,bool can_reshape)2898 Node *StoreBNode::Ideal(PhaseGVN *phase, bool can_reshape){
2899 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFF);
2900 if( progress != NULL ) return progress;
2901
2902 progress = StoreNode::Ideal_sign_extended_input(phase, 24);
2903 if( progress != NULL ) return progress;
2904
2905 // Finally check the default case
2906 return StoreNode::Ideal(phase, can_reshape);
2907 }
2908
2909 //=============================================================================
2910 //------------------------------Ideal------------------------------------------
2911 // If the store is from an AND mask that leaves the low bits untouched, then
2912 // we can skip the AND operation
Ideal(PhaseGVN * phase,bool can_reshape)2913 Node *StoreCNode::Ideal(PhaseGVN *phase, bool can_reshape){
2914 Node *progress = StoreNode::Ideal_masked_input(phase, 0xFFFF);
2915 if( progress != NULL ) return progress;
2916
2917 progress = StoreNode::Ideal_sign_extended_input(phase, 16);
2918 if( progress != NULL ) return progress;
2919
2920 // Finally check the default case
2921 return StoreNode::Ideal(phase, can_reshape);
2922 }
2923
2924 //=============================================================================
2925 //------------------------------Identity---------------------------------------
Identity(PhaseGVN * phase)2926 Node* StoreCMNode::Identity(PhaseGVN* phase) {
2927 // No need to card mark when storing a null ptr
2928 Node* my_store = in(MemNode::OopStore);
2929 if (my_store->is_Store()) {
2930 const Type *t1 = phase->type( my_store->in(MemNode::ValueIn) );
2931 if( t1 == TypePtr::NULL_PTR ) {
2932 return in(MemNode::Memory);
2933 }
2934 }
2935 return this;
2936 }
2937
2938 //=============================================================================
2939 //------------------------------Ideal---------------------------------------
Ideal(PhaseGVN * phase,bool can_reshape)2940 Node *StoreCMNode::Ideal(PhaseGVN *phase, bool can_reshape){
2941 Node* progress = StoreNode::Ideal(phase, can_reshape);
2942 if (progress != NULL) return progress;
2943
2944 Node* my_store = in(MemNode::OopStore);
2945 if (my_store->is_MergeMem()) {
2946 Node* mem = my_store->as_MergeMem()->memory_at(oop_alias_idx());
2947 set_req(MemNode::OopStore, mem);
2948 return this;
2949 }
2950
2951 return NULL;
2952 }
2953
2954 //------------------------------Value-----------------------------------------
Value(PhaseGVN * phase) const2955 const Type* StoreCMNode::Value(PhaseGVN* phase) const {
2956 // Either input is TOP ==> the result is TOP
2957 const Type *t = phase->type( in(MemNode::Memory) );
2958 if( t == Type::TOP ) return Type::TOP;
2959 t = phase->type( in(MemNode::Address) );
2960 if( t == Type::TOP ) return Type::TOP;
2961 t = phase->type( in(MemNode::ValueIn) );
2962 if( t == Type::TOP ) return Type::TOP;
2963 // If extra input is TOP ==> the result is TOP
2964 t = phase->type( in(MemNode::OopStore) );
2965 if( t == Type::TOP ) return Type::TOP;
2966
2967 return StoreNode::Value( phase );
2968 }
2969
2970
2971 //=============================================================================
2972 //----------------------------------SCMemProjNode------------------------------
Value(PhaseGVN * phase) const2973 const Type* SCMemProjNode::Value(PhaseGVN* phase) const
2974 {
2975 return bottom_type();
2976 }
2977
2978 //=============================================================================
2979 //----------------------------------LoadStoreNode------------------------------
LoadStoreNode(Node * c,Node * mem,Node * adr,Node * val,const TypePtr * at,const Type * rt,uint required)2980 LoadStoreNode::LoadStoreNode( Node *c, Node *mem, Node *adr, Node *val, const TypePtr* at, const Type* rt, uint required )
2981 : Node(required),
2982 _type(rt),
2983 _adr_type(at),
2984 _barrier_data(0)
2985 {
2986 init_req(MemNode::Control, c );
2987 init_req(MemNode::Memory , mem);
2988 init_req(MemNode::Address, adr);
2989 init_req(MemNode::ValueIn, val);
2990 init_class_id(Class_LoadStore);
2991 }
2992
ideal_reg() const2993 uint LoadStoreNode::ideal_reg() const {
2994 return _type->ideal_reg();
2995 }
2996
result_not_used() const2997 bool LoadStoreNode::result_not_used() const {
2998 for( DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++ ) {
2999 Node *x = fast_out(i);
3000 if (x->Opcode() == Op_SCMemProj) continue;
3001 return false;
3002 }
3003 return true;
3004 }
3005
trailing_membar() const3006 MemBarNode* LoadStoreNode::trailing_membar() const {
3007 MemBarNode* trailing = NULL;
3008 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
3009 Node* u = fast_out(i);
3010 if (u->is_MemBar()) {
3011 if (u->as_MemBar()->trailing_load_store()) {
3012 assert(u->Opcode() == Op_MemBarAcquire, "");
3013 assert(trailing == NULL, "only one");
3014 trailing = u->as_MemBar();
3015 #ifdef ASSERT
3016 Node* leading = trailing->leading_membar();
3017 assert(support_IRIW_for_not_multiple_copy_atomic_cpu || leading->Opcode() == Op_MemBarRelease, "incorrect membar");
3018 assert(leading->as_MemBar()->leading_load_store(), "incorrect membar pair");
3019 assert(leading->as_MemBar()->trailing_membar() == trailing, "incorrect membar pair");
3020 #endif
3021 } else {
3022 assert(u->as_MemBar()->standalone(), "wrong barrier kind");
3023 }
3024 }
3025 }
3026
3027 return trailing;
3028 }
3029
size_of() const3030 uint LoadStoreNode::size_of() const { return sizeof(*this); }
3031
3032 //=============================================================================
3033 //----------------------------------LoadStoreConditionalNode--------------------
LoadStoreConditionalNode(Node * c,Node * mem,Node * adr,Node * val,Node * ex)3034 LoadStoreConditionalNode::LoadStoreConditionalNode( Node *c, Node *mem, Node *adr, Node *val, Node *ex ) : LoadStoreNode(c, mem, adr, val, NULL, TypeInt::BOOL, 5) {
3035 init_req(ExpectedIn, ex );
3036 }
3037
3038 //=============================================================================
3039 //-------------------------------adr_type--------------------------------------
adr_type() const3040 const TypePtr* ClearArrayNode::adr_type() const {
3041 Node *adr = in(3);
3042 if (adr == NULL) return NULL; // node is dead
3043 return MemNode::calculate_adr_type(adr->bottom_type());
3044 }
3045
3046 //------------------------------match_edge-------------------------------------
3047 // Do we Match on this edge index or not? Do not match memory
match_edge(uint idx) const3048 uint ClearArrayNode::match_edge(uint idx) const {
3049 return idx > 1;
3050 }
3051
3052 //------------------------------Identity---------------------------------------
3053 // Clearing a zero length array does nothing
Identity(PhaseGVN * phase)3054 Node* ClearArrayNode::Identity(PhaseGVN* phase) {
3055 return phase->type(in(2))->higher_equal(TypeX::ZERO) ? in(1) : this;
3056 }
3057
3058 //------------------------------Idealize---------------------------------------
3059 // Clearing a short array is faster with stores
Ideal(PhaseGVN * phase,bool can_reshape)3060 Node *ClearArrayNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3061 // Already know this is a large node, do not try to ideal it
3062 if (!IdealizeClearArrayNode || _is_large) return NULL;
3063
3064 const int unit = BytesPerLong;
3065 const TypeX* t = phase->type(in(2))->isa_intptr_t();
3066 if (!t) return NULL;
3067 if (!t->is_con()) return NULL;
3068 intptr_t raw_count = t->get_con();
3069 intptr_t size = raw_count;
3070 if (!Matcher::init_array_count_is_in_bytes) size *= unit;
3071 // Clearing nothing uses the Identity call.
3072 // Negative clears are possible on dead ClearArrays
3073 // (see jck test stmt114.stmt11402.val).
3074 if (size <= 0 || size % unit != 0) return NULL;
3075 intptr_t count = size / unit;
3076 // Length too long; communicate this to matchers and assemblers.
3077 // Assemblers are responsible to produce fast hardware clears for it.
3078 if (size > InitArrayShortSize) {
3079 return new ClearArrayNode(in(0), in(1), in(2), in(3), true);
3080 }
3081 Node *mem = in(1);
3082 if( phase->type(mem)==Type::TOP ) return NULL;
3083 Node *adr = in(3);
3084 const Type* at = phase->type(adr);
3085 if( at==Type::TOP ) return NULL;
3086 const TypePtr* atp = at->isa_ptr();
3087 // adjust atp to be the correct array element address type
3088 if (atp == NULL) atp = TypePtr::BOTTOM;
3089 else atp = atp->add_offset(Type::OffsetBot);
3090 // Get base for derived pointer purposes
3091 if( adr->Opcode() != Op_AddP ) Unimplemented();
3092 Node *base = adr->in(1);
3093
3094 Node *zero = phase->makecon(TypeLong::ZERO);
3095 Node *off = phase->MakeConX(BytesPerLong);
3096 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3097 count--;
3098 while( count-- ) {
3099 mem = phase->transform(mem);
3100 adr = phase->transform(new AddPNode(base,adr,off));
3101 mem = new StoreLNode(in(0),mem,adr,atp,zero,MemNode::unordered,false);
3102 }
3103 return mem;
3104 }
3105
3106 //----------------------------step_through----------------------------------
3107 // Return allocation input memory edge if it is different instance
3108 // or itself if it is the one we are looking for.
step_through(Node ** np,uint instance_id,PhaseTransform * phase)3109 bool ClearArrayNode::step_through(Node** np, uint instance_id, PhaseTransform* phase) {
3110 Node* n = *np;
3111 assert(n->is_ClearArray(), "sanity");
3112 intptr_t offset;
3113 AllocateNode* alloc = AllocateNode::Ideal_allocation(n->in(3), phase, offset);
3114 // This method is called only before Allocate nodes are expanded
3115 // during macro nodes expansion. Before that ClearArray nodes are
3116 // only generated in PhaseMacroExpand::generate_arraycopy() (before
3117 // Allocate nodes are expanded) which follows allocations.
3118 assert(alloc != NULL, "should have allocation");
3119 if (alloc->_idx == instance_id) {
3120 // Can not bypass initialization of the instance we are looking for.
3121 return false;
3122 }
3123 // Otherwise skip it.
3124 InitializeNode* init = alloc->initialization();
3125 if (init != NULL)
3126 *np = init->in(TypeFunc::Memory);
3127 else
3128 *np = alloc->in(TypeFunc::Memory);
3129 return true;
3130 }
3131
3132 //----------------------------clear_memory-------------------------------------
3133 // Generate code to initialize object storage to zero.
clear_memory(Node * ctl,Node * mem,Node * dest,intptr_t start_offset,Node * end_offset,PhaseGVN * phase)3134 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3135 intptr_t start_offset,
3136 Node* end_offset,
3137 PhaseGVN* phase) {
3138 intptr_t offset = start_offset;
3139
3140 int unit = BytesPerLong;
3141 if ((offset % unit) != 0) {
3142 Node* adr = new AddPNode(dest, dest, phase->MakeConX(offset));
3143 adr = phase->transform(adr);
3144 const TypePtr* atp = TypeRawPtr::BOTTOM;
3145 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3146 mem = phase->transform(mem);
3147 offset += BytesPerInt;
3148 }
3149 assert((offset % unit) == 0, "");
3150
3151 // Initialize the remaining stuff, if any, with a ClearArray.
