1 /*
2 * Copyright (c) 2015, 2019, Red Hat, Inc. All rights reserved.
3 *
4 * This code is free software; you can redistribute it and/or modify it
5 * under the terms of the GNU General Public License version 2 only, as
6 * published by the Free Software Foundation.
7 *
8 * This code is distributed in the hope that it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
11 * version 2 for more details (a copy is included in the LICENSE file that
12 * accompanied this code).
13 *
14 * You should have received a copy of the GNU General Public License version
15 * 2 along with this work; if not, write to the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
17 *
18 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
19 * or visit www.oracle.com if you need additional information or have any
20 * questions.
21 *
22 */
23
24 #include "precompiled.hpp"
25
26 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
27 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
28 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
29 #include "gc/shenandoah/shenandoahHeap.hpp"
30 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
31 #include "gc/shenandoah/shenandoahRuntime.hpp"
32 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/block.hpp"
35 #include "opto/callnode.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/movenode.hpp"
38 #include "opto/phaseX.hpp"
39 #include "opto/rootnode.hpp"
40 #include "opto/runtime.hpp"
41 #include "opto/subnode.hpp"
42
43 bool ShenandoahBarrierC2Support::expand(Compile* C, PhaseIterGVN& igvn) {
44 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
45 if ((state->enqueue_barriers_count() +
46 state->load_reference_barriers_count()) > 0) {
47 bool attempt_more_loopopts = ShenandoahLoopOptsAfterExpansion;
48 C->clear_major_progress();
49 PhaseIdealLoop ideal_loop(igvn, LoopOptsShenandoahExpand);
50 if (C->failing()) return false;
51 PhaseIdealLoop::verify(igvn);
52 DEBUG_ONLY(verify_raw_mem(C->root());)
53 if (attempt_more_loopopts) {
54 C->set_major_progress();
55 if (!C->optimize_loops(igvn, LoopOptsShenandoahPostExpand)) {
56 return false;
57 }
58 C->clear_major_progress();
59 }
60 }
61 return true;
62 }
63
64 bool ShenandoahBarrierC2Support::is_heap_state_test(Node* iff, int mask) {
65 if (!UseShenandoahGC) {
66 return false;
67 }
68 assert(iff->is_If(), "bad input");
69 if (iff->Opcode() != Op_If) {
70 return false;
71 }
72 Node* bol = iff->in(1);
73 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
74 return false;
75 }
76 Node* cmp = bol->in(1);
77 if (cmp->Opcode() != Op_CmpI) {
78 return false;
79 }
80 Node* in1 = cmp->in(1);
81 Node* in2 = cmp->in(2);
82 if (in2->find_int_con(-1) != 0) {
83 return false;
84 }
85 if (in1->Opcode() != Op_AndI) {
86 return false;
87 }
88 in2 = in1->in(2);
89 if (in2->find_int_con(-1) != mask) {
90 return false;
91 }
92 in1 = in1->in(1);
93
94 return is_gc_state_load(in1);
95 }
96
97 bool ShenandoahBarrierC2Support::is_heap_stable_test(Node* iff) {
98 return is_heap_state_test(iff, ShenandoahHeap::HAS_FORWARDED);
99 }
100
101 bool ShenandoahBarrierC2Support::is_gc_state_load(Node *n) {
102 if (!UseShenandoahGC) {
103 return false;
104 }
105 if (n->Opcode() != Op_LoadB && n->Opcode() != Op_LoadUB) {
106 return false;
107 }
108 Node* addp = n->in(MemNode::Address);
109 if (!addp->is_AddP()) {
110 return false;
111 }
112 Node* base = addp->in(AddPNode::Address);
113 Node* off = addp->in(AddPNode::Offset);
114 if (base->Opcode() != Op_ThreadLocal) {
115 return false;
116 }
117 if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) {
118 return false;
119 }
120 return true;
121 }
122
123 bool ShenandoahBarrierC2Support::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) {
124 assert(phase->is_dominator(stop, start), "bad inputs");
125 ResourceMark rm;
126 Unique_Node_List wq;
127 wq.push(start);
128 for (uint next = 0; next < wq.size(); next++) {
129 Node *m = wq.at(next);
130 if (m == stop) {
131 continue;
132 }
133 if (m->is_SafePoint() && !m->is_CallLeaf()) {
134 return true;
135 }
136 if (m->is_Region()) {
137 for (uint i = 1; i < m->req(); i++) {
138 wq.push(m->in(i));
139 }
140 } else {
141 wq.push(m->in(0));
142 }
143 }
144 return false;
145 }
146
147 bool ShenandoahBarrierC2Support::try_common_gc_state_load(Node *n, PhaseIdealLoop *phase) {
148 assert(is_gc_state_load(n), "inconsistent");
149 Node* addp = n->in(MemNode::Address);
150 Node* dominator = NULL;
151 for (DUIterator_Fast imax, i = addp->fast_outs(imax); i < imax; i++) {
152 Node* u = addp->fast_out(i);
153 assert(is_gc_state_load(u), "inconsistent");
154 if (u != n && phase->is_dominator(u->in(0), n->in(0))) {
155 if (dominator == NULL) {
156 dominator = u;
157 } else {
158 if (phase->dom_depth(u->in(0)) < phase->dom_depth(dominator->in(0))) {
159 dominator = u;
160 }
161 }
162 }
163 }
164 if (dominator == NULL || has_safepoint_between(n->in(0), dominator->in(0), phase)) {
165 return false;
166 }
167 phase->igvn().replace_node(n, dominator);
168
169 return true;
170 }
171
172 #ifdef ASSERT
173 bool ShenandoahBarrierC2Support::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) {
174 assert(phis.size() == 0, "");
175
176 while (true) {
177 if (in->bottom_type() == TypePtr::NULL_PTR) {
178 if (trace) {tty->print_cr("NULL");}
179 } else if (!in->bottom_type()->make_ptr()->make_oopptr()) {
180 if (trace) {tty->print_cr("Non oop");}
181 } else if (t == ShenandoahLoad && ShenandoahOptimizeStableFinals &&
182 in->bottom_type()->make_ptr()->isa_aryptr() &&
183 in->bottom_type()->make_ptr()->is_aryptr()->is_stable()) {
184 if (trace) {tty->print_cr("Stable array load");}
185 } else {
186 if (in->is_ConstraintCast()) {
187 in = in->in(1);
188 continue;
189 } else if (in->is_AddP()) {
190 assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access");
191 in = in->in(AddPNode::Address);
192 continue;
193 } else if (in->is_Con()) {
194 if (trace) {
195 tty->print("Found constant");
196 in->dump();
197 }
198 } else if (in->Opcode() == Op_Parm) {
199 if (trace) {
200 tty->print("Found argument");
201 }
202 } else if (in->Opcode() == Op_CreateEx) {
203 if (trace) {
204 tty->print("Found create-exception");
205 }
206 } else if (in->Opcode() == Op_LoadP && in->adr_type() == TypeRawPtr::BOTTOM) {
207 if (trace) {
208 tty->print("Found raw LoadP (OSR argument?)");
209 }
210 } else if (in->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
211 if (t == ShenandoahOopStore) {
212 uint i = 0;
213 for (; i < phis.size(); i++) {
214 Node* n = phis.node_at(i);
215 if (n->Opcode() == Op_ShenandoahEnqueueBarrier) {
216 break;
217 }
218 }
219 if (i == phis.size()) {
220 return false;
221 }
222 }
223 barriers_used.push(in);
224 if (trace) {tty->print("Found barrier"); in->dump();}
225 } else if (in->Opcode() == Op_ShenandoahEnqueueBarrier) {
226 if (t != ShenandoahOopStore) {
227 in = in->in(1);
228 continue;
229 }
230 if (trace) {tty->print("Found enqueue barrier"); in->dump();}
231 phis.push(in, in->req());
232 in = in->in(1);
233 continue;
234 } else if (in->is_Proj() && in->in(0)->is_Allocate()) {
235 if (trace) {
236 tty->print("Found alloc");
237 in->in(0)->dump();
238 }
239 } else if (in->is_Proj() && (in->in(0)->Opcode() == Op_CallStaticJava || in->in(0)->Opcode() == Op_CallDynamicJava)) {
240 if (trace) {
241 tty->print("Found Java call");
242 }
243 } else if (in->is_Phi()) {
244 if (!visited.test_set(in->_idx)) {
245 if (trace) {tty->print("Pushed phi:"); in->dump();}
246 phis.push(in, 2);
247 in = in->in(1);
248 continue;
249 }
250 if (trace) {tty->print("Already seen phi:"); in->dump();}
251 } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) {
252 if (!visited.test_set(in->_idx)) {
253 if (trace) {tty->print("Pushed cmovep:"); in->dump();}
254 phis.push(in, CMoveNode::IfTrue);
255 in = in->in(CMoveNode::IfFalse);
256 continue;
257 }
258 if (trace) {tty->print("Already seen cmovep:"); in->dump();}
259 } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) {
260 in = in->in(1);
261 continue;
262 } else {
263 return false;
264 }
265 }
266 bool cont = false;
267 while (phis.is_nonempty()) {
268 uint idx = phis.index();
269 Node* phi = phis.node();
270 if (idx >= phi->req()) {
271 if (trace) {tty->print("Popped phi:"); phi->dump();}
272 phis.pop();
273 continue;
274 }
275 if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();}
276 in = phi->in(idx);
277 phis.set_index(idx+1);
278 cont = true;
279 break;
280 }
281 if (!cont) {
282 break;
283 }
284 }
285 return true;
286 }
287
288 void ShenandoahBarrierC2Support::report_verify_failure(const char* msg, Node* n1, Node* n2) {
