1 /*
2 * Copyright (c) 2015, Red Hat, Inc. and/or its affiliates.
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 #include "gc/shenandoah/brooksPointer.hpp"
26 #include "gc/shenandoah/shenandoahHeap.hpp"
27 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
28 #include "opto/arraycopynode.hpp"
29 #include "opto/block.hpp"
30 #include "opto/callnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/movenode.hpp"
33 #include "opto/phaseX.hpp"
34 #include "opto/rootnode.hpp"
35 #include "opto/runtime.hpp"
36 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
37 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
38 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
39 #include "opto/subnode.hpp"
40
41 Node* ShenandoahBarrierNode::skip_through_barrier(Node* n) {
42 if (n == NULL) {
43 return NULL;
44 }
45 if (n->Opcode() == Op_ShenandoahEnqueueBarrier) {
46 n = n->in(1);
47 }
48
49 if (n->is_ShenandoahBarrier()) {
50 return n->in(ValueIn);
51 } else if (n->is_Phi() &&
52 n->req() == 3 &&
53 n->in(1) != NULL &&
54 n->in(1)->is_ShenandoahBarrier() &&
55 n->in(2) != NULL &&
56 n->in(2)->bottom_type() == TypePtr::NULL_PTR &&
57 n->in(0) != NULL &&
58 n->in(0)->in(1) != NULL &&
59 n->in(0)->in(1)->is_IfProj() &&
60 n->in(0)->in(2) != NULL &&
61 n->in(0)->in(2)->is_IfProj() &&
62 n->in(0)->in(1)->in(0) != NULL &&
63 n->in(0)->in(1)->in(0) == n->in(0)->in(2)->in(0) &&
64 n->in(1)->in(ValueIn)->Opcode() == Op_CastPP) {
65 Node* iff = n->in(0)->in(1)->in(0);
66 Node* res = n->in(1)->in(ValueIn)->in(1);
67 if (iff->is_If() &&
68 iff->in(1) != NULL &&
69 iff->in(1)->is_Bool() &&
70 iff->in(1)->as_Bool()->_test._test == BoolTest::ne &&
71 iff->in(1)->in(1) != NULL &&
72 iff->in(1)->in(1)->Opcode() == Op_CmpP &&
73 iff->in(1)->in(1)->in(1) != NULL &&
74 iff->in(1)->in(1)->in(1) == res &&
75 iff->in(1)->in(1)->in(2) != NULL &&
76 iff->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
77 return res;
78 }
79 }
80 return n;
81 }
82
83 bool ShenandoahBarrierNode::needs_barrier(PhaseGVN* phase, ShenandoahBarrierNode* orig, Node* n, Node* rb_mem, bool allow_fromspace) {
84 Unique_Node_List visited;
85 return needs_barrier_impl(phase, orig, n, rb_mem, allow_fromspace, visited);
86 }
87
88 bool ShenandoahBarrierNode::needs_barrier_impl(PhaseGVN* phase, ShenandoahBarrierNode* orig, Node* n, Node* rb_mem, bool allow_fromspace, Unique_Node_List &visited) {
89
90 if (visited.member(n)) {
91 return false; // Been there.
92 }
93 visited.push(n);
94
95 if (n->is_Allocate()) {
96 // tty->print_cr("killed barrier for newly allocated object");
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 // tty->print_cr("killed barrier for NULL object");
110 return false;
111 }
112 if (type->make_oopptr() && type->make_oopptr()->const_oop() != NULL) {
113 // tty->print_cr("killed barrier for constant object");
114 return false;
115 }
116
117 if (ShenandoahOptimizeStableFinals) {
118 const TypeAryPtr* ary = type->isa_aryptr();
119 if (ary && ary->is_stable() && allow_fromspace) {
120 return false;
121 }
122 }
123
124 if (n->is_CheckCastPP() || n->is_ConstraintCast() || n->Opcode() == Op_ShenandoahEnqueueBarrier) {
125 return needs_barrier_impl(phase, orig, n->in(1), rb_mem, allow_fromspace, visited);
126 }
127 if (n->is_Parm()) {
128 return true;
129 }
130 if (n->is_Proj()) {
131 return needs_barrier_impl(phase, orig, n->in(0), rb_mem, allow_fromspace, visited);
132 }
133 if (n->is_Phi()) {
134 bool need_barrier = false;
135 for (uint i = 1; i < n->req() && ! need_barrier; i++) {
136 Node* input = n->in(i);
137 if (input == NULL) {
138 need_barrier = true; // Phi not complete yet?
139 } else if (needs_barrier_impl(phase, orig, input, rb_mem, allow_fromspace, visited)) {
140 need_barrier = true;
141 }
142 }
143 return need_barrier;
144 }
145 if (n->is_CMove()) {
146 return needs_barrier_impl(phase, orig, n->in(CMoveNode::IfFalse), rb_mem, allow_fromspace, visited) ||
147 needs_barrier_impl(phase, orig, n->in(CMoveNode::IfTrue ), rb_mem, allow_fromspace, visited);
148 }
149 if (n->Opcode() == Op_CreateEx) {
150 return true;
151 }
152 if (n->Opcode() == Op_ShenandoahWriteBarrier) {
153 // tty->print_cr("skipped barrier for chained write barrier object");
154 return false;
155 }
156 if (n->Opcode() == Op_ShenandoahReadBarrier) {
157 if (rb_mem == n->in(Memory)) {
158 // tty->print_cr("Eliminated chained read barrier");
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_GetAndSetN ||
170 n->Opcode() == Op_CompareAndExchangeN) {
171 return true;
172 }
173 if (n->Opcode() == Op_DecodeN ||
174 n->Opcode() == Op_EncodeP) {
175 return needs_barrier_impl(phase, orig, n->in(1), rb_mem, allow_fromspace, visited);
176 }
177
178 #ifdef ASSERT
179 tty->print("need barrier on?: "); n->dump();
180 ShouldNotReachHere();
181 #endif
182 return true;
183 }
184
185 bool ShenandoahReadBarrierNode::dominates_memory_rb_impl(PhaseGVN* phase,
186 Node* b1,
187 Node* b2,
188 Node* current,
189 bool linear) {
190 ResourceMark rm;
191 VectorSet visited(Thread::current()->resource_area());
192 Node_Stack phis(0);
193
194 for(int i = 0; i < 10; i++) {
195 if (current == NULL) {
196 return false;
197 } else if (visited.test_set(current->_idx) || current->is_top() || current == b1) {
198 current = NULL;
199 while (phis.is_nonempty() && current == NULL) {
200 uint idx = phis.index();
201 Node* phi = phis.node();
202 if (idx >= phi->req()) {
203 phis.pop();
204 } else {
205 current = phi->in(idx);
206 phis.set_index(idx+1);
207 }
208 }
209 if (current == NULL) {
210 return true;
211 }
212 } else if (current == phase->C->immutable_memory()) {
213 return false;
214 } else if (current->isa_Phi()) {
215 if (!linear) {
216 return false;
217 }
218 phis.push(current, 2);
219 current = current->in(1);
220 } else if (current->Opcode() == Op_ShenandoahWriteBarrier) {
221 const Type* in_type = current->bottom_type();
222 const Type* this_type = b2->bottom_type();
223 if (is_independent(in_type, this_type)) {
224 current = current->in(Memory);
225 } else {
226 return false;
227 }
228 } else if (current->Opcode() == Op_ShenandoahWBMemProj) {
229 current = current->in(0);
230 } else if (current->is_Proj()) {
231 current = current->in(0);
232 } else if (current->is_Call()) {
233 return false; // TODO: Maybe improve by looking at the call's memory effects?
234 } else if (current->is_MemBar()) {
235 return false; // TODO: Do we need to stop at *any* membar?
236 } else if (current->is_MergeMem()) {
237 // if (true) return false;
238 // tty->print_cr("current == mergemem: "); current->dump();
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 // tty->print_cr("what else can we see here:");
244 #ifdef ASSERT
245 current->dump();
246 #endif
247 ShouldNotReachHere();
248 return false;
249 }
250 }
251 return false;
252 }
253
254 bool ShenandoahReadBarrierNode::is_independent(Node* mem) {
255 if (mem->is_Phi() || mem->is_Proj() || mem->is_MergeMem()) {
256 return true;
257 } else if (mem->Opcode() == Op_ShenandoahWriteBarrier) {
258 const Type* mem_type = mem->bottom_type();
259 const Type* this_type = bottom_type();
260 if (is_independent(mem_type, this_type)) {
261 return true;
262 } else {
263 return false;
264 }
265 } else if (mem->is_Call() || mem->is_MemBar()) {
266 return false;
267 }
268 #ifdef ASSERT
269 mem->dump();
270 #endif
271 ShouldNotReachHere();
272 return true;
273 }
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 if ((! in_type->isa_oopptr()) || (! this_type->isa_oopptr())) {
285 #ifdef ASSERT
286 tty->print_cr("not oopptr");
287 tty->print("in: "); in_type->dump(); tty->print_cr(" ");
288 tty->print("this: "); this_type->dump(); tty->print_cr(" ");
289 #endif
290 return false;
291 }
292 */
293
294 ciKlass* in_kls = in_type->is_oopptr()->klass();
295 ciKlass* this_kls = this_type->is_oopptr()->klass();
296 if (in_kls != NULL && this_kls != NULL &&
297 in_kls->is_loaded() && this_kls->is_loaded() &&
298 (!in_kls->is_subclass_of(this_kls)) &&
299 (!this_kls->is_subclass_of(in_kls))) {
300 #ifdef ASSERT
301 // tty->print_cr("independent: ");
302 // tty->print("in: "); in_kls->print(); tty->print_cr(" ");
303 // tty->print("this: "); this_kls->print(); tty->print_cr(" ");
304 #endif
305 return true;
306 }
307 #ifdef ASSERT
308 // tty->print_cr("possibly dependend?");
309 // tty->print("in: "); in_type->dump(); tty->print_cr(" ");
310 // tty->print("this: "); this_type->dump(); tty->print_cr(" ");
311 #endif
312 return false;
313 }
314
315 Node* ShenandoahReadBarrierNode::Ideal(PhaseGVN *phase, bool can_reshape) {
316
317 if (! can_reshape) {
318 return NULL;
319 }
320
321 if (in(Memory) == phase->C->immutable_memory()) return NULL;
322
323 // If memory input is a MergeMem, take the appropriate slice out of it.
324 Node* mem_in = in(Memory);
325 if (mem_in->isa_MergeMem()) {
326 const TypePtr* adr_type = brooks_pointer_type(bottom_type());
327 uint alias_idx = phase->C->get_alias_index(adr_type);
328 mem_in = mem_in->as_MergeMem()->memory_at(alias_idx);
329 set_req(Memory, mem_in);
330 return this;
331 }
332
333 Node* input = in(Memory);
334 if (input->Opcode() == Op_ShenandoahWBMemProj) {
335 ResourceMark rm;
336 VectorSet seen(Thread::current()->resource_area());
337 Node* n = in(Memory);
338 while (n->Opcode() == Op_ShenandoahWBMemProj &&
339 n->in(0) != NULL &&
340 n->in(0)->Opcode() == Op_ShenandoahWriteBarrier &&
341 n->in(0)->in(Memory) != NULL) {
342 if (seen.test_set(n->_idx)) {
343 return NULL; // loop
344 }
345 n = n->in(0)->in(Memory);
346 }
347
348 Node* wb = input->in(0);
349 const Type* in_type = phase->type(wb);
350 // is_top() test not sufficient here: we can come here after CCP
351 // in a dead branch of the graph that has not yet been removed.
352 if (in_type == Type::TOP) return NULL; // Dead path.
353 assert(wb->Opcode() == Op_ShenandoahWriteBarrier, "expect write barrier");
354 if (is_independent(in_type, _type)) {
355 phase->igvn_rehash_node_delayed(wb);
356 set_req(Memory, wb->in(Memory));
357 if (can_reshape && input->outcnt() == 0) {
358 phase->is_IterGVN()->_worklist.push(input);
359 }
360 return this;
361 }
362 }
363 return NULL;
364 }
365
366 ShenandoahWriteBarrierNode::ShenandoahWriteBarrierNode(Compile* C, Node* ctrl, Node* mem, Node* obj)
367 : ShenandoahBarrierNode(ctrl, mem, obj, false) {
368 assert(UseShenandoahGC && ShenandoahWriteBarrier, "should be enabled");
369 ShenandoahBarrierSetC2::bsc2()->state()->add_shenandoah_barrier(this);
370 }
371
372
373 Node* ShenandoahWriteBarrierNode::Identity(PhaseGVN* phase) {
374 assert(in(0) != NULL, "should have control");
375 PhaseIterGVN* igvn = phase->is_IterGVN();
376 Node* mem_in = in(Memory);
377 Node* mem_proj = NULL;
378
379 if (igvn != NULL) {
380 mem_proj = find_out_with(Op_ShenandoahWBMemProj);
381 if (mem_proj == NULL || mem_in == mem_proj) {
382 return this;
383 }
384 }
385
386 Node* replacement = Identity_impl(phase);
387 if (igvn != NULL) {
388 if (replacement != NULL && replacement != this) {
389 igvn->replace_node(mem_proj, mem_in);
390 }
391 }
392 return replacement;
393 }
394
395
396 Node* ShenandoahWriteBarrierNode::Ideal(PhaseGVN *phase, bool can_reshape) {
397 assert(in(0) != NULL, "should have control");
398 if (!can_reshape) {
399 return NULL;
400 }
401
402 PhaseIterGVN* igvn = phase->is_IterGVN();
403 Node* mem_in = in(Memory);
404
405 if (mem_in->isa_MergeMem()) {
406 const TypePtr* adr_type = brooks_pointer_type(bottom_type());
407 uint alias_idx = phase->C->get_alias_index(adr_type);
408 mem_in = mem_in->as_MergeMem()->memory_at(alias_idx);
409 set_req(Memory, mem_in);
410 return this;
411 }
412
413 Node* val = in(ValueIn);
414 if (val->is_ShenandoahBarrier()) {
415 set_req(ValueIn, val->in(ValueIn));
416 return this;
417 }
418
419 return NULL;
420 }
421
422 bool ShenandoahWriteBarrierNode::expand(Compile* C, PhaseIterGVN& igvn, int& loop_opts_cnt) {
423 if (UseShenandoahGC) {
424 if (ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count() > 0 || (!ShenandoahWriteBarrier && ShenandoahStoreValEnqueueBarrier)) {
425 bool attempt_more_loopopts = ShenandoahLoopOptsAfterExpansion;
426 C->clear_major_progress();
427 PhaseIdealLoop ideal_loop(igvn, LoopOptsShenandoahExpand);
428 if (C->failing()) return false;
429 PhaseIdealLoop::verify(igvn);
430 DEBUG_ONLY(ShenandoahBarrierNode::verify_raw_mem(C->root());)
431 if (attempt_more_loopopts) {
432 C->set_major_progress();
433 if (!C->optimize_loops(loop_opts_cnt, igvn, LoopOptsShenandoahPostExpand)) {
434 return false;
435 }
436 C->clear_major_progress();
437 }
438 }
439 }
440 return true;
441 }
442
443 bool ShenandoahWriteBarrierNode::is_heap_state_test(Node* iff, int mask) {
444 if (!UseShenandoahGC) {
445 return false;
446 }
447 assert(iff->is_If(), "bad input");
448 if (iff->Opcode() != Op_If) {
449 return false;
450 }
451 Node* bol = iff->in(1);
452 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
453 return false;
454 }
455 Node* cmp = bol->in(1);
456 if (cmp->Opcode() != Op_CmpI) {
457 return false;
458 }
459 Node* in1 = cmp->in(1);
460 Node* in2 = cmp->in(2);
461 if (in2->find_int_con(-1) != 0) {
462 return false;
463 }
464 if (in1->Opcode() != Op_AndI) {
465 return false;
466 }
467 in2 = in1->in(2);
468 if (in2->find_int_con(-1) != mask) {
469 return false;
470 }
471 in1 = in1->in(1);
472
473 return is_gc_state_load(in1);
474 }
475
476
477 bool ShenandoahWriteBarrierNode::is_evacuation_in_progress_test(Node* iff) {
478 return is_heap_state_test(iff, ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL);
479 }
480
481 bool ShenandoahWriteBarrierNode::is_heap_stable_test(Node* iff) {
482 return is_heap_state_test(iff, ShenandoahHeap::HAS_FORWARDED);
483 }
484
485 bool ShenandoahWriteBarrierNode::is_gc_state_load(Node *n) {
486 if (!UseShenandoahGC) {
487 return false;
488 }
489 if (n->Opcode() != Op_LoadB) {
490 return false;
491 }
492 Node* addp = n->in(MemNode::Address);
493 if (!addp->is_AddP()) {
494 return false;
495 }
496 Node* base = addp->in(AddPNode::Address);
497 Node* off = addp->in(AddPNode::Offset);
498 if (base->Opcode() != Op_ThreadLocal) {
499 return false;
500 }
501 if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) {
502 return false;
503 }
504 return true;
505 }
506
507 bool ShenandoahWriteBarrierNode::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) {
508 assert(phase->is_dominator(stop, start), "bad inputs");
509 ResourceMark rm;
510 Unique_Node_List wq;
511 wq.push(start);
512 for (uint next = 0; next < wq.size(); next++) {
513 Node *m = wq.at(next);
514 if (m == stop) {
515 continue;
516 }
517 if (m->is_SafePoint() && !m->is_CallLeaf()) {
518 return true;
519 }
520 if (m->is_Region()) {
521 for (uint i = 1; i < m->req(); i++) {
522 wq.push(m->in(i));
523 }
524 } else {
525 wq.push(m->in(0));
526 }
527 }
528 return false;
529 }
530
531 bool ShenandoahWriteBarrierNode::try_common_gc_state_load(Node *n, PhaseIdealLoop *phase) {
532 assert(is_gc_state_load(n), "inconsistent");
533 Node* addp = n->in(MemNode::Address);
534 Node* dominator = NULL;
535 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) {
536 Node* u = addp->fast_out(i);
537 assert(is_gc_state_load(u), "inconsistent");
538 if (u != n && phase->is_dominator(u->in(0), n->in(0))) {
539 if (dominator == NULL) {
540 dominator = u;
541 } else {
542 if (phase->dom_depth(u->in(0)) < phase->dom_depth(dominator->in(0))) {
543 dominator = u;
544 }
545 }
546 }
547 }
548 if (dominator == NULL || has_safepoint_between(n->in(0), dominator->in(0), phase)) {
549 return false;
550 }
551 phase->igvn().replace_node(n, dominator);
552
553 return true;
554 }
555
556 Node* ShenandoahWriteBarrierNode::evacuation_in_progress_test_ctrl(Node* iff) {
557 assert(is_evacuation_in_progress_test(iff), "bad input");
558 return iff->in(0);
559 }
560
561 bool ShenandoahBarrierNode::dominates_memory_impl(PhaseGVN* phase,
562 Node* b1,
563 Node* b2,
564 Node* current,
565 bool linear) {
566 ResourceMark rm;
567 VectorSet visited(Thread::current()->resource_area());
568 Node_Stack phis(0);
569
570
571 for(int i = 0; i < 10; i++) {
572 if (current == NULL) {
573 return false;
574 } else if (visited.test_set(current->_idx) || current->is_top() || current == b1) {
575 current = NULL;
576 while (phis.is_nonempty() && current == NULL) {
577 uint idx = phis.index();
578 Node* phi = phis.node();
579 if (idx >= phi->req()) {
580 phis.pop();
581 } else {
582 current = phi->in(idx);
583 phis.set_index(idx+1);
584 }
585 }
586 if (current == NULL) {
587 return true;
588 }
589 } else if (current == b2) {
590 return false;
591 } else if (current == phase->C->immutable_memory()) {
592 return false;
593 } else if (current->isa_Phi()) {
594 if (!linear) {
595 return false;
596 }
597 phis.push(current, 2);
598 current = current->in(1);
599 } else if (current->Opcode() == Op_ShenandoahWriteBarrier) {
600 current = current->in(Memory);
601 } else if (current->Opcode() == Op_ShenandoahWBMemProj) {
602 current = current->in(0);
603 } else if (current->is_Proj()) {
604 current = current->in(0);
605 } else if (current->is_Call()) {
606 current = current->in(TypeFunc::Memory);
607 } else if (current->is_MemBar()) {
608 current = current->in(TypeFunc::Memory);
609 } else if (current->is_MergeMem()) {
610 const TypePtr* adr_type = brooks_pointer_type(phase->type(b2));
611 uint alias_idx = phase->C->get_alias_index(adr_type);
612 current = current->as_MergeMem()->memory_at(alias_idx);
613 } else {
614 #ifdef ASSERT
615 current->dump();
616 #endif
617 ShouldNotReachHere();
618 return false;
619 }
620 }
621 return false;
622 }
623
624 /**
625 * Determines if b1 dominates b2 through memory inputs. It returns true if:
626 * - b1 can be reached by following each branch in b2's memory input (through phis, etc)
627 * - or we get back to b2 (i.e. through a loop) without seeing b1
628 * In all other cases, (in particular, if we reach immutable_memory without having seen b1)
629 * we return false.
