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 ShenandoahBarriersForConst;
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 /**
186 * In Shenandoah, we need barriers on acmp (and similar instructions that compare two
187 * oops) to avoid false negatives. If it compares a from-space and a to-space
188 * copy of an object, a regular acmp would return false, even though both are
189 * the same. The acmp barrier compares the two objects, and when they are
190 * *not equal* it does a read-barrier on both, and compares them again. When it
191 * failed because of different copies of the object, we know that the object
192 * must already have been evacuated (and therefore doesn't require a write-barrier).
193 */
194 void ShenandoahBarrierNode::do_cmpp_if(GraphKit& kit, Node*& taken_branch, Node*& untaken_branch, Node*& taken_memory, Node*& untaken_memory) {
195 assert(taken_memory == NULL && untaken_memory == NULL, "unexpected memory inputs");
196 if (!UseShenandoahGC || !ShenandoahAcmpBarrier || ShenandoahVerifyOptoBarriers) {
197 return;
198 }
199 if (taken_branch->is_top() || untaken_branch->is_top()) {
200 // one of the branches is known to be untaken
201 return;
202 }
203 assert(taken_branch->is_IfProj() && untaken_branch->is_IfProj(), "if projections only");
204 assert(taken_branch->in(0) == untaken_branch->in(0), "should come from same if");
205 IfNode* iff = taken_branch->in(0)->as_If();
206 BoolNode* bol = iff->in(1)->as_Bool();
207 Node* cmp = bol->in(1);
208 if (cmp->Opcode() != Op_CmpP) {
209 return;
210 }
211 Node* a = cmp->in(1);
212 Node* b = cmp->in(2);
213 const Type* a_type = kit.gvn().type(a);
214 const Type* b_type = kit.gvn().type(b);
215 if (a_type->higher_equal(TypePtr::NULL_PTR) || b_type->higher_equal(TypePtr::NULL_PTR)) {
216 // We know one arg is gonna be null. No need for barriers.
217 return;
218 }
219
220 const TypePtr* a_adr_type = ShenandoahBarrierNode::brooks_pointer_type(a_type);
221 const TypePtr* b_adr_type = ShenandoahBarrierNode::brooks_pointer_type(b_type);
222 if ((! ShenandoahBarrierNode::needs_barrier(&kit.gvn(), NULL, a, kit.memory(a_adr_type), false)) &&
223 (! ShenandoahBarrierNode::needs_barrier(&kit.gvn(), NULL, b, kit.memory(b_adr_type), false))) {
224 // We know both args are in to-space already. No acmp barrier needed.
225 return;
226 }
227
228 Node* equal_path = iff->proj_out(true);
229 Node* not_equal_path = iff->proj_out(false);
230
231 if (bol->_test._test == BoolTest::ne) {
232 swap(equal_path, not_equal_path);
233 }
234
235 Node* init_equal_path = equal_path;
236 Node* init_not_equal_path = not_equal_path;
237
238 uint alias_a = kit.C->get_alias_index(a_adr_type);
239 uint alias_b = kit.C->get_alias_index(b_adr_type);
240
241 Node* equal_memory = NULL;
242 Node* not_equal_memory = NULL;
243
244 RegionNode* region = new RegionNode(3);
245 region->init_req(1, equal_path);
246 PhiNode* mem_phi = NULL;
247 if (alias_a == alias_b) {
248 mem_phi = PhiNode::make(region, kit.memory(alias_a), Type::MEMORY, kit.C->get_adr_type(alias_a));
249 } else {
250 Node* mem = kit.reset_memory();
251 mem_phi = PhiNode::make(region, mem, Type::MEMORY, TypePtr::BOTTOM);
252 kit.set_all_memory(mem);
253 }
254
255 kit.set_control(not_equal_path);
256
257 Node* mb = NULL;
258 if (alias_a == alias_b) {
259 Node* mem = kit.reset_memory();
260 mb = MemBarNode::make(kit.C, Op_MemBarAcquire, alias_a);
261 mb->init_req(TypeFunc::Control, kit.control());
262 mb->init_req(TypeFunc::Memory, mem);
263 Node* membar = kit.gvn().transform(mb);
264 kit.set_control(kit.gvn().transform(new ProjNode(membar, TypeFunc::Control)));
265 Node* newmem = kit.gvn().transform(new ProjNode(membar, TypeFunc::Memory));
266 kit.set_all_memory(mem);
267 kit.set_memory(newmem, alias_a);
268 } else {
269 mb = kit.insert_mem_bar(Op_MemBarAcquire);
270 }
271
272 ShenandoahBarrierSetC2* bs = (ShenandoahBarrierSetC2*) BarrierSet::barrier_set()->barrier_set_c2();
273 a = bs->shenandoah_read_barrier_acmp(&kit, a);
274 b = bs->shenandoah_read_barrier_acmp(&kit, b);
275
276 Node* cmp2 = kit.gvn().transform(new CmpPNode(a, b));
277 Node* bol2 = bol->clone();
278 bol2->set_req(1, cmp2);
279 bol2 = kit.gvn().transform(bol2);
280 Node* iff2 = iff->clone();
281 iff2->set_req(0, kit.control());
282 iff2->set_req(1, bol2);
283 kit.gvn().set_type(iff2, kit.gvn().type(iff));
284 Node* equal_path2 = equal_path->clone();
285 equal_path2->set_req(0, iff2);
286 equal_path2 = kit.gvn().transform(equal_path2);
287 Node* not_equal_path2 = not_equal_path->clone();
288 not_equal_path2->set_req(0, iff2);
289 not_equal_path2 = kit.gvn().transform(not_equal_path2);
290
291 region->init_req(2, equal_path2);
292 not_equal_memory = kit.reset_memory();
293 not_equal_path = not_equal_path2;
294
295 kit.set_all_memory(not_equal_memory);
296
297 if (alias_a == alias_b) {
298 mem_phi->init_req(2, kit.memory(alias_a));
299 kit.set_memory(mem_phi, alias_a);
300 } else {
301 mem_phi->init_req(2, kit.reset_memory());
302 }
303
304 kit.record_for_igvn(mem_phi);
305 kit.gvn().set_type(mem_phi, Type::MEMORY);
306
307 if (alias_a == alias_b) {
308 equal_memory = kit.reset_memory();
309 } else {
310 equal_memory = mem_phi;
311 }
312
313 assert(kit.map()->memory() == NULL, "no live memory state");
314 equal_path = kit.gvn().transform(region);
315
316 if (taken_branch == init_equal_path) {
317 assert(untaken_branch == init_not_equal_path, "inconsistent");
318 taken_branch = equal_path;
319 untaken_branch = not_equal_path;
320 taken_memory = equal_memory;
321 untaken_memory = not_equal_memory;
322 } else {
323 assert(taken_branch == init_not_equal_path, "inconsistent");
324 assert(untaken_branch == init_equal_path, "inconsistent");
325 taken_branch = not_equal_path;
326 untaken_branch = equal_path;
327 taken_memory = not_equal_memory;
328 untaken_memory = equal_memory;
329 }
330 }
331
332 bool ShenandoahReadBarrierNode::dominates_memory_rb_impl(PhaseGVN* phase,
333 Node* b1,
334 Node* b2,
335 Node* current,
336 bool linear) {
337 ResourceMark rm;
338 VectorSet visited(Thread::current()->resource_area());
339 Node_Stack phis(0);
340
341 for(int i = 0; i < 10; i++) {
342 if (current == NULL) {
343 return false;
344 } else if (visited.test_set(current->_idx) || current->is_top() || current == b1) {
345 current = NULL;
346 while (phis.is_nonempty() && current == NULL) {
347 uint idx = phis.index();
348 Node* phi = phis.node();
349 if (idx >= phi->req()) {
350 phis.pop();
351 } else {
352 current = phi->in(idx);
353 phis.set_index(idx+1);
354 }
355 }
356 if (current == NULL) {
357 return true;
358 }
359 } else if (current == phase->C->immutable_memory()) {
360 return false;
361 } else if (current->isa_Phi()) {
362 if (!linear) {
363 return false;
364 }
365 phis.push(current, 2);
366 current = current->in(1);
367 } else if (current->Opcode() == Op_ShenandoahWriteBarrier) {
368 const Type* in_type = current->bottom_type();
369 const Type* this_type = b2->bottom_type();
370 if (is_independent(in_type, this_type)) {
371 current = current->in(Memory);
372 } else {
373 return false;
374 }
375 } else if (current->Opcode() == Op_ShenandoahWBMemProj) {
376 current = current->in(0);
377 } else if (current->is_Proj()) {
378 current = current->in(0);
379 } else if (current->is_Call()) {
380 return false; // TODO: Maybe improve by looking at the call's memory effects?
381 } else if (current->is_MemBar()) {
382 return false; // TODO: Do we need to stop at *any* membar?
383 } else if (current->is_MergeMem()) {
384 // if (true) return false;
385 // tty->print_cr("current == mergemem: "); current->dump();
386 const TypePtr* adr_type = brooks_pointer_type(phase->type(b2));
387 uint alias_idx = phase->C->get_alias_index(adr_type);
388 current = current->as_MergeMem()->memory_at(alias_idx);
389 } else {
390 // tty->print_cr("what else can we see here:");
391 #ifdef ASSERT
392 current->dump();
393 #endif
394 ShouldNotReachHere();
395 return false;
396 }
397 }
398 return false;
399 }
400
401 bool ShenandoahReadBarrierNode::is_independent(Node* mem) {
402 if (mem->is_Phi() || mem->is_Proj() || mem->is_MergeMem()) {
403 return true;
404 } else if (mem->Opcode() == Op_ShenandoahWriteBarrier) {
405 const Type* mem_type = mem->bottom_type();
406 const Type* this_type = bottom_type();
407 if (is_independent(mem_type, this_type)) {
408 return true;
409 } else {
410 return false;
411 }
412 } else if (mem->is_Call() || mem->is_MemBar()) {
413 return false;
414 }
415 #ifdef ASSERT
416 mem->dump();
417 #endif
418 ShouldNotReachHere();
419 return true;
420 }
421
422
423 bool ShenandoahReadBarrierNode::dominates_memory_rb(PhaseGVN* phase, Node* b1, Node* b2, bool linear) {
424 return dominates_memory_rb_impl(phase, b1->in(Memory), b2, b2->in(Memory), linear);
425 }
426
427 bool ShenandoahReadBarrierNode::is_independent(const Type* in_type, const Type* this_type) {
428 assert(in_type->isa_oopptr(), "expect oop ptr");
429 assert(this_type->isa_oopptr(), "expect oop ptr");
430 /*
431 if ((! in_type->isa_oopptr()) || (! this_type->isa_oopptr())) {
432 #ifdef ASSERT
433 tty->print_cr("not oopptr");
434 tty->print("in: "); in_type->dump(); tty->print_cr(" ");
435 tty->print("this: "); this_type->dump(); tty->print_cr(" ");
436 #endif
437 return false;
438 }
439 */
440
441 ciKlass* in_kls = in_type->is_oopptr()->klass();
442 ciKlass* this_kls = this_type->is_oopptr()->klass();
443 if (in_kls != NULL && this_kls != NULL &&
444 in_kls->is_loaded() && this_kls->is_loaded() &&
445 (!in_kls->is_subclass_of(this_kls)) &&
446 (!this_kls->is_subclass_of(in_kls))) {
447 #ifdef ASSERT
448 // tty->print_cr("independent: ");
449 // tty->print("in: "); in_kls->print(); tty->print_cr(" ");
450 // tty->print("this: "); this_kls->print(); tty->print_cr(" ");
451 #endif
452 return true;
453 }
454 #ifdef ASSERT
455 // tty->print_cr("possibly dependend?");
456 // tty->print("in: "); in_type->dump(); tty->print_cr(" ");
457 // tty->print("this: "); this_type->dump(); tty->print_cr(" ");
458 #endif
459 return false;
460 }
461
462 Node* ShenandoahReadBarrierNode::Ideal(PhaseGVN *phase, bool can_reshape) {
463
464 if (! can_reshape) {
465 return NULL;
466 }
467
468 if (in(Memory) == phase->C->immutable_memory()) return NULL;
469
470 // If memory input is a MergeMem, take the appropriate slice out of it.
471 Node* mem_in = in(Memory);
472 if (mem_in->isa_MergeMem()) {
473 const TypePtr* adr_type = brooks_pointer_type(bottom_type());
474 uint alias_idx = phase->C->get_alias_index(adr_type);
475 mem_in = mem_in->as_MergeMem()->memory_at(alias_idx);
476 set_req(Memory, mem_in);
477 return this;
478 }
479
480 Node* input = in(Memory);
481 if (input->Opcode() == Op_ShenandoahWBMemProj) {
482 ResourceMark rm;
483 VectorSet seen(Thread::current()->resource_area());
484 Node* n = in(Memory);
485 while (n->Opcode() == Op_ShenandoahWBMemProj &&
486 n->in(0) != NULL &&
487 n->in(0)->Opcode() == Op_ShenandoahWriteBarrier &&
488 n->in(0)->in(Memory) != NULL) {
489 if (seen.test_set(n->_idx)) {
490 return NULL; // loop
491 }
492 n = n->in(0)->in(Memory);
493 }
494
495 Node* wb = input->in(0);
496 const Type* in_type = phase->type(wb);
497 // is_top() test not sufficient here: we can come here after CCP
498 // in a dead branch of the graph that has not yet been removed.
499 if (in_type == Type::TOP) return NULL; // Dead path.
500 assert(wb->Opcode() == Op_ShenandoahWriteBarrier, "expect write barrier");
501 if (is_independent(in_type, _type)) {
502 phase->igvn_rehash_node_delayed(wb);
503 set_req(Memory, wb->in(Memory));
504 if (can_reshape && input->outcnt() == 0) {
505 phase->is_IterGVN()->_worklist.push(input);
506 }
507 return this;
508 }
509 }
510 return NULL;
511 }
512
513 ShenandoahWriteBarrierNode::ShenandoahWriteBarrierNode(Compile* C, Node* ctrl, Node* mem, Node* obj)
514 : ShenandoahBarrierNode(ctrl, mem, obj, false) {
515 assert(UseShenandoahGC && ShenandoahWriteBarrier, "should be enabled");
516 ShenandoahBarrierSetC2::bsc2()->state()->add_shenandoah_barrier(this);
517 }
518
519
520 Node* ShenandoahWriteBarrierNode::Identity(PhaseGVN* phase) {
521 assert(in(0) != NULL, "should have control");
522 PhaseIterGVN* igvn = phase->is_IterGVN();
523 Node* mem_in = in(Memory);
524 Node* mem_proj = NULL;
525
526 if (igvn != NULL) {
527 mem_proj = find_out_with(Op_ShenandoahWBMemProj);
528 if (mem_proj == NULL || mem_in == mem_proj) {
529 return this;
530 }
531 }
532
533 Node* replacement = Identity_impl(phase);
534 if (igvn != NULL) {
535 if (replacement != NULL && replacement != this) {
536 igvn->replace_node(mem_proj, mem_in);
537 }
538 }
539 return replacement;
540 }
541
542
543 Node* ShenandoahWriteBarrierNode::Ideal(PhaseGVN *phase, bool can_reshape) {
544 assert(in(0) != NULL, "should have control");
545 if (!can_reshape) {
546 return NULL;
547 }
548
549 PhaseIterGVN* igvn = phase->is_IterGVN();
550 Node* mem_in = in(Memory);
551
552 if (mem_in->isa_MergeMem()) {
553 const TypePtr* adr_type = brooks_pointer_type(bottom_type());
554 uint alias_idx = phase->C->get_alias_index(adr_type);
555 mem_in = mem_in->as_MergeMem()->memory_at(alias_idx);
556 set_req(Memory, mem_in);
557 return this;
558 }
559
560 Node* val = in(ValueIn);
561 if (val->is_ShenandoahBarrier()) {
562 set_req(ValueIn, val->in(ValueIn));
563 return this;
564 }
565
566 return NULL;
567 }
568
569 bool ShenandoahWriteBarrierNode::expand(Compile* C, PhaseIterGVN& igvn, int& loop_opts_cnt) {
570 if (UseShenandoahGC) {
571 if (ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count() > 0 || (!ShenandoahWriteBarrier && ShenandoahStoreValEnqueueBarrier)) {
572 bool attempt_more_loopopts = ShenandoahLoopOptsAfterExpansion;
573 C->clear_major_progress();
574 PhaseIdealLoop ideal_loop(igvn, LoopOptsShenandoahExpand);
575 if (C->failing()) return false;
576 PhaseIdealLoop::verify(igvn);
577 DEBUG_ONLY(ShenandoahBarrierNode::verify_raw_mem(C->root());)
578 if (attempt_more_loopopts) {
579 C->set_major_progress();
580 if (!C->optimize_loops(loop_opts_cnt, igvn, LoopOptsShenandoahPostExpand)) {
581 return false;
582 }
583 C->clear_major_progress();
584 }
585 }
586 }
587 return true;
588 }
589
590 bool ShenandoahWriteBarrierNode::is_heap_state_test(Node* iff, int mask) {
591 if (!UseShenandoahGC) {
592 return false;
593 }
594 assert(iff->is_If(), "bad input");
595 if (iff->Opcode() != Op_If) {
596 return false;
597 }
598 Node* bol = iff->in(1);
599 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
600 return false;
601 }
602 Node* cmp = bol->in(1);
603 if (cmp->Opcode() != Op_CmpI) {
604 return false;
605 }
606 Node* in1 = cmp->in(1);
607 Node* in2 = cmp->in(2);
608 if (in2->find_int_con(-1) != 0) {
609 return false;
610 }
611 if (in1->Opcode() != Op_AndI) {
612 return false;
613 }
614 in2 = in1->in(2);
615 if (in2->find_int_con(-1) != mask) {
616 return false;
617 }
618 in1 = in1->in(1);
619
620 return is_gc_state_load(in1);
621 }
622
623
624 bool ShenandoahWriteBarrierNode::is_evacuation_in_progress_test(Node* iff) {
625 return is_heap_state_test(iff, ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL);
626 }
627
628 bool ShenandoahWriteBarrierNode::is_heap_stable_test(Node* iff) {
629 if (!UseShenandoahGC) {
630 return false;
631 }
632 assert(iff->is_If(), "bad input");
633 if (iff->Opcode() != Op_If) {
634 return false;
635 }
636 Node* bol = iff->in(1);
637 if (bol == NULL || !bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
638 return false;
639 }
640 Node* cmp = bol->in(1);
641 if (cmp->Opcode() != Op_CmpI) {
642 return false;
643 }
644 Node* in1 = cmp->in(1);
645 Node* in2 = cmp->in(2);
646 if (in2->find_int_con(-1) != 0) {
647 return false;
648 }
649 return is_gc_state_load(in1);
650 }
651
652 bool ShenandoahWriteBarrierNode::is_gc_state_load(Node *n) {
653 if (!UseShenandoahGC) {
654 return false;
655 }
656 if (n->Opcode() != Op_LoadB) {
657 return false;
658 }
659 Node* addp = n->in(MemNode::Address);
660 if (!addp->is_AddP()) {
661 return false;
662 }
663 Node* base = addp->in(AddPNode::Address);
664 Node* off = addp->in(AddPNode::Offset);
665 if (base->Opcode() != Op_ThreadLocal) {
666 return false;
667 }
668 if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) {
669 return false;
670 }
671 return true;
672 }
673
674 bool ShenandoahWriteBarrierNode::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) {
675 assert(phase->is_dominator(stop, start), "bad inputs");
676 ResourceMark rm;
677 Unique_Node_List wq;
678 wq.push(start);
679 for (uint next = 0; next < wq.size(); next++) {
680 Node *m = wq.at(next);
681 if (m == stop) {
682 continue;
683 }
684 if (m->is_SafePoint() && !m->is_CallLeaf()) {
685 return true;
686 }
687 if (m->is_Region()) {
688 for (uint i = 1; i < m->req(); i++) {
689 wq.push(m->in(i));
690 }
691 } else {
692 wq.push(m->in(0));
693 }
694 }
695 return false;
696 }
697
698 bool ShenandoahWriteBarrierNode::try_common_gc_state_load(Node *n, PhaseIdealLoop *phase) {
699 assert(is_gc_state_load(n), "inconsistent");
700 Node* addp = n->in(MemNode::Address);
701 Node* dominator = NULL;
702 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) {
703 Node* u = addp->fast_out(i);
704 assert(is_gc_state_load(u), "inconsistent");
705 if (u != n && phase->is_dominator(u->in(0), n->in(0))) {
706 if (dominator == NULL) {
707 dominator = u;
708 } else {
709 if (phase->dom_depth(u->in(0)) < phase->dom_depth(dominator->in(0))) {
710 dominator = u;
711 }
712 }
713 }
714 }
715 if (dominator == NULL || has_safepoint_between(n->in(0), dominator->in(0), phase)) {
716 return false;
717 }
718 phase->igvn().replace_node(n, dominator);
719
720 return true;
721 }
722
723 Node* ShenandoahWriteBarrierNode::evacuation_in_progress_test_ctrl(Node* iff) {
724 assert(is_evacuation_in_progress_test(iff), "bad input");
725 return iff->in(0);
726 }
727
728 bool ShenandoahBarrierNode::dominates_memory_impl(PhaseGVN* phase,
729 Node* b1,
730 Node* b2,
731 Node* current,
732 bool linear) {
733 ResourceMark rm;
734 VectorSet visited(Thread::current()->resource_area());
735 Node_Stack phis(0);
736
737
738 for(int i = 0; i < 10; i++) {
739 if (current == NULL) {
740 return false;
741 } else if (visited.test_set(current->_idx) || current->is_top() || current == b1) {
742 current = NULL;
743 while (phis.is_nonempty() && current == NULL) {
744 uint idx = phis.index();
745 Node* phi = phis.node();
746 if (idx >= phi->req()) {
747 phis.pop();
748 } else {
749 current = phi->in(idx);
750 phis.set_index(idx+1);
751 }
752 }
753 if (current == NULL) {
754 return true;
755 }
756 } else if (current == b2) {
757 return false;
758 } else if (current == phase->C->immutable_memory()) {
759 return false;
760 } else if (current->isa_Phi()) {
761 if (!linear) {
762 return false;
763 }
764 phis.push(current, 2);
765 current = current->in(1);
766 } else if (current->Opcode() == Op_ShenandoahWriteBarrier) {
767 current = current->in(Memory);
768 } else if (current->Opcode() == Op_ShenandoahWBMemProj) {
769 current = current->in(0);
770 } else if (current->is_Proj()) {
771 current = current->in(0);
772 } else if (current->is_Call()) {
773 current = current->in(TypeFunc::Memory);
774 } else if (current->is_MemBar()) {
775 current = current->in(TypeFunc::Memory);
776 } else if (current->is_MergeMem()) {
777 const TypePtr* adr_type = brooks_pointer_type(phase->type(b2));
778 uint alias_idx = phase->C->get_alias_index(adr_type);
779 current = current->as_MergeMem()->memory_at(alias_idx);
780 } else {
781 #ifdef ASSERT
782 current->dump();
783 #endif
784 ShouldNotReachHere();
785 return false;
786 }
787 }
788 return false;
789 }
790
791 /**
792 * Determines if b1 dominates b2 through memory inputs. It returns true if:
793 * - b1 can be reached by following each branch in b2's memory input (through phis, etc)
794 * - or we get back to b2 (i.e. through a loop) without seeing b1
795 * In all other cases, (in particular, if we reach immutable_memory without having seen b1)
796 * we return false.
