1 /*
2 * Copyright (c) 2015, 2018, Red Hat, Inc. All rights reserved.
3 *
4 * This code is free software; you can redistribute it and/or modify it
5 * under the terms of the GNU General Public License version 2 only, as
6 * published by the Free Software Foundation.
7 *
8 * This code is distributed in the hope that it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
11 * version 2 for more details (a copy is included in the LICENSE file that
12 * accompanied this code).
13 *
14 * You should have received a copy of the GNU General Public License version
15 * 2 along with this work; if not, write to the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
17 *
18 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
19 * or visit www.oracle.com if you need additional information or have any
20 * questions.
21 *
22 */
23
24 #include "precompiled.hpp"
25
26 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
27 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
28 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
29 #include "gc/shenandoah/shenandoahBrooksPointer.hpp"
30 #include "gc/shenandoah/shenandoahHeap.hpp"
31 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
32 #include "gc/shenandoah/shenandoahRuntime.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/block.hpp"
35 #include "opto/callnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/movenode.hpp"
38 #include "opto/phaseX.hpp"
39 #include "opto/rootnode.hpp"
40 #include "opto/runtime.hpp"
41 #include "opto/subnode.hpp"
42
43 Node* ShenandoahBarrierNode::skip_through_barrier(Node* n) {
44 if (n == NULL) {
45 return NULL;
46 }
47 if (n->Opcode() == Op_ShenandoahEnqueueBarrier) {
48 n = n->in(1);
49 }
50
51 if (n->is_ShenandoahBarrier()) {
52 return n->in(ValueIn);
53 } else if (n->is_Phi() &&
54 n->req() == 3 &&
55 n->in(1) != NULL &&
56 n->in(1)->is_ShenandoahBarrier() &&
57 n->in(2) != NULL &&
58 n->in(2)->bottom_type() == TypePtr::NULL_PTR &&
59 n->in(0) != NULL &&
60 n->in(0)->in(1) != NULL &&
61 n->in(0)->in(1)->is_IfProj() &&
62 n->in(0)->in(2) != NULL &&
63 n->in(0)->in(2)->is_IfProj() &&
64 n->in(0)->in(1)->in(0) != NULL &&
65 n->in(0)->in(1)->in(0) == n->in(0)->in(2)->in(0) &&
66 n->in(1)->in(ValueIn)->Opcode() == Op_CastPP) {
67 Node* iff = n->in(0)->in(1)->in(0);
68 Node* res = n->in(1)->in(ValueIn)->in(1);
69 if (iff->is_If() &&
70 iff->in(1) != NULL &&
71 iff->in(1)->is_Bool() &&
72 iff->in(1)->as_Bool()->_test._test == BoolTest::ne &&
73 iff->in(1)->in(1) != NULL &&
74 iff->in(1)->in(1)->Opcode() == Op_CmpP &&
75 iff->in(1)->in(1)->in(1) != NULL &&
76 iff->in(1)->in(1)->in(1) == res &&
77 iff->in(1)->in(1)->in(2) != NULL &&
78 iff->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
79 return res;
80 }
81 }
82 return n;
83 }
84
85 bool ShenandoahBarrierNode::needs_barrier(PhaseGVN* phase, ShenandoahBarrierNode* orig, Node* n, Node* rb_mem, bool allow_fromspace) {
86 Unique_Node_List visited;
87 return needs_barrier_impl(phase, orig, n, rb_mem, allow_fromspace, visited);
88 }
89
90 bool ShenandoahBarrierNode::needs_barrier_impl(PhaseGVN* phase, ShenandoahBarrierNode* orig, Node* n, Node* rb_mem, bool allow_fromspace, Unique_Node_List &visited) {
91 if (visited.member(n)) {
92 return false; // Been there.
93 }
94 visited.push(n);
95
96 if (n->is_Allocate()) {
97 return false;
98 }
99
100 if (n->is_CallJava() || n->Opcode() == Op_CallLeafNoFP) {
101 return true;
102 }
103
104 const Type* type = phase->type(n);
105 if (type == Type::TOP) {
106 return false;
107 }
108 if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) {
109 return false;
110 }
111 if (type->make_oopptr() && type->make_oopptr()->const_oop() != NULL) {
112 return false;
113 }
114
115 if (ShenandoahOptimizeStableFinals) {
116 const TypeAryPtr* ary = type->isa_aryptr();
117 if (ary && ary->is_stable() && allow_fromspace) {
118 return false;
119 }
120 }
121
122 if (n->is_CheckCastPP() || n->is_ConstraintCast() || n->Opcode() == Op_ShenandoahEnqueueBarrier) {
123 return needs_barrier_impl(phase, orig, n->in(1), rb_mem, allow_fromspace, visited);
124 }
125 if (n->is_Parm()) {
126 return true;
127 }
128 if (n->is_Proj()) {
129 return needs_barrier_impl(phase, orig, n->in(0), rb_mem, allow_fromspace, visited);
130 }
131
132 if (n->Opcode() == Op_ShenandoahWBMemProj) {
133 return needs_barrier_impl(phase, orig, n->in(ShenandoahWBMemProjNode::WriteBarrier), rb_mem, allow_fromspace, visited);
134 }
135 if (n->is_Phi()) {
136 bool need_barrier = false;
137 for (uint i = 1; i < n->req() && ! need_barrier; i++) {
138 Node* input = n->in(i);
139 if (input == NULL) {
140 need_barrier = true; // Phi not complete yet?
141 } else if (needs_barrier_impl(phase, orig, input, rb_mem, allow_fromspace, visited)) {
142 need_barrier = true;
143 }
144 }
145 return need_barrier;
146 }
147 if (n->is_CMove()) {
148 return needs_barrier_impl(phase, orig, n->in(CMoveNode::IfFalse), rb_mem, allow_fromspace, visited) ||
149 needs_barrier_impl(phase, orig, n->in(CMoveNode::IfTrue ), rb_mem, allow_fromspace, visited);
150 }
151 if (n->Opcode() == Op_CreateEx) {
152 return true;
153 }
154 if (n->Opcode() == Op_ShenandoahWriteBarrier) {
155 return false;
156 }
157 if (n->Opcode() == Op_ShenandoahReadBarrier) {
158 if (rb_mem == n->in(Memory)) {
159 return false;
160 } else {
161 return true;
162 }
163 }
164
165 if (n->Opcode() == Op_LoadP ||
166 n->Opcode() == Op_LoadN ||
167 n->Opcode() == Op_GetAndSetP ||
168 n->Opcode() == Op_CompareAndExchangeP ||
169 n->Opcode() == Op_ShenandoahCompareAndExchangeP ||
170 n->Opcode() == Op_GetAndSetN ||
171 n->Opcode() == Op_CompareAndExchangeN ||
172 n->Opcode() == Op_ShenandoahCompareAndExchangeN) {
173 return true;
174 }
175 if (n->Opcode() == Op_DecodeN ||
176 n->Opcode() == Op_EncodeP) {
177 return needs_barrier_impl(phase, orig, n->in(1), rb_mem, allow_fromspace, visited);
178 }
179
180 #ifdef ASSERT
181 tty->print("need barrier on?: "); n->dump();
182 ShouldNotReachHere();
183 #endif
184 return true;
185 }
186
187 bool ShenandoahReadBarrierNode::dominates_memory_rb_impl(PhaseGVN* phase,
188 Node* b1,
189 Node* b2,
190 Node* current,
191 bool linear) {
192 ResourceMark rm;
193 VectorSet visited(Thread::current()->resource_area());
194 Node_Stack phis(0);
195
196 for(int i = 0; i < 10; i++) {
197 if (current == NULL) {
198 return false;
199 } else if (visited.test_set(current->_idx) || current->is_top() || current == b1) {
200 current = NULL;
201 while (phis.is_nonempty() && current == NULL) {
202 uint idx = phis.index();
203 Node* phi = phis.node();
204 if (idx >= phi->req()) {
205 phis.pop();
206 } else {
207 current = phi->in(idx);
208 phis.set_index(idx+1);
209 }
210 }
211 if (current == NULL) {
212 return true;
213 }
214 } else if (current == phase->C->immutable_memory()) {
215 return false;
216 } else if (current->isa_Phi()) {
217 if (!linear) {
218 return false;
219 }
220 phis.push(current, 2);
221 current = current->in(1);
222 } else if (current->Opcode() == Op_ShenandoahWriteBarrier) {
223 const Type* in_type = current->bottom_type();
224 const Type* this_type = b2->bottom_type();
225 if (is_independent(in_type, this_type)) {
226 current = current->in(Memory);
227 } else {
228 return false;
229 }
230 } else if (current->Opcode() == Op_ShenandoahWBMemProj) {
231 current = current->in(ShenandoahWBMemProjNode::WriteBarrier);
232 } else if (current->is_Proj()) {
233 current = current->in(0);
234 } else if (current->is_Call()) {
235 return false; // TODO: Maybe improve by looking at the call's memory effects?
236 } else if (current->is_MemBar()) {
237 return false; // TODO: Do we need to stop at *any* membar?
238 } else if (current->is_MergeMem()) {
239 const TypePtr* adr_type = brooks_pointer_type(phase->type(b2));
240 uint alias_idx = phase->C->get_alias_index(adr_type);
241 current = current->as_MergeMem()->memory_at(alias_idx);
242 } else {
243 #ifdef ASSERT
244 current->dump();
245 #endif
246 ShouldNotReachHere();
247 return false;
248 }
249 }
250 return false;
251 }
252
253 bool ShenandoahReadBarrierNode::is_independent(Node* mem) {
254 if (mem->is_Phi() || mem->is_Proj() || mem->is_MergeMem()) {
255 return true;
256 } else if (mem->Opcode() == Op_ShenandoahWBMemProj) {
257 return true;
258 } else if (mem->Opcode() == Op_ShenandoahWriteBarrier) {
259 const Type* mem_type = mem->bottom_type();
260 const Type* this_type = bottom_type();
261 if (is_independent(mem_type, this_type)) {
262 return true;
263 } else {
264 return false;
265 }
266 } else if (mem->is_Call() || mem->is_MemBar()) {
267 return false;
268 }
269 #ifdef ASSERT
270 mem->dump();
271 #endif
272 ShouldNotReachHere();
273 return true;
274 }
275
276 bool ShenandoahReadBarrierNode::dominates_memory_rb(PhaseGVN* phase, Node* b1, Node* b2, bool linear) {
277 return dominates_memory_rb_impl(phase, b1->in(Memory), b2, b2->in(Memory), linear);
278 }
279
280 bool ShenandoahReadBarrierNode::is_independent(const Type* in_type, const Type* this_type) {
281 assert(in_type->isa_oopptr(), "expect oop ptr");
282 assert(this_type->isa_oopptr(), "expect oop ptr");
283
284 ciKlass* in_kls = in_type->is_oopptr()->klass();
285 ciKlass* this_kls = this_type->is_oopptr()->klass();
286 if (in_kls != NULL && this_kls != NULL &&
287 in_kls->is_loaded() && this_kls->is_loaded() &&
288 (!in_kls->is_subclass_of(this_kls)) &&
289 (!this_kls->is_subclass_of(in_kls))) {
290 return true;
291 }
292 return false;
293 }
294
295 Node* ShenandoahReadBarrierNode::Ideal(PhaseGVN *phase, bool can_reshape) {
296 if (! can_reshape) {
297 return NULL;
298 }
299
300 if (in(Memory) == phase->C->immutable_memory()) return NULL;
301
302 // If memory input is a MergeMem, take the appropriate slice out of it.
303 Node* mem_in = in(Memory);
304 if (mem_in->isa_MergeMem()) {
305 const TypePtr* adr_type = brooks_pointer_type(bottom_type());
306 uint alias_idx = phase->C->get_alias_index(adr_type);
307 mem_in = mem_in->as_MergeMem()->memory_at(alias_idx);
308 set_req(Memory, mem_in);
309 return this;
310 }
311
312 Node* input = in(Memory);
313 if (input->Opcode() == Op_ShenandoahWBMemProj) {
314 ResourceMark rm;
315 VectorSet seen(Thread::current()->resource_area());
316 Node* n = in(Memory);
317 while (n->Opcode() == Op_ShenandoahWBMemProj &&
318 n->in(ShenandoahWBMemProjNode::WriteBarrier) != NULL &&
319 n->in(ShenandoahWBMemProjNode::WriteBarrier)->Opcode() == Op_ShenandoahWriteBarrier &&
320 n->in(ShenandoahWBMemProjNode::WriteBarrier)->in(Memory) != NULL) {
321 if (seen.test_set(n->_idx)) {
322 return NULL; // loop
323 }
324 n = n->in(ShenandoahWBMemProjNode::WriteBarrier)->in(Memory);
325 }
326
327 Node* wb = input->in(ShenandoahWBMemProjNode::WriteBarrier);
328 const Type* in_type = phase->type(wb);
329 // is_top() test not sufficient here: we can come here after CCP
330 // in a dead branch of the graph that has not yet been removed.
331 if (in_type == Type::TOP) return NULL; // Dead path.
332 assert(wb->Opcode() == Op_ShenandoahWriteBarrier, "expect write barrier");
333 if (is_independent(in_type, _type)) {
334 phase->igvn_rehash_node_delayed(wb);
335 set_req(Memory, wb->in(Memory));
336 if (can_reshape && input->outcnt() == 0) {
337 phase->is_IterGVN()->_worklist.push(input);
338 }
339 return this;
340 }
341 }
342 return NULL;
343 }
344
345 ShenandoahWriteBarrierNode::ShenandoahWriteBarrierNode(Compile* C, Node* ctrl, Node* mem, Node* obj)
346 : ShenandoahBarrierNode(ctrl, mem, obj, false) {
347 assert(UseShenandoahGC && ShenandoahWriteBarrier, "should be enabled");
348 ShenandoahBarrierSetC2::bsc2()->state()->add_shenandoah_barrier(this);
349 }
350
351 Node* ShenandoahWriteBarrierNode::Identity(PhaseGVN* phase) {
352 assert(in(0) != NULL, "should have control");
353 PhaseIterGVN* igvn = phase->is_IterGVN();
354 Node* mem_in = in(Memory);
355 Node* mem_proj = NULL;
356
357 if (igvn != NULL) {
358 mem_proj = find_out_with(Op_ShenandoahWBMemProj);
359 if (mem_in == mem_proj) {
360 return this;
361 }
362 }
363
364 Node* replacement = Identity_impl(phase);
365 if (igvn != NULL) {
366 if (replacement != NULL && replacement != this && mem_proj != NULL) {
367 igvn->replace_node(mem_proj, mem_in);
368 }
369 }
370 return replacement;
371 }
372
373 Node* ShenandoahWriteBarrierNode::Ideal(PhaseGVN *phase, bool can_reshape) {
374 assert(in(0) != NULL, "should have control");
375 if (!can_reshape) {
376 return NULL;
377 }
378
379 Node* mem_in = in(Memory);
380
381 if (mem_in->isa_MergeMem()) {
382 const TypePtr* adr_type = brooks_pointer_type(bottom_type());
383 uint alias_idx = phase->C->get_alias_index(adr_type);
384 mem_in = mem_in->as_MergeMem()->memory_at(alias_idx);
385 set_req(Memory, mem_in);
386 return this;
387 }
388
389 Node* val = in(ValueIn);
390 if (val->is_ShenandoahBarrier()) {
391 set_req(ValueIn, val->in(ValueIn));
392 return this;
393 }
394
395 return NULL;
396 }
397
398 bool ShenandoahWriteBarrierNode::expand(Compile* C, PhaseIterGVN& igvn, int& loop_opts_cnt) {
399 if (UseShenandoahGC) {
400 if (ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count() > 0 || (!ShenandoahWriteBarrier && ShenandoahStoreValEnqueueBarrier)) {
401 bool attempt_more_loopopts = ShenandoahLoopOptsAfterExpansion;
402 C->clear_major_progress();
403 PhaseIdealLoop ideal_loop(igvn, LoopOptsShenandoahExpand);
404 if (C->failing()) return false;
405 PhaseIdealLoop::verify(igvn);
406 DEBUG_ONLY(ShenandoahBarrierNode::verify_raw_mem(C->root());)
407 if (attempt_more_loopopts) {
408 C->set_major_progress();
409 if (!C->optimize_loops(loop_opts_cnt, igvn, LoopOptsShenandoahPostExpand)) {
410 return false;
411 }
412 C->clear_major_progress();
413 }
414 }
415 }
416 return true;
417 }
418
419 bool ShenandoahWriteBarrierNode::is_heap_state_test(Node* iff, int mask) {
420 if (!UseShenandoahGC) {
421 return false;
422 }
423 assert(iff->is_If(), "bad input");
424 if (iff->Opcode() != Op_If) {
425 return false;
426 }
427 Node* bol = iff->in(1);
428 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
429 return false;
430 }
431 Node* cmp = bol->in(1);
432 if (cmp->Opcode() != Op_CmpI) {
433 return false;
434 }
435 Node* in1 = cmp->in(1);
436 Node* in2 = cmp->in(2);
437 if (in2->find_int_con(-1) != 0) {
438 return false;
439 }
440 if (in1->Opcode() != Op_AndI) {
441 return false;
442 }
443 in2 = in1->in(2);
444 if (in2->find_int_con(-1) != mask) {
445 return false;
446 }
447 in1 = in1->in(1);
448
449 return is_gc_state_load(in1);
450 }
451
452 bool ShenandoahWriteBarrierNode::is_heap_stable_test(Node* iff) {
453 return is_heap_state_test(iff, ShenandoahHeap::HAS_FORWARDED);
454 }
455
456 bool ShenandoahWriteBarrierNode::is_gc_state_load(Node *n) {
457 if (!UseShenandoahGC) {
458 return false;
459 }
460 if (n->Opcode() != Op_LoadB && n->Opcode() != Op_LoadUB) {
461 return false;
462 }
463 Node* addp = n->in(MemNode::Address);
464 if (!addp->is_AddP()) {
465 return false;
466 }
467 Node* base = addp->in(AddPNode::Address);
468 Node* off = addp->in(AddPNode::Offset);
469 if (base->Opcode() != Op_ThreadLocal) {
470 return false;
471 }
472 if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) {
473 return false;
474 }
475 return true;
476 }
477
478 bool ShenandoahWriteBarrierNode::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) {
479 assert(phase->is_dominator(stop, start), "bad inputs");
480 ResourceMark rm;
481 Unique_Node_List wq;
482 wq.push(start);
483 for (uint next = 0; next < wq.size(); next++) {
484 Node *m = wq.at(next);
485 if (m == stop) {
486 continue;
487 }
488 if (m->is_SafePoint() && !m->is_CallLeaf()) {
489 return true;
490 }
491 if (m->is_Region()) {
492 for (uint i = 1; i < m->req(); i++) {
493 wq.push(m->in(i));
494 }
495 } else {
496 wq.push(m->in(0));
497 }
498 }
499 return false;
500 }
501
502 bool ShenandoahWriteBarrierNode::try_common_gc_state_load(Node *n, PhaseIdealLoop *phase) {
503 assert(is_gc_state_load(n), "inconsistent");
504 Node* addp = n->in(MemNode::Address);
505 Node* dominator = NULL;
506 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) {
507 Node* u = addp->fast_out(i);
508 assert(is_gc_state_load(u), "inconsistent");
509 if (u != n && phase->is_dominator(u->in(0), n->in(0))) {
510 if (dominator == NULL) {
511 dominator = u;
512 } else {
513 if (phase->dom_depth(u->in(0)) < phase->dom_depth(dominator->in(0))) {
514 dominator = u;
515 }
516 }
517 }
518 }
519 if (dominator == NULL || has_safepoint_between(n->in(0), dominator->in(0), phase)) {
520 return false;
521 }
522 phase->igvn().replace_node(n, dominator);
523
524 return true;
525 }
526
527 bool ShenandoahBarrierNode::dominates_memory_impl(PhaseGVN* phase,
528 Node* b1,
529 Node* b2,
530 Node* current,
531 bool linear) {
532 ResourceMark rm;
533 VectorSet visited(Thread::current()->resource_area());
534 Node_Stack phis(0);
535
536 for(int i = 0; i < 10; i++) {
537 if (current == NULL) {
538 return false;
539 } else if (visited.test_set(current->_idx) || current->is_top() || current == b1) {
540 current = NULL;
541 while (phis.is_nonempty() && current == NULL) {
542 uint idx = phis.index();
543 Node* phi = phis.node();
544 if (idx >= phi->req()) {
545 phis.pop();
546 } else {
547 current = phi->in(idx);
548 phis.set_index(idx+1);
549 }
550 }
551 if (current == NULL) {
552 return true;
553 }
554 } else if (current == b2) {
555 return false;
556 } else if (current == phase->C->immutable_memory()) {
557 return false;
558 } else if (current->isa_Phi()) {
559 if (!linear) {
560 return false;
561 }
562 phis.push(current, 2);
563 current = current->in(1);
564 } else if (current->Opcode() == Op_ShenandoahWriteBarrier) {
565 current = current->in(Memory);
566 } else if (current->Opcode() == Op_ShenandoahWBMemProj) {
567 current = current->in(ShenandoahWBMemProjNode::WriteBarrier);
568 } else if (current->is_Proj()) {
569 current = current->in(0);
570 } else if (current->is_Call()) {
571 current = current->in(TypeFunc::Memory);
572 } else if (current->is_MemBar()) {
573 current = current->in(TypeFunc::Memory);
574 } else if (current->is_MergeMem()) {
575 const TypePtr* adr_type = brooks_pointer_type(phase->type(b2));
576 uint alias_idx = phase->C->get_alias_index(adr_type);
577 current = current->as_MergeMem()->memory_at(alias_idx);
578 } else {
579 #ifdef ASSERT
580 current->dump();
581 #endif
582 ShouldNotReachHere();
583 return false;
584 }
585 }
586 return false;
587 }
588
589 /**
590 * Determines if b1 dominates b2 through memory inputs. It returns true if:
591 * - b1 can be reached by following each branch in b2's memory input (through phis, etc)
592 * - or we get back to b2 (i.e. through a loop) without seeing b1
593 * In all other cases, (in particular, if we reach immutable_memory without having seen b1)
594 * we return false.