630 */
631 bool ShenandoahBarrierNode::dominates_memory(PhaseGVN* phase, Node* b1, Node* b2, bool linear) {
632 return dominates_memory_impl(phase, b1, b2, b2->in(Memory), linear);
633 }
634
635 Node* ShenandoahBarrierNode::Identity_impl(PhaseGVN* phase) {
636 Node* n = in(ValueIn);
637
638 Node* rb_mem = Opcode() == Op_ShenandoahReadBarrier ? in(Memory) : NULL;
639 if (! needs_barrier(phase, this, n, rb_mem, _allow_fromspace)) {
640 return n;
641 }
642
643 // tty->print_cr("find sibling for: "); dump(2);
644 // Try to find a write barrier sibling with identical inputs that we can fold into.
645 for (DUIterator i = n->outs(); n->has_out(i); i++) {
646 Node* sibling = n->out(i);
647 if (sibling == this) {
648 continue;
649 }
650 /*
651 assert(sibling->Opcode() != Op_ShenandoahWriteBarrier ||
652 Opcode() != Op_ShenandoahWriteBarrier || hash() == sibling->hash(),
653 "if this is a write barrier, then sibling can't be write barrier too");
654 */
655 if (sibling->Opcode() != Op_ShenandoahWriteBarrier) {
656 continue;
657 }
658 /*
659 if (sibling->outcnt() == 0) {
660 // Some dead node.
661 continue;
662 }
663 */
664 assert(sibling->in(ValueIn) == in(ValueIn), "sanity");
665 assert(sibling->Opcode() == Op_ShenandoahWriteBarrier, "sanity");
666 // tty->print_cr("candidate: "); sibling->dump();
667
668 if (dominates_memory(phase, sibling, this, phase->is_IterGVN() == NULL)) {
669
670 /*
671 tty->print_cr("matched barrier:");
672 sibling->dump();
673 tty->print_cr("for: ");
674 dump();
675 */
676 return sibling;
677 }
678
679 /*
680 tty->print_cr("couldn't match candidate:");
681 sibling->dump(2);
682 */
683 }
684 /*
685 tty->print_cr("couldn't match barrier to any:");
686 dump();
687 */
688 return this;
689 }
690
691 #ifndef PRODUCT
692 void ShenandoahBarrierNode::dump_spec(outputStream *st) const {
693 const TypePtr* adr = adr_type();
694 if (adr == NULL) {
695 return;
696 }
697 st->print(" @");
698 adr->dump_on(st);
699 st->print(" (");
700 Compile::current()->alias_type(adr)->adr_type()->dump_on(st);
701 st->print(") ");
702 }
703 #endif
704
705 Node* ShenandoahReadBarrierNode::Identity(PhaseGVN* phase) {
706
707 // if (true) return this;
708
709 // tty->print("optimizing rb: "); dump();
710 Node* id = Identity_impl(phase);
711
712 if (id == this && phase->is_IterGVN()) {
713 Node* n = in(ValueIn);
714 // No success in super call. Try to combine identical read barriers.
715 for (DUIterator i = n->outs(); n->has_out(i); i++) {
716 Node* sibling = n->out(i);
717 if (sibling == this || sibling->Opcode() != Op_ShenandoahReadBarrier) {
718 continue;
719 }
720 assert(sibling->in(ValueIn) == in(ValueIn), "sanity");
721 if (phase->is_IterGVN()->hash_find(sibling) &&
722 sibling->bottom_type() == bottom_type() &&
723 sibling->in(Control) == in(Control) &&
724 dominates_memory_rb(phase, sibling, this, phase->is_IterGVN() == NULL)) {
725 /*
726 if (in(Memory) != sibling->in(Memory)) {
727 tty->print_cr("interesting rb-fold");
728 dump();
729 sibling->dump();
730 }
731 */
732 return sibling;
733 }
734 }
735 }
736 return id;
737 }
738
739 const Type* ShenandoahBarrierNode::Value(PhaseGVN* phase) const {
740 // Either input is TOP ==> the result is TOP
741 const Type *t1 = phase->type(in(Memory));
742 if (t1 == Type::TOP) return Type::TOP;
743 const Type *t2 = phase->type(in(ValueIn));
744 if( t2 == Type::TOP ) return Type::TOP;
745
746 if (t2 == TypePtr::NULL_PTR) {
747 return _type;
748 }
749
750 const Type* type = t2->is_oopptr()->cast_to_nonconst();
751 return type;
752 }
753
754 uint ShenandoahBarrierNode::hash() const {
755 return TypeNode::hash() + _allow_fromspace;
756 }
757
758 uint ShenandoahBarrierNode::cmp(const Node& n) const {
759 return _allow_fromspace == ((ShenandoahBarrierNode&) n)._allow_fromspace
760 && TypeNode::cmp(n);
761 }
762
763 uint ShenandoahBarrierNode::size_of() const {
764 return sizeof(*this);
765 }
766
767 Node* ShenandoahWBMemProjNode::Identity(PhaseGVN* phase) {
768
769 Node* wb = in(0);
770 if (wb->is_top()) return phase->C->top(); // Dead path.
771
772 assert(wb->Opcode() == Op_ShenandoahWriteBarrier, "expect write barrier");
773 PhaseIterGVN* igvn = phase->is_IterGVN();
774 // We can't do the below unless the graph is fully constructed.
775 if (igvn == NULL) {
776 return this;
777 }
778
779 // If the mem projection has no barrier users, it's not needed anymore.
780 if (wb->outcnt() == 1) {
781 return wb->in(ShenandoahBarrierNode::Memory);
782 }
783
784 return this;
785 }
786
787 #ifdef ASSERT
788 bool ShenandoahBarrierNode::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) {
789 assert(phis.size() == 0, "");
790
791 while (true) {
792 if (in->bottom_type() == TypePtr::NULL_PTR) {
793 if (trace) {tty->print_cr("NULL");}
794 } else if (!in->bottom_type()->make_ptr()->make_oopptr()) {
795 if (trace) {tty->print_cr("Non oop");}
796 } else if (t == ShenandoahLoad && ShenandoahOptimizeStableFinals &&
797 in->bottom_type()->make_ptr()->isa_aryptr() &&
798 in->bottom_type()->make_ptr()->is_aryptr()->is_stable()) {
799 if (trace) {tty->print_cr("Stable array load");}
800 } else {
801 if (in->is_ConstraintCast()) {
802 in = in->in(1);
803 continue;
804 } else if (in->is_AddP()) {
805 assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access");
806 in = in->in(AddPNode::Address);
807 continue;
808 } else if (in->is_Con()) {
809 if (trace) {tty->print("Found constant"); in->dump();}
810 } else if (in->is_ShenandoahBarrier()) {
811 if (t == ShenandoahOopStore) {
812 if (in->Opcode() != Op_ShenandoahWriteBarrier) {
813 return false;
814 }
815 uint i = 0;
816 for (; i < phis.size(); i++) {
817 Node* n = phis.node_at(i);
818 if (n->Opcode() == Op_ShenandoahEnqueueBarrier) {
819 break;
820 }
821 }
822 if (i == phis.size()) {
823 return false;
824 }
825 } else if (t == ShenandoahStore && in->Opcode() != Op_ShenandoahWriteBarrier) {
826 return false;
827 }
828 barriers_used.push(in);
829 if (trace) {tty->print("Found barrier"); in->dump();}
830 } else if (in->Opcode() == Op_ShenandoahEnqueueBarrier) {
831 if (t != ShenandoahOopStore) {
832 return false;
833 }
834 if (trace) {tty->print("Found enqueue barrier"); in->dump();}
835 phis.push(in, in->req());
836 in = in->in(1);
837 continue;
838 } else if (in->is_Proj() && in->in(0)->is_Allocate()) {
839 if (trace) {tty->print("Found alloc"); in->in(0)->dump();}
840 } else if (in->is_Phi()) {
841 if (!visited.test_set(in->_idx)) {
842 if (trace) {tty->print("Pushed phi:"); in->dump();}
843 phis.push(in, 2);
844 in = in->in(1);
845 continue;
846 }
847 if (trace) {tty->print("Already seen phi:"); in->dump();}
848 } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) {
849 if (!visited.test_set(in->_idx)) {
850 if (trace) {tty->print("Pushed cmovep:"); in->dump();}
851 phis.push(in, CMoveNode::IfTrue);
852 in = in->in(CMoveNode::IfFalse);
853 continue;
854 }
855 if (trace) {tty->print("Already seen cmovep:"); in->dump();}
856 } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) {
857 in = in->in(1);
858 continue;
859 } else {
860 return false;
861 }
862 }
863 bool cont = false;
864 while (phis.is_nonempty()) {
865 uint idx = phis.index();
866 Node* phi = phis.node();
867 if (idx >= phi->req()) {
868 if (trace) {tty->print("Popped phi:"); phi->dump();}
869 phis.pop();
870 continue;
871 }
872 if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();}
873 in = phi->in(idx);
874 phis.set_index(idx+1);
875 cont = true;
876 break;
877 }
878 if (!cont) {
879 break;
880 }
881 }
882 return true;
883 }
884
885 void ShenandoahBarrierNode::report_verify_failure(const char *msg, Node *n1, Node *n2) {
886 if (n1 != NULL) {
887 n1->dump(+10);
888 }
889 if (n2 != NULL) {
890 n2->dump(+10);
891 }
892 fatal("%s", msg);
893 }
894
895 void ShenandoahBarrierNode::verify(RootNode* root) {
896 ResourceMark rm;
897 Unique_Node_List wq;
898 GrowableArray<Node*> barriers;
899 Unique_Node_List barriers_used;
900 Node_Stack phis(0);
901 VectorSet visited(Thread::current()->resource_area());
902 const bool trace = false;
903 const bool verify_no_useless_barrier = false;
904
905 wq.push(root);
906 for (uint next = 0; next < wq.size(); next++) {
907 Node *n = wq.at(next);
908 if (n->is_Load()) {
909 const bool trace = false;
910 if (trace) {tty->print("Verifying"); n->dump();}
911 if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) {
912 if (trace) {tty->print_cr("Load range/klass");}
913 } else {
914 const TypePtr* adr_type = n->as_Load()->adr_type();
915
916 if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) {
917 if (trace) {tty->print_cr("Mark load");}
918 } else if (adr_type->isa_instptr() &&
919 adr_type->is_instptr()->klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
920 adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset) {
921 if (trace) {tty->print_cr("Reference.get()");}
922 } else {
923 bool verify = true;
924 if (adr_type->isa_instptr()) {
925 const TypeInstPtr* tinst = adr_type->is_instptr();
926 ciKlass* k = tinst->klass();
927 assert(k->is_instance_klass(), "");
928 ciInstanceKlass* ik = (ciInstanceKlass*)k;
929 int offset = adr_type->offset();
930
931 if ((ik->debug_final_field_at(offset) && ShenandoahOptimizeInstanceFinals) ||
932 (ik->debug_stable_field_at(offset) && ShenandoahOptimizeStableFinals)) {
933 if (trace) {tty->print_cr("Final/stable");}
934 verify = false;
935 } else if (k == ciEnv::current()->Class_klass() &&
936 tinst->const_oop() != NULL &&
937 tinst->offset() >= (ik->size_helper() * wordSize)) {
938 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
939 ciField* field = k->get_field_by_offset(tinst->offset(), true);
940 if ((ShenandoahOptimizeStaticFinals && field->is_final()) ||
941 (ShenandoahOptimizeStableFinals && field->is_stable())) {
942 verify = false;
943 }
944 }
945 }
946
947 if (verify && !ShenandoahBarrierNode::verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) {
948 report_verify_failure("Shenandoah verification: Load should have barriers", n);
949 }
950 }
951 }
952 } else if (n->is_Store()) {
953 const bool trace = false;
954
955 if (trace) {tty->print("Verifying"); n->dump();}
956 if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) {
957 Node* adr = n->in(MemNode::Address);
958 bool verify = true;
959
960 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) {
961 adr = adr->in(AddPNode::Address);
962 if (adr->is_AddP()) {
963 assert(adr->in(AddPNode::Base)->is_top(), "");
964 adr = adr->in(AddPNode::Address);
965 if (adr->Opcode() == Op_LoadP &&
966 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() &&
967 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
968 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) {
969 if (trace) {tty->print_cr("SATB prebarrier");}
970 verify = false;
971 }
972 }
973 }
974
975 if (verify && !ShenandoahBarrierNode::verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahStoreValEnqueueBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
976 report_verify_failure("Shenandoah verification: Store should have barriers", n);
977 }
978 }
979 if (!ShenandoahBarrierNode::verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
980 report_verify_failure("Shenandoah verification: Store (address) should have barriers", n);
981 }
982 } else if (n->Opcode() == Op_CmpP) {
983 const bool trace = false;
984
985 Node* in1 = n->in(1);
986 Node* in2 = n->in(2);
987 if (in1->bottom_type()->isa_oopptr()) {
988 if (trace) {tty->print("Verifying"); n->dump();}
989
990 bool mark_inputs = false;
991 if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR ||
992 (in1->is_Con() || in2->is_Con())) {
993 if (trace) {tty->print_cr("Comparison against a constant");}
994 mark_inputs = true;
995 } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) ||
996 (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) {
997 if (trace) {tty->print_cr("Comparison with newly alloc'ed object");}
998 mark_inputs = true;
999 } else {
1000 assert(in2->bottom_type()->isa_oopptr(), "");
1001
1002 if (!ShenandoahBarrierNode::verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) ||
1003 !ShenandoahBarrierNode::verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) {
1004 report_verify_failure("Shenandoah verification: Cmp should have barriers", n);
1005 }
1006 }
1007 if (verify_no_useless_barrier &&
1008 mark_inputs &&
1009 (!ShenandoahBarrierNode::verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) ||
1010 !ShenandoahBarrierNode::verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) {
1011 phis.clear();
1012 visited.Reset();
1013 }
1014 }
1015 } else if (n->is_LoadStore()) {
1016 if (n->in(MemNode::ValueIn)->bottom_type()->isa_ptr() &&
1017 !ShenandoahBarrierNode::verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahLoad, trace, barriers_used)) {
1018 report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n);
1019 }
1020
1021 if (n->in(MemNode::Address)->bottom_type()->isa_oopptr() && !ShenandoahBarrierNode::verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
1022 report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n);
1023 }
1024 } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) {
1025 CallNode* call = n->as_Call();
1026
1027 static struct {
1028 const char* name;
1029 struct {
1030 int pos;
1031 verify_type t;
1032 } args[6];
1033 } calls[] = {
1034 "aescrypt_encryptBlock",
1035 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
1036 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1037 "aescrypt_decryptBlock",
1038 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
1039 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1040 "multiplyToLen",
1041 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { TypeFunc::Parms+4, ShenandoahStore },
1042 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1043 "squareToLen",
1044 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { -1, ShenandoahNone},
1045 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1046 "montgomery_multiply",
1047 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
1048 { TypeFunc::Parms+6, ShenandoahStore }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1049 "montgomery_square",
1050 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+5, ShenandoahStore },
1051 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1052 "mulAdd",
1053 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
1054 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1055 "vectorizedMismatch",
1056 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
1057 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1058 "updateBytesCRC32",
1059 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
1060 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1061 "updateBytesAdler32",
1062 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
1063 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1064 "updateBytesCRC32C",
1065 { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad}, { -1, ShenandoahNone},
1066 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1067 "counterMode_AESCrypt",
1068 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
1069 { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } },
1070 "cipherBlockChaining_encryptAESCrypt",
1071 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
1072 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1073 "cipherBlockChaining_decryptAESCrypt",
1074 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
1075 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1076 "shenandoah_clone_barrier",
1077 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
1078 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1079 "ghash_processBlocks",
1080 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
1081 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1082 "sha1_implCompress",
1083 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1084 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1085 "sha256_implCompress",
1086 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1087 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1088 "sha512_implCompress",
1089 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1090 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1091 "sha1_implCompressMB",
1092 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1093 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1094 "sha256_implCompressMB",
1095 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1096 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1097 "sha512_implCompressMB",
1098 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1099 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1100 };
1101
1102 if (call->is_call_to_arraycopystub()) {
1103 Node* dest = NULL;
1104 const TypeTuple* args = n->as_Call()->_tf->domain();
1105 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
1106 if (args->field_at(i)->isa_ptr()) {
1107 j++;
1108 if (j == 2) {
1109 dest = n->in(i);
1110 break;
1111 }
1112 }
1113 }
1114 if (!ShenandoahBarrierNode::verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) ||
1115 !ShenandoahBarrierNode::verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) {
1116 report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n);
1117 }
1118 } else if (strlen(call->_name) > 5 &&
1119 !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) {
1120 if (!ShenandoahBarrierNode::verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) {
1121 report_verify_failure("Shenandoah verification: _fill should have barriers", n);
1122 }
1123 } else if (!strcmp(call->_name, "shenandoah_wb_pre")) {
1124 // skip
1125 } else {
1126 const int calls_len = sizeof(calls) / sizeof(calls[0]);
1127 int i = 0;
1128 for (; i < calls_len; i++) {
1129 if (!strcmp(calls[i].name, call->_name)) {
1130 break;
1131 }
1132 }
1133 if (i != calls_len) {
1134 const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]);
1135 for (uint j = 0; j < args_len; j++) {
1136 int pos = calls[i].args[j].pos;
1137 if (pos == -1) {
1138 break;
1139 }
1140 if (!ShenandoahBarrierNode::verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) {
1141 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
1142 }
1143 }
1144 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
1145 if (call->in(j)->bottom_type()->make_ptr() &&
1146 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
1147 uint k = 0;
1148 for (; k < args_len && calls[i].args[k].pos != (int)j; k++);
1149 if (k == args_len) {
1150 fatal("arg %d for call %s not covered", j, call->_name);
1151 }
1152 }
1153 }
1154 } else {
1155 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
1156 if (call->in(j)->bottom_type()->make_ptr() &&
1157 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
1158 fatal("%s not covered", call->_name);
1159 }
1160 }
1161 }
1162 }
1163 } else if (n->is_ShenandoahBarrier()) {
1164 assert(!barriers.contains(n), "");
1165 assert(n->Opcode() != Op_ShenandoahWriteBarrier || n->find_out_with(Op_ShenandoahWBMemProj) != NULL, "bad shenandoah write barrier");
1166 assert(n->Opcode() != Op_ShenandoahWriteBarrier || n->outcnt() > 1, "bad shenandoah write barrier");
1167 barriers.push(n);
1168 } else if (n->Opcode() == Op_ShenandoahEnqueueBarrier) {
1169 // skip
1170 } else if (n->is_AddP()
1171 || n->is_Phi()
1172 || n->is_ConstraintCast()
1173 || n->Opcode() == Op_Return
1174 || n->Opcode() == Op_CMoveP
1175 || n->Opcode() == Op_CMoveN
1176 || n->Opcode() == Op_Rethrow
1177 || n->is_MemBar()
1178 || n->Opcode() == Op_Conv2B
1179 || n->Opcode() == Op_SafePoint
1180 || n->is_CallJava()
1181 || n->Opcode() == Op_Unlock
1182 || n->Opcode() == Op_EncodeP
1183 || n->Opcode() == Op_DecodeN) {
1184 // nothing to do
1185 } else {
1186 static struct {
1187 int opcode;
1188 struct {
1189 int pos;
1190 verify_type t;
1191 } inputs[2];
1192 } others[] = {
1193 Op_FastLock,
1194 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
1195 Op_Lock,
1196 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone} },
1197 Op_ArrayCopy,
1198 { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } },
1199 Op_StrCompressedCopy,
1200 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
1201 Op_StrInflatedCopy,
1202 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
1203 Op_AryEq,
1204 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
1205 Op_StrIndexOf,
1206 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
1207 Op_StrComp,
1208 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
1209 Op_StrEquals,
1210 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
1211 Op_EncodeISOArray,
1212 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
1213 Op_HasNegatives,
1214 { { 2, ShenandoahLoad }, { -1, ShenandoahNone} },
1215 Op_CastP2X,
1216 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
1217 Op_StrIndexOfChar,
1218 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } },
1219 };
1220
1221 const int others_len = sizeof(others) / sizeof(others[0]);
1222 int i = 0;
1223 for (; i < others_len; i++) {
1224 if (others[i].opcode == n->Opcode()) {
1225 break;
1226 }
1227 }
1228 uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req();
1229 if (i != others_len) {
1230 const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]);
1231 for (uint j = 0; j < inputs_len; j++) {
1232 int pos = others[i].inputs[j].pos;
1233 if (pos == -1) {
1234 break;
1235 }
1236 if (!ShenandoahBarrierNode::verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) {
1237 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
1238 }
1239 }
1240 for (uint j = 1; j < stop; j++) {
1241 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
1242 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
1243 uint k = 0;
1244 for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++);
1245 if (k == inputs_len) {
1246 fatal("arg %d for node %s not covered", j, n->Name());
1247 }
1248 }
1249 }
1250 } else {
1251 for (uint j = 1; j < stop; j++) {
1252 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
1253 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
1254 fatal("%s not covered", n->Name());
1255 }
1256 }
1257 }
1258 }
1259
1260 if (n->is_SafePoint()) {
1261 SafePointNode* sfpt = n->as_SafePoint();
1262 if (verify_no_useless_barrier && sfpt->jvms() != NULL) {
1263 for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) {
1264 if (!ShenandoahBarrierNode::verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) {
1265 phis.clear();
1266 visited.Reset();
1267 }
1268 }
1269 }
1270 }
1271 for( uint i = 0; i < n->len(); ++i ) {
1272 Node *m = n->in(i);
1273 if (m == NULL) continue;
1274
1275 // In most cases, inputs should be known to be non null. If it's
1276 // not the case, it could be a missing cast_not_null() in an
1277 // intrinsic or support might be needed in AddPNode::Ideal() to
1278 // avoid a NULL+offset input.