797 */
798 bool ShenandoahBarrierNode::dominates_memory(PhaseGVN* phase, Node* b1, Node* b2, bool linear) {
799 return dominates_memory_impl(phase, b1, b2, b2->in(Memory), linear);
800 }
801
802 Node* ShenandoahBarrierNode::Identity_impl(PhaseGVN* phase) {
803 Node* n = in(ValueIn);
804
805 Node* rb_mem = Opcode() == Op_ShenandoahReadBarrier ? in(Memory) : NULL;
806 if (! needs_barrier(phase, this, n, rb_mem, _allow_fromspace)) {
807 return n;
808 }
809
810 // tty->print_cr("find sibling for: "); dump(2);
811 // Try to find a write barrier sibling with identical inputs that we can fold into.
812 for (DUIterator i = n->outs(); n->has_out(i); i++) {
813 Node* sibling = n->out(i);
814 if (sibling == this) {
815 continue;
816 }
817 /*
818 assert(sibling->Opcode() != Op_ShenandoahWriteBarrier ||
819 Opcode() != Op_ShenandoahWriteBarrier || hash() == sibling->hash(),
820 "if this is a write barrier, then sibling can't be write barrier too");
821 */
822 if (sibling->Opcode() != Op_ShenandoahWriteBarrier) {
823 continue;
824 }
825 /*
826 if (sibling->outcnt() == 0) {
827 // Some dead node.
828 continue;
829 }
830 */
831 assert(sibling->in(ValueIn) == in(ValueIn), "sanity");
832 assert(sibling->Opcode() == Op_ShenandoahWriteBarrier, "sanity");
833 // tty->print_cr("candidate: "); sibling->dump();
834
835 if (dominates_memory(phase, sibling, this, phase->is_IterGVN() == NULL)) {
836
837 /*
838 tty->print_cr("matched barrier:");
839 sibling->dump();
840 tty->print_cr("for: ");
841 dump();
842 */
843 return sibling;
844 }
845
846 /*
847 tty->print_cr("couldn't match candidate:");
848 sibling->dump(2);
849 */
850 }
851 /*
852 tty->print_cr("couldn't match barrier to any:");
853 dump();
854 */
855 return this;
856 }
857
858 #ifndef PRODUCT
859 void ShenandoahBarrierNode::dump_spec(outputStream *st) const {
860 const TypePtr* adr = adr_type();
861 if (adr == NULL) {
862 return;
863 }
864 st->print(" @");
865 adr->dump_on(st);
866 st->print(" (");
867 Compile::current()->alias_type(adr)->adr_type()->dump_on(st);
868 st->print(") ");
869 }
870 #endif
871
872 Node* ShenandoahReadBarrierNode::Identity(PhaseGVN* phase) {
873
874 // if (true) return this;
875
876 // tty->print("optimizing rb: "); dump();
877 Node* id = Identity_impl(phase);
878
879 if (id == this && phase->is_IterGVN()) {
880 Node* n = in(ValueIn);
881 // No success in super call. Try to combine identical read barriers.
882 for (DUIterator i = n->outs(); n->has_out(i); i++) {
883 Node* sibling = n->out(i);
884 if (sibling == this || sibling->Opcode() != Op_ShenandoahReadBarrier) {
885 continue;
886 }
887 assert(sibling->in(ValueIn) == in(ValueIn), "sanity");
888 if (phase->is_IterGVN()->hash_find(sibling) &&
889 sibling->bottom_type() == bottom_type() &&
890 sibling->in(Control) == in(Control) &&
891 dominates_memory_rb(phase, sibling, this, phase->is_IterGVN() == NULL)) {
892 /*
893 if (in(Memory) != sibling->in(Memory)) {
894 tty->print_cr("interesting rb-fold");
895 dump();
896 sibling->dump();
897 }
898 */
899 return sibling;
900 }
901 }
902 }
903 return id;
904 }
905
906 const Type* ShenandoahBarrierNode::Value(PhaseGVN* phase) const {
907 // Either input is TOP ==> the result is TOP
908 const Type *t1 = phase->type(in(Memory));
909 if (t1 == Type::TOP) return Type::TOP;
910 const Type *t2 = phase->type(in(ValueIn));
911 if( t2 == Type::TOP ) return Type::TOP;
912
913 if (t2 == TypePtr::NULL_PTR) {
914 return _type;
915 }
916
917 const Type* type = t2->is_oopptr()->cast_to_nonconst();
918 return type;
919 }
920
921 uint ShenandoahBarrierNode::hash() const {
922 return TypeNode::hash() + _allow_fromspace;
923 }
924
925 uint ShenandoahBarrierNode::cmp(const Node& n) const {
926 return _allow_fromspace == ((ShenandoahBarrierNode&) n)._allow_fromspace
927 && TypeNode::cmp(n);
928 }
929
930 uint ShenandoahBarrierNode::size_of() const {
931 return sizeof(*this);
932 }
933
934 Node* ShenandoahWBMemProjNode::Identity(PhaseGVN* phase) {
935
936 Node* wb = in(0);
937 if (wb->is_top()) return phase->C->top(); // Dead path.
938
939 assert(wb->Opcode() == Op_ShenandoahWriteBarrier, "expect write barrier");
940 PhaseIterGVN* igvn = phase->is_IterGVN();
941 // We can't do the below unless the graph is fully constructed.
942 if (igvn == NULL) {
943 return this;
944 }
945
946 // If the mem projection has no barrier users, it's not needed anymore.
947 if (wb->outcnt() == 1) {
948 return wb->in(ShenandoahBarrierNode::Memory);
949 }
950
951 return this;
952 }
953
954 #ifdef ASSERT
955 bool ShenandoahBarrierNode::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) {
956 assert(phis.size() == 0, "");
957
958 while (true) {
959 if (in->bottom_type() == TypePtr::NULL_PTR) {
960 if (trace) {tty->print_cr("NULL");}
961 } else if (!in->bottom_type()->make_ptr()->make_oopptr()) {
962 if (trace) {tty->print_cr("Non oop");}
963 } else if (t == ShenandoahLoad && ShenandoahOptimizeStableFinals &&
964 in->bottom_type()->make_ptr()->isa_aryptr() &&
965 in->bottom_type()->make_ptr()->is_aryptr()->is_stable()) {
966 if (trace) {tty->print_cr("Stable array load");}
967 } else {
968 if (in->is_ConstraintCast()) {
969 in = in->in(1);
970 continue;
971 } else if (in->is_AddP()) {
972 assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access");
973 in = in->in(AddPNode::Address);
974 continue;
975 } else if (in->is_Con() && !ShenandoahBarriersForConst) {
976 if (trace) {tty->print("Found constant"); in->dump();}
977 } else if (in->is_ShenandoahBarrier()) {
978 if (t == ShenandoahOopStore) {
979 if (in->Opcode() != Op_ShenandoahWriteBarrier) {
980 return false;
981 }
982 uint i = 0;
983 for (; i < phis.size(); i++) {
984 Node* n = phis.node_at(i);
985 if (n->Opcode() == Op_ShenandoahEnqueueBarrier) {
986 break;
987 }
988 }
989 if (i == phis.size()) {
990 return false;
991 }
992 } else if (t == ShenandoahStore && in->Opcode() != Op_ShenandoahWriteBarrier) {
993 return false;
994 }
995 barriers_used.push(in);
996 if (trace) {tty->print("Found barrier"); in->dump();}
997 } else if (in->Opcode() == Op_ShenandoahEnqueueBarrier) {
998 if (t != ShenandoahOopStore) {
999 return false;
1000 }
1001 if (trace) {tty->print("Found enqueue barrier"); in->dump();}
1002 phis.push(in, in->req());
1003 in = in->in(1);
1004 continue;
1005 } else if (in->is_Proj() && in->in(0)->is_Allocate()) {
1006 if (trace) {tty->print("Found alloc"); in->in(0)->dump();}
1007 } else if (in->is_Phi()) {
1008 if (!visited.test_set(in->_idx)) {
1009 if (trace) {tty->print("Pushed phi:"); in->dump();}
1010 phis.push(in, 2);
1011 in = in->in(1);
1012 continue;
1013 }
1014 if (trace) {tty->print("Already seen phi:"); in->dump();}
1015 } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) {
1016 if (!visited.test_set(in->_idx)) {
1017 if (trace) {tty->print("Pushed cmovep:"); in->dump();}
1018 phis.push(in, CMoveNode::IfTrue);
1019 in = in->in(CMoveNode::IfFalse);
1020 continue;
1021 }
1022 if (trace) {tty->print("Already seen cmovep:"); in->dump();}
1023 } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) {
1024 in = in->in(1);
1025 continue;
1026 } else {
1027 return false;
1028 }
1029 }
1030 bool cont = false;
1031 while (phis.is_nonempty()) {
1032 uint idx = phis.index();
1033 Node* phi = phis.node();
1034 if (idx >= phi->req()) {
1035 if (trace) {tty->print("Popped phi:"); phi->dump();}
1036 phis.pop();
1037 continue;
1038 }
1039 if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();}
1040 in = phi->in(idx);
1041 phis.set_index(idx+1);
1042 cont = true;
1043 break;
1044 }
1045 if (!cont) {
1046 break;
1047 }
1048 }
1049 return true;
1050 }
1051
1052 void ShenandoahBarrierNode::report_verify_failure(const char *msg, Node *n1, Node *n2) {
1053 if (n1 != NULL) {
1054 n1->dump(+10);
1055 }
1056 if (n2 != NULL) {
1057 n2->dump(+10);
1058 }
1059 fatal("%s", msg);
1060 }
1061
1062 void ShenandoahBarrierNode::verify(RootNode* root) {
1063 ResourceMark rm;
1064 Unique_Node_List wq;
1065 GrowableArray<Node*> barriers;
1066 Unique_Node_List barriers_used;
1067 Node_Stack phis(0);
1068 VectorSet visited(Thread::current()->resource_area());
1069 const bool trace = false;
1070 const bool verify_no_useless_barrier = false;
1071
1072 wq.push(root);
1073 for (uint next = 0; next < wq.size(); next++) {
1074 Node *n = wq.at(next);
1075 if (n->is_Load()) {
1076 const bool trace = false;
1077 if (trace) {tty->print("Verifying"); n->dump();}
1078 if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) {
1079 if (trace) {tty->print_cr("Load range/klass");}
1080 } else {
1081 const TypePtr* adr_type = n->as_Load()->adr_type();
1082
1083 if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) {
1084 if (trace) {tty->print_cr("Mark load");}
1085 } else if (adr_type->isa_instptr() &&
1086 adr_type->is_instptr()->klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
1087 adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset) {
1088 if (trace) {tty->print_cr("Reference.get()");}
1089 } else {
1090 bool verify = true;
1091 if (adr_type->isa_instptr()) {
1092 const TypeInstPtr* tinst = adr_type->is_instptr();
1093 ciKlass* k = tinst->klass();
1094 assert(k->is_instance_klass(), "");
1095 ciInstanceKlass* ik = (ciInstanceKlass*)k;
1096 int offset = adr_type->offset();
1097
1098 if ((ik->debug_final_field_at(offset) && ShenandoahOptimizeInstanceFinals) ||
1099 (ik->debug_stable_field_at(offset) && ShenandoahOptimizeStableFinals)) {
1100 if (trace) {tty->print_cr("Final/stable");}
1101 verify = false;
1102 } else if (k == ciEnv::current()->Class_klass() &&
1103 tinst->const_oop() != NULL &&
1104 tinst->offset() >= (ik->size_helper() * wordSize)) {
1105 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
1106 ciField* field = k->get_field_by_offset(tinst->offset(), true);
1107 if ((ShenandoahOptimizeStaticFinals && field->is_final()) ||
1108 (ShenandoahOptimizeStableFinals && field->is_stable())) {
1109 verify = false;
1110 }
1111 }
1112 }
1113
1114 if (verify && !ShenandoahBarrierNode::verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) {
1115 report_verify_failure("Shenandoah verification: Load should have barriers", n);
1116 }
1117 }
1118 }
1119 } else if (n->is_Store()) {
1120 const bool trace = false;
1121
1122 if (trace) {tty->print("Verifying"); n->dump();}
1123 if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) {
1124 Node* adr = n->in(MemNode::Address);
1125 bool verify = true;
1126
1127 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) {
1128 adr = adr->in(AddPNode::Address);
1129 if (adr->is_AddP()) {
1130 assert(adr->in(AddPNode::Base)->is_top(), "");
1131 adr = adr->in(AddPNode::Address);
1132 if (adr->Opcode() == Op_LoadP &&
1133 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() &&
1134 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
1135 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) {
1136 if (trace) {tty->print_cr("SATB prebarrier");}
1137 verify = false;
1138 }
1139 }
1140 }
1141
1142 if (verify && !ShenandoahBarrierNode::verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahStoreValEnqueueBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
1143 report_verify_failure("Shenandoah verification: Store should have barriers", n);
1144 }
1145 }
1146 if (!ShenandoahBarrierNode::verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
1147 report_verify_failure("Shenandoah verification: Store (address) should have barriers", n);
1148 }
1149 } else if (n->Opcode() == Op_CmpP) {
1150 const bool trace = false;
1151
1152 Node* in1 = n->in(1);
1153 Node* in2 = n->in(2);
1154 if (in1->bottom_type()->isa_oopptr()) {
1155 if (trace) {tty->print("Verifying"); n->dump();}
1156
1157 bool mark_inputs = false;
1158 if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR ||
1159 ((in1->is_Con() || in2->is_Con()) && !ShenandoahBarriersForConst)) {
1160 if (trace) {tty->print_cr("Comparison against a constant");}
1161 mark_inputs = true;
1162 } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) ||
1163 (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) {
1164 if (trace) {tty->print_cr("Comparison with newly alloc'ed object");}
1165 mark_inputs = true;
1166 } else {
1167 assert(in2->bottom_type()->isa_oopptr(), "");
1168
1169 if (!ShenandoahBarrierNode::verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) ||
1170 !ShenandoahBarrierNode::verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) {
1171 report_verify_failure("Shenandoah verification: Cmp should have barriers", n);
1172 }
1173 }
1174 if (verify_no_useless_barrier &&
1175 mark_inputs &&
1176 (!ShenandoahBarrierNode::verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) ||
1177 !ShenandoahBarrierNode::verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) {
1178 phis.clear();
1179 visited.Reset();
1180 }
1181 }
1182 } else if (n->is_LoadStore()) {
1183 if (n->in(MemNode::ValueIn)->bottom_type()->isa_ptr() &&
1184 !ShenandoahBarrierNode::verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahLoad, trace, barriers_used)) {
1185 report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n);
1186 }
1187
1188 if (n->in(MemNode::Address)->bottom_type()->isa_oopptr() && !ShenandoahBarrierNode::verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
1189 report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n);
1190 }
1191 } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) {
1192 CallNode* call = n->as_Call();
1193
1194 static struct {
1195 const char* name;
1196 struct {
1197 int pos;
1198 verify_type t;
1199 } args[6];
1200 } calls[] = {
1201 "aescrypt_encryptBlock",
1202 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
1203 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1204 "aescrypt_decryptBlock",
1205 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
1206 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1207 "multiplyToLen",
1208 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { TypeFunc::Parms+4, ShenandoahStore },
1209 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1210 "squareToLen",
1211 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { -1, ShenandoahNone},
1212 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1213 "montgomery_multiply",
1214 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
1215 { TypeFunc::Parms+6, ShenandoahStore }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1216 "montgomery_square",
1217 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+5, ShenandoahStore },
1218 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1219 "mulAdd",
1220 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
1221 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1222 "vectorizedMismatch",
1223 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
1224 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1225 "updateBytesCRC32",
1226 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
1227 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1228 "updateBytesAdler32",
1229 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
1230 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1231 "updateBytesCRC32C",
1232 { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad}, { -1, ShenandoahNone},
1233 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1234 "counterMode_AESCrypt",
1235 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
1236 { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } },
1237 "cipherBlockChaining_encryptAESCrypt",
1238 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
1239 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1240 "cipherBlockChaining_decryptAESCrypt",
1241 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
1242 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1243 "shenandoah_clone_barrier",
1244 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
1245 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1246 "ghash_processBlocks",
1247 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
1248 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1249 "sha1_implCompress",
1250 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1251 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1252 "sha256_implCompress",
1253 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1254 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1255 "sha512_implCompress",
1256 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1257 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1258 "sha1_implCompressMB",
1259 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1260 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1261 "sha256_implCompressMB",
1262 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1263 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1264 "sha512_implCompressMB",
1265 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
1266 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
1267 };
1268
1269 if (call->is_call_to_arraycopystub()) {
1270 Node* dest = NULL;
1271 const TypeTuple* args = n->as_Call()->_tf->domain();
1272 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
1273 if (args->field_at(i)->isa_ptr()) {
1274 j++;
1275 if (j == 2) {
1276 dest = n->in(i);
1277 break;
1278 }
1279 }
1280 }
1281 if (!ShenandoahBarrierNode::verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) ||
1282 !ShenandoahBarrierNode::verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) {
1283 report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n);
1284 }
1285 } else if (strlen(call->_name) > 5 &&
1286 !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) {
1287 if (!ShenandoahBarrierNode::verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) {
1288 report_verify_failure("Shenandoah verification: _fill should have barriers", n);
1289 }
1290 } else if (!strcmp(call->_name, "shenandoah_wb_pre")) {
1291 // skip
1292 } else {
1293 const int calls_len = sizeof(calls) / sizeof(calls[0]);
1294 int i = 0;
1295 for (; i < calls_len; i++) {
1296 if (!strcmp(calls[i].name, call->_name)) {
1297 break;
1298 }
1299 }
1300 if (i != calls_len) {
1301 const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]);
1302 for (uint j = 0; j < args_len; j++) {
1303 int pos = calls[i].args[j].pos;
1304 if (pos == -1) {
1305 break;
1306 }
1307 if (!ShenandoahBarrierNode::verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) {
1308 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
1309 }
1310 }
1311 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
1312 if (call->in(j)->bottom_type()->make_ptr() &&
1313 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
1314 uint k = 0;
1315 for (; k < args_len && calls[i].args[k].pos != (int)j; k++);
1316 if (k == args_len) {
1317 fatal("arg %d for call %s not covered", j, call->_name);
1318 }
1319 }
1320 }
1321 } else {
1322 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
1323 if (call->in(j)->bottom_type()->make_ptr() &&
1324 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
1325 fatal("%s not covered", call->_name);
1326 }
1327 }
1328 }
1329 }
1330 } else if (n->is_ShenandoahBarrier()) {
1331 assert(!barriers.contains(n), "");
1332 assert(n->Opcode() != Op_ShenandoahWriteBarrier || n->find_out_with(Op_ShenandoahWBMemProj) != NULL, "bad shenandoah write barrier");
1333 assert(n->Opcode() != Op_ShenandoahWriteBarrier || n->outcnt() > 1, "bad shenandoah write barrier");
1334 barriers.push(n);
1335 } else if (n->Opcode() == Op_ShenandoahEnqueueBarrier) {
1336 // skip
1337 } else if (n->is_AddP()
1338 || n->is_Phi()
1339 || n->is_ConstraintCast()
1340 || n->Opcode() == Op_Return
1341 || n->Opcode() == Op_CMoveP
1342 || n->Opcode() == Op_CMoveN
1343 || n->Opcode() == Op_Rethrow
1344 || n->is_MemBar()
1345 || n->Opcode() == Op_Conv2B
1346 || n->Opcode() == Op_SafePoint
1347 || n->is_CallJava()
1348 || n->Opcode() == Op_Unlock
1349 || n->Opcode() == Op_EncodeP
1350 || n->Opcode() == Op_DecodeN
1351 || (n->Opcode() == Op_CastP2X && UseShenandoahMatrix)) {
1352 // nothing to do
1353 } else {
1354 static struct {
1355 int opcode;
1356 struct {
1357 int pos;
1358 verify_type t;
1359 } inputs[2];
1360 } others[] = {
1361 Op_FastLock,
1362 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
1363 Op_Lock,
1364 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone} },
1365 Op_ArrayCopy,
1366 { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } },
1367 Op_StrCompressedCopy,
1368 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
1369 Op_StrInflatedCopy,
1370 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
1371 Op_AryEq,
1372 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
1373 Op_StrIndexOf,
1374 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
1375 Op_StrComp,
1376 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
1377 Op_StrEquals,
1378 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
1379 Op_EncodeISOArray,
1380 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
1381 Op_HasNegatives,
1382 { { 2, ShenandoahLoad }, { -1, ShenandoahNone} },
1383 Op_CastP2X,
1384 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
1385 Op_StrIndexOfChar,
1386 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } },
1387 };
1388
1389 const int others_len = sizeof(others) / sizeof(others[0]);
1390 int i = 0;
1391 for (; i < others_len; i++) {
1392 if (others[i].opcode == n->Opcode()) {
1393 break;
1394 }
1395 }
1396 uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req();
1397 if (i != others_len) {
1398 const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]);
1399 for (uint j = 0; j < inputs_len; j++) {
1400 int pos = others[i].inputs[j].pos;
1401 if (pos == -1) {
1402 break;
1403 }
1404 if (!ShenandoahBarrierNode::verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) {
1405 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
1406 }
1407 }
1408 for (uint j = 1; j < stop; j++) {
1409 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
1410 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
1411 uint k = 0;
1412 for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++);
1413 if (k == inputs_len) {
1414 fatal("arg %d for node %s not covered", j, n->Name());
1415 }
1416 }
1417 }
1418 } else {
1419 for (uint j = 1; j < stop; j++) {
1420 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
1421 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
1422 fatal("%s not covered", n->Name());
1423 }
1424 }
1425 }
1426 }
1427
1428 if (n->is_SafePoint()) {
1429 SafePointNode* sfpt = n->as_SafePoint();
1430 if (verify_no_useless_barrier && sfpt->jvms() != NULL) {
1431 for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) {
1432 if (!ShenandoahBarrierNode::verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) {
1433 phis.clear();
1434 visited.Reset();
1435 }
1436 }
1437 }
1438 }
1439 for( uint i = 0; i < n->len(); ++i ) {
1440 Node *m = n->in(i);
1441 if (m == NULL) continue;
1442
1443 // In most cases, inputs should be known to be non null. If it's
1444 // not the case, it could be a missing cast_not_null() in an
1445 // intrinsic or support might be needed in AddPNode::Ideal() to
1446 // avoid a NULL+offset input.