3152 return clear_memory(ctl, mem, dest, phase->MakeConX(offset), end_offset, phase);
3153 }
3154
clear_memory(Node * ctl,Node * mem,Node * dest,Node * start_offset,Node * end_offset,PhaseGVN * phase)3155 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3156 Node* start_offset,
3157 Node* end_offset,
3158 PhaseGVN* phase) {
3159 if (start_offset == end_offset) {
3160 // nothing to do
3161 return mem;
3162 }
3163
3164 int unit = BytesPerLong;
3165 Node* zbase = start_offset;
3166 Node* zend = end_offset;
3167
3168 // Scale to the unit required by the CPU:
3169 if (!Matcher::init_array_count_is_in_bytes) {
3170 Node* shift = phase->intcon(exact_log2(unit));
3171 zbase = phase->transform(new URShiftXNode(zbase, shift) );
3172 zend = phase->transform(new URShiftXNode(zend, shift) );
3173 }
3174
3175 // Bulk clear double-words
3176 Node* zsize = phase->transform(new SubXNode(zend, zbase) );
3177 Node* adr = phase->transform(new AddPNode(dest, dest, start_offset) );
3178 mem = new ClearArrayNode(ctl, mem, zsize, adr, false);
3179 return phase->transform(mem);
3180 }
3181
clear_memory(Node * ctl,Node * mem,Node * dest,intptr_t start_offset,intptr_t end_offset,PhaseGVN * phase)3182 Node* ClearArrayNode::clear_memory(Node* ctl, Node* mem, Node* dest,
3183 intptr_t start_offset,
3184 intptr_t end_offset,
3185 PhaseGVN* phase) {
3186 if (start_offset == end_offset) {
3187 // nothing to do
3188 return mem;
3189 }
3190
3191 assert((end_offset % BytesPerInt) == 0, "odd end offset");
3192 intptr_t done_offset = end_offset;
3193 if ((done_offset % BytesPerLong) != 0) {
3194 done_offset -= BytesPerInt;
3195 }
3196 if (done_offset > start_offset) {
3197 mem = clear_memory(ctl, mem, dest,
3198 start_offset, phase->MakeConX(done_offset), phase);
3199 }
3200 if (done_offset < end_offset) { // emit the final 32-bit store
3201 Node* adr = new AddPNode(dest, dest, phase->MakeConX(done_offset));
3202 adr = phase->transform(adr);
3203 const TypePtr* atp = TypeRawPtr::BOTTOM;
3204 mem = StoreNode::make(*phase, ctl, mem, adr, atp, phase->zerocon(T_INT), T_INT, MemNode::unordered);
3205 mem = phase->transform(mem);
3206 done_offset += BytesPerInt;
3207 }
3208 assert(done_offset == end_offset, "");
3209 return mem;
3210 }
3211
3212 //=============================================================================
MemBarNode(Compile * C,int alias_idx,Node * precedent)3213 MemBarNode::MemBarNode(Compile* C, int alias_idx, Node* precedent)
3214 : MultiNode(TypeFunc::Parms + (precedent == NULL? 0: 1)),
3215 _adr_type(C->get_adr_type(alias_idx)), _kind(Standalone)
3216 #ifdef ASSERT
3217 , _pair_idx(0)
3218 #endif
3219 {
3220 init_class_id(Class_MemBar);
3221 Node* top = C->top();
3222 init_req(TypeFunc::I_O,top);
3223 init_req(TypeFunc::FramePtr,top);
3224 init_req(TypeFunc::ReturnAdr,top);
3225 if (precedent != NULL)
3226 init_req(TypeFunc::Parms, precedent);
3227 }
3228
3229 //------------------------------cmp--------------------------------------------
hash() const3230 uint MemBarNode::hash() const { return NO_HASH; }
cmp(const Node & n) const3231 bool MemBarNode::cmp( const Node &n ) const {
3232 return (&n == this); // Always fail except on self
3233 }
3234
3235 //------------------------------make-------------------------------------------
make(Compile * C,int opcode,int atp,Node * pn)3236 MemBarNode* MemBarNode::make(Compile* C, int opcode, int atp, Node* pn) {
3237 switch (opcode) {
3238 case Op_MemBarAcquire: return new MemBarAcquireNode(C, atp, pn);
3239 case Op_LoadFence: return new LoadFenceNode(C, atp, pn);
3240 case Op_MemBarRelease: return new MemBarReleaseNode(C, atp, pn);
3241 case Op_StoreFence: return new StoreFenceNode(C, atp, pn);
3242 case Op_MemBarAcquireLock: return new MemBarAcquireLockNode(C, atp, pn);
3243 case Op_MemBarReleaseLock: return new MemBarReleaseLockNode(C, atp, pn);
3244 case Op_MemBarVolatile: return new MemBarVolatileNode(C, atp, pn);
3245 case Op_MemBarCPUOrder: return new MemBarCPUOrderNode(C, atp, pn);
3246 case Op_OnSpinWait: return new OnSpinWaitNode(C, atp, pn);
3247 case Op_Initialize: return new InitializeNode(C, atp, pn);
3248 case Op_MemBarStoreStore: return new MemBarStoreStoreNode(C, atp, pn);
3249 default: ShouldNotReachHere(); return NULL;
3250 }
3251 }
3252
remove(PhaseIterGVN * igvn)3253 void MemBarNode::remove(PhaseIterGVN *igvn) {
3254 if (outcnt() != 2) {
3255 return;
3256 }
3257 if (trailing_store() || trailing_load_store()) {
3258 MemBarNode* leading = leading_membar();
3259 if (leading != NULL) {
3260 assert(leading->trailing_membar() == this, "inconsistent leading/trailing membars");
3261 leading->remove(igvn);
3262 }
3263 }
3264 igvn->replace_node(proj_out(TypeFunc::Memory), in(TypeFunc::Memory));
3265 igvn->replace_node(proj_out(TypeFunc::Control), in(TypeFunc::Control));
3266 }
3267
3268 //------------------------------Ideal------------------------------------------
3269 // Return a node which is more "ideal" than the current node. Strip out
3270 // control copies
Ideal(PhaseGVN * phase,bool can_reshape)3271 Node *MemBarNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3272 if (remove_dead_region(phase, can_reshape)) return this;
3273 // Don't bother trying to transform a dead node
3274 if (in(0) && in(0)->is_top()) {
3275 return NULL;
3276 }
3277
3278 bool progress = false;
3279 // Eliminate volatile MemBars for scalar replaced objects.
3280 if (can_reshape && req() == (Precedent+1)) {
3281 bool eliminate = false;
3282 int opc = Opcode();
3283 if ((opc == Op_MemBarAcquire || opc == Op_MemBarVolatile)) {
3284 // Volatile field loads and stores.
3285 Node* my_mem = in(MemBarNode::Precedent);
3286 // The MembarAquire may keep an unused LoadNode alive through the Precedent edge
3287 if ((my_mem != NULL) && (opc == Op_MemBarAcquire) && (my_mem->outcnt() == 1)) {
3288 // if the Precedent is a decodeN and its input (a Load) is used at more than one place,
3289 // replace this Precedent (decodeN) with the Load instead.
3290 if ((my_mem->Opcode() == Op_DecodeN) && (my_mem->in(1)->outcnt() > 1)) {
3291 Node* load_node = my_mem->in(1);
3292 set_req(MemBarNode::Precedent, load_node);
3293 phase->is_IterGVN()->_worklist.push(my_mem);
3294 my_mem = load_node;
3295 } else {
3296 assert(my_mem->unique_out() == this, "sanity");
3297 del_req(Precedent);
3298 phase->is_IterGVN()->_worklist.push(my_mem); // remove dead node later
3299 my_mem = NULL;
3300 }
3301 progress = true;
3302 }
3303 if (my_mem != NULL && my_mem->is_Mem()) {
3304 const TypeOopPtr* t_oop = my_mem->in(MemNode::Address)->bottom_type()->isa_oopptr();
3305 // Check for scalar replaced object reference.
3306 if( t_oop != NULL && t_oop->is_known_instance_field() &&
3307 t_oop->offset() != Type::OffsetBot &&
3308 t_oop->offset() != Type::OffsetTop) {
3309 eliminate = true;
3310 }
3311 }
3312 } else if (opc == Op_MemBarRelease) {
3313 // Final field stores.
3314 Node* alloc = AllocateNode::Ideal_allocation(in(MemBarNode::Precedent), phase);
3315 if ((alloc != NULL) && alloc->is_Allocate() &&
3316 alloc->as_Allocate()->does_not_escape_thread()) {
3317 // The allocated object does not escape.
3318 eliminate = true;
3319 }
3320 }
3321 if (eliminate) {
3322 // Replace MemBar projections by its inputs.
3323 PhaseIterGVN* igvn = phase->is_IterGVN();
3324 remove(igvn);
3325 // Must return either the original node (now dead) or a new node
3326 // (Do not return a top here, since that would break the uniqueness of top.)
3327 return new ConINode(TypeInt::ZERO);
3328 }
3329 }
3330 return progress ? this : NULL;
3331 }
3332
3333 //------------------------------Value------------------------------------------
Value(PhaseGVN * phase) const3334 const Type* MemBarNode::Value(PhaseGVN* phase) const {
3335 if( !in(0) ) return Type::TOP;
3336 if( phase->type(in(0)) == Type::TOP )
3337 return Type::TOP;
3338 return TypeTuple::MEMBAR;
3339 }
3340
3341 //------------------------------match------------------------------------------
3342 // Construct projections for memory.
match(const ProjNode * proj,const Matcher * m)3343 Node *MemBarNode::match( const ProjNode *proj, const Matcher *m ) {
3344 switch (proj->_con) {
3345 case TypeFunc::Control:
3346 case TypeFunc::Memory:
3347 return new MachProjNode(this,proj->_con,RegMask::Empty,MachProjNode::unmatched_proj);
3348 }
3349 ShouldNotReachHere();
3350 return NULL;
3351 }
3352
set_store_pair(MemBarNode * leading,MemBarNode * trailing)3353 void MemBarNode::set_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3354 trailing->_kind = TrailingStore;
3355 leading->_kind = LeadingStore;
3356 #ifdef ASSERT
3357 trailing->_pair_idx = leading->_idx;
3358 leading->_pair_idx = leading->_idx;
3359 #endif
3360 }
3361
set_load_store_pair(MemBarNode * leading,MemBarNode * trailing)3362 void MemBarNode::set_load_store_pair(MemBarNode* leading, MemBarNode* trailing) {
3363 trailing->_kind = TrailingLoadStore;
3364 leading->_kind = LeadingLoadStore;
3365 #ifdef ASSERT
3366 trailing->_pair_idx = leading->_idx;
3367 leading->_pair_idx = leading->_idx;
3368 #endif
3369 }
3370
trailing_membar() const3371 MemBarNode* MemBarNode::trailing_membar() const {
3372 ResourceMark rm;
3373 Node* trailing = (Node*)this;
3374 VectorSet seen;
3375 Node_Stack multis(0);
3376 do {
3377 Node* c = trailing;
3378 uint i = 0;
3379 do {
3380 trailing = NULL;
3381 for (; i < c->outcnt(); i++) {
3382 Node* next = c->raw_out(i);
3383 if (next != c && next->is_CFG()) {
3384 if (c->is_MultiBranch()) {
3385 if (multis.node() == c) {
3386 multis.set_index(i+1);
3387 } else {
3388 multis.push(c, i+1);
3389 }
3390 }
3391 trailing = next;
3392 break;
3393 }
3394 }
3395 if (trailing != NULL && !seen.test_set(trailing->_idx)) {
3396 break;
3397 }
3398 while (multis.size() > 0) {
3399 c = multis.node();
3400 i = multis.index();
3401 if (i < c->req()) {
3402 break;
3403 }
3404 multis.pop();
3405 }
3406 } while (multis.size() > 0);
3407 } while (!trailing->is_MemBar() || !trailing->as_MemBar()->trailing());
3408
3409 MemBarNode* mb = trailing->as_MemBar();
3410 assert((mb->_kind == TrailingStore && _kind == LeadingStore) ||
3411 (mb->_kind == TrailingLoadStore && _kind == LeadingLoadStore), "bad trailing membar");
3412 assert(mb->_pair_idx == _pair_idx, "bad trailing membar");
3413 return mb;
3414 }
3415
leading_membar() const3416 MemBarNode* MemBarNode::leading_membar() const {
3417 ResourceMark rm;
3418 VectorSet seen;
3419 Node_Stack regions(0);
3420 Node* leading = in(0);
3421 while (leading != NULL && (!leading->is_MemBar() || !leading->as_MemBar()->leading())) {
3422 while (leading == NULL || leading->is_top() || seen.test_set(leading->_idx)) {
3423 leading = NULL;
3424 while (regions.size() > 0 && leading == NULL) {
3425 Node* r = regions.node();
3426 uint i = regions.index();
3427 if (i < r->req()) {
3428 leading = r->in(i);
3429 regions.set_index(i+1);
3430 } else {
3431 regions.pop();
3432 }
3433 }
3434 if (leading == NULL) {
3435 assert(regions.size() == 0, "all paths should have been tried");
3436 return NULL;
3437 }
3438 }
3439 if (leading->is_Region()) {
3440 regions.push(leading, 2);
3441 leading = leading->in(1);
3442 } else {
3443 leading = leading->in(0);
3444 }
3445 }
3446 #ifdef ASSERT
3447 Unique_Node_List wq;
3448 wq.push((Node*)this);
3449 uint found = 0;
3450 for (uint i = 0; i < wq.size(); i++) {
3451 Node* n = wq.at(i);
3452 if (n->is_Region()) {
3453 for (uint j = 1; j < n->req(); j++) {
3454 Node* in = n->in(j);
3455 if (in != NULL && !in->is_top()) {
3456 wq.push(in);
3457 }
3458 }
3459 } else {
3460 if (n->is_MemBar() && n->as_MemBar()->leading()) {
3461 assert(n == leading, "consistency check failed");
3462 found++;
3463 } else {
3464 Node* in = n->in(0);
3465 if (in != NULL && !in->is_top()) {
3466 wq.push(in);
3467 }
3468 }
3469 }
3470 }
3471 assert(found == 1 || (found == 0 && leading == NULL), "consistency check failed");
3472 #endif
3473 if (leading == NULL) {
3474 return NULL;
3475 }
3476 MemBarNode* mb = leading->as_MemBar();
3477 assert((mb->_kind == LeadingStore && _kind == TrailingStore) ||
3478 (mb->_kind == LeadingLoadStore && _kind == TrailingLoadStore), "bad leading membar");
3479 assert(mb->_pair_idx == _pair_idx, "bad leading membar");
3480 return mb;
3481 }
3482
3483 //===========================InitializeNode====================================
3484 // SUMMARY:
3485 // This node acts as a memory barrier on raw memory, after some raw stores.