289 if (n1 != NULL) {
290 n1->dump(+10);
291 }
292 if (n2 != NULL) {
293 n2->dump(+10);
294 }
295 fatal("%s", msg);
296 }
297
298 void ShenandoahBarrierC2Support::verify(RootNode* root) {
299 ResourceMark rm;
300 Unique_Node_List wq;
301 GrowableArray<Node*> barriers;
302 Unique_Node_List barriers_used;
303 Node_Stack phis(0);
304 VectorSet visited(Thread::current()->resource_area());
305 const bool trace = false;
306 const bool verify_no_useless_barrier = false;
307
308 wq.push(root);
309 for (uint next = 0; next < wq.size(); next++) {
310 Node *n = wq.at(next);
311 if (n->is_Load()) {
312 const bool trace = false;
313 if (trace) {tty->print("Verifying"); n->dump();}
314 if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) {
315 if (trace) {tty->print_cr("Load range/klass");}
316 } else {
317 const TypePtr* adr_type = n->as_Load()->adr_type();
318
319 if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) {
320 if (trace) {tty->print_cr("Mark load");}
321 } else if (adr_type->isa_instptr() &&
322 adr_type->is_instptr()->klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
323 adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset) {
324 if (trace) {tty->print_cr("Reference.get()");}
325 } else {
326 bool verify = true;
327 if (adr_type->isa_instptr()) {
328 const TypeInstPtr* tinst = adr_type->is_instptr();
329 ciKlass* k = tinst->klass();
330 assert(k->is_instance_klass(), "");
331 ciInstanceKlass* ik = (ciInstanceKlass*)k;
332 int offset = adr_type->offset();
333
334 if ((ik->debug_final_field_at(offset) && ShenandoahOptimizeInstanceFinals) ||
335 (ik->debug_stable_field_at(offset) && ShenandoahOptimizeStableFinals)) {
336 if (trace) {tty->print_cr("Final/stable");}
337 verify = false;
338 } else if (k == ciEnv::current()->Class_klass() &&
339 tinst->const_oop() != NULL &&
340 tinst->offset() >= (ik->size_helper() * wordSize)) {
341 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
342 ciField* field = k->get_field_by_offset(tinst->offset(), true);
343 if ((ShenandoahOptimizeStaticFinals && field->is_final()) ||
344 (ShenandoahOptimizeStableFinals && field->is_stable())) {
345 verify = false;
346 }
347 }
348 }
349
350 if (verify && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) {
351 report_verify_failure("Shenandoah verification: Load should have barriers", n);
352 }
353 }
354 }
355 } else if (n->is_Store()) {
356 const bool trace = false;
357
358 if (trace) {tty->print("Verifying"); n->dump();}
359 if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) {
360 Node* adr = n->in(MemNode::Address);
361 bool verify = true;
362
363 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) {
364 adr = adr->in(AddPNode::Address);
365 if (adr->is_AddP()) {
366 assert(adr->in(AddPNode::Base)->is_top(), "");
367 adr = adr->in(AddPNode::Address);
368 if (adr->Opcode() == Op_LoadP &&
369 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() &&
370 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
371 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) {
372 if (trace) {tty->print_cr("SATB prebarrier");}
373 verify = false;
374 }
375 }
376 }
377
378 if (verify && !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahStoreValEnqueueBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
379 report_verify_failure("Shenandoah verification: Store should have barriers", n);
380 }
381 }
382 if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
383 report_verify_failure("Shenandoah verification: Store (address) should have barriers", n);
384 }
385 } else if (n->Opcode() == Op_CmpP) {
386 const bool trace = false;
387
388 Node* in1 = n->in(1);
389 Node* in2 = n->in(2);
390 if (in1->bottom_type()->isa_oopptr()) {
391 if (trace) {tty->print("Verifying"); n->dump();}
392
393 bool mark_inputs = false;
394 if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR ||
395 (in1->is_Con() || in2->is_Con())) {
396 if (trace) {tty->print_cr("Comparison against a constant");}
397 mark_inputs = true;
398 } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) ||
399 (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) {
400 if (trace) {tty->print_cr("Comparison with newly alloc'ed object");}
401 mark_inputs = true;
402 } else {
403 assert(in2->bottom_type()->isa_oopptr(), "");
404
405 if (!verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) ||
406 !verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) {
407 report_verify_failure("Shenandoah verification: Cmp should have barriers", n);
408 }
409 }
410 if (verify_no_useless_barrier &&
411 mark_inputs &&
412 (!verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) ||
413 !verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) {
414 phis.clear();
415 visited.Reset();
416 }
417 }
418 } else if (n->is_LoadStore()) {
419 if (n->in(MemNode::ValueIn)->bottom_type()->make_ptr() &&
420 !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahStoreValEnqueueBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
421 report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n);
422 }
423
424 if (n->in(MemNode::Address)->bottom_type()->make_oopptr() && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
425 report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n);
426 }
427 } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) {
428 CallNode* call = n->as_Call();
429
430 static struct {
431 const char* name;
432 struct {
433 int pos;
434 verify_type t;
435 } args[6];
436 } calls[] = {
437 "aescrypt_encryptBlock",
438 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
439 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
440 "aescrypt_decryptBlock",
441 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
442 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
443 "multiplyToLen",
444 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { TypeFunc::Parms+4, ShenandoahStore },
445 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
446 "squareToLen",
447 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { -1, ShenandoahNone},
448 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
449 "montgomery_multiply",
450 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
451 { TypeFunc::Parms+6, ShenandoahStore }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
452 "montgomery_square",
453 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+5, ShenandoahStore },
454 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
455 "mulAdd",
456 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
457 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
458 "vectorizedMismatch",
459 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
460 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
461 "updateBytesCRC32",
462 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
463 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
464 "updateBytesAdler32",
465 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
466 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
467 "updateBytesCRC32C",
468 { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad}, { -1, ShenandoahNone},
469 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
470 "counterMode_AESCrypt",
471 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
472 { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } },
473 "cipherBlockChaining_encryptAESCrypt",
474 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
475 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
476 "cipherBlockChaining_decryptAESCrypt",
477 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
478 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
479 "shenandoah_clone_barrier",
480 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
481 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
482 "ghash_processBlocks",
483 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
484 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
485 "sha1_implCompress",
486 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
487 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
488 "sha256_implCompress",
489 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
490 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
491 "sha512_implCompress",
492 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
493 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
494 "sha1_implCompressMB",
495 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
496 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
497 "sha256_implCompressMB",