1447 if (!(n->is_Phi() ||
1448 (n->is_SafePoint() && (!n->is_CallRuntime() || !strcmp(n->as_Call()->_name, "shenandoah_wb_pre") || !strcmp(n->as_Call()->_name, "unsafe_arraycopy"))) ||
1449 n->Opcode() == Op_CmpP ||
1450 n->Opcode() == Op_CmpN ||
1451 (n->Opcode() == Op_StoreP && i == StoreNode::ValueIn) ||
1452 (n->Opcode() == Op_StoreN && i == StoreNode::ValueIn) ||
1453 n->is_ConstraintCast() ||
1454 n->Opcode() == Op_Return ||
1455 n->Opcode() == Op_Conv2B ||
1456 n->is_AddP() ||
1457 n->Opcode() == Op_CMoveP ||
1458 n->Opcode() == Op_CMoveN ||
1459 n->Opcode() == Op_Rethrow ||
1460 n->is_MemBar() ||
1461 n->is_Mem() ||
1462 n->Opcode() == Op_AryEq ||
1463 n->Opcode() == Op_SCMemProj ||
1464 n->Opcode() == Op_EncodeP ||
1465 n->Opcode() == Op_DecodeN ||
1466 n->Opcode() == Op_ShenandoahWriteBarrier ||
1467 n->Opcode() == Op_ShenandoahWBMemProj ||
1468 n->Opcode() == Op_ShenandoahEnqueueBarrier)) {
1469 if (m->bottom_type()->make_oopptr() && m->bottom_type()->make_oopptr()->meet(TypePtr::NULL_PTR) == m->bottom_type()) {
1470 report_verify_failure("Shenandoah verification: null input", n, m);
1471 }
1472 }
1473
1474 wq.push(m);
1475 }
1476 }
1477
1478 if (verify_no_useless_barrier) {
1479 for (int i = 0; i < barriers.length(); i++) {
1480 Node* n = barriers.at(i);
1481 if (!barriers_used.member(n)) {
1482 tty->print("XXX useless barrier"); n->dump(-2);
1483 ShouldNotReachHere();
1484 }
1485 }
1486 }
1487 }
1488 #endif
1489
1490 bool ShenandoahBarrierNode::is_dominator_same_ctrl(Node*c, Node* d, Node* n, PhaseIdealLoop* phase) {
1491 // That both nodes have the same control is not sufficient to prove
1492 // domination, verify that there's no path from d to n
1493 ResourceMark rm;
1494 Unique_Node_List wq;
1495 wq.push(d);
1496 for (uint next = 0; next < wq.size(); next++) {
1497 Node *m = wq.at(next);
1498 if (m == n) {
1499 return false;
1500 }
1501 if (m->is_Phi() && m->in(0)->is_Loop()) {
1502 assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control");
1503 } else {
1504 for (uint i = 0; i < m->req(); i++) {
1505 if (m->in(i) != NULL && phase->ctrl_or_self(m->in(i)) == c) {
1506 wq.push(m->in(i));
1507 }
1508 }
1509 }
1510 }
1511 return true;
1512 }
1513
1514 bool ShenandoahBarrierNode::is_dominator(Node *d_c, Node *n_c, Node* d, Node* n, PhaseIdealLoop* phase) {
1515 if (d_c != n_c) {
1516 return phase->is_dominator(d_c, n_c);
1517 }
1518 return is_dominator_same_ctrl(d_c, d, n, phase);
1519 }
1520
1521 Node* next_mem(Node* mem, int alias) {
1522 Node* res = NULL;
1523 if (mem->is_Proj()) {
1524 res = mem->in(0);
1525 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
1526 res = mem->in(TypeFunc::Memory);
1527 } else if (mem->is_Phi()) {
1528 res = mem->in(1);
1529 } else if (mem->is_ShenandoahBarrier()) {
1530 res = mem->in(ShenandoahBarrierNode::Memory);
1531 } else if (mem->is_MergeMem()) {
1532 res = mem->as_MergeMem()->memory_at(alias);
1533 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
1534 assert(alias = Compile::AliasIdxRaw, "following raw memory can't lead to a barrier");
1535 res = mem->in(MemNode::Memory);
1536 } else {
1537 #ifdef ASSERT
1538 mem->dump();
1539 #endif
1540 ShouldNotReachHere();
1541 }
1542 return res;
1543 }
1544
1545 Node* ShenandoahBarrierNode::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) {
1546 Node* iffproj = NULL;
1547 while (c != dom) {
1548 Node* next = phase->idom(c);
1549 assert(next->unique_ctrl_out() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?");
1550 if (c->is_Region()) {
1551 ResourceMark rm;
1552 Unique_Node_List wq;
1553 wq.push(c);
1554 for (uint i = 0; i < wq.size(); i++) {
1555 Node *n = wq.at(i);
1556 if (n == next) {
1557 continue;
1558 }
1559 if (n->is_Region()) {
1560 for (uint j = 1; j < n->req(); j++) {
1561 wq.push(n->in(j));
1562 }
1563 } else {
1564 wq.push(n->in(0));
1565 }
1566 }
1567 for (uint i = 0; i < wq.size(); i++) {
1568 Node *n = wq.at(i);
1569 assert(n->is_CFG(), "");
1570 if (n->is_Multi()) {
1571 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1572 Node* u = n->fast_out(j);
1573 if (u->is_CFG()) {
1574 if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
1575 return NodeSentinel;
1576 }
1577 }
1578 }
1579 }
1580 }
1581 } else if (c->is_Proj()) {
1582 if (c->is_IfProj()) {
1583 if (c->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) != NULL) {
1584 // continue;
1585 } else {
1586 if (!allow_one_proj) {
1587 return NodeSentinel;
1588 }
1589 if (iffproj == NULL) {
1590 iffproj = c;
1591 } else {
1592 return NodeSentinel;
1593 }
1594 }
1595 } else if (c->Opcode() == Op_JumpProj) {
1596 return NodeSentinel; // unsupported
1597 } else if (c->Opcode() == Op_CatchProj) {
1598 return NodeSentinel; // unsupported
1599 } else if (c->Opcode() == Op_CProj && next->Opcode() == Op_NeverBranch) {
1600 return NodeSentinel; // unsupported
1601 } else {
1602 assert(next->unique_ctrl_out() == c, "unsupported branch pattern");
1603 }
1604 }
1605 c = next;
1606 }
1607 return iffproj;
1608 }
1609
1610 #ifdef ASSERT
1611 void ShenandoahWriteBarrierNode::memory_dominates_all_paths_helper(Node* c, Node* rep_ctrl, Unique_Node_List& controls, PhaseIdealLoop* phase) {
1612 const bool trace = false;
1613 if (trace) { tty->print("X control is"); c->dump(); }
1614
1615 uint start = controls.size();
1616 controls.push(c);
1617 for (uint i = start; i < controls.size(); i++) {
1618 Node *n = controls.at(i);
1619
1620 if (trace) { tty->print("X from"); n->dump(); }
1621
1622 if (n == rep_ctrl) {
1623 continue;
1624 }
1625
1626 if (n->is_Proj()) {
1627 Node* n_dom = n->in(0);
1628 IdealLoopTree* n_dom_loop = phase->get_loop(n_dom);
1629 if (n->is_IfProj() && n_dom->outcnt() == 2) {
1630 n_dom_loop = phase->get_loop(n_dom->as_If()->proj_out(n->as_Proj()->_con == 0 ? 1 : 0));
1631 }
1632 if (n_dom_loop != phase->ltree_root()) {
1633 Node* tail = n_dom_loop->tail();
1634 if (tail->is_Region()) {
1635 for (uint j = 1; j < tail->req(); j++) {
1636 if (phase->is_dominator(n_dom, tail->in(j)) && !phase->is_dominator(n, tail->in(j))) {
1637 assert(phase->is_dominator(rep_ctrl, tail->in(j)), "why are we here?");
1638 // entering loop from below, mark backedge
1639 if (trace) { tty->print("X pushing backedge"); tail->in(j)->dump(); }
1640 controls.push(tail->in(j));
1641 //assert(n->in(0) == n_dom, "strange flow control");
1642 }
1643 }
1644 } else if (phase->get_loop(n) != n_dom_loop && phase->is_dominator(n_dom, tail)) {
1645 // entering loop from below, mark backedge
1646 if (trace) { tty->print("X pushing backedge"); tail->dump(); }
1647 controls.push(tail);
1648 //assert(n->in(0) == n_dom, "strange flow control");
1649 }
1650 }
1651 }
1652
1653 if (n->is_Loop()) {
1654 Node* c = n->in(LoopNode::EntryControl);
1655 if (trace) { tty->print("X pushing"); c->dump(); }
1656 controls.push(c);
1657 } else if (n->is_Region()) {
1658 for (uint i = 1; i < n->req(); i++) {
1659 Node* c = n->in(i);
1660 if (trace) { tty->print("X pushing"); c->dump(); }
1661 controls.push(c);
1662 }
1663 } else {
1664 Node* c = n->in(0);
1665 if (trace) { tty->print("X pushing"); c->dump(); }
1666 controls.push(c);
1667 }
1668 }
1669 }
1670
1671 bool ShenandoahWriteBarrierNode::memory_dominates_all_paths(Node* mem, Node* rep_ctrl, int alias, PhaseIdealLoop* phase) {
1672 const bool trace = false;
1673 if (trace) {
1674 tty->print("XXX mem is"); mem->dump();
1675 tty->print("XXX rep ctrl is"); rep_ctrl->dump();
1676 tty->print_cr("XXX alias is %d", alias);
1677 }
1678 ResourceMark rm;
1679 Unique_Node_List wq;
1680 Unique_Node_List controls;
1681 wq.push(mem);
1682 for (uint next = 0; next < wq.size(); next++) {
1683 Node *nn = wq.at(next);
1684 if (trace) { tty->print("XX from mem"); nn->dump(); }
1685 assert(nn->bottom_type() == Type::MEMORY, "memory only");
1686
1687 if (nn->is_Phi()) {
1688 Node* r = nn->in(0);
1689 for (DUIterator_Fast jmax, j = r->fast_outs(jmax); j < jmax; j++) {
1690 Node* u = r->fast_out(j);
1691 if (u->is_Phi() && u->bottom_type() == Type::MEMORY && u != nn &&
1692 (u->adr_type() == TypePtr::BOTTOM || phase->C->get_alias_index(u->adr_type()) == alias)) {
1693 if (trace) { tty->print("XX Next mem (other phi)"); u->dump(); }
1694 wq.push(u);
1695 }
1696 }
1697 }
1698
1699 for (DUIterator_Fast imax, i = nn->fast_outs(imax); i < imax; i++) {
1700 Node* use = nn->fast_out(i);
1701
1702 if (trace) { tty->print("XX use %p", use->adr_type()); use->dump(); }
1703 if (use->is_CFG() && use->in(TypeFunc::Memory) == nn) {
1704 Node* c = use->in(0);
1705 if (phase->is_dominator(rep_ctrl, c)) {
1706 memory_dominates_all_paths_helper(c, rep_ctrl, controls, phase);
1707 } else if (use->is_CallStaticJava() && use->as_CallStaticJava()->uncommon_trap_request() != 0 && c->is_Region()) {
1708 Node* region = c;
1709 if (trace) { tty->print("XX unc region"); region->dump(); }
1710 for (uint j = 1; j < region->req(); j++) {
1711 if (phase->is_dominator(rep_ctrl, region->in(j))) {
1712 if (trace) { tty->print("XX unc follows"); region->in(j)->dump(); }
1713 memory_dominates_all_paths_helper(region->in(j), rep_ctrl, controls, phase);
1714 }
1715 }
1716 }
1717 //continue;
1718 } else if (use->is_Phi()) {
1719 assert(use->bottom_type() == Type::MEMORY, "bad phi");
1720 if ((use->adr_type() == TypePtr::BOTTOM /*&& !shenandoah_has_alias_phi(C, use, alias)*/) ||
1721 phase->C->get_alias_index(use->adr_type()) == alias) {
1722 for (uint j = 1; j < use->req(); j++) {
1723 if (use->in(j) == nn) {
1724 Node* c = use->in(0)->in(j);
1725 if (phase->is_dominator(rep_ctrl, c)) {
1726 memory_dominates_all_paths_helper(c, rep_ctrl, controls, phase);
1727 }
1728 }
1729 }
1730 }
1731 // continue;
1732 }
1733
1734 if (use->is_MergeMem()) {
1735 if (use->as_MergeMem()->memory_at(alias) == nn) {
1736 if (trace) { tty->print("XX Next mem"); use->dump(); }
1737 // follow the memory edges
1738 wq.push(use);
1739 }
1740 } else if (use->is_Phi()) {
1741 assert(use->bottom_type() == Type::MEMORY, "bad phi");
1742 if ((use->adr_type() == TypePtr::BOTTOM /*&& !shenandoah_has_alias_phi(C, use, alias)*/) ||
1743 phase->C->get_alias_index(use->adr_type()) == alias) {
1744 if (trace) { tty->print("XX Next mem"); use->dump(); }
1745 // follow the memory edges
1746 wq.push(use);
1747 }
1748 } else if (use->bottom_type() == Type::MEMORY &&
1749 (use->adr_type() == TypePtr::BOTTOM || phase->C->get_alias_index(use->adr_type()) == alias)) {
1750 if (trace) { tty->print("XX Next mem"); use->dump(); }
1751 // follow the memory edges
1752 wq.push(use);
1753 } else if ((use->is_SafePoint() || use->is_MemBar()) &&
1754 (use->adr_type() == TypePtr::BOTTOM || phase->C->get_alias_index(use->adr_type()) == alias)) {
1755 for (DUIterator_Fast jmax, j = use->fast_outs(jmax); j < jmax; j++) {
1756 Node* u = use->fast_out(j);
1757 if (u->bottom_type() == Type::MEMORY) {
1758 if (trace) { tty->print("XX Next mem"); u->dump(); }
1759 // follow the memory edges
1760 wq.push(u);
1761 }
1762 }
1763 } else if (use->Opcode() == Op_ShenandoahWriteBarrier && phase->C->get_alias_index(use->adr_type()) == alias) {
1764 Node* m = use->find_out_with(Op_ShenandoahWBMemProj);
1765 if (m != NULL) {
1766 if (trace) { tty->print("XX Next mem"); m->dump(); }
1767 // follow the memory edges
1768 wq.push(m);
1769 }
1770 }
1771 }
1772 }
1773
1774 if (controls.size() == 0) {
1775 return false;
1776 }
1777
1778 for (uint i = 0; i < controls.size(); i++) {
1779 Node *n = controls.at(i);
1780
1781 if (trace) { tty->print("X checking"); n->dump(); }
1782
1783 if (n->unique_ctrl_out() != NULL) {
1784 continue;
1785 }
1786
1787 if (n->Opcode() == Op_NeverBranch) {
1788 Node* taken = n->as_Multi()->proj_out(0);
1789 if (!controls.member(taken)) {
1790 if (trace) { tty->print("X not seen"); taken->dump(); }
1791 return false;
1792 }
1793 continue;
1794 }
1795
1796 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1797 Node* u = n->fast_out(j);
1798
1799 if (u->is_CFG()) {
1800 if (!controls.member(u)) {
1801 if (u->is_Proj() && u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
1802 if (trace) { tty->print("X not seen but unc"); u->dump(); }
1803 } else {
1804 Node* c = u;
1805 do {
1806 c = c->unique_ctrl_out();
1807 } while (c != NULL && c->is_Region());
1808 if (c != NULL && c->Opcode() == Op_Halt) {
1809 if (trace) { tty->print("X not seen but halt"); c->dump(); }
1810 } else {
1811 if (trace) { tty->print("X not seen"); u->dump(); }
1812 return false;
1813 }
1814 }
1815 } else {
1816 if (trace) { tty->print("X seen"); u->dump(); }
1817 }
1818 }
1819 }
1820 }
1821 return true;
1822 }
1823 #endif
1824
1825 Node* ShenandoahBarrierNode::dom_mem(Node* mem, Node*& mem_ctrl, Node* n, Node* rep_ctrl, int alias, PhaseIdealLoop* phase) {
1826 ResourceMark rm;
1827 VectorSet wq(Thread::current()->resource_area());
1828 wq.set(mem->_idx);
1829 mem_ctrl = phase->get_ctrl(mem);
1830 while (!is_dominator(mem_ctrl, rep_ctrl, mem, n, phase)) {
1831 mem = next_mem(mem, alias);
1832 if (wq.test_set(mem->_idx)) {
1833 return NULL; // hit an unexpected loop
1834 }
1835 mem_ctrl = phase->ctrl_or_self(mem);
1836 }
1837 if (mem->is_MergeMem()) {
1838 mem = mem->as_MergeMem()->memory_at(alias);
1839 mem_ctrl = phase->ctrl_or_self(mem);
1840 }
1841 return mem;
1842 }
1843
1844 Node* ShenandoahBarrierNode::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) {
1845 ResourceMark rm;
1846 VectorSet wq(Thread::current()->resource_area());
1847 wq.set(mem->_idx);
1848 mem_ctrl = phase->ctrl_or_self(mem);
1849 while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) {
1850 mem = next_mem(mem, alias);
1851 if (wq.test_set(mem->_idx)) {
1852 return NULL;
1853 }
1854 mem_ctrl = phase->ctrl_or_self(mem);
1855 }
1856 if (mem->is_MergeMem()) {
1857 mem = mem->as_MergeMem()->memory_at(alias);
1858 mem_ctrl = phase->ctrl_or_self(mem);
1859 }
1860 return mem;
1861 }
1862
1863 const TypePtr* ShenandoahBarrierNode::fix_addp_type(const TypePtr* res, Node* base) {
1864 if (UseShenandoahGC && ShenandoahBarriersForConst) {
1865 // With barriers on constant oops, if a field being accessed is a
1866 // static field, correct alias analysis requires that we look
1867 // beyond the barriers (that hide the constant) to find the actual
1868 // java class mirror constant.