595 */
596 bool ShenandoahBarrierNode::dominates_memory(PhaseGVN* phase, Node* b1, Node* b2, bool linear) {
597 return dominates_memory_impl(phase, b1, b2, b2->in(Memory), linear);
598 }
599
600 Node* ShenandoahBarrierNode::Identity_impl(PhaseGVN* phase) {
601 Node* n = in(ValueIn);
602
603 Node* rb_mem = Opcode() == Op_ShenandoahReadBarrier ? in(Memory) : NULL;
604 if (! needs_barrier(phase, this, n, rb_mem, _allow_fromspace)) {
605 return n;
606 }
607
608 // Try to find a write barrier sibling with identical inputs that we can fold into.
609 for (DUIterator i = n->outs(); n->has_out(i); i++) {
610 Node* sibling = n->out(i);
611 if (sibling == this) {
612 continue;
613 }
614 if (sibling->Opcode() != Op_ShenandoahWriteBarrier) {
615 continue;
616 }
617
618 assert(sibling->in(ValueIn) == in(ValueIn), "sanity");
619 assert(sibling->Opcode() == Op_ShenandoahWriteBarrier, "sanity");
620
621 if (dominates_memory(phase, sibling, this, phase->is_IterGVN() == NULL)) {
622 return sibling;
623 }
624 }
625 return this;
626 }
627
628 #ifndef PRODUCT
629 void ShenandoahBarrierNode::dump_spec(outputStream *st) const {
630 const TypePtr* adr = adr_type();
631 if (adr == NULL) {
632 return;
633 }
634 st->print(" @");
635 adr->dump_on(st);
636 st->print(" (");
637 Compile::current()->alias_type(adr)->adr_type()->dump_on(st);
638 st->print(") ");
639 }
640 #endif
641
642 Node* ShenandoahReadBarrierNode::Identity(PhaseGVN* phase) {
643 Node* id = Identity_impl(phase);
644
645 if (id == this && phase->is_IterGVN()) {
646 Node* n = in(ValueIn);
647 // No success in super call. Try to combine identical read barriers.
648 for (DUIterator i = n->outs(); n->has_out(i); i++) {
649 Node* sibling = n->out(i);
650 if (sibling == this || sibling->Opcode() != Op_ShenandoahReadBarrier) {
651 continue;
652 }
653 assert(sibling->in(ValueIn) == in(ValueIn), "sanity");
654 if (phase->is_IterGVN()->hash_find(sibling) &&
655 sibling->bottom_type() == bottom_type() &&
656 sibling->in(Control) == in(Control) &&
657 dominates_memory_rb(phase, sibling, this, phase->is_IterGVN() == NULL)) {
658 return sibling;
659 }
660 }
661 }
662 return id;
663 }
664
665 const Type* ShenandoahBarrierNode::Value(PhaseGVN* phase) const {
666 // Either input is TOP ==> the result is TOP
667 const Type *t1 = phase->type(in(Memory));
668 if (t1 == Type::TOP) return Type::TOP;
669 const Type *t2 = phase->type(in(ValueIn));
670 if( t2 == Type::TOP ) return Type::TOP;
671
672 if (t2 == TypePtr::NULL_PTR) {
673 return _type;
674 }
675
676 const Type* type = t2->is_oopptr()->cast_to_nonconst();
677 return type;
678 }
679
680 uint ShenandoahBarrierNode::hash() const {
681 return TypeNode::hash() + _allow_fromspace;
682 }
683
684 uint ShenandoahBarrierNode::cmp(const Node& n) const {
685 return _allow_fromspace == ((ShenandoahBarrierNode&) n)._allow_fromspace
686 && TypeNode::cmp(n);
687 }
688
689 uint ShenandoahBarrierNode::size_of() const {
690 return sizeof(*this);
691 }
692
693 Node* ShenandoahWBMemProjNode::Identity(PhaseGVN* phase) {
694 Node* wb = in(WriteBarrier);
695 if (wb->is_top()) return phase->C->top(); // Dead path.
696
697 assert(wb->Opcode() == Op_ShenandoahWriteBarrier, "expect write barrier");
698 PhaseIterGVN* igvn = phase->is_IterGVN();
699 // We can't do the below unless the graph is fully constructed.
700 if (igvn == NULL) {
701 return this;
702 }
703
704 // If the mem projection has no barrier users, it's not needed anymore.
705 if (wb->outcnt() == 1) {
706 return wb->in(ShenandoahBarrierNode::Memory);
707 }
708
709 return this;
710 }
711
712 #ifdef ASSERT
713 bool ShenandoahBarrierNode::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) {
714 assert(phis.size() == 0, "");
715
716 while (true) {
717 if (in->bottom_type() == TypePtr::NULL_PTR) {
718 if (trace) {tty->print_cr("NULL");}
719 } else if (!in->bottom_type()->make_ptr()->make_oopptr()) {
720 if (trace) {tty->print_cr("Non oop");}
721 } else if (t == ShenandoahLoad && ShenandoahOptimizeStableFinals &&
722 in->bottom_type()->make_ptr()->isa_aryptr() &&
723 in->bottom_type()->make_ptr()->is_aryptr()->is_stable()) {
724 if (trace) {tty->print_cr("Stable array load");}
725 } else {
726 if (in->is_ConstraintCast()) {
727 in = in->in(1);
728 continue;
729 } else if (in->is_AddP()) {
730 assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access");
731 in = in->in(AddPNode::Address);
732 continue;
733 } else if (in->is_Con()) {
734 if (trace) {tty->print("Found constant"); in->dump();}
735 } else if (in->is_ShenandoahBarrier()) {
736 if (t == ShenandoahOopStore) {
737 if (in->Opcode() != Op_ShenandoahWriteBarrier) {
738 return false;
739 }
740 uint i = 0;
741 for (; i < phis.size(); i++) {
742 Node* n = phis.node_at(i);
743 if (n->Opcode() == Op_ShenandoahEnqueueBarrier) {
744 break;
745 }
746 }
747 if (i == phis.size()) {
748 return false;
749 }
750 } else if (t == ShenandoahStore && in->Opcode() != Op_ShenandoahWriteBarrier) {
751 return false;
752 }
753 barriers_used.push(in);
754 if (trace) {tty->print("Found barrier"); in->dump();}
755 } else if (in->Opcode() == Op_ShenandoahEnqueueBarrier) {
756 if (t != ShenandoahOopStore) {
757 return false;
758 }
759 if (trace) {tty->print("Found enqueue barrier"); in->dump();}
760 phis.push(in, in->req());
761 in = in->in(1);
762 continue;
763 } else if (in->is_Proj() && in->in(0)->is_Allocate()) {
764 if (trace) {tty->print("Found alloc"); in->in(0)->dump();}
765 } else if (in->is_Phi()) {
766 if (!visited.test_set(in->_idx)) {
767 if (trace) {tty->print("Pushed phi:"); in->dump();}
768 phis.push(in, 2);
769 in = in->in(1);
770 continue;
771 }
772 if (trace) {tty->print("Already seen phi:"); in->dump();}
773 } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) {
774 if (!visited.test_set(in->_idx)) {
775 if (trace) {tty->print("Pushed cmovep:"); in->dump();}
776 phis.push(in, CMoveNode::IfTrue);
777 in = in->in(CMoveNode::IfFalse);
778 continue;
779 }
780 if (trace) {tty->print("Already seen cmovep:"); in->dump();}
781 } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) {
782 in = in->in(1);
783 continue;
784 } else {
785 return false;
786 }
787 }
788 bool cont = false;
789 while (phis.is_nonempty()) {
790 uint idx = phis.index();
791 Node* phi = phis.node();
792 if (idx >= phi->req()) {
793 if (trace) {tty->print("Popped phi:"); phi->dump();}
794 phis.pop();
795 continue;
796 }
797 if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();}
798 in = phi->in(idx);
799 phis.set_index(idx+1);
800 cont = true;
801 break;
802 }
803 if (!cont) {
804 break;
805 }
806 }
807 return true;
808 }
809
810 void ShenandoahBarrierNode::report_verify_failure(const char *msg, Node *n1, Node *n2) {
811 if (n1 != NULL) {
812 n1->dump(+10);
813 }
814 if (n2 != NULL) {
815 n2->dump(+10);
816 }
817 fatal("%s", msg);
818 }
819
820 void ShenandoahBarrierNode::verify(RootNode* root) {
821 ResourceMark rm;
822 Unique_Node_List wq;
823 GrowableArray<Node*> barriers;
824 Unique_Node_List barriers_used;
825 Node_Stack phis(0);
826 VectorSet visited(Thread::current()->resource_area());
827 const bool trace = false;
828 const bool verify_no_useless_barrier = false;
829
830 wq.push(root);
831 for (uint next = 0; next < wq.size(); next++) {
832 Node *n = wq.at(next);
833 if (n->is_Load()) {
834 const bool trace = false;
835 if (trace) {tty->print("Verifying"); n->dump();}
836 if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) {
837 if (trace) {tty->print_cr("Load range/klass");}
838 } else {
839 const TypePtr* adr_type = n->as_Load()->adr_type();
840
841 if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) {
842 if (trace) {tty->print_cr("Mark load");}
843 } else if (adr_type->isa_instptr() &&
844 adr_type->is_instptr()->klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
845 adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset) {
846 if (trace) {tty->print_cr("Reference.get()");}
847 } else {
848 bool verify = true;
849 if (adr_type->isa_instptr()) {
850 const TypeInstPtr* tinst = adr_type->is_instptr();
851 ciKlass* k = tinst->klass();
852 assert(k->is_instance_klass(), "");
853 ciInstanceKlass* ik = (ciInstanceKlass*)k;
854 int offset = adr_type->offset();
855
856 if ((ik->debug_final_field_at(offset) && ShenandoahOptimizeInstanceFinals) ||
857 (ik->debug_stable_field_at(offset) && ShenandoahOptimizeStableFinals)) {
858 if (trace) {tty->print_cr("Final/stable");}
859 verify = false;
860 } else if (k == ciEnv::current()->Class_klass() &&
861 tinst->const_oop() != NULL &&
862 tinst->offset() >= (ik->size_helper() * wordSize)) {
863 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
864 ciField* field = k->get_field_by_offset(tinst->offset(), true);
865 if ((ShenandoahOptimizeStaticFinals && field->is_final()) ||
866 (ShenandoahOptimizeStableFinals && field->is_stable())) {
867 verify = false;
868 }
869 }
870 }
871
872 if (verify && !ShenandoahBarrierNode::verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) {
873 report_verify_failure("Shenandoah verification: Load should have barriers", n);
874 }
875 }
876 }
877 } else if (n->is_Store()) {
878 const bool trace = false;
879
880 if (trace) {tty->print("Verifying"); n->dump();}
881 if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) {
882 Node* adr = n->in(MemNode::Address);
883 bool verify = true;
884
885 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) {
886 adr = adr->in(AddPNode::Address);
887 if (adr->is_AddP()) {
888 assert(adr->in(AddPNode::Base)->is_top(), "");
889 adr = adr->in(AddPNode::Address);
890 if (adr->Opcode() == Op_LoadP &&
891 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() &&
892 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
893 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) {
894 if (trace) {tty->print_cr("SATB prebarrier");}
895 verify = false;
896 }
897 }
898 }
899
900 if (verify && !ShenandoahBarrierNode::verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahStoreValEnqueueBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
901 report_verify_failure("Shenandoah verification: Store should have barriers", n);
902 }
903 }
904 if (!ShenandoahBarrierNode::verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
905 report_verify_failure("Shenandoah verification: Store (address) should have barriers", n);
906 }
907 } else if (n->Opcode() == Op_CmpP) {
908 const bool trace = false;
909
910 Node* in1 = n->in(1);
911 Node* in2 = n->in(2);
912 if (in1->bottom_type()->isa_oopptr()) {
913 if (trace) {tty->print("Verifying"); n->dump();}
914
915 bool mark_inputs = false;
916 if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR ||
917 (in1->is_Con() || in2->is_Con())) {
918 if (trace) {tty->print_cr("Comparison against a constant");}
919 mark_inputs = true;
920 } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) ||
921 (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) {
922 if (trace) {tty->print_cr("Comparison with newly alloc'ed object");}
923 mark_inputs = true;
924 } else {
925 assert(in2->bottom_type()->isa_oopptr(), "");
926
927 if (!ShenandoahBarrierNode::verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) ||
928 !ShenandoahBarrierNode::verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) {
929 report_verify_failure("Shenandoah verification: Cmp should have barriers", n);
930 }
931 }
932 if (verify_no_useless_barrier &&
933 mark_inputs &&
934 (!ShenandoahBarrierNode::verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) ||
935 !ShenandoahBarrierNode::verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) {
936 phis.clear();
937 visited.Reset();
938 }
939 }
940 } else if (n->is_LoadStore()) {
941 if (n->in(MemNode::ValueIn)->bottom_type()->make_ptr() &&
942 !ShenandoahBarrierNode::verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahStoreValEnqueueBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
943 report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n);
944 }
945
946 if (n->in(MemNode::Address)->bottom_type()->make_oopptr() && !ShenandoahBarrierNode::verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
947 report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n);
948 }
949 } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) {
950 CallNode* call = n->as_Call();
951
952 static struct {
953 const char* name;
954 struct {
955 int pos;
956 verify_type t;
957 } args[6];
958 } calls[] = {
959 "aescrypt_encryptBlock",
960 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
961 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
962 "aescrypt_decryptBlock",
963 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
964 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
965 "multiplyToLen",
966 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { TypeFunc::Parms+4, ShenandoahStore },
967 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
968 "squareToLen",
969 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { -1, ShenandoahNone},
970 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
971 "montgomery_multiply",
972 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
973 { TypeFunc::Parms+6, ShenandoahStore }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
974 "montgomery_square",
975 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+5, ShenandoahStore },
976 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
977 "mulAdd",
978 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
979 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
980 "vectorizedMismatch",
981 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
982 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
983 "updateBytesCRC32",
984 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
985 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
986 "updateBytesAdler32",
987 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
988 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
989 "updateBytesCRC32C",
990 { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad}, { -1, ShenandoahNone},
991 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
992 "counterMode_AESCrypt",
993 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
994 { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } },
995 "cipherBlockChaining_encryptAESCrypt",
996 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
997 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
998 "cipherBlockChaining_decryptAESCrypt",
999 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
1000 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1001 "shenandoah_clone_barrier",
1002 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
1003 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1004 "ghash_processBlocks",
1005 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
1006 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1007 "sha1_implCompress",
1008 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1009 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1010 "sha256_implCompress",
1011 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1012 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1013 "sha512_implCompress",
1014 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1015 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1016 "sha1_implCompressMB",
1017 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1018 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1019 "sha256_implCompressMB",
1020 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1021 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1022 "sha512_implCompressMB",
1023 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1024 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1025 };
1026
1027 if (call->is_call_to_arraycopystub()) {
1028 Node* dest = NULL;
1029 const TypeTuple* args = n->as_Call()->_tf->domain();
1030 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
1031 if (args->field_at(i)->isa_ptr()) {
1032 j++;
1033 if (j == 2) {
1034 dest = n->in(i);
1035 break;
1036 }
1037 }
1038 }
1039 if (!ShenandoahBarrierNode::verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) ||
1040 !ShenandoahBarrierNode::verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) {
1041 report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n);
1042 }
1043 } else if (strlen(call->_name) > 5 &&
1044 !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) {
1045 if (!ShenandoahBarrierNode::verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) {
1046 report_verify_failure("Shenandoah verification: _fill should have barriers", n);
1047 }
1048 } else if (!strcmp(call->_name, "shenandoah_wb_pre")) {
1049 // skip
1050 } else {
1051 const int calls_len = sizeof(calls) / sizeof(calls[0]);
1052 int i = 0;
1053 for (; i < calls_len; i++) {
1054 if (!strcmp(calls[i].name, call->_name)) {
1055 break;
1056 }
1057 }
1058 if (i != calls_len) {
1059 const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]);
1060 for (uint j = 0; j < args_len; j++) {
1061 int pos = calls[i].args[j].pos;
1062 if (pos == -1) {
1063 break;
1064 }
1065 if (!ShenandoahBarrierNode::verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) {
1066 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
1067 }
1068 }
1069 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
1070 if (call->in(j)->bottom_type()->make_ptr() &&
1071 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
1072 uint k = 0;
1073 for (; k < args_len && calls[i].args[k].pos != (int)j; k++);
1074 if (k == args_len) {
1075 fatal("arg %d for call %s not covered", j, call->_name);
1076 }
1077 }
1078 }
1079 } else {
1080 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
1081 if (call->in(j)->bottom_type()->make_ptr() &&
1082 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
1083 fatal("%s not covered", call->_name);
1084 }
1085 }
1086 }
1087 }
1088 } else if (n->is_ShenandoahBarrier()) {
1089 assert(!barriers.contains(n), "");
1090 assert(n->Opcode() != Op_ShenandoahWriteBarrier || n->find_out_with(Op_ShenandoahWBMemProj) != NULL, "bad shenandoah write barrier");
1091 assert(n->Opcode() != Op_ShenandoahWriteBarrier || n->outcnt() > 1, "bad shenandoah write barrier");
1092 barriers.push(n);
1093 } else if (n->Opcode() == Op_ShenandoahEnqueueBarrier) {
1094 // skip
1095 } else if (n->Opcode() == Op_ShenandoahWBMemProj) {
1096 assert(n->in(0) == NULL && n->in(ShenandoahWBMemProjNode::WriteBarrier)->Opcode() == Op_ShenandoahWriteBarrier, "strange ShenandoahWBMemProj");
1097 } else if (n->is_AddP()
1098 || n->is_Phi()
1099 || n->is_ConstraintCast()
1100 || n->Opcode() == Op_Return
1101 || n->Opcode() == Op_CMoveP
1102 || n->Opcode() == Op_CMoveN
1103 || n->Opcode() == Op_Rethrow
1104 || n->is_MemBar()
1105 || n->Opcode() == Op_Conv2B
1106 || n->Opcode() == Op_SafePoint
1107 || n->is_CallJava()
1108 || n->Opcode() == Op_Unlock
1109 || n->Opcode() == Op_EncodeP
1110 || n->Opcode() == Op_DecodeN) {
1111 // nothing to do
1112 } else {
1113 static struct {
1114 int opcode;
1115 struct {
1116 int pos;
1117 verify_type t;
1118 } inputs[2];
1119 } others[] = {
1120 Op_FastLock,
1121 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
1122 Op_Lock,
1123 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone} },
1124 Op_ArrayCopy,
1125 { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } },
1126 Op_StrCompressedCopy,
1127 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
1128 Op_StrInflatedCopy,
1129 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
1130 Op_AryEq,
1131 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
1132 Op_StrIndexOf,
1133 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
1134 Op_StrComp,
1135 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
1136 Op_StrEquals,
1137 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
1138 Op_EncodeISOArray,
1139 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
1140 Op_HasNegatives,
1141 { { 2, ShenandoahLoad }, { -1, ShenandoahNone} },
1142 Op_CastP2X,
1143 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
1144 Op_StrIndexOfChar,
1145 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } },
1146 };
1147
1148 const int others_len = sizeof(others) / sizeof(others[0]);
1149 int i = 0;
1150 for (; i < others_len; i++) {
1151 if (others[i].opcode == n->Opcode()) {
1152 break;
1153 }
1154 }
1155 uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req();
1156 if (i != others_len) {
1157 const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]);
1158 for (uint j = 0; j < inputs_len; j++) {
1159 int pos = others[i].inputs[j].pos;
1160 if (pos == -1) {
1161 break;
1162 }
1163 if (!ShenandoahBarrierNode::verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) {
1164 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
1165 }
1166 }
1167 for (uint j = 1; j < stop; j++) {
1168 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
1169 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
1170 uint k = 0;
1171 for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++);
1172 if (k == inputs_len) {
1173 fatal("arg %d for node %s not covered", j, n->Name());
1174 }
1175 }
1176 }
1177 } else {
1178 for (uint j = 1; j < stop; j++) {
1179 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
1180 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
1181 fatal("%s not covered", n->Name());
1182 }
1183 }
1184 }
1185 }
1186
1187 if (n->is_SafePoint()) {
1188 SafePointNode* sfpt = n->as_SafePoint();
1189 if (verify_no_useless_barrier && sfpt->jvms() != NULL) {
1190 for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) {
1191 if (!ShenandoahBarrierNode::verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) {
1192 phis.clear();
1193 visited.Reset();
1194 }
1195 }
1196 }
1197 }
1198 for( uint i = 0; i < n->len(); ++i ) {
1199 Node *m = n->in(i);
1200 if (m == NULL) continue;
1201
1202 // In most cases, inputs should be known to be non null. If it's
1203 // not the case, it could be a missing cast_not_null() in an
1204 // intrinsic or support might be needed in AddPNode::Ideal() to
1205 // avoid a NULL+offset input.