3486 // The 'cooked' oop value feeds from the Initialize, not the Allocation.
3487 // The Initialize can 'capture' suitably constrained stores as raw inits.
3488 // It can coalesce related raw stores into larger units (called 'tiles').
3489 // It can avoid zeroing new storage for memory units which have raw inits.
3490 // At macro-expansion, it is marked 'complete', and does not optimize further.
3491 //
3492 // EXAMPLE:
3493 // The object 'new short[2]' occupies 16 bytes in a 32-bit machine.
3494 // ctl = incoming control; mem* = incoming memory
3495 // (Note: A star * on a memory edge denotes I/O and other standard edges.)
3496 // First allocate uninitialized memory and fill in the header:
3497 // alloc = (Allocate ctl mem* 16 #short[].klass ...)
3498 // ctl := alloc.Control; mem* := alloc.Memory*
3499 // rawmem = alloc.Memory; rawoop = alloc.RawAddress
3500 // Then initialize to zero the non-header parts of the raw memory block:
3501 // init = (Initialize alloc.Control alloc.Memory* alloc.RawAddress)
3502 // ctl := init.Control; mem.SLICE(#short[*]) := init.Memory
3503 // After the initialize node executes, the object is ready for service:
3504 // oop := (CheckCastPP init.Control alloc.RawAddress #short[])
3505 // Suppose its body is immediately initialized as {1,2}:
3506 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3507 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3508 // mem.SLICE(#short[*]) := store2
3509 //
3510 // DETAILS:
3511 // An InitializeNode collects and isolates object initialization after
3512 // an AllocateNode and before the next possible safepoint. As a
3513 // memory barrier (MemBarNode), it keeps critical stores from drifting
3514 // down past any safepoint or any publication of the allocation.
3515 // Before this barrier, a newly-allocated object may have uninitialized bits.
3516 // After this barrier, it may be treated as a real oop, and GC is allowed.
3517 //
3518 // The semantics of the InitializeNode include an implicit zeroing of
3519 // the new object from object header to the end of the object.
3520 // (The object header and end are determined by the AllocateNode.)
3521 //
3522 // Certain stores may be added as direct inputs to the InitializeNode.
3523 // These stores must update raw memory, and they must be to addresses
3524 // derived from the raw address produced by AllocateNode, and with
3525 // a constant offset. They must be ordered by increasing offset.
3526 // The first one is at in(RawStores), the last at in(req()-1).
3527 // Unlike most memory operations, they are not linked in a chain,
3528 // but are displayed in parallel as users of the rawmem output of
3529 // the allocation.
3530 //
3531 // (See comments in InitializeNode::capture_store, which continue
3532 // the example given above.)
3533 //
3534 // When the associated Allocate is macro-expanded, the InitializeNode
3535 // may be rewritten to optimize collected stores. A ClearArrayNode
3536 // may also be created at that point to represent any required zeroing.
3537 // The InitializeNode is then marked 'complete', prohibiting further
3538 // capturing of nearby memory operations.
3539 //
3540 // During macro-expansion, all captured initializations which store
3541 // constant values of 32 bits or smaller are coalesced (if advantageous)
3542 // into larger 'tiles' 32 or 64 bits. This allows an object to be
3543 // initialized in fewer memory operations. Memory words which are
3544 // covered by neither tiles nor non-constant stores are pre-zeroed
3545 // by explicit stores of zero. (The code shape happens to do all
3546 // zeroing first, then all other stores, with both sequences occurring
3547 // in order of ascending offsets.)
3548 //
3549 // Alternatively, code may be inserted between an AllocateNode and its
3550 // InitializeNode, to perform arbitrary initialization of the new object.
3551 // E.g., the object copying intrinsics insert complex data transfers here.
3552 // The initialization must then be marked as 'complete' disable the
3553 // built-in zeroing semantics and the collection of initializing stores.
3554 //
3555 // While an InitializeNode is incomplete, reads from the memory state
3556 // produced by it are optimizable if they match the control edge and
3557 // new oop address associated with the allocation/initialization.
3558 // They return a stored value (if the offset matches) or else zero.
3559 // A write to the memory state, if it matches control and address,
3560 // and if it is to a constant offset, may be 'captured' by the
3561 // InitializeNode. It is cloned as a raw memory operation and rewired
3562 // inside the initialization, to the raw oop produced by the allocation.
3563 // Operations on addresses which are provably distinct (e.g., to
3564 // other AllocateNodes) are allowed to bypass the initialization.
3565 //
3566 // The effect of all this is to consolidate object initialization
3567 // (both arrays and non-arrays, both piecewise and bulk) into a
3568 // single location, where it can be optimized as a unit.
3569 //
3570 // Only stores with an offset less than TrackedInitializationLimit words
3571 // will be considered for capture by an InitializeNode. This puts a
3572 // reasonable limit on the complexity of optimized initializations.
3573
3574 //---------------------------InitializeNode------------------------------------
InitializeNode(Compile * C,int adr_type,Node * rawoop)3575 InitializeNode::InitializeNode(Compile* C, int adr_type, Node* rawoop)
3576 : MemBarNode(C, adr_type, rawoop),
3577 _is_complete(Incomplete), _does_not_escape(false)
3578 {
3579 init_class_id(Class_Initialize);
3580
3581 assert(adr_type == Compile::AliasIdxRaw, "only valid atp");
3582 assert(in(RawAddress) == rawoop, "proper init");
3583 // Note: allocation() can be NULL, for secondary initialization barriers
3584 }
3585
3586 // Since this node is not matched, it will be processed by the
3587 // register allocator. Declare that there are no constraints
3588 // on the allocation of the RawAddress edge.
in_RegMask(uint idx) const3589 const RegMask &InitializeNode::in_RegMask(uint idx) const {
3590 // This edge should be set to top, by the set_complete. But be conservative.
3591 if (idx == InitializeNode::RawAddress)
3592 return *(Compile::current()->matcher()->idealreg2spillmask[in(idx)->ideal_reg()]);
3593 return RegMask::Empty;
3594 }
3595
memory(uint alias_idx)3596 Node* InitializeNode::memory(uint alias_idx) {
3597 Node* mem = in(Memory);
3598 if (mem->is_MergeMem()) {
3599 return mem->as_MergeMem()->memory_at(alias_idx);
3600 } else {
3601 // incoming raw memory is not split
3602 return mem;
3603 }
3604 }
3605
is_non_zero()3606 bool InitializeNode::is_non_zero() {
3607 if (is_complete()) return false;
3608 remove_extra_zeroes();
3609 return (req() > RawStores);
3610 }
3611
set_complete(PhaseGVN * phase)3612 void InitializeNode::set_complete(PhaseGVN* phase) {
3613 assert(!is_complete(), "caller responsibility");
3614 _is_complete = Complete;
3615
3616 // After this node is complete, it contains a bunch of
3617 // raw-memory initializations. There is no need for
3618 // it to have anything to do with non-raw memory effects.
3619 // Therefore, tell all non-raw users to re-optimize themselves,
3620 // after skipping the memory effects of this initialization.
3621 PhaseIterGVN* igvn = phase->is_IterGVN();
3622 if (igvn) igvn->add_users_to_worklist(this);
3623 }
3624
3625 // convenience function
3626 // return false if the init contains any stores already
maybe_set_complete(PhaseGVN * phase)3627 bool AllocateNode::maybe_set_complete(PhaseGVN* phase) {
3628 InitializeNode* init = initialization();
3629 if (init == NULL || init->is_complete()) return false;
3630 init->remove_extra_zeroes();
3631 // for now, if this allocation has already collected any inits, bail:
3632 if (init->is_non_zero()) return false;
3633 init->set_complete(phase);
3634 return true;
3635 }
3636
remove_extra_zeroes()3637 void InitializeNode::remove_extra_zeroes() {
3638 if (req() == RawStores) return;
3639 Node* zmem = zero_memory();
3640 uint fill = RawStores;
3641 for (uint i = fill; i < req(); i++) {
3642 Node* n = in(i);
3643 if (n->is_top() || n == zmem) continue; // skip
3644 if (fill < i) set_req(fill, n); // compact
3645 ++fill;
3646 }
3647 // delete any empty spaces created:
3648 while (fill < req()) {
3649 del_req(fill);
3650 }
3651 }
3652
3653 // Helper for remembering which stores go with which offsets.
get_store_offset(Node * st,PhaseTransform * phase)3654 intptr_t InitializeNode::get_store_offset(Node* st, PhaseTransform* phase) {
3655 if (!st->is_Store()) return -1; // can happen to dead code via subsume_node
3656 intptr_t offset = -1;
3657 Node* base = AddPNode::Ideal_base_and_offset(st->in(MemNode::Address),
3658 phase, offset);
3659 if (base == NULL) return -1; // something is dead,
3660 if (offset < 0) return -1; // dead, dead
3661 return offset;
3662 }
3663
3664 // Helper for proving that an initialization expression is
3665 // "simple enough" to be folded into an object initialization.
3666 // Attempts to prove that a store's initial value 'n' can be captured
3667 // within the initialization without creating a vicious cycle, such as:
3668 // { Foo p = new Foo(); p.next = p; }
3669 // True for constants and parameters and small combinations thereof.
detect_init_independence(Node * value,PhaseGVN * phase)3670 bool InitializeNode::detect_init_independence(Node* value, PhaseGVN* phase) {
3671 ResourceMark rm;
3672 Unique_Node_List worklist;
3673 worklist.push(value);
3674
3675 uint complexity_limit = 20;
3676 for (uint j = 0; j < worklist.size(); j++) {
3677 if (j >= complexity_limit) {
3678 return false; // Bail out if processed too many nodes
3679 }
3680
3681 Node* n = worklist.at(j);
3682 if (n == NULL) continue; // (can this really happen?)