1869 const TypeInstPtr* ti = res->isa_instptr();
1870 if (ti != NULL &&
1871 ti->const_oop() == NULL &&
1872 ti->klass() == ciEnv::current()->Class_klass() &&
1873 ti->offset() >= (ti->klass()->as_instance_klass()->size_helper() * wordSize)) {
1874 ResourceMark rm;
1875 Unique_Node_List wq;
1876 ciObject* const_oop = NULL;
1877 wq.push(base);
1878 for (uint i = 0; i < wq.size(); i++) {
1879 Node *n = wq.at(i);
1880 if (n->is_ShenandoahBarrier() ||
1881 (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_ShenandoahReadBarrier)) {
1882 Node* m = n->in(ShenandoahBarrierNode::ValueIn);
1883 if (m != NULL) {
1884 wq.push(m);
1885 }
1886 } else if (n->is_Phi()) {
1887 for (uint j = 1; j < n->req(); j++) {
1888 Node* m = n->in(j);
1889 if (m != NULL) {
1890 wq.push(m);
1891 }
1892 }
1893 } else if (n->is_ConstraintCast() || (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_CheckCastPP) ||
1894 n->Opcode() == Op_ShenandoahEnqueueBarrier || n->is_MachSpillCopy()) {
1895 assert(n->Opcode() != Op_ShenandoahEnqueueBarrier, "");
1896 Node* m = n->in(1);
1897 if (m != NULL) {
1898 wq.push(m);
1899 }
1900 } else {
1901 const TypeInstPtr* tn = n->bottom_type()->isa_instptr();
1902 if (tn != NULL) {
1903 if (tn->const_oop() != NULL) {
1904 if (const_oop == NULL) {
1905 const_oop = tn->const_oop();
1906 } else if (const_oop != tn->const_oop()) {
1907 const_oop = NULL;
1908 break;
1909 }
1910 } else {
1911 if (n->is_Proj()) {
1912 if (n->in(0)->Opcode() == Op_CallLeafNoFP) {
1913 if (!ShenandoahBarrierSetAssembler::is_shenandoah_wb_C_call(n->in(0)->as_Call()->entry_point())) {
1914 const_oop = NULL;
1915 break;
1916 }
1917 } else if (n->in(0)->is_MachCallLeaf()) {
1918 if (!ShenandoahBarrierSetAssembler::is_shenandoah_wb_C_call(n->in(0)->as_MachCall()->entry_point())) {
1919 const_oop = NULL;
1920 break;
1921 }
1922 }
1923 } else {
1924 fatal("2 different static fields being accessed with a single AddP");
1925 const_oop = NULL;
1926 break;
1927 }
1928 }
1929 } else {
1930 assert(n->bottom_type() == Type::TOP, "not an instance ptr?");
1931 }
1932 }
1933 }
1934 if (const_oop != NULL) {
1935 res = ti->cast_to_const(const_oop);
1936 }
1937 }
1938 }
1939 return res;
1940 }
1941
1942 static void disconnect_barrier_mem(Node* wb, PhaseIterGVN& igvn) {
1943 Node* mem_in = wb->in(ShenandoahBarrierNode::Memory);
1944 Node* proj = wb->find_out_with(Op_ShenandoahWBMemProj);
1945
1946 for (DUIterator_Last imin, i = proj->last_outs(imin); i >= imin; ) {
1947 Node* u = proj->last_out(i);
1948 igvn.rehash_node_delayed(u);
1949 int nb = u->replace_edge(proj, mem_in);
1950 assert(nb > 0, "no replacement?");
1951 i -= nb;
1952 }
1953 }
1954
1955 Node* ShenandoahWriteBarrierNode::move_above_predicates(LoopNode* cl, Node* val_ctrl, PhaseIdealLoop* phase) {
1956 Node* entry = cl->skip_strip_mined(-1)->in(LoopNode::EntryControl);
1957 Node* above_pred = phase->skip_all_loop_predicates(entry);
1958 Node* ctrl = entry;
1959 while (ctrl != above_pred) {
1960 Node* next = ctrl->in(0);
1961 if (!phase->is_dominator(val_ctrl, next)) {
1962 break;
1963 }
1964 ctrl = next;
1965 }
1966 return ctrl;
1967 }
1968
1969 static MemoryGraphFixer* find_fixer(GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, int alias) {
1970 for (int i = 0; i < memory_graph_fixers.length(); i++) {
1971 if (memory_graph_fixers.at(i)->alias() == alias) {
1972 return memory_graph_fixers.at(i);
1973 }
1974 }
1975 return NULL;
1976 }
1977
1978 static MemoryGraphFixer* create_fixer(GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, int alias, PhaseIdealLoop* phase, bool include_lsm) {
1979 assert(find_fixer(memory_graph_fixers, alias) == NULL, "none should exist yet");
1980 MemoryGraphFixer* fixer = new MemoryGraphFixer(alias, include_lsm, phase);
1981 memory_graph_fixers.push(fixer);
1982 return fixer;
1983 }
1984
1985 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) {
1986 assert(cl->is_Loop(), "bad control");
1987 Node* ctrl = move_above_predicates(cl, val_ctrl, phase);
1988 Node* mem_ctrl = NULL;
1989 int alias = phase->C->get_alias_index(adr_type());
1990
1991 MemoryGraphFixer* fixer = find_fixer(memory_graph_fixers, alias);
1992 if (fixer == NULL) {
1993 fixer = create_fixer(memory_graph_fixers, alias, phase, include_lsm);
1994 }
1995
1996 Node* proj = find_out_with(Op_ShenandoahWBMemProj);
1997
1998 fixer->remove(proj);
1999 Node* mem = fixer->find_mem(ctrl, NULL);
2000
2001 assert(!ShenandoahVerifyOptoBarriers || memory_dominates_all_paths(mem, ctrl, alias, phase), "can't fix the memory graph");
2002
2003 phase->set_ctrl_and_loop(this, ctrl);
2004 phase->igvn().replace_input_of(this, Control, ctrl);
2005
2006 disconnect_barrier_mem(this, phase->igvn());
2007
2008 phase->igvn().replace_input_of(this, Memory, mem);
2009 phase->set_ctrl_and_loop(proj, ctrl);
2010
2011 fixer->fix_mem(ctrl, ctrl, mem, mem, proj, uses);
2012 assert(proj->outcnt() > 0, "disconnected write barrier");
2013 }
2014
2015 LoopNode* ShenandoahWriteBarrierNode::try_move_before_pre_loop(Node* c, Node* val_ctrl, PhaseIdealLoop* phase) {
2016 // A write barrier between a pre and main loop can get in the way of
2017 // vectorization. Move it above the pre loop if possible
2018 CountedLoopNode* cl = NULL;
2019 if (c->is_IfFalse() &&
2020 c->in(0)->is_CountedLoopEnd()) {
2021 cl = c->in(0)->as_CountedLoopEnd()->loopnode();
2022 } else if (c->is_IfProj() &&
2023 c->in(0)->is_If() &&
2024 c->in(0)->in(0)->is_IfFalse() &&
2025 c->in(0)->in(0)->in(0)->is_CountedLoopEnd()) {
2026 cl = c->in(0)->in(0)->in(0)->as_CountedLoopEnd()->loopnode();
2027 }
2028 if (cl != NULL &&
2029 cl->is_pre_loop() &&
2030 val_ctrl != cl &&
2031 phase->is_dominator(val_ctrl, cl)) {
2032 return cl;
2033 }
2034 return NULL;
2035 }
2036
2037 void ShenandoahWriteBarrierNode::try_move_before_loop(GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, PhaseIdealLoop* phase, bool include_lsm, Unique_Node_List& uses) {
2038 Node *n_ctrl = phase->get_ctrl(this);
2039 IdealLoopTree *n_loop = phase->get_loop(n_ctrl);
2040 Node* val = in(ValueIn);
2041 Node* val_ctrl = phase->get_ctrl(val);
2042 if (n_loop != phase->ltree_root() && !n_loop->_irreducible) {
2043 IdealLoopTree *val_loop = phase->get_loop(val_ctrl);
2044 Node* mem = in(Memory);
2045 IdealLoopTree *mem_loop = phase->get_loop(phase->get_ctrl(mem));
2046 if (!n_loop->is_member(val_loop) &&
2047 n_loop->is_member(mem_loop)) {
2048 Node* n_loop_head = n_loop->_head;
2049
2050 if (n_loop_head->is_Loop()) {
2051 LoopNode* loop = n_loop_head->as_Loop();
2052 if (n_loop_head->is_CountedLoop() && n_loop_head->as_CountedLoop()->is_main_loop()) {
2053 LoopNode* res = try_move_before_pre_loop(n_loop_head->in(LoopNode::EntryControl), val_ctrl, phase);
2054 if (res != NULL) {
2055 loop = res;
2056 }
2057 }
2058
2059 try_move_before_loop_helper(loop, val_ctrl, memory_graph_fixers, phase, include_lsm, uses);
2060 }
2061 }
2062 }
2063 LoopNode* ctrl = try_move_before_pre_loop(in(0), val_ctrl, phase);
2064 if (ctrl != NULL) {
2065 try_move_before_loop_helper(ctrl, val_ctrl, memory_graph_fixers, phase, include_lsm, uses);
2066 }
2067 }
2068
2069 Node* ShenandoahWriteBarrierNode::would_subsume(ShenandoahBarrierNode* other, PhaseIdealLoop* phase) {
2070 Node* val = in(ValueIn);
2071 Node* val_ctrl = phase->get_ctrl(val);
2072 Node* other_mem = other->in(Memory);
2073 Node* other_ctrl = phase->get_ctrl(other);
2074 Node* this_ctrl = phase->get_ctrl(this);
2075 IdealLoopTree* this_loop = phase->get_loop(this_ctrl);
2076 IdealLoopTree* other_loop = phase->get_loop(other_ctrl);
2077
2078 Node* ctrl = phase->dom_lca(other_ctrl, this_ctrl);
2079
2080 if (ctrl->is_Proj() &&
2081 ctrl->in(0)->is_Call() &&
2082 ctrl->unique_ctrl_out() != NULL &&
2083 ctrl->unique_ctrl_out()->Opcode() == Op_Catch &&
2084 !phase->is_dominator(val_ctrl, ctrl->in(0)->in(0))) {
2085 return NULL;
2086 }
2087
2088 IdealLoopTree* loop = phase->get_loop(ctrl);
2089
2090 // We don't want to move a write barrier in a loop
2091 // If the LCA is in a inner loop, try a control out of loop if possible
2092 while (!loop->is_member(this_loop) && (other->Opcode() != Op_ShenandoahWriteBarrier || !loop->is_member(other_loop))) {
2093 ctrl = phase->idom(ctrl);
2094 if (ctrl->is_MultiBranch()) {
2095 ctrl = ctrl->in(0);
2096 }
2097 if (ctrl != val_ctrl && phase->is_dominator(ctrl, val_ctrl)) {
2098 return NULL;
2099 }
2100 loop = phase->get_loop(ctrl);
2101 }
2102
2103 if (ShenandoahDontIncreaseWBFreq) {
2104 Node* this_iffproj = no_branches(this_ctrl, ctrl, true, phase);
2105 if (other->Opcode() == Op_ShenandoahWriteBarrier) {
2106 Node* other_iffproj = no_branches(other_ctrl, ctrl, true, phase);
2107 if (other_iffproj == NULL || this_iffproj == NULL) {
2108 return ctrl;
2109 } else if (other_iffproj != NodeSentinel && this_iffproj != NodeSentinel &&
2110 other_iffproj->in(0) == this_iffproj->in(0)) {
2111 return ctrl;
2112 }
2113 } else if (this_iffproj == NULL) {
2114 return ctrl;
2115 }
2116 return NULL;
2117 }
2118
2119 return ctrl;
2120 }
2121
2122 void ShenandoahWriteBarrierNode::optimize_before_expansion(PhaseIdealLoop* phase, GrowableArray<MemoryGraphFixer*> memory_graph_fixers, bool include_lsm) {
2123 bool progress = false;
2124 Unique_Node_List uses;
2125 do {
2126 progress = false;
2127 for (int i = 0; i < ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count(); i++) {
2128 ShenandoahWriteBarrierNode* wb = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barrier(i);
2129
2130 wb->try_move_before_loop(memory_graph_fixers, phase, include_lsm, uses);
2131
2132 Node* val = wb->in(ValueIn);
2133
2134 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
2135 Node* u = val->fast_out(j);
2136 if (u != wb && u->is_ShenandoahBarrier()) {
2137 Node* rep_ctrl = wb->would_subsume(u->as_ShenandoahBarrier(), phase);
2138
2139 if (rep_ctrl != NULL) {
2140 Node* other = u;
2141 Node* val_ctrl = phase->get_ctrl(val);
2142 if (rep_ctrl->is_Proj() &&
2143 rep_ctrl->in(0)->is_Call() &&
2144 rep_ctrl->unique_ctrl_out() != NULL &&
2145 rep_ctrl->unique_ctrl_out()->Opcode() == Op_Catch) {
2146 rep_ctrl = rep_ctrl->in(0)->in(0);
2147
2148 assert(phase->is_dominator(val_ctrl, rep_ctrl), "bad control");
2149 } else {
2150 LoopNode* c = ShenandoahWriteBarrierNode::try_move_before_pre_loop(rep_ctrl, val_ctrl, phase);
2151 if (c != NULL) {
2152 rep_ctrl = ShenandoahWriteBarrierNode::move_above_predicates(c, val_ctrl, phase);
2153 } else {
2154 while (rep_ctrl->is_IfProj()) {
2155 CallStaticJavaNode* unc = rep_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
2156 if (unc != NULL) {
2157 int req = unc->uncommon_trap_request();
2158 Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
2159 if ((trap_reason == Deoptimization::Reason_loop_limit_check ||
2160 trap_reason == Deoptimization::Reason_predicate ||
2161 trap_reason == Deoptimization::Reason_profile_predicate) &&
2162 phase->is_dominator(val_ctrl, rep_ctrl->in(0)->in(0))) {
2163 rep_ctrl = rep_ctrl->in(0)->in(0);
2164 continue;
2165 }
2166 }
2167 break;
2168 }
2169 }
2170 }
2171
2172 Node* wb_ctrl = phase->get_ctrl(wb);
2173 Node* other_ctrl = phase->get_ctrl(other);
2174 int alias = phase->C->get_alias_index(wb->adr_type());
2175 MemoryGraphFixer* fixer = find_fixer(memory_graph_fixers, alias);;
2176 if (!is_dominator(wb_ctrl, other_ctrl, wb, other, phase)) {
2177 if (fixer == NULL) {
2178 fixer = create_fixer(memory_graph_fixers, alias, phase, include_lsm);
2179 }
2180 Node* mem = fixer->find_mem(rep_ctrl, phase->get_ctrl(other) == rep_ctrl ? other : NULL);
2181
2182 if (mem->has_out_with(Op_Lock) || mem->has_out_with(Op_Unlock)) {
2183 continue;
2184 }
2185
2186 Node* wb_proj = wb->find_out_with(Op_ShenandoahWBMemProj);
2187 fixer->remove(wb_proj);
2188 Node* mem_for_ctrl = fixer->find_mem(rep_ctrl, NULL);
2189
2190 if (wb->in(Memory) != mem) {
2191 disconnect_barrier_mem(wb, phase->igvn());
2192 phase->igvn().replace_input_of(wb, Memory, mem);
2193 }
2194 if (rep_ctrl != wb_ctrl) {
2195 phase->set_ctrl_and_loop(wb, rep_ctrl);
2196 phase->igvn().replace_input_of(wb, Control, rep_ctrl);
2197 phase->set_ctrl_and_loop(wb_proj, rep_ctrl);
2198 progress = true;
2199 }
2200
2201 fixer->fix_mem(rep_ctrl, rep_ctrl, mem, mem_for_ctrl, wb_proj, uses);
2202
2203 assert(!ShenandoahVerifyOptoBarriers || ShenandoahWriteBarrierNode::memory_dominates_all_paths(mem, rep_ctrl, alias, phase), "can't fix the memory graph");
2204 }
2205
2206 if (other->Opcode() == Op_ShenandoahWriteBarrier) {
2207 Node* other_proj = other->find_out_with(Op_ShenandoahWBMemProj);
2208 if (fixer != NULL) {
2209 fixer->remove(other_proj);
2210 }
2211 phase->igvn().replace_node(other_proj, other->in(Memory));
2212 }
2213 phase->igvn().replace_node(other, wb);
2214 --j; --jmax;
2215 }
2216 }
2217 }
2218 }
2219 } while(progress);
2220 }
2221
2222 void ShenandoahReadBarrierNode::try_move(Node *n_ctrl, PhaseIdealLoop* phase) {
2223 Node* mem = in(MemNode::Memory);
2224 int alias = phase->C->get_alias_index(adr_type());
2225 const bool trace = false;
2226
2227 #ifdef ASSERT
2228 if (trace) { tty->print("Trying to move mem of"); dump(); }
2229 #endif
2230
2231 Node* new_mem = mem;
2232
2233 ResourceMark rm;
2234 VectorSet seen(Thread::current()->resource_area());
2235 Node_List phis;
2236
2237 for (;;) {
2238 #ifdef ASSERT
2239 if (trace) { tty->print("Looking for dominator from"); mem->dump(); }
2240 #endif
2241 if (mem->is_Proj() && mem->in(0)->is_Start()) {
2242 if (new_mem != in(MemNode::Memory)) {
2243 #ifdef ASSERT
2244 if (trace) { tty->print("XXX Setting mem to"); new_mem->dump(); tty->print(" for "); dump(); }
2245 #endif
2246 phase->igvn().replace_input_of(this, MemNode::Memory, new_mem);
2247 }
2248 return;
2249 }
2250
2251 Node* candidate = mem;
2252 do {
2253 if (!is_independent(mem)) {
2254 if (trace) { tty->print_cr("Not independent"); }
2255 if (new_mem != in(MemNode::Memory)) {
2256 #ifdef ASSERT
2257 if (trace) { tty->print("XXX Setting mem to"); new_mem->dump(); tty->print(" for "); dump(); }
2258 #endif
2259 phase->igvn().replace_input_of(this, MemNode::Memory, new_mem);
2260 }
2261 return;
2262 }
2263 if (seen.test_set(mem->_idx)) {
2264 if (trace) { tty->print_cr("Already seen"); }
2265 ShouldNotReachHere();
2266 // Strange graph
2267 if (new_mem != in(MemNode::Memory)) {
2268 #ifdef ASSERT
2269 if (trace) { tty->print("XXX Setting mem to"); new_mem->dump(); tty->print(" for "); dump(); }
2270 #endif
2271 phase->igvn().replace_input_of(this, MemNode::Memory, new_mem);
2272 }
2273 return;
2274 }
2275 if (mem->is_Phi()) {
2276 phis.push(mem);
2277 }
2278 mem = next_mem(mem, alias);
2279 if (mem->bottom_type() == Type::MEMORY) {
2280 candidate = mem;
2281 }
2282 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");
2283 #ifdef ASSERT
2284 if (trace) { tty->print("Next mem is"); mem->dump(); }
2285 #endif
2286 } while (mem->bottom_type() != Type::MEMORY || !phase->is_dominator(phase->ctrl_or_self(mem), n_ctrl));
2287
2288 assert(mem->bottom_type() == Type::MEMORY, "bad mem");
2289
2290 bool not_dom = false;
2291 for (uint i = 0; i < phis.size() && !not_dom; i++) {
2292 Node* nn = phis.at(i);
2293
2294 #ifdef ASSERT
2295 if (trace) { tty->print("Looking from phi"); nn->dump(); }
2296 #endif
2297 assert(nn->is_Phi(), "phis only");
2298 for (uint j = 2; j < nn->req() && !not_dom; j++) {
2299 Node* m = nn->in(j);
2300 #ifdef ASSERT
2301 if (trace) { tty->print("Input %d is", j); m->dump(); }
2302 #endif
2303 while (m != mem && !seen.test_set(m->_idx)) {
2304 if (is_dominator(phase->ctrl_or_self(m), phase->ctrl_or_self(mem), m, mem, phase)) {
2305 not_dom = true;
2306 // Scheduling anomaly
2307 #ifdef ASSERT
2308 if (trace) { tty->print("Giving up"); m->dump(); }
2309 #endif
2310 break;
2311 }
2312 if (!is_independent(m)) {
2313 if (trace) { tty->print_cr("Not independent"); }
2314 if (new_mem != in(MemNode::Memory)) {
2315 #ifdef ASSERT
2316 if (trace) { tty->print("XXX Setting mem to"); new_mem->dump(); tty->print(" for "); dump(); }
2317 #endif
2318 phase->igvn().replace_input_of(this, MemNode::Memory, new_mem);
2319 }
2320 return;
2321 }
2322 if (m->is_Phi()) {
2323 phis.push(m);
2324 }
2325 m = next_mem(m, alias);
2326 #ifdef ASSERT
2327 if (trace) { tty->print("Next mem is"); m->dump(); }
2328 #endif
2329 }
2330 }
2331 }
2332 if (!not_dom) {
2333 new_mem = mem;
2334 phis.clear();
2335 } else {
2336 seen.Clear();
2337 }
2338 }
2339 }
2340
2341 CallStaticJavaNode* ShenandoahWriteBarrierNode::pin_and_expand_null_check(PhaseIterGVN& igvn) {
2342 Node* val = in(ValueIn);
2343
2344 const Type* val_t = igvn.type(val);
2345
2346 if (val_t->meet(TypePtr::NULL_PTR) != val_t &&
2347 val->Opcode() == Op_CastPP &&
2348 val->in(0) != NULL &&
2349 val->in(0)->Opcode() == Op_IfTrue &&
2350 val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
2351 val->in(0)->in(0)->is_If() &&
2352 val->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
2353 val->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
2354 val->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
2355 val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1) &&
2356 val->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
2357 assert(val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1), "");
2358 CallStaticJavaNode* unc = val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
2359 return unc;
2360 }
2361 return NULL;
2362 }
2363
2364 void ShenandoahWriteBarrierNode::pin_and_expand_move_barrier(PhaseIdealLoop* phase, GrowableArray<MemoryGraphFixer*>& memory_graph_fixers, Unique_Node_List& uses) {
2365 Node* unc = pin_and_expand_null_check(phase->igvn());
2366 Node* val = in(ValueIn);
2367
2368 if (unc != NULL) {
2369 Node* ctrl = phase->get_ctrl(this);
2370 Node* unc_ctrl = val->in(0);
2371
2372 // Don't move write barrier in a loop
2373 IdealLoopTree* loop = phase->get_loop(ctrl);
2374 IdealLoopTree* unc_loop = phase->get_loop(unc_ctrl);
2375
2376 if (!unc_loop->is_member(loop)) {
2377 return;
2378 }
2379
2380 Node* branch = no_branches(ctrl, unc_ctrl, false, phase);
2381 assert(branch == NULL || branch == NodeSentinel, "was not looking for a branch");
2382 if (branch == NodeSentinel) {
2383 return;
2384 }
2385
2386
2387 RegionNode* r = new RegionNode(3);
2388 IfNode* iff = unc_ctrl->in(0)->as_If();
2389
2390 Node* ctrl_use = unc_ctrl->unique_ctrl_out();
2391 Node* unc_ctrl_clone = unc_ctrl->clone();
2392 phase->register_control(unc_ctrl_clone, loop, iff);
2393 Node* c = unc_ctrl_clone;
2394 Node* new_cast = clone_null_check(c, val, unc_ctrl_clone, r, 1, phase);
2395
2396 phase->igvn().replace_input_of(unc_ctrl, 0, c->in(0));
2397 phase->set_idom(unc_ctrl, c->in(0), phase->dom_depth(unc_ctrl));
2398 phase->lazy_replace(c, unc_ctrl);
2399 c = NULL;;
2400 phase->igvn().replace_input_of(val, 0, unc_ctrl_clone);
2401 phase->set_ctrl(val, unc_ctrl_clone);
2402
2403 IfNode* new_iff = new_cast->in(0)->in(0)->as_If();
2404 fix_null_check(iff, unc, unc_ctrl_clone, r, uses, phase);
2405 Node* iff_proj = iff->proj_out(0);
2406 r->init_req(2, iff_proj);
2407
2408 Node* new_bol = new_iff->in(1)->clone();
2409 Node* new_cmp = new_bol->in(1)->clone();
2410 assert(new_cmp->Opcode() == Op_CmpP, "broken");
2411 assert(new_cmp->in(1) == val->in(1), "broken");
2412 new_bol->set_req(1, new_cmp);
2413 new_cmp->set_req(1, this);
2414 phase->register_new_node(new_bol, new_iff->in(0));
2415 phase->register_new_node(new_cmp, new_iff->in(0));
2416 phase->igvn().replace_input_of(new_iff, 1, new_bol);
2417 phase->igvn().replace_input_of(new_cast, 1, this);
2418
2419 for (DUIterator_Fast imax, i = this->fast_outs(imax); i < imax; i++) {
2420 Node* u = this->fast_out(i);
2421 if (u == new_cast || u->Opcode() == Op_ShenandoahWBMemProj || u == new_cmp) {
2422 continue;
2423 }
2424 phase->igvn().rehash_node_delayed(u);
2425 int nb = u->replace_edge(this, new_cast);
2426 assert(nb > 0, "no update?");
2427 --i; imax -= nb;
2428 }
2429
2430 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
2431 Node* u = val->fast_out(i);
2432 if (u == this) {
2433 continue;
2434 }
2435 phase->igvn().rehash_node_delayed(u);
2436 int nb = u->replace_edge(val, new_cast);
2437 assert(nb > 0, "no update?");
2438 --i; imax -= nb;
2439 }
2440
2441 Node* new_ctrl = unc_ctrl_clone;
2442
2443 int alias = phase->C->get_alias_index(adr_type());
2444 MemoryGraphFixer* fixer = find_fixer(memory_graph_fixers, alias);
2445 if (fixer == NULL) {
2446 fixer = create_fixer(memory_graph_fixers, alias, phase, true);
2447 }
2448
2449 Node* proj = find_out_with(Op_ShenandoahWBMemProj);
2450 fixer->remove(proj);
2451 Node* mem = fixer->find_mem(new_ctrl, NULL);
2452
2453 if (in(Memory) != mem) {
2454 disconnect_barrier_mem(this, phase->igvn());
2455 phase->igvn().replace_input_of(this, Memory, mem);
2456 }
2457
2458 phase->set_ctrl_and_loop(this, new_ctrl);
2459 phase->igvn().replace_input_of(this, Control, new_ctrl);
2460 phase->set_ctrl_and_loop(proj, new_ctrl);
2461
2462 fixer->fix_mem(new_ctrl, new_ctrl, mem, mem, proj, uses);
2463 }
2464 }
2465
2466 void ShenandoahWriteBarrierNode::pin_and_expand_helper(PhaseIdealLoop* phase) {
2467 Node* val = in(ValueIn);
2468 CallStaticJavaNode* unc = pin_and_expand_null_check(phase->igvn());
2469 Node* rep = this;
2470 Node* ctrl = phase->get_ctrl(this);
2471 if (unc != NULL && val->in(0) == ctrl) {
2472 Node* unc_ctrl = val->in(0);
2473 IfNode* other_iff = unc_ctrl->unique_ctrl_out()->as_If();
2474 ProjNode* other_unc_ctrl = other_iff->proj_out(1);
2475 Node* cast = NULL;
2476 for (DUIterator_Fast imax, i = other_unc_ctrl->fast_outs(imax); i < imax && cast == NULL; i++) {
2477 Node* u = other_unc_ctrl->fast_out(i);
2478 if (u->Opcode() == Op_CastPP && u->in(1) == this) {
2479 cast = u;
2480 }
2481 }
2482 assert(other_unc_ctrl->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) == unc, "broken");
2483 rep = cast;
2484 }
2485
2486 // Replace all uses of barrier's input that are dominated by ctrl
2487 // with the value returned by the barrier: no need to keep both
2488 // live.