1279 if (!(n->is_Phi() ||
1280 (n->is_SafePoint() && (!n->is_CallRuntime() || !strcmp(n->as_Call()->_name, "shenandoah_wb_pre") || !strcmp(n->as_Call()->_name, "unsafe_arraycopy"))) ||
1281 n->Opcode() == Op_CmpP ||
1282 n->Opcode() == Op_CmpN ||
1283 (n->Opcode() == Op_StoreP && i == StoreNode::ValueIn) ||
1284 (n->Opcode() == Op_StoreN && i == StoreNode::ValueIn) ||
1285 n->is_ConstraintCast() ||
1286 n->Opcode() == Op_Return ||
1287 n->Opcode() == Op_Conv2B ||
1288 n->is_AddP() ||
1289 n->Opcode() == Op_CMoveP ||
1290 n->Opcode() == Op_CMoveN ||
1291 n->Opcode() == Op_Rethrow ||
1292 n->is_MemBar() ||
1293 n->is_Mem() ||
1294 n->Opcode() == Op_AryEq ||
1295 n->Opcode() == Op_SCMemProj ||
1296 n->Opcode() == Op_EncodeP ||
1297 n->Opcode() == Op_DecodeN ||
1298 n->Opcode() == Op_ShenandoahWriteBarrier ||
1299 n->Opcode() == Op_ShenandoahWBMemProj ||
1300 n->Opcode() == Op_ShenandoahEnqueueBarrier)) {
1301 if (m->bottom_type()->make_oopptr() && m->bottom_type()->make_oopptr()->meet(TypePtr::NULL_PTR) == m->bottom_type()) {
1302 report_verify_failure("Shenandoah verification: null input", n, m);
1303 }
1304 }
1305
1306 wq.push(m);
1307 }
1308 }
1309
1310 if (verify_no_useless_barrier) {
1311 for (int i = 0; i < barriers.length(); i++) {
1312 Node* n = barriers.at(i);
1313 if (!barriers_used.member(n)) {
1314 tty->print("XXX useless barrier"); n->dump(-2);
1315 ShouldNotReachHere();
1316 }
1317 }
1318 }
1319 }
1320 #endif
1321
1322 bool ShenandoahBarrierNode::is_dominator_same_ctrl(Node*c, Node* d, Node* n, PhaseIdealLoop* phase) {
1323 // That both nodes have the same control is not sufficient to prove
1324 // domination, verify that there's no path from d to n
1325 ResourceMark rm;
1326 Unique_Node_List wq;
1327 wq.push(d);
1328 for (uint next = 0; next < wq.size(); next++) {
1329 Node *m = wq.at(next);
1330 if (m == n) {
1331 return false;
1332 }
1333 if (m->is_Phi() && m->in(0)->is_Loop()) {
1334 assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control");
1335 } else {
1336 for (uint i = 0; i < m->req(); i++) {
1337 if (m->in(i) != NULL && phase->ctrl_or_self(m->in(i)) == c) {
1338 wq.push(m->in(i));
1339 }
1340 }
1341 }
1342 }
1343 return true;
1344 }
1345
1346 bool ShenandoahBarrierNode::is_dominator(Node *d_c, Node *n_c, Node* d, Node* n, PhaseIdealLoop* phase) {
1347 if (d_c != n_c) {
1348 return phase->is_dominator(d_c, n_c);
1349 }
1350 return is_dominator_same_ctrl(d_c, d, n, phase);
1351 }
1352
1353 Node* next_mem(Node* mem, int alias) {
1354 Node* res = NULL;
1355 if (mem->is_Proj()) {
1356 res = mem->in(0);
1357 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
1358 res = mem->in(TypeFunc::Memory);
1359 } else if (mem->is_Phi()) {
1360 res = mem->in(1);
1361 } else if (mem->is_ShenandoahBarrier()) {
1362 res = mem->in(ShenandoahBarrierNode::Memory);
1363 } else if (mem->is_MergeMem()) {
1364 res = mem->as_MergeMem()->memory_at(alias);
1365 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
1366 assert(alias = Compile::AliasIdxRaw, "following raw memory can't lead to a barrier");
1367 res = mem->in(MemNode::Memory);
1368 } else {
1369 #ifdef ASSERT
1370 mem->dump();
1371 #endif
1372 ShouldNotReachHere();
1373 }
1374 return res;
1375 }
1376
1377 Node* ShenandoahBarrierNode::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) {
1378 Node* iffproj = NULL;
1379 while (c != dom) {
1380 Node* next = phase->idom(c);
1381 assert(next->unique_ctrl_out() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?");
1382 if (c->is_Region()) {
1383 ResourceMark rm;
1384 Unique_Node_List wq;
1385 wq.push(c);
1386 for (uint i = 0; i < wq.size(); i++) {
1387 Node *n = wq.at(i);
1388 if (n == next) {
1389 continue;
1390 }
1391 if (n->is_Region()) {
1392 for (uint j = 1; j < n->req(); j++) {
1393 wq.push(n->in(j));
1394 }
1395 } else {
1396 wq.push(n->in(0));
1397 }
1398 }
1399 for (uint i = 0; i < wq.size(); i++) {
1400 Node *n = wq.at(i);
1401 assert(n->is_CFG(), "");
1402 if (n->is_Multi()) {
1403 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1404 Node* u = n->fast_out(j);
1405 if (u->is_CFG()) {
1406 if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
1407 return NodeSentinel;
1408 }
1409 }
1410 }
1411 }
1412 }
1413 } else if (c->is_Proj()) {
1414 if (c->is_IfProj()) {
1415 if (c->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) != NULL) {
1416 // continue;
1417 } else {
1418 if (!allow_one_proj) {
1419 return NodeSentinel;
1420 }
1421 if (iffproj == NULL) {
1422 iffproj = c;
1423 } else {
1424 return NodeSentinel;
1425 }
1426 }
1427 } else if (c->Opcode() == Op_JumpProj) {
1428 return NodeSentinel; // unsupported
1429 } else if (c->Opcode() == Op_CatchProj) {
1430 return NodeSentinel; // unsupported
1431 } else if (c->Opcode() == Op_CProj && next->Opcode() == Op_NeverBranch) {
1432 return NodeSentinel; // unsupported
1433 } else {
1434 assert(next->unique_ctrl_out() == c, "unsupported branch pattern");
1435 }
1436 }
1437 c = next;
1438 }
1439 return iffproj;
1440 }
1441
1442 #ifdef ASSERT
1443 void ShenandoahWriteBarrierNode::memory_dominates_all_paths_helper(Node* c, Node* rep_ctrl, Unique_Node_List& controls, PhaseIdealLoop* phase) {
1444 const bool trace = false;
1445 if (trace) { tty->print("X control is"); c->dump(); }
1446
1447 uint start = controls.size();
1448 controls.push(c);
1449 for (uint i = start; i < controls.size(); i++) {
1450 Node *n = controls.at(i);
1451
1452 if (trace) { tty->print("X from"); n->dump(); }
1453
1454 if (n == rep_ctrl) {
1455 continue;
1456 }
1457
1458 if (n->is_Proj()) {
1459 Node* n_dom = n->in(0);
1460 IdealLoopTree* n_dom_loop = phase->get_loop(n_dom);
1461 if (n->is_IfProj() && n_dom->outcnt() == 2) {
1462 n_dom_loop = phase->get_loop(n_dom->as_If()->proj_out(n->as_Proj()->_con == 0 ? 1 : 0));
1463 }
1464 if (n_dom_loop != phase->ltree_root()) {
1465 Node* tail = n_dom_loop->tail();
1466 if (tail->is_Region()) {
1467 for (uint j = 1; j < tail->req(); j++) {
1468 if (phase->is_dominator(n_dom, tail->in(j)) && !phase->is_dominator(n, tail->in(j))) {
1469 assert(phase->is_dominator(rep_ctrl, tail->in(j)), "why are we here?");
1470 // entering loop from below, mark backedge
1471 if (trace) { tty->print("X pushing backedge"); tail->in(j)->dump(); }
1472 controls.push(tail->in(j));
1473 //assert(n->in(0) == n_dom, "strange flow control");
1474 }
1475 }
1476 } else if (phase->get_loop(n) != n_dom_loop && phase->is_dominator(n_dom, tail)) {
1477 // entering loop from below, mark backedge
1478 if (trace) { tty->print("X pushing backedge"); tail->dump(); }
1479 controls.push(tail);
1480 //assert(n->in(0) == n_dom, "strange flow control");
1481 }
1482 }
1483 }
1484
1485 if (n->is_Loop()) {
1486 Node* c = n->in(LoopNode::EntryControl);
1487 if (trace) { tty->print("X pushing"); c->dump(); }
1488 controls.push(c);
1489 } else if (n->is_Region()) {
1490 for (uint i = 1; i < n->req(); i++) {
1491 Node* c = n->in(i);
1492 if (trace) { tty->print("X pushing"); c->dump(); }
1493 controls.push(c);
1494 }
1495 } else {
1496 Node* c = n->in(0);
1497 if (trace) { tty->print("X pushing"); c->dump(); }
1498 controls.push(c);
1499 }
1500 }
1501 }
1502
1503 bool ShenandoahWriteBarrierNode::memory_dominates_all_paths(Node* mem, Node* rep_ctrl, int alias, PhaseIdealLoop* phase) {
1504 const bool trace = false;
1505 if (trace) {
1506 tty->print("XXX mem is"); mem->dump();
1507 tty->print("XXX rep ctrl is"); rep_ctrl->dump();
1508 tty->print_cr("XXX alias is %d", alias);
1509 }
1510 ResourceMark rm;
1511 Unique_Node_List wq;
1512 Unique_Node_List controls;
1513 wq.push(mem);
1514 for (uint next = 0; next < wq.size(); next++) {
1515 Node *nn = wq.at(next);
1516 if (trace) { tty->print("XX from mem"); nn->dump(); }
1517 assert(nn->bottom_type() == Type::MEMORY, "memory only");
1518
1519 if (nn->is_Phi()) {
1520 Node* r = nn->in(0);
1521 for (DUIterator_Fast jmax, j = r->fast_outs(jmax); j < jmax; j++) {
1522 Node* u = r->fast_out(j);
1523 if (u->is_Phi() && u->bottom_type() == Type::MEMORY && u != nn &&
1524 (u->adr_type() == TypePtr::BOTTOM || phase->C->get_alias_index(u->adr_type()) == alias)) {
1525 if (trace) { tty->print("XX Next mem (other phi)"); u->dump(); }
1526 wq.push(u);
1527 }
1528 }
1529 }
1530
1531 for (DUIterator_Fast imax, i = nn->fast_outs(imax); i < imax; i++) {
1532 Node* use = nn->fast_out(i);
1533
1534 if (trace) { tty->print("XX use %p", use->adr_type()); use->dump(); }
1535 if (use->is_CFG() && use->in(TypeFunc::Memory) == nn) {
1536 Node* c = use->in(0);
1537 if (phase->is_dominator(rep_ctrl, c)) {
1538 memory_dominates_all_paths_helper(c, rep_ctrl, controls, phase);
1539 } else if (use->is_CallStaticJava() && use->as_CallStaticJava()->uncommon_trap_request() != 0 && c->is_Region()) {
1540 Node* region = c;
1541 if (trace) { tty->print("XX unc region"); region->dump(); }
1542 for (uint j = 1; j < region->req(); j++) {
1543 if (phase->is_dominator(rep_ctrl, region->in(j))) {
1544 if (trace) { tty->print("XX unc follows"); region->in(j)->dump(); }
1545 memory_dominates_all_paths_helper(region->in(j), rep_ctrl, controls, phase);
1546 }
1547 }
1548 }
1549 //continue;
1550 } else if (use->is_Phi()) {
1551 assert(use->bottom_type() == Type::MEMORY, "bad phi");
1552 if ((use->adr_type() == TypePtr::BOTTOM /*&& !shenandoah_has_alias_phi(C, use, alias)*/) ||
1553 phase->C->get_alias_index(use->adr_type()) == alias) {
1554 for (uint j = 1; j < use->req(); j++) {
1555 if (use->in(j) == nn) {
1556 Node* c = use->in(0)->in(j);
1557 if (phase->is_dominator(rep_ctrl, c)) {
1558 memory_dominates_all_paths_helper(c, rep_ctrl, controls, phase);
1559 }
1560 }
1561 }
1562 }
1563 // continue;
1564 }
1565
1566 if (use->is_MergeMem()) {
1567 if (use->as_MergeMem()->memory_at(alias) == nn) {
1568 if (trace) { tty->print("XX Next mem"); use->dump(); }
1569 // follow the memory edges
1570 wq.push(use);
1571 }
1572 } else if (use->is_Phi()) {
1573 assert(use->bottom_type() == Type::MEMORY, "bad phi");
1574 if ((use->adr_type() == TypePtr::BOTTOM /*&& !shenandoah_has_alias_phi(C, use, alias)*/) ||
1575 phase->C->get_alias_index(use->adr_type()) == alias) {
1576 if (trace) { tty->print("XX Next mem"); use->dump(); }
1577 // follow the memory edges
1578 wq.push(use);
1579 }
1580 } else if (use->bottom_type() == Type::MEMORY &&
1581 (use->adr_type() == TypePtr::BOTTOM || phase->C->get_alias_index(use->adr_type()) == alias)) {
1582 if (trace) { tty->print("XX Next mem"); use->dump(); }
1583 // follow the memory edges
1584 wq.push(use);
1585 } else if ((use->is_SafePoint() || use->is_MemBar()) &&
1586 (use->adr_type() == TypePtr::BOTTOM || phase->C->get_alias_index(use->adr_type()) == alias)) {
1587 for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) {
1588 Node* u = use->fast_out(j);
1589 if (u->bottom_type() == Type::MEMORY) {
1590 if (trace) { tty->print("XX Next mem"); u->dump(); }
1591 // follow the memory edges
1592 wq.push(u);
1593 }
1594 }
1595 } else if (use->Opcode() == Op_ShenandoahWriteBarrier && phase->C->get_alias_index(use->adr_type()) == alias) {
1596 Node* m = use->find_out_with(Op_ShenandoahWBMemProj);
1597 if (m != NULL) {
1598 if (trace) { tty->print("XX Next mem"); m->dump(); }
1599 // follow the memory edges
1600 wq.push(m);
1601 }
1602 }
1603 }
1604 }
1605
1606 if (controls.size() == 0) {
1607 return false;
1608 }
1609
1610 for (uint i = 0; i < controls.size(); i++) {
1611 Node *n = controls.at(i);
1612
1613 if (trace) { tty->print("X checking"); n->dump(); }
1614
1615 if (n->unique_ctrl_out() != NULL) {
1616 continue;
1617 }
1618
1619 if (n->Opcode() == Op_NeverBranch) {
1620 Node* taken = n->as_Multi()->proj_out(0);
1621 if (!controls.member(taken)) {
1622 if (trace) { tty->print("X not seen"); taken->dump(); }
1623 return false;
1624 }
1625 continue;
1626 }
1627
1628 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1629 Node* u = n->fast_out(j);
1630
1631 if (u->is_CFG()) {
1632 if (!controls.member(u)) {
1633 if (u->is_Proj() && u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
1634 if (trace) { tty->print("X not seen but unc"); u->dump(); }
1635 } else {
1636 Node* c = u;
1637 do {
1638 c = c->unique_ctrl_out();
1639 } while (c != NULL && c->is_Region());
1640 if (c != NULL && c->Opcode() == Op_Halt) {
1641 if (trace) { tty->print("X not seen but halt"); c->dump(); }
1642 } else {
1643 if (trace) { tty->print("X not seen"); u->dump(); }
1644 return false;
1645 }
1646 }
1647 } else {
1648 if (trace) { tty->print("X seen"); u->dump(); }
1649 }
1650 }
1651 }
1652 }
1653 return true;
1654 }
1655 #endif
1656
1657 Node* ShenandoahBarrierNode::dom_mem(Node* mem, Node*& mem_ctrl, Node* n, Node* rep_ctrl, int alias, PhaseIdealLoop* phase) {
1658 ResourceMark rm;
1659 VectorSet wq(Thread::current()->resource_area());
1660 wq.set(mem->_idx);
1661 mem_ctrl = phase->get_ctrl(mem);
1662 while (!is_dominator(mem_ctrl, rep_ctrl, mem, n, phase)) {
1663 mem = next_mem(mem, alias);
1664 if (wq.test_set(mem->_idx)) {
1665 return NULL; // hit an unexpected loop
1666 }
1667 mem_ctrl = phase->ctrl_or_self(mem);
1668 }
1669 if (mem->is_MergeMem()) {
1670 mem = mem->as_MergeMem()->memory_at(alias);
1671 mem_ctrl = phase->ctrl_or_self(mem);
1672 }
1673 return mem;
1674 }
1675
1676 Node* ShenandoahBarrierNode::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) {
1677 ResourceMark rm;
1678 VectorSet wq(Thread::current()->resource_area());
1679 wq.set(mem->_idx);
1680 mem_ctrl = phase->ctrl_or_self(mem);
1681 while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) {
1682 mem = next_mem(mem, alias);
1683 if (wq.test_set(mem->_idx)) {
1684 return NULL;
1685 }
1686 mem_ctrl = phase->ctrl_or_self(mem);
1687 }
1688 if (mem->is_MergeMem()) {
1689 mem = mem->as_MergeMem()->memory_at(alias);
1690 mem_ctrl = phase->ctrl_or_self(mem);
1691 }
1692 return mem;
1693 }
1694
1695 static void disconnect_barrier_mem(Node* wb, PhaseIterGVN& igvn) {
1696 Node* mem_in = wb->in(ShenandoahBarrierNode::Memory);
1697 Node* proj = wb->find_out_with(Op_ShenandoahWBMemProj);
1698
1699 for (DUIterator_Last imin, i = proj->last_outs(imin); i >= imin; ) {
1700 Node* u = proj->last_out(i);
1701 igvn.rehash_node_delayed(u);
1702 int nb = u->replace_edge(proj, mem_in);
1703 assert(nb > 0, "no replacement?");
1704 i -= nb;
1705 }
1706 }
1707
1708 Node* ShenandoahWriteBarrierNode::move_above_predicates(LoopNode* cl, Node* val_ctrl, PhaseIdealLoop* phase) {
1709 Node* entry = cl->skip_strip_mined(-1)->in(LoopNode::EntryControl);
1710 Node* above_pred = phase->skip_all_loop_predicates(entry);
1711 Node* ctrl = entry;
1712 while (ctrl != above_pred) {
1713 Node* next = ctrl->in(0);
1714 if (!phase->is_dominator(val_ctrl, next)) {
1715 break;
1716 }
1717 ctrl = next;
1718 }
1719 return ctrl;
1720 }
1721
1722 static MemoryGraphFixer* find_fixer(GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, int alias) {
1723 for (int i = 0; i < memory_graph_fixers.length(); i++) {
1724 if (memory_graph_fixers.at(i)->alias() == alias) {
1725 return memory_graph_fixers.at(i);
1726 }
1727 }
1728 return NULL;
1729 }
1730
1731 static MemoryGraphFixer* create_fixer(GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, int alias, PhaseIdealLoop* phase, bool include_lsm) {
1732 assert(find_fixer(memory_graph_fixers, alias) == NULL, "none should exist yet");
1733 MemoryGraphFixer* fixer = new MemoryGraphFixer(alias, include_lsm, phase);
1734 memory_graph_fixers.push(fixer);
1735 return fixer;
1736 }
1737
1738 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) {
1739 assert(cl->is_Loop(), "bad control");
1740 Node* ctrl = move_above_predicates(cl, val_ctrl, phase);
1741 Node* mem_ctrl = NULL;
1742 int alias = phase->C->get_alias_index(adr_type());
1743
1744 MemoryGraphFixer* fixer = find_fixer(memory_graph_fixers, alias);
1745 if (fixer == NULL) {
1746 fixer = create_fixer(memory_graph_fixers, alias, phase, include_lsm);
1747 }
1748
1749 Node* proj = find_out_with(Op_ShenandoahWBMemProj);
1750
1751 fixer->remove(proj);
1752 Node* mem = fixer->find_mem(ctrl, NULL);
1753
1754 assert(!ShenandoahVerifyOptoBarriers || memory_dominates_all_paths(mem, ctrl, alias, phase), "can't fix the memory graph");
1755
1756 phase->set_ctrl_and_loop(this, ctrl);
1757 phase->igvn().replace_input_of(this, Control, ctrl);
1758
1759 disconnect_barrier_mem(this, phase->igvn());
1760
1761 phase->igvn().replace_input_of(this, Memory, mem);
1762 phase->set_ctrl_and_loop(proj, ctrl);
1763
1764 fixer->fix_mem(ctrl, ctrl, mem, mem, proj, uses);
1765 assert(proj->outcnt() > 0, "disconnected write barrier");
1766 }
1767
1768 LoopNode* ShenandoahWriteBarrierNode::try_move_before_pre_loop(Node* c, Node* val_ctrl, PhaseIdealLoop* phase) {
1769 // A write barrier between a pre and main loop can get in the way of
1770 // vectorization. Move it above the pre loop if possible