498 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
499 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
500 "sha512_implCompressMB",
501 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
502 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
503 "encodeBlock",
504 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahStore }, { -1, ShenandoahNone },
505 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
506 };
507
508 if (call->is_call_to_arraycopystub()) {
509 Node* dest = NULL;
510 const TypeTuple* args = n->as_Call()->_tf->domain();
511 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
512 if (args->field_at(i)->isa_ptr()) {
513 j++;
514 if (j == 2) {
515 dest = n->in(i);
516 break;
517 }
518 }
519 }
520 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) ||
521 !verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) {
522 report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n);
523 }
524 } else if (strlen(call->_name) > 5 &&
525 !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) {
526 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) {
527 report_verify_failure("Shenandoah verification: _fill should have barriers", n);
528 }
529 } else if (!strcmp(call->_name, "shenandoah_wb_pre")) {
530 // skip
531 } else {
532 const int calls_len = sizeof(calls) / sizeof(calls[0]);
533 int i = 0;
534 for (; i < calls_len; i++) {
535 if (!strcmp(calls[i].name, call->_name)) {
536 break;
537 }
538 }
539 if (i != calls_len) {
540 const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]);
541 for (uint j = 0; j < args_len; j++) {
542 int pos = calls[i].args[j].pos;
543 if (pos == -1) {
544 break;
545 }
546 if (!verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) {
547 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
548 }
549 }
550 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
551 if (call->in(j)->bottom_type()->make_ptr() &&
552 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
553 uint k = 0;
554 for (; k < args_len && calls[i].args[k].pos != (int)j; k++);
555 if (k == args_len) {
556 fatal("arg %d for call %s not covered", j, call->_name);
557 }
558 }
559 }
560 } else {
561 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
562 if (call->in(j)->bottom_type()->make_ptr() &&
563 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
564 fatal("%s not covered", call->_name);
565 }
566 }
567 }
568 }
569 } else if (n->Opcode() == Op_ShenandoahEnqueueBarrier || n->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
570 // skip
571 } else if (n->is_AddP()
572 || n->is_Phi()
573 || n->is_ConstraintCast()
574 || n->Opcode() == Op_Return
575 || n->Opcode() == Op_CMoveP
576 || n->Opcode() == Op_CMoveN
577 || n->Opcode() == Op_Rethrow
578 || n->is_MemBar()
579 || n->Opcode() == Op_Conv2B
580 || n->Opcode() == Op_SafePoint
581 || n->is_CallJava()
582 || n->Opcode() == Op_Unlock
583 || n->Opcode() == Op_EncodeP
584 || n->Opcode() == Op_DecodeN) {
585 // nothing to do
586 } else {
587 static struct {
588 int opcode;
589 struct {
590 int pos;
591 verify_type t;
592 } inputs[2];
593 } others[] = {
594 Op_FastLock,
595 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
596 Op_Lock,
597 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone} },
598 Op_ArrayCopy,
599 { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } },
600 Op_StrCompressedCopy,
601 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
602 Op_StrInflatedCopy,
603 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
604 Op_AryEq,
605 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
606 Op_StrIndexOf,
607 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
608 Op_StrComp,
609 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
610 Op_StrEquals,
611 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
612 Op_EncodeISOArray,
613 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
614 Op_HasNegatives,
615 { { 2, ShenandoahLoad }, { -1, ShenandoahNone} },
616 Op_CastP2X,
617 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
618 Op_StrIndexOfChar,
619 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } },
620 };
621
622 const int others_len = sizeof(others) / sizeof(others[0]);
623 int i = 0;
624 for (; i < others_len; i++) {
625 if (others[i].opcode == n->Opcode()) {
626 break;
627 }
628 }
629 uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req();
630 if (i != others_len) {
631 const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]);
632 for (uint j = 0; j < inputs_len; j++) {
633 int pos = others[i].inputs[j].pos;
634 if (pos == -1) {
635 break;
636 }
637 if (!verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) {
638 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
639 }
640 }
641 for (uint j = 1; j < stop; j++) {
642 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
643 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
644 uint k = 0;
645 for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++);
646 if (k == inputs_len) {
647 fatal("arg %d for node %s not covered", j, n->Name());
648 }
649 }
650 }
651 } else {
652 for (uint j = 1; j < stop; j++) {
653 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
654 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
655 fatal("%s not covered", n->Name());
656 }
657 }
658 }
659 }
660
661 if (n->is_SafePoint()) {
662 SafePointNode* sfpt = n->as_SafePoint();
663 if (verify_no_useless_barrier && sfpt->jvms() != NULL) {
664 for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) {
665 if (!verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) {
666 phis.clear();
667 visited.Reset();
668 }
669 }
670 }
671 }
672 for( uint i = 0; i < n->len(); ++i ) {
673 Node *m = n->in(i);
674 if (m == NULL) continue;
675
676 // In most cases, inputs should be known to be non null. If it's
677 // not the case, it could be a missing cast_not_null() in an
678 // intrinsic or support might be needed in AddPNode::Ideal() to
679 // avoid a NULL+offset input.
680 if (!(n->is_Phi() ||
681 (n->is_SafePoint() && (!n->is_CallRuntime() || !strcmp(n->as_Call()->_name, "shenandoah_wb_pre") || !strcmp(n->as_Call()->_name, "unsafe_arraycopy"))) ||
682 n->Opcode() == Op_CmpP ||
683 n->Opcode() == Op_CmpN ||
684 (n->Opcode() == Op_StoreP && i == StoreNode::ValueIn) ||
685 (n->Opcode() == Op_StoreN && i == StoreNode::ValueIn) ||
686 n->is_ConstraintCast() ||
687 n->Opcode() == Op_Return ||
688 n->Opcode() == Op_Conv2B ||
689 n->is_AddP() ||
690 n->Opcode() == Op_CMoveP ||
691 n->Opcode() == Op_CMoveN ||
692 n->Opcode() == Op_Rethrow ||
693 n->is_MemBar() ||
694 n->is_Mem() ||
695 n->Opcode() == Op_AryEq ||
696 n->Opcode() == Op_SCMemProj ||
697 n->Opcode() == Op_EncodeP ||
698 n->Opcode() == Op_DecodeN ||
699 n->Opcode() == Op_ShenandoahEnqueueBarrier ||
700 n->Opcode() == Op_ShenandoahLoadReferenceBarrier)) {
701 if (m->bottom_type()->make_oopptr() && m->bottom_type()->make_oopptr()->meet(TypePtr::NULL_PTR) == m->bottom_type()) {
702 report_verify_failure("Shenandoah verification: null input", n, m);
703 }
704 }
705
706 wq.push(m);
707 }
708 }
709
710 if (verify_no_useless_barrier) {
711 for (int i = 0; i < barriers.length(); i++) {
712 Node* n = barriers.at(i);
713 if (!barriers_used.member(n)) {
714 tty->print("XXX useless barrier"); n->dump(-2);
715 ShouldNotReachHere();
716 }
717 }
718 }
719 }
720 #endif
721
722 bool ShenandoahBarrierC2Support::is_dominator_same_ctrl(Node* c, Node* d, Node* n, PhaseIdealLoop* phase) {
723 // That both nodes have the same control is not sufficient to prove
724 // domination, verify that there's no path from d to n
725 ResourceMark rm;
726 Unique_Node_List wq;
727 wq.push(d);
728 for (uint next = 0; next < wq.size(); next++) {
729 Node *m = wq.at(next);
730 if (m == n) {
731 return false;
732 }
733 if (m->is_Phi() && m->in(0)->is_Loop()) {
734 assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control");
735 } else {
736 for (uint i = 0; i < m->req(); i++) {
737 if (m->in(i) != NULL && phase->ctrl_or_self(m->in(i)) == c) {
738 wq.push(m->in(i));
739 }
740 }
741 }
742 }
743 return true;
744 }
745
746 bool ShenandoahBarrierC2Support::is_dominator(Node* d_c, Node* n_c, Node* d, Node* n, PhaseIdealLoop* phase) {
747 if (d_c != n_c) {
748 return phase->is_dominator(d_c, n_c);
749 }
750 return is_dominator_same_ctrl(d_c, d, n, phase);
751 }
752
753 Node* next_mem(Node* mem, int alias) {
754 Node* res = NULL;
755 if (mem->is_Proj()) {
756 res = mem->in(0);
757 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
758 res = mem->in(TypeFunc::Memory);