1206 if (!(n->is_Phi() ||
1207 (n->is_SafePoint() && (!n->is_CallRuntime() || !strcmp(n->as_Call()->_name, "shenandoah_wb_pre") || !strcmp(n->as_Call()->_name, "unsafe_arraycopy"))) ||
1208 n->Opcode() == Op_CmpP ||
1209 n->Opcode() == Op_CmpN ||
1210 (n->Opcode() == Op_StoreP && i == StoreNode::ValueIn) ||
1211 (n->Opcode() == Op_StoreN && i == StoreNode::ValueIn) ||
1212 n->is_ConstraintCast() ||
1213 n->Opcode() == Op_Return ||
1214 n->Opcode() == Op_Conv2B ||
1215 n->is_AddP() ||
1216 n->Opcode() == Op_CMoveP ||
1217 n->Opcode() == Op_CMoveN ||
1218 n->Opcode() == Op_Rethrow ||
1219 n->is_MemBar() ||
1220 n->is_Mem() ||
1221 n->Opcode() == Op_AryEq ||
1222 n->Opcode() == Op_SCMemProj ||
1223 n->Opcode() == Op_EncodeP ||
1224 n->Opcode() == Op_DecodeN ||
1225 n->Opcode() == Op_ShenandoahWriteBarrier ||
1226 n->Opcode() == Op_ShenandoahWBMemProj ||
1227 n->Opcode() == Op_ShenandoahEnqueueBarrier)) {
1228 if (m->bottom_type()->make_oopptr() && m->bottom_type()->make_oopptr()->meet(TypePtr::NULL_PTR) == m->bottom_type()) {
1229 report_verify_failure("Shenandoah verification: null input", n, m);
1230 }
1231 }
1232
1233 wq.push(m);
1234 }
1235 }
1236
1237 if (verify_no_useless_barrier) {
1238 for (int i = 0; i < barriers.length(); i++) {
1239 Node* n = barriers.at(i);
1240 if (!barriers_used.member(n)) {
1241 tty->print("XXX useless barrier"); n->dump(-2);
1242 ShouldNotReachHere();
1243 }
1244 }
1245 }
1246 }
1247 #endif
1248
1249 bool ShenandoahBarrierNode::is_dominator_same_ctrl(Node*c, Node* d, Node* n, PhaseIdealLoop* phase) {
1250 // That both nodes have the same control is not sufficient to prove
1251 // domination, verify that there's no path from d to n
1252 ResourceMark rm;
1253 Unique_Node_List wq;
1254 wq.push(d);
1255 for (uint next = 0; next < wq.size(); next++) {
1256 Node *m = wq.at(next);
1257 if (m == n) {
1258 return false;
1259 }
1260 if (m->is_Phi() && m->in(0)->is_Loop()) {
1261 assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control");
1262 } else {
1263 for (uint i = 0; i < m->req(); i++) {
1264 if (m->in(i) != NULL && phase->ctrl_or_self(m->in(i)) == c) {
1265 wq.push(m->in(i));
1266 }
1267 }
1268 }
1269 }
1270 return true;
1271 }
1272
1273 bool ShenandoahBarrierNode::is_dominator(Node *d_c, Node *n_c, Node* d, Node* n, PhaseIdealLoop* phase) {
1274 if (d_c != n_c) {
1275 return phase->is_dominator(d_c, n_c);
1276 }
1277 return is_dominator_same_ctrl(d_c, d, n, phase);
1278 }
1279
1280 Node* next_mem(Node* mem, int alias) {
1281 Node* res = NULL;
1282 if (mem->is_Proj()) {
1283 res = mem->in(0);
1284 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
1285 res = mem->in(TypeFunc::Memory);
1286 } else if (mem->is_Phi()) {
1287 res = mem->in(1);
1288 } else if (mem->is_ShenandoahBarrier()) {
1289 res = mem->in(ShenandoahBarrierNode::Memory);
1290 } else if (mem->is_MergeMem()) {
1291 res = mem->as_MergeMem()->memory_at(alias);
1292 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
1293 assert(alias = Compile::AliasIdxRaw, "following raw memory can't lead to a barrier");
1294 res = mem->in(MemNode::Memory);
1295 } else if (mem->Opcode() == Op_ShenandoahWBMemProj) {
1296 res = mem->in(ShenandoahWBMemProjNode::WriteBarrier);
1297 } else {
1298 #ifdef ASSERT
1299 mem->dump();
1300 #endif
1301 ShouldNotReachHere();
1302 }
1303 return res;
1304 }
1305
1306 Node* ShenandoahBarrierNode::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) {
1307 Node* iffproj = NULL;
1308 while (c != dom) {
1309 Node* next = phase->idom(c);
1310 assert(next->unique_ctrl_out() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?");
1311 if (c->is_Region()) {
1312 ResourceMark rm;
1313 Unique_Node_List wq;
1314 wq.push(c);
1315 for (uint i = 0; i < wq.size(); i++) {
1316 Node *n = wq.at(i);
1317 if (n == next) {
1318 continue;
1319 }
1320 if (n->is_Region()) {
1321 for (uint j = 1; j < n->req(); j++) {
1322 wq.push(n->in(j));
1323 }
1324 } else {
1325 wq.push(n->in(0));
1326 }
1327 }
1328 for (uint i = 0; i < wq.size(); i++) {
1329 Node *n = wq.at(i);
1330 assert(n->is_CFG(), "");
1331 if (n->is_Multi()) {
1332 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1333 Node* u = n->fast_out(j);
1334 if (u->is_CFG()) {
1335 if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
1336 return NodeSentinel;
1337 }
1338 }
1339 }
1340 }
1341 }
1342 } else if (c->is_Proj()) {
1343 if (c->is_IfProj()) {
1344 if (c->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) != NULL) {
1345 // continue;
1346 } else {
1347 if (!allow_one_proj) {
1348 return NodeSentinel;
1349 }
1350 if (iffproj == NULL) {
1351 iffproj = c;
1352 } else {
1353 return NodeSentinel;
1354 }
1355 }
1356 } else if (c->Opcode() == Op_JumpProj) {
1357 return NodeSentinel; // unsupported
1358 } else if (c->Opcode() == Op_CatchProj) {
1359 return NodeSentinel; // unsupported
1360 } else if (c->Opcode() == Op_CProj && next->Opcode() == Op_NeverBranch) {
1361 return NodeSentinel; // unsupported
1362 } else {
1363 assert(next->unique_ctrl_out() == c, "unsupported branch pattern");
1364 }
1365 }
1366 c = next;
1367 }
1368 return iffproj;
1369 }
1370
1371 #ifdef ASSERT
1372 void ShenandoahWriteBarrierNode::memory_dominates_all_paths_helper(Node* c, Node* rep_ctrl, Unique_Node_List& controls, PhaseIdealLoop* phase) {
1373 const bool trace = false;
1374 if (trace) { tty->print("X control is"); c->dump(); }
1375
1376 uint start = controls.size();
1377 controls.push(c);
1378 for (uint i = start; i < controls.size(); i++) {
1379 Node *n = controls.at(i);
1380
1381 if (trace) { tty->print("X from"); n->dump(); }
1382
1383 if (n == rep_ctrl) {
1384 continue;
1385 }
1386
1387 if (n->is_Proj()) {
1388 Node* n_dom = n->in(0);
1389 IdealLoopTree* n_dom_loop = phase->get_loop(n_dom);
1390 if (n->is_IfProj() && n_dom->outcnt() == 2) {
1391 n_dom_loop = phase->get_loop(n_dom->as_If()->proj_out(n->as_Proj()->_con == 0 ? 1 : 0));
1392 }
1393 if (n_dom_loop != phase->ltree_root()) {
1394 Node* tail = n_dom_loop->tail();
1395 if (tail->is_Region()) {
1396 for (uint j = 1; j < tail->req(); j++) {
1397 if (phase->is_dominator(n_dom, tail->in(j)) && !phase->is_dominator(n, tail->in(j))) {
1398 assert(phase->is_dominator(rep_ctrl, tail->in(j)), "why are we here?");
1399 // entering loop from below, mark backedge
1400 if (trace) { tty->print("X pushing backedge"); tail->in(j)->dump(); }
1401 controls.push(tail->in(j));
1402 //assert(n->in(0) == n_dom, "strange flow control");
1403 }
1404 }
1405 } else if (phase->get_loop(n) != n_dom_loop && phase->is_dominator(n_dom, tail)) {
1406 // entering loop from below, mark backedge
1407 if (trace) { tty->print("X pushing backedge"); tail->dump(); }
1408 controls.push(tail);
1409 //assert(n->in(0) == n_dom, "strange flow control");
1410 }
1411 }
1412 }
1413
1414 if (n->is_Loop()) {
1415 Node* c = n->in(LoopNode::EntryControl);
1416 if (trace) { tty->print("X pushing"); c->dump(); }
1417 controls.push(c);
1418 } else if (n->is_Region()) {
1419 for (uint i = 1; i < n->req(); i++) {
1420 Node* c = n->in(i);
1421 if (trace) { tty->print("X pushing"); c->dump(); }
1422 controls.push(c);
1423 }
1424 } else {
1425 Node* c = n->in(0);
1426 if (trace) { tty->print("X pushing"); c->dump(); }
1427 controls.push(c);
1428 }
1429 }
1430 }
1431
1432 bool ShenandoahWriteBarrierNode::memory_dominates_all_paths(Node* mem, Node* rep_ctrl, int alias, PhaseIdealLoop* phase) {
1433 const bool trace = false;
1434 if (trace) {
1435 tty->print("XXX mem is"); mem->dump();
1436 tty->print("XXX rep ctrl is"); rep_ctrl->dump();
1437 tty->print_cr("XXX alias is %d", alias);
1438 }
1439 ResourceMark rm;
1440 Unique_Node_List wq;
1441 Unique_Node_List controls;
1442 wq.push(mem);
1443 for (uint next = 0; next < wq.size(); next++) {
1444 Node *nn = wq.at(next);
1445 if (trace) { tty->print("XX from mem"); nn->dump(); }
1446 assert(nn->bottom_type() == Type::MEMORY, "memory only");
1447
1448 if (nn->is_Phi()) {
1449 Node* r = nn->in(0);
1450 for (DUIterator_Fast jmax, j = r->fast_outs(jmax); j < jmax; j++) {
1451 Node* u = r->fast_out(j);
1452 if (u->is_Phi() && u->bottom_type() == Type::MEMORY && u != nn &&
1453 (u->adr_type() == TypePtr::BOTTOM || phase->C->get_alias_index(u->adr_type()) == alias)) {
1454 if (trace) { tty->print("XX Next mem (other phi)"); u->dump(); }
1455 wq.push(u);
1456 }
1457 }
1458 }
1459
1460 for (DUIterator_Fast imax, i = nn->fast_outs(imax); i < imax; i++) {
1461 Node* use = nn->fast_out(i);
1462
1463 if (trace) { tty->print("XX use %p", use->adr_type()); use->dump(); }
1464 if (use->is_CFG() && use->in(TypeFunc::Memory) == nn) {
1465 Node* c = use->in(0);
1466 if (phase->is_dominator(rep_ctrl, c)) {
1467 memory_dominates_all_paths_helper(c, rep_ctrl, controls, phase);
1468 } else if (use->is_CallStaticJava() && use->as_CallStaticJava()->uncommon_trap_request() != 0 && c->is_Region()) {
1469 Node* region = c;
1470 if (trace) { tty->print("XX unc region"); region->dump(); }
1471 for (uint j = 1; j < region->req(); j++) {
1472 if (phase->is_dominator(rep_ctrl, region->in(j))) {
1473 if (trace) { tty->print("XX unc follows"); region->in(j)->dump(); }
1474 memory_dominates_all_paths_helper(region->in(j), rep_ctrl, controls, phase);
1475 }
1476 }
1477 }
1478 //continue;
1479 } else if (use->is_Phi()) {
1480 assert(use->bottom_type() == Type::MEMORY, "bad phi");
1481 if ((use->adr_type() == TypePtr::BOTTOM) ||
1482 phase->C->get_alias_index(use->adr_type()) == alias) {
1483 for (uint j = 1; j < use->req(); j++) {
1484 if (use->in(j) == nn) {
1485 Node* c = use->in(0)->in(j);
1486 if (phase->is_dominator(rep_ctrl, c)) {
1487 memory_dominates_all_paths_helper(c, rep_ctrl, controls, phase);
1488 }
1489 }
1490 }
1491 }
1492 // continue;
1493 }
1494
1495 if (use->is_MergeMem()) {
1496 if (use->as_MergeMem()->memory_at(alias) == nn) {
1497 if (trace) { tty->print("XX Next mem"); use->dump(); }
1498 // follow the memory edges
1499 wq.push(use);
1500 }
1501 } else if (use->is_Phi()) {
1502 assert(use->bottom_type() == Type::MEMORY, "bad phi");
1503 if ((use->adr_type() == TypePtr::BOTTOM) ||
1504 phase->C->get_alias_index(use->adr_type()) == alias) {
1505 if (trace) { tty->print("XX Next mem"); use->dump(); }
1506 // follow the memory edges
1507 wq.push(use);
1508 }
1509 } else if (use->bottom_type() == Type::MEMORY &&
1510 (use->adr_type() == TypePtr::BOTTOM || phase->C->get_alias_index(use->adr_type()) == alias)) {
1511 if (trace) { tty->print("XX Next mem"); use->dump(); }
1512 // follow the memory edges
1513 wq.push(use);
1514 } else if ((use->is_SafePoint() || use->is_MemBar()) &&
1515 (use->adr_type() == TypePtr::BOTTOM || phase->C->get_alias_index(use->adr_type()) == alias)) {
1516 for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) {
1517 Node* u = use->fast_out(j);
1518 if (u->bottom_type() == Type::MEMORY) {
1519 if (trace) { tty->print("XX Next mem"); u->dump(); }
1520 // follow the memory edges
1521 wq.push(u);
1522 }
1523 }
1524 } else if (use->Opcode() == Op_ShenandoahWriteBarrier && phase->C->get_alias_index(use->adr_type()) == alias) {
1525 Node* m = use->find_out_with(Op_ShenandoahWBMemProj);
1526 if (m != NULL) {
1527 if (trace) { tty->print("XX Next mem"); m->dump(); }
1528 // follow the memory edges
1529 wq.push(m);
1530 }
1531 }
1532 }
1533 }
1534
1535 if (controls.size() == 0) {
1536 return false;
1537 }
1538
1539 for (uint i = 0; i < controls.size(); i++) {
1540 Node *n = controls.at(i);
1541
1542 if (trace) { tty->print("X checking"); n->dump(); }
1543
1544 if (n->unique_ctrl_out() != NULL) {
1545 continue;
1546 }
1547
1548 if (n->Opcode() == Op_NeverBranch) {
1549 Node* taken = n->as_Multi()->proj_out(0);
1550 if (!controls.member(taken)) {
1551 if (trace) { tty->print("X not seen"); taken->dump(); }
1552 return false;
1553 }
1554 continue;
1555 }
1556
1557 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1558 Node* u = n->fast_out(j);
1559
1560 if (u->is_CFG()) {
1561 if (!controls.member(u)) {
1562 if (u->is_Proj() && u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
1563 if (trace) { tty->print("X not seen but unc"); u->dump(); }
1564 } else {
1565 Node* c = u;
1566 do {
1567 c = c->unique_ctrl_out();
1568 } while (c != NULL && c->is_Region());
1569 if (c != NULL && c->Opcode() == Op_Halt) {
1570 if (trace) { tty->print("X not seen but halt"); c->dump(); }
1571 } else {
1572 if (trace) { tty->print("X not seen"); u->dump(); }
1573 return false;
1574 }
1575 }
1576 } else {
1577 if (trace) { tty->print("X seen"); u->dump(); }
1578 }
1579 }
1580 }
1581 }
1582 return true;
1583 }
1584 #endif
1585
1586 Node* ShenandoahBarrierNode::dom_mem(Node* mem, Node*& mem_ctrl, Node* n, Node* rep_ctrl, int alias, PhaseIdealLoop* phase) {
1587 ResourceMark rm;
1588 VectorSet wq(Thread::current()->resource_area());
1589 wq.set(mem->_idx);
1590 mem_ctrl = phase->get_ctrl(mem);
1591 while (!is_dominator(mem_ctrl, rep_ctrl, mem, n, phase)) {
1592 mem = next_mem(mem, alias);
1593 if (wq.test_set(mem->_idx)) {
1594 return NULL; // hit an unexpected loop
1595 }
1596 mem_ctrl = phase->ctrl_or_self(mem);
1597 }
1598 if (mem->is_MergeMem()) {
1599 mem = mem->as_MergeMem()->memory_at(alias);
1600 mem_ctrl = phase->ctrl_or_self(mem);
1601 }
1602 return mem;
1603 }
1604
1605 Node* ShenandoahBarrierNode::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) {
1606 ResourceMark rm;
1607 VectorSet wq(Thread::current()->resource_area());
1608 wq.set(mem->_idx);
1609 mem_ctrl = phase->ctrl_or_self(mem);
1610 while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) {
1611 mem = next_mem(mem, alias);
1612 if (wq.test_set(mem->_idx)) {
1613 return NULL;
1614 }
1615 mem_ctrl = phase->ctrl_or_self(mem);
1616 }
1617 if (mem->is_MergeMem()) {
1618 mem = mem->as_MergeMem()->memory_at(alias);
1619 mem_ctrl = phase->ctrl_or_self(mem);
1620 }
1621 return mem;
1622 }
1623
1624 static void disconnect_barrier_mem(Node* wb, PhaseIterGVN& igvn) {
1625 Node* mem_in = wb->in(ShenandoahBarrierNode::Memory);
1626 Node* proj = wb->find_out_with(Op_ShenandoahWBMemProj);
1627
1628 for (DUIterator_Last imin, i = proj->last_outs(imin); i >= imin; ) {
1629 Node* u = proj->last_out(i);
1630 igvn.rehash_node_delayed(u);
1631 int nb = u->replace_edge(proj, mem_in);
1632 assert(nb > 0, "no replacement?");
1633 i -= nb;
1634 }
1635 }
1636
1637 Node* ShenandoahWriteBarrierNode::move_above_predicates(LoopNode* cl, Node* val_ctrl, PhaseIdealLoop* phase) {
1638 Node* entry = cl->skip_strip_mined(-1)->in(LoopNode::EntryControl);
1639 Node* above_pred = phase->skip_all_loop_predicates(entry);
1640 Node* ctrl = entry;
1641 while (ctrl != above_pred) {
1642 Node* next = ctrl->in(0);
1643 if (!phase->is_dominator(val_ctrl, next)) {
1644 break;
1645 }
1646 ctrl = next;
1647 }
1648 return ctrl;
1649 }
1650
1651 static MemoryGraphFixer* find_fixer(GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, int alias) {
1652 for (int i = 0; i < memory_graph_fixers.length(); i++) {
1653 if (memory_graph_fixers.at(i)->alias() == alias) {
1654 return memory_graph_fixers.at(i);
1655 }
1656 }
1657 return NULL;
1658 }
1659
1660 static MemoryGraphFixer* create_fixer(GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, int alias, PhaseIdealLoop* phase, bool include_lsm) {
1661 assert(find_fixer(memory_graph_fixers, alias) == NULL, "none should exist yet");
1662 MemoryGraphFixer* fixer = new MemoryGraphFixer(alias, include_lsm, phase);
1663 memory_graph_fixers.push(fixer);
1664 return fixer;
1665 }
1666
1667 void ShenandoahWriteBarrierNode::try_move_before_loop_helper(LoopNode* cl, Node* val_ctrl, GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, PhaseIdealLoop* phase, bool include_lsm, Unique_Node_List& uses) {
1668 assert(cl->is_Loop(), "bad control");
1669 Node* ctrl = move_above_predicates(cl, val_ctrl, phase);
1670 Node* mem_ctrl = NULL;
1671 int alias = phase->C->get_alias_index(adr_type());
1672
1673 MemoryGraphFixer* fixer = find_fixer(memory_graph_fixers, alias);
1674 if (fixer == NULL) {
1675 fixer = create_fixer(memory_graph_fixers, alias, phase, include_lsm);
1676 }
1677
1678 Node* proj = find_out_with(Op_ShenandoahWBMemProj);
1679
1680 fixer->remove(proj);
1681 Node* mem = fixer->find_mem(ctrl, NULL);
1682
1683 assert(!ShenandoahVerifyOptoBarriers || memory_dominates_all_paths(mem, ctrl, alias, phase), "can't fix the memory graph");
1684
1685 phase->set_ctrl_and_loop(this, ctrl);
1686 phase->igvn().replace_input_of(this, Control, ctrl);
1687
1688 disconnect_barrier_mem(this, phase->igvn());
1689
1690 phase->igvn().replace_input_of(this, Memory, mem);
1691 phase->set_ctrl_and_loop(proj, ctrl);
1692
1693 fixer->fix_mem(ctrl, ctrl, mem, mem, proj, uses);
1694 assert(proj->outcnt() > 0, "disconnected write barrier");
1695 }
1696
1697 LoopNode* ShenandoahWriteBarrierNode::try_move_before_pre_loop(Node* c, Node* val_ctrl, PhaseIdealLoop* phase) {
1698 // A write barrier between a pre and main loop can get in the way of
1699 // vectorization. Move it above the pre loop if possible