1771 CountedLoopNode* cl = NULL;
1772 if (c->is_IfFalse() &&
1773 c->in(0)->is_CountedLoopEnd()) {
1774 cl = c->in(0)->as_CountedLoopEnd()->loopnode();
1775 } else if (c->is_IfProj() &&
1776 c->in(0)->is_If() &&
1777 c->in(0)->in(0)->is_IfFalse() &&
1778 c->in(0)->in(0)->in(0)->is_CountedLoopEnd()) {
1779 cl = c->in(0)->in(0)->in(0)->as_CountedLoopEnd()->loopnode();
1780 }
1781 if (cl != NULL &&
1782 cl->is_pre_loop() &&
1783 val_ctrl != cl &&
1784 phase->is_dominator(val_ctrl, cl)) {
1785 return cl;
1786 }
1787 return NULL;
1788 }
1789
1790 void ShenandoahWriteBarrierNode::try_move_before_loop(GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, PhaseIdealLoop* phase, bool include_lsm, Unique_Node_List& uses) {
1791 Node *n_ctrl = phase->get_ctrl(this);
1792 IdealLoopTree *n_loop = phase->get_loop(n_ctrl);
1793 Node* val = in(ValueIn);
1794 Node* val_ctrl = phase->get_ctrl(val);
1795 if (n_loop != phase->ltree_root() && !n_loop->_irreducible) {
1796 IdealLoopTree *val_loop = phase->get_loop(val_ctrl);
1797 Node* mem = in(Memory);
1798 IdealLoopTree *mem_loop = phase->get_loop(phase->get_ctrl(mem));
1799 if (!n_loop->is_member(val_loop) &&
1800 n_loop->is_member(mem_loop)) {
1801 Node* n_loop_head = n_loop->_head;
1802
1803 if (n_loop_head->is_Loop()) {
1804 LoopNode* loop = n_loop_head->as_Loop();
1805 if (n_loop_head->is_CountedLoop() && n_loop_head->as_CountedLoop()->is_main_loop()) {
1806 LoopNode* res = try_move_before_pre_loop(n_loop_head->in(LoopNode::EntryControl), val_ctrl, phase);
1807 if (res != NULL) {
1808 loop = res;
1809 }
1810 }
1811
1812 try_move_before_loop_helper(loop, val_ctrl, memory_graph_fixers, phase, include_lsm, uses);
1813 }
1814 }
1815 }
1816 LoopNode* ctrl = try_move_before_pre_loop(in(0), val_ctrl, phase);
1817 if (ctrl != NULL) {
1818 try_move_before_loop_helper(ctrl, val_ctrl, memory_graph_fixers, phase, include_lsm, uses);
1819 }
1820 }
1821
1822 Node* ShenandoahWriteBarrierNode::would_subsume(ShenandoahBarrierNode* other, PhaseIdealLoop* phase) {
1823 Node* val = in(ValueIn);
1824 Node* val_ctrl = phase->get_ctrl(val);
1825 Node* other_mem = other->in(Memory);
1826 Node* other_ctrl = phase->get_ctrl(other);
1827 Node* this_ctrl = phase->get_ctrl(this);
1828 IdealLoopTree* this_loop = phase->get_loop(this_ctrl);
1829 IdealLoopTree* other_loop = phase->get_loop(other_ctrl);
1830
1831 Node* ctrl = phase->dom_lca(other_ctrl, this_ctrl);
1832
1833 if (ctrl->is_Proj() &&
1834 ctrl->in(0)->is_Call() &&
1835 ctrl->unique_ctrl_out() != NULL &&
1836 ctrl->unique_ctrl_out()->Opcode() == Op_Catch &&
1837 !phase->is_dominator(val_ctrl, ctrl->in(0)->in(0))) {
1838 return NULL;
1839 }
1840
1841 IdealLoopTree* loop = phase->get_loop(ctrl);
1842
1843 // We don't want to move a write barrier in a loop
1844 // If the LCA is in a inner loop, try a control out of loop if possible
1845 while (!loop->is_member(this_loop) && (other->Opcode() != Op_ShenandoahWriteBarrier || !loop->is_member(other_loop))) {
1846 ctrl = phase->idom(ctrl);
1847 if (ctrl->is_MultiBranch()) {
1848 ctrl = ctrl->in(0);
1849 }
1850 if (ctrl != val_ctrl && phase->is_dominator(ctrl, val_ctrl)) {
1851 return NULL;
1852 }
1853 loop = phase->get_loop(ctrl);
1854 }
1855
1856 if (ShenandoahDontIncreaseWBFreq) {
1857 Node* this_iffproj = no_branches(this_ctrl, ctrl, true, phase);
1858 if (other->Opcode() == Op_ShenandoahWriteBarrier) {
1859 Node* other_iffproj = no_branches(other_ctrl, ctrl, true, phase);
1860 if (other_iffproj == NULL || this_iffproj == NULL) {
1861 return ctrl;
1862 } else if (other_iffproj != NodeSentinel && this_iffproj != NodeSentinel &&
1863 other_iffproj->in(0) == this_iffproj->in(0)) {
1864 return ctrl;
1865 }
1866 } else if (this_iffproj == NULL) {
1867 return ctrl;
1868 }
1869 return NULL;
1870 }
1871
1872 return ctrl;
1873 }
1874
1875 void ShenandoahWriteBarrierNode::optimize_before_expansion(PhaseIdealLoop* phase, GrowableArray<MemoryGraphFixer*> memory_graph_fixers, bool include_lsm) {
1876 bool progress = false;
1877 Unique_Node_List uses;
1878 do {
1879 progress = false;
1880 for (int i = 0; i < ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count(); i++) {
1881 ShenandoahWriteBarrierNode* wb = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barrier(i);
1882
1883 wb->try_move_before_loop(memory_graph_fixers, phase, include_lsm, uses);
1884
1885 Node* val = wb->in(ValueIn);
1886
1887 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
1888 Node* u = val->fast_out(j);
1889 if (u != wb && u->is_ShenandoahBarrier()) {
1890 Node* rep_ctrl = wb->would_subsume(u->as_ShenandoahBarrier(), phase);
1891
1892 if (rep_ctrl != NULL) {
1893 Node* other = u;
1894 Node* val_ctrl = phase->get_ctrl(val);
1895 if (rep_ctrl->is_Proj() &&
1896 rep_ctrl->in(0)->is_Call() &&
1897 rep_ctrl->unique_ctrl_out() != NULL &&
1898 rep_ctrl->unique_ctrl_out()->Opcode() == Op_Catch) {
1899 rep_ctrl = rep_ctrl->in(0)->in(0);
1900
1901 assert(phase->is_dominator(val_ctrl, rep_ctrl), "bad control");
1902 } else {
1903 LoopNode* c = ShenandoahWriteBarrierNode::try_move_before_pre_loop(rep_ctrl, val_ctrl, phase);
1904 if (c != NULL) {
1905 rep_ctrl = ShenandoahWriteBarrierNode::move_above_predicates(c, val_ctrl, phase);
1906 } else {
1907 while (rep_ctrl->is_IfProj()) {
1908 CallStaticJavaNode* unc = rep_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
1909 if (unc != NULL) {
1910 int req = unc->uncommon_trap_request();
1911 Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
1912 if ((trap_reason == Deoptimization::Reason_loop_limit_check ||
1913 trap_reason == Deoptimization::Reason_predicate ||
1914 trap_reason == Deoptimization::Reason_profile_predicate) &&
1915 phase->is_dominator(val_ctrl, rep_ctrl->in(0)->in(0))) {
1916 rep_ctrl = rep_ctrl->in(0)->in(0);
1917 continue;
1918 }
1919 }
1920 break;
1921 }
1922 }
1923 }
1924
1925 Node* wb_ctrl = phase->get_ctrl(wb);
1926 Node* other_ctrl = phase->get_ctrl(other);
1927 int alias = phase->C->get_alias_index(wb->adr_type());
1928 MemoryGraphFixer* fixer = find_fixer(memory_graph_fixers, alias);;
1929 if (!is_dominator(wb_ctrl, other_ctrl, wb, other, phase)) {
1930 if (fixer == NULL) {
1931 fixer = create_fixer(memory_graph_fixers, alias, phase, include_lsm);
1932 }
1933 Node* mem = fixer->find_mem(rep_ctrl, phase->get_ctrl(other) == rep_ctrl ? other : NULL);
1934
1935 if (mem->has_out_with(Op_Lock) || mem->has_out_with(Op_Unlock)) {
1936 continue;
1937 }
1938
1939 Node* wb_proj = wb->find_out_with(Op_ShenandoahWBMemProj);
1940 fixer->remove(wb_proj);
1941 Node* mem_for_ctrl = fixer->find_mem(rep_ctrl, NULL);
1942
1943 if (wb->in(Memory) != mem) {
1944 disconnect_barrier_mem(wb, phase->igvn());
1945 phase->igvn().replace_input_of(wb, Memory, mem);
1946 }
1947 if (rep_ctrl != wb_ctrl) {
1948 phase->set_ctrl_and_loop(wb, rep_ctrl);
1949 phase->igvn().replace_input_of(wb, Control, rep_ctrl);
1950 phase->set_ctrl_and_loop(wb_proj, rep_ctrl);
1951 progress = true;
1952 }
1953
1954 fixer->fix_mem(rep_ctrl, rep_ctrl, mem, mem_for_ctrl, wb_proj, uses);
1955
1956 assert(!ShenandoahVerifyOptoBarriers || ShenandoahWriteBarrierNode::memory_dominates_all_paths(mem, rep_ctrl, alias, phase), "can't fix the memory graph");
1957 }
1958
1959 if (other->Opcode() == Op_ShenandoahWriteBarrier) {
1960 Node* other_proj = other->find_out_with(Op_ShenandoahWBMemProj);
1961 if (fixer != NULL) {
1962 fixer->remove(other_proj);
1963 }
1964 phase->igvn().replace_node(other_proj, other->in(Memory));
1965 }
1966 phase->igvn().replace_node(other, wb);
1967 --j; --jmax;
1968 }
1969 }
1970 }
1971 }
1972 } while(progress);
1973 }
1974
1975 void ShenandoahReadBarrierNode::try_move(Node *n_ctrl, PhaseIdealLoop* phase) {
1976 Node* mem = in(MemNode::Memory);
1977 int alias = phase->C->get_alias_index(adr_type());
1978 const bool trace = false;
1979
1980 #ifdef ASSERT
1981 if (trace) { tty->print("Trying to move mem of"); dump(); }
1982 #endif
1983
1984 Node* new_mem = mem;
1985
1986 ResourceMark rm;
1987 VectorSet seen(Thread::current()->resource_area());
1988 Node_List phis;
1989
1990 for (;;) {
1991 #ifdef ASSERT
1992 if (trace) { tty->print("Looking for dominator from"); mem->dump(); }
1993 #endif
1994 if (mem->is_Proj() && mem->in(0)->is_Start()) {
1995 if (new_mem != in(MemNode::Memory)) {
1996 #ifdef ASSERT
1997 if (trace) { tty->print("XXX Setting mem to"); new_mem->dump(); tty->print(" for "); dump(); }
1998 #endif
1999 phase->igvn().replace_input_of(this, MemNode::Memory, new_mem);
2000 }
2001 return;
2002 }
2003
2004 Node* candidate = mem;
2005 do {
2006 if (!is_independent(mem)) {
2007 if (trace) { tty->print_cr("Not independent"); }
2008 if (new_mem != in(MemNode::Memory)) {
2009 #ifdef ASSERT
2010 if (trace) { tty->print("XXX Setting mem to"); new_mem->dump(); tty->print(" for "); dump(); }
2011 #endif
2012 phase->igvn().replace_input_of(this, MemNode::Memory, new_mem);
2013 }
2014 return;
2015 }
2016 if (seen.test_set(mem->_idx)) {
2017 if (trace) { tty->print_cr("Already seen"); }
2018 ShouldNotReachHere();
2019 // Strange graph
2020 if (new_mem != in(MemNode::Memory)) {
2021 #ifdef ASSERT
2022 if (trace) { tty->print("XXX Setting mem to"); new_mem->dump(); tty->print(" for "); dump(); }
2023 #endif
2024 phase->igvn().replace_input_of(this, MemNode::Memory, new_mem);
2025 }
2026 return;
2027 }
2028 if (mem->is_Phi()) {
2029 phis.push(mem);
2030 }
2031 mem = next_mem(mem, alias);
2032 if (mem->bottom_type() == Type::MEMORY) {
2033 candidate = mem;
2034 }
2035 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");
2036 #ifdef ASSERT
2037 if (trace) { tty->print("Next mem is"); mem->dump(); }
2038 #endif
2039 } while (mem->bottom_type() != Type::MEMORY || !phase->is_dominator(phase->ctrl_or_self(mem), n_ctrl));
2040
2041 assert(mem->bottom_type() == Type::MEMORY, "bad mem");
2042
2043 bool not_dom = false;
2044 for (uint i = 0; i < phis.size() && !not_dom; i++) {
2045 Node* nn = phis.at(i);
2046
2047 #ifdef ASSERT
2048 if (trace) { tty->print("Looking from phi"); nn->dump(); }
2049 #endif
2050 assert(nn->is_Phi(), "phis only");
2051 for (uint j = 2; j < nn->req() && !not_dom; j++) {
2052 Node* m = nn->in(j);
2053 #ifdef ASSERT
2054 if (trace) { tty->print("Input %d is", j); m->dump(); }
2055 #endif
2056 while (m != mem && !seen.test_set(m->_idx)) {
2057 if (is_dominator(phase->ctrl_or_self(m), phase->ctrl_or_self(mem), m, mem, phase)) {
2058 not_dom = true;
2059 // Scheduling anomaly
2060 #ifdef ASSERT
2061 if (trace) { tty->print("Giving up"); m->dump(); }
2062 #endif
2063 break;
2064 }
2065 if (!is_independent(m)) {
2066 if (trace) { tty->print_cr("Not independent"); }
2067 if (new_mem != in(MemNode::Memory)) {
2068 #ifdef ASSERT
2069 if (trace) { tty->print("XXX Setting mem to"); new_mem->dump(); tty->print(" for "); dump(); }
2070 #endif
2071 phase->igvn().replace_input_of(this, MemNode::Memory, new_mem);
2072 }
2073 return;
2074 }
2075 if (m->is_Phi()) {
2076 phis.push(m);
2077 }
2078 m = next_mem(m, alias);
2079 #ifdef ASSERT
2080 if (trace) { tty->print("Next mem is"); m->dump(); }
2081 #endif
2082 }
2083 }
2084 }
2085 if (!not_dom) {
2086 new_mem = mem;
2087 phis.clear();
2088 } else {
2089 seen.Clear();
2090 }
2091 }
2092 }
2093
2094 CallStaticJavaNode* ShenandoahWriteBarrierNode::pin_and_expand_null_check(PhaseIterGVN& igvn) {
2095 Node* val = in(ValueIn);
2096
2097 const Type* val_t = igvn.type(val);
2098
2099 if (val_t->meet(TypePtr::NULL_PTR) != val_t &&
2100 val->Opcode() == Op_CastPP &&
2101 val->in(0) != NULL &&
2102 val->in(0)->Opcode() == Op_IfTrue &&
2103 val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
2104 val->in(0)->in(0)->is_If() &&
2105 val->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
2106 val->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
2107 val->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
2108 val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1) &&
2109 val->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
2110 assert(val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1), "");
2111 CallStaticJavaNode* unc = val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
2112 return unc;
2113 }
2114 return NULL;
2115 }
2116
2117 void ShenandoahWriteBarrierNode::pin_and_expand_move_barrier(PhaseIdealLoop* phase, GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, Unique_Node_List& uses) {
2118 Node* unc = pin_and_expand_null_check(phase->igvn());
2119 Node* val = in(ValueIn);
2120
2121 if (unc != NULL) {
2122 Node* ctrl = phase->get_ctrl(this);
2123 Node* unc_ctrl = val->in(0);
2124
2125 // Don't move write barrier in a loop
2126 IdealLoopTree* loop = phase->get_loop(ctrl);
2127 IdealLoopTree* unc_loop = phase->get_loop(unc_ctrl);
2128
2129 if (!unc_loop->is_member(loop)) {
2130 return;
2131 }
2132
2133 Node* branch = no_branches(ctrl, unc_ctrl, false, phase);
2134 assert(branch == NULL || branch == NodeSentinel, "was not looking for a branch");
2135 if (branch == NodeSentinel) {
2136 return;
2137 }
2138
2139
2140 RegionNode* r = new RegionNode(3);
2141 IfNode* iff = unc_ctrl->in(0)->as_If();
2142
2143 Node* ctrl_use = unc_ctrl->unique_ctrl_out();
2144 Node* unc_ctrl_clone = unc_ctrl->clone();
2145 phase->register_control(unc_ctrl_clone, loop, iff);
2146 Node* c = unc_ctrl_clone;
2147 Node* new_cast = clone_null_check(c, val, unc_ctrl_clone, r, 1, phase);
2148
2149 phase->igvn().replace_input_of(unc_ctrl, 0, c->in(0));
2150 phase->set_idom(unc_ctrl, c->in(0), phase->dom_depth(unc_ctrl));
2151 phase->lazy_replace(c, unc_ctrl);
2152 c = NULL;;
2153 phase->igvn().replace_input_of(val, 0, unc_ctrl_clone);
2154 phase->set_ctrl(val, unc_ctrl_clone);
2155
2156 IfNode* new_iff = new_cast->in(0)->in(0)->as_If();
2157 fix_null_check(iff, unc, unc_ctrl_clone, r, uses, phase);
2158 Node* iff_proj = iff->proj_out(0);
2159 r->init_req(2, iff_proj);
2160
2161 Node* new_bol = new_iff->in(1)->clone();
2162 Node* new_cmp = new_bol->in(1)->clone();
2163 assert(new_cmp->Opcode() == Op_CmpP, "broken");
2164 assert(new_cmp->in(1) == val->in(1), "broken");
2165 new_bol->set_req(1, new_cmp);
2166 new_cmp->set_req(1, this);
2167 phase->register_new_node(new_bol, new_iff->in(0));
2168 phase->register_new_node(new_cmp, new_iff->in(0));
2169 phase->igvn().replace_input_of(new_iff, 1, new_bol);
2170 phase->igvn().replace_input_of(new_cast, 1, this);
2171
2172 for (DUIterator_Fast imax, i = this->fast_outs(imax); i < imax; i++) {
2173 Node* u = this->fast_out(i);
2174 if (u == new_cast || u->Opcode() == Op_ShenandoahWBMemProj || u == new_cmp) {
2175 continue;
2176 }
2177 phase->igvn().rehash_node_delayed(u);
2178 int nb = u->replace_edge(this, new_cast);
2179 assert(nb > 0, "no update?");
2180 --i; imax -= nb;
2181 }
2182
2183 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
2184 Node* u = val->fast_out(i);
2185 if (u == this) {
2186 continue;
2187 }
2188 phase->igvn().rehash_node_delayed(u);
2189 int nb = u->replace_edge(val, new_cast);
2190 assert(nb > 0, "no update?");
2191 --i; imax -= nb;
2192 }
2193
2194 Node* new_ctrl = unc_ctrl_clone;
2195
2196 int alias = phase->C->get_alias_index(adr_type());
2197 MemoryGraphFixer* fixer = find_fixer(memory_graph_fixers, alias);
2198 if (fixer == NULL) {
2199 fixer = create_fixer(memory_graph_fixers, alias, phase, true);
2200 }
2201
2202 Node* proj = find_out_with(Op_ShenandoahWBMemProj);
2203 fixer->remove(proj);
2204 Node* mem = fixer->find_mem(new_ctrl, NULL);
2205
2206 if (in(Memory) != mem) {
2207 disconnect_barrier_mem(this, phase->igvn());
2208 phase->igvn().replace_input_of(this, Memory, mem);
2209 }
2210
2211 phase->set_ctrl_and_loop(this, new_ctrl);
2212 phase->igvn().replace_input_of(this, Control, new_ctrl);
2213 phase->set_ctrl_and_loop(proj, new_ctrl);
2214
2215 fixer->fix_mem(new_ctrl, new_ctrl, mem, mem, proj, uses);
2216 }
2217 }
2218
2219 void ShenandoahWriteBarrierNode::pin_and_expand_helper(PhaseIdealLoop* phase) {
2220 Node* val = in(ValueIn);
2221 CallStaticJavaNode* unc = pin_and_expand_null_check(phase->igvn());
2222 Node* rep = this;
2223 Node* ctrl = phase->get_ctrl(this);
2224 if (unc != NULL && val->in(0) == ctrl) {
2225 Node* unc_ctrl = val->in(0);
2226 IfNode* other_iff = unc_ctrl->unique_ctrl_out()->as_If();
2227 ProjNode* other_unc_ctrl = other_iff->proj_out(1);
2228 Node* cast = NULL;
2229 for (DUIterator_Fast imax, i = other_unc_ctrl->fast_outs(imax); i < imax && cast == NULL; i++) {
2230 Node* u = other_unc_ctrl->fast_out(i);
2231 if (u->Opcode() == Op_CastPP && u->in(1) == this) {
2232 cast = u;
2233 }
2234 }
2235 assert(other_unc_ctrl->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) == unc, "broken");
2236 rep = cast;
2237 }
2238
2239 // Replace all uses of barrier's input that are dominated by ctrl
2240 // with the value returned by the barrier: no need to keep both
2241 // live.