759 } else if (mem->is_Phi()) {
760 res = mem->in(1);
761 } else if (mem->is_MergeMem()) {
762 res = mem->as_MergeMem()->memory_at(alias);
763 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
764 assert(alias = Compile::AliasIdxRaw, "following raw memory can't lead to a barrier");
765 res = mem->in(MemNode::Memory);
766 } else {
767 #ifdef ASSERT
768 mem->dump();
769 #endif
770 ShouldNotReachHere();
771 }
772 return res;
773 }
774
775 Node* ShenandoahBarrierC2Support::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) {
776 Node* iffproj = NULL;
777 while (c != dom) {
778 Node* next = phase->idom(c);
779 assert(next->unique_ctrl_out() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?");
780 if (c->is_Region()) {
781 ResourceMark rm;
782 Unique_Node_List wq;
783 wq.push(c);
784 for (uint i = 0; i < wq.size(); i++) {
785 Node *n = wq.at(i);
786 if (n == next) {
787 continue;
788 }
789 if (n->is_Region()) {
790 for (uint j = 1; j < n->req(); j++) {
791 wq.push(n->in(j));
792 }
793 } else {
794 wq.push(n->in(0));
795 }
796 }
797 for (uint i = 0; i < wq.size(); i++) {
798 Node *n = wq.at(i);
799 assert(n->is_CFG(), "");
800 if (n->is_Multi()) {
801 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
802 Node* u = n->fast_out(j);
803 if (u->is_CFG()) {
804 if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
805 return NodeSentinel;
806 }
807 }
808 }
809 }
810 }
811 } else if (c->is_Proj()) {
812 if (c->is_IfProj()) {
813 if (c->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) != NULL) {
814 // continue;
815 } else {
816 if (!allow_one_proj) {
817 return NodeSentinel;
818 }
819 if (iffproj == NULL) {
820 iffproj = c;
821 } else {
822 return NodeSentinel;
823 }
824 }
825 } else if (c->Opcode() == Op_JumpProj) {
826 return NodeSentinel; // unsupported
827 } else if (c->Opcode() == Op_CatchProj) {
828 return NodeSentinel; // unsupported
829 } else if (c->Opcode() == Op_CProj && next->Opcode() == Op_NeverBranch) {
830 return NodeSentinel; // unsupported
831 } else {
832 assert(next->unique_ctrl_out() == c, "unsupported branch pattern");
833 }
834 }
835 c = next;
836 }
837 return iffproj;
838 }
839
840 Node* ShenandoahBarrierC2Support::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) {
841 ResourceMark rm;
842 VectorSet wq(Thread::current()->resource_area());
843 wq.set(mem->_idx);
844 mem_ctrl = phase->ctrl_or_self(mem);
845 while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) {
846 mem = next_mem(mem, alias);
847 if (wq.test_set(mem->_idx)) {
848 return NULL;
849 }
850 mem_ctrl = phase->ctrl_or_self(mem);
851 }
852 if (mem->is_MergeMem()) {
853 mem = mem->as_MergeMem()->memory_at(alias);
854 mem_ctrl = phase->ctrl_or_self(mem);
855 }
856 return mem;
857 }
858
859 Node* ShenandoahBarrierC2Support::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) {
860 Node* mem = NULL;
861 Node* c = ctrl;
862 do {
863 if (c->is_Region()) {
864 Node* phi_bottom = NULL;
865 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == NULL; i++) {
866 Node* u = c->fast_out(i);
867 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
868 if (u->adr_type() == TypePtr::BOTTOM) {
869 mem = u;
870 }
871 }
872 }
873 } else {
874 if (c->is_Call() && c->as_Call()->adr_type() != NULL) {
875 CallProjections projs;
876 c->as_Call()->extract_projections(&projs, true, false);
877 if (projs.fallthrough_memproj != NULL) {
878 if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
879 if (projs.catchall_memproj == NULL) {
880 mem = projs.fallthrough_memproj;
881 } else {
882 if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) {
883 mem = projs.fallthrough_memproj;
884 } else {
885 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier");
886 mem = projs.catchall_memproj;
887 }
888 }
889 }
890 } else {
891 Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
892 if (proj != NULL &&
893 proj->adr_type() == TypePtr::BOTTOM) {
894 mem = proj;
895 }
896 }
897 } else {
898 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
899 Node* u = c->fast_out(i);
900 if (u->is_Proj() &&
901 u->bottom_type() == Type::MEMORY &&
902 u->adr_type() == TypePtr::BOTTOM) {
903 assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
904 assert(mem == NULL, "only one proj");
905 mem = u;
906 }
907 }
908 assert(!c->is_Call() || c->as_Call()->adr_type() != NULL || mem == NULL, "no mem projection expected");
909 }
910 }
911 c = phase->idom(c);
912 } while (mem == NULL);
913 return mem;
914 }
915
916 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
917 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
918 Node* u = n->fast_out(i);
919 if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
920 uses.push(u);
921 }
922 }
923 }
924
925 static void hide_strip_mined_loop(OuterStripMinedLoopNode* outer, CountedLoopNode* inner, PhaseIdealLoop* phase) {
926 OuterStripMinedLoopEndNode* le = inner->outer_loop_end();
927 Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl));
928 phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl));
929 Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt);
930 phase->register_control(new_le, phase->get_loop(le), le->in(0));
931 phase->lazy_replace(outer, new_outer);
932 phase->lazy_replace(le, new_le);
933 inner->clear_strip_mined();
934 }
935
936 void ShenandoahBarrierC2Support::test_heap_stable(Node*& ctrl, Node* raw_mem, Node*& heap_stable_ctrl,
937 PhaseIdealLoop* phase) {
938 IdealLoopTree* loop = phase->get_loop(ctrl);
939 Node* thread = new ThreadLocalNode();
940 phase->register_new_node(thread, ctrl);
941 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
942 phase->set_ctrl(offset, phase->C->root());
943 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, offset);
944 phase->register_new_node(gc_state_addr, ctrl);
945 uint gc_state_idx = Compile::AliasIdxRaw;
946 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument
947 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
948
949 Node* gc_state = new LoadBNode(ctrl, raw_mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered);
950 phase->register_new_node(gc_state, ctrl);
951 Node* heap_stable_and = new AndINode(gc_state, phase->igvn().intcon(ShenandoahHeap::HAS_FORWARDED));
952 phase->register_new_node(heap_stable_and, ctrl);
953 Node* heap_stable_cmp = new CmpINode(heap_stable_and, phase->igvn().zerocon(T_INT));
954 phase->register_new_node(heap_stable_cmp, ctrl);
955 Node* heap_stable_test = new BoolNode(heap_stable_cmp, BoolTest::ne);
956 phase->register_new_node(heap_stable_test, ctrl);
957 IfNode* heap_stable_iff = new IfNode(ctrl, heap_stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
958 phase->register_control(heap_stable_iff, loop, ctrl);
959
960 heap_stable_ctrl = new IfFalseNode(heap_stable_iff);
961 phase->register_control(heap_stable_ctrl, loop, heap_stable_iff);
962 ctrl = new IfTrueNode(heap_stable_iff);
963 phase->register_control(ctrl, loop, heap_stable_iff);
964
965 assert(is_heap_stable_test(heap_stable_iff), "Should match the shape");
966 }
967
968 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) {
969 const Type* val_t = phase->igvn().type(val);
970 if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
971 IdealLoopTree* loop = phase->get_loop(ctrl);
972 Node* null_cmp = new CmpPNode(val, phase->igvn().zerocon(T_OBJECT));
973 phase->register_new_node(null_cmp, ctrl);
974 Node* null_test = new BoolNode(null_cmp, BoolTest::ne);
975 phase->register_new_node(null_test, ctrl);
976 IfNode* null_iff = new IfNode(ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
977 phase->register_control(null_iff, loop, ctrl);
978 ctrl = new IfTrueNode(null_iff);
979 phase->register_control(ctrl, loop, null_iff);
980 null_ctrl = new IfFalseNode(null_iff);
981 phase->register_control(null_ctrl, loop, null_iff);
982 }
983 }
984
985 Node* ShenandoahBarrierC2Support::clone_null_check(Node*& c, Node* val, Node* unc_ctrl, PhaseIdealLoop* phase) {
986 IdealLoopTree *loop = phase->get_loop(c);
987 Node* iff = unc_ctrl->in(0);
988 assert(iff->is_If(), "broken");
989 Node* new_iff = iff->clone();
990 new_iff->set_req(0, c);
991 phase->register_control(new_iff, loop, c);
992 Node* iffalse = new IfFalseNode(new_iff->as_If());
993 phase->register_control(iffalse, loop, new_iff);
994 Node* iftrue = new IfTrueNode(new_iff->as_If());
995 phase->register_control(iftrue, loop, new_iff);
996 c = iftrue;
997 const Type *t = phase->igvn().type(val);
998 assert(val->Opcode() == Op_CastPP, "expect cast to non null here");
999 Node* uncasted_val = val->in(1);
1000 val = new CastPPNode(uncasted_val, t);