1700 CountedLoopNode* cl = NULL;
1701 if (c->is_IfFalse() &&
1702 c->in(0)->is_CountedLoopEnd()) {
1703 cl = c->in(0)->as_CountedLoopEnd()->loopnode();
1704 } else if (c->is_IfProj() &&
1705 c->in(0)->is_If() &&
1706 c->in(0)->in(0)->is_IfFalse() &&
1707 c->in(0)->in(0)->in(0)->is_CountedLoopEnd()) {
1708 cl = c->in(0)->in(0)->in(0)->as_CountedLoopEnd()->loopnode();
1709 }
1710 if (cl != NULL &&
1711 cl->is_pre_loop() &&
1712 val_ctrl != cl &&
1713 phase->is_dominator(val_ctrl, cl)) {
1714 return cl;
1715 }
1716 return NULL;
1717 }
1718
1719 void ShenandoahWriteBarrierNode::try_move_before_loop(GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, PhaseIdealLoop* phase, bool include_lsm, Unique_Node_List& uses) {
1720 Node *n_ctrl = phase->get_ctrl(this);
1721 IdealLoopTree *n_loop = phase->get_loop(n_ctrl);
1722 Node* val = in(ValueIn);
1723 Node* val_ctrl = phase->get_ctrl(val);
1724 if (n_loop != phase->ltree_root() && !n_loop->_irreducible) {
1725 IdealLoopTree *val_loop = phase->get_loop(val_ctrl);
1726 Node* mem = in(Memory);
1727 IdealLoopTree *mem_loop = phase->get_loop(phase->get_ctrl(mem));
1728 if (!n_loop->is_member(val_loop) &&
1729 n_loop->is_member(mem_loop)) {
1730 Node* n_loop_head = n_loop->_head;
1731
1732 if (n_loop_head->is_Loop()) {
1733 LoopNode* loop = n_loop_head->as_Loop();
1734 if (n_loop_head->is_CountedLoop() && n_loop_head->as_CountedLoop()->is_main_loop()) {
1735 LoopNode* res = try_move_before_pre_loop(n_loop_head->in(LoopNode::EntryControl), val_ctrl, phase);
1736 if (res != NULL) {
1737 loop = res;
1738 }
1739 }
1740
1741 try_move_before_loop_helper(loop, val_ctrl, memory_graph_fixers, phase, include_lsm, uses);
1742 }
1743 }
1744 }
1745 LoopNode* ctrl = try_move_before_pre_loop(in(0), val_ctrl, phase);
1746 if (ctrl != NULL) {
1747 try_move_before_loop_helper(ctrl, val_ctrl, memory_graph_fixers, phase, include_lsm, uses);
1748 }
1749 }
1750
1751 Node* ShenandoahWriteBarrierNode::would_subsume(ShenandoahBarrierNode* other, PhaseIdealLoop* phase) {
1752 Node* val = in(ValueIn);
1753 Node* val_ctrl = phase->get_ctrl(val);
1754 Node* other_mem = other->in(Memory);
1755 Node* other_ctrl = phase->get_ctrl(other);
1756 Node* this_ctrl = phase->get_ctrl(this);
1757 IdealLoopTree* this_loop = phase->get_loop(this_ctrl);
1758 IdealLoopTree* other_loop = phase->get_loop(other_ctrl);
1759
1760 Node* ctrl = phase->dom_lca(other_ctrl, this_ctrl);
1761
1762 if (ctrl->is_Proj() &&
1763 ctrl->in(0)->is_Call() &&
1764 ctrl->unique_ctrl_out() != NULL &&
1765 ctrl->unique_ctrl_out()->Opcode() == Op_Catch &&
1766 !phase->is_dominator(val_ctrl, ctrl->in(0)->in(0))) {
1767 return NULL;
1768 }
1769
1770 IdealLoopTree* loop = phase->get_loop(ctrl);
1771
1772 // We don't want to move a write barrier in a loop
1773 // If the LCA is in a inner loop, try a control out of loop if possible
1774 while (!loop->is_member(this_loop) && (other->Opcode() != Op_ShenandoahWriteBarrier || !loop->is_member(other_loop))) {
1775 ctrl = phase->idom(ctrl);
1776 if (ctrl->is_MultiBranch()) {
1777 ctrl = ctrl->in(0);
1778 }
1779 if (ctrl != val_ctrl && phase->is_dominator(ctrl, val_ctrl)) {
1780 return NULL;
1781 }
1782 loop = phase->get_loop(ctrl);
1783 }
1784
1785 if (ShenandoahDontIncreaseWBFreq) {
1786 Node* this_iffproj = no_branches(this_ctrl, ctrl, true, phase);
1787 if (other->Opcode() == Op_ShenandoahWriteBarrier) {
1788 Node* other_iffproj = no_branches(other_ctrl, ctrl, true, phase);
1789 if (other_iffproj == NULL || this_iffproj == NULL) {
1790 return ctrl;
1791 } else if (other_iffproj != NodeSentinel && this_iffproj != NodeSentinel &&
1792 other_iffproj->in(0) == this_iffproj->in(0)) {
1793 return ctrl;
1794 }
1795 } else if (this_iffproj == NULL) {
1796 return ctrl;
1797 }
1798 return NULL;
1799 }
1800
1801 return ctrl;
1802 }
1803
1804 void ShenandoahWriteBarrierNode::optimize_before_expansion(PhaseIdealLoop* phase, GrowableArray<MemoryGraphFixer*> memory_graph_fixers, bool include_lsm) {
1805 bool progress = false;
1806 Unique_Node_List uses;
1807 do {
1808 progress = false;
1809 for (int i = 0; i < ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count(); i++) {
1810 ShenandoahWriteBarrierNode* wb = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barrier(i);
1811
1812 wb->try_move_before_loop(memory_graph_fixers, phase, include_lsm, uses);
1813
1814 Node* val = wb->in(ValueIn);
1815
1816 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
1817 Node* u = val->fast_out(j);
1818 if (u != wb && u->is_ShenandoahBarrier()) {
1819 Node* rep_ctrl = wb->would_subsume(u->as_ShenandoahBarrier(), phase);
1820
1821 if (rep_ctrl != NULL) {
1822 Node* other = u;
1823 Node* val_ctrl = phase->get_ctrl(val);
1824 if (rep_ctrl->is_Proj() &&
1825 rep_ctrl->in(0)->is_Call() &&
1826 rep_ctrl->unique_ctrl_out() != NULL &&
1827 rep_ctrl->unique_ctrl_out()->Opcode() == Op_Catch) {
1828 rep_ctrl = rep_ctrl->in(0)->in(0);
1829
1830 assert(phase->is_dominator(val_ctrl, rep_ctrl), "bad control");
1831 } else {
1832 LoopNode* c = ShenandoahWriteBarrierNode::try_move_before_pre_loop(rep_ctrl, val_ctrl, phase);
1833 if (c != NULL) {
1834 rep_ctrl = ShenandoahWriteBarrierNode::move_above_predicates(c, val_ctrl, phase);
1835 } else {
1836 while (rep_ctrl->is_IfProj()) {
1837 CallStaticJavaNode* unc = rep_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
1838 if (unc != NULL) {
1839 int req = unc->uncommon_trap_request();
1840 Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
1841 if ((trap_reason == Deoptimization::Reason_loop_limit_check ||
1842 trap_reason == Deoptimization::Reason_predicate ||
1843 trap_reason == Deoptimization::Reason_profile_predicate) &&
1844 phase->is_dominator(val_ctrl, rep_ctrl->in(0)->in(0))) {
1845 rep_ctrl = rep_ctrl->in(0)->in(0);
1846 continue;
1847 }
1848 }
1849 break;
1850 }
1851 }
1852 }
1853
1854 Node* wb_ctrl = phase->get_ctrl(wb);
1855 Node* other_ctrl = phase->get_ctrl(other);
1856 int alias = phase->C->get_alias_index(wb->adr_type());
1857 MemoryGraphFixer* fixer = find_fixer(memory_graph_fixers, alias);;
1858 if (!is_dominator(wb_ctrl, other_ctrl, wb, other, phase)) {
1859 if (fixer == NULL) {
1860 fixer = create_fixer(memory_graph_fixers, alias, phase, include_lsm);
1861 }
1862 Node* mem = fixer->find_mem(rep_ctrl, phase->get_ctrl(other) == rep_ctrl ? other : NULL);
1863
1864 if (mem->has_out_with(Op_Lock) || mem->has_out_with(Op_Unlock)) {
1865 continue;
1866 }
1867
1868 Node* wb_proj = wb->find_out_with(Op_ShenandoahWBMemProj);
1869 fixer->remove(wb_proj);
1870 Node* mem_for_ctrl = fixer->find_mem(rep_ctrl, NULL);
1871
1872 if (wb->in(Memory) != mem) {
1873 disconnect_barrier_mem(wb, phase->igvn());
1874 phase->igvn().replace_input_of(wb, Memory, mem);
1875 }
1876 if (rep_ctrl != wb_ctrl) {
1877 phase->set_ctrl_and_loop(wb, rep_ctrl);
1878 phase->igvn().replace_input_of(wb, Control, rep_ctrl);
1879 phase->set_ctrl_and_loop(wb_proj, rep_ctrl);
1880 progress = true;
1881 }
1882
1883 fixer->fix_mem(rep_ctrl, rep_ctrl, mem, mem_for_ctrl, wb_proj, uses);
1884
1885 assert(!ShenandoahVerifyOptoBarriers || ShenandoahWriteBarrierNode::memory_dominates_all_paths(mem, rep_ctrl, alias, phase), "can't fix the memory graph");
1886 }
1887
1888 if (other->Opcode() == Op_ShenandoahWriteBarrier) {
1889 Node* other_proj = other->find_out_with(Op_ShenandoahWBMemProj);
1890 if (fixer != NULL) {
1891 fixer->remove(other_proj);
1892 }
1893 phase->igvn().replace_node(other_proj, other->in(Memory));
1894 }
1895 phase->igvn().replace_node(other, wb);
1896 --j; --jmax;
1897 }
1898 }
1899 }
1900 }
1901 } while(progress);
1902 }
1903
1904 void ShenandoahReadBarrierNode::try_move(Node *n_ctrl, PhaseIdealLoop* phase) {
1905 Node* mem = in(MemNode::Memory);
1906 int alias = phase->C->get_alias_index(adr_type());
1907 const bool trace = false;
1908
1909 #ifdef ASSERT
1910 if (trace) { tty->print("Trying to move mem of"); dump(); }
1911 #endif
1912
1913 Node* new_mem = mem;
1914
1915 ResourceMark rm;
1916 VectorSet seen(Thread::current()->resource_area());
1917 Node_List phis;
1918
1919 for (;;) {
1920 #ifdef ASSERT
1921 if (trace) { tty->print("Looking for dominator from"); mem->dump(); }
1922 #endif
1923 if (mem->is_Proj() && mem->in(0)->is_Start()) {
1924 if (new_mem != in(MemNode::Memory)) {
1925 #ifdef ASSERT
1926 if (trace) { tty->print("XXX Setting mem to"); new_mem->dump(); tty->print(" for "); dump(); }
1927 #endif
1928 phase->igvn().replace_input_of(this, MemNode::Memory, new_mem);
1929 }
1930 return;
1931 }
1932
1933 Node* candidate = mem;
1934 do {
1935 if (!is_independent(mem)) {
1936 if (trace) { tty->print_cr("Not independent"); }
1937 if (new_mem != in(MemNode::Memory)) {
1938 #ifdef ASSERT
1939 if (trace) { tty->print("XXX Setting mem to"); new_mem->dump(); tty->print(" for "); dump(); }
1940 #endif
1941 phase->igvn().replace_input_of(this, MemNode::Memory, new_mem);
1942 }
1943 return;
1944 }
1945 if (seen.test_set(mem->_idx)) {
1946 if (trace) { tty->print_cr("Already seen"); }
1947 ShouldNotReachHere();
1948 // Strange graph
1949 if (new_mem != in(MemNode::Memory)) {
1950 #ifdef ASSERT
1951 if (trace) { tty->print("XXX Setting mem to"); new_mem->dump(); tty->print(" for "); dump(); }
1952 #endif
1953 phase->igvn().replace_input_of(this, MemNode::Memory, new_mem);
1954 }
1955 return;
1956 }
1957 if (mem->is_Phi()) {
1958 phis.push(mem);
1959 }
1960 mem = next_mem(mem, alias);
1961 if (mem->bottom_type() == Type::MEMORY) {
1962 candidate = mem;
1963 }
1964 assert(is_dominator(phase->ctrl_or_self(mem), n_ctrl, mem, this, phase) == phase->is_dominator(phase->ctrl_or_self(mem), n_ctrl), "strange dominator");
1965 #ifdef ASSERT
1966 if (trace) { tty->print("Next mem is"); mem->dump(); }
1967 #endif
1968 } while (mem->bottom_type() != Type::MEMORY || !phase->is_dominator(phase->ctrl_or_self(mem), n_ctrl));
1969
1970 assert(mem->bottom_type() == Type::MEMORY, "bad mem");
1971
1972 bool not_dom = false;
1973 for (uint i = 0; i < phis.size() && !not_dom; i++) {
1974 Node* nn = phis.at(i);
1975
1976 #ifdef ASSERT
1977 if (trace) { tty->print("Looking from phi"); nn->dump(); }
1978 #endif
1979 assert(nn->is_Phi(), "phis only");
1980 for (uint j = 2; j < nn->req() && !not_dom; j++) {
1981 Node* m = nn->in(j);
1982 #ifdef ASSERT
1983 if (trace) { tty->print("Input %d is", j); m->dump(); }
1984 #endif
1985 while (m != mem && !seen.test_set(m->_idx)) {
1986 if (is_dominator(phase->ctrl_or_self(m), phase->ctrl_or_self(mem), m, mem, phase)) {
1987 not_dom = true;
1988 // Scheduling anomaly
1989 #ifdef ASSERT
1990 if (trace) { tty->print("Giving up"); m->dump(); }
1991 #endif
1992 break;
1993 }
1994 if (!is_independent(m)) {
1995 if (trace) { tty->print_cr("Not independent"); }
1996 if (new_mem != in(MemNode::Memory)) {
1997 #ifdef ASSERT
1998 if (trace) { tty->print("XXX Setting mem to"); new_mem->dump(); tty->print(" for "); dump(); }
1999 #endif
2000 phase->igvn().replace_input_of(this, MemNode::Memory, new_mem);
2001 }
2002 return;
2003 }
2004 if (m->is_Phi()) {
2005 phis.push(m);
2006 }
2007 m = next_mem(m, alias);
2008 #ifdef ASSERT
2009 if (trace) { tty->print("Next mem is"); m->dump(); }
2010 #endif
2011 }
2012 }
2013 }
2014 if (!not_dom) {
2015 new_mem = mem;
2016 phis.clear();
2017 } else {
2018 seen.Clear();
2019 }
2020 }
2021 }
2022
2023 CallStaticJavaNode* ShenandoahWriteBarrierNode::pin_and_expand_null_check(PhaseIterGVN& igvn) {
2024 Node* val = in(ValueIn);
2025
2026 const Type* val_t = igvn.type(val);
2027
2028 if (val_t->meet(TypePtr::NULL_PTR) != val_t &&
2029 val->Opcode() == Op_CastPP &&
2030 val->in(0) != NULL &&
2031 val->in(0)->Opcode() == Op_IfTrue &&
2032 val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
2033 val->in(0)->in(0)->is_If() &&
2034 val->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
2035 val->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
2036 val->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
2037 val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1) &&
2038 val->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
2039 assert(val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1), "");
2040 CallStaticJavaNode* unc = val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
2041 return unc;
2042 }
2043 return NULL;
2044 }
2045
2046 void ShenandoahWriteBarrierNode::pin_and_expand_move_barrier(PhaseIdealLoop* phase, GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, Unique_Node_List& uses) {
2047 Node* unc = pin_and_expand_null_check(phase->igvn());
2048 Node* val = in(ValueIn);
2049
2050 if (unc != NULL) {
2051 Node* ctrl = phase->get_ctrl(this);
2052 Node* unc_ctrl = val->in(0);
2053
2054 // Don't move write barrier in a loop
2055 IdealLoopTree* loop = phase->get_loop(ctrl);
2056 IdealLoopTree* unc_loop = phase->get_loop(unc_ctrl);
2057
2058 if (!unc_loop->is_member(loop)) {
2059 return;
2060 }
2061
2062 Node* branch = no_branches(ctrl, unc_ctrl, false, phase);
2063 assert(branch == NULL || branch == NodeSentinel, "was not looking for a branch");
2064 if (branch == NodeSentinel) {
2065 return;
2066 }
2067
2068 RegionNode* r = new RegionNode(3);
2069 IfNode* iff = unc_ctrl->in(0)->as_If();
2070
2071 Node* ctrl_use = unc_ctrl->unique_ctrl_out();
2072 Node* unc_ctrl_clone = unc_ctrl->clone();
2073 phase->register_control(unc_ctrl_clone, loop, iff);
2074 Node* c = unc_ctrl_clone;
2075 Node* new_cast = clone_null_check(c, val, unc_ctrl_clone, phase);
2076 r->init_req(1, new_cast->in(0)->in(0)->as_If()->proj_out(0));
2077
2078 phase->igvn().replace_input_of(unc_ctrl, 0, c->in(0));
2079 phase->set_idom(unc_ctrl, c->in(0), phase->dom_depth(unc_ctrl));
2080 phase->lazy_replace(c, unc_ctrl);
2081 c = NULL;;
2082 phase->igvn().replace_input_of(val, 0, unc_ctrl_clone);
2083 phase->set_ctrl(val, unc_ctrl_clone);
2084
2085 IfNode* new_iff = new_cast->in(0)->in(0)->as_If();
2086 fix_null_check(unc, unc_ctrl_clone, r, uses, phase);
2087 Node* iff_proj = iff->proj_out(0);
2088 r->init_req(2, iff_proj);
2089 phase->register_control(r, phase->ltree_root(), iff);
2090
2091 Node* new_bol = new_iff->in(1)->clone();
2092 Node* new_cmp = new_bol->in(1)->clone();
2093 assert(new_cmp->Opcode() == Op_CmpP, "broken");
2094 assert(new_cmp->in(1) == val->in(1), "broken");
2095 new_bol->set_req(1, new_cmp);
2096 new_cmp->set_req(1, this);
2097 phase->register_new_node(new_bol, new_iff->in(0));
2098 phase->register_new_node(new_cmp, new_iff->in(0));
2099 phase->igvn().replace_input_of(new_iff, 1, new_bol);
2100 phase->igvn().replace_input_of(new_cast, 1, this);
2101
2102 for (DUIterator_Fast imax, i = this->fast_outs(imax); i < imax; i++) {
2103 Node* u = this->fast_out(i);
2104 if (u == new_cast || u->Opcode() == Op_ShenandoahWBMemProj || u == new_cmp) {
2105 continue;
2106 }
2107 phase->igvn().rehash_node_delayed(u);
2108 int nb = u->replace_edge(this, new_cast);
2109 assert(nb > 0, "no update?");
2110 --i; imax -= nb;
2111 }
2112
2113 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
2114 Node* u = val->fast_out(i);
2115 if (u == this) {
2116 continue;
2117 }
2118 phase->igvn().rehash_node_delayed(u);
2119 int nb = u->replace_edge(val, new_cast);
2120 assert(nb > 0, "no update?");
2121 --i; imax -= nb;
2122 }
2123
2124 Node* new_ctrl = unc_ctrl_clone;
2125
2126 int alias = phase->C->get_alias_index(adr_type());
2127 MemoryGraphFixer* fixer = find_fixer(memory_graph_fixers, alias);
2128 if (fixer == NULL) {
2129 fixer = create_fixer(memory_graph_fixers, alias, phase, true);
2130 }
2131
2132 Node* proj = find_out_with(Op_ShenandoahWBMemProj);
2133 fixer->remove(proj);
2134 Node* mem = fixer->find_mem(new_ctrl, NULL);
2135
2136 if (in(Memory) != mem) {
2137 disconnect_barrier_mem(this, phase->igvn());
2138 phase->igvn().replace_input_of(this, Memory, mem);
2139 }
2140
2141 phase->set_ctrl_and_loop(this, new_ctrl);
2142 phase->igvn().replace_input_of(this, Control, new_ctrl);
2143 phase->set_ctrl_and_loop(proj, new_ctrl);
2144
2145 fixer->fix_mem(new_ctrl, new_ctrl, mem, mem, proj, uses);
2146 }
2147 }
2148
2149 void ShenandoahWriteBarrierNode::pin_and_expand_helper(PhaseIdealLoop* phase) {
2150 Node* val = in(ValueIn);
2151 CallStaticJavaNode* unc = pin_and_expand_null_check(phase->igvn());
2152 Node* rep = this;
2153 Node* ctrl = phase->get_ctrl(this);
2154 if (unc != NULL && val->in(0) == ctrl) {
2155 Node* unc_ctrl = val->in(0);
2156 IfNode* other_iff = unc_ctrl->unique_ctrl_out()->as_If();
2157 ProjNode* other_unc_ctrl = other_iff->proj_out(1);
2158 Node* cast = NULL;
2159 for (DUIterator_Fast imax, i = other_unc_ctrl->fast_outs(imax); i < imax && cast == NULL; i++) {
2160 Node* u = other_unc_ctrl->fast_out(i);
2161 if (u->Opcode() == Op_CastPP && u->in(1) == this) {
2162 cast = u;
2163 }
2164 }
2165 assert(other_unc_ctrl->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) == unc, "broken");
2166 rep = cast;
2167 }
2168
2169 // Replace all uses of barrier's input that are dominated by ctrl
2170 // with the value returned by the barrier: no need to keep both
2171 // live.