2489 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
2490 Node* u = val->fast_out(i);
2491 if (u != this) {
2492 if (u->is_Phi()) {
2493 int nb = 0;
2494 for (uint j = 1; j < u->req(); j++) {
2495 if (u->in(j) == val) {
2496 Node* c = u->in(0)->in(j);
2497 if (phase->is_dominator(ctrl, c)) {
2498 phase->igvn().replace_input_of(u, j, rep);
2499 nb++;
2500 }
2501 }
2502 }
2503 if (nb > 0) {
2504 imax -= nb;
2505 --i;
2506 }
2507 } else {
2508 Node* c = phase->ctrl_or_self(u);
2509 if (is_dominator(ctrl, c, this, u, phase)) {
2510 phase->igvn().rehash_node_delayed(u);
2511 int nb = u->replace_edge(val, rep);
2512 assert(nb > 0, "no update?");
2513 --i, imax -= nb;
2514 }
2515 }
2516 }
2517 }
2518 }
2519
2520 Node* ShenandoahWriteBarrierNode::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) {
2521 Node* mem = NULL;
2522 Node* c = ctrl;
2523 do {
2524 if (c->is_Region()) {
2525 Node* phi_bottom = NULL;
2526 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == NULL; i++) {
2527 Node* u = c->fast_out(i);
2528 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
2529 if (u->adr_type() == TypePtr::BOTTOM) {
2530 mem = u;
2531 }
2532 }
2533 }
2534 } else {
2535 if (c->is_Call() && c->as_Call()->adr_type() != NULL) {
2536 CallProjections projs;
2537 c->as_Call()->extract_projections(&projs, true, false);
2538 if (projs.fallthrough_memproj != NULL) {
2539 if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
2540 if (projs.catchall_memproj == NULL) {
2541 mem = projs.fallthrough_memproj;
2542 } else {
2543 if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) {
2544 mem = projs.fallthrough_memproj;
2545 } else {
2546 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier");
2547 mem = projs.catchall_memproj;
2548 }
2549 }
2550 }
2551 } else {
2552 Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
2553 if (proj != NULL &&
2554 proj->adr_type() == TypePtr::BOTTOM) {
2555 mem = proj;
2556 }
2557 }
2558 } else {
2559 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
2560 Node* u = c->fast_out(i);
2561 if (u->is_Proj() &&
2562 u->bottom_type() == Type::MEMORY &&
2563 u->adr_type() == TypePtr::BOTTOM) {
2564 assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
2565 assert(mem == NULL, "only one proj");
2566 mem = u;
2567 }
2568 }
2569 assert(!c->is_Call() || c->as_Call()->adr_type() != NULL || mem == NULL, "no mem projection expected");
2570 }
2571 }
2572 c = phase->idom(c);
2573 } while (mem == NULL);
2574 return mem;
2575 }
2576
2577 void ShenandoahWriteBarrierNode::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
2578 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2579 Node* u = n->fast_out(i);
2580 if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
2581 uses.push(u);
2582 }
2583 }
2584 }
2585
2586 void ShenandoahWriteBarrierNode::test_heap_stable(Node* ctrl, Node* raw_mem, Node*& gc_state, Node*& heap_stable,
2587 Node*& heap_not_stable, PhaseIdealLoop* phase) {
2588 IdealLoopTree *loop = phase->get_loop(ctrl);
2589 Node* thread = new ThreadLocalNode();
2590 phase->register_new_node(thread, ctrl);
2591 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
2592 phase->set_ctrl(offset, phase->C->root());
2593 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, offset);
2594 phase->register_new_node(gc_state_addr, ctrl);
2595 uint gc_state_idx = Compile::AliasIdxRaw;
2596 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument
2597 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
2598
2599 gc_state = new LoadBNode(ctrl, raw_mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered);
2600 phase->register_new_node(gc_state, ctrl);
2601
2602 Node* heap_stable_cmp = new CmpINode(gc_state, phase->igvn().zerocon(T_INT));
2603 phase->register_new_node(heap_stable_cmp, ctrl);
2604 Node* heap_stable_test = new BoolNode(heap_stable_cmp, BoolTest::ne);
2605 phase->register_new_node(heap_stable_test, ctrl);
2606 IfNode* heap_stable_iff = new IfNode(ctrl, heap_stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
2607 phase->register_control(heap_stable_iff, loop, ctrl);
2608
2609 heap_stable = new IfFalseNode(heap_stable_iff);
2610 phase->register_control(heap_stable, loop, heap_stable_iff);
2611 heap_not_stable = new IfTrueNode(heap_stable_iff);
2612 phase->register_control(heap_not_stable, loop, heap_stable_iff);
2613
2614 assert(is_heap_stable_test(heap_stable_iff), "Should match the shape");
2615 }
2616
2617
2618 void ShenandoahWriteBarrierNode::test_evacuation_in_progress(Node* ctrl, Node* val, Node*& raw_mem,
2619 Node*& evac_in_progress, Node*& evac_not_in_progress,
2620 Node*& heap_stable, Node*& null_val,
2621 PhaseIdealLoop* phase) {
2622 IdealLoopTree *loop = phase->get_loop(ctrl);
2623 Node* heap_not_stable = NULL;
2624 Node* unused_gc_state = NULL;
2625
2626 test_heap_stable(ctrl, raw_mem, unused_gc_state, heap_stable, heap_not_stable, phase);
2627
2628 ctrl = heap_not_stable;
2629
2630 const Type* val_t = phase->igvn().type(val);
2631
2632 if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
2633 Node* null_cmp = new CmpPNode(val, phase->igvn().zerocon(T_OBJECT));
2634 phase->register_new_node(null_cmp, ctrl);
2635 Node* null_test = new BoolNode(null_cmp, BoolTest::ne);
2636 phase->register_new_node(null_test, ctrl);
2637 IfNode* null_iff = new IfNode(ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
2638 phase->register_control(null_iff, loop, ctrl);
2639 Node* not_null = new IfTrueNode(null_iff);
2640 phase->register_control(not_null, loop, null_iff);
2641 Node* null = new IfFalseNode(null_iff);
2642 phase->register_control(null, loop, null_iff);
2643 null_val = null;
2644 ctrl = not_null;
2645 }
2646
2647 Node* thread = new ThreadLocalNode();
2648 phase->register_new_node(thread, ctrl);
2649 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
2650 phase->set_ctrl(offset, phase->C->root());
2651 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, offset);
2652 phase->register_new_node(gc_state_addr, ctrl);
2653 uint gc_state_idx = Compile::AliasIdxRaw;
2654 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument
2655 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
2656
2657 Node* gc_state = new LoadBNode(ctrl, raw_mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered);
2658 phase->register_new_node(gc_state, ctrl);
2659
2660 Node* evacuation_in_progress = new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL));
2661 phase->register_new_node(evacuation_in_progress, ctrl);
2662 Node* evacuation_in_progress_cmp = new CmpINode(evacuation_in_progress, phase->igvn().zerocon(T_INT));
2663 phase->register_new_node(evacuation_in_progress_cmp, ctrl);
2664 Node* evacuation_in_progress_test = new BoolNode(evacuation_in_progress_cmp, BoolTest::ne);
2665 phase->register_new_node(evacuation_in_progress_test, ctrl);
2666 IfNode* evacuation_iff = new IfNode(ctrl, evacuation_in_progress_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
2667 phase->register_control(evacuation_iff, loop, ctrl);
2668
2669 assert(is_evacuation_in_progress_test(evacuation_iff), "Should match the shape");
2670 assert(is_gc_state_load(gc_state), "Should match the shape");
2671
2672 evac_not_in_progress = new IfFalseNode(evacuation_iff);
2673 phase->register_control(evac_not_in_progress, loop, evacuation_iff);
2674 evac_in_progress = new IfTrueNode(evacuation_iff);
2675 phase->register_control(evac_in_progress, loop, evacuation_iff);
2676 }
2677
2678 Node* ShenandoahWriteBarrierNode::clone_null_check(Node*& c, Node* val, Node* unc_ctrl,
2679 Node* unc_region, uint input, PhaseIdealLoop* phase) {
2680 IdealLoopTree *loop = phase->get_loop(c);
2681 Node* iff = unc_ctrl->in(0);
2682 assert(iff->is_If(), "broken");
2683 Node* new_iff = iff->clone();
2684 new_iff->set_req(0, c);
2685 phase->register_control(new_iff, loop, c);
2686 Node* iffalse = new IfFalseNode(new_iff->as_If());
2687 phase->register_control(iffalse, loop, new_iff);
2688 Node* iftrue = new IfTrueNode(new_iff->as_If());
2689 phase->register_control(iftrue, loop, new_iff);
2690 c = iftrue;
2691 const Type *t = phase->igvn().type(val);
2692 assert(val->Opcode() == Op_CastPP, "expect cast to non null here");
2693 Node* uncasted_val = val->in(1);
2694 val = new CastPPNode(uncasted_val, t);
2695 val->init_req(0, c);
2696 phase->register_new_node(val, c);
2697 unc_region->init_req(input, iffalse);
2698 return val;
2699 }
2700
2701 void ShenandoahWriteBarrierNode::fix_null_check(Node* dom, Node* unc, Node* unc_ctrl, Node* unc_region,
2702 Unique_Node_List& uses, PhaseIdealLoop* phase) {
2703 IfNode* iff = unc_ctrl->in(0)->as_If();
2704 Node* proj = iff->proj_out(0);
2705 assert(proj != unc_ctrl, "bad projection");
2706 Node* use = proj->unique_ctrl_out();
2707
2708 assert(use == unc || use->is_Region(), "what else?");
2709
2710 uses.clear();
2711 if (use == unc) {
2712 phase->set_idom(use, unc_region, phase->dom_depth(use));
2713 for (uint i = 1; i < unc->req(); i++) {
2714 Node* n = unc->in(i);
2715 if (phase->has_ctrl(n) && phase->get_ctrl(n) == proj) {
2716 uses.push(n);
2717 }
2718 }
2719 } else {
2720 assert(use->is_Region(), "what else?");
2721 uint idx = 1;
2722 for (; use->in(idx) != proj; idx++);
2723 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
2724 Node* u = use->fast_out(i);
2725 if (u->is_Phi() && phase->get_ctrl(u->in(idx)) == proj) {
2726 uses.push(u->in(idx));
2727 }
2728 }
2729 }
2730 for(uint next = 0; next < uses.size(); next++ ) {
2731 Node *n = uses.at(next);
2732 assert(phase->get_ctrl(n) == proj, "bad control");
2733 phase->set_ctrl_and_loop(n, unc_region);
2734 if (n->in(0) == proj) {
2735 phase->igvn().replace_input_of(n, 0, unc_region);
2736 }
2737 for (uint i = 0; i < n->req(); i++) {
2738 Node* m = n->in(i);
2739 if (m != NULL && phase->has_ctrl(m) && phase->get_ctrl(m) == proj) {
2740 uses.push(m);
2741 }
2742 }
2743 }
2744
2745 phase->igvn().rehash_node_delayed(use);
2746 int nb = use->replace_edge(proj, unc_region);
2747 assert(nb == 1, "only use expected");
2748 phase->register_control(unc_region, phase->ltree_root(), dom);
2749 }
2750
2751 void ShenandoahWriteBarrierNode::evacuation_not_in_progress_null_check(Node*& c, Node*& val, Node* unc_ctrl, Node*& unc_region, PhaseIdealLoop* phase) {
2752 if (unc_ctrl != NULL) {
2753 // Clone the null check in this branch to allow implicit null check
2754 unc_region = new RegionNode(3);
2755 val = clone_null_check(c, val, unc_ctrl, unc_region, 1, phase);
2756 }
2757 }
2758
2759 void ShenandoahWriteBarrierNode::evacuation_not_in_progress(Node* c, Node* val, Node* unc_ctrl, Node* raw_mem, Node* wb_mem, Node* region,
2760 Node* val_phi, Node* mem_phi, Node* raw_mem_phi, Node*& unc_region, PhaseIdealLoop* phase) {
2761 evacuation_not_in_progress_null_check(c, val, unc_ctrl, unc_region, phase);
2762 region->init_req(1, c);
2763 if (ShenandoahWriteBarrierRB) {
2764 Node* rbfalse = new ShenandoahReadBarrierNode(c, wb_mem, val);
2765 phase->register_new_node(rbfalse, c);
2766 val_phi->init_req(1, rbfalse);
2767 } else {
2768 val_phi->init_req(1, val);
2769 }
2770 mem_phi->init_req(1, wb_mem);
2771 raw_mem_phi->init_req(1, raw_mem);
2772 }
2773
2774 void ShenandoahWriteBarrierNode::heap_stable(Node* c, Node* val, Node* unc_ctrl, Node* raw_mem, Node* wb_mem, Node* region,
2775 Node* val_phi, Node* mem_phi, Node* raw_mem_phi, Node* unc_region, PhaseIdealLoop* phase) {
2776 region->init_req(1, c);
2777 if (unc_ctrl != NULL) {
2778 val = val->in(1);
2779 }
2780 val_phi->init_req(1, val);
2781 mem_phi->init_req(1, wb_mem);
2782 raw_mem_phi->init_req(1, raw_mem);
2783 }
2784
2785 void ShenandoahWriteBarrierNode::evacuation_in_progress_null_check(Node*& c, Node*& val, Node* evacuation_iff, Node* unc, Node* unc_ctrl,
2786 Node* unc_region, Unique_Node_List& uses, PhaseIdealLoop* phase) {
2787 if (unc != NULL) {
2788 // Clone the null check in this branch to allow implicit null check
2789 val = clone_null_check(c, val, unc_ctrl, unc_region, 2, phase);
2790
2791 fix_null_check(evacuation_iff, unc, unc_ctrl, unc_region, uses, phase);
2792
2793 IfNode* iff = unc_ctrl->in(0)->as_If();
2794 phase->igvn().replace_input_of(iff, 1, phase->igvn().intcon(1));
2795 }
2796 }
2797
2798 void ShenandoahWriteBarrierNode::in_cset_fast_test(Node*& c, Node* rbtrue, Node* raw_mem, Node* wb_mem, Node* region, Node* val_phi, Node* mem_phi,
2799 Node* raw_mem_phi, PhaseIdealLoop* phase) {
2800 if (ShenandoahWriteBarrierCsetTestInIR) {
2801 IdealLoopTree *loop = phase->get_loop(c);
2802 Node* raw_rbtrue = new CastP2XNode(c, rbtrue);
2803 phase->register_new_node(raw_rbtrue, c);
2804 Node* cset_offset = new URShiftXNode(raw_rbtrue, phase->igvn().intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
2805 phase->register_new_node(cset_offset, c);
2806 Node* in_cset_fast_test_base_addr = phase->igvn().makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
2807 phase->set_ctrl(in_cset_fast_test_base_addr, phase->C->root());
2808 Node* in_cset_fast_test_adr = new AddPNode(phase->C->top(), in_cset_fast_test_base_addr, cset_offset);
2809 phase->register_new_node(in_cset_fast_test_adr, c);
2810 uint in_cset_fast_test_idx = Compile::AliasIdxRaw;
2811 const TypePtr* in_cset_fast_test_adr_type = NULL; // debug-mode-only argument
2812 debug_only(in_cset_fast_test_adr_type = phase->C->get_adr_type(in_cset_fast_test_idx));
2813 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);
2814 phase->register_new_node(in_cset_fast_test_load, c);
2815 Node* in_cset_fast_test_cmp = new CmpINode(in_cset_fast_test_load, phase->igvn().zerocon(T_INT));
2816 phase->register_new_node(in_cset_fast_test_cmp, c);
2817 Node* in_cset_fast_test_test = new BoolNode(in_cset_fast_test_cmp, BoolTest::ne);
2818 phase->register_new_node(in_cset_fast_test_test, c);
2819 IfNode* in_cset_fast_test_iff = new IfNode(c, in_cset_fast_test_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
2820 phase->register_control(in_cset_fast_test_iff, loop, c);
2821
2822 Node* in_cset_fast_test_success = new IfFalseNode(in_cset_fast_test_iff);
2823 phase->register_control(in_cset_fast_test_success, loop, in_cset_fast_test_iff);
2824
2825 region->init_req(3, in_cset_fast_test_success);
2826 val_phi->init_req(3, rbtrue);
2827 mem_phi->init_req(3, wb_mem);
2828 raw_mem_phi->init_req(3, raw_mem);
2829
2830 Node* in_cset_fast_test_failure = new IfTrueNode(in_cset_fast_test_iff);
2831 phase->register_control(in_cset_fast_test_failure, loop, in_cset_fast_test_iff);
2832
2833 c = in_cset_fast_test_failure;
2834 }
2835 }
2836
2837 void ShenandoahWriteBarrierNode::evacuation_in_progress(Node* c, Node* val, Node* evacuation_iff, Node* unc, Node* unc_ctrl,
2838 Node* raw_mem, Node* wb_mem, Node* region, Node* val_phi, Node* mem_phi,
2839 Node* raw_mem_phi, Node* unc_region, int alias, Unique_Node_List& uses,
2840 PhaseIdealLoop* phase) {
2841 evacuation_in_progress_null_check(c, val, evacuation_iff, unc, unc_ctrl, unc_region, uses, phase);
2842
2843 IdealLoopTree *loop = phase->get_loop(c);
2844 Node* rbtrue = new ShenandoahReadBarrierNode(c, wb_mem, val);
2845 phase->register_new_node(rbtrue, c);
2846
2847 Node* in_cset_fast_test_failure = NULL;
2848 in_cset_fast_test(c, rbtrue, raw_mem, wb_mem, region, val_phi, mem_phi, raw_mem_phi, phase);
2849
2850 // The slow path stub consumes and produces raw memory in addition
2851 // to the existing memory edges
2852 Node* base = find_bottom_mem(c, phase);
2853
2854 MergeMemNode* mm = MergeMemNode::make(base);
2855 mm->set_memory_at(alias, wb_mem);
2856 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
2857 phase->register_new_node(mm, c);
2858
2859 Node* call = new CallLeafNoFPNode(ShenandoahBarrierSetC2::shenandoah_write_barrier_Type(), ShenandoahBarrierSetAssembler::shenandoah_wb_C(), "shenandoah_write_barrier", TypeRawPtr::BOTTOM);
2860 call->init_req(TypeFunc::Control, c);
2861 call->init_req(TypeFunc::I_O, phase->C->top());
2862 call->init_req(TypeFunc::Memory, mm);
2863 call->init_req(TypeFunc::FramePtr, phase->C->top());
2864 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
2865 call->init_req(TypeFunc::Parms, rbtrue);
2866 phase->register_control(call, loop, c);
2867 Node* ctrl_proj = new ProjNode(call, TypeFunc::Control);
2868 phase->register_control(ctrl_proj, loop, call);
2869 Node* mem_proj = new ProjNode(call, TypeFunc::Memory);
2870 phase->register_new_node(mem_proj, call);
2871 Node* res_proj = new ProjNode(call, TypeFunc::Parms);
2872 phase->register_new_node(res_proj, call);
2873 Node* res = new CheckCastPPNode(ctrl_proj, res_proj, phase->igvn().type(val)->is_oopptr()->cast_to_nonconst());
2874 phase->register_new_node(res, ctrl_proj);
2875 region->init_req(2, ctrl_proj);
2876 val_phi->init_req(2, res);
2877 mem_phi->init_req(2, mem_proj);
2878 raw_mem_phi->init_req(2, mem_proj);
2879 }
2880
2881 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) {
2882 Node* ctrl = phase->get_ctrl(barrier);
2883 Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
2884
2885 // Update the control of all nodes that should be after the
2886 // barrier control flow
2887 uses.clear();
2888 // Every node that is control dependent on the barrier's input
2889 // control will be after the expanded barrier. The raw memory (if
2890 // its memory is control dependent on the barrier's input control)
2891 // must stay above the barrier.