1001 val->init_req(0, c);
1002 phase->register_new_node(val, c);
1003 return val;
1004 }
1005
1006 void ShenandoahBarrierC2Support::fix_null_check(Node* unc, Node* unc_ctrl, Node* new_unc_ctrl,
1007 Unique_Node_List& uses, PhaseIdealLoop* phase) {
1008 IfNode* iff = unc_ctrl->in(0)->as_If();
1009 Node* proj = iff->proj_out(0);
1010 assert(proj != unc_ctrl, "bad projection");
1011 Node* use = proj->unique_ctrl_out();
1012
1013 assert(use == unc || use->is_Region(), "what else?");
1014
1015 uses.clear();
1016 if (use == unc) {
1017 phase->set_idom(use, new_unc_ctrl, phase->dom_depth(use));
1018 for (uint i = 1; i < unc->req(); i++) {
1019 Node* n = unc->in(i);
1020 if (phase->has_ctrl(n) && phase->get_ctrl(n) == proj) {
1021 uses.push(n);
1022 }
1023 }
1024 } else {
1025 assert(use->is_Region(), "what else?");
1026 uint idx = 1;
1027 for (; use->in(idx) != proj; idx++);
1028 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
1029 Node* u = use->fast_out(i);
1030 if (u->is_Phi() && phase->get_ctrl(u->in(idx)) == proj) {
1031 uses.push(u->in(idx));
1032 }
1033 }
1034 }
1035 for(uint next = 0; next < uses.size(); next++ ) {
1036 Node *n = uses.at(next);
1037 assert(phase->get_ctrl(n) == proj, "bad control");
1038 phase->set_ctrl_and_loop(n, new_unc_ctrl);
1039 if (n->in(0) == proj) {
1040 phase->igvn().replace_input_of(n, 0, new_unc_ctrl);
1041 }
1042 for (uint i = 0; i < n->req(); i++) {
1043 Node* m = n->in(i);
1044 if (m != NULL && phase->has_ctrl(m) && phase->get_ctrl(m) == proj) {
1045 uses.push(m);
1046 }
1047 }
1048 }
1049
1050 phase->igvn().rehash_node_delayed(use);
1051 int nb = use->replace_edge(proj, new_unc_ctrl);
1052 assert(nb == 1, "only use expected");
1053 }
1054
1055 void ShenandoahBarrierC2Support::in_cset_fast_test(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) {
1056 IdealLoopTree *loop = phase->get_loop(ctrl);
1057 Node* raw_rbtrue = new CastP2XNode(ctrl, val);
1058 phase->register_new_node(raw_rbtrue, ctrl);
1059 Node* cset_offset = new URShiftXNode(raw_rbtrue, phase->igvn().intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
1060 phase->register_new_node(cset_offset, ctrl);
1061 Node* in_cset_fast_test_base_addr = phase->igvn().makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
1062 phase->set_ctrl(in_cset_fast_test_base_addr, phase->C->root());
1063 Node* in_cset_fast_test_adr = new AddPNode(phase->C->top(), in_cset_fast_test_base_addr, cset_offset);
1064 phase->register_new_node(in_cset_fast_test_adr, ctrl);
1065 uint in_cset_fast_test_idx = Compile::AliasIdxRaw;
1066 const TypePtr* in_cset_fast_test_adr_type = NULL; // debug-mode-only argument
1067 debug_only(in_cset_fast_test_adr_type = phase->C->get_adr_type(in_cset_fast_test_idx));
1068 Node* in_cset_fast_test_load = new LoadBNode(ctrl, raw_mem, in_cset_fast_test_adr, in_cset_fast_test_adr_type, TypeInt::BYTE, MemNode::unordered);
1069 phase->register_new_node(in_cset_fast_test_load, ctrl);
1070 Node* in_cset_fast_test_cmp = new CmpINode(in_cset_fast_test_load, phase->igvn().zerocon(T_INT));
1071 phase->register_new_node(in_cset_fast_test_cmp, ctrl);
1072 Node* in_cset_fast_test_test = new BoolNode(in_cset_fast_test_cmp, BoolTest::eq);
1073 phase->register_new_node(in_cset_fast_test_test, ctrl);
1074 IfNode* in_cset_fast_test_iff = new IfNode(ctrl, in_cset_fast_test_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
1075 phase->register_control(in_cset_fast_test_iff, loop, ctrl);
1076
1077 not_cset_ctrl = new IfTrueNode(in_cset_fast_test_iff);
1078 phase->register_control(not_cset_ctrl, loop, in_cset_fast_test_iff);
1079
1080 ctrl = new IfFalseNode(in_cset_fast_test_iff);
1081 phase->register_control(ctrl, loop, in_cset_fast_test_iff);
1082 }
1083
1084 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node*& result_mem, Node* raw_mem, PhaseIdealLoop* phase) {
1085 IdealLoopTree*loop = phase->get_loop(ctrl);
1086 const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr()->cast_to_nonconst();
1087
1088 // The slow path stub consumes and produces raw memory in addition
1089 // to the existing memory edges
1090 Node* base = find_bottom_mem(ctrl, phase);
1091 MergeMemNode* mm = MergeMemNode::make(base);
1092 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
1093 phase->register_new_node(mm, ctrl);
1094
1095 Node* call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_write_barrier_Type(), CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_JRT), "shenandoah_write_barrier", TypeRawPtr::BOTTOM);
1096 call->init_req(TypeFunc::Control, ctrl);
1097 call->init_req(TypeFunc::I_O, phase->C->top());
1098 call->init_req(TypeFunc::Memory, mm);
1099 call->init_req(TypeFunc::FramePtr, phase->C->top());
1100 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1101 call->init_req(TypeFunc::Parms, val);
1102 phase->register_control(call, loop, ctrl);
1103 ctrl = new ProjNode(call, TypeFunc::Control);
1104 phase->register_control(ctrl, loop, call);
1105 result_mem = new ProjNode(call, TypeFunc::Memory);
1106 phase->register_new_node(result_mem, call);
1107 val = new ProjNode(call, TypeFunc::Parms);
1108 phase->register_new_node(val, call);
1109 val = new CheckCastPPNode(ctrl, val, obj_type);
1110 phase->register_new_node(val, ctrl);
1111 }
1112
1113 void ShenandoahBarrierC2Support::fix_ctrl(Node* barrier, Node* region, const MemoryGraphFixer& fixer, Unique_Node_List& uses, Unique_Node_List& uses_to_ignore, uint last, PhaseIdealLoop* phase) {
1114 Node* ctrl = phase->get_ctrl(barrier);
1115 Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
1116
1117 // Update the control of all nodes that should be after the
1118 // barrier control flow
1119 uses.clear();
1120 // Every node that is control dependent on the barrier's input
1121 // control will be after the expanded barrier. The raw memory (if
1122 // its memory is control dependent on the barrier's input control)
1123 // must stay above the barrier.
1124 uses_to_ignore.clear();
1125 if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
1126 uses_to_ignore.push(init_raw_mem);
1127 }
1128 for (uint next = 0; next < uses_to_ignore.size(); next++) {
1129 Node *n = uses_to_ignore.at(next);
1130 for (uint i = 0; i < n->req(); i++) {
1131 Node* in = n->in(i);
1132 if (in != NULL && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
1133 uses_to_ignore.push(in);
1134 }
1135 }
1136 }
1137 for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
1138 Node* u = ctrl->fast_out(i);
1139 if (u->_idx < last &&
1140 u != barrier &&
1141 !uses_to_ignore.member(u) &&
1142 (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) &&
1143 (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) {
1144 Node* old_c = phase->ctrl_or_self(u);
1145 Node* c = old_c;
1146 if (c != ctrl ||
1147 is_dominator_same_ctrl(old_c, barrier, u, phase) ||
1148 ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) {
1149 phase->igvn().rehash_node_delayed(u);
1150 int nb = u->replace_edge(ctrl, region);
1151 if (u->is_CFG()) {
1152 if (phase->idom(u) == ctrl) {
1153 phase->set_idom(u, region, phase->dom_depth(region));
1154 }
1155 } else if (phase->get_ctrl(u) == ctrl) {
1156 assert(u != init_raw_mem, "should leave input raw mem above the barrier");
1157 uses.push(u);
1158 }
1159 assert(nb == 1, "more than 1 ctrl input?");
1160 --i, imax -= nb;
1161 }
1162 }
1163 }
1164 }
1165
1166 static Node* create_phis_on_call_return(Node* ctrl, Node* c, Node* n, Node* n_clone, const CallProjections& projs, PhaseIdealLoop* phase) {
1167 Node* region = NULL;
1168 while (c != ctrl) {
1169 if (c->is_Region()) {
1170 region = c;
1171 }
1172 c = phase->idom(c);
1173 }
1174 assert(region != NULL, "");
1175 Node* phi = new PhiNode(region, n->bottom_type());
1176 for (uint j = 1; j < region->req(); j++) {
1177 Node* in = region->in(j);
1178 if (phase->is_dominator(projs.fallthrough_catchproj, in)) {
1179 phi->init_req(j, n);
1180 } else if (phase->is_dominator(projs.catchall_catchproj, in)) {
1181 phi->init_req(j, n_clone);
1182 } else {
1183 phi->init_req(j, create_phis_on_call_return(ctrl, in, n, n_clone, projs, phase));
1184 }
1185 }
1186 phase->register_new_node(phi, region);
1187 return phi;
1188 }
1189
1190 void ShenandoahBarrierC2Support::pin_and_expand(PhaseIdealLoop* phase) {
1191 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
1192
1193 // Collect raw memory state at CFG points in the entire graph and
1194 // record it in memory_nodes. Optimize the raw memory graph in the
1195 // process. Optimizing the memory graph also makes the memory graph
1196 // simpler.