2172 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
2173 Node* u = val->fast_out(i);
2174 if (u != this) {
2175 if (u->is_Phi()) {
2176 int nb = 0;
2177 for (uint j = 1; j < u->req(); j++) {
2178 if (u->in(j) == val) {
2179 Node* c = u->in(0)->in(j);
2180 if (phase->is_dominator(ctrl, c)) {
2181 phase->igvn().replace_input_of(u, j, rep);
2182 nb++;
2183 }
2184 }
2185 }
2186 if (nb > 0) {
2187 imax -= nb;
2188 --i;
2189 }
2190 } else {
2191 Node* c = phase->ctrl_or_self(u);
2192 if (is_dominator(ctrl, c, this, u, phase)) {
2193 phase->igvn().rehash_node_delayed(u);
2194 int nb = u->replace_edge(val, rep);
2195 assert(nb > 0, "no update?");
2196 --i, imax -= nb;
2197 }
2198 }
2199 }
2200 }
2201 }
2202
2203 Node* ShenandoahWriteBarrierNode::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) {
2204 Node* mem = NULL;
2205 Node* c = ctrl;
2206 do {
2207 if (c->is_Region()) {
2208 Node* phi_bottom = NULL;
2209 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == NULL; i++) {
2210 Node* u = c->fast_out(i);
2211 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
2212 if (u->adr_type() == TypePtr::BOTTOM) {
2213 mem = u;
2214 }
2215 }
2216 }
2217 } else {
2218 if (c->is_Call() && c->as_Call()->adr_type() != NULL) {
2219 CallProjections projs;
2220 c->as_Call()->extract_projections(&projs, true, false);
2221 if (projs.fallthrough_memproj != NULL) {
2222 if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
2223 if (projs.catchall_memproj == NULL) {
2224 mem = projs.fallthrough_memproj;
2225 } else {
2226 if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) {
2227 mem = projs.fallthrough_memproj;
2228 } else {
2229 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier");
2230 mem = projs.catchall_memproj;
2231 }
2232 }
2233 }
2234 } else {
2235 Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
2236 if (proj != NULL &&
2237 proj->adr_type() == TypePtr::BOTTOM) {
2238 mem = proj;
2239 }
2240 }
2241 } else {
2242 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
2243 Node* u = c->fast_out(i);
2244 if (u->is_Proj() &&
2245 u->bottom_type() == Type::MEMORY &&
2246 u->adr_type() == TypePtr::BOTTOM) {
2247 assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
2248 assert(mem == NULL, "only one proj");
2249 mem = u;
2250 }
2251 }
2252 assert(!c->is_Call() || c->as_Call()->adr_type() != NULL || mem == NULL, "no mem projection expected");
2253 }
2254 }
2255 c = phase->idom(c);
2256 } while (mem == NULL);
2257 return mem;
2258 }
2259
2260 void ShenandoahWriteBarrierNode::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
2261 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2262 Node* u = n->fast_out(i);
2263 if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
2264 uses.push(u);
2265 }
2266 }
2267 }
2268
2269 void ShenandoahWriteBarrierNode::test_heap_stable(Node*& ctrl, Node* raw_mem, Node*& heap_stable_ctrl,
2270 PhaseIdealLoop* phase) {
2271 IdealLoopTree* loop = phase->get_loop(ctrl);
2272 Node* thread = new ThreadLocalNode();
2273 phase->register_new_node(thread, ctrl);
2274 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
2275 phase->set_ctrl(offset, phase->C->root());
2276 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, offset);
2277 phase->register_new_node(gc_state_addr, ctrl);
2278 uint gc_state_idx = Compile::AliasIdxRaw;
2279 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument
2280 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
2281
2282 Node* gc_state = new LoadBNode(ctrl, raw_mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered);
2283 phase->register_new_node(gc_state, ctrl);
2284 Node* heap_stable_and = new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::HAS_FORWARDED));
2285 phase->register_new_node(heap_stable_and, ctrl);
2286 Node* heap_stable_cmp = new CmpINode(heap_stable_and, phase->igvn().zerocon(T_INT));
2287 phase->register_new_node(heap_stable_cmp, ctrl);
2288 Node* heap_stable_test = new BoolNode(heap_stable_cmp, BoolTest::ne);
2289 phase->register_new_node(heap_stable_test, ctrl);
2290 IfNode* heap_stable_iff = new IfNode(ctrl, heap_stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
2291 phase->register_control(heap_stable_iff, loop, ctrl);
2292
2293 heap_stable_ctrl = new IfFalseNode(heap_stable_iff);
2294 phase->register_control(heap_stable_ctrl, loop, heap_stable_iff);
2295 ctrl = new IfTrueNode(heap_stable_iff);
2296 phase->register_control(ctrl, loop, heap_stable_iff);
2297
2298 assert(is_heap_stable_test(heap_stable_iff), "Should match the shape");
2299 }
2300
2301 void ShenandoahWriteBarrierNode::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) {
2302 const Type* val_t = phase->igvn().type(val);
2303 if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
2304 IdealLoopTree* loop = phase->get_loop(ctrl);
2305 Node* null_cmp = new CmpPNode(val, phase->igvn().zerocon(T_OBJECT));
2306 phase->register_new_node(null_cmp, ctrl);
2307 Node* null_test = new BoolNode(null_cmp, BoolTest::ne);
2308 phase->register_new_node(null_test, ctrl);
2309 IfNode* null_iff = new IfNode(ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
2310 phase->register_control(null_iff, loop, ctrl);
2311 ctrl = new IfTrueNode(null_iff);
2312 phase->register_control(ctrl, loop, null_iff);
2313 null_ctrl = new IfFalseNode(null_iff);
2314 phase->register_control(null_ctrl, loop, null_iff);
2315 }
2316 }
2317
2318 Node* ShenandoahWriteBarrierNode::clone_null_check(Node*& c, Node* val, Node* unc_ctrl, PhaseIdealLoop* phase) {
2319 IdealLoopTree *loop = phase->get_loop(c);
2320 Node* iff = unc_ctrl->in(0);
2321 assert(iff->is_If(), "broken");
2322 Node* new_iff = iff->clone();
2323 new_iff->set_req(0, c);
2324 phase->register_control(new_iff, loop, c);
2325 Node* iffalse = new IfFalseNode(new_iff->as_If());
2326 phase->register_control(iffalse, loop, new_iff);
2327 Node* iftrue = new IfTrueNode(new_iff->as_If());
2328 phase->register_control(iftrue, loop, new_iff);
2329 c = iftrue;
2330 const Type *t = phase->igvn().type(val);
2331 assert(val->Opcode() == Op_CastPP, "expect cast to non null here");
2332 Node* uncasted_val = val->in(1);
2333 val = new CastPPNode(uncasted_val, t);
2334 val->init_req(0, c);
2335 phase->register_new_node(val, c);
2336 return val;
2337 }
2338
2339 void ShenandoahWriteBarrierNode::fix_null_check(Node* unc, Node* unc_ctrl, Node* new_unc_ctrl,
2340 Unique_Node_List& uses, PhaseIdealLoop* phase) {
2341 IfNode* iff = unc_ctrl->in(0)->as_If();
2342 Node* proj = iff->proj_out(0);
2343 assert(proj != unc_ctrl, "bad projection");
2344 Node* use = proj->unique_ctrl_out();
2345
2346 assert(use == unc || use->is_Region(), "what else?");
2347
2348 uses.clear();
2349 if (use == unc) {
2350 phase->set_idom(use, new_unc_ctrl, phase->dom_depth(use));
2351 for (uint i = 1; i < unc->req(); i++) {
2352 Node* n = unc->in(i);
2353 if (phase->has_ctrl(n) && phase->get_ctrl(n) == proj) {
2354 uses.push(n);
2355 }
2356 }
2357 } else {
2358 assert(use->is_Region(), "what else?");
2359 uint idx = 1;
2360 for (; use->in(idx) != proj; idx++);
2361 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2362 Node* u = use->fast_out(i);
2363 if (u->is_Phi() && phase->get_ctrl(u->in(idx)) == proj) {
2364 uses.push(u->in(idx));
2365 }
2366 }
2367 }
2368 for(uint next = 0; next < uses.size(); next++ ) {
2369 Node *n = uses.at(next);
2370 assert(phase->get_ctrl(n) == proj, "bad control");
2371 phase->set_ctrl_and_loop(n, new_unc_ctrl);
2372 if (n->in(0) == proj) {
2373 phase->igvn().replace_input_of(n, 0, new_unc_ctrl);
2374 }
2375 for (uint i = 0; i < n->req(); i++) {
2376 Node* m = n->in(i);
2377 if (m != NULL && phase->has_ctrl(m) && phase->get_ctrl(m) == proj) {
2378 uses.push(m);
2379 }
2380 }
2381 }
2382
2383 phase->igvn().rehash_node_delayed(use);
2384 int nb = use->replace_edge(proj, new_unc_ctrl);
2385 assert(nb == 1, "only use expected");
2386 }
2387
2388 void ShenandoahWriteBarrierNode::in_cset_fast_test(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) {
2389 IdealLoopTree *loop = phase->get_loop(ctrl);
2390 Node* raw_rbtrue = new CastP2XNode(ctrl, val);
2391 phase->register_new_node(raw_rbtrue, ctrl);
2392 Node* cset_offset = new URShiftXNode(raw_rbtrue, phase->igvn().intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
2393 phase->register_new_node(cset_offset, ctrl);
2394 Node* in_cset_fast_test_base_addr = phase->igvn().makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
2395 phase->set_ctrl(in_cset_fast_test_base_addr, phase->C->root());
2396 Node* in_cset_fast_test_adr = new AddPNode(phase->C->top(), in_cset_fast_test_base_addr, cset_offset);
2397 phase->register_new_node(in_cset_fast_test_adr, ctrl);
2398 uint in_cset_fast_test_idx = Compile::AliasIdxRaw;
2399 const TypePtr* in_cset_fast_test_adr_type = NULL; // debug-mode-only argument
2400 debug_only(in_cset_fast_test_adr_type = phase->C->get_adr_type(in_cset_fast_test_idx));
2401 Node* in_cset_fast_test_load = new LoadBNode(ctrl, raw_mem, in_cset_fast_test_adr, in_cset_fast_test_adr_type, TypeInt::BYTE, MemNode::unordered);
2402 phase->register_new_node(in_cset_fast_test_load, ctrl);
2403 Node* in_cset_fast_test_cmp = new CmpINode(in_cset_fast_test_load, phase->igvn().zerocon(T_INT));
2404 phase->register_new_node(in_cset_fast_test_cmp, ctrl);
2405 Node* in_cset_fast_test_test = new BoolNode(in_cset_fast_test_cmp, BoolTest::eq);
2406 phase->register_new_node(in_cset_fast_test_test, ctrl);
2407 IfNode* in_cset_fast_test_iff = new IfNode(ctrl, in_cset_fast_test_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
2408 phase->register_control(in_cset_fast_test_iff, loop, ctrl);
2409
2410 not_cset_ctrl = new IfTrueNode(in_cset_fast_test_iff);
2411 phase->register_control(not_cset_ctrl, loop, in_cset_fast_test_iff);
2412
2413 ctrl = new IfFalseNode(in_cset_fast_test_iff);
2414 phase->register_control(ctrl, loop, in_cset_fast_test_iff);
2415 }
2416
2417 void ShenandoahWriteBarrierNode::call_wb_stub(Node*& ctrl, Node*& val, Node*& result_mem,
2418 Node* raw_mem, Node* wb_mem,
2419 int alias,
2420 PhaseIdealLoop* phase) {
2421 IdealLoopTree*loop = phase->get_loop(ctrl);
2422 const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr()->cast_to_nonconst();
2423
2424 // The slow path stub consumes and produces raw memory in addition
2425 // to the existing memory edges
2426 Node* base = find_bottom_mem(ctrl, phase);
2427
2428 MergeMemNode* mm = MergeMemNode::make(base);
2429 mm->set_memory_at(alias, wb_mem);
2430 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
2431 phase->register_new_node(mm, ctrl);
2432
2433 Node* call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_write_barrier_Type(), CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_JRT), "shenandoah_write_barrier", TypeRawPtr::BOTTOM);
2434 call->init_req(TypeFunc::Control, ctrl);
2435 call->init_req(TypeFunc::I_O, phase->C->top());
2436 call->init_req(TypeFunc::Memory, mm);
2437 call->init_req(TypeFunc::FramePtr, phase->C->top());
2438 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
2439 call->init_req(TypeFunc::Parms, val);
2440 phase->register_control(call, loop, ctrl);
2441 ctrl = new ProjNode(call, TypeFunc::Control);
2442 phase->register_control(ctrl, loop, call);
2443 result_mem = new ProjNode(call, TypeFunc::Memory);
2444 phase->register_new_node(result_mem, call);
2445 val = new ProjNode(call, TypeFunc::Parms);
2446 phase->register_new_node(val, call);
2447 val = new CheckCastPPNode(ctrl, val, obj_type);
2448 phase->register_new_node(val, ctrl);
2449 }
2450
2451 void ShenandoahWriteBarrierNode::fix_ctrl(Node* barrier, Node* region, const MemoryGraphFixer& fixer, Unique_Node_List& uses, Unique_Node_List& uses_to_ignore, uint last, PhaseIdealLoop* phase) {
2452 Node* ctrl = phase->get_ctrl(barrier);
2453 Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
2454
2455 // Update the control of all nodes that should be after the
2456 // barrier control flow
2457 uses.clear();
2458 // Every node that is control dependent on the barrier's input
2459 // control will be after the expanded barrier. The raw memory (if
2460 // its memory is control dependent on the barrier's input control)
2461 // must stay above the barrier.
2462 uses_to_ignore.clear();
2463 if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
2464 uses_to_ignore.push(init_raw_mem);
2465 }
2466 for (uint next = 0; next < uses_to_ignore.size(); next++) {
2467 Node *n = uses_to_ignore.at(next);
2468 for (uint i = 0; i < n->req(); i++) {
2469 Node* in = n->in(i);
2470 if (in != NULL && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
2471 uses_to_ignore.push(in);
2472 }
2473 }
2474 }
2475 for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
2476 Node* u = ctrl->fast_out(i);
2477 if (u->_idx < last &&
2478 u != barrier &&
2479 !uses_to_ignore.member(u) &&
2480 (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) &&
2481 (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) {
2482 Node* old_c = phase->ctrl_or_self(u);
2483 Node* c = old_c;
2484 if (c != ctrl ||
2485 is_dominator_same_ctrl(old_c, barrier, u, phase) ||
2486 ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) {
2487 phase->igvn().rehash_node_delayed(u);
2488 int nb = u->replace_edge(ctrl, region);
2489 if (u->is_CFG()) {
2490 if (phase->idom(u) == ctrl) {
2491 phase->set_idom(u, region, phase->dom_depth(region));
2492 }
2493 } else if (phase->get_ctrl(u) == ctrl) {
2494 assert(u != init_raw_mem, "should leave input raw mem above the barrier");
2495 uses.push(u);
2496 }
2497 assert(nb == 1, "more than 1 ctrl input?");
2498 --i, imax -= nb;
2499 }
2500 }
2501 }
2502 }
2503
2504 void ShenandoahWriteBarrierNode::pin_and_expand(PhaseIdealLoop* phase) {
2505 Node_List enqueue_barriers;
2506 if (ShenandoahStoreValEnqueueBarrier) {
2507 Unique_Node_List wq;
2508 wq.push(phase->C->root());
2509 for (uint i = 0; i < wq.size(); i++) {
2510 Node* n = wq.at(i);
2511 if (n->Opcode() == Op_ShenandoahEnqueueBarrier) {
2512 enqueue_barriers.push(n);
2513 }
2514 for (uint i = 0; i < n->req(); i++) {
2515 Node* in = n->in(i);
2516 if (in != NULL) {
2517 wq.push(in);
2518 }
2519 }
2520 }
2521 }
2522
2523 const bool trace = false;
2524
2525 // Collect raw memory state at CFG points in the entire graph and
2526 // record it in memory_nodes. Optimize the raw memory graph in the
2527 // process. Optimizing the memory graph also makes the memory graph
2528 // simpler.