3683 if (n->is_Proj()) n = n->in(0);
3684 if (n == this) return false; // found a cycle
3685 if (n->is_Con()) continue;
3686 if (n->is_Start()) continue; // params, etc., are OK
3687 if (n->is_Root()) continue; // even better
3688
3689 // There cannot be any dependency if 'n' is a CFG node that dominates the current allocation
3690 if (n->is_CFG() && phase->is_dominator(n, allocation())) {
3691 continue;
3692 }
3693
3694 Node* ctl = n->in(0);
3695 if (ctl != NULL && !ctl->is_top()) {
3696 if (ctl->is_Proj()) ctl = ctl->in(0);
3697 if (ctl == this) return false;
3698
3699 // If we already know that the enclosing memory op is pinned right after
3700 // the init, then any control flow that the store has picked up
3701 // must have preceded the init, or else be equal to the init.
3702 // Even after loop optimizations (which might change control edges)
3703 // a store is never pinned *before* the availability of its inputs.
3704 if (!MemNode::all_controls_dominate(n, this))
3705 return false; // failed to prove a good control
3706 }
3707
3708 // Check data edges for possible dependencies on 'this'.
3709 for (uint i = 1; i < n->req(); i++) {
3710 Node* m = n->in(i);
3711 if (m == NULL || m == n || m->is_top()) continue;
3712
3713 // Only process data inputs once
3714 worklist.push(m);
3715 }
3716 }
3717
3718 return true;
3719 }
3720
3721 // Here are all the checks a Store must pass before it can be moved into
3722 // an initialization. Returns zero if a check fails.
3723 // On success, returns the (constant) offset to which the store applies,
3724 // within the initialized memory.
can_capture_store(StoreNode * st,PhaseGVN * phase,bool can_reshape)3725 intptr_t InitializeNode::can_capture_store(StoreNode* st, PhaseGVN* phase, bool can_reshape) {
3726 const int FAIL = 0;
3727 if (st->req() != MemNode::ValueIn + 1)
3728 return FAIL; // an inscrutable StoreNode (card mark?)
3729 Node* ctl = st->in(MemNode::Control);
3730 if (!(ctl != NULL && ctl->is_Proj() && ctl->in(0) == this))
3731 return FAIL; // must be unconditional after the initialization
3732 Node* mem = st->in(MemNode::Memory);
3733 if (!(mem->is_Proj() && mem->in(0) == this))
3734 return FAIL; // must not be preceded by other stores
3735 Node* adr = st->in(MemNode::Address);
3736 intptr_t offset;
3737 AllocateNode* alloc = AllocateNode::Ideal_allocation(adr, phase, offset);
3738 if (alloc == NULL)
3739 return FAIL; // inscrutable address
3740 if (alloc != allocation())
3741 return FAIL; // wrong allocation! (store needs to float up)
3742 int size_in_bytes = st->memory_size();
3743 if ((size_in_bytes != 0) && (offset % size_in_bytes) != 0) {
3744 return FAIL; // mismatched access
3745 }
3746 Node* val = st->in(MemNode::ValueIn);
3747
3748 if (!detect_init_independence(val, phase))
3749 return FAIL; // stored value must be 'simple enough'
3750
3751 // The Store can be captured only if nothing after the allocation
3752 // and before the Store is using the memory location that the store
3753 // overwrites.
3754 bool failed = false;
3755 // If is_complete_with_arraycopy() is true the shape of the graph is
3756 // well defined and is safe so no need for extra checks.
3757 if (!is_complete_with_arraycopy()) {
3758 // We are going to look at each use of the memory state following
3759 // the allocation to make sure nothing reads the memory that the
3760 // Store writes.
3761 const TypePtr* t_adr = phase->type(adr)->isa_ptr();
3762 int alias_idx = phase->C->get_alias_index(t_adr);
3763 ResourceMark rm;
3764 Unique_Node_List mems;
3765 mems.push(mem);
3766 Node* unique_merge = NULL;
3767 for (uint next = 0; next < mems.size(); ++next) {
3768 Node *m = mems.at(next);
3769 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) {
3770 Node *n = m->fast_out(j);
3771 if (n->outcnt() == 0) {
3772 continue;
3773 }
3774 if (n == st) {
3775 continue;
3776 } else if (n->in(0) != NULL && n->in(0) != ctl) {
3777 // If the control of this use is different from the control
3778 // of the Store which is right after the InitializeNode then
3779 // this node cannot be between the InitializeNode and the
3780 // Store.
3781 continue;
3782 } else if (n->is_MergeMem()) {
3783 if (n->as_MergeMem()->memory_at(alias_idx) == m) {
3784 // We can hit a MergeMemNode (that will likely go away
3785 // later) that is a direct use of the memory state
3786 // following the InitializeNode on the same slice as the
3787 // store node that we'd like to capture. We need to check
3788 // the uses of the MergeMemNode.
3789 mems.push(n);
3790 }
3791 } else if (n->is_Mem()) {
3792 Node* other_adr = n->in(MemNode::Address);
3793 if (other_adr == adr) {
3794 failed = true;
3795 break;
3796 } else {
3797 const TypePtr* other_t_adr = phase->type(other_adr)->isa_ptr();
3798 if (other_t_adr != NULL) {
3799 int other_alias_idx = phase->C->get_alias_index(other_t_adr);
3800 if (other_alias_idx == alias_idx) {
3801 // A load from the same memory slice as the store right
3802 // after the InitializeNode. We check the control of the
3803 // object/array that is loaded from. If it's the same as
3804 // the store control then we cannot capture the store.
3805 assert(!n->is_Store(), "2 stores to same slice on same control?");
3806 Node* base = other_adr;
3807 assert(base->is_AddP(), "should be addp but is %s", base->Name());
3808 base = base->in(AddPNode::Base);
3809 if (base != NULL) {
3810 base = base->uncast();
3811 if (base->is_Proj() && base->in(0) == alloc) {
3812 failed = true;
3813 break;
3814 }
3815 }
3816 }
3817 }
3818 }
3819 } else {
3820 failed = true;
3821 break;
3822 }
3823 }
3824 }
3825 }
3826 if (failed) {
3827 if (!can_reshape) {
3828 // We decided we couldn't capture the store during parsing. We
3829 // should try again during the next IGVN once the graph is
3830 // cleaner.
3831 phase->C->record_for_igvn(st);
3832 }
3833 return FAIL;
3834 }
3835
3836 return offset; // success
3837 }
3838
3839 // Find the captured store in(i) which corresponds to the range
3840 // [start..start+size) in the initialized object.
3841 // If there is one, return its index i. If there isn't, return the
3842 // negative of the index where it should be inserted.
3843 // Return 0 if the queried range overlaps an initialization boundary
3844 // or if dead code is encountered.
3845 // If size_in_bytes is zero, do not bother with overlap checks.
captured_store_insertion_point(intptr_t start,int size_in_bytes,PhaseTransform * phase)3846 int InitializeNode::captured_store_insertion_point(intptr_t start,
3847 int size_in_bytes,
3848 PhaseTransform* phase) {
3849 const int FAIL = 0, MAX_STORE = MAX2(BytesPerLong, (int)MaxVectorSize);
3850
3851 if (is_complete())
3852 return FAIL; // arraycopy got here first; punt
3853
3854 assert(allocation() != NULL, "must be present");
3855
3856 // no negatives, no header fields:
3857 if (start < (intptr_t) allocation()->minimum_header_size()) return FAIL;
3858
3859 // after a certain size, we bail out on tracking all the stores:
3860 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
3861 if (start >= ti_limit) return FAIL;
3862
3863 for (uint i = InitializeNode::RawStores, limit = req(); ; ) {
3864 if (i >= limit) return -(int)i; // not found; here is where to put it
3865
3866 Node* st = in(i);
3867 intptr_t st_off = get_store_offset(st, phase);
3868 if (st_off < 0) {
3869 if (st != zero_memory()) {
3870 return FAIL; // bail out if there is dead garbage
3871 }
3872 } else if (st_off > start) {
3873 // ...we are done, since stores are ordered
3874 if (st_off < start + size_in_bytes) {
3875 return FAIL; // the next store overlaps
3876 }
3877 return -(int)i; // not found; here is where to put it
3878 } else if (st_off < start) {
3879 assert(st->as_Store()->memory_size() <= MAX_STORE, "");
3880 if (size_in_bytes != 0 &&
3881 start < st_off + MAX_STORE &&
3882 start < st_off + st->as_Store()->memory_size()) {
3883 return FAIL; // the previous store overlaps
3884 }
3885 } else {
3886 if (size_in_bytes != 0 &&
3887 st->as_Store()->memory_size() != size_in_bytes) {
3888 return FAIL; // mismatched store size
3889 }
3890 return i;
3891 }
3892
3893 ++i;
3894 }
3895 }
3896
3897 // Look for a captured store which initializes at the offset 'start'
3898 // with the given size. If there is no such store, and no other
3899 // initialization interferes, then return zero_memory (the memory
3900 // projection of the AllocateNode).
find_captured_store(intptr_t start,int size_in_bytes,PhaseTransform * phase)3901 Node* InitializeNode::find_captured_store(intptr_t start, int size_in_bytes,
3902 PhaseTransform* phase) {
3903 assert(stores_are_sane(phase), "");
3904 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3905 if (i == 0) {
3906 return NULL; // something is dead
3907 } else if (i < 0) {
3908 return zero_memory(); // just primordial zero bits here
3909 } else {
3910 Node* st = in(i); // here is the store at this position
3911 assert(get_store_offset(st->as_Store(), phase) == start, "sanity");
3912 return st;
3913 }
3914 }
3915
3916 // Create, as a raw pointer, an address within my new object at 'offset'.
make_raw_address(intptr_t offset,PhaseTransform * phase)3917 Node* InitializeNode::make_raw_address(intptr_t offset,
3918 PhaseTransform* phase) {
3919 Node* addr = in(RawAddress);
3920 if (offset != 0) {
3921 Compile* C = phase->C;
3922 addr = phase->transform( new AddPNode(C->top(), addr,
3923 phase->MakeConX(offset)) );
3924 }
3925 return addr;
3926 }
3927
3928 // Clone the given store, converting it into a raw store
3929 // initializing a field or element of my new object.
3930 // Caller is responsible for retiring the original store,
3931 // with subsume_node or the like.
3932 //
3933 // From the example above InitializeNode::InitializeNode,
3934 // here are the old stores to be captured:
3935 // store1 = (StoreC init.Control init.Memory (+ oop 12) 1)
3936 // store2 = (StoreC init.Control store1 (+ oop 14) 2)
3937 //
3938 // Here is the changed code; note the extra edges on init:
3939 // alloc = (Allocate ...)
3940 // rawoop = alloc.RawAddress
3941 // rawstore1 = (StoreC alloc.Control alloc.Memory (+ rawoop 12) 1)
3942 // rawstore2 = (StoreC alloc.Control alloc.Memory (+ rawoop 14) 2)
3943 // init = (Initialize alloc.Control alloc.Memory rawoop
3944 // rawstore1 rawstore2)
3945 //
capture_store(StoreNode * st,intptr_t start,PhaseGVN * phase,bool can_reshape)3946 Node* InitializeNode::capture_store(StoreNode* st, intptr_t start,
3947 PhaseGVN* phase, bool can_reshape) {
3948 assert(stores_are_sane(phase), "");
3949
3950 if (start < 0) return NULL;
3951 assert(can_capture_store(st, phase, can_reshape) == start, "sanity");
3952
3953 Compile* C = phase->C;
3954 int size_in_bytes = st->memory_size();
3955 int i = captured_store_insertion_point(start, size_in_bytes, phase);
3956 if (i == 0) return NULL; // bail out
3957 Node* prev_mem = NULL; // raw memory for the captured store
3958 if (i > 0) {
3959 prev_mem = in(i); // there is a pre-existing store under this one
3960 set_req(i, C->top()); // temporarily disconnect it
3961 // See StoreNode::Ideal 'st->outcnt() == 1' for the reason to disconnect.