1197 GrowableArray<MemoryGraphFixer*> memory_graph_fixers;
1198
1199 Unique_Node_List uses;
1200 for (int i = 0; i < state->enqueue_barriers_count(); i++) {
1201 Node* barrier = state->enqueue_barrier(i);
1202 Node* ctrl = phase->get_ctrl(barrier);
1203 IdealLoopTree* loop = phase->get_loop(ctrl);
1204 if (loop->_head->is_OuterStripMinedLoop()) {
1205 // Expanding a barrier here will break loop strip mining
1206 // verification. Transform the loop so the loop nest doesn't
1207 // appear as strip mined.
1208 OuterStripMinedLoopNode* outer = loop->_head->as_OuterStripMinedLoop();
1209 hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1210 }
1211 }
1212
1213 Node_Stack stack(0);
1214 Node_List clones;
1215 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1216 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1217 if (lrb->get_barrier_strength() == ShenandoahLoadReferenceBarrierNode::NONE) {
1218 continue;
1219 }
1220
1221 Node* ctrl = phase->get_ctrl(lrb);
1222 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1223
1224 CallStaticJavaNode* unc = NULL;
1225 Node* unc_ctrl = NULL;
1226 Node* uncasted_val = val;
1227
1228 for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1229 Node* u = lrb->fast_out(i);
1230 if (u->Opcode() == Op_CastPP &&
1231 u->in(0) != NULL &&
1232 phase->is_dominator(u->in(0), ctrl)) {
1233 const Type* u_t = phase->igvn().type(u);
1234
1235 if (u_t->meet(TypePtr::NULL_PTR) != u_t &&
1236 u->in(0)->Opcode() == Op_IfTrue &&
1237 u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
1238 u->in(0)->in(0)->is_If() &&
1239 u->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
1240 u->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
1241 u->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
1242 u->in(0)->in(0)->in(1)->in(1)->in(1) == val &&
1243 u->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
1244 IdealLoopTree* loop = phase->get_loop(ctrl);
1245 IdealLoopTree* unc_loop = phase->get_loop(u->in(0));
1246
1247 if (!unc_loop->is_member(loop)) {
1248 continue;
1249 }
1250
1251 Node* branch = no_branches(ctrl, u->in(0), false, phase);
1252 assert(branch == NULL || branch == NodeSentinel, "was not looking for a branch");
1253 if (branch == NodeSentinel) {
1254 continue;
1255 }
1256
1257 phase->igvn().replace_input_of(u, 1, val);
1258 phase->igvn().replace_input_of(lrb, ShenandoahLoadReferenceBarrierNode::ValueIn, u);
1259 phase->set_ctrl(u, u->in(0));
1260 phase->set_ctrl(lrb, u->in(0));
1261 unc = u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
1262 unc_ctrl = u->in(0);
1263 val = u;
1264
1265 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
1266 Node* u = val->fast_out(j);
1267 if (u == lrb) continue;
1268 phase->igvn().rehash_node_delayed(u);
1269 int nb = u->replace_edge(val, lrb);
1270 --j; jmax -= nb;
1271 }
1272
1273 RegionNode* r = new RegionNode(3);
1274 IfNode* iff = unc_ctrl->in(0)->as_If();
1275
1276 Node* ctrl_use = unc_ctrl->unique_ctrl_out();
1277 Node* unc_ctrl_clone = unc_ctrl->clone();
1278 phase->register_control(unc_ctrl_clone, loop, iff);
1279 Node* c = unc_ctrl_clone;
1280 Node* new_cast = clone_null_check(c, val, unc_ctrl_clone, phase);
1281 r->init_req(1, new_cast->in(0)->in(0)->as_If()->proj_out(0));
1282
1283 phase->igvn().replace_input_of(unc_ctrl, 0, c->in(0));
1284 phase->set_idom(unc_ctrl, c->in(0), phase->dom_depth(unc_ctrl));
1285 phase->lazy_replace(c, unc_ctrl);
1286 c = NULL;;
1287 phase->igvn().replace_input_of(val, 0, unc_ctrl_clone);
1288 phase->set_ctrl(val, unc_ctrl_clone);
1289
1290 IfNode* new_iff = new_cast->in(0)->in(0)->as_If();
1291 fix_null_check(unc, unc_ctrl_clone, r, uses, phase);
1292 Node* iff_proj = iff->proj_out(0);
1293 r->init_req(2, iff_proj);
1294 phase->register_control(r, phase->ltree_root(), iff);
1295
1296 Node* new_bol = new_iff->in(1)->clone();
1297 Node* new_cmp = new_bol->in(1)->clone();
1298 assert(new_cmp->Opcode() == Op_CmpP, "broken");
1299 assert(new_cmp->in(1) == val->in(1), "broken");
1300 new_bol->set_req(1, new_cmp);
1301 new_cmp->set_req(1, lrb);
1302 phase->register_new_node(new_bol, new_iff->in(0));
1303 phase->register_new_node(new_cmp, new_iff->in(0));
1304 phase->igvn().replace_input_of(new_iff, 1, new_bol);
1305 phase->igvn().replace_input_of(new_cast, 1, lrb);
1306
1307 for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1308 Node* u = lrb->fast_out(i);
1309 if (u == new_cast || u == new_cmp) {
1310 continue;
1311 }
1312 phase->igvn().rehash_node_delayed(u);
1313 int nb = u->replace_edge(lrb, new_cast);
1314 assert(nb > 0, "no update?");
1315 --i; imax -= nb;
1316 }
1317
1318 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
1319 Node* u = val->fast_out(i);
1320 if (u == lrb) {
1321 continue;
1322 }
1323 phase->igvn().rehash_node_delayed(u);
1324 int nb = u->replace_edge(val, new_cast);
1325 assert(nb > 0, "no update?");
1326 --i; imax -= nb;
1327 }
1328
1329 ctrl = unc_ctrl_clone;
1330 phase->set_ctrl_and_loop(lrb, ctrl);
1331 break;
1332 }
1333 }
1334 }
1335 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) {
1336 CallNode* call = ctrl->in(0)->as_CallJava();
1337 CallProjections projs;
1338 call->extract_projections(&projs, false, false);
1339
1340 Node* lrb_clone = lrb->clone();
1341 phase->register_new_node(lrb_clone, projs.catchall_catchproj);
1342 phase->set_ctrl(lrb, projs.fallthrough_catchproj);
1343
1344 stack.push(lrb, 0);
1345 clones.push(lrb_clone);
1346
1347 do {
1348 assert(stack.size() == clones.size(), "");
1349 Node* n = stack.node();
1350 #ifdef ASSERT
1351 if (n->is_Load()) {
1352 Node* mem = n->in(MemNode::Memory);
1353 for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
1354 Node* u = mem->fast_out(j);
1355 assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?");
1356 }
1357 }
1358 #endif
1359 uint idx = stack.index();
1360 Node* n_clone = clones.at(clones.size()-1);
1361 if (idx < n->outcnt()) {
1362 Node* u = n->raw_out(idx);
1363 Node* c = phase->ctrl_or_self(u);
1364 if (c == ctrl) {
1365 stack.set_index(idx+1);
1366 assert(!u->is_CFG(), "");
1367 stack.push(u, 0);
1368 Node* u_clone = u->clone();
1369 int nb = u_clone->replace_edge(n, n_clone);
1370 assert(nb > 0, "should have replaced some uses");
1371 phase->register_new_node(u_clone, projs.catchall_catchproj);
1372 clones.push(u_clone);
1373 phase->set_ctrl(u, projs.fallthrough_catchproj);
1374 } else {
1375 bool replaced = false;
1376 if (u->is_Phi()) {
1377 for (uint k = 1; k < u->req(); k++) {
1378 if (u->in(k) == n) {