2529 GrowableArray<MemoryGraphFixer*> memory_graph_fixers;
2530
2531 // Let's try to common write barriers again
2532 optimize_before_expansion(phase, memory_graph_fixers, true);
2533
2534 Unique_Node_List uses;
2535 for (int i = 0; i < ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count(); i++) {
2536 ShenandoahWriteBarrierNode* wb = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barrier(i);
2537 Node* ctrl = phase->get_ctrl(wb);
2538
2539 Node* val = wb->in(ValueIn);
2540 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) {
2541 assert(is_dominator(phase->get_ctrl(val), ctrl->in(0)->in(0), val, ctrl->in(0), phase), "can't move");
2542 phase->set_ctrl(wb, ctrl->in(0)->in(0));
2543 } else if (ctrl->is_CallRuntime()) {
2544 assert(is_dominator(phase->get_ctrl(val), ctrl->in(0), val, ctrl, phase), "can't move");
2545 phase->set_ctrl(wb, ctrl->in(0));
2546 }
2547
2548 assert(wb->Opcode() == Op_ShenandoahWriteBarrier, "only for write barriers");
2549 // Look for a null check that dominates this barrier and move the
2550 // barrier right after the null check to enable implicit null
2551 // checks
2552 wb->pin_and_expand_move_barrier(phase, memory_graph_fixers, uses);
2553
2554 wb->pin_and_expand_helper(phase);
2555 }
2556
2557 MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
2558 Unique_Node_List uses_to_ignore;
2559 for (uint i = 0; i < enqueue_barriers.size(); i++) {
2560 Node* barrier = enqueue_barriers.at(i);
2561 Node* pre_val = barrier->in(1);
2562
2563 if (phase->igvn().type(pre_val)->higher_equal(TypePtr::NULL_PTR)) {
2564 ShouldNotReachHere();
2565 continue;
2566 }
2567
2568 Node* ctrl = phase->get_ctrl(barrier);
2569
2570 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) {
2571 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0)->in(0), pre_val, ctrl->in(0), phase), "can't move");
2572 ctrl = ctrl->in(0)->in(0);
2573 phase->set_ctrl(barrier, ctrl);
2574 } else if (ctrl->is_CallRuntime()) {
2575 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0), pre_val, ctrl, phase), "can't move");
2576 ctrl = ctrl->in(0);
2577 phase->set_ctrl(barrier, ctrl);
2578 }
2579
2580 Node* init_ctrl = ctrl;
2581 IdealLoopTree* loop = phase->get_loop(ctrl);
2582 Node* raw_mem = fixer.find_mem(ctrl, barrier);
2583 Node* init_raw_mem = raw_mem;
2584 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL);
2585 Node* heap_stable_ctrl = NULL;
2586 Node* null_ctrl = NULL;
2587 uint last = phase->C->unique();
2588
2589 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
2590 Node* region = new RegionNode(PATH_LIMIT);
2591 Node* phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
2592
2593 enum { _fast_path = 1, _slow_path, _null_path, PATH_LIMIT2 };
2594 Node* region2 = new RegionNode(PATH_LIMIT2);
2595 Node* phi2 = PhiNode::make(region2, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
2596
2597 // Stable path.
2598 test_heap_stable(ctrl, raw_mem, heap_stable_ctrl, phase);
2599 region->init_req(_heap_stable, heap_stable_ctrl);
2600 phi->init_req(_heap_stable, raw_mem);
2601
2602 // Null path
2603 Node* reg2_ctrl = NULL;
2604 test_null(ctrl, pre_val, null_ctrl, phase);
2605 if (null_ctrl != NULL) {
2606 reg2_ctrl = null_ctrl->in(0);
2607 region2->init_req(_null_path, null_ctrl);
2608 phi2->init_req(_null_path, raw_mem);
2609 } else {
2610 region2->del_req(_null_path);
2611 phi2->del_req(_null_path);
2612 }
2613
2614 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
2615 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
2616 Node* thread = new ThreadLocalNode();
2617 phase->register_new_node(thread, ctrl);
2618 Node* buffer_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(buffer_offset));
2619 phase->register_new_node(buffer_adr, ctrl);
2620 Node* index_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(index_offset));
2621 phase->register_new_node(index_adr, ctrl);
2622
2623 BasicType index_bt = TypeX_X->basic_type();
2624 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
2625 const TypePtr* adr_type = TypeRawPtr::BOTTOM;
2626 Node* index = new LoadXNode(ctrl, raw_mem, index_adr, adr_type, TypeX_X, MemNode::unordered);
2627 phase->register_new_node(index, ctrl);
2628 Node* index_cmp = new CmpXNode(index, phase->igvn().MakeConX(0));
2629 phase->register_new_node(index_cmp, ctrl);
2630 Node* index_test = new BoolNode(index_cmp, BoolTest::ne);
2631 phase->register_new_node(index_test, ctrl);
2632 IfNode* queue_full_iff = new IfNode(ctrl, index_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
2633 if (reg2_ctrl == NULL) reg2_ctrl = queue_full_iff;
2634 phase->register_control(queue_full_iff, loop, ctrl);
2635 Node* not_full = new IfTrueNode(queue_full_iff);
2636 phase->register_control(not_full, loop, queue_full_iff);
2637 Node* full = new IfFalseNode(queue_full_iff);
2638 phase->register_control(full, loop, queue_full_iff);
2639
2640 ctrl = not_full;
2641
2642 Node* next_index = new SubXNode(index, phase->igvn().MakeConX(sizeof(intptr_t)));
2643 phase->register_new_node(next_index, ctrl);
2644
2645 Node* buffer = new LoadPNode(ctrl, raw_mem, buffer_adr, adr_type, TypeRawPtr::NOTNULL, MemNode::unordered);
2646 phase->register_new_node(buffer, ctrl);
2647 Node *log_addr = new AddPNode(phase->C->top(), buffer, next_index);
2648 phase->register_new_node(log_addr, ctrl);
2649 Node* log_store = new StorePNode(ctrl, raw_mem, log_addr, adr_type, pre_val, MemNode::unordered);
2650 phase->register_new_node(log_store, ctrl);
2651 // update the index
2652 Node* index_update = new StoreXNode(ctrl, log_store, index_adr, adr_type, next_index, MemNode::unordered);
2653 phase->register_new_node(index_update, ctrl);
2654
2655 // Fast-path case
2656 region2->init_req(_fast_path, ctrl);
2657 phi2->init_req(_fast_path, index_update);
2658
2659 ctrl = full;
2660
2661 Node* base = find_bottom_mem(ctrl, phase);
2662
2663 MergeMemNode* mm = MergeMemNode::make(base);
2664 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
2665 phase->register_new_node(mm, ctrl);
2666
2667 Node* call = new CallLeafNode(ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type(), CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", TypeRawPtr::BOTTOM);
2668 call->init_req(TypeFunc::Control, ctrl);
2669 call->init_req(TypeFunc::I_O, phase->C->top());
2670 call->init_req(TypeFunc::Memory, mm);
2671 call->init_req(TypeFunc::FramePtr, phase->C->top());
2672 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
2673 call->init_req(TypeFunc::Parms, pre_val);
2674 call->init_req(TypeFunc::Parms+1, thread);
2675 phase->register_control(call, loop, ctrl);
2676
2677 Node* ctrl_proj = new ProjNode(call, TypeFunc::Control);
2678 phase->register_control(ctrl_proj, loop, call);
2679 Node* mem_proj = new ProjNode(call, TypeFunc::Memory);
2680 phase->register_new_node(mem_proj, call);
2681
2682 // Slow-path case
2683 region2->init_req(_slow_path, ctrl_proj);
2684 phi2->init_req(_slow_path, mem_proj);
2685
2686 phase->register_control(region2, loop, reg2_ctrl);
2687 phase->register_new_node(phi2, region2);
2688
2689 region->init_req(_heap_unstable, region2);
2690 phi->init_req(_heap_unstable, phi2);
2691
2692 phase->register_control(region, loop, heap_stable_ctrl->in(0));
2693 phase->register_new_node(phi, region);
2694
2695 fix_ctrl(barrier, region, fixer, uses, uses_to_ignore, last, phase);
2696 for(uint next = 0; next < uses.size(); next++ ) {
2697 Node *n = uses.at(next);
2698 assert(phase->get_ctrl(n) == init_ctrl, "bad control");
2699 assert(n != init_raw_mem, "should leave input raw mem above the barrier");
2700 phase->set_ctrl(n, region);
2701 follow_barrier_uses(n, init_ctrl, uses, phase);
2702 }
2703 fixer.fix_mem(init_ctrl, region, init_raw_mem, raw_mem_for_ctrl, phi, uses);
2704
2705 phase->igvn().replace_node(barrier, pre_val);
2706 }
2707
2708 for (int i = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count(); i > 0; i--) {
2709 int cnt = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count();
2710 ShenandoahWriteBarrierNode* wb = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barrier(i-1);
2711
2712 uint last = phase->C->unique();
2713 Node* ctrl = phase->get_ctrl(wb);
2714 Node* orig_ctrl = ctrl;
2715
2716 Node* raw_mem = fixer.find_mem(ctrl, wb);
2717 Node* init_raw_mem = raw_mem;
2718 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL);
2719 int alias = phase->C->get_alias_index(wb->adr_type());
2720 Node* wb_mem = wb->in(Memory);
2721 Node* init_wb_mem = wb_mem;
2722
2723 Node* val = wb->in(ValueIn);
2724 Node* wbproj = wb->find_out_with(Op_ShenandoahWBMemProj);
2725 IdealLoopTree *loop = phase->get_loop(ctrl);
2726
2727 assert(val->Opcode() != Op_ShenandoahWriteBarrier, "No chain of write barriers");
2728
2729 CallStaticJavaNode* unc = wb->pin_and_expand_null_check(phase->igvn());
2730 Node* unc_ctrl = NULL;
2731 if (unc != NULL) {
2732 if (val->in(0) != ctrl) {
2733 unc = NULL;
2734 } else {
2735 unc_ctrl = val->in(0);
2736 }
2737 }
2738
2739 Node* uncasted_val = val;
2740 if (unc != NULL) {
2741 uncasted_val = val->in(1);
2742 }
2743
2744 Node* heap_stable_ctrl = NULL;
2745 Node* null_ctrl = NULL;
2746
2747 assert(val->bottom_type()->make_oopptr(), "need oop");
2748 assert(val->bottom_type()->make_oopptr()->const_oop() == NULL, "expect non-constant");
2749
2750 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
2751 Node* region = new RegionNode(PATH_LIMIT);
2752 Node* val_phi = new PhiNode(region, uncasted_val->bottom_type()->is_oopptr());
2753 Node* mem_phi = PhiNode::make(region, wb_mem, Type::MEMORY, phase->C->alias_type(wb->adr_type())->adr_type());
2754 Node* raw_mem_phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
2755
2756 enum { _not_cset = 1, _not_equal, _evac_path, _null_path, PATH_LIMIT2 };
2757 Node* region2 = new RegionNode(PATH_LIMIT2);
2758 Node* val_phi2 = new PhiNode(region2, uncasted_val->bottom_type()->is_oopptr());
2759 Node* mem_phi2 = PhiNode::make(region2, wb_mem, Type::MEMORY, phase->C->alias_type(wb->adr_type())->adr_type());
2760 Node* raw_mem_phi2 = PhiNode::make(region2, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
2761
2762 // Stable path.
2763 test_heap_stable(ctrl, raw_mem, heap_stable_ctrl, phase);
2764 IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
2765
2766 // Heap stable case
2767 region->init_req(_heap_stable, heap_stable_ctrl);
2768 val_phi->init_req(_heap_stable, uncasted_val);
2769 mem_phi->init_req(_heap_stable, wb_mem);
2770 raw_mem_phi->init_req(_heap_stable, raw_mem);
2771
2772 Node* reg2_ctrl = NULL;
2773 // Null case
2774 test_null(ctrl, val, null_ctrl, phase);
2775 if (null_ctrl != NULL) {
2776 reg2_ctrl = null_ctrl->in(0);
2777 region2->init_req(_null_path, null_ctrl);
2778 val_phi2->init_req(_null_path, uncasted_val);
2779 mem_phi2->init_req(_null_path, wb_mem);
2780 raw_mem_phi2->init_req(_null_path, raw_mem);
2781 } else {
2782 region2->del_req(_null_path);
2783 val_phi2->del_req(_null_path);
2784 mem_phi2->del_req(_null_path);
2785 raw_mem_phi2->del_req(_null_path);
2786 }
2787
2788 // Test for in-cset.
2789 // Wires !in_cset(obj) to slot 2 of region and phis
2790 Node* not_cset_ctrl = NULL;
2791 in_cset_fast_test(ctrl, not_cset_ctrl, uncasted_val, raw_mem, phase);
2792 if (not_cset_ctrl != NULL) {
2793 if (reg2_ctrl == NULL) reg2_ctrl = not_cset_ctrl->in(0);
2794 region2->init_req(_not_cset, not_cset_ctrl);
2795 val_phi2->init_req(_not_cset, uncasted_val);
2796 mem_phi2->init_req(_not_cset, wb_mem);
2797 raw_mem_phi2->init_req(_not_cset, raw_mem);
2798 }
2799
2800 // Resolve object when orig-value is in cset.
2801 // Make the unconditional resolve for fwdptr, not the read barrier.
2802 Node* new_val = uncasted_val;
2803 if (unc_ctrl != NULL) {
2804 // Clone the null check in this branch to allow implicit null check
2805 new_val = clone_null_check(ctrl, val, unc_ctrl, phase);
2806 fix_null_check(unc, unc_ctrl, ctrl->in(0)->as_If()->proj_out(0), uses, phase);
2807
2808 IfNode* iff = unc_ctrl->in(0)->as_If();
2809 phase->igvn().replace_input_of(iff, 1, phase->igvn().intcon(1));
2810 }
2811 Node* addr = new AddPNode(new_val, uncasted_val, phase->igvn().MakeConX(ShenandoahBrooksPointer::byte_offset()));
2812 phase->register_new_node(addr, ctrl);
2813 assert(val->bottom_type()->isa_oopptr(), "what else?");
2814 const TypePtr* obj_type = val->bottom_type()->is_oopptr();
2815 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(obj_type);
2816 Node* fwd = new LoadPNode(ctrl, wb_mem, addr, adr_type, obj_type, MemNode::unordered);
2817 phase->register_new_node(fwd, ctrl);
2818
2819 // Only branch to WB stub if object is not forwarded; otherwise reply with fwd ptr
2820 Node* cmp = new CmpPNode(fwd, new_val);
2821 phase->register_new_node(cmp, ctrl);
2822 Node* bol = new BoolNode(cmp, BoolTest::eq);
2823 phase->register_new_node(bol, ctrl);
2824
2825 IfNode* iff = new IfNode(ctrl, bol, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
2826 if (reg2_ctrl == NULL) reg2_ctrl = iff;
2827 phase->register_control(iff, loop, ctrl);
2828 Node* if_not_eq = new IfFalseNode(iff);
2829 phase->register_control(if_not_eq, loop, iff);
2830 Node* if_eq = new IfTrueNode(iff);
2831 phase->register_control(if_eq, loop, iff);
2832
2833 // Wire up not-equal-path in slots 3.
2834 region2->init_req(_not_equal, if_not_eq);
2835 val_phi2->init_req(_not_equal, fwd);
2836 mem_phi2->init_req(_not_equal, wb_mem);
2837 raw_mem_phi2->init_req(_not_equal, raw_mem);
2838
2839 // Call wb-stub and wire up that path in slots 4
2840 Node* result_mem = NULL;
2841 ctrl = if_eq;
2842 call_wb_stub(ctrl, new_val, result_mem,
2843 raw_mem, wb_mem,
2844 alias, phase);
2845 region2->init_req(_evac_path, ctrl);
2846 val_phi2->init_req(_evac_path, new_val);
2847 mem_phi2->init_req(_evac_path, result_mem);
2848 raw_mem_phi2->init_req(_evac_path, result_mem);
2849
2850 phase->register_control(region2, loop, reg2_ctrl);
2851 phase->register_new_node(val_phi2, region2);
2852 phase->register_new_node(mem_phi2, region2);
2853 phase->register_new_node(raw_mem_phi2, region2);
2854
2855 region->init_req(_heap_unstable, region2);
2856 val_phi->init_req(_heap_unstable, val_phi2);
2857 mem_phi->init_req(_heap_unstable, mem_phi2);
2858 raw_mem_phi->init_req(_heap_unstable, raw_mem_phi2);
2859
2860 phase->register_control(region, loop, heap_stable_iff);
2861 Node* out_val = val_phi;
2862 phase->register_new_node(val_phi, region);
2863 phase->register_new_node(mem_phi, region);
2864 phase->register_new_node(raw_mem_phi, region);
2865
2866 fix_ctrl(wb, region, fixer, uses, uses_to_ignore, last, phase);
2867
2868 ctrl = orig_ctrl;
2869
2870 phase->igvn().replace_input_of(wbproj, ShenandoahWBMemProjNode::WriteBarrier, phase->C->top());
2871 phase->igvn().replace_node(wbproj, mem_phi);
2872 if (unc != NULL) {
2873 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
2874 Node* u = val->fast_out(i);
2875 Node* c = phase->ctrl_or_self(u);
2876 if (u != wb && (c != ctrl || is_dominator_same_ctrl(c, wb, u, phase))) {
2877 phase->igvn().rehash_node_delayed(u);
2878 int nb = u->replace_edge(val, out_val);
2879 --i, imax -= nb;
2880 }
2881 }
2882 if (val->outcnt() == 0) {
2883 phase->igvn()._worklist.push(val);
2884 }
2885 }
2886 phase->igvn().replace_node(wb, out_val);
2887
2888 follow_barrier_uses(mem_phi, ctrl, uses, phase);
2889 follow_barrier_uses(out_val, ctrl, uses, phase);
2890
2891 for(uint next = 0; next < uses.size(); next++ ) {
2892 Node *n = uses.at(next);
2893 assert(phase->get_ctrl(n) == ctrl, "bad control");
2894 assert(n != init_raw_mem, "should leave input raw mem above the barrier");
2895 phase->set_ctrl(n, region);
2896 follow_barrier_uses(n, ctrl, uses, phase);
2897 }
2898
2899 // The slow path call produces memory: hook the raw memory phi
2900 // from the expanded write barrier with the rest of the graph
2901 // which may require adding memory phis at every post dominated
2902 // region and at enclosing loop heads. Use the memory state
2903 // collected in memory_nodes to fix the memory graph. Update that
2904 // memory state as we go.
2905 fixer.fix_mem(ctrl, region, init_raw_mem, raw_mem_for_ctrl, raw_mem_phi, uses);
2906 assert(ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count() == cnt - 1, "not replaced");
2907 }
2908
2909 assert(ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count() == 0, "all write barrier nodes should have been replaced");
2910 }
2911
2912 void ShenandoahWriteBarrierNode::move_heap_stable_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
2913 IdealLoopTree *loop = phase->get_loop(iff);
2914 Node* loop_head = loop->_head;
2915 Node* entry_c = loop_head->in(LoopNode::EntryControl);
2916
2917 Node* bol = iff->in(1);
2918 Node* cmp = bol->in(1);
2919 Node* andi = cmp->in(1);
2920 Node* load = andi->in(1);
2921
2922 assert(is_gc_state_load(load), "broken");
2923 if (!phase->is_dominator(load->in(0), entry_c)) {
2924 Node* mem_ctrl = NULL;
2925 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
2926 load = load->clone();
2927 load->set_req(MemNode::Memory, mem);
2928 load->set_req(0, entry_c);
2929 phase->register_new_node(load, entry_c);
2930 andi = andi->clone();
2931 andi->set_req(1, load);
2932 phase->register_new_node(andi, entry_c);
2933 cmp = cmp->clone();
2934 cmp->set_req(1, andi);
2935 phase->register_new_node(cmp, entry_c);
2936 bol = bol->clone();
2937 bol->set_req(1, cmp);
2938 phase->register_new_node(bol, entry_c);
2939
2940 Node* old_bol =iff->in(1);
2941 phase->igvn().replace_input_of(iff, 1, bol);
2942 }
2943 }
2944
2945 bool ShenandoahWriteBarrierNode::identical_backtoback_ifs(Node *n, PhaseIdealLoop* phase) {
2946 if (!n->is_If() || n->is_CountedLoopEnd()) {
2947 return false;
2948 }
2949 Node* region = n->in(0);
2950
2951 if (!region->is_Region()) {
2952 return false;
2953 }
2954 Node* dom = phase->idom(region);
2955 if (!dom->is_If()) {
2956 return false;
2957 }
2958
2959 if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) {
2960 return false;
2961 }
2962
2963 IfNode* dom_if = dom->as_If();
2964 Node* proj_true = dom_if->proj_out(1);
2965 Node* proj_false = dom_if->proj_out(0);
2966
2967 for (uint i = 1; i < region->req(); i++) {
2968 if (phase->is_dominator(proj_true, region->in(i))) {
2969 continue;
2970 }
2971 if (phase->is_dominator(proj_false, region->in(i))) {
2972 continue;
2973 }
2974 return false;
2975 }
2976
2977 return true;
2978 }
2979
2980 void ShenandoahWriteBarrierNode::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) {
2981 assert(is_heap_stable_test(n), "no other tests");
2982 if (identical_backtoback_ifs(n, phase)) {
2983 Node* n_ctrl = n->in(0);
2984 if (phase->can_split_if(n_ctrl)) {
2985 IfNode* dom_if = phase->idom(n_ctrl)->as_If();
2986 if (is_heap_stable_test(n)) {
2987 Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1);
2988 assert(is_gc_state_load(gc_state_load), "broken");
2989 Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1);
2990 assert(is_gc_state_load(dom_gc_state_load), "broken");
2991 if (gc_state_load != dom_gc_state_load) {
2992 phase->igvn().replace_node(gc_state_load, dom_gc_state_load);
2993 }
2994 }
2995 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
2996 Node* proj_true = dom_if->proj_out(1);
2997 Node* proj_false = dom_if->proj_out(0);
2998 Node* con_true = phase->igvn().makecon(TypeInt::ONE);
2999 Node* con_false = phase->igvn().makecon(TypeInt::ZERO);
3000
3001 for (uint i = 1; i < n_ctrl->req(); i++) {
3002 if (phase->is_dominator(proj_true, n_ctrl->in(i))) {
3003 bolphi->init_req(i, con_true);
3004 } else {
3005 assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if");
3006 bolphi->init_req(i, con_false);
3007 }
3008 }
3009 phase->register_new_node(bolphi, n_ctrl);
3010 phase->igvn().replace_input_of(n, 1, bolphi);
3011 phase->do_split_if(n);
3012 }
3013 }
3014 }
3015
3016 IfNode* ShenandoahWriteBarrierNode::find_unswitching_candidate(const IdealLoopTree *loop, PhaseIdealLoop* phase) {
3017 // Find first invariant test that doesn't exit the loop
3018 LoopNode *head = loop->_head->as_Loop();
3019 IfNode* unswitch_iff = NULL;
3020 Node* n = head->in(LoopNode::LoopBackControl);
3021 int loop_has_sfpts = -1;
3022 while (n != head) {
3023 Node* n_dom = phase->idom(n);
3024 if (n->is_Region()) {
3025 if (n_dom->is_If()) {
3026 IfNode* iff = n_dom->as_If();
3027 if (iff->in(1)->is_Bool()) {
3028 BoolNode* bol = iff->in(1)->as_Bool();
3029 if (bol->in(1)->is_Cmp()) {
3030 // If condition is invariant and not a loop exit,
3031 // then found reason to unswitch.