2242 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
2243 Node* u = val->fast_out(i);
2244 if (u != this) {
2245 if (u->is_Phi()) {
2246 int nb = 0;
2247 for (uint j = 1; j < u->req(); j++) {
2248 if (u->in(j) == val) {
2249 Node* c = u->in(0)->in(j);
2250 if (phase->is_dominator(ctrl, c)) {
2251 phase->igvn().replace_input_of(u, j, rep);
2252 nb++;
2253 }
2254 }
2255 }
2256 if (nb > 0) {
2257 imax -= nb;
2258 --i;
2259 }
2260 } else {
2261 Node* c = phase->ctrl_or_self(u);
2262 if (is_dominator(ctrl, c, this, u, phase)) {
2263 phase->igvn().rehash_node_delayed(u);
2264 int nb = u->replace_edge(val, rep);
2265 assert(nb > 0, "no update?");
2266 --i, imax -= nb;
2267 }
2268 }
2269 }
2270 }
2271 }
2272
2273 Node* ShenandoahWriteBarrierNode::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) {
2274 Node* mem = NULL;
2275 Node* c = ctrl;
2276 do {
2277 if (c->is_Region()) {
2278 Node* phi_bottom = NULL;
2279 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == NULL; i++) {
2280 Node* u = c->fast_out(i);
2281 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
2282 if (u->adr_type() == TypePtr::BOTTOM) {
2283 mem = u;
2284 }
2285 }
2286 }
2287 } else {
2288 if (c->is_Call() && c->as_Call()->adr_type() != NULL) {
2289 CallProjections projs;
2290 c->as_Call()->extract_projections(&projs, true, false);
2291 if (projs.fallthrough_memproj != NULL) {
2292 if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
2293 if (projs.catchall_memproj == NULL) {
2294 mem = projs.fallthrough_memproj;
2295 } else {
2296 if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) {
2297 mem = projs.fallthrough_memproj;
2298 } else {
2299 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier");
2300 mem = projs.catchall_memproj;
2301 }
2302 }
2303 }
2304 } else {
2305 Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
2306 if (proj != NULL &&
2307 proj->adr_type() == TypePtr::BOTTOM) {
2308 mem = proj;
2309 }
2310 }
2311 } else {
2312 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
2313 Node* u = c->fast_out(i);
2314 if (u->is_Proj() &&
2315 u->bottom_type() == Type::MEMORY &&
2316 u->adr_type() == TypePtr::BOTTOM) {
2317 assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
2318 assert(mem == NULL, "only one proj");
2319 mem = u;
2320 }
2321 }
2322 assert(!c->is_Call() || c->as_Call()->adr_type() != NULL || mem == NULL, "no mem projection expected");
2323 }
2324 }
2325 c = phase->idom(c);
2326 } while (mem == NULL);
2327 return mem;
2328 }
2329
2330 void ShenandoahWriteBarrierNode::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
2331 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2332 Node* u = n->fast_out(i);
2333 if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
2334 uses.push(u);
2335 }
2336 }
2337 }
2338
2339 void ShenandoahWriteBarrierNode::test_heap_stable(Node* ctrl, Node* raw_mem, Node*& gc_state, Node*& heap_stable,
2340 Node*& heap_not_stable, PhaseIdealLoop* phase) {
2341 IdealLoopTree *loop = phase->get_loop(ctrl);
2342 Node* thread = new ThreadLocalNode();
2343 phase->register_new_node(thread, ctrl);
2344 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
2345 phase->set_ctrl(offset, phase->C->root());
2346 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, offset);
2347 phase->register_new_node(gc_state_addr, ctrl);
2348 uint gc_state_idx = Compile::AliasIdxRaw;
2349 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument
2350 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
2351
2352 gc_state = new LoadBNode(ctrl, raw_mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered);
2353 phase->register_new_node(gc_state, ctrl);
2354 Node* heap_stable_and = new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::HAS_FORWARDED));
2355 phase->register_new_node(heap_stable_and, ctrl);
2356 Node* heap_stable_cmp = new CmpINode(heap_stable_and, phase->igvn().zerocon(T_INT));
2357 phase->register_new_node(heap_stable_cmp, ctrl);
2358 Node* heap_stable_test = new BoolNode(heap_stable_cmp, BoolTest::ne);
2359 phase->register_new_node(heap_stable_test, ctrl);
2360 IfNode* heap_stable_iff = new IfNode(ctrl, heap_stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
2361 phase->register_control(heap_stable_iff, loop, ctrl);
2362
2363 heap_stable = new IfFalseNode(heap_stable_iff);
2364 phase->register_control(heap_stable, loop, heap_stable_iff);
2365 heap_not_stable = new IfTrueNode(heap_stable_iff);
2366 phase->register_control(heap_not_stable, loop, heap_stable_iff);
2367
2368 assert(is_heap_stable_test(heap_stable_iff), "Should match the shape");
2369 }
2370
2371
2372 void ShenandoahWriteBarrierNode::test_evacuation_in_progress(Node* ctrl, Node* val, Node*& raw_mem,
2373 Node*& evac_in_progress, Node*& evac_not_in_progress,
2374 Node*& heap_stable, Node*& null_val,
2375 PhaseIdealLoop* phase) {
2376 IdealLoopTree *loop = phase->get_loop(ctrl);
2377 Node* heap_not_stable = NULL;
2378 Node* unused_gc_state = NULL;
2379
2380 test_heap_stable(ctrl, raw_mem, unused_gc_state, heap_stable, heap_not_stable, phase);
2381
2382 ctrl = heap_not_stable;
2383
2384 const Type* val_t = phase->igvn().type(val);
2385
2386 if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
2387 Node* null_cmp = new CmpPNode(val, phase->igvn().zerocon(T_OBJECT));
2388 phase->register_new_node(null_cmp, ctrl);
2389 Node* null_test = new BoolNode(null_cmp, BoolTest::ne);
2390 phase->register_new_node(null_test, ctrl);
2391 IfNode* null_iff = new IfNode(ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
2392 phase->register_control(null_iff, loop, ctrl);
2393 Node* not_null = new IfTrueNode(null_iff);
2394 phase->register_control(not_null, loop, null_iff);
2395 Node* null = new IfFalseNode(null_iff);
2396 phase->register_control(null, loop, null_iff);
2397 null_val = null;
2398 ctrl = not_null;
2399 }
2400
2401 Node* thread = new ThreadLocalNode();
2402 phase->register_new_node(thread, ctrl);
2403 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
2404 phase->set_ctrl(offset, phase->C->root());
2405 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, offset);
2406 phase->register_new_node(gc_state_addr, ctrl);
2407 uint gc_state_idx = Compile::AliasIdxRaw;
2408 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument
2409 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
2410
2411 Node* gc_state = new LoadBNode(ctrl, raw_mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered);
2412 phase->register_new_node(gc_state, ctrl);
2413
2414 Node* evacuation_in_progress = new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL));
2415 phase->register_new_node(evacuation_in_progress, ctrl);
2416 Node* evacuation_in_progress_cmp = new CmpINode(evacuation_in_progress, phase->igvn().zerocon(T_INT));
2417 phase->register_new_node(evacuation_in_progress_cmp, ctrl);
2418 Node* evacuation_in_progress_test = new BoolNode(evacuation_in_progress_cmp, BoolTest::ne);
2419 phase->register_new_node(evacuation_in_progress_test, ctrl);
2420 IfNode* evacuation_iff = new IfNode(ctrl, evacuation_in_progress_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
2421 phase->register_control(evacuation_iff, loop, ctrl);
2422
2423 assert(is_evacuation_in_progress_test(evacuation_iff), "Should match the shape");
2424 assert(is_gc_state_load(gc_state), "Should match the shape");
2425
2426 evac_not_in_progress = new IfFalseNode(evacuation_iff);
2427 phase->register_control(evac_not_in_progress, loop, evacuation_iff);
2428 evac_in_progress = new IfTrueNode(evacuation_iff);
2429 phase->register_control(evac_in_progress, loop, evacuation_iff);
2430 }
2431
2432 Node* ShenandoahWriteBarrierNode::clone_null_check(Node*& c, Node* val, Node* unc_ctrl,
2433 Node* unc_region, uint input, PhaseIdealLoop* phase) {
2434 IdealLoopTree *loop = phase->get_loop(c);
2435 Node* iff = unc_ctrl->in(0);
2436 assert(iff->is_If(), "broken");
2437 Node* new_iff = iff->clone();
2438 new_iff->set_req(0, c);
2439 phase->register_control(new_iff, loop, c);
2440 Node* iffalse = new IfFalseNode(new_iff->as_If());
2441 phase->register_control(iffalse, loop, new_iff);
2442 Node* iftrue = new IfTrueNode(new_iff->as_If());
2443 phase->register_control(iftrue, loop, new_iff);
2444 c = iftrue;
2445 const Type *t = phase->igvn().type(val);
2446 assert(val->Opcode() == Op_CastPP, "expect cast to non null here");
2447 Node* uncasted_val = val->in(1);
2448 val = new CastPPNode(uncasted_val, t);
2449 val->init_req(0, c);
2450 phase->register_new_node(val, c);
2451 unc_region->init_req(input, iffalse);
2452 return val;
2453 }
2454
2455 void ShenandoahWriteBarrierNode::fix_null_check(Node* dom, Node* unc, Node* unc_ctrl, Node* unc_region,
2456 Unique_Node_List& uses, PhaseIdealLoop* phase) {
2457 IfNode* iff = unc_ctrl->in(0)->as_If();
2458 Node* proj = iff->proj_out(0);
2459 assert(proj != unc_ctrl, "bad projection");
2460 Node* use = proj->unique_ctrl_out();
2461
2462 assert(use == unc || use->is_Region(), "what else?");
2463
2464 uses.clear();
2465 if (use == unc) {
2466 phase->set_idom(use, unc_region, phase->dom_depth(use));
2467 for (uint i = 1; i < unc->req(); i++) {
2468 Node* n = unc->in(i);
2469 if (phase->has_ctrl(n) && phase->get_ctrl(n) == proj) {
2470 uses.push(n);
2471 }
2472 }
2473 } else {
2474 assert(use->is_Region(), "what else?");
2475 uint idx = 1;
2476 for (; use->in(idx) != proj; idx++);
2477 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2478 Node* u = use->fast_out(i);
2479 if (u->is_Phi() && phase->get_ctrl(u->in(idx)) == proj) {
2480 uses.push(u->in(idx));
2481 }
2482 }
2483 }
2484 for(uint next = 0; next < uses.size(); next++ ) {
2485 Node *n = uses.at(next);
2486 assert(phase->get_ctrl(n) == proj, "bad control");
2487 phase->set_ctrl_and_loop(n, unc_region);
2488 if (n->in(0) == proj) {
2489 phase->igvn().replace_input_of(n, 0, unc_region);
2490 }
2491 for (uint i = 0; i < n->req(); i++) {
2492 Node* m = n->in(i);
2493 if (m != NULL && phase->has_ctrl(m) && phase->get_ctrl(m) == proj) {
2494 uses.push(m);
2495 }
2496 }
2497 }
2498
2499 phase->igvn().rehash_node_delayed(use);
2500 int nb = use->replace_edge(proj, unc_region);
2501 assert(nb == 1, "only use expected");
2502 phase->register_control(unc_region, phase->ltree_root(), dom);
2503 }
2504
2505 void ShenandoahWriteBarrierNode::evacuation_not_in_progress_null_check(Node*& c, Node*& val, Node* unc_ctrl, Node*& unc_region, PhaseIdealLoop* phase) {
2506 if (unc_ctrl != NULL) {
2507 // Clone the null check in this branch to allow implicit null check
2508 unc_region = new RegionNode(3);
2509 val = clone_null_check(c, val, unc_ctrl, unc_region, 1, phase);
2510 }
2511 }
2512
2513 void ShenandoahWriteBarrierNode::evacuation_not_in_progress(Node* c, Node* val, Node* unc_ctrl, Node* raw_mem, Node* wb_mem, Node* region,
2514 Node* val_phi, Node* mem_phi, Node* raw_mem_phi, Node*& unc_region, PhaseIdealLoop* phase) {
2515 evacuation_not_in_progress_null_check(c, val, unc_ctrl, unc_region, phase);
2516 region->init_req(1, c);
2517 if (ShenandoahWriteBarrierRB) {
2518 Node* rbfalse = new ShenandoahReadBarrierNode(c, wb_mem, val);
2519 phase->register_new_node(rbfalse, c);
2520 val_phi->init_req(1, rbfalse);
2521 } else {
2522 val_phi->init_req(1, val);
2523 }
2524 mem_phi->init_req(1, wb_mem);
2525 raw_mem_phi->init_req(1, raw_mem);
2526 }
2527
2528 void ShenandoahWriteBarrierNode::heap_stable(Node* c, Node* val, Node* unc_ctrl, Node* raw_mem, Node* wb_mem, Node* region,
2529 Node* val_phi, Node* mem_phi, Node* raw_mem_phi, Node* unc_region, PhaseIdealLoop* phase) {
2530 region->init_req(1, c);
2531 if (unc_ctrl != NULL) {
2532 val = val->in(1);
2533 }
2534 val_phi->init_req(1, val);
2535 mem_phi->init_req(1, wb_mem);
2536 raw_mem_phi->init_req(1, raw_mem);
2537 }
2538
2539 void ShenandoahWriteBarrierNode::evacuation_in_progress_null_check(Node*& c, Node*& val, Node* evacuation_iff, Node* unc, Node* unc_ctrl,
2540 Node* unc_region, Unique_Node_List& uses, PhaseIdealLoop* phase) {
2541 if (unc != NULL) {
2542 // Clone the null check in this branch to allow implicit null check
2543 val = clone_null_check(c, val, unc_ctrl, unc_region, 2, phase);
2544
2545 fix_null_check(evacuation_iff, unc, unc_ctrl, unc_region, uses, phase);
2546
2547 IfNode* iff = unc_ctrl->in(0)->as_If();
2548 phase->igvn().replace_input_of(iff, 1, phase->igvn().intcon(1));
2549 }
2550 }
2551
2552 void ShenandoahWriteBarrierNode::in_cset_fast_test(Node*& c, Node* rbtrue, Node* raw_mem, Node* wb_mem, Node* region, Node* val_phi, Node* mem_phi,
2553 Node* raw_mem_phi, PhaseIdealLoop* phase) {
2554 if (ShenandoahWriteBarrierCsetTestInIR) {
2555 IdealLoopTree *loop = phase->get_loop(c);
2556 Node* raw_rbtrue = new CastP2XNode(c, rbtrue);
2557 phase->register_new_node(raw_rbtrue, c);
2558 Node* cset_offset = new URShiftXNode(raw_rbtrue, phase->igvn().intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
2559 phase->register_new_node(cset_offset, c);
2560 Node* in_cset_fast_test_base_addr = phase->igvn().makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
2561 phase->set_ctrl(in_cset_fast_test_base_addr, phase->C->root());
2562 Node* in_cset_fast_test_adr = new AddPNode(phase->C->top(), in_cset_fast_test_base_addr, cset_offset);
2563 phase->register_new_node(in_cset_fast_test_adr, c);
2564 uint in_cset_fast_test_idx = Compile::AliasIdxRaw;
2565 const TypePtr* in_cset_fast_test_adr_type = NULL; // debug-mode-only argument
2566 debug_only(in_cset_fast_test_adr_type = phase->C->get_adr_type(in_cset_fast_test_idx));
2567 Node* in_cset_fast_test_load = new LoadBNode(c, raw_mem, in_cset_fast_test_adr, in_cset_fast_test_adr_type, TypeInt::BYTE, MemNode::unordered);
2568 phase->register_new_node(in_cset_fast_test_load, c);
2569 Node* in_cset_fast_test_cmp = new CmpINode(in_cset_fast_test_load, phase->igvn().zerocon(T_INT));
2570 phase->register_new_node(in_cset_fast_test_cmp, c);
2571 Node* in_cset_fast_test_test = new BoolNode(in_cset_fast_test_cmp, BoolTest::ne);
2572 phase->register_new_node(in_cset_fast_test_test, c);
2573 IfNode* in_cset_fast_test_iff = new IfNode(c, in_cset_fast_test_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
2574 phase->register_control(in_cset_fast_test_iff, loop, c);
2575
2576 Node* in_cset_fast_test_success = new IfFalseNode(in_cset_fast_test_iff);
2577 phase->register_control(in_cset_fast_test_success, loop, in_cset_fast_test_iff);
2578
2579 region->init_req(3, in_cset_fast_test_success);
2580 val_phi->init_req(3, rbtrue);
2581 mem_phi->init_req(3, wb_mem);
2582 raw_mem_phi->init_req(3, raw_mem);
2583
2584 Node* in_cset_fast_test_failure = new IfTrueNode(in_cset_fast_test_iff);
2585 phase->register_control(in_cset_fast_test_failure, loop, in_cset_fast_test_iff);
2586
2587 c = in_cset_fast_test_failure;
2588 }
2589 }
2590
2591 void ShenandoahWriteBarrierNode::evacuation_in_progress(Node* c, Node* val, Node* evacuation_iff, Node* unc, Node* unc_ctrl,
2592 Node* raw_mem, Node* wb_mem, Node* region, Node* val_phi, Node* mem_phi,
2593 Node* raw_mem_phi, Node* unc_region, int alias, Unique_Node_List& uses,
2594 PhaseIdealLoop* phase) {
2595 evacuation_in_progress_null_check(c, val, evacuation_iff, unc, unc_ctrl, unc_region, uses, phase);
2596
2597 IdealLoopTree *loop = phase->get_loop(c);
2598
2599 // Important to perform resolve here, before doing cset check, because that would
2600 // capture forwarded objects we do not need to evacuate again.
2601 Node* rbtrue = new ShenandoahReadBarrierNode(c, wb_mem, val);
2602 phase->register_new_node(rbtrue, c);
2603
2604 in_cset_fast_test(c, rbtrue, raw_mem, wb_mem, region, val_phi, mem_phi, raw_mem_phi, phase);
2605
2606 // The slow path stub consumes and produces raw memory in addition
2607 // to the existing memory edges
2608 Node* base = find_bottom_mem(c, phase);
2609
2610 MergeMemNode* mm = MergeMemNode::make(base);
2611 mm->set_memory_at(alias, wb_mem);
2612 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
2613 phase->register_new_node(mm, c);
2614
2615 Node* call = new CallLeafNoFPNode(ShenandoahBarrierSetC2::shenandoah_write_barrier_Type(), ShenandoahBarrierSetAssembler::shenandoah_wb_C(), "shenandoah_write_barrier", TypeRawPtr::BOTTOM);
2616 call->init_req(TypeFunc::Control, c);
2617 call->init_req(TypeFunc::I_O, phase->C->top());
2618 call->init_req(TypeFunc::Memory, mm);
2619 call->init_req(TypeFunc::FramePtr, phase->C->top());
2620 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
2621 call->init_req(TypeFunc::Parms, rbtrue);
2622 phase->register_control(call, loop, c);
2623 Node* ctrl_proj = new ProjNode(call, TypeFunc::Control);
2624 phase->register_control(ctrl_proj, loop, call);
2625 Node* mem_proj = new ProjNode(call, TypeFunc::Memory);
2626 phase->register_new_node(mem_proj, call);
2627 Node* res_proj = new ProjNode(call, TypeFunc::Parms);
2628 phase->register_new_node(res_proj, call);
2629 Node* res = new CheckCastPPNode(ctrl_proj, res_proj, phase->igvn().type(val)->is_oopptr()->cast_to_nonconst());
2630 phase->register_new_node(res, ctrl_proj);
2631 region->init_req(2, ctrl_proj);
2632 val_phi->init_req(2, res);
2633 mem_phi->init_req(2, mem_proj);
2634 raw_mem_phi->init_req(2, mem_proj);
2635 }
2636
2637 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) {
2638 Node* ctrl = phase->get_ctrl(barrier);
2639 Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
2640
2641 // Update the control of all nodes that should be after the
2642 // barrier control flow
2643 uses.clear();
2644 // Every node that is control dependent on the barrier's input
2645 // control will be after the expanded barrier. The raw memory (if
2646 // its memory is control dependent on the barrier's input control)
2647 // must stay above the barrier.