1379 if (phase->is_dominator(projs.catchall_catchproj, u->in(0)->in(k))) {
1380 phase->igvn().replace_input_of(u, k, n_clone);
1381 replaced = true;
1382 } else if (!phase->is_dominator(projs.fallthrough_catchproj, u->in(0)->in(k))) {
1383 phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase));
1384 replaced = true;
1385 }
1386 }
1387 }
1388 } else {
1389 if (phase->is_dominator(projs.catchall_catchproj, c)) {
1390 phase->igvn().rehash_node_delayed(u);
1391 int nb = u->replace_edge(n, n_clone);
1392 assert(nb > 0, "should have replaced some uses");
1393 replaced = true;
1394 } else if (!phase->is_dominator(projs.fallthrough_catchproj, c)) {
1395 phase->igvn().rehash_node_delayed(u);
1396 int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase));
1397 assert(nb > 0, "should have replaced some uses");
1398 replaced = true;
1399 }
1400 }
1401 if (!replaced) {
1402 stack.set_index(idx+1);
1403 }
1404 }
1405 } else {
1406 // assert(n_clone->outcnt() > 0, "");
1407 // assert(n->outcnt() > 0, "");
1408 stack.pop();
1409 clones.pop();
1410 }
1411 } while (stack.size() > 0);
1412 assert(stack.size() == 0 && clones.size() == 0, "");
1413 ctrl = projs.fallthrough_catchproj;
1414 }
1415 }
1416
1417 // Expand load-reference-barriers
1418 MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
1419 Unique_Node_List uses_to_ignore;
1420 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1421 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1422 if (lrb->get_barrier_strength() == ShenandoahLoadReferenceBarrierNode::NONE) {
1423 phase->igvn().replace_node(lrb, lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1424 continue;
1425 }
1426 uint last = phase->C->unique();
1427 Node* ctrl = phase->get_ctrl(lrb);
1428 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1429
1430
1431 Node* orig_ctrl = ctrl;
1432
1433 Node* raw_mem = fixer.find_mem(ctrl, lrb);
1434 Node* init_raw_mem = raw_mem;
1435 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL);
1436 // int alias = phase->C->get_alias_index(lrb->adr_type());
1437
1438 IdealLoopTree *loop = phase->get_loop(ctrl);
1439 CallStaticJavaNode* unc = lrb->pin_and_expand_null_check(phase->igvn());
1440 Node* unc_ctrl = NULL;
1441 if (unc != NULL) {
1442 if (val->in(ShenandoahLoadReferenceBarrierNode::Control) != ctrl) {
1443 unc = NULL;
1444 } else {
1445 unc_ctrl = val->in(ShenandoahLoadReferenceBarrierNode::Control);
1446 }
1447 }
1448
1449 Node* uncasted_val = val;
1450 if (unc != NULL) {
1451 uncasted_val = val->in(1);
1452 }
1453
1454 Node* heap_stable_ctrl = NULL;
1455 Node* null_ctrl = NULL;
1456
1457 assert(val->bottom_type()->make_oopptr(), "need oop");
1458 assert(val->bottom_type()->make_oopptr()->const_oop() == NULL, "expect non-constant");
1459
1460 enum { _heap_stable = 1, _not_cset, _fwded, _evac_path, _null_path, PATH_LIMIT };
1461 Node* region = new RegionNode(PATH_LIMIT);
1462 Node* val_phi = new PhiNode(region, uncasted_val->bottom_type()->is_oopptr());
1463 Node* raw_mem_phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1464
1465 // Stable path.
1466 test_heap_stable(ctrl, raw_mem, heap_stable_ctrl, phase);
1467 IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
1468
1469 // Heap stable case
1470 region->init_req(_heap_stable, heap_stable_ctrl);
1471 val_phi->init_req(_heap_stable, uncasted_val);
1472 raw_mem_phi->init_req(_heap_stable, raw_mem);
1473
1474 Node* reg2_ctrl = NULL;
1475 // Null case
1476 test_null(ctrl, val, null_ctrl, phase);
1477 if (null_ctrl != NULL) {
1478 reg2_ctrl = null_ctrl->in(0);
1479 region->init_req(_null_path, null_ctrl);
1480 val_phi->init_req(_null_path, uncasted_val);
1481 raw_mem_phi->init_req(_null_path, raw_mem);
1482 } else {
1483 region->del_req(_null_path);
1484 val_phi->del_req(_null_path);
1485 raw_mem_phi->del_req(_null_path);
1486 }
1487
1488 // Test for in-cset.
1489 // Wires !in_cset(obj) to slot 2 of region and phis
1490 Node* not_cset_ctrl = NULL;
1491 in_cset_fast_test(ctrl, not_cset_ctrl, uncasted_val, raw_mem, phase);
1492 if (not_cset_ctrl != NULL) {
1493 if (reg2_ctrl == NULL) reg2_ctrl = not_cset_ctrl->in(0);
1494 region->init_req(_not_cset, not_cset_ctrl);
1495 val_phi->init_req(_not_cset, uncasted_val);
1496 raw_mem_phi->init_req(_not_cset, raw_mem);
1497 }
1498
1499 // Resolve object when orig-value is in cset.
1500 // Make the unconditional resolve for fwdptr.
1501 Node* new_val = uncasted_val;
1502 if (unc_ctrl != NULL) {
1503 // Clone the null check in this branch to allow implicit null check
1504 new_val = clone_null_check(ctrl, val, unc_ctrl, phase);
1505 fix_null_check(unc, unc_ctrl, ctrl->in(0)->as_If()->proj_out(0), uses, phase);
1506
1507 IfNode* iff = unc_ctrl->in(0)->as_If();
1508 phase->igvn().replace_input_of(iff, 1, phase->igvn().intcon(1));
1509 }
1510 Node* addr = new AddPNode(new_val, new_val, phase->igvn().MakeConX(oopDesc::mark_offset_in_bytes()));
1511 phase->register_new_node(addr, ctrl);
1512 assert(new_val->bottom_type()->isa_oopptr(), "what else?");
1513 Node* markword = new LoadXNode(ctrl, raw_mem, addr, TypeRawPtr::BOTTOM, TypeX_X, MemNode::unordered);
1514 phase->register_new_node(markword, ctrl);
1515
1516 // Test if object is forwarded. This is the case if lowest two bits are set.
1517 Node* masked = new AndXNode(markword, phase->igvn().MakeConX(markOopDesc::lock_mask_in_place));
1518 phase->register_new_node(masked, ctrl);
1519 Node* cmp = new CmpXNode(masked, phase->igvn().MakeConX(markOopDesc::marked_value));
1520 phase->register_new_node(cmp, ctrl);
1521
1522 // Only branch to LRB stub if object is not forwarded; otherwise reply with fwd ptr
1523 Node* bol = new BoolNode(cmp, BoolTest::eq); // Equals 3 means it's forwarded
1524 phase->register_new_node(bol, ctrl);
1525
1526 IfNode* iff = new IfNode(ctrl, bol, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
1527 phase->register_control(iff, loop, ctrl);
1528 Node* if_fwd = new IfTrueNode(iff);
1529 phase->register_control(if_fwd, loop, iff);
1530 Node* if_not_fwd = new IfFalseNode(iff);
1531 phase->register_control(if_not_fwd, loop, iff);
1532
1533 // Decode forward pointer.
1534 Node* masked2 = new AndXNode(markword, phase->igvn().MakeConX(~markOopDesc::lock_mask_in_place));
1535 phase->register_new_node(masked2, if_fwd);
1536 Node* fwdraw = new CastX2PNode(masked2);
1537 fwdraw->init_req(0, if_fwd);
1538 phase->register_new_node(fwdraw, if_fwd);
1539 Node* fwd = new CheckCastPPNode(NULL, fwdraw, val->bottom_type());
1540 phase->register_new_node(fwd, if_fwd);
1541
1542 // Wire up not-equal-path in slots 3.