3962 } else {
3963 i = -i; // no pre-existing store
3964 prev_mem = zero_memory(); // a slice of the newly allocated object
3965 if (i > InitializeNode::RawStores && in(i-1) == prev_mem)
3966 set_req(--i, C->top()); // reuse this edge; it has been folded away
3967 else
3968 ins_req(i, C->top()); // build a new edge
3969 }
3970 Node* new_st = st->clone();
3971 new_st->set_req(MemNode::Control, in(Control));
3972 new_st->set_req(MemNode::Memory, prev_mem);
3973 new_st->set_req(MemNode::Address, make_raw_address(start, phase));
3974 new_st = phase->transform(new_st);
3975
3976 // At this point, new_st might have swallowed a pre-existing store
3977 // at the same offset, or perhaps new_st might have disappeared,
3978 // if it redundantly stored the same value (or zero to fresh memory).
3979
3980 // In any case, wire it in:
3981 PhaseIterGVN* igvn = phase->is_IterGVN();
3982 if (igvn) {
3983 igvn->rehash_node_delayed(this);
3984 }
3985 set_req(i, new_st);
3986
3987 // The caller may now kill the old guy.
3988 DEBUG_ONLY(Node* check_st = find_captured_store(start, size_in_bytes, phase));
3989 assert(check_st == new_st || check_st == NULL, "must be findable");
3990 assert(!is_complete(), "");
3991 return new_st;
3992 }
3993
store_constant(jlong * tiles,int num_tiles,intptr_t st_off,int st_size,jlong con)3994 static bool store_constant(jlong* tiles, int num_tiles,
3995 intptr_t st_off, int st_size,
3996 jlong con) {
3997 if ((st_off & (st_size-1)) != 0)
3998 return false; // strange store offset (assume size==2**N)
3999 address addr = (address)tiles + st_off;
4000 assert(st_off >= 0 && addr+st_size <= (address)&tiles[num_tiles], "oob");
4001 switch (st_size) {
4002 case sizeof(jbyte): *(jbyte*) addr = (jbyte) con; break;
4003 case sizeof(jchar): *(jchar*) addr = (jchar) con; break;
4004 case sizeof(jint): *(jint*) addr = (jint) con; break;
4005 case sizeof(jlong): *(jlong*) addr = (jlong) con; break;
4006 default: return false; // strange store size (detect size!=2**N here)
4007 }
4008 return true; // return success to caller
4009 }
4010
4011 // Coalesce subword constants into int constants and possibly
4012 // into long constants. The goal, if the CPU permits,
4013 // is to initialize the object with a small number of 64-bit tiles.
4014 // Also, convert floating-point constants to bit patterns.
4015 // Non-constants are not relevant to this pass.
4016 //
4017 // In terms of the running example on InitializeNode::InitializeNode
4018 // and InitializeNode::capture_store, here is the transformation
4019 // of rawstore1 and rawstore2 into rawstore12:
4020 // alloc = (Allocate ...)
4021 // rawoop = alloc.RawAddress
4022 // tile12 = 0x00010002
4023 // rawstore12 = (StoreI alloc.Control alloc.Memory (+ rawoop 12) tile12)
4024 // init = (Initialize alloc.Control alloc.Memory rawoop rawstore12)
4025 //
4026 void
coalesce_subword_stores(intptr_t header_size,Node * size_in_bytes,PhaseGVN * phase)4027 InitializeNode::coalesce_subword_stores(intptr_t header_size,
4028 Node* size_in_bytes,
4029 PhaseGVN* phase) {
4030 Compile* C = phase->C;
4031
4032 assert(stores_are_sane(phase), "");
4033 // Note: After this pass, they are not completely sane,
4034 // since there may be some overlaps.
4035
4036 int old_subword = 0, old_long = 0, new_int = 0, new_long = 0;
4037
4038 intptr_t ti_limit = (TrackedInitializationLimit * HeapWordSize);
4039 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, ti_limit);
4040 size_limit = MIN2(size_limit, ti_limit);
4041 size_limit = align_up(size_limit, BytesPerLong);
4042 int num_tiles = size_limit / BytesPerLong;
4043
4044 // allocate space for the tile map:
4045 const int small_len = DEBUG_ONLY(true ? 3 :) 30; // keep stack frames small
4046 jlong tiles_buf[small_len];
4047 Node* nodes_buf[small_len];
4048 jlong inits_buf[small_len];
4049 jlong* tiles = ((num_tiles <= small_len) ? &tiles_buf[0]
4050 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4051 Node** nodes = ((num_tiles <= small_len) ? &nodes_buf[0]
4052 : NEW_RESOURCE_ARRAY(Node*, num_tiles));
4053 jlong* inits = ((num_tiles <= small_len) ? &inits_buf[0]
4054 : NEW_RESOURCE_ARRAY(jlong, num_tiles));
4055 // tiles: exact bitwise model of all primitive constants
4056 // nodes: last constant-storing node subsumed into the tiles model
4057 // inits: which bytes (in each tile) are touched by any initializations
4058
4059 //// Pass A: Fill in the tile model with any relevant stores.
4060
4061 Copy::zero_to_bytes(tiles, sizeof(tiles[0]) * num_tiles);
4062 Copy::zero_to_bytes(nodes, sizeof(nodes[0]) * num_tiles);
4063 Copy::zero_to_bytes(inits, sizeof(inits[0]) * num_tiles);
4064 Node* zmem = zero_memory(); // initially zero memory state
4065 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4066 Node* st = in(i);
4067 intptr_t st_off = get_store_offset(st, phase);
4068
4069 // Figure out the store's offset and constant value:
4070 if (st_off < header_size) continue; //skip (ignore header)
4071 if (st->in(MemNode::Memory) != zmem) continue; //skip (odd store chain)
4072 int st_size = st->as_Store()->memory_size();
4073 if (st_off + st_size > size_limit) break;
4074
4075 // Record which bytes are touched, whether by constant or not.
4076 if (!store_constant(inits, num_tiles, st_off, st_size, (jlong) -1))
4077 continue; // skip (strange store size)
4078
4079 const Type* val = phase->type(st->in(MemNode::ValueIn));
4080 if (!val->singleton()) continue; //skip (non-con store)
4081 BasicType type = val->basic_type();
4082
4083 jlong con = 0;
4084 switch (type) {
4085 case T_INT: con = val->is_int()->get_con(); break;
4086 case T_LONG: con = val->is_long()->get_con(); break;
4087 case T_FLOAT: con = jint_cast(val->getf()); break;
4088 case T_DOUBLE: con = jlong_cast(val->getd()); break;
4089 default: continue; //skip (odd store type)
4090 }
4091
4092 if (type == T_LONG && Matcher::isSimpleConstant64(con) &&
4093 st->Opcode() == Op_StoreL) {
4094 continue; // This StoreL is already optimal.
4095 }
4096
4097 // Store down the constant.
4098 store_constant(tiles, num_tiles, st_off, st_size, con);
4099
4100 intptr_t j = st_off >> LogBytesPerLong;
4101
4102 if (type == T_INT && st_size == BytesPerInt
4103 && (st_off & BytesPerInt) == BytesPerInt) {
4104 jlong lcon = tiles[j];
4105 if (!Matcher::isSimpleConstant64(lcon) &&
4106 st->Opcode() == Op_StoreI) {
4107 // This StoreI is already optimal by itself.
4108 jint* intcon = (jint*) &tiles[j];
4109 intcon[1] = 0; // undo the store_constant()
4110
4111 // If the previous store is also optimal by itself, back up and
4112 // undo the action of the previous loop iteration... if we can.
4113 // But if we can't, just let the previous half take care of itself.
4114 st = nodes[j];
4115 st_off -= BytesPerInt;
4116 con = intcon[0];
4117 if (con != 0 && st != NULL && st->Opcode() == Op_StoreI) {
4118 assert(st_off >= header_size, "still ignoring header");
4119 assert(get_store_offset(st, phase) == st_off, "must be");
4120 assert(in(i-1) == zmem, "must be");
4121 DEBUG_ONLY(const Type* tcon = phase->type(st->in(MemNode::ValueIn)));
4122 assert(con == tcon->is_int()->get_con(), "must be");
4123 // Undo the effects of the previous loop trip, which swallowed st:
4124 intcon[0] = 0; // undo store_constant()
4125 set_req(i-1, st); // undo set_req(i, zmem)
4126 nodes[j] = NULL; // undo nodes[j] = st
4127 --old_subword; // undo ++old_subword
4128 }
4129 continue; // This StoreI is already optimal.
4130 }
4131 }
4132
4133 // This store is not needed.
4134 set_req(i, zmem);
4135 nodes[j] = st; // record for the moment
4136 if (st_size < BytesPerLong) // something has changed
4137 ++old_subword; // includes int/float, but who's counting...
4138 else ++old_long;
4139 }
4140
4141 if ((old_subword + old_long) == 0)
4142 return; // nothing more to do
4143
4144 //// Pass B: Convert any non-zero tiles into optimal constant stores.
4145 // Be sure to insert them before overlapping non-constant stores.
4146 // (E.g., byte[] x = { 1,2,y,4 } => x[int 0] = 0x01020004, x[2]=y.)
4147 for (int j = 0; j < num_tiles; j++) {
4148 jlong con = tiles[j];
4149 jlong init = inits[j];
4150 if (con == 0) continue;
4151 jint con0, con1; // split the constant, address-wise
4152 jint init0, init1; // split the init map, address-wise
4153 { union { jlong con; jint intcon[2]; } u;
4154 u.con = con;
4155 con0 = u.intcon[0];
4156 con1 = u.intcon[1];
4157 u.con = init;
4158 init0 = u.intcon[0];
4159 init1 = u.intcon[1];
4160 }
4161
4162 Node* old = nodes[j];
4163 assert(old != NULL, "need the prior store");
4164 intptr_t offset = (j * BytesPerLong);
4165
4166 bool split = !Matcher::isSimpleConstant64(con);
4167
4168 if (offset < header_size) {
4169 assert(offset + BytesPerInt >= header_size, "second int counts");
4170 assert(*(jint*)&tiles[j] == 0, "junk in header");
4171 split = true; // only the second word counts
4172 // Example: int a[] = { 42 ... }
4173 } else if (con0 == 0 && init0 == -1) {
4174 split = true; // first word is covered by full inits
4175 // Example: int a[] = { ... foo(), 42 ... }
4176 } else if (con1 == 0 && init1 == -1) {
4177 split = true; // second word is covered by full inits
4178 // Example: int a[] = { ... 42, foo() ... }
4179 }
4180
4181 // Here's a case where init0 is neither 0 nor -1:
4182 // byte a[] = { ... 0,0,foo(),0, 0,0,0,42 ... }
4183 // Assuming big-endian memory, init0, init1 are 0x0000FF00, 0x000000FF.
4184 // In this case the tile is not split; it is (jlong)42.
4185 // The big tile is stored down, and then the foo() value is inserted.
4186 // (If there were foo(),foo() instead of foo(),0, init0 would be -1.)
4187
4188 Node* ctl = old->in(MemNode::Control);
4189 Node* adr = make_raw_address(offset, phase);
4190 const TypePtr* atp = TypeRawPtr::BOTTOM;
4191
4192 // One or two coalesced stores to plop down.
4193 Node* st[2];
4194 intptr_t off[2];
4195 int nst = 0;
4196 if (!split) {
4197 ++new_long;
4198 off[nst] = offset;
4199 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4200 phase->longcon(con), T_LONG, MemNode::unordered);
4201 } else {
4202 // Omit either if it is a zero.
4203 if (con0 != 0) {
4204 ++new_int;
4205 off[nst] = offset;
4206 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4207 phase->intcon(con0), T_INT, MemNode::unordered);
4208 }
4209 if (con1 != 0) {
4210 ++new_int;
4211 offset += BytesPerInt;
4212 adr = make_raw_address(offset, phase);
4213 off[nst] = offset;
4214 st[nst++] = StoreNode::make(*phase, ctl, zmem, adr, atp,
4215 phase->intcon(con1), T_INT, MemNode::unordered);
4216 }
4217 }
4218
4219 // Insert second store first, then the first before the second.