2648 uses_to_ignore.clear();
2649 if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
2650 uses_to_ignore.push(init_raw_mem);
2651 }
2652 for (uint next = 0; next < uses_to_ignore.size(); next++) {
2653 Node *n = uses_to_ignore.at(next);
2654 for (uint i = 0; i < n->req(); i++) {
2655 Node* in = n->in(i);
2656 if (in != NULL && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
2657 uses_to_ignore.push(in);
2658 }
2659 }
2660 }
2661 for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
2662 Node* u = ctrl->fast_out(i);
2663 if (u->_idx < last &&
2664 u != barrier &&
2665 !uses_to_ignore.member(u) &&
2666 (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) &&
2667 (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) {
2668 Node* old_c = phase->ctrl_or_self(u);
2669 Node* c = old_c;
2670 if (c != ctrl ||
2671 is_dominator_same_ctrl(old_c, barrier, u, phase) ||
2672 u->is_shenandoah_state_load()) {
2673 phase->igvn().rehash_node_delayed(u);
2674 int nb = u->replace_edge(ctrl, region);
2675 if (u->is_CFG()) {
2676 if (phase->idom(u) == ctrl) {
2677 phase->set_idom(u, region, phase->dom_depth(region));
2678 }
2679 } else if (phase->get_ctrl(u) == ctrl) {
2680 assert(u != init_raw_mem, "should leave input raw mem above the barrier");
2681 uses.push(u);
2682 }
2683 assert(nb == 1, "more than 1 ctrl input?");
2684 --i, imax -= nb;
2685 }
2686 }
2687 }
2688 }
2689
2690
2691 void ShenandoahWriteBarrierNode::pin_and_expand(PhaseIdealLoop* phase) {
2692 Node_List enqueue_barriers;
2693 if (ShenandoahStoreValEnqueueBarrier) {
2694 Unique_Node_List wq;
2695 wq.push(phase->C->root());
2696 for (uint i = 0; i < wq.size(); i++) {
2697 Node* n = wq.at(i);
2698 if (n->Opcode() == Op_ShenandoahEnqueueBarrier) {
2699 enqueue_barriers.push(n);
2700 }
2701 for (uint i = 0; i < n->req(); i++) {
2702 Node* in = n->in(i);
2703 if (in != NULL) {
2704 wq.push(in);
2705 }
2706 }
2707 }
2708 }
2709
2710 const bool trace = false;
2711
2712 // Collect raw memory state at CFG points in the entire graph and
2713 // record it in memory_nodes. Optimize the raw memory graph in the
2714 // process. Optimizing the memory graph also makes the memory graph
2715 // simpler.
2716 GrowableArray<MemoryGraphFixer*> memory_graph_fixers;
2717
2718 // Let's try to common write barriers again
2719 optimize_before_expansion(phase, memory_graph_fixers, true);
2720
2721 Unique_Node_List uses;
2722 for (int i = 0; i < ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count(); i++) {
2723 ShenandoahWriteBarrierNode* wb = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barrier(i);
2724 Node* ctrl = phase->get_ctrl(wb);
2725
2726 Node* val = wb->in(ValueIn);
2727 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) {
2728 assert(is_dominator(phase->get_ctrl(val), ctrl->in(0)->in(0), val, ctrl->in(0), phase), "can't move");
2729 phase->set_ctrl(wb, ctrl->in(0)->in(0));
2730 } else if (ctrl->is_CallRuntime()) {
2731 assert(is_dominator(phase->get_ctrl(val), ctrl->in(0), val, ctrl, phase), "can't move");
2732 phase->set_ctrl(wb, ctrl->in(0));
2733 }
2734
2735 assert(wb->Opcode() == Op_ShenandoahWriteBarrier, "only for write barriers");
2736 // Look for a null check that dominates this barrier and move the
2737 // barrier right after the null check to enable implicit null
2738 // checks
2739 wb->pin_and_expand_move_barrier(phase, memory_graph_fixers, uses);
2740
2741 wb->pin_and_expand_helper(phase);
2742 }
2743
2744 MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
2745 Unique_Node_List uses_to_ignore;
2746 for (uint i = 0; i < enqueue_barriers.size(); i++) {
2747 Node* barrier = enqueue_barriers.at(i);
2748 Node* pre_val = barrier->in(1);
2749
2750 if (phase->igvn().type(pre_val)->higher_equal(TypePtr::NULL_PTR)) {
2751 ShouldNotReachHere();
2752 continue;
2753 }
2754
2755 Node* ctrl = phase->get_ctrl(barrier);
2756
2757 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) {
2758 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0)->in(0), pre_val, ctrl->in(0), phase), "can't move");
2759 ctrl = ctrl->in(0)->in(0);
2760 phase->set_ctrl(barrier, ctrl);
2761 } else if (ctrl->is_CallRuntime()) {
2762 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0), pre_val, ctrl, phase), "can't move");
2763 ctrl = ctrl->in(0);
2764 phase->set_ctrl(barrier, ctrl);
2765 }
2766
2767 Node* init_ctrl = ctrl;
2768 IdealLoopTree* loop = phase->get_loop(ctrl);
2769 Node* raw_mem = fixer.find_mem(ctrl, barrier);
2770 Node* init_raw_mem = raw_mem;
2771 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL);
2772 Node* evac_in_progress = NULL;
2773 Node* evac_not_in_progress = NULL;
2774 Node* heap_stable = NULL;
2775 Node* null_val = NULL;
2776 uint last = phase->C->unique();
2777
2778 Node* stable_test_region = new RegionNode(3);
2779 Node* stable_test_phi = PhiNode::make(stable_test_region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
2780 Node* unstable_region = new RegionNode(5);
2781 Node* unstable_phi = PhiNode::make(unstable_region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
2782
2783 test_evacuation_in_progress(ctrl, pre_val, raw_mem, evac_in_progress, evac_not_in_progress, heap_stable, null_val, phase);
2784
2785 stable_test_region->init_req(1, heap_stable);
2786 stable_test_region->init_req(2, unstable_region);
2787 stable_test_phi->init_req(1, raw_mem);
2788 stable_test_phi->init_req(2, unstable_phi);
2789
2790 if (null_val != NULL) {
2791 unstable_region->init_req(1, null_val);
2792 unstable_phi->init_req(1, raw_mem);
2793 }
2794 unstable_region->init_req(2, evac_not_in_progress);
2795 unstable_phi->init_req(2, raw_mem);
2796
2797 ctrl = evac_in_progress;
2798
2799 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
2800 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
2801 Node* thread = new ThreadLocalNode();
2802 phase->register_new_node(thread, ctrl);
2803 Node* buffer_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(buffer_offset));
2804 phase->register_new_node(buffer_adr, ctrl);
2805 Node* index_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(index_offset));
2806 phase->register_new_node(index_adr, ctrl);
2807
2808 BasicType index_bt = TypeX_X->basic_type();
2809 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
2810 const TypePtr* adr_type = TypeRawPtr::BOTTOM;
2811 Node* index = new LoadXNode(ctrl, raw_mem, index_adr, adr_type, TypeX_X, MemNode::unordered);
2812 phase->register_new_node(index, ctrl);
2813 Node* index_cmp = new CmpXNode(index, phase->igvn().MakeConX(0));
2814 phase->register_new_node(index_cmp, ctrl);
2815 Node* index_test = new BoolNode(index_cmp, BoolTest::ne);
2816 phase->register_new_node(index_test, ctrl);
2817 IfNode* queue_full_iff = new IfNode(ctrl, index_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
2818 phase->register_control(queue_full_iff, loop, ctrl);
2819 Node* not_full = new IfTrueNode(queue_full_iff);
2820 phase->register_control(not_full, loop, queue_full_iff);
2821 Node* full = new IfFalseNode(queue_full_iff);
2822 phase->register_control(full, loop, queue_full_iff);
2823
2824 ctrl = not_full;
2825
2826 Node* next_index = new SubXNode(index, phase->igvn().MakeConX(sizeof(intptr_t)));
2827 phase->register_new_node(next_index, ctrl);
2828
2829 Node* buffer = new LoadPNode(ctrl, raw_mem, buffer_adr, adr_type, TypeRawPtr::NOTNULL, MemNode::unordered);
2830 phase->register_new_node(buffer, ctrl);
2831 Node *log_addr = new AddPNode(phase->C->top(), buffer, next_index);
2832 phase->register_new_node(log_addr, ctrl);
2833 Node* log_store = new StorePNode(ctrl, raw_mem, log_addr, adr_type, pre_val, MemNode::unordered);
2834 phase->register_new_node(log_store, ctrl);
2835 // update the index
2836 Node* index_update = new StoreXNode(ctrl, log_store, index_adr, adr_type, next_index, MemNode::unordered);
2837 phase->register_new_node(index_update, ctrl);
2838
2839 unstable_region->init_req(3, ctrl);
2840 unstable_phi->init_req(3, index_update);
2841
2842 ctrl = full;
2843
2844 Node* base = find_bottom_mem(ctrl, phase);
2845
2846 MergeMemNode* mm = MergeMemNode::make(base);
2847 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
2848 phase->register_new_node(mm, ctrl);
2849
2850 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);
2851 call->init_req(TypeFunc::Control, ctrl);
2852 call->init_req(TypeFunc::I_O, phase->C->top());
2853 call->init_req(TypeFunc::Memory, mm);
2854 call->init_req(TypeFunc::FramePtr, phase->C->top());
2855 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
2856 call->init_req(TypeFunc::Parms, pre_val);
2857 call->init_req(TypeFunc::Parms+1, thread);
2858 phase->register_control(call, loop, ctrl);
2859
2860 Node* ctrl_proj = new ProjNode(call, TypeFunc::Control);
2861 phase->register_control(ctrl_proj, loop, call);
2862 Node* mem_proj = new ProjNode(call, TypeFunc::Memory);
2863 phase->register_new_node(mem_proj, call);
2864
2865 unstable_region->init_req(4, ctrl_proj);
2866 unstable_phi->init_req(4, mem_proj);
2867
2868 phase->register_control(unstable_region, loop, null_val != NULL ? null_val->in(0) : evac_in_progress->in(0));
2869 phase->register_new_node(unstable_phi, unstable_region);
2870 phase->register_control(stable_test_region, loop, heap_stable->in(0));
2871 phase->register_new_node(stable_test_phi, stable_test_region);
2872
2873 fix_ctrl(barrier, stable_test_region, fixer, uses, uses_to_ignore, last, phase);
2874 for(uint next = 0; next < uses.size(); next++ ) {
2875 Node *n = uses.at(next);
2876 assert(phase->get_ctrl(n) == init_ctrl, "bad control");
2877 assert(n != init_raw_mem, "should leave input raw mem above the barrier");
2878 phase->set_ctrl(n, stable_test_region);
2879 follow_barrier_uses(n, init_ctrl, uses, phase);
2880 }
2881 fixer.fix_mem(init_ctrl, stable_test_region, init_raw_mem, raw_mem_for_ctrl, stable_test_phi, uses);
2882
2883 phase->igvn().replace_node(barrier, pre_val);
2884 }
2885
2886 for (int i = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count(); i > 0; i--) {
2887 int cnt = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count();
2888 ShenandoahWriteBarrierNode* wb = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barrier(i-1);
2889
2890 uint last = phase->C->unique();
2891 Node* ctrl = phase->get_ctrl(wb);
2892
2893 Node* raw_mem = fixer.find_mem(ctrl, wb);
2894 Node* init_raw_mem = raw_mem;
2895 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL);
2896 int alias = phase->C->get_alias_index(wb->adr_type());
2897 Node* wb_mem = wb->in(Memory);
2898 Node* init_wb_mem = wb_mem;
2899
2900 Node* val = wb->in(ValueIn);
2901 Node* wbproj = wb->find_out_with(Op_ShenandoahWBMemProj);
2902 IdealLoopTree *loop = phase->get_loop(ctrl);
2903
2904 assert(val->Opcode() != Op_ShenandoahWriteBarrier, "No chain of write barriers");
2905
2906 CallStaticJavaNode* unc = wb->pin_and_expand_null_check(phase->igvn());
2907 Node* unc_ctrl = NULL;
2908 if (unc != NULL) {
2909 if (val->in(0) != ctrl) {
2910 unc = NULL;
2911 } else {
2912 unc_ctrl = val->in(0);
2913 }
2914 }
2915
2916 Node* uncasted_val = val;
2917 if (unc != NULL) {
2918 uncasted_val = val->in(1);
2919 }
2920
2921 Node* evac_in_progress = NULL;
2922 Node* evac_not_in_progress = NULL;
2923 Node* heap_stable_ctrl = NULL;
2924 Node* null_ctrl = NULL;
2925 test_evacuation_in_progress(ctrl, val, raw_mem, evac_in_progress, evac_not_in_progress, heap_stable_ctrl, null_ctrl, phase);
2926 IfNode* evacuation_iff = evac_in_progress->in(0)->as_If();
2927 IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
2928
2929 Node* evacuation_region = new RegionNode(5);
2930 Node* evacuation_val_phi = new PhiNode(evacuation_region, uncasted_val->bottom_type()->is_oopptr()->cast_to_nonconst());
2931 Node* evacuation_mem_phi = PhiNode::make(evacuation_region, wb_mem, Type::MEMORY, phase->C->alias_type(wb->adr_type())->adr_type());
2932 Node* evacuation_raw_mem_phi = PhiNode::make(evacuation_region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
2933 Node* region = new RegionNode(3);
2934 Node* val_phi = new PhiNode(region, uncasted_val->bottom_type()->is_oopptr()->cast_to_nonconst());
2935 Node* mem_phi = PhiNode::make(region, wb_mem, Type::MEMORY, phase->C->alias_type(wb->adr_type())->adr_type());
2936 Node* raw_mem_phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
2937
2938 if (null_ctrl != NULL) {
2939 evacuation_region->init_req(4, null_ctrl);
2940 evacuation_val_phi->init_req(4, phase->igvn().zerocon(T_OBJECT));
2941 evacuation_mem_phi->init_req(4, wb_mem);
2942 evacuation_raw_mem_phi->init_req(4, raw_mem);
2943 } else {
2944 evacuation_region->del_req(4);
2945 evacuation_val_phi->del_req(4);
2946 evacuation_mem_phi->del_req(4);
2947 evacuation_raw_mem_phi->del_req(4);
2948 }
2949
2950 Node* unc_region = NULL;
2951 evacuation_not_in_progress(evac_not_in_progress, val, unc_ctrl, raw_mem, wb_mem,
2952 evacuation_region, evacuation_val_phi, evacuation_mem_phi, evacuation_raw_mem_phi, unc_region,
2953 phase);
2954
2955 heap_stable(heap_stable_ctrl, val, unc_ctrl, init_raw_mem, init_wb_mem, region, val_phi, mem_phi, raw_mem_phi,
2956 unc_region, phase);
2957
2958 evacuation_in_progress(evac_in_progress, val, evacuation_iff, unc, unc_ctrl,
2959 raw_mem, wb_mem, evacuation_region, evacuation_val_phi, evacuation_mem_phi, evacuation_raw_mem_phi,
2960 unc_region, alias, uses,
2961 phase);
2962 region->init_req(2, evacuation_region);
2963 val_phi->init_req(2, evacuation_val_phi);
2964 mem_phi->init_req(2, evacuation_mem_phi);
2965 raw_mem_phi->init_req(2, evacuation_raw_mem_phi);
2966 phase->register_control(evacuation_region, loop, evacuation_iff);
2967 phase->register_new_node(evacuation_val_phi, evacuation_region);
2968 phase->register_new_node(evacuation_mem_phi, evacuation_region);
2969 phase->register_new_node(evacuation_raw_mem_phi, evacuation_region);
2970
2971 phase->register_control(region, loop, heap_stable_iff);
2972
2973 Node* out_val = val_phi;
2974 phase->register_new_node(val_phi, region);
2975 phase->register_new_node(mem_phi, region);
2976 phase->register_new_node(raw_mem_phi, region);
2977
2978 fix_ctrl(wb, region, fixer, uses, uses_to_ignore, last, phase);
2979
2980 phase->igvn().replace_input_of(wbproj, 0, phase->C->top());
2981 phase->lazy_replace(wbproj, mem_phi);
2982
2983 if (unc != NULL) {
2984 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
2985 Node* u = val->fast_out(i);
2986 Node* c = phase->ctrl_or_self(u);
2987 if (u != wb && (c != ctrl || is_dominator_same_ctrl(c, wb, u, phase))) {
2988 phase->igvn().rehash_node_delayed(u);
2989 int nb = u->replace_edge(val, out_val);
2990 --i, imax -= nb;
2991 }
2992 }
2993 if (val->outcnt() == 0) {
2994 phase->lazy_update(val, out_val);
2995 phase->igvn()._worklist.push(val);
2996 }
2997 }
2998 phase->lazy_replace(wb, out_val);
2999
3000 follow_barrier_uses(mem_phi, ctrl, uses, phase);
3001 follow_barrier_uses(out_val, ctrl, uses, phase);
3002
3003 for(uint next = 0; next < uses.size(); next++ ) {
3004 Node *n = uses.at(next);
3005 assert(phase->get_ctrl(n) == ctrl, "bad control");
3006 assert(n != init_raw_mem, "should leave input raw mem above the barrier");
3007 phase->set_ctrl(n, region);
3008 follow_barrier_uses(n, ctrl, uses, phase);
3009 }
3010
3011 // The slow path call produces memory: hook the raw memory phi
3012 // from the expanded write barrier with the rest of the graph
3013 // which may require adding memory phis at every post dominated
3014 // region and at enclosing loop heads. Use the memory state
3015 // collected in memory_nodes to fix the memory graph. Update that
3016 // memory state as we go.