2892 uses_to_ignore.clear();
2893 if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
2894 uses_to_ignore.push(init_raw_mem);
2895 }
2896 for (uint next = 0; next < uses_to_ignore.size(); next++) {
2897 Node *n = uses_to_ignore.at(next);
2898 for (uint i = 0; i < n->req(); i++) {
2899 Node* in = n->in(i);
2900 if (in != NULL && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
2901 uses_to_ignore.push(in);
2902 }
2903 }
2904 }
2905 for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
2906 Node* u = ctrl->fast_out(i);
2907 if (u->_idx < last &&
2908 u != barrier &&
2909 !uses_to_ignore.member(u) &&
2910 (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) &&
2911 (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) {
2912 Node* old_c = phase->ctrl_or_self(u);
2913 Node* c = old_c;
2914 if (c != ctrl ||
2915 is_dominator_same_ctrl(old_c, barrier, u, phase) ||
2916 u->is_g1_marking_load()) {
2917 phase->igvn().rehash_node_delayed(u);
2918 int nb = u->replace_edge(ctrl, region);
2919 if (u->is_CFG()) {
2920 if (phase->idom(u) == ctrl) {
2921 phase->set_idom(u, region, phase->dom_depth(region));
2922 }
2923 } else if (phase->get_ctrl(u) == ctrl) {
2924 assert(u != init_raw_mem, "should leave input raw mem above the barrier");
2925 uses.push(u);
2926 }
2927 assert(nb == 1, "more than 1 ctrl input?");
2928 --i, imax -= nb;
2929 }
2930 }
2931 }
2932 }
2933
2934
2935 void ShenandoahWriteBarrierNode::pin_and_expand(PhaseIdealLoop* phase) {
2936 Node_List enqueue_barriers;
2937 if (ShenandoahStoreValEnqueueBarrier) {
2938 Unique_Node_List wq;
2939 wq.push(phase->C->root());
2940 for (uint i = 0; i < wq.size(); i++) {
2941 Node* n = wq.at(i);
2942 if (n->Opcode() == Op_ShenandoahEnqueueBarrier) {
2943 enqueue_barriers.push(n);
2944 }
2945 for (uint i = 0; i < n->req(); i++) {
2946 Node* in = n->in(i);
2947 if (in != NULL) {
2948 wq.push(in);
2949 }
2950 }
2951 }
2952 }
2953
2954 const bool trace = false;
2955
2956 // Collect raw memory state at CFG points in the entire graph and
2957 // record it in memory_nodes. Optimize the raw memory graph in the
2958 // process. Optimizing the memory graph also makes the memory graph
2959 // simpler.
2960 GrowableArray<MemoryGraphFixer*> memory_graph_fixers;
2961
2962 // Let's try to common write barriers again
2963 optimize_before_expansion(phase, memory_graph_fixers, true);
2964
2965 Unique_Node_List uses;
2966 for (int i = 0; i < ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count(); i++) {
2967 ShenandoahWriteBarrierNode* wb = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barrier(i);
2968 Node* ctrl = phase->get_ctrl(wb);
2969
2970 Node* val = wb->in(ValueIn);
2971 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) {
2972 assert(is_dominator(phase->get_ctrl(val), ctrl->in(0)->in(0), val, ctrl->in(0), phase), "can't move");
2973 phase->set_ctrl(wb, ctrl->in(0)->in(0));
2974 } else if (ctrl->is_CallRuntime()) {
2975 assert(is_dominator(phase->get_ctrl(val), ctrl->in(0), val, ctrl, phase), "can't move");
2976 phase->set_ctrl(wb, ctrl->in(0));
2977 }
2978
2979 assert(wb->Opcode() == Op_ShenandoahWriteBarrier, "only for write barriers");
2980 // Look for a null check that dominates this barrier and move the
2981 // barrier right after the null check to enable implicit null
2982 // checks
2983 wb->pin_and_expand_move_barrier(phase, memory_graph_fixers, uses);
2984
2985 wb->pin_and_expand_helper(phase);
2986 }
2987
2988 MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
2989 Unique_Node_List uses_to_ignore;
2990 for (uint i = 0; i < enqueue_barriers.size(); i++) {
2991 Node* barrier = enqueue_barriers.at(i);
2992 Node* pre_val = barrier->in(1);
2993
2994 if (phase->igvn().type(pre_val)->higher_equal(TypePtr::NULL_PTR)) {
2995 ShouldNotReachHere();
2996 continue;
2997 }
2998
2999 Node* ctrl = phase->get_ctrl(barrier);
3000
3001 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) {
3002 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0)->in(0), pre_val, ctrl->in(0), phase), "can't move");
3003 ctrl = ctrl->in(0)->in(0);
3004 phase->set_ctrl(barrier, ctrl);
3005 } else if (ctrl->is_CallRuntime()) {
3006 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0), pre_val, ctrl, phase), "can't move");
3007 ctrl = ctrl->in(0);
3008 phase->set_ctrl(barrier, ctrl);
3009 }
3010
3011 Node* init_ctrl = ctrl;
3012 IdealLoopTree* loop = phase->get_loop(ctrl);
3013 Node* raw_mem = fixer.find_mem(ctrl, barrier);
3014 Node* init_raw_mem = raw_mem;
3015 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL);
3016 Node* evac_in_progress = NULL;
3017 Node* evac_not_in_progress = NULL;
3018 Node* heap_stable = NULL;
3019 Node* null_val = NULL;
3020 uint last = phase->C->unique();
3021
3022 Node* stable_test_region = new RegionNode(3);
3023 Node* stable_test_phi = PhiNode::make(stable_test_region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
3024 Node* unstable_region = new RegionNode(5);
3025 Node* unstable_phi = PhiNode::make(unstable_region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
3026
3027 test_evacuation_in_progress(ctrl, pre_val, raw_mem, evac_in_progress, evac_not_in_progress, heap_stable, null_val, phase);
3028
3029 stable_test_region->init_req(1, heap_stable);
3030 stable_test_region->init_req(2, unstable_region);
3031 stable_test_phi->init_req(1, raw_mem);
3032 stable_test_phi->init_req(2, unstable_phi);
3033
3034 if (null_val != NULL) {
3035 unstable_region->init_req(1, null_val);
3036 unstable_phi->init_req(1, raw_mem);
3037 }
3038 unstable_region->init_req(2, evac_not_in_progress);
3039 unstable_phi->init_req(2, raw_mem);
3040
3041 ctrl = evac_in_progress;
3042
3043 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
3044 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
3045 Node* thread = new ThreadLocalNode();
3046 phase->register_new_node(thread, ctrl);
3047 Node* buffer_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(buffer_offset));
3048 phase->register_new_node(buffer_adr, ctrl);
3049 Node* index_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(index_offset));
3050 phase->register_new_node(index_adr, ctrl);
3051
3052 BasicType index_bt = TypeX_X->basic_type();
3053 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
3054 const TypePtr* adr_type = TypeRawPtr::BOTTOM;
3055 Node* index = new LoadXNode(ctrl, raw_mem, index_adr, adr_type, TypeX_X, MemNode::unordered);
3056 phase->register_new_node(index, ctrl);
3057 Node* index_cmp = new CmpXNode(index, phase->igvn().MakeConX(0));
3058 phase->register_new_node(index_cmp, ctrl);
3059 Node* index_test = new BoolNode(index_cmp, BoolTest::ne);
3060 phase->register_new_node(index_test, ctrl);
3061 IfNode* queue_full_iff = new IfNode(ctrl, index_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
3062 phase->register_control(queue_full_iff, loop, ctrl);
3063 Node* not_full = new IfTrueNode(queue_full_iff);
3064 phase->register_control(not_full, loop, queue_full_iff);
3065 Node* full = new IfFalseNode(queue_full_iff);
3066 phase->register_control(full, loop, queue_full_iff);
3067
3068 ctrl = not_full;
3069
3070 Node* next_index = new SubXNode(index, phase->igvn().MakeConX(sizeof(intptr_t)));
3071 phase->register_new_node(next_index, ctrl);
3072
3073 Node* buffer = new LoadPNode(ctrl, raw_mem, buffer_adr, adr_type, TypeRawPtr::NOTNULL, MemNode::unordered);
3074 phase->register_new_node(buffer, ctrl);
3075 Node *log_addr = new AddPNode(phase->C->top(), buffer, next_index);
3076 phase->register_new_node(log_addr, ctrl);
3077 Node* log_store = new StorePNode(ctrl, raw_mem, log_addr, adr_type, pre_val, MemNode::unordered);
3078 phase->register_new_node(log_store, ctrl);
3079 // update the index
3080 Node* index_update = new StoreXNode(ctrl, log_store, index_adr, adr_type, next_index, MemNode::unordered);
3081 phase->register_new_node(index_update, ctrl);
3082
3083 unstable_region->init_req(3, ctrl);
3084 unstable_phi->init_req(3, index_update);
3085
3086 ctrl = full;
3087
3088 Node* base = find_bottom_mem(ctrl, phase);
3089
3090 MergeMemNode* mm = MergeMemNode::make(base);
3091 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
3092 phase->register_new_node(mm, ctrl);
3093
3094 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);
3095 call->init_req(TypeFunc::Control, ctrl);
3096 call->init_req(TypeFunc::I_O, phase->C->top());
3097 call->init_req(TypeFunc::Memory, mm);
3098 call->init_req(TypeFunc::FramePtr, phase->C->top());
3099 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
3100 call->init_req(TypeFunc::Parms, pre_val);
3101 call->init_req(TypeFunc::Parms+1, thread);
3102 phase->register_control(call, loop, ctrl);
3103
3104 Node* ctrl_proj = new ProjNode(call, TypeFunc::Control);
3105 phase->register_control(ctrl_proj, loop, call);
3106 Node* mem_proj = new ProjNode(call, TypeFunc::Memory);
3107 phase->register_new_node(mem_proj, call);
3108
3109 unstable_region->init_req(4, ctrl_proj);
3110 unstable_phi->init_req(4, mem_proj);
3111
3112 phase->register_control(unstable_region, loop, null_val != NULL ? null_val->in(0) : evac_in_progress->in(0));
3113 phase->register_new_node(unstable_phi, unstable_region);
3114 phase->register_control(stable_test_region, loop, heap_stable->in(0));
3115 phase->register_new_node(stable_test_phi, stable_test_region);
3116
3117 fix_ctrl(barrier, stable_test_region, fixer, uses, uses_to_ignore, last, phase);
3118 for(uint next = 0; next < uses.size(); next++ ) {
3119 Node *n = uses.at(next);
3120 assert(phase->get_ctrl(n) == init_ctrl, "bad control");
3121 assert(n != init_raw_mem, "should leave input raw mem above the barrier");
3122 phase->set_ctrl(n, stable_test_region);
3123 follow_barrier_uses(n, init_ctrl, uses, phase);
3124 }
3125 fixer.fix_mem(init_ctrl, stable_test_region, init_raw_mem, raw_mem_for_ctrl, stable_test_phi, uses);
3126
3127 phase->igvn().replace_node(barrier, pre_val);
3128 }
3129
3130 for (int i = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count(); i > 0; i--) {
3131 int cnt = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count();
3132 ShenandoahWriteBarrierNode* wb = ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barrier(i-1);
3133
3134 uint last = phase->C->unique();
3135 Node* ctrl = phase->get_ctrl(wb);
3136
3137 Node* raw_mem = fixer.find_mem(ctrl, wb);
3138 Node* init_raw_mem = raw_mem;
3139 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL);
3140 int alias = phase->C->get_alias_index(wb->adr_type());
3141 Node* wb_mem = wb->in(Memory);
3142 Node* init_wb_mem = wb_mem;
3143
3144 Node* val = wb->in(ValueIn);
3145 Node* wbproj = wb->find_out_with(Op_ShenandoahWBMemProj);
3146 IdealLoopTree *loop = phase->get_loop(ctrl);
3147
3148 assert(val->Opcode() != Op_ShenandoahWriteBarrier, "No chain of write barriers");
3149
3150 CallStaticJavaNode* unc = wb->pin_and_expand_null_check(phase->igvn());
3151 Node* unc_ctrl = NULL;
3152 if (unc != NULL) {
3153 if (val->in(0) != ctrl) {
3154 unc = NULL;
3155 } else {
3156 unc_ctrl = val->in(0);
3157 }
3158 }
3159
3160 Node* uncasted_val = val;
3161 if (unc != NULL) {
3162 uncasted_val = val->in(1);
3163 }
3164
3165 Node* evac_in_progress = NULL;
3166 Node* evac_not_in_progress = NULL;
3167 Node* heap_stable_ctrl = NULL;
3168 Node* null_ctrl = NULL;
3169 test_evacuation_in_progress(ctrl, val, raw_mem, evac_in_progress, evac_not_in_progress, heap_stable_ctrl, null_ctrl, phase);
3170 IfNode* evacuation_iff = evac_in_progress->in(0)->as_If();
3171 IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
3172
3173 Node* evacuation_region = new RegionNode(5);
3174 Node* evacuation_val_phi = new PhiNode(evacuation_region, uncasted_val->bottom_type()->is_oopptr()->cast_to_nonconst());
3175 Node* evacuation_mem_phi = PhiNode::make(evacuation_region, wb_mem, Type::MEMORY, phase->C->alias_type(wb->adr_type())->adr_type());
3176 Node* evacuation_raw_mem_phi = PhiNode::make(evacuation_region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
3177 Node* region = new RegionNode(3);
3178 Node* val_phi = new PhiNode(region, uncasted_val->bottom_type()->is_oopptr()->cast_to_nonconst());
3179 Node* mem_phi = PhiNode::make(region, wb_mem, Type::MEMORY, phase->C->alias_type(wb->adr_type())->adr_type());
3180 Node* raw_mem_phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
3181
3182 if (null_ctrl != NULL) {
3183 evacuation_region->init_req(4, null_ctrl);
3184 evacuation_val_phi->init_req(4, phase->igvn().zerocon(T_OBJECT));
3185 evacuation_mem_phi->init_req(4, wb_mem);
3186 evacuation_raw_mem_phi->init_req(4, raw_mem);
3187 } else {
3188 evacuation_region->del_req(4);
3189 evacuation_val_phi->del_req(4);
3190 evacuation_mem_phi->del_req(4);
3191 evacuation_raw_mem_phi->del_req(4);
3192 }
3193
3194 Node* unc_region = NULL;
3195 evacuation_not_in_progress(evac_not_in_progress, val, unc_ctrl, raw_mem, wb_mem,
3196 evacuation_region, evacuation_val_phi, evacuation_mem_phi, evacuation_raw_mem_phi, unc_region,
3197 phase);
3198
3199 heap_stable(heap_stable_ctrl, val, unc_ctrl, init_raw_mem, init_wb_mem, region, val_phi, mem_phi, raw_mem_phi,
3200 unc_region, phase);
3201
3202 evacuation_in_progress(evac_in_progress, val, evacuation_iff, unc, unc_ctrl,
3203 raw_mem, wb_mem, evacuation_region, evacuation_val_phi, evacuation_mem_phi, evacuation_raw_mem_phi,
3204 unc_region, alias, uses,
3205 phase);
3206 region->init_req(2, evacuation_region);
3207 val_phi->init_req(2, evacuation_val_phi);
3208 mem_phi->init_req(2, evacuation_mem_phi);
3209 raw_mem_phi->init_req(2, evacuation_raw_mem_phi);
3210 phase->register_control(evacuation_region, loop, evacuation_iff);
3211 phase->register_new_node(evacuation_val_phi, evacuation_region);
3212 phase->register_new_node(evacuation_mem_phi, evacuation_region);
3213 phase->register_new_node(evacuation_raw_mem_phi, evacuation_region);
3214
3215 phase->register_control(region, loop, heap_stable_iff);
3216
3217 Node* out_val = val_phi;
3218 phase->register_new_node(val_phi, region);
3219 phase->register_new_node(mem_phi, region);
3220 phase->register_new_node(raw_mem_phi, region);
3221
3222 fix_ctrl(wb, region, fixer, uses, uses_to_ignore, last, phase);
3223
3224 phase->igvn().replace_input_of(wbproj, 0, phase->C->top());
3225 phase->lazy_replace(wbproj, mem_phi);
3226
3227 if (unc != NULL) {
3228 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
3229 Node* u = val->fast_out(i);
3230 Node* c = phase->ctrl_or_self(u);
3231 if (u != wb && (c != ctrl || is_dominator_same_ctrl(c, wb, u, phase))) {
3232 phase->igvn().rehash_node_delayed(u);
3233 int nb = u->replace_edge(val, out_val);
3234 --i, imax -= nb;
3235 }
3236 }
3237 if (val->outcnt() == 0) {
3238 phase->lazy_update(val, out_val);
3239 phase->igvn()._worklist.push(val);
3240 }
3241 }
3242 phase->lazy_replace(wb, out_val);
3243
3244 follow_barrier_uses(mem_phi, ctrl, uses, phase);
3245 follow_barrier_uses(out_val, ctrl, uses, phase);
3246
3247 for(uint next = 0; next < uses.size(); next++ ) {
3248 Node *n = uses.at(next);
3249 assert(phase->get_ctrl(n) == ctrl, "bad control");
3250 assert(n != init_raw_mem, "should leave input raw mem above the barrier");
3251 phase->set_ctrl(n, region);
3252 follow_barrier_uses(n, ctrl, uses, phase);
3253 }
3254
3255 // The slow path call produces memory: hook the raw memory phi
3256 // from the expanded write barrier with the rest of the graph
3257 // which may require adding memory phis at every post dominated
3258 // region and at enclosing loop heads. Use the memory state
3259 // collected in memory_nodes to fix the memory graph. Update that
3260 // memory state as we go.