4220 // Insert each one just before any overlapping non-constant stores.
4221 while (nst > 0) {
4222 Node* st1 = st[--nst];
4223 C->copy_node_notes_to(st1, old);
4224 st1 = phase->transform(st1);
4225 offset = off[nst];
4226 assert(offset >= header_size, "do not smash header");
4227 int ins_idx = captured_store_insertion_point(offset, /*size:*/0, phase);
4228 guarantee(ins_idx != 0, "must re-insert constant store");
4229 if (ins_idx < 0) ins_idx = -ins_idx; // never overlap
4230 if (ins_idx > InitializeNode::RawStores && in(ins_idx-1) == zmem)
4231 set_req(--ins_idx, st1);
4232 else
4233 ins_req(ins_idx, st1);
4234 }
4235 }
4236
4237 if (PrintCompilation && WizardMode)
4238 tty->print_cr("Changed %d/%d subword/long constants into %d/%d int/long",
4239 old_subword, old_long, new_int, new_long);
4240 if (C->log() != NULL)
4241 C->log()->elem("comment that='%d/%d subword/long to %d/%d int/long'",
4242 old_subword, old_long, new_int, new_long);
4243
4244 // Clean up any remaining occurrences of zmem:
4245 remove_extra_zeroes();
4246 }
4247
4248 // Explore forward from in(start) to find the first fully initialized
4249 // word, and return its offset. Skip groups of subword stores which
4250 // together initialize full words. If in(start) is itself part of a
4251 // fully initialized word, return the offset of in(start). If there
4252 // are no following full-word stores, or if something is fishy, return
4253 // a negative value.
find_next_fullword_store(uint start,PhaseGVN * phase)4254 intptr_t InitializeNode::find_next_fullword_store(uint start, PhaseGVN* phase) {
4255 int int_map = 0;
4256 intptr_t int_map_off = 0;
4257 const int FULL_MAP = right_n_bits(BytesPerInt); // the int_map we hope for
4258
4259 for (uint i = start, limit = req(); i < limit; i++) {
4260 Node* st = in(i);
4261
4262 intptr_t st_off = get_store_offset(st, phase);
4263 if (st_off < 0) break; // return conservative answer
4264
4265 int st_size = st->as_Store()->memory_size();
4266 if (st_size >= BytesPerInt && (st_off % BytesPerInt) == 0) {
4267 return st_off; // we found a complete word init
4268 }
4269
4270 // update the map:
4271
4272 intptr_t this_int_off = align_down(st_off, BytesPerInt);
4273 if (this_int_off != int_map_off) {
4274 // reset the map:
4275 int_map = 0;
4276 int_map_off = this_int_off;
4277 }
4278
4279 int subword_off = st_off - this_int_off;
4280 int_map |= right_n_bits(st_size) << subword_off;
4281 if ((int_map & FULL_MAP) == FULL_MAP) {
4282 return this_int_off; // we found a complete word init
4283 }
4284
4285 // Did this store hit or cross the word boundary?
4286 intptr_t next_int_off = align_down(st_off + st_size, BytesPerInt);
4287 if (next_int_off == this_int_off + BytesPerInt) {
4288 // We passed the current int, without fully initializing it.
4289 int_map_off = next_int_off;
4290 int_map >>= BytesPerInt;
4291 } else if (next_int_off > this_int_off + BytesPerInt) {
4292 // We passed the current and next int.
4293 return this_int_off + BytesPerInt;
4294 }
4295 }
4296
4297 return -1;
4298 }
4299
4300
4301 // Called when the associated AllocateNode is expanded into CFG.
4302 // At this point, we may perform additional optimizations.
4303 // Linearize the stores by ascending offset, to make memory
4304 // activity as coherent as possible.
complete_stores(Node * rawctl,Node * rawmem,Node * rawptr,intptr_t header_size,Node * size_in_bytes,PhaseIterGVN * phase)4305 Node* InitializeNode::complete_stores(Node* rawctl, Node* rawmem, Node* rawptr,
4306 intptr_t header_size,
4307 Node* size_in_bytes,
4308 PhaseIterGVN* phase) {
4309 assert(!is_complete(), "not already complete");
4310 assert(stores_are_sane(phase), "");
4311 assert(allocation() != NULL, "must be present");
4312
4313 remove_extra_zeroes();
4314
4315 if (ReduceFieldZeroing || ReduceBulkZeroing)
4316 // reduce instruction count for common initialization patterns
4317 coalesce_subword_stores(header_size, size_in_bytes, phase);
4318
4319 Node* zmem = zero_memory(); // initially zero memory state
4320 Node* inits = zmem; // accumulating a linearized chain of inits
4321 #ifdef ASSERT
4322 intptr_t first_offset = allocation()->minimum_header_size();
4323 intptr_t last_init_off = first_offset; // previous init offset
4324 intptr_t last_init_end = first_offset; // previous init offset+size
4325 intptr_t last_tile_end = first_offset; // previous tile offset+size
4326 #endif
4327 intptr_t zeroes_done = header_size;
4328
4329 bool do_zeroing = true; // we might give up if inits are very sparse
4330 int big_init_gaps = 0; // how many large gaps have we seen?
4331
4332 if (UseTLAB && ZeroTLAB) do_zeroing = false;
4333 if (!ReduceFieldZeroing && !ReduceBulkZeroing) do_zeroing = false;
4334
4335 for (uint i = InitializeNode::RawStores, limit = req(); i < limit; i++) {
4336 Node* st = in(i);
4337 intptr_t st_off = get_store_offset(st, phase);
4338 if (st_off < 0)
4339 break; // unknown junk in the inits
4340 if (st->in(MemNode::Memory) != zmem)
4341 break; // complicated store chains somehow in list
4342
4343 int st_size = st->as_Store()->memory_size();
4344 intptr_t next_init_off = st_off + st_size;
4345
4346 if (do_zeroing && zeroes_done < next_init_off) {
4347 // See if this store needs a zero before it or under it.
4348 intptr_t zeroes_needed = st_off;
4349
4350 if (st_size < BytesPerInt) {
4351 // Look for subword stores which only partially initialize words.
4352 // If we find some, we must lay down some word-level zeroes first,
4353 // underneath the subword stores.
4354 //
4355 // Examples:
4356 // byte[] a = { p,q,r,s } => a[0]=p,a[1]=q,a[2]=r,a[3]=s
4357 // byte[] a = { x,y,0,0 } => a[0..3] = 0, a[0]=x,a[1]=y
4358 // byte[] a = { 0,0,z,0 } => a[0..3] = 0, a[2]=z
4359 //
4360 // Note: coalesce_subword_stores may have already done this,
4361 // if it was prompted by constant non-zero subword initializers.
4362 // But this case can still arise with non-constant stores.
4363
4364 intptr_t next_full_store = find_next_fullword_store(i, phase);
4365
4366 // In the examples above:
4367 // in(i) p q r s x y z
4368 // st_off 12 13 14 15 12 13 14
4369 // st_size 1 1 1 1 1 1 1
4370 // next_full_s. 12 16 16 16 16 16 16
4371 // z's_done 12 16 16 16 12 16 12
4372 // z's_needed 12 16 16 16 16 16 16
4373 // zsize 0 0 0 0 4 0 4
4374 if (next_full_store < 0) {
4375 // Conservative tack: Zero to end of current word.
4376 zeroes_needed = align_up(zeroes_needed, BytesPerInt);
4377 } else {
4378 // Zero to beginning of next fully initialized word.
4379 // Or, don't zero at all, if we are already in that word.
4380 assert(next_full_store >= zeroes_needed, "must go forward");
4381 assert((next_full_store & (BytesPerInt-1)) == 0, "even boundary");
4382 zeroes_needed = next_full_store;
4383 }
4384 }
4385
4386 if (zeroes_needed > zeroes_done) {
4387 intptr_t zsize = zeroes_needed - zeroes_done;
4388 // Do some incremental zeroing on rawmem, in parallel with inits.
4389 zeroes_done = align_down(zeroes_done, BytesPerInt);
4390 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4391 zeroes_done, zeroes_needed,
4392 phase);
4393 zeroes_done = zeroes_needed;
4394 if (zsize > InitArrayShortSize && ++big_init_gaps > 2)
4395 do_zeroing = false; // leave the hole, next time
4396 }
4397 }
4398
4399 // Collect the store and move on:
4400 phase->replace_input_of(st, MemNode::Memory, inits);
4401 inits = st; // put it on the linearized chain
4402 set_req(i, zmem); // unhook from previous position
4403
4404 if (zeroes_done == st_off)
4405 zeroes_done = next_init_off;
4406
4407 assert(!do_zeroing || zeroes_done >= next_init_off, "don't miss any");
4408
4409 #ifdef ASSERT
4410 // Various order invariants. Weaker than stores_are_sane because
4411 // a large constant tile can be filled in by smaller non-constant stores.
4412 assert(st_off >= last_init_off, "inits do not reverse");
4413 last_init_off = st_off;
4414 const Type* val = NULL;
4415 if (st_size >= BytesPerInt &&
4416 (val = phase->type(st->in(MemNode::ValueIn)))->singleton() &&
4417 (int)val->basic_type() < (int)T_OBJECT) {
4418 assert(st_off >= last_tile_end, "tiles do not overlap");
4419 assert(st_off >= last_init_end, "tiles do not overwrite inits");
4420 last_tile_end = MAX2(last_tile_end, next_init_off);
4421 } else {
4422 intptr_t st_tile_end = align_up(next_init_off, BytesPerLong);
4423 assert(st_tile_end >= last_tile_end, "inits stay with tiles");
4424 assert(st_off >= last_init_end, "inits do not overlap");
4425 last_init_end = next_init_off; // it's a non-tile
4426 }
4427 #endif //ASSERT
4428 }
4429
4430 remove_extra_zeroes(); // clear out all the zmems left over
4431 add_req(inits);
4432
4433 if (!(UseTLAB && ZeroTLAB)) {
4434 // If anything remains to be zeroed, zero it all now.
4435 zeroes_done = align_down(zeroes_done, BytesPerInt);
4436 // if it is the last unused 4 bytes of an instance, forget about it
4437 intptr_t size_limit = phase->find_intptr_t_con(size_in_bytes, max_jint);
4438 if (zeroes_done + BytesPerLong >= size_limit) {
4439 AllocateNode* alloc = allocation();
4440 assert(alloc != NULL, "must be present");
4441 if (alloc != NULL && alloc->Opcode() == Op_Allocate) {
4442 Node* klass_node = alloc->in(AllocateNode::KlassNode);
4443 ciKlass* k = phase->type(klass_node)->is_klassptr()->klass();
4444 if (zeroes_done == k->layout_helper())
4445 zeroes_done = size_limit;
4446 }
4447 }
4448 if (zeroes_done < size_limit) {
4449 rawmem = ClearArrayNode::clear_memory(rawctl, rawmem, rawptr,
4450 zeroes_done, size_in_bytes, phase);
4451 }
4452 }
4453
4454 set_complete(phase);
4455 return rawmem;
4456 }
4457
4458
4459 #ifdef ASSERT
stores_are_sane(PhaseTransform * phase)4460 bool InitializeNode::stores_are_sane(PhaseTransform* phase) {
4461 if (is_complete())
4462 return true; // stores could be anything at this point
4463 assert(allocation() != NULL, "must be present");
4464 intptr_t last_off = allocation()->minimum_header_size();
4465 for (uint i = InitializeNode::RawStores; i < req(); i++) {
4466 Node* st = in(i);
4467 intptr_t st_off = get_store_offset(st, phase);
4468 if (st_off < 0) continue; // ignore dead garbage
4469 if (last_off > st_off) {
4470 tty->print_cr("*** bad store offset at %d: " INTX_FORMAT " > " INTX_FORMAT, i, last_off, st_off);
4471 this->dump(2);
4472 assert(false, "ascending store offsets");
4473 return false;
4474 }
4475 last_off = st_off + st->as_Store()->memory_size();
4476 }
4477 return true;
4478 }
4479 #endif //ASSERT
4480
4481
4482
4483
4484 //============================MergeMemNode=====================================
4485 //
4486 // SEMANTICS OF MEMORY MERGES: A MergeMem is a memory state assembled from several
4487 // contributing store or call operations. Each contributor provides the memory
4488 // state for a particular "alias type" (see Compile::alias_type). For example,
4489 // if a MergeMem has an input X for alias category #6, then any memory reference
4490 // to alias category #6 may use X as its memory state input, as an exact equivalent
4491 // to using the MergeMem as a whole.