3017 fixer.fix_mem(ctrl, region, init_raw_mem, raw_mem_for_ctrl, raw_mem_phi, uses);
3018 assert(ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count() == cnt - 1, "not replaced");
3019 }
3020
3021 assert(ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count() == 0, "all write barrier nodes should have been replaced");
3022 }
3023
3024 void ShenandoahWriteBarrierNode::move_evacuation_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
3025 // move test and its mem barriers out of the loop
3026 assert(is_evacuation_in_progress_test(iff), "inconsistent");
3027
3028 IdealLoopTree *loop = phase->get_loop(iff);
3029 Node* loop_head = loop->_head;
3030 Node* entry_c = loop_head->in(LoopNode::EntryControl);
3031
3032 Node* load = iff->in(1)->in(1)->in(1)->in(1);
3033 assert(is_gc_state_load(load), "broken");
3034 if (!phase->is_dominator(load->in(0), entry_c)) {
3035 Node* mem_ctrl = NULL;
3036 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
3037 phase->igvn().replace_input_of(load, MemNode::Memory, mem);
3038 phase->igvn().replace_input_of(load, 0, entry_c);
3039 phase->set_ctrl_and_loop(load, entry_c);
3040 }
3041 }
3042
3043 void ShenandoahWriteBarrierNode::move_heap_stable_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
3044 IdealLoopTree *loop = phase->get_loop(iff);
3045 Node* loop_head = loop->_head;
3046 Node* entry_c = loop_head->in(LoopNode::EntryControl);
3047
3048 Node* load = iff->in(1)->in(1)->in(1)->in(1);
3049 assert(is_gc_state_load(load), "broken");
3050 if (!phase->is_dominator(load->in(0), entry_c)) {
3051 Node* mem_ctrl = NULL;
3052 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
3053 phase->igvn().replace_input_of(load, MemNode::Memory, mem);
3054 phase->igvn().replace_input_of(load, 0, entry_c);
3055 phase->set_ctrl_and_loop(load, entry_c);
3056 }
3057 }
3058
3059 void ShenandoahWriteBarrierNode::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) {
3060 assert(is_evacuation_in_progress_test(n) || is_heap_stable_test(n), "no other tests");
3061 if (phase->identical_backtoback_ifs(n)) {
3062 Node* n_ctrl = is_evacuation_in_progress_test(n) ? ShenandoahWriteBarrierNode::evacuation_in_progress_test_ctrl(n) : n->in(0);
3063 if (phase->can_split_if(n_ctrl)) {
3064 IfNode* dom_if = phase->idom(n_ctrl)->as_If();
3065 if (is_heap_stable_test(n)) {
3066 Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1);
3067 assert(is_gc_state_load(gc_state_load), "broken");
3068 Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1);
3069 assert(is_gc_state_load(dom_gc_state_load), "broken");
3070 if (gc_state_load != dom_gc_state_load) {
3071 phase->igvn().replace_node(gc_state_load, dom_gc_state_load);
3072 }
3073 }
3074 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
3075 Node* proj_true = dom_if->proj_out(1);
3076 Node* proj_false = dom_if->proj_out(0);
3077 Node* con_true = phase->igvn().makecon(TypeInt::ONE);
3078 Node* con_false = phase->igvn().makecon(TypeInt::ZERO);
3079
3080 for (uint i = 1; i < n_ctrl->req(); i++) {
3081 if (phase->is_dominator(proj_true, n_ctrl->in(i))) {
3082 bolphi->init_req(i, con_true);
3083 } else {
3084 assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if");
3085 bolphi->init_req(i, con_false);
3086 }
3087 }
3088 phase->register_new_node(bolphi, n_ctrl);
3089 phase->igvn().replace_input_of(n, 1, bolphi);
3090 phase->do_split_if(n);
3091 }
3092 }
3093 }
3094
3095 void ShenandoahWriteBarrierNode::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) {
3096 Node_List heap_stable_tests;
3097 Node_List evacuation_tests;
3098 Node_List gc_state_loads;
3099
3100 stack.push(phase->C->start(), 0);
3101 do {
3102 Node* n = stack.node();
3103 uint i = stack.index();
3104
3105 if (i < n->outcnt()) {
3106 Node* u = n->raw_out(i);
3107 stack.set_index(i+1);
3108 if (!visited.test_set(u->_idx)) {
3109 stack.push(u, 0);
3110 }
3111 } else {
3112 stack.pop();
3113 if (n->is_If() && ShenandoahWriteBarrierNode::is_evacuation_in_progress_test(n)) {
3114 evacuation_tests.push(n);
3115 }
3116 if (ShenandoahCommonGCStateLoads && ShenandoahWriteBarrierNode::is_gc_state_load(n)) {
3117 gc_state_loads.push(n);
3118 }
3119 if (n->is_If() && ShenandoahWriteBarrierNode::is_heap_stable_test(n)) {
3120 heap_stable_tests.push(n);
3121 }
3122 }
3123 } while (stack.size() > 0);
3124
3125 bool progress;
3126 do {
3127 progress = false;
3128 for (uint i = 0; i < gc_state_loads.size(); i++) {
3129 Node* n = gc_state_loads.at(i);
3130 if (n->outcnt() != 0) {
3131 progress |= ShenandoahWriteBarrierNode::try_common_gc_state_load(n, phase);
3132 }
3133 }
3134 } while (progress);
3135
3136 for (uint i = 0; i < heap_stable_tests.size(); i++) {
3137 Node* n = heap_stable_tests.at(i);
3138 assert(is_heap_stable_test(n), "only evacuation test");
3139 merge_back_to_back_tests(n, phase);
3140 }
3141
3142 if (!phase->C->major_progress()) {
3143 for (uint i = 0; i < evacuation_tests.size(); i++) {
3144 Node* n = evacuation_tests.at(i);
3145 assert(is_evacuation_in_progress_test(n), "only evacuation test");
3146 merge_back_to_back_tests(n, phase);
3147 }
3148 }
3149
3150 if (!phase->C->major_progress()) {
3151 VectorSet seen(Thread::current()->resource_area());
3152 for (uint i = 0; i < heap_stable_tests.size(); i++) {
3153 Node* n = heap_stable_tests.at(i);
3154 IdealLoopTree* loop = phase->get_loop(n);
3155 if (loop != phase->ltree_root() &&
3156 loop->_child == NULL &&
3157 !loop->_irreducible) {
3158 LoopNode* head = loop->_head->as_Loop();
3159 if ((!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) &&
3160 !seen.test_set(head->_idx) &&
3161 loop->policy_unswitching(phase, true)) {
3162 IfNode* iff = phase->find_unswitching_candidate(loop, true);
3163 if (iff != NULL && (is_evacuation_in_progress_test(iff) || is_heap_stable_test(iff))) {
3164 if (head->is_strip_mined()) {
3165 head->verify_strip_mined(0);
3166 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop();
3167 OuterStripMinedLoopEndNode* le = head->outer_loop_end();
3168 Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl));
3169 phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl));
3170 Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt);
3171 phase->register_control(new_le, phase->get_loop(le), le->in(0));
3172 phase->lazy_replace(outer, new_outer);
3173 phase->lazy_replace(le, new_le);
3174 head->clear_strip_mined();
3175 }
3176 phase->do_unswitching(loop, old_new, true);
3177 }
3178 }
3179 }
3180 }
3181 }
3182 }
3183
3184 #ifdef ASSERT
3185 void ShenandoahBarrierNode::verify_raw_mem(RootNode* root) {
3186 const bool trace = false;
3187 ResourceMark rm;
3188 Unique_Node_List nodes;
3189 Unique_Node_List controls;
3190 Unique_Node_List memories;
3191
3192 nodes.push(root);
3193 for (uint next = 0; next < nodes.size(); next++) {
3194 Node *n = nodes.at(next);
3195 if (n->Opcode() == Op_CallLeafNoFP &&
3196 ShenandoahBarrierSetAssembler::is_shenandoah_wb_C_call(n->as_Call()->entry_point())) {
3197 controls.push(n);
3198 if (trace) { tty->print("XXXXXX verifying"); n->dump(); }
3199 for (uint next2 = 0; next2 < controls.size(); next2++) {
3200 Node *m = controls.at(next2);
3201 if (!m->is_Loop() || controls.member(m->in(LoopNode::EntryControl)) || 1) {
3202 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3203 Node* u = m->fast_out(i);
3204 if (u->is_CFG() && !u->is_Root() &&
3205 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1) &&
3206 !(u->is_Region() && u->unique_ctrl_out()->Opcode() == Op_Halt)) {
3207 if (trace) { tty->print("XXXXXX pushing control"); u->dump(); }
3208 controls.push(u);
3209 }
3210 }
3211 }
3212 }
3213 memories.push(n->as_Call()->proj_out(TypeFunc::Memory));
3214 for (uint next2 = 0; next2 < memories.size(); next2++) {
3215 Node *m = memories.at(next2);
3216 assert(m->bottom_type() == Type::MEMORY, "");
3217 if (!m->is_Phi() || !m->in(0)->is_Loop() || controls.member(m->in(0)->in(LoopNode::EntryControl)) || 1) {
3218 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3219 Node* u = m->fast_out(i);
3220 if (u->bottom_type() == Type::MEMORY && (u->is_Mem() || u->is_ClearArray())) {
3221 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
3222 memories.push(u);
3223 } else if (u->is_LoadStore()) {
3224 if (trace) { tty->print("XXXXXX pushing memory"); u->find_out_with(Op_SCMemProj)->dump(); }
3225 memories.push(u->find_out_with(Op_SCMemProj));
3226 } else if (u->is_MergeMem() && u->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == m) {
3227 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
3228 memories.push(u);
3229 } else if (u->is_Phi()) {
3230 assert(u->bottom_type() == Type::MEMORY, "");
3231 if (u->adr_type() == TypeRawPtr::BOTTOM || u->adr_type() == TypePtr::BOTTOM) {
3232 assert(controls.member(u->in(0)), "");
3233 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
3234 memories.push(u);
3235 }
3236 } else if (u->is_SafePoint() || u->is_MemBar()) {
3237 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
3238 Node* uu = u->fast_out(j);
3239 if (uu->bottom_type() == Type::MEMORY) {
3240 if (trace) { tty->print("XXXXXX pushing memory"); uu->dump(); }
3241 memories.push(uu);
3242 }
3243 }
3244 }
3245 }
3246 }
3247 }
3248 for (uint next2 = 0; next2 < controls.size(); next2++) {
3249 Node *m = controls.at(next2);
3250 if (m->is_Region()) {
3251 bool all_in = true;
3252 for (uint i = 1; i < m->req(); i++) {
3253 if (!controls.member(m->in(i))) {
3254 all_in = false;
3255 break;
3256 }
3257 }
3258 if (trace) { tty->print("XXX verifying %s", all_in ? "all in" : ""); m->dump(); }
3259 bool found_phi = false;
3260 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax && !found_phi; j++) {
3261 Node* u = m->fast_out(j);
3262 if (u->is_Phi() && memories.member(u)) {
3263 found_phi = true;
3264 for (uint i = 1; i < u->req() && found_phi; i++) {
3265 Node* k = u->in(i);
3266 if (memories.member(k) != controls.member(m->in(i))) {
3267 found_phi = false;
3268 }
3269 }
3270 }
3271 }
3272 assert(found_phi || all_in, "");
3273 }
3274 }
3275 controls.clear();
3276 memories.clear();
3277 }
3278 for( uint i = 0; i < n->len(); ++i ) {
3279 Node *m = n->in(i);
3280 if (m != NULL) {
3281 nodes.push(m);
3282 }
3283 }
3284 }
3285 }
3286 #endif
3287
3288 const Type* ShenandoahEnqueueBarrierNode::bottom_type() const {
3289 if (in(1) == NULL || in(1)->is_top()) {
3290 return Type::TOP;
3291 }
3292 const Type* t = in(1)->bottom_type();
3293 if (t == TypePtr::NULL_PTR) {
3294 return t;
3295 }
3296 return t->is_oopptr()->cast_to_nonconst();
3297 }
3298
3299 const Type* ShenandoahEnqueueBarrierNode::Value(PhaseGVN* phase) const {
3300 if (in(1) == NULL) {
3301 return Type::TOP;
3302 }
3303 const Type* t = phase->type(in(1));
3304 if (t == Type::TOP) {
3305 return Type::TOP;
3306 }
3307 if (t == TypePtr::NULL_PTR) {
3308 return t;
3309 }
3310 return t->is_oopptr()->cast_to_nonconst();
3311 }
3312
3313 int ShenandoahEnqueueBarrierNode::needed(Node* n) {
3314 if (n == NULL ||
3315 n->is_Allocate() ||
3316 n->bottom_type() == TypePtr::NULL_PTR ||
3317 n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL) {
3318 return NotNeeded;
3319 }
3320 if (n->is_Phi() ||
3321 n->is_CMove()) {
3322 return MaybeNeeded;
3323 }
3324 return Needed;
3325 }
3326
3327 Node* ShenandoahEnqueueBarrierNode::next(Node* n) {
3328 for (;;) {
3329 if (n == NULL) {
3330 return n;
3331 } else if (n->bottom_type() == TypePtr::NULL_PTR) {
3332 return n;
3333 } else if (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL) {
3334 return n;
3335 } else if (n->is_ConstraintCast() ||
3336 n->Opcode() == Op_DecodeN ||
3337 n->Opcode() == Op_EncodeP) {
3338 n = n->in(1);
3339 } else if (n->is_Proj()) {
3340 n = n->in(0);
3341 } else {
3342 return n;
3343 }
3344 }
3345 ShouldNotReachHere();
3346 return NULL;
3347 }
3348
3349
3350 Node* ShenandoahEnqueueBarrierNode::Identity(PhaseGVN* phase) {
3351 PhaseIterGVN* igvn = phase->is_IterGVN();
3352
3353 Node* n = next(in(1));
3354
3355 int cont = needed(n);
3356
3357 if (cont == NotNeeded) {
3358 return in(1);
3359 } else if (cont == MaybeNeeded) {
3360 if (igvn == NULL) {
3361 phase->record_for_igvn(this);
3362 return this;
3363 } else {
3364 ResourceMark rm;
3365 Unique_Node_List wq;
3366 uint wq_i = 0;
3367
3368 for (;;) {
3369 if (n->is_Phi()) {
3370 for (uint i = 1; i < n->req(); i++) {
3371 Node* m = n->in(i);
3372 if (m != NULL) {
3373 wq.push(m);
3374 }
3375 }
3376 } else {
3377 assert(n->is_CMove(), "nothing else here");
3378 Node* m = n->in(CMoveNode::IfFalse);
3379 wq.push(m);
3380 m = n->in(CMoveNode::IfTrue);
3381 wq.push(m);
3382 }
3383 Node* orig_n = NULL;
3384 do {
3385 if (wq_i >= wq.size()) {
3386 return in(1);
3387 }
3388 n = wq.at(wq_i);
3389 wq_i++;
3390 orig_n = n;
3391 n = next(n);
3392 cont = needed(n);
3393 if (cont == Needed) {
3394 return this;
3395 }
3396 } while (cont != MaybeNeeded || (orig_n != n && wq.member(n)));
3397 }
3398 }
3399 }
3400
3401 return this;
3402 }
3403
3404 #ifdef ASSERT
3405 static bool has_never_branch(Node* root) {
3406 for (uint i = 1; i < root->req(); i++) {
3407 Node* in = root->in(i);
3408 if (in != NULL && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->Opcode() == Op_NeverBranch) {
3409 return true;
3410 }
3411 }
3412 return false;
3413 }
3414 #endif
3415
3416 void MemoryGraphFixer::collect_memory_nodes() {
3417 Node_Stack stack(0);
3418 VectorSet visited(Thread::current()->resource_area());
3419 Node_List regions;
3420
3421 // Walk the raw memory graph and create a mapping from CFG node to
3422 // memory node. Exclude phis for now.
3423 stack.push(_phase->C->root(), 1);
3424 do {
3425 Node* n = stack.node();
3426 int opc = n->Opcode();
3427 uint i = stack.index();
3428 if (i < n->req()) {
3429 Node* mem = NULL;
3430 if (opc == Op_Root) {
3431 Node* in = n->in(i);
3432 int in_opc = in->Opcode();
3433 if (in_opc == Op_Return || in_opc == Op_Rethrow) {
3434 mem = in->in(TypeFunc::Memory);
3435 } else if (in_opc == Op_Halt) {
3436 if (!in->in(0)->is_Region()) {
3437 Node* proj = in->in(0);
3438 assert(proj->is_Proj(), "");
3439 Node* in = proj->in(0);
3440 assert(in->is_CallStaticJava() || in->Opcode() == Op_NeverBranch || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
3441 if (in->is_CallStaticJava()) {
3442 mem = in->in(TypeFunc::Memory);
3443 } else if (in->Opcode() == Op_Catch) {
3444 Node* call = in->in(0)->in(0);
3445 assert(call->is_Call(), "");
3446 mem = call->in(TypeFunc::Memory);
3447 }
3448 }
3449 } else {
3450 #ifdef ASSERT
3451 n->dump();
3452 in->dump();
3453 #endif
3454 ShouldNotReachHere();
3455 }
3456 } else {
3457 assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
3458 assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
3459 mem = n->in(i);
3460 }
3461 i++;
3462 stack.set_index(i);
3463 if (mem == NULL) {
3464 continue;
3465 }
3466 for (;;) {
3467 if (visited.test_set(mem->_idx) || mem->is_Start()) {
3468 break;
3469 }
3470 if (mem->is_Phi()) {
3471 stack.push(mem, 2);
3472 mem = mem->in(1);
3473 } else if (mem->is_Proj()) {
3474 stack.push(mem, mem->req());
3475 mem = mem->in(0);
3476 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
3477 mem = mem->in(TypeFunc::Memory);
3478 } else if (mem->is_MergeMem()) {
3479 MergeMemNode* mm = mem->as_MergeMem();
3480 mem = mm->memory_at(_alias);
3481 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
3482 assert(_alias == Compile::AliasIdxRaw, "");
3483 stack.push(mem, mem->req());
3484 mem = mem->in(MemNode::Memory);
3485 } else if (mem->Opcode() == Op_ShenandoahWriteBarrier) {
3486 assert(_alias != Compile::AliasIdxRaw, "");
3487 mem = mem->in(ShenandoahBarrierNode::Memory);
3488 } else {
3489 #ifdef ASSERT
3490 mem->dump();
3491 #endif
3492 ShouldNotReachHere();
3493 }
3494 }
3495 } else {
3496 if (n->is_Phi()) {
3497 // Nothing
3498 } else if (!n->is_Root()) {
3499 Node* c = get_ctrl(n);
3500 _memory_nodes.map(c->_idx, n);
3501 }
3502 stack.pop();
3503 }
3504 } while(stack.is_nonempty());
3505
3506 // Iterate over CFG nodes in rpo and propagate memory state to
3507 // compute memory state at regions, creating new phis if needed.
3508 Node_List rpo_list;
3509 visited.Clear();
3510 _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
3511 Node* root = rpo_list.pop();
3512 assert(root == _phase->C->root(), "");
3513
3514 const bool trace = false;
3515 #ifdef ASSERT
3516 if (trace) {
3517 for (int i = rpo_list.size() - 1; i >= 0; i--) {
3518 Node* c = rpo_list.at(i);
3519 if (_memory_nodes[c->_idx] != NULL) {
3520 tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump();
3521 }
3522 }
3523 }
3524 #endif
3525 uint last = _phase->C->unique();
3526
3527 #ifdef ASSERT
3528 uint8_t max_depth = 0;
3529 for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
3530 IdealLoopTree* lpt = iter.current();
3531 max_depth = MAX2(max_depth, lpt->_nest);
3532 }
3533 #endif
3534
3535 bool progress = true;
3536 int iteration = 0;
3537 Node_List dead_phis;
3538 while (progress) {
3539 progress = false;
3540 iteration++;
3541 assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop(), "");
3542 if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
3543 IdealLoopTree* last_updated_ilt = NULL;
3544 for (int i = rpo_list.size() - 1; i >= 0; i--) {
3545 Node* c = rpo_list.at(i);
3546
3547 Node* prev_mem = _memory_nodes[c->_idx];
3548 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
3549 Node* prev_region = regions[c->_idx];
3550 Node* unique = NULL;
3551 for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
3552 Node* m = _memory_nodes[c->in(j)->_idx];
3553 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");
3554 if (m != NULL) {
3555 if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
3556 assert(c->is_Loop() && j == LoopNode::LoopBackControl || _phase->C->has_irreducible_loop(), "");
3557 // continue
3558 } else if (unique == NULL) {
3559 unique = m;
3560 } else if (m == unique) {
3561 // continue
3562 } else {
3563 unique = NodeSentinel;
3564 }
3565 }
3566 }
3567 assert(unique != NULL, "empty phi???");
3568 if (unique != NodeSentinel) {
3569 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c) {
3570 dead_phis.push(prev_region);
3571 }
3572 regions.map(c->_idx, unique);
3573 } else {
3574 Node* phi = NULL;
3575 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
3576 phi = prev_region;
3577 for (uint k = 1; k < c->req(); k++) {
3578 Node* m = _memory_nodes[c->in(k)->_idx];
3579 assert(m != NULL, "expect memory state");
3580 phi->set_req(k, m);
3581 }
3582 } else {
3583 for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == NULL; j++) {
3584 Node* u = c->fast_out(j);
3585 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
3586 (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
3587 phi = u;
3588 for (uint k = 1; k < c->req() && phi != NULL; k++) {
3589 Node* m = _memory_nodes[c->in(k)->_idx];
3590 assert(m != NULL, "expect memory state");
3591 if (u->in(k) != m) {
3592 phi = NULL;
3593 }
3594 }
3595 }
3596 }
3597 if (phi == NULL) {
3598 phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
3599 for (uint k = 1; k < c->req(); k++) {
3600 Node* m = _memory_nodes[c->in(k)->_idx];
3601 assert(m != NULL, "expect memory state");
3602 phi->init_req(k, m);
3603 }
3604 }
3605 }
3606 assert(phi != NULL, "");
3607 regions.map(c->_idx, phi);
3608 }
3609 Node* current_region = regions[c->_idx];
3610 if (current_region != prev_region) {
3611 progress = true;
3612 if (prev_region == prev_mem) {
3613 _memory_nodes.map(c->_idx, current_region);
3614 }
3615 }
3616 } else if (prev_mem == NULL || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
3617 Node* m = _memory_nodes[_phase->idom(c)->_idx];
3618 assert(m != NULL, "expect memory state");
3619 if (m != prev_mem) {
3620 _memory_nodes.map(c->_idx, m);
3621 progress = true;
3622 }
3623 }
3624 #ifdef ASSERT
3625 if (trace) { tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); }
3626 #endif
3627 }
3628 }
3629
3630 // Replace existing phi with computed memory state for that region
3631 // if different (could be a new phi or a dominating memory node if
3632 // that phi was found to be useless).