3261 fixer.fix_mem(ctrl, region, init_raw_mem, raw_mem_for_ctrl, raw_mem_phi, uses);
3262 assert(ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count() == cnt - 1, "not replaced");
3263 }
3264
3265 assert(ShenandoahBarrierSetC2::bsc2()->state()->shenandoah_barriers_count() == 0, "all write barrier nodes should have been replaced");
3266 }
3267
3268 void ShenandoahWriteBarrierNode::move_evacuation_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
3269 // move test and its mem barriers out of the loop
3270 assert(is_evacuation_in_progress_test(iff), "inconsistent");
3271
3272 IdealLoopTree *loop = phase->get_loop(iff);
3273 Node* loop_head = loop->_head;
3274 Node* entry_c = loop_head->in(LoopNode::EntryControl);
3275
3276 Node* load = iff->in(1)->in(1)->in(1)->in(1);
3277 assert(is_gc_state_load(load), "broken");
3278 if (!phase->is_dominator(load->in(0), entry_c)) {
3279 Node* mem_ctrl = NULL;
3280 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
3281 phase->igvn().replace_input_of(load, MemNode::Memory, mem);
3282 phase->igvn().replace_input_of(load, 0, entry_c);
3283 phase->set_ctrl_and_loop(load, entry_c);
3284 }
3285 }
3286
3287 void ShenandoahWriteBarrierNode::move_heap_stable_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
3288 IdealLoopTree *loop = phase->get_loop(iff);
3289 Node* loop_head = loop->_head;
3290 Node* entry_c = loop_head->in(LoopNode::EntryControl);
3291
3292 Node* load = iff->in(1)->in(1)->in(1);
3293 assert(is_gc_state_load(load), "broken");
3294 if (!phase->is_dominator(load->in(0), entry_c)) {
3295 Node* mem_ctrl = NULL;
3296 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
3297 phase->igvn().replace_input_of(load, MemNode::Memory, mem);
3298 phase->igvn().replace_input_of(load, 0, entry_c);
3299 phase->set_ctrl_and_loop(load, entry_c);
3300 }
3301 }
3302
3303 void ShenandoahWriteBarrierNode::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) {
3304 assert(is_evacuation_in_progress_test(n) || is_heap_stable_test(n), "no other tests");
3305 if (phase->identical_backtoback_ifs(n)) {
3306 Node* n_ctrl = is_evacuation_in_progress_test(n) ? ShenandoahWriteBarrierNode::evacuation_in_progress_test_ctrl(n) : n->in(0);
3307 if (phase->can_split_if(n_ctrl)) {
3308 IfNode* dom_if = phase->idom(n_ctrl)->as_If();
3309 if (is_heap_stable_test(n)) {
3310 Node* gc_state_load = n->in(1)->in(1)->in(1);
3311 assert(is_gc_state_load(gc_state_load), "broken");
3312 Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1);
3313 assert(is_gc_state_load(dom_gc_state_load), "broken");
3314 if (gc_state_load != dom_gc_state_load) {
3315 phase->igvn().replace_node(gc_state_load, dom_gc_state_load);
3316 }
3317 }
3318 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
3319 Node* proj_true = dom_if->proj_out(1);
3320 Node* proj_false = dom_if->proj_out(0);
3321 Node* con_true = phase->igvn().makecon(TypeInt::ONE);
3322 Node* con_false = phase->igvn().makecon(TypeInt::ZERO);
3323
3324 for (uint i = 1; i < n_ctrl->req(); i++) {
3325 if (phase->is_dominator(proj_true, n_ctrl->in(i))) {
3326 bolphi->init_req(i, con_true);
3327 } else {
3328 assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if");
3329 bolphi->init_req(i, con_false);
3330 }
3331 }
3332 phase->register_new_node(bolphi, n_ctrl);
3333 phase->igvn().replace_input_of(n, 1, bolphi);
3334 phase->do_split_if(n);
3335 }
3336 }
3337 }
3338
3339 void ShenandoahWriteBarrierNode::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) {
3340 Node_List heap_stable_tests;
3341 Node_List evacuation_tests;
3342 Node_List gc_state_loads;
3343
3344 stack.push(phase->C->start(), 0);
3345 do {
3346 Node* n = stack.node();
3347 uint i = stack.index();
3348
3349 if (i < n->outcnt()) {
3350 Node* u = n->raw_out(i);
3351 stack.set_index(i+1);
3352 if (!visited.test_set(u->_idx)) {
3353 stack.push(u, 0);
3354 }
3355 } else {
3356 stack.pop();
3357 if (n->is_If() && ShenandoahWriteBarrierNode::is_evacuation_in_progress_test(n)) {
3358 evacuation_tests.push(n);
3359 }
3360 if (ShenandoahCommonGCStateLoads && ShenandoahWriteBarrierNode::is_gc_state_load(n)) {
3361 gc_state_loads.push(n);
3362 }
3363 if (n->is_If() && ShenandoahWriteBarrierNode::is_heap_stable_test(n)) {
3364 heap_stable_tests.push(n);
3365 }
3366 }
3367 } while (stack.size() > 0);
3368
3369 bool progress;
3370 do {
3371 progress = false;
3372 for (uint i = 0; i < gc_state_loads.size(); i++) {
3373 Node* n = gc_state_loads.at(i);
3374 if (n->outcnt() != 0) {
3375 progress |= ShenandoahWriteBarrierNode::try_common_gc_state_load(n, phase);
3376 }
3377 }
3378 } while (progress);
3379
3380 for (uint i = 0; i < heap_stable_tests.size(); i++) {
3381 Node* n = heap_stable_tests.at(i);
3382 assert(is_heap_stable_test(n), "only evacuation test");
3383 merge_back_to_back_tests(n, phase);
3384 }
3385
3386 if (!phase->C->major_progress()) {
3387 for (uint i = 0; i < evacuation_tests.size(); i++) {
3388 Node* n = evacuation_tests.at(i);
3389 assert(is_evacuation_in_progress_test(n), "only evacuation test");
3390 merge_back_to_back_tests(n, phase);
3391 }
3392 }
3393
3394 if (!phase->C->major_progress()) {
3395 VectorSet seen(Thread::current()->resource_area());
3396 for (uint i = 0; i < heap_stable_tests.size(); i++) {
3397 Node* n = heap_stable_tests.at(i);
3398 IdealLoopTree* loop = phase->get_loop(n);
3399 if (loop != phase->ltree_root() &&
3400 loop->_child == NULL &&
3401 !loop->_irreducible) {
3402 LoopNode* head = loop->_head->as_Loop();
3403 if ((!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) &&
3404 !seen.test_set(head->_idx) &&
3405 loop->policy_unswitching(phase, true)) {
3406 IfNode* iff = phase->find_unswitching_candidate(loop, true);
3407 if (iff != NULL && (is_evacuation_in_progress_test(iff) || is_heap_stable_test(iff))) {
3408 if (head->is_strip_mined()) {
3409 head->verify_strip_mined(0);
3410 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop();
3411 OuterStripMinedLoopEndNode* le = head->outer_loop_end();
3412 Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl));
3413 phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl));
3414 Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt);
3415 phase->register_control(new_le, phase->get_loop(le), le->in(0));
3416 phase->lazy_replace(outer, new_outer);
3417 phase->lazy_replace(le, new_le);
3418 head->clear_strip_mined();
3419 }
3420 phase->do_unswitching(loop, old_new, true);
3421 }
3422 }
3423 }
3424 }
3425 }
3426 }
3427
3428 #ifdef ASSERT
3429 void ShenandoahBarrierNode::verify_raw_mem(RootNode* root) {
3430 const bool trace = false;
3431 ResourceMark rm;
3432 Unique_Node_List nodes;
3433 Unique_Node_List controls;
3434 Unique_Node_List memories;
3435
3436 nodes.push(root);
3437 for (uint next = 0; next < nodes.size(); next++) {
3438 Node *n = nodes.at(next);
3439 if (n->Opcode() == Op_CallLeafNoFP &&
3440 ShenandoahBarrierSetAssembler::is_shenandoah_wb_C_call(n->as_Call()->entry_point())) {
3441 controls.push(n);
3442 if (trace) { tty->print("XXXXXX verifying"); n->dump(); }
3443 for (uint next2 = 0; next2 < controls.size(); next2++) {
3444 Node *m = controls.at(next2);
3445 if (!m->is_Loop() || controls.member(m->in(LoopNode::EntryControl)) || 1) {
3446 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3447 Node* u = m->fast_out(i);
3448 if (u->is_CFG() && !u->is_Root() &&
3449 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1) &&
3450 !(u->is_Region() && u->unique_ctrl_out()->Opcode() == Op_Halt)) {
3451 if (trace) { tty->print("XXXXXX pushing control"); u->dump(); }
3452 controls.push(u);
3453 }
3454 }
3455 }
3456 }
3457 memories.push(n->as_Call()->proj_out(TypeFunc::Memory));
3458 for (uint next2 = 0; next2 < memories.size(); next2++) {
3459 Node *m = memories.at(next2);
3460 assert(m->bottom_type() == Type::MEMORY, "");
3461 if (!m->is_Phi() || !m->in(0)->is_Loop() || controls.member(m->in(0)->in(LoopNode::EntryControl)) || 1) {
3462 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
3463 Node* u = m->fast_out(i);
3464 if (u->bottom_type() == Type::MEMORY && (u->is_Mem() || u->is_ClearArray())) {
3465 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
3466 memories.push(u);
3467 } else if (u->is_LoadStore()) {
3468 if (trace) { tty->print("XXXXXX pushing memory"); u->find_out_with(Op_SCMemProj)->dump(); }
3469 memories.push(u->find_out_with(Op_SCMemProj));
3470 } else if (u->is_MergeMem() && u->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == m) {
3471 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
3472 memories.push(u);
3473 } else if (u->is_Phi()) {
3474 assert(u->bottom_type() == Type::MEMORY, "");
3475 if (u->adr_type() == TypeRawPtr::BOTTOM || u->adr_type() == TypePtr::BOTTOM) {
3476 assert(controls.member(u->in(0)), "");
3477 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
3478 memories.push(u);
3479 }
3480 } else if (u->is_SafePoint() || u->is_MemBar()) {
3481 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
3482 Node* uu = u->fast_out(j);
3483 if (uu->bottom_type() == Type::MEMORY) {
3484 if (trace) { tty->print("XXXXXX pushing memory"); uu->dump(); }
3485 memories.push(uu);
3486 }
3487 }
3488 }
3489 }
3490 }
3491 }
3492 for (uint next2 = 0; next2 < controls.size(); next2++) {
3493 Node *m = controls.at(next2);
3494 if (m->is_Region()) {
3495 bool all_in = true;
3496 for (uint i = 1; i < m->req(); i++) {
3497 if (!controls.member(m->in(i))) {
3498 all_in = false;
3499 break;
3500 }
3501 }
3502 if (trace) { tty->print("XXX verifying %s", all_in ? "all in" : ""); m->dump(); }
3503 bool found_phi = false;
3504 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax && !found_phi; j++) {
3505 Node* u = m->fast_out(j);
3506 if (u->is_Phi() && memories.member(u)) {
3507 found_phi = true;
3508 for (uint i = 1; i < u->req() && found_phi; i++) {
3509 Node* k = u->in(i);
3510 if (memories.member(k) != controls.member(m->in(i))) {
3511 found_phi = false;
3512 }
3513 }
3514 }
3515 }
3516 assert(found_phi || all_in, "");
3517 }
3518 }
3519 controls.clear();
3520 memories.clear();
3521 }
3522 for( uint i = 0; i < n->len(); ++i ) {
3523 Node *m = n->in(i);
3524 if (m != NULL) {
3525 nodes.push(m);
3526 }
3527 }
3528 }
3529 }
3530 #endif
3531
3532 const Type* ShenandoahEnqueueBarrierNode::bottom_type() const {
3533 if (in(1) == NULL || in(1)->is_top()) {
3534 return Type::TOP;
3535 }
3536 const Type* t = in(1)->bottom_type();
3537 if (t == TypePtr::NULL_PTR) {
3538 return t;
3539 }
3540 return t->is_oopptr()->cast_to_nonconst();
3541 }
3542
3543 const Type* ShenandoahEnqueueBarrierNode::Value(PhaseGVN* phase) const {
3544 if (in(1) == NULL) {
3545 return Type::TOP;
3546 }
3547 const Type* t = phase->type(in(1));
3548 if (t == Type::TOP) {
3549 return Type::TOP;
3550 }
3551 if (t == TypePtr::NULL_PTR) {
3552 return t;
3553 }
3554 return t->is_oopptr()->cast_to_nonconst();
3555 }
3556
3557 int ShenandoahEnqueueBarrierNode::needed(Node* n) {
3558 if (n == NULL ||
3559 n->is_Allocate() ||
3560 n->bottom_type() == TypePtr::NULL_PTR ||
3561 n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL && !ShenandoahBarriersForConst) {
3562 return NotNeeded;
3563 }
3564 if (n->is_Phi() ||
3565 n->is_CMove()) {
3566 return MaybeNeeded;
3567 }
3568 return Needed;
3569 }
3570
3571 Node* ShenandoahEnqueueBarrierNode::next(Node* n) {
3572 for (;;) {
3573 if (n == NULL) {
3574 return n;
3575 } else if (n->bottom_type() == TypePtr::NULL_PTR) {
3576 return n;
3577 } else if (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL && !ShenandoahBarriersForConst) {
3578 return n;
3579 } else if (n->is_ConstraintCast() ||
3580 n->Opcode() == Op_DecodeN ||
3581 n->Opcode() == Op_EncodeP) {
3582 n = n->in(1);
3583 } else if (n->is_Proj()) {
3584 n = n->in(0);
3585 } else {
3586 return n;
3587 }
3588 }
3589 ShouldNotReachHere();
3590 return NULL;
3591 }
3592
3593
3594 Node* ShenandoahEnqueueBarrierNode::Identity(PhaseGVN* phase) {
3595 PhaseIterGVN* igvn = phase->is_IterGVN();
3596
3597 Node* n = next(in(1));
3598
3599 int cont = needed(n);
3600
3601 if (cont == NotNeeded) {
3602 return in(1);
3603 } else if (cont == MaybeNeeded) {
3604 if (igvn == NULL) {
3605 phase->record_for_igvn(this);
3606 return this;
3607 } else {
3608 ResourceMark rm;
3609 Unique_Node_List wq;
3610 uint wq_i = 0;
3611
3612 for (;;) {
3613 if (n->is_Phi()) {
3614 for (uint i = 1; i < n->req(); i++) {
3615 Node* m = n->in(i);
3616 if (m != NULL) {
3617 wq.push(m);
3618 }
3619 }
3620 } else {
3621 assert(n->is_CMove(), "nothing else here");
3622 Node* m = n->in(CMoveNode::IfFalse);
3623 wq.push(m);
3624 m = n->in(CMoveNode::IfTrue);
3625 wq.push(m);
3626 }
3627 Node* orig_n = NULL;
3628 do {
3629 if (wq_i >= wq.size()) {
3630 return in(1);
3631 }
3632 n = wq.at(wq_i);
3633 wq_i++;
3634 orig_n = n;
3635 n = next(n);
3636 cont = needed(n);
3637 if (cont == Needed) {
3638 return this;
3639 }
3640 } while (cont != MaybeNeeded || (orig_n != n && wq.member(n)));
3641 }
3642 }
3643 }
3644
3645 return this;
3646 }
3647
3648 #ifdef ASSERT
3649 static bool has_never_branch(Node* root) {
3650 for (uint i = 1; i < root->req(); i++) {
3651 Node* in = root->in(i);
3652 if (in != NULL && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->Opcode() == Op_NeverBranch) {
3653 return true;
3654 }
3655 }
3656 return false;
3657 }
3658 #endif
3659
3660 void MemoryGraphFixer::collect_memory_nodes() {
3661 Node_Stack stack(0);
3662 VectorSet visited(Thread::current()->resource_area());
3663 Node_List regions;
3664
3665 // Walk the raw memory graph and create a mapping from CFG node to
3666 // memory node. Exclude phis for now.
3667 stack.push(_phase->C->root(), 1);
3668 do {
3669 Node* n = stack.node();
3670 int opc = n->Opcode();
3671 uint i = stack.index();
3672 if (i < n->req()) {
3673 Node* mem = NULL;
3674 if (opc == Op_Root) {
3675 Node* in = n->in(i);
3676 int in_opc = in->Opcode();
3677 if (in_opc == Op_Return || in_opc == Op_Rethrow) {
3678 mem = in->in(TypeFunc::Memory);
3679 } else if (in_opc == Op_Halt) {
3680 if (!in->in(0)->is_Region()) {
3681 Node* proj = in->in(0);
3682 assert(proj->is_Proj(), "");
3683 Node* in = proj->in(0);
3684 assert(in->is_CallStaticJava() || in->Opcode() == Op_NeverBranch || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
3685 if (in->is_CallStaticJava()) {
3686 mem = in->in(TypeFunc::Memory);
3687 } else if (in->Opcode() == Op_Catch) {
3688 Node* call = in->in(0)->in(0);
3689 assert(call->is_Call(), "");
3690 mem = call->in(TypeFunc::Memory);
3691 }
3692 }
3693 } else {
3694 #ifdef ASSERT
3695 n->dump();
3696 in->dump();
3697 #endif
3698 ShouldNotReachHere();
3699 }
3700 } else {
3701 assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
3702 assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
3703 mem = n->in(i);
3704 }
3705 i++;
3706 stack.set_index(i);
3707 if (mem == NULL) {
3708 continue;
3709 }
3710 for (;;) {
3711 if (visited.test_set(mem->_idx) || mem->is_Start()) {
3712 break;
3713 }
3714 if (mem->is_Phi()) {
3715 stack.push(mem, 2);
3716 mem = mem->in(1);
3717 } else if (mem->is_Proj()) {
3718 stack.push(mem, mem->req());
3719 mem = mem->in(0);
3720 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
3721 mem = mem->in(TypeFunc::Memory);
3722 } else if (mem->is_MergeMem()) {
3723 MergeMemNode* mm = mem->as_MergeMem();
3724 mem = mm->memory_at(_alias);
3725 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
3726 assert(_alias == Compile::AliasIdxRaw, "");
3727 stack.push(mem, mem->req());
3728 mem = mem->in(MemNode::Memory);
3729 } else if (mem->Opcode() == Op_ShenandoahWriteBarrier) {
3730 assert(_alias != Compile::AliasIdxRaw, "");
3731 mem = mem->in(ShenandoahBarrierNode::Memory);
3732 } else {
3733 #ifdef ASSERT
3734 mem->dump();
3735 #endif
3736 ShouldNotReachHere();
3737 }
3738 }
3739 } else {
3740 if (n->is_Phi()) {
3741 // Nothing
3742 } else if (!n->is_Root()) {
3743 Node* c = get_ctrl(n);
3744 _memory_nodes.map(c->_idx, n);
3745 }
3746 stack.pop();
3747 }
3748 } while(stack.is_nonempty());
3749
3750 // Iterate over CFG nodes in rpo and propagate memory state to
3751 // compute memory state at regions, creating new phis if needed.
3752 Node_List rpo_list;
3753 visited.Clear();
3754 _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
3755 Node* root = rpo_list.pop();
3756 assert(root == _phase->C->root(), "");
3757
3758 const bool trace = false;
3759 #ifdef ASSERT
3760 if (trace) {
3761 for (int i = rpo_list.size() - 1; i >= 0; i--) {
3762 Node* c = rpo_list.at(i);
3763 if (_memory_nodes[c->_idx] != NULL) {
3764 tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump();
3765 }
3766 }
3767 }
3768 #endif
3769 uint last = _phase->C->unique();
3770
3771 #ifdef ASSERT
3772 uint8_t max_depth = 0;
3773 for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
3774 IdealLoopTree* lpt = iter.current();
3775 max_depth = MAX2(max_depth, lpt->_nest);
3776 }
3777 #endif
3778
3779 bool progress = true;
3780 int iteration = 0;
3781 Node_List dead_phis;
3782 while (progress) {
3783 progress = false;
3784 iteration++;
3785 assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop(), "");
3786 if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
3787 IdealLoopTree* last_updated_ilt = NULL;
3788 for (int i = rpo_list.size() - 1; i >= 0; i--) {
3789 Node* c = rpo_list.at(i);
3790
3791 Node* prev_mem = _memory_nodes[c->_idx];
3792 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
3793 Node* prev_region = regions[c->_idx];
3794 Node* unique = NULL;
3795 for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
3796 Node* m = _memory_nodes[c->in(j)->_idx];
3797 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");
3798 if (m != NULL) {
3799 if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
3800 assert(c->is_Loop() && j == LoopNode::LoopBackControl || _phase->C->has_irreducible_loop(), "");
3801 // continue
3802 } else if (unique == NULL) {
3803 unique = m;
3804 } else if (m == unique) {
3805 // continue
3806 } else {
3807 unique = NodeSentinel;
3808 }
3809 }
3810 }
3811 assert(unique != NULL, "empty phi???");
3812 if (unique != NodeSentinel) {
3813 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c) {
3814 dead_phis.push(prev_region);
3815 }
3816 regions.map(c->_idx, unique);
3817 } else {
3818 Node* phi = NULL;
3819 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
3820 phi = prev_region;
3821 for (uint k = 1; k < c->req(); k++) {
3822 Node* m = _memory_nodes[c->in(k)->_idx];
3823 assert(m != NULL, "expect memory state");
3824 phi->set_req(k, m);
3825 }
3826 } else {
3827 for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == NULL; j++) {
3828 Node* u = c->fast_out(j);
3829 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
3830 (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
3831 phi = u;
3832 for (uint k = 1; k < c->req() && phi != NULL; k++) {
3833 Node* m = _memory_nodes[c->in(k)->_idx];
3834 assert(m != NULL, "expect memory state");
3835 if (u->in(k) != m) {
3836 phi = NULL;
3837 }
3838 }
3839 }
3840 }
3841 if (phi == NULL) {
3842 phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
3843 for (uint k = 1; k < c->req(); k++) {
3844 Node* m = _memory_nodes[c->in(k)->_idx];
3845 assert(m != NULL, "expect memory state");
3846 phi->init_req(k, m);
3847 }
3848 }
3849 }
3850 assert(phi != NULL, "");
3851 regions.map(c->_idx, phi);
3852 }
3853 Node* current_region = regions[c->_idx];
3854 if (current_region != prev_region) {
3855 progress = true;
3856 if (prev_region == prev_mem) {
3857 _memory_nodes.map(c->_idx, current_region);
3858 }
3859 }
3860 } else if (prev_mem == NULL || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
3861 Node* m = _memory_nodes[_phase->idom(c)->_idx];
3862 assert(m != NULL, "expect memory state");
3863 if (m != prev_mem) {
3864 _memory_nodes.map(c->_idx, m);
3865 progress = true;
3866 }
3867 }
3868 #ifdef ASSERT
3869 if (trace) { tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); }
3870 #endif
3871 }
3872 }
3873
3874 // Replace existing phi with computed memory state for that region
3875 // if different (could be a new phi or a dominating memory node if
3876 // that phi was found to be useless).