4492 // Load<6>( MergeMem(<6>: X, ...), p ) <==> Load<6>(X,p)
4493 //
4494 // (Here, the <N> notation gives the index of the relevant adr_type.)
4495 //
4496 // In one special case (and more cases in the future), alias categories overlap.
4497 // The special alias category "Bot" (Compile::AliasIdxBot) includes all memory
4498 // states. Therefore, if a MergeMem has only one contributing input W for Bot,
4499 // it is exactly equivalent to that state W:
4500 // MergeMem(<Bot>: W) <==> W
4501 //
4502 // Usually, the merge has more than one input. In that case, where inputs
4503 // overlap (i.e., one is Bot), the narrower alias type determines the memory
4504 // state for that type, and the wider alias type (Bot) fills in everywhere else:
4505 // Load<5>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<5>(W,p)
4506 // Load<6>( MergeMem(<Bot>: W, <6>: X), p ) <==> Load<6>(X,p)
4507 //
4508 // A merge can take a "wide" memory state as one of its narrow inputs.
4509 // This simply means that the merge observes out only the relevant parts of
4510 // the wide input. That is, wide memory states arriving at narrow merge inputs
4511 // are implicitly "filtered" or "sliced" as necessary. (This is rare.)
4512 //
4513 // These rules imply that MergeMem nodes may cascade (via their <Bot> links),
4514 // and that memory slices "leak through":
4515 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y)) <==> MergeMem(<Bot>: W, <7>: Y)
4516 //
4517 // But, in such a cascade, repeated memory slices can "block the leak":
4518 // MergeMem(<Bot>: MergeMem(<Bot>: W, <7>: Y), <7>: Y') <==> MergeMem(<Bot>: W, <7>: Y')
4519 //
4520 // In the last example, Y is not part of the combined memory state of the
4521 // outermost MergeMem. The system must, of course, prevent unschedulable
4522 // memory states from arising, so you can be sure that the state Y is somehow
4523 // a precursor to state Y'.
4524 //
4525 //
4526 // REPRESENTATION OF MEMORY MERGES: The indexes used to address the Node::in array
4527 // of each MergeMemNode array are exactly the numerical alias indexes, including
4528 // but not limited to AliasIdxTop, AliasIdxBot, and AliasIdxRaw. The functions
4529 // Compile::alias_type (and kin) produce and manage these indexes.
4530 //
4531 // By convention, the value of in(AliasIdxTop) (i.e., in(1)) is always the top node.
4532 // (Note that this provides quick access to the top node inside MergeMem methods,
4533 // without the need to reach out via TLS to Compile::current.)
4534 //
4535 // As a consequence of what was just described, a MergeMem that represents a full
4536 // memory state has an edge in(AliasIdxBot) which is a "wide" memory state,
4537 // containing all alias categories.
4538 //
4539 // MergeMem nodes never (?) have control inputs, so in(0) is NULL.
4540 //
4541 // All other edges in(N) (including in(AliasIdxRaw), which is in(3)) are either
4542 // a memory state for the alias type <N>, or else the top node, meaning that
4543 // there is no particular input for that alias type. Note that the length of
4544 // a MergeMem is variable, and may be extended at any time to accommodate new
4545 // memory states at larger alias indexes. When merges grow, they are of course
4546 // filled with "top" in the unused in() positions.
4547 //
4548 // This use of top is named "empty_memory()", or "empty_mem" (no-memory) as a variable.
4549 // (Top was chosen because it works smoothly with passes like GCM.)
4550 //
4551 // For convenience, we hardwire the alias index for TypeRawPtr::BOTTOM. (It is
4552 // the type of random VM bits like TLS references.) Since it is always the
4553 // first non-Bot memory slice, some low-level loops use it to initialize an
4554 // index variable: for (i = AliasIdxRaw; i < req(); i++).
4555 //
4556 //
4557 // ACCESSORS: There is a special accessor MergeMemNode::base_memory which returns
4558 // the distinguished "wide" state. The accessor MergeMemNode::memory_at(N) returns
4559 // the memory state for alias type <N>, or (if there is no particular slice at <N>,
4560 // it returns the base memory. To prevent bugs, memory_at does not accept <Top>
4561 // or <Bot> indexes. The iterator MergeMemStream provides robust iteration over
4562 // MergeMem nodes or pairs of such nodes, ensuring that the non-top edges are visited.
4563 //
4564 // %%%% We may get rid of base_memory as a separate accessor at some point; it isn't
4565 // really that different from the other memory inputs. An abbreviation called
4566 // "bot_memory()" for "memory_at(AliasIdxBot)" would keep code tidy.
4567 //
4568 //
4569 // PARTIAL MEMORY STATES: During optimization, MergeMem nodes may arise that represent
4570 // partial memory states. When a Phi splits through a MergeMem, the copy of the Phi
4571 // that "emerges though" the base memory will be marked as excluding the alias types
4572 // of the other (narrow-memory) copies which "emerged through" the narrow edges:
4573 //
4574 // Phi<Bot>(U, MergeMem(<Bot>: W, <8>: Y))
4575 // ==Ideal=> MergeMem(<Bot>: Phi<Bot-8>(U, W), Phi<8>(U, Y))
4576 //
4577 // This strange "subtraction" effect is necessary to ensure IGVN convergence.
4578 // (It is currently unimplemented.) As you can see, the resulting merge is
4579 // actually a disjoint union of memory states, rather than an overlay.
4580 //
4581
4582 //------------------------------MergeMemNode-----------------------------------
make_empty_memory()4583 Node* MergeMemNode::make_empty_memory() {
4584 Node* empty_memory = (Node*) Compile::current()->top();
4585 assert(empty_memory->is_top(), "correct sentinel identity");
4586 return empty_memory;
4587 }
4588
MergeMemNode(Node * new_base)4589 MergeMemNode::MergeMemNode(Node *new_base) : Node(1+Compile::AliasIdxRaw) {
4590 init_class_id(Class_MergeMem);
4591 // all inputs are nullified in Node::Node(int)
4592 // set_input(0, NULL); // no control input
4593
4594 // Initialize the edges uniformly to top, for starters.
4595 Node* empty_mem = make_empty_memory();
4596 for (uint i = Compile::AliasIdxTop; i < req(); i++) {
4597 init_req(i,empty_mem);
4598 }
4599 assert(empty_memory() == empty_mem, "");
4600
4601 if( new_base != NULL && new_base->is_MergeMem() ) {
4602 MergeMemNode* mdef = new_base->as_MergeMem();
4603 assert(mdef->empty_memory() == empty_mem, "consistent sentinels");
4604 for (MergeMemStream mms(this, mdef); mms.next_non_empty2(); ) {
4605 mms.set_memory(mms.memory2());
4606 }
4607 assert(base_memory() == mdef->base_memory(), "");
4608 } else {
4609 set_base_memory(new_base);
4610 }
4611 }
4612
4613 // Make a new, untransformed MergeMem with the same base as 'mem'.
4614 // If mem is itself a MergeMem, populate the result with the same edges.
make(Node * mem)4615 MergeMemNode* MergeMemNode::make(Node* mem) {
4616 return new MergeMemNode(mem);
4617 }
4618
4619 //------------------------------cmp--------------------------------------------
hash() const4620 uint MergeMemNode::hash() const { return NO_HASH; }
cmp(const Node & n) const4621 bool MergeMemNode::cmp( const Node &n ) const {
4622 return (&n == this); // Always fail except on self
4623 }
4624
4625 //------------------------------Identity---------------------------------------
Identity(PhaseGVN * phase)4626 Node* MergeMemNode::Identity(PhaseGVN* phase) {
4627 // Identity if this merge point does not record any interesting memory
4628 // disambiguations.
4629 Node* base_mem = base_memory();
4630 Node* empty_mem = empty_memory();
4631 if (base_mem != empty_mem) { // Memory path is not dead?
4632 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4633 Node* mem = in(i);
4634 if (mem != empty_mem && mem != base_mem) {
4635 return this; // Many memory splits; no change
4636 }
4637 }
4638 }
4639 return base_mem; // No memory splits; ID on the one true input
4640 }
4641
4642 //------------------------------Ideal------------------------------------------
4643 // This method is invoked recursively on chains of MergeMem nodes
Ideal(PhaseGVN * phase,bool can_reshape)4644 Node *MergeMemNode::Ideal(PhaseGVN *phase, bool can_reshape) {
4645 // Remove chain'd MergeMems
4646 //
4647 // This is delicate, because the each "in(i)" (i >= Raw) is interpreted
4648 // relative to the "in(Bot)". Since we are patching both at the same time,
4649 // we have to be careful to read each "in(i)" relative to the old "in(Bot)",
4650 // but rewrite each "in(i)" relative to the new "in(Bot)".
4651 Node *progress = NULL;
4652
4653
4654 Node* old_base = base_memory();
4655 Node* empty_mem = empty_memory();
4656 if (old_base == empty_mem)
4657 return NULL; // Dead memory path.
4658
4659 MergeMemNode* old_mbase;
4660 if (old_base != NULL && old_base->is_MergeMem())
4661 old_mbase = old_base->as_MergeMem();
4662 else
4663 old_mbase = NULL;
4664 Node* new_base = old_base;
4665
4666 // simplify stacked MergeMems in base memory
4667 if (old_mbase) new_base = old_mbase->base_memory();
4668
4669 // the base memory might contribute new slices beyond my req()
4670 if (old_mbase) grow_to_match(old_mbase);
4671
4672 // Look carefully at the base node if it is a phi.
4673 PhiNode* phi_base;
4674 if (new_base != NULL && new_base->is_Phi())
4675 phi_base = new_base->as_Phi();
4676 else
4677 phi_base = NULL;
4678
4679 Node* phi_reg = NULL;
4680 uint phi_len = (uint)-1;
4681 if (phi_base != NULL) {
4682 phi_reg = phi_base->region();
4683 phi_len = phi_base->req();
4684 // see if the phi is unfinished
4685 for (uint i = 1; i < phi_len; i++) {
4686 if (phi_base->in(i) == NULL) {
4687 // incomplete phi; do not look at it yet!
4688 phi_reg = NULL;
4689 phi_len = (uint)-1;
4690 break;
4691 }
4692 }
4693 }
4694
4695 // Note: We do not call verify_sparse on entry, because inputs
4696 // can normalize to the base_memory via subsume_node or similar
4697 // mechanisms. This method repairs that damage.
4698
4699 assert(!old_mbase || old_mbase->is_empty_memory(empty_mem), "consistent sentinels");
4700
4701 // Look at each slice.
4702 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4703 Node* old_in = in(i);
4704 // calculate the old memory value
4705 Node* old_mem = old_in;
4706 if (old_mem == empty_mem) old_mem = old_base;
4707 assert(old_mem == memory_at(i), "");
4708
4709 // maybe update (reslice) the old memory value
4710
4711 // simplify stacked MergeMems
4712 Node* new_mem = old_mem;
4713 MergeMemNode* old_mmem;
4714 if (old_mem != NULL && old_mem->is_MergeMem())
4715 old_mmem = old_mem->as_MergeMem();
4716 else
4717 old_mmem = NULL;
4718 if (old_mmem == this) {
4719 // This can happen if loops break up and safepoints disappear.
4720 // A merge of BotPtr (default) with a RawPtr memory derived from a
4721 // safepoint can be rewritten to a merge of the same BotPtr with
4722 // the BotPtr phi coming into the loop. If that phi disappears
4723 // also, we can end up with a self-loop of the mergemem.
4724 // In general, if loops degenerate and memory effects disappear,
4725 // a mergemem can be left looking at itself. This simply means
4726 // that the mergemem's default should be used, since there is
4727 // no longer any apparent effect on this slice.