3633 while (dead_phis.size() > 0) {
3634 Node* n = dead_phis.pop();
3635 n->replace_by(_phase->C->top());
3636 n->destruct();
3637 }
3638 for (int i = rpo_list.size() - 1; i >= 0; i--) {
3639 Node* c = rpo_list.at(i);
3640 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
3641 Node* n = regions[c->_idx];
3642 if (n->is_Phi() && n->_idx >= last && n->in(0) == c) {
3643 _phase->register_new_node(n, c);
3644 }
3645 }
3646 }
3647 for (int i = rpo_list.size() - 1; i >= 0; i--) {
3648 Node* c = rpo_list.at(i);
3649 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
3650 Node* n = regions[c->_idx];
3651 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
3652 Node* u = c->fast_out(i);
3653 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
3654 u != n) {
3655 if (u->adr_type() == TypePtr::BOTTOM) {
3656 fix_memory_uses(u, n, n, c);
3657 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
3658 _phase->lazy_replace(u, n);
3659 --i; --imax;
3660 }
3661 }
3662 }
3663 }
3664 }
3665 }
3666
3667 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
3668 Node* c = _phase->get_ctrl(n);
3669 if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Call()) {
3670 assert(c == n->in(0), "");
3671 CallNode* call = c->as_Call();
3672 CallProjections projs;
3673 call->extract_projections(&projs, true, false);
3674 if (projs.catchall_memproj != NULL) {
3675 if (projs.fallthrough_memproj == n) {
3676 c = projs.fallthrough_catchproj;
3677 } else {
3678 assert(projs.catchall_memproj == n, "");
3679 c = projs.catchall_catchproj;
3680 }
3681 }
3682 }
3683 return c;
3684 }
3685
3686 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
3687 if (_phase->has_ctrl(n))
3688 return get_ctrl(n);
3689 else {
3690 assert (n->is_CFG(), "must be a CFG node");
3691 return n;
3692 }
3693 }
3694
3695 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
3696 return m != NULL && get_ctrl(m) == c;
3697 }
3698
3699 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
3700 assert(n == NULL || _phase->ctrl_or_self(n) == ctrl, "");
3701 Node* mem = _memory_nodes[ctrl->_idx];
3702 Node* c = ctrl;
3703 while (!mem_is_valid(mem, c) &&
3704 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
3705 c = _phase->idom(c);
3706 mem = _memory_nodes[c->_idx];
3707 }
3708 if (n != NULL && mem_is_valid(mem, c)) {
3709 while (!ShenandoahWriteBarrierNode::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
3710 mem = next_mem(mem, _alias);
3711 }
3712 if (mem->is_MergeMem()) {
3713 mem = mem->as_MergeMem()->memory_at(_alias);
3714 }
3715 if (!mem_is_valid(mem, c)) {
3716 do {
3717 c = _phase->idom(c);
3718 mem = _memory_nodes[c->_idx];
3719 } while (!mem_is_valid(mem, c) &&
3720 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
3721 }
3722 }
3723 assert(mem->bottom_type() == Type::MEMORY, "");
3724 return mem;
3725 }
3726
3727 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
3728 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
3729 Node* use = region->fast_out(i);
3730 if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
3731 (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
3732 return true;
3733 }
3734 }
3735 return false;
3736 }
3737
3738 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
3739 assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
3740 const bool trace = false;
3741 DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
3742 DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
3743 GrowableArray<Node*> phis;
3744 if (mem_for_ctrl != mem) {
3745 Node* old = mem_for_ctrl;
3746 Node* prev = NULL;
3747 while (old != mem) {
3748 prev = old;
3749 if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
3750 assert(_alias == Compile::AliasIdxRaw, "");
3751 old = old->in(MemNode::Memory);
3752 } else if (old->Opcode() == Op_SCMemProj) {
3753 assert(_alias == Compile::AliasIdxRaw, "");
3754 old = old->in(0);
3755 } else if (old->Opcode() == Op_ShenandoahWBMemProj) {
3756 assert(_alias != Compile::AliasIdxRaw, "");
3757 old = old->in(0);
3758 } else if (old->Opcode() == Op_ShenandoahWriteBarrier) {
3759 assert(_alias != Compile::AliasIdxRaw, "");
3760 old = old->in(ShenandoahBarrierNode::Memory);
3761 } else {
3762 ShouldNotReachHere();
3763 }
3764 }
3765 assert(prev != NULL, "");
3766 if (new_ctrl != ctrl) {
3767 _memory_nodes.map(ctrl->_idx, mem);
3768 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
3769 }
3770 uint input = prev->Opcode() == Op_ShenandoahWriteBarrier ? (uint)ShenandoahBarrierNode::Memory : (uint)MemNode::Memory;
3771 _phase->igvn().replace_input_of(prev, input, new_mem);
3772 } else {
3773 uses.clear();
3774 _memory_nodes.map(new_ctrl->_idx, new_mem);
3775 uses.push(new_ctrl);
3776 for(uint next = 0; next < uses.size(); next++ ) {
3777 Node *n = uses.at(next);
3778 assert(n->is_CFG(), "");
3779 DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
3780 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3781 Node* u = n->fast_out(i);
3782 if (!u->is_Root() && u->is_CFG() && u != n) {
3783 Node* m = _memory_nodes[u->_idx];
3784 if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
3785 !has_mem_phi(u) &&
3786 u->unique_ctrl_out()->Opcode() != Op_Halt) {
3787 DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
3788 DEBUG_ONLY(if (trace && m != NULL) { tty->print("ZZZ mem"); m->dump(); });
3789
3790 if (!mem_is_valid(m, u) || !m->is_Phi()) {
3791 bool push = true;
3792 bool create_phi = true;
3793 if (_phase->is_dominator(new_ctrl, u)) {
3794 create_phi = false;
3795 } else if (!_phase->C->has_irreducible_loop()) {
3796 IdealLoopTree* loop = _phase->get_loop(ctrl);
3797 bool do_check = true;
3798 IdealLoopTree* l = loop;
3799 create_phi = false;
3800 while (l != _phase->ltree_root()) {
3801 if (_phase->is_dominator(l->_head, u) && _phase->is_dominator(_phase->idom(u), l->_head)) {
3802 create_phi = true;
3803 do_check = false;
3804 break;
3805 }
3806 l = l->_parent;
3807 }
3808
3809 if (do_check) {
3810 assert(!create_phi, "");
3811 IdealLoopTree* u_loop = _phase->get_loop(u);
3812 if (u_loop != _phase->ltree_root() && u_loop->is_member(loop)) {
3813 Node* c = ctrl;
3814 while (!_phase->is_dominator(c, u_loop->tail())) {
3815 c = _phase->idom(c);
3816 }
3817 if (!_phase->is_dominator(c, u)) {
3818 do_check = false;
3819 }
3820 }
3821 }
3822
3823 if (do_check && _phase->is_dominator(_phase->idom(u), new_ctrl)) {
3824 create_phi = true;
3825 }
3826 }
3827 if (create_phi) {
3828 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
3829 _phase->register_new_node(phi, u);
3830 phis.push(phi);
3831 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
3832 if (!mem_is_valid(m, u)) {
3833 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
3834 _memory_nodes.map(u->_idx, phi);
3835 } else {
3836 DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
3837 for (;;) {
3838 assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj() || m->Opcode() == Op_ShenandoahWriteBarrier /*|| m->is_MergeMem()*/, "");
3839 Node* next = NULL;
3840 if (m->is_Proj()) {
3841 next = m->in(0);
3842 } else if (m->is_Mem() || m->is_LoadStore()) {
3843 assert(_alias == Compile::AliasIdxRaw, "");
3844 next = m->in(MemNode::Memory);
3845 } else {
3846 assert(_alias != Compile::AliasIdxRaw, "");
3847 assert (m->Opcode() == Op_ShenandoahWriteBarrier, "");
3848 next = m->in(ShenandoahBarrierNode::Memory);
3849 }
3850 if (_phase->get_ctrl(next) != u) {
3851 break;
3852 }
3853 if (next->is_MergeMem()) {
3854 assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
3855 break;
3856 }
3857 if (next->is_Phi()) {
3858 assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
3859 break;
3860 }
3861 m = next;
3862 }
3863
3864 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
3865 assert(m->is_Mem() || m->is_LoadStore() || m->Opcode() == Op_ShenandoahWriteBarrier, "");
3866 uint input = (m->is_Mem() || m->is_LoadStore()) ? (uint)MemNode::Memory : (uint)ShenandoahBarrierNode::Memory;
3867 _phase->igvn().replace_input_of(m, input, phi);
3868 push = false;
3869 }
3870 } else {
3871 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
3872 }
3873 if (push) {
3874 uses.push(u);
3875 }
3876 }
3877 } else if (!mem_is_valid(m, u) &&
3878 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1)) {
3879 uses.push(u);
3880 }
3881 }
3882 }
3883 }
3884 for (int i = 0; i < phis.length(); i++) {
3885 Node* n = phis.at(i);
3886 Node* r = n->in(0);
3887 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
3888 for (uint j = 1; j < n->req(); j++) {
3889 Node* m = find_mem(r->in(j), NULL);
3890 _phase->igvn().replace_input_of(n, j, m);
3891 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
3892 }
3893 }
3894 }
3895 uint last = _phase->C->unique();
3896 MergeMemNode* mm = NULL;
3897 int alias = _alias;
3898 DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
3899 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
3900 Node* u = mem->out(i);
3901 if (u->_idx < last) {
3902 if (u->is_Mem()) {
3903 if (_phase->C->get_alias_index(u->adr_type()) == alias) {
3904 Node* m = find_mem(_phase->get_ctrl(u), u);
3905 if (m != mem) {
3906 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
3907 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
3908 --i;
3909 }
3910 }
3911 } else if (u->is_MergeMem()) {
3912 MergeMemNode* u_mm = u->as_MergeMem();
3913 if (u_mm->memory_at(alias) == mem) {
3914 MergeMemNode* newmm = NULL;
3915 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
3916 Node* uu = u->fast_out(j);
3917 assert(!uu->is_MergeMem(), "chain of MergeMems?");
3918 if (uu->is_Phi()) {
3919 assert(uu->adr_type() == TypePtr::BOTTOM, "");
3920 Node* region = uu->in(0);
3921 int nb = 0;
3922 for (uint k = 1; k < uu->req(); k++) {
3923 if (uu->in(k) == u) {
3924 Node* m = find_mem(region->in(k), NULL);
3925 if (m != mem) {
3926 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
3927 if (newmm == NULL || 1) {
3928 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
3929 }
3930 if (newmm != u) {
3931 _phase->igvn().replace_input_of(uu, k, newmm);
3932 nb++;
3933 --jmax;
3934 }
3935 }
3936 }
3937 }
3938 if (nb > 0) {
3939 --j;
3940 }
3941 } else {
3942 Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
3943 if (m != mem) {
3944 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
3945 if (newmm == NULL || 1) {
3946 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
3947 }
3948 if (newmm != u) {
3949 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
3950 --j, --jmax;
3951 }
3952 }
3953 }
3954 }
3955 }
3956 } else if (u->is_Phi()) {
3957 assert(u->bottom_type() == Type::MEMORY, "what else?");
3958 if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
3959 Node* region = u->in(0);
3960 bool replaced = false;
3961 for (uint j = 1; j < u->req(); j++) {
3962 if (u->in(j) == mem) {
3963 Node* m = find_mem(region->in(j), NULL);
3964 Node* nnew = m;
3965 if (m != mem) {
3966 if (u->adr_type() == TypePtr::BOTTOM) {
3967 mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
3968 nnew = mm;
3969 }
3970 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
3971 _phase->igvn().replace_input_of(u, j, nnew);
3972 replaced = true;
3973 }
3974 }
3975 }
3976 if (replaced) {
3977 --i;
3978 }
3979 }
3980 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
3981 u->adr_type() == NULL) {
3982 assert(u->adr_type() != NULL ||
3983 u->Opcode() == Op_Rethrow ||
3984 u->Opcode() == Op_Return ||
3985 u->Opcode() == Op_SafePoint ||
3986 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
3987 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
3988 u->Opcode() == Op_CallLeaf, "");
3989 Node* m = find_mem(_phase->ctrl_or_self(u), u);
3990 if (m != mem) {
3991 mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
3992 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
3993 --i;
3994 }
3995 } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
3996 Node* m = find_mem(_phase->ctrl_or_self(u), u);
3997 if (m != mem) {
3998 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
3999 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
4000 --i;
4001 }
4002 } else if (u->adr_type() != TypePtr::BOTTOM &&
4003 _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
4004 Node* m = find_mem(_phase->ctrl_or_self(u), u);
4005 assert(m != mem, "");
4006 // u is on the wrong slice...
4007 assert(u->is_ClearArray(), "");
4008 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
4009 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
4010 --i;
4011 }
4012 }
4013 }
4014 #ifdef ASSERT
4015 assert(new_mem->outcnt() > 0, "");
4016 for (int i = 0; i < phis.length(); i++) {
4017 Node* n = phis.at(i);
4018 assert(n->outcnt() > 0, "new phi must have uses now");
4019 }
4020 #endif
4021 }
4022
4023 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
4024 MergeMemNode* mm = MergeMemNode::make(mem);
4025 mm->set_memory_at(_alias, rep_proj);
4026 _phase->register_new_node(mm, rep_ctrl);
4027 return mm;
4028 }
4029
4030 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
4031 MergeMemNode* newmm = NULL;
4032 MergeMemNode* u_mm = u->as_MergeMem();
4033 Node* c = _phase->get_ctrl(u);
4034 if (_phase->is_dominator(c, rep_ctrl)) {
4035 c = rep_ctrl;
4036 } else {
4037 assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
4038 }
4039 if (u->outcnt() == 1) {
4040 if (u->req() > (uint)_alias && u->in(_alias) == mem) {
4041 _phase->igvn().replace_input_of(u, _alias, rep_proj);
4042 --i;
4043 } else {
4044 _phase->igvn().rehash_node_delayed(u);
4045 u_mm->set_memory_at(_alias, rep_proj);
4046 }
4047 newmm = u_mm;
4048 _phase->set_ctrl_and_loop(u, c);
4049 } else {
4050 // can't simply clone u and then change one of its input because
4051 // it adds and then removes an edge which messes with the
4052 // DUIterator
4053 newmm = MergeMemNode::make(u_mm->base_memory());
4054 for (uint j = 0; j < u->req(); j++) {
4055 if (j < newmm->req()) {
4056 if (j == (uint)_alias) {
4057 newmm->set_req(j, rep_proj);
4058 } else if (newmm->in(j) != u->in(j)) {
4059 newmm->set_req(j, u->in(j));
4060 }
4061 } else if (j == (uint)_alias) {
4062 newmm->add_req(rep_proj);
4063 } else {
4064 newmm->add_req(u->in(j));
4065 }
4066 }
4067 if ((uint)_alias >= u->req()) {
4068 newmm->set_memory_at(_alias, rep_proj);
4069 }
4070 _phase->register_new_node(newmm, c);
4071 }
4072 return newmm;
4073 }
4074
4075 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
4076 if (phi->adr_type() == TypePtr::BOTTOM) {
4077 Node* region = phi->in(0);
4078 for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
4079 Node* uu = region->fast_out(j);
4080 if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
4081 return false;
4082 }
4083 }
4084 return true;
4085 }
4086 return _phase->C->get_alias_index(phi->adr_type()) == _alias;
4087 }
4088
4089
4090 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
4091 uint last = _phase-> C->unique();
4092 MergeMemNode* mm = NULL;
4093 assert(mem->bottom_type() == Type::MEMORY, "");
4094 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
4095 Node* u = mem->out(i);
4096 if (u != replacement && u->_idx < last) {
4097 if (u->is_ShenandoahBarrier() && _alias != Compile::AliasIdxRaw) {
4098 if (_phase->C->get_alias_index(u->adr_type()) == _alias && ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
4099 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
4100 assert(u->find_edge(mem) == -1, "only one edge");
4101 --i;
4102 }
4103 } else if (u->is_Mem()) {
4104 if (_phase->C->get_alias_index(u->adr_type()) == _alias && ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
4105 assert(_alias == Compile::AliasIdxRaw , "only raw memory can lead to a memory operation");
4106 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
4107 assert(u->find_edge(mem) == -1, "only one edge");
4108 --i;
4109 }
4110 } else if (u->is_MergeMem()) {
4111 MergeMemNode* u_mm = u->as_MergeMem();
4112 if (u_mm->memory_at(_alias) == mem) {
4113 MergeMemNode* newmm = NULL;
4114 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
4115 Node* uu = u->fast_out(j);
4116 assert(!uu->is_MergeMem(), "chain of MergeMems?");
4117 if (uu->is_Phi()) {
4118 if (should_process_phi(uu)) {
4119 Node* region = uu->in(0);
4120 int nb = 0;
4121 for (uint k = 1; k < uu->req(); k++) {
4122 if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
4123 if (newmm == NULL) {
4124 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
4125 }
4126 if (newmm != u) {
4127 _phase->igvn().replace_input_of(uu, k, newmm);
4128 nb++;
4129 --jmax;
4130 }
4131 }
4132 }
4133 if (nb > 0) {
4134 --j;
4135 }
4136 }
4137 } else {
4138 if (rep_ctrl != uu && ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
4139 if (newmm == NULL) {
4140 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
4141 }
4142 if (newmm != u) {
4143 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
4144 --j, --jmax;
4145 }
4146 }
4147 }
4148 }
4149 }
4150 } else if (u->is_Phi()) {
4151 assert(u->bottom_type() == Type::MEMORY, "what else?");
4152 Node* region = u->in(0);
4153 if (should_process_phi(u)) {
4154 bool replaced = false;
4155 for (uint j = 1; j < u->req(); j++) {
4156 if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
4157 Node* nnew = rep_proj;
4158 if (u->adr_type() == TypePtr::BOTTOM) {
4159 if (mm == NULL) {
4160 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
4161 }
4162 nnew = mm;
4163 }
4164 _phase->igvn().replace_input_of(u, j, nnew);
4165 replaced = true;
4166 }
4167 }
4168 if (replaced) {
4169 --i;
4170 }
4171
4172 }
4173 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
4174 u->adr_type() == NULL) {
4175 assert(u->adr_type() != NULL ||
4176 u->Opcode() == Op_Rethrow ||
4177 u->Opcode() == Op_Return ||
4178 u->Opcode() == Op_SafePoint ||
4179 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
4180 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
4181 u->Opcode() == Op_CallLeaf, "");
4182 if (ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
4183 if (mm == NULL) {
4184 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
4185 }
4186 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
4187 --i;
4188 }
4189 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
4190 if (ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
4191 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
4192 --i;
4193 }
4194 }
4195 }
4196 }
4197 }
4198
4199 void MemoryGraphFixer::remove(Node* n) {
4200 assert(n->Opcode() == Op_ShenandoahWBMemProj, "");
4201 Node* c = _phase->get_ctrl(n);
4202 Node* mem = find_mem(c, NULL);
4203 if (mem == n) {
4204 _memory_nodes.map(c->_idx, mem->in(0)->in(ShenandoahBarrierNode::Memory));
4205 }
4206 }
4207
4208 static bool is_on_null_check_path(Block* b, Block* null_check_block) {
4209 if (null_check_block == NULL) {
4210 return false;
4211 }
4212 do {
4213 assert(null_check_block->_num_succs == 1, "only one succ on the path to unc");
4214 if (b == null_check_block) {
4215 return true;
4216 }
4217 null_check_block = null_check_block->_succs[0];
4218 } while(!null_check_block->head()->is_Root());
4219
4220 return false;
4221 }
4222
4223 int PhaseCFG::replace_uses_with_shenandoah_barrier_helper(Node* n, Node* use, Node* val, Block* block, Block* null_check_block) {
4224 int nb = 0;
4225 Block* buse = get_block_for_node(use);
4226 if (is_on_null_check_path(buse, null_check_block)) {
4227 return 0;
4228 }
4229 if (use->is_Phi()) {
4230 for (uint j = 1; j < use->req(); j++) {
4231 if (use->in(j) == val) {
4232 Block* b = get_block_for_node(use->in(0)->in(j));
4233 if ((block != b && block->dom_lca(b) == block) ||
4234 block == b) {
4235 use->set_req(j, n);
4236 nb++;
4237 }
4238 }
4239 }
4240 } else {
4241 if ((block != buse && block->dom_lca(buse) == block) ||
4242 (block == buse && !use->is_scheduled())) {
4243 // Let precedence edges alone (can confuse anti-dependence verification code)
4244 for (uint i = 0; i < use->req(); i++) {
4245 if (use->in(i) == val) {
4246 use->set_req(i, n);
4247 nb++;
4248 }
4249 }
4250 assert(nb > 0 || use->find_prec_edge(val) != -1, "no replacement?");
4251 }
4252 }
4253
4254 return nb;
4255 }
4256
4257 void PhaseCFG::replace_uses_with_shenandoah_barrier(Node* n, Block* block, Node_List& worklist, GrowableArray<int>& ready_cnt, uint max_idx, uint& phi_cnt) {
4258 // Replace all uses of barrier's input that are dominated by the
4259 // barrier with the value returned by the barrier: no need to keep
4260 // both live.
4261 if (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_ShenandoahReadBarrier) {
4262 MachNullCheckNode* null_check = NULL;
4263 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && null_check == NULL; i++) {
4264 Node* use = n->fast_out(i);
4265 if (use->is_MachNullCheck()) {
4266 null_check = use->as_MachNullCheck();
4267 }
4268 }
4269 Block* null_check_block = NULL;
4270 if (null_check != NULL) {
4271 Node* proj = null_check->find_out_with(Op_IfTrue);
4272 Node* head = proj->unique_out();
4273 null_check_block = get_block_for_node(head);
4274 }
4275
4276 Node* val = n->in(ShenandoahBarrierNode::ValueIn);
4277 if (!val->bottom_type()->isa_narrowoop()) {
4278 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
4279 Node* use = val->fast_out(i);
4280 if (use != n) {
4281 int nb = replace_uses_with_shenandoah_barrier_helper(n, use, val, block, null_check_block);
4282 if (nb > 0) {
4283 --i; imax -= nb;
4284 }
4285 }
4286 }
4287 } else {
4288 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
4289 Node* u = val->fast_out(i);
4290 if (u->is_Mach() && u->as_Mach()->ideal_Opcode() == Op_DecodeN) {
4291 int projs = 0;
4292 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
4293 Node* uu = u->fast_out(j);
4294 assert(!uu->is_MachTemp(), "");
4295 if (uu->is_MachProj() && uu->outcnt() == 0) {
4296 projs++;
4297 } else {
4298 int nb = replace_uses_with_shenandoah_barrier_helper(n, uu, u, block, null_check_block);
4299 if (nb > 0) {
4300 if (!u->is_scheduled()) {
4301 push_ready_nodes(n, uu, block, ready_cnt, worklist, max_idx, nb);
4302 }
4303 --j; jmax -= nb;
4304 }
4305 }
4306 }
4307 // The DecodeN may have gone dead
4308 if (u->outcnt() - projs == 0) {
4309 u->disconnect_inputs(NULL, C);
4310 Block* bu = get_block_for_node(u);
4311 unmap_node_from_block(u);
4312 if (bu == block) {
4313 if (u->is_scheduled()) {
4314 block->find_remove(u);
4315 phi_cnt--;
4316 } else {
4317 worklist.yank(u);
4318 block->remove_node(block->end_idx()-1);
4319 }
4320 } else {
4321 bu->find_remove(u);
4322 }
4323 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
4324 Node* uu = u->fast_out(j);
4325 assert(uu->is_MachProj() && uu->outcnt() == 0, "");
4326 assert(bu == get_block_for_node(uu), "");
4327 uu->disconnect_inputs(NULL, C);
4328 --j; --jmax;
4329 unmap_node_from_block(uu);
4330 if (bu == block) {
4331 if (u->is_scheduled()) {
4332 block->find_remove(uu);
4333 phi_cnt--;
4334 } else {
4335 worklist.yank(uu);
4336 block->remove_node(block->end_idx()-1);
4337 }
4338 } else {
4339 bu->find_remove(uu);
4340 }
4341 assert(uu->is_scheduled() == u->is_scheduled(), "");
4342 }
4343 --i; --imax;
4344 }
4345 }
4346 }
4347 }
4348 }
4349 }