3032 if (is_heap_stable_test(iff) &&
3033 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) {
3034 assert(!loop->is_loop_exit(iff), "both branches should be in the loop");
3035 if (loop_has_sfpts == -1) {
3036 for(uint i = 0; i < loop->_body.size(); i++) {
3037 Node *m = loop->_body[i];
3038 if (m->is_SafePoint() && !m->is_CallLeaf()) {
3039 loop_has_sfpts = 1;
3040 break;
3041 }
3042 }
3043 if (loop_has_sfpts == -1) {
3044 loop_has_sfpts = 0;
3045 }
3046 }
3047 if (!loop_has_sfpts) {
3048 unswitch_iff = iff;
3049 }
3050 }
3051 }
3052 }
3053 }
3054 }
3055 n = n_dom;
3056 }
3057 return unswitch_iff;
3058 }
3059
3060
3061 void ShenandoahWriteBarrierNode::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) {
3062 Node_List heap_stable_tests;
3063 Node_List gc_state_loads;
3064
3065 stack.push(phase->C->start(), 0);
3066 do {
3067 Node* n = stack.node();
3068 uint i = stack.index();
3069
3070 if (i < n->outcnt()) {
3071 Node* u = n->raw_out(i);
3072 stack.set_index(i+1);
3073 if (!visited.test_set(u->_idx)) {
3074 stack.push(u, 0);
3075 }
3076 } else {
3077 stack.pop();
3078 if (ShenandoahCommonGCStateLoads && is_gc_state_load(n)) {
3079 gc_state_loads.push(n);
3080 }
3081 if (n->is_If() && is_heap_stable_test(n)) {
3082 heap_stable_tests.push(n);
3083 }
3084 }
3085 } while (stack.size() > 0);
3086
3087 bool progress;
3088 do {
3089 progress = false;
3090 for (uint i = 0; i < gc_state_loads.size(); i++) {
3091 Node* n = gc_state_loads.at(i);
3092 if (n->outcnt() != 0) {
3093 progress |= try_common_gc_state_load(n, phase);
3094 }
3095 }
3096 } while (progress);
3097
3098 for (uint i = 0; i < heap_stable_tests.size(); i++) {
3099 Node* n = heap_stable_tests.at(i);
3100 assert(is_heap_stable_test(n), "only evacuation test");
3101 merge_back_to_back_tests(n, phase);
3102 }
3103
3104 if (!phase->C->major_progress()) {
3105 VectorSet seen(Thread::current()->resource_area());
3106 for (uint i = 0; i < heap_stable_tests.size(); i++) {
3107 Node* n = heap_stable_tests.at(i);
3108 IdealLoopTree* loop = phase->get_loop(n);
3109 if (loop != phase->ltree_root() &&
3110 loop->_child == NULL &&
3111 !loop->_irreducible) {
3112 LoopNode* head = loop->_head->as_Loop();
3113 if ((!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) &&
3114 !seen.test_set(head->_idx)) {
3115 IfNode* iff = find_unswitching_candidate(loop, phase);
3116 if (iff != NULL) {
3117 Node* bol = iff->in(1);
3118 if (head->is_strip_mined()) {
3119 head->verify_strip_mined(0);
3120 }
3121 move_heap_stable_test_out_of_loop(iff, phase);
3122 if (loop->policy_unswitching(phase)) {
3123 if (head->is_strip_mined()) {
3124 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop();
3125 OuterStripMinedLoopEndNode* le = head->outer_loop_end();
3126 Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl));
3127 phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl));
3128 Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt);
3129 phase->register_control(new_le, phase->get_loop(le), le->in(0));
3130 phase->lazy_replace(outer, new_outer);
3131 phase->lazy_replace(le, new_le);
3132 head->clear_strip_mined();
3133 }
3134 phase->do_unswitching(loop, old_new);
3135 } else {
3136 // Not proceeding with unswitching. Move load back in
3137 // the loop.
3138 phase->igvn().replace_input_of(iff, 1, bol);
3139 }
3140 }
3141 }
3142 }
3143 }
3144 }
3145 }
3146
3147 #ifdef ASSERT
3148 void ShenandoahBarrierNode::verify_raw_mem(RootNode* root) {
3149 const bool trace = false;
3150 ResourceMark rm;
3151 Unique_Node_List nodes;
3152 Unique_Node_List controls;
3153 Unique_Node_List memories;
3154
3155 nodes.push(root);
3156 for (uint next = 0; next < nodes.size(); next++) {
3157 Node *n = nodes.at(next);
3158 if (ShenandoahBarrierSetC2::is_shenandoah_wb_call(n)) {
3159 controls.push(n);
3160 if (trace) { tty->print("XXXXXX verifying"); n->dump(); }
3161 for (uint next2 = 0; next2 < controls.size(); next2++) {
3162 Node *m = controls.at(next2);
3163 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3164 Node* u = m->fast_out(i);
3165 if (u->is_CFG() && !u->is_Root() &&
3166 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1) &&
3167 !(u->is_Region() && u->unique_ctrl_out()->Opcode() == Op_Halt)) {
3168 if (trace) { tty->print("XXXXXX pushing control"); u->dump(); }
3169 controls.push(u);
3170 }
3171 }
3172 }
3173 memories.push(n->as_Call()->proj_out(TypeFunc::Memory));
3174 for (uint next2 = 0; next2 < memories.size(); next2++) {
3175 Node *m = memories.at(next2);
3176 assert(m->bottom_type() == Type::MEMORY, "");
3177 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3178 Node* u = m->fast_out(i);
3179 if (u->bottom_type() == Type::MEMORY && (u->is_Mem() || u->is_ClearArray())) {
3180 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
3181 memories.push(u);
3182 } else if (u->is_LoadStore()) {
3183 if (trace) { tty->print("XXXXXX pushing memory"); u->find_out_with(Op_SCMemProj)->dump(); }
3184 memories.push(u->find_out_with(Op_SCMemProj));
3185 } else if (u->is_MergeMem() && u->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == m) {
3186 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
3187 memories.push(u);
3188 } else if (u->is_Phi()) {
3189 assert(u->bottom_type() == Type::MEMORY, "");
3190 if (u->adr_type() == TypeRawPtr::BOTTOM || u->adr_type() == TypePtr::BOTTOM) {
3191 assert(controls.member(u->in(0)), "");
3192 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
3193 memories.push(u);
3194 }
3195 } else if (u->is_SafePoint() || u->is_MemBar()) {
3196 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
3197 Node* uu = u->fast_out(j);
3198 if (uu->bottom_type() == Type::MEMORY) {
3199 if (trace) { tty->print("XXXXXX pushing memory"); uu->dump(); }
3200 memories.push(uu);
3201 }
3202 }
3203 }
3204 }
3205 }
3206 for (uint next2 = 0; next2 < controls.size(); next2++) {
3207 Node *m = controls.at(next2);
3208 if (m->is_Region()) {
3209 bool all_in = true;
3210 for (uint i = 1; i < m->req(); i++) {
3211 if (!controls.member(m->in(i))) {
3212 all_in = false;
3213 break;
3214 }
3215 }
3216 if (trace) { tty->print("XXX verifying %s", all_in ? "all in" : ""); m->dump(); }
3217 bool found_phi = false;
3218 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax && !found_phi; j++) {
3219 Node* u = m->fast_out(j);
3220 if (u->is_Phi() && memories.member(u)) {
3221 found_phi = true;
3222 for (uint i = 1; i < u->req() && found_phi; i++) {
3223 Node* k = u->in(i);
3224 if (memories.member(k) != controls.member(m->in(i))) {
3225 found_phi = false;
3226 }
3227 }
3228 }
3229 }
3230 assert(found_phi || all_in, "");
3231 }
3232 }
3233 controls.clear();
3234 memories.clear();
3235 }
3236 for( uint i = 0; i < n->len(); ++i ) {
3237 Node *m = n->in(i);
3238 if (m != NULL) {
3239 nodes.push(m);
3240 }
3241 }
3242 }
3243 }
3244 #endif
3245
3246 const Type* ShenandoahEnqueueBarrierNode::bottom_type() const {
3247 if (in(1) == NULL || in(1)->is_top()) {
3248 return Type::TOP;
3249 }
3250 const Type* t = in(1)->bottom_type();
3251 if (t == TypePtr::NULL_PTR) {
3252 return t;
3253 }
3254 return t->is_oopptr()->cast_to_nonconst();
3255 }
3256
3257 const Type* ShenandoahEnqueueBarrierNode::Value(PhaseGVN* phase) const {
3258 if (in(1) == NULL) {
3259 return Type::TOP;
3260 }
3261 const Type* t = phase->type(in(1));
3262 if (t == Type::TOP) {
3263 return Type::TOP;
3264 }
3265 if (t == TypePtr::NULL_PTR) {
3266 return t;
3267 }
3268 return t->is_oopptr()->cast_to_nonconst();
3269 }
3270
3271 int ShenandoahEnqueueBarrierNode::needed(Node* n) {
3272 if (n == NULL ||
3273 n->is_Allocate() ||
3274 n->bottom_type() == TypePtr::NULL_PTR ||
3275 (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL)) {
3276 return NotNeeded;
3277 }
3278 if (n->is_Phi() ||
3279 n->is_CMove()) {
3280 return MaybeNeeded;
3281 }
3282 return Needed;
3283 }
3284
3285 Node* ShenandoahEnqueueBarrierNode::next(Node* n) {
3286 for (;;) {
3287 if (n == NULL) {
3288 return n;
3289 } else if (n->bottom_type() == TypePtr::NULL_PTR) {
3290 return n;
3291 } else if (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL) {
3292 return n;
3293 } else if (n->is_ConstraintCast() ||
3294 n->Opcode() == Op_DecodeN ||
3295 n->Opcode() == Op_EncodeP) {
3296 n = n->in(1);
3297 } else if (n->is_Proj()) {
3298 n = n->in(0);
3299 } else {
3300 return n;
3301 }
3302 }
3303 ShouldNotReachHere();
3304 return NULL;
3305 }
3306
3307 Node* ShenandoahEnqueueBarrierNode::Identity(PhaseGVN* phase) {
3308 PhaseIterGVN* igvn = phase->is_IterGVN();
3309
3310 Node* n = next(in(1));
3311
3312 int cont = needed(n);
3313
3314 if (cont == NotNeeded) {
3315 return in(1);
3316 } else if (cont == MaybeNeeded) {
3317 if (igvn == NULL) {
3318 phase->record_for_igvn(this);
3319 return this;
3320 } else {
3321 ResourceMark rm;
3322 Unique_Node_List wq;
3323 uint wq_i = 0;
3324
3325 for (;;) {
3326 if (n->is_Phi()) {
3327 for (uint i = 1; i < n->req(); i++) {
3328 Node* m = n->in(i);
3329 if (m != NULL) {
3330 wq.push(m);
3331 }
3332 }
3333 } else {
3334 assert(n->is_CMove(), "nothing else here");
3335 Node* m = n->in(CMoveNode::IfFalse);
3336 wq.push(m);
3337 m = n->in(CMoveNode::IfTrue);
3338 wq.push(m);
3339 }
3340 Node* orig_n = NULL;
3341 do {
3342 if (wq_i >= wq.size()) {
3343 return in(1);
3344 }
3345 n = wq.at(wq_i);
3346 wq_i++;
3347 orig_n = n;
3348 n = next(n);
3349 cont = needed(n);
3350 if (cont == Needed) {
3351 return this;
3352 }
3353 } while (cont != MaybeNeeded || (orig_n != n && wq.member(n)));
3354 }
3355 }
3356 }
3357
3358 return this;
3359 }
3360
3361 #ifdef ASSERT
3362 static bool has_never_branch(Node* root) {
3363 for (uint i = 1; i < root->req(); i++) {
3364 Node* in = root->in(i);
3365 if (in != NULL && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->Opcode() == Op_NeverBranch) {
3366 return true;
3367 }
3368 }
3369 return false;
3370 }
3371 #endif
3372
3373 void MemoryGraphFixer::collect_memory_nodes() {
3374 Node_Stack stack(0);
3375 VectorSet visited(Thread::current()->resource_area());
3376 Node_List regions;
3377
3378 // Walk the raw memory graph and create a mapping from CFG node to
3379 // memory node. Exclude phis for now.
3380 stack.push(_phase->C->root(), 1);
3381 do {
3382 Node* n = stack.node();
3383 int opc = n->Opcode();
3384 uint i = stack.index();
3385 if (i < n->req()) {
3386 Node* mem = NULL;
3387 if (opc == Op_Root) {
3388 Node* in = n->in(i);
3389 int in_opc = in->Opcode();
3390 if (in_opc == Op_Return || in_opc == Op_Rethrow) {
3391 mem = in->in(TypeFunc::Memory);
3392 } else if (in_opc == Op_Halt) {
3393 if (!in->in(0)->is_Region()) {
3394 Node* proj = in->in(0);
3395 assert(proj->is_Proj(), "");
3396 Node* in = proj->in(0);
3397 assert(in->is_CallStaticJava() || in->Opcode() == Op_NeverBranch || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
3398 if (in->is_CallStaticJava()) {
3399 mem = in->in(TypeFunc::Memory);
3400 } else if (in->Opcode() == Op_Catch) {
3401 Node* call = in->in(0)->in(0);
3402 assert(call->is_Call(), "");
3403 mem = call->in(TypeFunc::Memory);
3404 }
3405 }
3406 } else {
3407 #ifdef ASSERT
3408 n->dump();
3409 in->dump();
3410 #endif
3411 ShouldNotReachHere();
3412 }
3413 } else {
3414 assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
3415 assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
3416 mem = n->in(i);
3417 }
3418 i++;
3419 stack.set_index(i);
3420 if (mem == NULL) {
3421 continue;
3422 }
3423 for (;;) {
3424 if (visited.test_set(mem->_idx) || mem->is_Start()) {
3425 break;
3426 }
3427 if (mem->is_Phi()) {
3428 stack.push(mem, 2);
3429 mem = mem->in(1);
3430 } else if (mem->is_Proj()) {
3431 stack.push(mem, mem->req());
3432 mem = mem->in(0);
3433 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
3434 mem = mem->in(TypeFunc::Memory);
3435 } else if (mem->is_MergeMem()) {
3436 MergeMemNode* mm = mem->as_MergeMem();
3437 mem = mm->memory_at(_alias);
3438 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
3439 assert(_alias == Compile::AliasIdxRaw, "");
3440 stack.push(mem, mem->req());
3441 mem = mem->in(MemNode::Memory);
3442 } else if (mem->Opcode() == Op_ShenandoahWriteBarrier) {
3443 assert(_alias != Compile::AliasIdxRaw, "");
3444 mem = mem->in(ShenandoahBarrierNode::Memory);
3445 } else if (mem->Opcode() == Op_ShenandoahWBMemProj) {
3446 stack.push(mem, mem->req());
3447 mem = mem->in(ShenandoahWBMemProjNode::WriteBarrier);
3448 } else {
3449 #ifdef ASSERT
3450 mem->dump();
3451 #endif
3452 ShouldNotReachHere();
3453 }
3454 }
3455 } else {
3456 if (n->is_Phi()) {
3457 // Nothing
3458 } else if (!n->is_Root()) {
3459 Node* c = get_ctrl(n);
3460 _memory_nodes.map(c->_idx, n);
3461 }
3462 stack.pop();
3463 }
3464 } while(stack.is_nonempty());
3465
3466 // Iterate over CFG nodes in rpo and propagate memory state to
3467 // compute memory state at regions, creating new phis if needed.
3468 Node_List rpo_list;
3469 visited.Clear();
3470 _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
3471 Node* root = rpo_list.pop();
3472 assert(root == _phase->C->root(), "");
3473
3474 const bool trace = false;
3475 #ifdef ASSERT
3476 if (trace) {
3477 for (int i = rpo_list.size() - 1; i >= 0; i--) {
3478 Node* c = rpo_list.at(i);
3479 if (_memory_nodes[c->_idx] != NULL) {
3480 tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump();
3481 }
3482 }
3483 }
3484 #endif
3485 uint last = _phase->C->unique();
3486
3487 #ifdef ASSERT
3488 uint8_t max_depth = 0;
3489 for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
3490 IdealLoopTree* lpt = iter.current();
3491 max_depth = MAX2(max_depth, lpt->_nest);
3492 }
3493 #endif
3494
3495 bool progress = true;
3496 int iteration = 0;
3497 Node_List dead_phis;
3498 while (progress) {
3499 progress = false;
3500 iteration++;
3501 assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop(), "");
3502 if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
3503 IdealLoopTree* last_updated_ilt = NULL;
3504 for (int i = rpo_list.size() - 1; i >= 0; i--) {
3505 Node* c = rpo_list.at(i);
3506
3507 Node* prev_mem = _memory_nodes[c->_idx];
3508 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
3509 Node* prev_region = regions[c->_idx];
3510 Node* unique = NULL;
3511 for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
3512 Node* m = _memory_nodes[c->in(j)->_idx];
3513 assert(m != NULL || (c->is_Loop() && j == LoopNode::LoopBackControl && iteration == 1) || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "expect memory state");
3514 if (m != NULL) {
3515 if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
3516 assert(c->is_Loop() && j == LoopNode::LoopBackControl || _phase->C->has_irreducible_loop(), "");
3517 // continue
3518 } else if (unique == NULL) {
3519 unique = m;
3520 } else if (m == unique) {
3521 // continue
3522 } else {
3523 unique = NodeSentinel;
3524 }
3525 }
3526 }
3527 assert(unique != NULL, "empty phi???");
3528 if (unique != NodeSentinel) {
3529 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c) {
3530 dead_phis.push(prev_region);
3531 }
3532 regions.map(c->_idx, unique);
3533 } else {
3534 Node* phi = NULL;
3535 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
3536 phi = prev_region;
3537 for (uint k = 1; k < c->req(); k++) {
3538 Node* m = _memory_nodes[c->in(k)->_idx];
3539 assert(m != NULL, "expect memory state");
3540 phi->set_req(k, m);
3541 }
3542 } else {
3543 for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == NULL; j++) {
3544 Node* u = c->fast_out(j);
3545 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
3546 (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
3547 phi = u;
3548 for (uint k = 1; k < c->req() && phi != NULL; k++) {
3549 Node* m = _memory_nodes[c->in(k)->_idx];
3550 assert(m != NULL, "expect memory state");
3551 if (u->in(k) != m) {
3552 phi = NULL;
3553 }
3554 }
3555 }
3556 }
3557 if (phi == NULL) {
3558 phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
3559 for (uint k = 1; k < c->req(); k++) {
3560 Node* m = _memory_nodes[c->in(k)->_idx];
3561 assert(m != NULL, "expect memory state");
3562 phi->init_req(k, m);
3563 }
3564 }
3565 }
3566 assert(phi != NULL, "");
3567 regions.map(c->_idx, phi);
3568 }
3569 Node* current_region = regions[c->_idx];
3570 if (current_region != prev_region) {
3571 progress = true;
3572 if (prev_region == prev_mem) {
3573 _memory_nodes.map(c->_idx, current_region);
3574 }
3575 }
3576 } else if (prev_mem == NULL || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
3577 Node* m = _memory_nodes[_phase->idom(c)->_idx];
3578 assert(m != NULL, "expect memory state");
3579 if (m != prev_mem) {
3580 _memory_nodes.map(c->_idx, m);
3581 progress = true;
3582 }
3583 }
3584 #ifdef ASSERT
3585 if (trace) { tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); }
3586 #endif
3587 }
3588 }
3589
3590 // Replace existing phi with computed memory state for that region
3591 // if different (could be a new phi or a dominating memory node if
3592 // that phi was found to be useless).