4728 // Note: If a memory slice is a MergeMem cycle, it is unreachable
4729 // from start. Update the input to TOP.
4730 new_mem = (new_base == this || new_base == empty_mem)? empty_mem : new_base;
4731 }
4732 else if (old_mmem != NULL) {
4733 new_mem = old_mmem->memory_at(i);
4734 }
4735 // else preceding memory was not a MergeMem
4736
4737 // maybe store down a new value
4738 Node* new_in = new_mem;
4739 if (new_in == new_base) new_in = empty_mem;
4740
4741 if (new_in != old_in) {
4742 // Warning: Do not combine this "if" with the previous "if"
4743 // A memory slice might have be be rewritten even if it is semantically
4744 // unchanged, if the base_memory value has changed.
4745 set_req(i, new_in);
4746 progress = this; // Report progress
4747 }
4748 }
4749
4750 if (new_base != old_base) {
4751 set_req(Compile::AliasIdxBot, new_base);
4752 // Don't use set_base_memory(new_base), because we need to update du.
4753 assert(base_memory() == new_base, "");
4754 progress = this;
4755 }
4756
4757 if( base_memory() == this ) {
4758 // a self cycle indicates this memory path is dead
4759 set_req(Compile::AliasIdxBot, empty_mem);
4760 }
4761
4762 // Resolve external cycles by calling Ideal on a MergeMem base_memory
4763 // Recursion must occur after the self cycle check above
4764 if( base_memory()->is_MergeMem() ) {
4765 MergeMemNode *new_mbase = base_memory()->as_MergeMem();
4766 Node *m = phase->transform(new_mbase); // Rollup any cycles
4767 if( m != NULL &&
4768 (m->is_top() ||
4769 (m->is_MergeMem() && m->as_MergeMem()->base_memory() == empty_mem)) ) {
4770 // propagate rollup of dead cycle to self
4771 set_req(Compile::AliasIdxBot, empty_mem);
4772 }
4773 }
4774
4775 if( base_memory() == empty_mem ) {
4776 progress = this;
4777 // Cut inputs during Parse phase only.
4778 // During Optimize phase a dead MergeMem node will be subsumed by Top.
4779 if( !can_reshape ) {
4780 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4781 if( in(i) != empty_mem ) { set_req(i, empty_mem); }
4782 }
4783 }
4784 }
4785
4786 if( !progress && base_memory()->is_Phi() && can_reshape ) {
4787 // Check if PhiNode::Ideal's "Split phis through memory merges"
4788 // transform should be attempted. Look for this->phi->this cycle.
4789 uint merge_width = req();
4790 if (merge_width > Compile::AliasIdxRaw) {
4791 PhiNode* phi = base_memory()->as_Phi();
4792 for( uint i = 1; i < phi->req(); ++i ) {// For all paths in
4793 if (phi->in(i) == this) {
4794 phase->is_IterGVN()->_worklist.push(phi);
4795 break;
4796 }
4797 }
4798 }
4799 }
4800
4801 assert(progress || verify_sparse(), "please, no dups of base");
4802 return progress;
4803 }
4804
4805 //-------------------------set_base_memory-------------------------------------
set_base_memory(Node * new_base)4806 void MergeMemNode::set_base_memory(Node *new_base) {
4807 Node* empty_mem = empty_memory();
4808 set_req(Compile::AliasIdxBot, new_base);
4809 assert(memory_at(req()) == new_base, "must set default memory");
4810 // Clear out other occurrences of new_base:
4811 if (new_base != empty_mem) {
4812 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4813 if (in(i) == new_base) set_req(i, empty_mem);
4814 }
4815 }
4816 }
4817
4818 //------------------------------out_RegMask------------------------------------
out_RegMask() const4819 const RegMask &MergeMemNode::out_RegMask() const {
4820 return RegMask::Empty;
4821 }
4822
4823 //------------------------------dump_spec--------------------------------------
4824 #ifndef PRODUCT
dump_spec(outputStream * st) const4825 void MergeMemNode::dump_spec(outputStream *st) const {
4826 st->print(" {");
4827 Node* base_mem = base_memory();
4828 for( uint i = Compile::AliasIdxRaw; i < req(); i++ ) {
4829 Node* mem = (in(i) != NULL) ? memory_at(i) : base_mem;
4830 if (mem == base_mem) { st->print(" -"); continue; }
4831 st->print( " N%d:", mem->_idx );
4832 Compile::current()->get_adr_type(i)->dump_on(st);
4833 }
4834 st->print(" }");
4835 }
4836 #endif // !PRODUCT
4837
4838
4839 #ifdef ASSERT
might_be_same(Node * a,Node * b)4840 static bool might_be_same(Node* a, Node* b) {
4841 if (a == b) return true;
4842 if (!(a->is_Phi() || b->is_Phi())) return false;
4843 // phis shift around during optimization
4844 return true; // pretty stupid...
4845 }
4846
4847 // verify a narrow slice (either incoming or outgoing)
verify_memory_slice(const MergeMemNode * m,int alias_idx,Node * n)4848 static void verify_memory_slice(const MergeMemNode* m, int alias_idx, Node* n) {
4849 if (!VerifyAliases) return; // don't bother to verify unless requested
4850 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
4851 if (Node::in_dump()) return; // muzzle asserts when printing
4852 assert(alias_idx >= Compile::AliasIdxRaw, "must not disturb base_memory or sentinel");
4853 assert(n != NULL, "");
4854 // Elide intervening MergeMem's
4855 while (n->is_MergeMem()) {
4856 n = n->as_MergeMem()->memory_at(alias_idx);
4857 }
4858 Compile* C = Compile::current();
4859 const TypePtr* n_adr_type = n->adr_type();
4860 if (n == m->empty_memory()) {
4861 // Implicit copy of base_memory()
4862 } else if (n_adr_type != TypePtr::BOTTOM) {
4863 assert(n_adr_type != NULL, "new memory must have a well-defined adr_type");
4864 assert(C->must_alias(n_adr_type, alias_idx), "new memory must match selected slice");
4865 } else {
4866 // A few places like make_runtime_call "know" that VM calls are narrow,
4867 // and can be used to update only the VM bits stored as TypeRawPtr::BOTTOM.
4868 bool expected_wide_mem = false;
4869 if (n == m->base_memory()) {
4870 expected_wide_mem = true;
4871 } else if (alias_idx == Compile::AliasIdxRaw ||
4872 n == m->memory_at(Compile::AliasIdxRaw)) {
4873 expected_wide_mem = true;
4874 } else if (!C->alias_type(alias_idx)->is_rewritable()) {
4875 // memory can "leak through" calls on channels that
4876 // are write-once. Allow this also.
4877 expected_wide_mem = true;
4878 }
4879 assert(expected_wide_mem, "expected narrow slice replacement");
4880 }
4881 }
4882 #else // !ASSERT
4883 #define verify_memory_slice(m,i,n) (void)(0) // PRODUCT version is no-op
4884 #endif
4885
4886
4887 //-----------------------------memory_at---------------------------------------
memory_at(uint alias_idx) const4888 Node* MergeMemNode::memory_at(uint alias_idx) const {
4889 assert(alias_idx >= Compile::AliasIdxRaw ||
4890 alias_idx == Compile::AliasIdxBot && Compile::current()->AliasLevel() == 0,
4891 "must avoid base_memory and AliasIdxTop");
4892
4893 // Otherwise, it is a narrow slice.
4894 Node* n = alias_idx < req() ? in(alias_idx) : empty_memory();
4895 Compile *C = Compile::current();
4896 if (is_empty_memory(n)) {
4897 // the array is sparse; empty slots are the "top" node
4898 n = base_memory();
4899 assert(Node::in_dump()
4900 || n == NULL || n->bottom_type() == Type::TOP
4901 || n->adr_type() == NULL // address is TOP
4902 || n->adr_type() == TypePtr::BOTTOM
4903 || n->adr_type() == TypeRawPtr::BOTTOM
4904 || Compile::current()->AliasLevel() == 0,
4905 "must be a wide memory");
4906 // AliasLevel == 0 if we are organizing the memory states manually.
4907 // See verify_memory_slice for comments on TypeRawPtr::BOTTOM.
4908 } else {
4909 // make sure the stored slice is sane
4910 #ifdef ASSERT
4911 if (VMError::is_error_reported() || Node::in_dump()) {
4912 } else if (might_be_same(n, base_memory())) {
4913 // Give it a pass: It is a mostly harmless repetition of the base.
4914 // This can arise normally from node subsumption during optimization.
4915 } else {
4916 verify_memory_slice(this, alias_idx, n);
4917 }
4918 #endif
4919 }
4920 return n;
4921 }
4922
4923 //---------------------------set_memory_at-------------------------------------
set_memory_at(uint alias_idx,Node * n)4924 void MergeMemNode::set_memory_at(uint alias_idx, Node *n) {
4925 verify_memory_slice(this, alias_idx, n);
4926 Node* empty_mem = empty_memory();
4927 if (n == base_memory()) n = empty_mem; // collapse default
4928 uint need_req = alias_idx+1;
4929 if (req() < need_req) {
4930 if (n == empty_mem) return; // already the default, so do not grow me
4931 // grow the sparse array
4932 do {
4933 add_req(empty_mem);
4934 } while (req() < need_req);
4935 }
4936 set_req( alias_idx, n );
4937 }
4938
4939
4940
4941 //--------------------------iteration_setup------------------------------------
iteration_setup(const MergeMemNode * other)4942 void MergeMemNode::iteration_setup(const MergeMemNode* other) {
4943 if (other != NULL) {
4944 grow_to_match(other);
4945 // invariant: the finite support of mm2 is within mm->req()
4946 #ifdef ASSERT
4947 for (uint i = req(); i < other->req(); i++) {
4948 assert(other->is_empty_memory(other->in(i)), "slice left uncovered");
4949 }
4950 #endif
4951 }
4952 // Replace spurious copies of base_memory by top.
4953 Node* base_mem = base_memory();
4954 if (base_mem != NULL && !base_mem->is_top()) {
4955 for (uint i = Compile::AliasIdxBot+1, imax = req(); i < imax; i++) {
4956 if (in(i) == base_mem)
4957 set_req(i, empty_memory());
4958 }
4959 }
4960 }
4961
4962 //---------------------------grow_to_match-------------------------------------
grow_to_match(const MergeMemNode * other)4963 void MergeMemNode::grow_to_match(const MergeMemNode* other) {
4964 Node* empty_mem = empty_memory();
4965 assert(other->is_empty_memory(empty_mem), "consistent sentinels");
4966 // look for the finite support of the other memory
4967 for (uint i = other->req(); --i >= req(); ) {
4968 if (other->in(i) != empty_mem) {
4969 uint new_len = i+1;
4970 while (req() < new_len) add_req(empty_mem);
4971 break;
4972 }
4973 }
4974 }
4975
4976 //---------------------------verify_sparse-------------------------------------
4977 #ifndef PRODUCT
verify_sparse() const4978 bool MergeMemNode::verify_sparse() const {
4979 assert(is_empty_memory(make_empty_memory()), "sane sentinel");
4980 Node* base_mem = base_memory();
4981 // The following can happen in degenerate cases, since empty==top.
4982 if (is_empty_memory(base_mem)) return true;
4983 for (uint i = Compile::AliasIdxRaw; i < req(); i++) {
4984 assert(in(i) != NULL, "sane slice");
4985 if (in(i) == base_mem) return false; // should have been the sentinel value!
4986 }
4987 return true;
4988 }
4989
match_memory(Node * mem,const MergeMemNode * mm,int idx)4990 bool MergeMemStream::match_memory(Node* mem, const MergeMemNode* mm, int idx) {
4991 Node* n;
4992 n = mm->in(idx);
4993 if (mem == n) return true; // might be empty_memory()
4994 n = (idx == Compile::AliasIdxBot)? mm->base_memory(): mm->memory_at(idx);
4995 if (mem == n) return true;
4996 return false;
4997 }
4998 #endif // !PRODUCT
4999