3877 while (dead_phis.size() > 0) {
3878 Node* n = dead_phis.pop();
3879 n->replace_by(_phase->C->top());
3880 n->destruct();
3881 }
3882 for (int i = rpo_list.size() - 1; i >= 0; i--) {
3883 Node* c = rpo_list.at(i);
3884 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
3885 Node* n = regions[c->_idx];
3886 if (n->is_Phi() && n->_idx >= last && n->in(0) == c) {
3887 _phase->register_new_node(n, c);
3888 }
3889 }
3890 }
3891 for (int i = rpo_list.size() - 1; i >= 0; i--) {
3892 Node* c = rpo_list.at(i);
3893 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
3894 Node* n = regions[c->_idx];
3895 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
3896 Node* u = c->fast_out(i);
3897 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
3898 u != n) {
3899 if (u->adr_type() == TypePtr::BOTTOM) {
3900 fix_memory_uses(u, n, n, c);
3901 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
3902 _phase->lazy_replace(u, n);
3903 --i; --imax;
3904 }
3905 }
3906 }
3907 }
3908 }
3909 }
3910
3911 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
3912 Node* c = _phase->get_ctrl(n);
3913 if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Call()) {
3914 assert(c == n->in(0), "");
3915 CallNode* call = c->as_Call();
3916 CallProjections projs;
3917 call->extract_projections(&projs, true, false);
3918 if (projs.catchall_memproj != NULL) {
3919 if (projs.fallthrough_memproj == n) {
3920 c = projs.fallthrough_catchproj;
3921 } else {
3922 assert(projs.catchall_memproj == n, "");
3923 c = projs.catchall_catchproj;
3924 }
3925 }
3926 }
3927 return c;
3928 }
3929
3930 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
3931 if (_phase->has_ctrl(n))
3932 return get_ctrl(n);
3933 else {
3934 assert (n->is_CFG(), "must be a CFG node");
3935 return n;
3936 }
3937 }
3938
3939 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
3940 return m != NULL && get_ctrl(m) == c;
3941 }
3942
3943 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
3944 assert(n == NULL || _phase->ctrl_or_self(n) == ctrl, "");
3945 Node* mem = _memory_nodes[ctrl->_idx];
3946 Node* c = ctrl;
3947 while (!mem_is_valid(mem, c) &&
3948 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
3949 c = _phase->idom(c);
3950 mem = _memory_nodes[c->_idx];
3951 }
3952 if (n != NULL && mem_is_valid(mem, c)) {
3953 while (!ShenandoahWriteBarrierNode::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
3954 mem = next_mem(mem, _alias);
3955 }
3956 if (mem->is_MergeMem()) {
3957 mem = mem->as_MergeMem()->memory_at(_alias);
3958 }
3959 if (!mem_is_valid(mem, c)) {
3960 do {
3961 c = _phase->idom(c);
3962 mem = _memory_nodes[c->_idx];
3963 } while (!mem_is_valid(mem, c) &&
3964 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
3965 }
3966 }
3967 assert(mem->bottom_type() == Type::MEMORY, "");
3968 return mem;
3969 }
3970
3971 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
3972 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
3973 Node* use = region->fast_out(i);
3974 if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
3975 (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
3976 return true;
3977 }
3978 }
3979 return false;
3980 }
3981
3982 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
3983 assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
3984 const bool trace = false;
3985 DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
3986 DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
3987 GrowableArray<Node*> phis;
3988 if (mem_for_ctrl != mem) {
3989 Node* old = mem_for_ctrl;
3990 Node* prev = NULL;
3991 while (old != mem) {
3992 prev = old;
3993 if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
3994 assert(_alias == Compile::AliasIdxRaw, "");
3995 old = old->in(MemNode::Memory);
3996 } else if (old->Opcode() == Op_SCMemProj) {
3997 assert(_alias == Compile::AliasIdxRaw, "");
3998 old = old->in(0);
3999 } else if (old->Opcode() == Op_ShenandoahWBMemProj) {
4000 assert(_alias != Compile::AliasIdxRaw, "");
4001 old = old->in(0);
4002 } else if (old->Opcode() == Op_ShenandoahWriteBarrier) {
4003 assert(_alias != Compile::AliasIdxRaw, "");
4004 old = old->in(ShenandoahBarrierNode::Memory);
4005 } else {
4006 ShouldNotReachHere();
4007 }
4008 }
4009 assert(prev != NULL, "");
4010 if (new_ctrl != ctrl) {
4011 _memory_nodes.map(ctrl->_idx, mem);
4012 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
4013 }
4014 uint input = prev->Opcode() == Op_ShenandoahWriteBarrier ? (uint)ShenandoahBarrierNode::Memory : (uint)MemNode::Memory;
4015 _phase->igvn().replace_input_of(prev, input, new_mem);
4016 } else {
4017 uses.clear();
4018 _memory_nodes.map(new_ctrl->_idx, new_mem);
4019 uses.push(new_ctrl);
4020 for(uint next = 0; next < uses.size(); next++ ) {
4021 Node *n = uses.at(next);
4022 assert(n->is_CFG(), "");
4023 DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
4024 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
4025 Node* u = n->fast_out(i);
4026 if (!u->is_Root() && u->is_CFG() && u != n) {
4027 Node* m = _memory_nodes[u->_idx];
4028 if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
4029 !has_mem_phi(u) &&
4030 u->unique_ctrl_out()->Opcode() != Op_Halt) {
4031 DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
4032 DEBUG_ONLY(if (trace && m != NULL) { tty->print("ZZZ mem"); m->dump(); });
4033
4034 if (!mem_is_valid(m, u) || !m->is_Phi()) {
4035 bool push = true;
4036 bool create_phi = true;
4037 if (_phase->is_dominator(new_ctrl, u)) {
4038 create_phi = false;
4039 } else if (!_phase->C->has_irreducible_loop()) {
4040 IdealLoopTree* loop = _phase->get_loop(ctrl);
4041 bool do_check = true;
4042 IdealLoopTree* l = loop;
4043 create_phi = false;
4044 while (l != _phase->ltree_root()) {
4045 if (_phase->is_dominator(l->_head, u) && _phase->is_dominator(_phase->idom(u), l->_head)) {
4046 create_phi = true;
4047 do_check = false;
4048 break;
4049 }
4050 l = l->_parent;
4051 }
4052
4053 if (do_check) {
4054 assert(!create_phi, "");
4055 IdealLoopTree* u_loop = _phase->get_loop(u);
4056 if (u_loop != _phase->ltree_root() && u_loop->is_member(loop)) {
4057 Node* c = ctrl;
4058 while (!_phase->is_dominator(c, u_loop->tail())) {
4059 c = _phase->idom(c);
4060 }
4061 if (!_phase->is_dominator(c, u)) {
4062 do_check = false;
4063 }
4064 }
4065 }
4066
4067 if (do_check && _phase->is_dominator(_phase->idom(u), new_ctrl)) {
4068 create_phi = true;
4069 }
4070 }
4071 if (create_phi) {
4072 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
4073 _phase->register_new_node(phi, u);
4074 phis.push(phi);
4075 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
4076 if (!mem_is_valid(m, u)) {
4077 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
4078 _memory_nodes.map(u->_idx, phi);
4079 } else {
4080 DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
4081 for (;;) {
4082 assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj() || m->Opcode() == Op_ShenandoahWriteBarrier /*|| m->is_MergeMem()*/, "");
4083 Node* next = NULL;
4084 if (m->is_Proj()) {
4085 next = m->in(0);
4086 } else if (m->is_Mem() || m->is_LoadStore()) {
4087 assert(_alias == Compile::AliasIdxRaw, "");
4088 next = m->in(MemNode::Memory);
4089 } else {
4090 assert(_alias != Compile::AliasIdxRaw, "");
4091 assert (m->Opcode() == Op_ShenandoahWriteBarrier, "");
4092 next = m->in(ShenandoahBarrierNode::Memory);
4093 }
4094 if (_phase->get_ctrl(next) != u) {
4095 break;
4096 }
4097 if (next->is_MergeMem()) {
4098 assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
4099 break;
4100 }
4101 if (next->is_Phi()) {
4102 assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
4103 break;
4104 }
4105 m = next;
4106 }
4107
4108 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
4109 assert(m->is_Mem() || m->is_LoadStore() || m->Opcode() == Op_ShenandoahWriteBarrier, "");
4110 uint input = (m->is_Mem() || m->is_LoadStore()) ? (uint)MemNode::Memory : (uint)ShenandoahBarrierNode::Memory;
4111 _phase->igvn().replace_input_of(m, input, phi);
4112 push = false;
4113 }
4114 } else {
4115 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
4116 }
4117 if (push) {
4118 uses.push(u);
4119 }
4120 }
4121 } else if (!mem_is_valid(m, u) &&
4122 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1)) {
4123 uses.push(u);
4124 }
4125 }
4126 }
4127 }
4128 for (int i = 0; i < phis.length(); i++) {
4129 Node* n = phis.at(i);
4130 Node* r = n->in(0);
4131 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
4132 for (uint j = 1; j < n->req(); j++) {
4133 Node* m = find_mem(r->in(j), NULL);
4134 _phase->igvn().replace_input_of(n, j, m);
4135 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
4136 }
4137 }
4138 }
4139 uint last = _phase->C->unique();
4140 MergeMemNode* mm = NULL;
4141 int alias = _alias;
4142 DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
4143 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
4144 Node* u = mem->out(i);
4145 if (u->_idx < last) {
4146 if (u->is_Mem()) {
4147 if (_phase->C->get_alias_index(u->adr_type()) == alias) {
4148 Node* m = find_mem(_phase->get_ctrl(u), u);
4149 if (m != mem) {
4150 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
4151 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
4152 --i;
4153 }
4154 }
4155 } else if (u->is_MergeMem()) {
4156 MergeMemNode* u_mm = u->as_MergeMem();
4157 if (u_mm->memory_at(alias) == mem) {
4158 MergeMemNode* newmm = NULL;
4159 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
4160 Node* uu = u->fast_out(j);
4161 assert(!uu->is_MergeMem(), "chain of MergeMems?");
4162 if (uu->is_Phi()) {
4163 assert(uu->adr_type() == TypePtr::BOTTOM, "");
4164 Node* region = uu->in(0);
4165 int nb = 0;
4166 for (uint k = 1; k < uu->req(); k++) {
4167 if (uu->in(k) == u) {
4168 Node* m = find_mem(region->in(k), NULL);
4169 if (m != mem) {
4170 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
4171 if (newmm == NULL || 1) {
4172 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
4173 }
4174 if (newmm != u) {
4175 _phase->igvn().replace_input_of(uu, k, newmm);
4176 nb++;
4177 --jmax;
4178 }
4179 }
4180 }
4181 }
4182 if (nb > 0) {
4183 --j;
4184 }
4185 } else {
4186 Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
4187 if (m != mem) {
4188 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
4189 if (newmm == NULL || 1) {
4190 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
4191 }
4192 if (newmm != u) {
4193 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
4194 --j, --jmax;
4195 }
4196 }
4197 }
4198 }
4199 }
4200 } else if (u->is_Phi()) {
4201 assert(u->bottom_type() == Type::MEMORY, "what else?");
4202 if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
4203 Node* region = u->in(0);
4204 bool replaced = false;
4205 for (uint j = 1; j < u->req(); j++) {
4206 if (u->in(j) == mem) {
4207 Node* m = find_mem(region->in(j), NULL);
4208 Node* nnew = m;
4209 if (m != mem) {
4210 if (u->adr_type() == TypePtr::BOTTOM) {
4211 mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
4212 nnew = mm;
4213 }
4214 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
4215 _phase->igvn().replace_input_of(u, j, nnew);
4216 replaced = true;
4217 }
4218 }
4219 }
4220 if (replaced) {
4221 --i;
4222 }
4223 }
4224 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
4225 u->adr_type() == NULL) {
4226 assert(u->adr_type() != NULL ||
4227 u->Opcode() == Op_Rethrow ||
4228 u->Opcode() == Op_Return ||
4229 u->Opcode() == Op_SafePoint ||
4230 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
4231 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
4232 u->Opcode() == Op_CallLeaf, "");
4233 Node* m = find_mem(_phase->ctrl_or_self(u), u);
4234 if (m != mem) {
4235 mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
4236 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
4237 --i;
4238 }
4239 } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
4240 Node* m = find_mem(_phase->ctrl_or_self(u), u);
4241 if (m != mem) {
4242 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
4243 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
4244 --i;
4245 }
4246 } else if (u->adr_type() != TypePtr::BOTTOM &&
4247 _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
4248 Node* m = find_mem(_phase->ctrl_or_self(u), u);
4249 assert(m != mem, "");
4250 // u is on the wrong slice...
4251 assert(u->is_ClearArray(), "");
4252 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
4253 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
4254 --i;
4255 }
4256 }
4257 }
4258 #ifdef ASSERT
4259 assert(new_mem->outcnt() > 0, "");
4260 for (int i = 0; i < phis.length(); i++) {
4261 Node* n = phis.at(i);
4262 assert(n->outcnt() > 0, "new phi must have uses now");
4263 }
4264 #endif
4265 }
4266
4267 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
4268 MergeMemNode* mm = MergeMemNode::make(mem);
4269 mm->set_memory_at(_alias, rep_proj);
4270 _phase->register_new_node(mm, rep_ctrl);
4271 return mm;
4272 }
4273
4274 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
4275 MergeMemNode* newmm = NULL;
4276 MergeMemNode* u_mm = u->as_MergeMem();
4277 Node* c = _phase->get_ctrl(u);
4278 if (_phase->is_dominator(c, rep_ctrl)) {
4279 c = rep_ctrl;
4280 } else {
4281 assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
4282 }
4283 if (u->outcnt() == 1) {
4284 if (u->req() > (uint)_alias && u->in(_alias) == mem) {
4285 _phase->igvn().replace_input_of(u, _alias, rep_proj);
4286 --i;
4287 } else {
4288 _phase->igvn().rehash_node_delayed(u);
4289 u_mm->set_memory_at(_alias, rep_proj);
4290 }
4291 newmm = u_mm;
4292 _phase->set_ctrl_and_loop(u, c);
4293 } else {
4294 // can't simply clone u and then change one of its input because
4295 // it adds and then removes an edge which messes with the
4296 // DUIterator
4297 newmm = MergeMemNode::make(u_mm->base_memory());
4298 for (uint j = 0; j < u->req(); j++) {
4299 if (j < newmm->req()) {
4300 if (j == (uint)_alias) {
4301 newmm->set_req(j, rep_proj);
4302 } else if (newmm->in(j) != u->in(j)) {
4303 newmm->set_req(j, u->in(j));
4304 }
4305 } else if (j == (uint)_alias) {
4306 newmm->add_req(rep_proj);
4307 } else {
4308 newmm->add_req(u->in(j));
4309 }
4310 }
4311 if ((uint)_alias >= u->req()) {
4312 newmm->set_memory_at(_alias, rep_proj);
4313 }
4314 _phase->register_new_node(newmm, c);
4315 }
4316 return newmm;
4317 }
4318
4319 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
4320 if (phi->adr_type() == TypePtr::BOTTOM) {
4321 Node* region = phi->in(0);
4322 for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
4323 Node* uu = region->fast_out(j);
4324 if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
4325 return false;
4326 }
4327 }
4328 return true;
4329 }
4330 return _phase->C->get_alias_index(phi->adr_type()) == _alias;
4331 }
4332
4333
4334 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
4335 uint last = _phase-> C->unique();
4336 MergeMemNode* mm = NULL;
4337 assert(mem->bottom_type() == Type::MEMORY, "");
4338 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
4339 Node* u = mem->out(i);
4340 if (u != replacement && u->_idx < last) {
4341 if (u->is_ShenandoahBarrier() && _alias != Compile::AliasIdxRaw) {
4342 if (_phase->C->get_alias_index(u->adr_type()) == _alias && ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
4343 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
4344 assert(u->find_edge(mem) == -1, "only one edge");
4345 --i;
4346 }
4347 } else if (u->is_Mem()) {
4348 if (_phase->C->get_alias_index(u->adr_type()) == _alias && ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
4349 assert(_alias == Compile::AliasIdxRaw , "only raw memory can lead to a memory operation");
4350 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
4351 assert(u->find_edge(mem) == -1, "only one edge");
4352 --i;
4353 }
4354 } else if (u->is_MergeMem()) {
4355 MergeMemNode* u_mm = u->as_MergeMem();
4356 if (u_mm->memory_at(_alias) == mem) {
4357 MergeMemNode* newmm = NULL;
4358 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
4359 Node* uu = u->fast_out(j);
4360 assert(!uu->is_MergeMem(), "chain of MergeMems?");
4361 if (uu->is_Phi()) {
4362 if (should_process_phi(uu)) {
4363 Node* region = uu->in(0);
4364 int nb = 0;
4365 for (uint k = 1; k < uu->req(); k++) {
4366 if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
4367 if (newmm == NULL) {
4368 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
4369 }
4370 if (newmm != u) {
4371 _phase->igvn().replace_input_of(uu, k, newmm);
4372 nb++;
4373 --jmax;
4374 }
4375 }
4376 }
4377 if (nb > 0) {
4378 --j;
4379 }
4380 }
4381 } else {
4382 if (rep_ctrl != uu && ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
4383 if (newmm == NULL) {
4384 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
4385 }
4386 if (newmm != u) {
4387 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
4388 --j, --jmax;
4389 }
4390 }
4391 }
4392 }
4393 }
4394 } else if (u->is_Phi()) {
4395 assert(u->bottom_type() == Type::MEMORY, "what else?");
4396 Node* region = u->in(0);
4397 if (should_process_phi(u)) {
4398 bool replaced = false;
4399 for (uint j = 1; j < u->req(); j++) {
4400 if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
4401 Node* nnew = rep_proj;
4402 if (u->adr_type() == TypePtr::BOTTOM) {
4403 if (mm == NULL) {
4404 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
4405 }
4406 nnew = mm;
4407 }
4408 _phase->igvn().replace_input_of(u, j, nnew);
4409 replaced = true;
4410 }
4411 }
4412 if (replaced) {
4413 --i;
4414 }
4415
4416 }
4417 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
4418 u->adr_type() == NULL) {
4419 assert(u->adr_type() != NULL ||
4420 u->Opcode() == Op_Rethrow ||
4421 u->Opcode() == Op_Return ||
4422 u->Opcode() == Op_SafePoint ||
4423 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
4424 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
4425 u->Opcode() == Op_CallLeaf, "");
4426 if (ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
4427 if (mm == NULL) {
4428 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
4429 }
4430 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
4431 --i;
4432 }
4433 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
4434 if (ShenandoahWriteBarrierNode::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
4435 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
4436 --i;
4437 }
4438 }
4439 }
4440 }
4441 }
4442
4443 void MemoryGraphFixer::remove(Node* n) {
4444 assert(n->Opcode() == Op_ShenandoahWBMemProj, "");
4445 Node* c = _phase->get_ctrl(n);
4446 Node* mem = find_mem(c, NULL);
4447 if (mem == n) {
4448 _memory_nodes.map(c->_idx, mem->in(0)->in(ShenandoahBarrierNode::Memory));
4449 }
4450 }
4451
4452 static bool is_on_null_check_path(Block* b, Block* null_check_block) {
4453 if (null_check_block == NULL) {
4454 return false;
4455 }
4456 do {
4457 assert(null_check_block->_num_succs == 1, "only one succ on the path to unc");
4458 if (b == null_check_block) {
4459 return true;
4460 }
4461 null_check_block = null_check_block->_succs[0];
4462 } while(!null_check_block->head()->is_Root());
4463
4464 return false;
4465 }
4466
4467 int PhaseCFG::replace_uses_with_shenandoah_barrier_helper(Node* n, Node* use, Node* val, Block* block, Block* null_check_block) {
4468 int nb = 0;
4469 Block* buse = get_block_for_node(use);
4470 if (is_on_null_check_path(buse, null_check_block)) {
4471 return 0;
4472 }
4473 if (use->is_Phi()) {
4474 for (uint j = 1; j < use->req(); j++) {
4475 if (use->in(j) == val) {
4476 Block* b = get_block_for_node(use->in(0)->in(j));
4477 if ((block != b && block->dom_lca(b) == block) ||
4478 block == b) {
4479 use->set_req(j, n);
4480 nb++;
4481 }
4482 }
4483 }
4484 } else {
4485 if ((block != buse && block->dom_lca(buse) == block) ||
4486 (block == buse && !use->is_scheduled())) {
4487 // Let precedence edges alone (can confuse anti-dependence verification code)
4488 for (uint i = 0; i < use->req(); i++) {
4489 if (use->in(i) == val) {
4490 use->set_req(i, n);
4491 nb++;
4492 }
4493 }
4494 assert(nb > 0 || use->find_prec_edge(val) != -1, "no replacement?");
4495 }
4496 }
4497
4498 return nb;
4499 }
4500
4501 void PhaseCFG::replace_uses_with_shenandoah_barrier(Node* n, Block* block, Node_List& worklist, GrowableArray<int>& ready_cnt, uint max_idx, uint& phi_cnt) {
4502 // Replace all uses of barrier's input that are dominated by the
4503 // barrier with the value returned by the barrier: no need to keep
4504 // both live.
4505 if (n->is_Mach() && n->as_Mach()->ideal_Opcode() == Op_ShenandoahReadBarrier) {
4506 MachNullCheckNode* null_check = NULL;
4507 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && null_check == NULL; i++) {
4508 Node* use = n->fast_out(i);
4509 if (use->is_MachNullCheck()) {
4510 null_check = use->as_MachNullCheck();
4511 }
4512 }
4513 Block* null_check_block = NULL;
4514 if (null_check != NULL) {
4515 Node* proj = null_check->find_out_with(Op_IfTrue);
4516 Node* head = proj->unique_out();
4517 null_check_block = get_block_for_node(head);
4518 }
4519
4520 Node* val = n->in(ShenandoahBarrierNode::ValueIn);
4521 if (!val->bottom_type()->isa_narrowoop()) {
4522 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
4523 Node* use = val->fast_out(i);
4524 if (use != n) {
4525 int nb = replace_uses_with_shenandoah_barrier_helper(n, use, val, block, null_check_block);
4526 if (nb > 0) {
4527 --i; imax -= nb;
4528 }
4529 }
4530 }
4531 } else {
4532 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
4533 Node* u = val->fast_out(i);
4534 if (u->is_Mach() && u->as_Mach()->ideal_Opcode() == Op_DecodeN) {
4535 int projs = 0;
4536 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
4537 Node* uu = u->fast_out(j);
4538 assert(!uu->is_MachTemp(), "");
4539 if (uu->is_MachProj() && uu->outcnt() == 0) {
4540 projs++;
4541 } else {
4542 int nb = replace_uses_with_shenandoah_barrier_helper(n, uu, u, block, null_check_block);
4543 if (nb > 0) {
4544 if (!u->is_scheduled()) {
4545 push_ready_nodes(n, uu, block, ready_cnt, worklist, max_idx, nb);
4546 }
4547 --j; jmax -= nb;
4548 }
4549 }
4550 }
4551 // The DecodeN may have gone dead
4552 if (u->outcnt() - projs == 0) {
4553 u->disconnect_inputs(NULL, C);
4554 Block* bu = get_block_for_node(u);
4555 unmap_node_from_block(u);
4556 if (bu == block) {
4557 if (u->is_scheduled()) {
4558 block->find_remove(u);
4559 phi_cnt--;
4560 } else {
4561 worklist.yank(u);
4562 block->remove_node(block->end_idx()-1);
4563 }
4564 } else {
4565 bu->find_remove(u);
4566 }
4567 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
4568 Node* uu = u->fast_out(j);
4569 assert(uu->is_MachProj() && uu->outcnt() == 0, "");
4570 assert(bu == get_block_for_node(uu), "");
4571 uu->disconnect_inputs(NULL, C);
4572 --j; --jmax;
4573 unmap_node_from_block(uu);
4574 if (bu == block) {
4575 if (u->is_scheduled()) {
4576 block->find_remove(uu);
4577 phi_cnt--;
4578 } else {
4579 worklist.yank(uu);
4580 block->remove_node(block->end_idx()-1);
4581 }
4582 } else {
4583 bu->find_remove(uu);
4584 }
4585 assert(uu->is_scheduled() == u->is_scheduled(), "");
4586 }
4587 --i; --imax;
4588 }
4589 }
4590 }
4591 }
4592 }
4593 }