3593 while (dead_phis.size() > 0) {
3594 Node* n = dead_phis.pop();
3595 n->replace_by(_phase->C->top());
3596 n->destruct();
3597 }
3598 for (int i = rpo_list.size() - 1; i >= 0; i--) {
3599 Node* c = rpo_list.at(i);
3600 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
3601 Node* n = regions[c->_idx];
3602 if (n->is_Phi() && n->_idx >= last && n->in(0) == c) {
3603 _phase->register_new_node(n, c);
3604 }
3605 }
3606 }
3607 for (int i = rpo_list.size() - 1; i >= 0; i--) {
3608 Node* c = rpo_list.at(i);
3609 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
3610 Node* n = regions[c->_idx];
3611 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
3612 Node* u = c->fast_out(i);
3613 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
3614 u != n) {
3615 if (u->adr_type() == TypePtr::BOTTOM) {
3616 fix_memory_uses(u, n, n, c);
3617 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
3618 _phase->lazy_replace(u, n);
3619 --i; --imax;
3620 }
3621 }
3622 }
3623 }
3624 }
3625 }
3626
3627 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
3628 Node* c = _phase->get_ctrl(n);
3629 if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Call()) {
3630 assert(c == n->in(0), "");
3631 CallNode* call = c->as_Call();
3632 CallProjections projs;
3633 call->extract_projections(&projs, true, false);
3634 if (projs.catchall_memproj != NULL) {
3635 if (projs.fallthrough_memproj == n) {
3636 c = projs.fallthrough_catchproj;
3637 } else {
3638 assert(projs.catchall_memproj == n, "");
3639 c = projs.catchall_catchproj;
3640 }
3641 }
3642 }
3643 return c;
3644 }
3645
3646 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
3647 if (_phase->has_ctrl(n))
3648 return get_ctrl(n);
3649 else {
3650 assert (n->is_CFG(), "must be a CFG node");
3651 return n;
3652 }
3653 }
3654
3655 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
3656 return m != NULL && get_ctrl(m) == c;
3657 }
3658
3659 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
3660 assert(n == NULL || _phase->ctrl_or_self(n) == ctrl, "");
3661 Node* mem = _memory_nodes[ctrl->_idx];
3662 Node* c = ctrl;
3663 while (!mem_is_valid(mem, c) &&
3664 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
3665 c = _phase->idom(c);
3666 mem = _memory_nodes[c->_idx];
3667 }
3668 if (n != NULL && mem_is_valid(mem, c)) {
3669 while (!ShenandoahWriteBarrierNode::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
3670 mem = next_mem(mem, _alias);
3671 }
3672 if (mem->is_MergeMem()) {
3673 mem = mem->as_MergeMem()->memory_at(_alias);
3674 }
3675 if (!mem_is_valid(mem, c)) {
3676 do {
3677 c = _phase->idom(c);
3678 mem = _memory_nodes[c->_idx];
3679 } while (!mem_is_valid(mem, c) &&
3680 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
3681 }
3682 }
3683 assert(mem->bottom_type() == Type::MEMORY, "");
3684 return mem;
3685 }
3686
3687 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
3688 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
3689 Node* use = region->fast_out(i);
3690 if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
3691 (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
3692 return true;
3693 }
3694 }
3695 return false;
3696 }
3697
3698 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
3699 assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
3700 const bool trace = false;
3701 DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
3702 DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
3703 GrowableArray<Node*> phis;
3704 if (mem_for_ctrl != mem) {
3705 Node* old = mem_for_ctrl;
3706 Node* prev = NULL;
3707 while (old != mem) {
3708 prev = old;
3709 if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
3710 assert(_alias == Compile::AliasIdxRaw, "");
3711 old = old->in(MemNode::Memory);
3712 } else if (old->Opcode() == Op_SCMemProj) {
3713 assert(_alias == Compile::AliasIdxRaw, "");
3714 old = old->in(0);
3715 } else if (old->Opcode() == Op_ShenandoahWBMemProj) {
3716 assert(_alias != Compile::AliasIdxRaw, "");
3717 old = old->in(ShenandoahWBMemProjNode::WriteBarrier);
3718 } else if (old->Opcode() == Op_ShenandoahWriteBarrier) {
3719 assert(_alias != Compile::AliasIdxRaw, "");
3720 old = old->in(ShenandoahBarrierNode::Memory);
3721 } else {
3722 ShouldNotReachHere();
3723 }
3724 }
3725 assert(prev != NULL, "");
3726 if (new_ctrl != ctrl) {
3727 _memory_nodes.map(ctrl->_idx, mem);
3728 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
3729 }
3730 uint input = prev->Opcode() == Op_ShenandoahWriteBarrier ? (uint)ShenandoahBarrierNode::Memory : (uint)MemNode::Memory;
3731 _phase->igvn().replace_input_of(prev, input, new_mem);
3732 } else {
3733 uses.clear();
3734 _memory_nodes.map(new_ctrl->_idx, new_mem);
3735 uses.push(new_ctrl);
3736 for(uint next = 0; next < uses.size(); next++ ) {
3737 Node *n = uses.at(next);
3738 assert(n->is_CFG(), "");
3739 DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
3740 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3741 Node* u = n->fast_out(i);
3742 if (!u->is_Root() && u->is_CFG() && u != n) {
3743 Node* m = _memory_nodes[u->_idx];
3744 if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
3745 !has_mem_phi(u) &&
3746 u->unique_ctrl_out()->Opcode() != Op_Halt) {
3747 DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
3748 DEBUG_ONLY(if (trace && m != NULL) { tty->print("ZZZ mem"); m->dump(); });
3749
3750 if (!mem_is_valid(m, u) || !m->is_Phi()) {
3751 bool push = true;
3752 bool create_phi = true;
3753 if (_phase->is_dominator(new_ctrl, u)) {
3754 create_phi = false;
3755 } else if (!_phase->C->has_irreducible_loop()) {
3756 IdealLoopTree* loop = _phase->get_loop(ctrl);
3757 bool do_check = true;
3758 IdealLoopTree* l = loop;
3759 create_phi = false;
3760 while (l != _phase->ltree_root()) {
3761 if (_phase->is_dominator(l->_head, u) && _phase->is_dominator(_phase->idom(u), l->_head)) {
3762 create_phi = true;
3763 do_check = false;
3764 break;
3765 }
3766 l = l->_parent;
3767 }
3768
3769 if (do_check) {
3770 assert(!create_phi, "");
3771 IdealLoopTree* u_loop = _phase->get_loop(u);
3772 if (u_loop != _phase->ltree_root() && u_loop->is_member(loop)) {
3773 Node* c = ctrl;
3774 while (!_phase->is_dominator(c, u_loop->tail())) {
3775 c = _phase->idom(c);
3776 }
3777 if (!_phase->is_dominator(c, u)) {
3778 do_check = false;
3779 }
3780 }
3781 }
3782
3783 if (do_check && _phase->is_dominator(_phase->idom(u), new_ctrl)) {
3784 create_phi = true;
3785 }
3786 }
3787 if (create_phi) {
3788 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
3789 _phase->register_new_node(phi, u);
3790 phis.push(phi);
3791 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
3792 if (!mem_is_valid(m, u)) {
3793 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
3794 _memory_nodes.map(u->_idx, phi);
3795 } else {
3796 DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
3797 for (;;) {
3798 assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj() || m->Opcode() == Op_ShenandoahWriteBarrier || m->Opcode() == Op_ShenandoahWBMemProj, "");
3799 Node* next = NULL;
3800 if (m->is_Proj()) {
3801 next = m->in(0);
3802 } else if (m->Opcode() == Op_ShenandoahWBMemProj) {
3803 next = m->in(ShenandoahWBMemProjNode::WriteBarrier);
3804 } else if (m->is_Mem() || m->is_LoadStore()) {
3805 assert(_alias == Compile::AliasIdxRaw, "");
3806 next = m->in(MemNode::Memory);
3807 } else {
3808 assert(_alias != Compile::AliasIdxRaw, "");
3809 assert (m->Opcode() == Op_ShenandoahWriteBarrier, "");
3810 next = m->in(ShenandoahBarrierNode::Memory);
3811 }
3812 if (_phase->get_ctrl(next) != u) {
3813 break;
3814 }
3815 if (next->is_MergeMem()) {
3816 assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
3817 break;
3818 }
3819 if (next->is_Phi()) {
3820 assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
3821 break;
3822 }
3823 m = next;
3824 }
3825
3826 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
3827 assert(m->is_Mem() || m->is_LoadStore() || m->Opcode() == Op_ShenandoahWriteBarrier, "");
3828 uint input = (m->is_Mem() || m->is_LoadStore()) ? (uint)MemNode::Memory : (uint)ShenandoahBarrierNode::Memory;
3829 _phase->igvn().replace_input_of(m, input, phi);
3830 push = false;
3831 }
3832 } else {
3833 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
3834 }
3835 if (push) {
3836 uses.push(u);
3837 }
3838 }
3839 } else if (!mem_is_valid(m, u) &&
3840 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1)) {
3841 uses.push(u);
3842 }
3843 }
3844 }
3845 }
3846 for (int i = 0; i < phis.length(); i++) {
3847 Node* n = phis.at(i);
3848 Node* r = n->in(0);
3849 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
3850 for (uint j = 1; j < n->req(); j++) {
3851 Node* m = find_mem(r->in(j), NULL);
3852 _phase->igvn().replace_input_of(n, j, m);
3853 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
3854 }
3855 }
3856 }
3857 uint last = _phase->C->unique();
3858 MergeMemNode* mm = NULL;
3859 int alias = _alias;
3860 DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
3861 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
3862 Node* u = mem->out(i);
3863 if (u->_idx < last) {
3864 if (u->is_Mem()) {
3865 if (_phase->C->get_alias_index(u->adr_type()) == alias) {
3866 Node* m = find_mem(_phase->get_ctrl(u), u);
3867 if (m != mem) {
3868 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
3869 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
3870 --i;
3871 }
3872 }
3873 } else if (u->is_MergeMem()) {
3874 MergeMemNode* u_mm = u->as_MergeMem();
3875 if (u_mm->memory_at(alias) == mem) {
3876 MergeMemNode* newmm = NULL;
3877 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
3878 Node* uu = u->fast_out(j);
3879 assert(!uu->is_MergeMem(), "chain of MergeMems?");
3880 if (uu->is_Phi()) {
3881 assert(uu->adr_type() == TypePtr::BOTTOM, "");
3882 Node* region = uu->in(0);
3883 int nb = 0;
3884 for (uint k = 1; k < uu->req(); k++) {
3885 if (uu->in(k) == u) {
3886 Node* m = find_mem(region->in(k), NULL);
3887 if (m != mem) {
3888 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
3889 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
3890 if (newmm != u) {
3891 _phase->igvn().replace_input_of(uu, k, newmm);
3892 nb++;
3893 --jmax;
3894 }
3895 }
3896 }
3897 }
3898 if (nb > 0) {
3899 --j;
3900 }
3901 } else {
3902 Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
3903 if (m != mem) {
3904 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
3905 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
3906 if (newmm != u) {
3907 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
3908 --j, --jmax;
3909 }
3910 }
3911 }
3912 }
3913 }
3914 } else if (u->is_Phi()) {
3915 assert(u->bottom_type() == Type::MEMORY, "what else?");
3916 if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
3917 Node* region = u->in(0);
3918 bool replaced = false;
3919 for (uint j = 1; j < u->req(); j++) {
3920 if (u->in(j) == mem) {
3921 Node* m = find_mem(region->in(j), NULL);
3922 Node* nnew = m;
3923 if (m != mem) {
3924 if (u->adr_type() == TypePtr::BOTTOM) {
3925 mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
3926 nnew = mm;
3927 }
3928 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
3929 _phase->igvn().replace_input_of(u, j, nnew);
3930 replaced = true;
3931 }
3932 }
3933 }
3934 if (replaced) {
3935 --i;
3936 }
3937 }
3938 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
3939 u->adr_type() == NULL) {
3940 assert(u->adr_type() != NULL ||
3941 u->Opcode() == Op_Rethrow ||
3942 u->Opcode() == Op_Return ||
3943 u->Opcode() == Op_SafePoint ||
3944 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
3945 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
3946 u->Opcode() == Op_CallLeaf, "");
3947 Node* m = find_mem(_phase->ctrl_or_self(u), u);
3948 if (m != mem) {
3949 mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
3950 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
3951 --i;
3952 }
3953 } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
3954 Node* m = find_mem(_phase->ctrl_or_self(u), u);
3955 if (m != mem) {
3956 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
3957 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
3958 --i;
3959 }
3960 } else if (u->adr_type() != TypePtr::BOTTOM &&
3961 _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
3962 Node* m = find_mem(_phase->ctrl_or_self(u), u);
3963 assert(m != mem, "");
3964 // u is on the wrong slice...
3965 assert(u->is_ClearArray(), "");
3966 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
3967 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
3968 --i;
3969 }
3970 }
3971 }
3972 #ifdef ASSERT
3973 assert(new_mem->outcnt() > 0, "");
3974 for (int i = 0; i < phis.length(); i++) {
3975 Node* n = phis.at(i);
3976 assert(n->outcnt() > 0, "new phi must have uses now");
3977 }
3978 #endif
3979 }
3980
3981 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
3982 MergeMemNode* mm = MergeMemNode::make(mem);
3983 mm->set_memory_at(_alias, rep_proj);
3984 _phase->register_new_node(mm, rep_ctrl);
3985 return mm;
3986 }
3987
3988 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
3989 MergeMemNode* newmm = NULL;
3990 MergeMemNode* u_mm = u->as_MergeMem();
3991 Node* c = _phase->get_ctrl(u);
3992 if (_phase->is_dominator(c, rep_ctrl)) {
3993 c = rep_ctrl;
3994 } else {
3995 assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
3996 }
3997 if (u->outcnt() == 1) {
3998 if (u->req() > (uint)_alias && u->in(_alias) == mem) {
3999 _phase->igvn().replace_input_of(u, _alias, rep_proj);
4000 --i;
4001 } else {
4002 _phase->igvn().rehash_node_delayed(u);
4003 u_mm->set_memory_at(_alias, rep_proj);
4004 }
4005 newmm = u_mm;
4006 _phase->set_ctrl_and_loop(u, c);
4007 } else {
4008 // can't simply clone u and then change one of its input because
4009 // it adds and then removes an edge which messes with the
4010 // DUIterator
4011 newmm = MergeMemNode::make(u_mm->base_memory());
4012 for (uint j = 0; j < u->req(); j++) {
4013 if (j < newmm->req()) {
4014 if (j == (uint)_alias) {
4015 newmm->set_req(j, rep_proj);
4016 } else if (newmm->in(j) != u->in(j)) {
4017 newmm->set_req(j, u->in(j));
4018 }
4019 } else if (j == (uint)_alias) {
4020 newmm->add_req(rep_proj);
4021 } else {
4022 newmm->add_req(u->in(j));
4023 }
4024 }
4025 if ((uint)_alias >= u->req()) {
4026 newmm->set_memory_at(_alias, rep_proj);
4027 }
4028 _phase->register_new_node(newmm, c);
4029 }
4030 return newmm;
4031 }
4032
4033 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
4034 if (phi->adr_type() == TypePtr::BOTTOM) {
4035 Node* region = phi->in(0);
4036 for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
4037 Node* uu = region->fast_out(j);
4038 if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
4039 return false;
4040 }
4041 }
4042 return true;
4043 }
4044 return _phase->C->get_alias_index(phi->adr_type()) == _alias;
4045 }
4046
4047 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
4048 uint last = _phase-> C->unique();
4049 MergeMemNode* mm = NULL;
4050 assert(mem->bottom_type() == Type::MEMORY, "");
4051 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
4052 Node* u = mem->out(i);
4053 if (u != replacement && u->_idx < last) {
4054 if (u->is_ShenandoahBarrier() && _alias != Compile::AliasIdxRaw) {
4055 if (_phase->C->get_alias_index(u->adr_type()) == _alias && ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
4056 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
4057 assert(u->find_edge(mem) == -1, "only one edge");
4058 --i;
4059 }
4060 } else if (u->is_Mem()) {
4061 if (_phase->C->get_alias_index(u->adr_type()) == _alias && ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
4062 assert(_alias == Compile::AliasIdxRaw , "only raw memory can lead to a memory operation");
4063 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
4064 assert(u->find_edge(mem) == -1, "only one edge");
4065 --i;
4066 }
4067 } else if (u->is_MergeMem()) {
4068 MergeMemNode* u_mm = u->as_MergeMem();
4069 if (u_mm->memory_at(_alias) == mem) {
4070 MergeMemNode* newmm = NULL;
4071 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
4072 Node* uu = u->fast_out(j);
4073 assert(!uu->is_MergeMem(), "chain of MergeMems?");
4074 if (uu->is_Phi()) {
4075 if (should_process_phi(uu)) {
4076 Node* region = uu->in(0);
4077 int nb = 0;
4078 for (uint k = 1; k < uu->req(); k++) {
4079 if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
4080 if (newmm == NULL) {
4081 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
4082 }
4083 if (newmm != u) {
4084 _phase->igvn().replace_input_of(uu, k, newmm);
4085 nb++;
4086 --jmax;
4087 }
4088 }
4089 }
4090 if (nb > 0) {
4091 --j;
4092 }
4093 }
4094 } else {
4095 if (rep_ctrl != uu && ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
4096 if (newmm == NULL) {
4097 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
4098 }
4099 if (newmm != u) {
4100 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
4101 --j, --jmax;
4102 }
4103 }
4104 }
4105 }
4106 }
4107 } else if (u->is_Phi()) {
4108 assert(u->bottom_type() == Type::MEMORY, "what else?");
4109 Node* region = u->in(0);
4110 if (should_process_phi(u)) {
4111 bool replaced = false;
4112 for (uint j = 1; j < u->req(); j++) {
4113 if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
4114 Node* nnew = rep_proj;
4115 if (u->adr_type() == TypePtr::BOTTOM) {
4116 if (mm == NULL) {
4117 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
4118 }
4119 nnew = mm;
4120 }
4121 _phase->igvn().replace_input_of(u, j, nnew);
4122 replaced = true;
4123 }
4124 }
4125 if (replaced) {
4126 --i;
4127 }
4128
4129 }
4130 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
4131 u->adr_type() == NULL) {
4132 assert(u->adr_type() != NULL ||
4133 u->Opcode() == Op_Rethrow ||
4134 u->Opcode() == Op_Return ||
4135 u->Opcode() == Op_SafePoint ||
4136 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
4137 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
4138 u->Opcode() == Op_CallLeaf, "");
4139 if (ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
4140 if (mm == NULL) {
4141 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
4142 }
4143 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
4144 --i;
4145 }
4146 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
4147 if (ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
4148 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
4149 --i;
4150 }
4151 }
4152 }
4153 }
4154 }
4155
4156 void MemoryGraphFixer::remove(Node* n) {
4157 assert(n->Opcode() == Op_ShenandoahWBMemProj, "");
4158 Node* c = _phase->get_ctrl(n);
4159 Node* mem = find_mem(c, NULL);
4160 if (mem == n) {
4161 _memory_nodes.map(c->_idx, mem->in(ShenandoahWBMemProjNode::WriteBarrier)->in(ShenandoahBarrierNode::Memory));
4162 }
4163 }
4164
4165