1543 region->init_req(_fwded, if_fwd);
1544 val_phi->init_req(_fwded, fwd);
1545 raw_mem_phi->init_req(_fwded, raw_mem);
1546
1547 // Call wb-stub and wire up that path in slots 4
1548 Node* result_mem = NULL;
1549 ctrl = if_not_fwd;
1550 fwd = new_val;
1551 call_lrb_stub(ctrl, fwd, result_mem, raw_mem, phase);
1552 region->init_req(_evac_path, ctrl);
1553 val_phi->init_req(_evac_path, fwd);
1554 raw_mem_phi->init_req(_evac_path, result_mem);
1555
1556 phase->register_control(region, loop, heap_stable_iff);
1557 Node* out_val = val_phi;
1558 phase->register_new_node(val_phi, region);
1559 phase->register_new_node(raw_mem_phi, region);
1560
1561 fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase);
1562
1563 ctrl = orig_ctrl;
1564
1565 if (unc != NULL) {
1566 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
1567 Node* u = val->fast_out(i);
1568 Node* c = phase->ctrl_or_self(u);
1569 if (u != lrb && (c != ctrl || is_dominator_same_ctrl(c, lrb, u, phase))) {
1570 phase->igvn().rehash_node_delayed(u);
1571 int nb = u->replace_edge(val, out_val);
1572 --i, imax -= nb;
1573 }
1574 }
1575 if (val->outcnt() == 0) {
1576 phase->igvn()._worklist.push(val);
1577 }
1578 }
1579 phase->igvn().replace_node(lrb, out_val);
1580
1581 follow_barrier_uses(out_val, ctrl, uses, phase);
1582
1583 for(uint next = 0; next < uses.size(); next++ ) {
1584 Node *n = uses.at(next);
1585 assert(phase->get_ctrl(n) == ctrl, "bad control");
1586 assert(n != init_raw_mem, "should leave input raw mem above the barrier");
1587 phase->set_ctrl(n, region);
1588 follow_barrier_uses(n, ctrl, uses, phase);
1589 }
1590
1591 // The slow path call produces memory: hook the raw memory phi
1592 // from the expanded load reference barrier with the rest of the graph
1593 // which may require adding memory phis at every post dominated
1594 // region and at enclosing loop heads. Use the memory state
1595 // collected in memory_nodes to fix the memory graph. Update that
1596 // memory state as we go.
1597 fixer.fix_mem(ctrl, region, init_raw_mem, raw_mem_for_ctrl, raw_mem_phi, uses);
1598 }
1599 // Done expanding load-reference-barriers.
1600 assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced");
1601
1602 for (int i = state->enqueue_barriers_count() - 1; i >= 0; i--) {
1603 Node* barrier = state->enqueue_barrier(i);
1604 Node* pre_val = barrier->in(1);
1605
1606 if (phase->igvn().type(pre_val)->higher_equal(TypePtr::NULL_PTR)) {
1607 ShouldNotReachHere();
1608 continue;
1609 }
1610
1611 Node* ctrl = phase->get_ctrl(barrier);
1612
1613 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) {
1614 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0)->in(0), pre_val, ctrl->in(0), phase), "can't move");
1615 ctrl = ctrl->in(0)->in(0);
1616 phase->set_ctrl(barrier, ctrl);
1617 } else if (ctrl->is_CallRuntime()) {
1618 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0), pre_val, ctrl, phase), "can't move");
1619 ctrl = ctrl->in(0);
1620 phase->set_ctrl(barrier, ctrl);
1621 }
1622
1623 Node* init_ctrl = ctrl;
1624 IdealLoopTree* loop = phase->get_loop(ctrl);
1625 Node* raw_mem = fixer.find_mem(ctrl, barrier);
1626 Node* init_raw_mem = raw_mem;
1627 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL);
1628 Node* heap_stable_ctrl = NULL;
1629 Node* null_ctrl = NULL;
1630 uint last = phase->C->unique();
1631
1632 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
1633 Node* region = new RegionNode(PATH_LIMIT);
1634 Node* phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1635
1636 enum { _fast_path = 1, _slow_path, _null_path, PATH_LIMIT2 };
1637 Node* region2 = new RegionNode(PATH_LIMIT2);
1638 Node* phi2 = PhiNode::make(region2, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1639
1640 // Stable path.
1641 test_heap_stable(ctrl, raw_mem, heap_stable_ctrl, phase);
1642 region->init_req(_heap_stable, heap_stable_ctrl);
1643 phi->init_req(_heap_stable, raw_mem);
1644
1645 // Null path
1646 Node* reg2_ctrl = NULL;
1647 test_null(ctrl, pre_val, null_ctrl, phase);
1648 if (null_ctrl != NULL) {
1649 reg2_ctrl = null_ctrl->in(0);
1650 region2->init_req(_null_path, null_ctrl);
1651 phi2->init_req(_null_path, raw_mem);
1652 } else {
1653 region2->del_req(_null_path);
1654 phi2->del_req(_null_path);
1655 }
1656
1657 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
1658 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1659 Node* thread = new ThreadLocalNode();
1660 phase->register_new_node(thread, ctrl);
1661 Node* buffer_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(buffer_offset));
1662 phase->register_new_node(buffer_adr, ctrl);
1663 Node* index_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(index_offset));
1664 phase->register_new_node(index_adr, ctrl);
1665
1666 BasicType index_bt = TypeX_X->basic_type();
1667 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
1668 const TypePtr* adr_type = TypeRawPtr::BOTTOM;
1669 Node* index = new LoadXNode(ctrl, raw_mem, index_adr, adr_type, TypeX_X, MemNode::unordered);
1670 phase->register_new_node(index, ctrl);
1671 Node* index_cmp = new CmpXNode(index, phase->igvn().MakeConX(0));
1672 phase->register_new_node(index_cmp, ctrl);
1673 Node* index_test = new BoolNode(index_cmp, BoolTest::ne);
1674 phase->register_new_node(index_test, ctrl);
1675 IfNode* queue_full_iff = new IfNode(ctrl, index_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
1676 if (reg2_ctrl == NULL) reg2_ctrl = queue_full_iff;
1677 phase->register_control(queue_full_iff, loop, ctrl);
1678 Node* not_full = new IfTrueNode(queue_full_iff);
1679 phase->register_control(not_full, loop, queue_full_iff);
1680 Node* full = new IfFalseNode(queue_full_iff);
1681 phase->register_control(full, loop, queue_full_iff);
1682
1683 ctrl = not_full;
1684
1685 Node* next_index = new SubXNode(index, phase->igvn().MakeConX(sizeof(intptr_t)));
1686 phase->register_new_node(next_index, ctrl);
1687
1688 Node* buffer = new LoadPNode(ctrl, raw_mem, buffer_adr, adr_type, TypeRawPtr::NOTNULL, MemNode::unordered);
1689 phase->register_new_node(buffer, ctrl);
1690 Node *log_addr = new AddPNode(phase->C->top(), buffer, next_index);
1691 phase->register_new_node(log_addr, ctrl);
1692 Node* log_store = new StorePNode(ctrl, raw_mem, log_addr, adr_type, pre_val, MemNode::unordered);
1693 phase->register_new_node(log_store, ctrl);
1694 // update the index
1695 Node* index_update = new StoreXNode(ctrl, log_store, index_adr, adr_type, next_index, MemNode::unordered);
1696 phase->register_new_node(index_update, ctrl);
1697
1698 // Fast-path case
1699 region2->init_req(_fast_path, ctrl);
1700 phi2->init_req(_fast_path, index_update);
1701
1702 ctrl = full;
1703
1704 Node* base = find_bottom_mem(ctrl, phase);
1705
1706 MergeMemNode* mm = MergeMemNode::make(base);
1707 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
1708 phase->register_new_node(mm, ctrl);
1709
1710 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);
1711 call->init_req(TypeFunc::Control, ctrl);
1712 call->init_req(TypeFunc::I_O, phase->C->top());
1713 call->init_req(TypeFunc::Memory, mm);
1714 call->init_req(TypeFunc::FramePtr, phase->C->top());
1715 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1716 call->init_req(TypeFunc::Parms, pre_val);
1717 call->init_req(TypeFunc::Parms+1, thread);
1718 phase->register_control(call, loop, ctrl);
1719
1720 Node* ctrl_proj = new ProjNode(call, TypeFunc::Control);
1721 phase->register_control(ctrl_proj, loop, call);
1722 Node* mem_proj = new ProjNode(call, TypeFunc::Memory);
1723 phase->register_new_node(mem_proj, call);
1724
1725 // Slow-path case
1726 region2->init_req(_slow_path, ctrl_proj);
1727 phi2->init_req(_slow_path, mem_proj);
1728
1729 phase->register_control(region2, loop, reg2_ctrl);
1730 phase->register_new_node(phi2, region2);
1731
1732 region->init_req(_heap_unstable, region2);
1733 phi->init_req(_heap_unstable, phi2);
1734
1735 phase->register_control(region, loop, heap_stable_ctrl->in(0));
1736 phase->register_new_node(phi, region);
1737
1738 fix_ctrl(barrier, region, fixer, uses, uses_to_ignore, last, phase);
1739 for(uint next = 0; next < uses.size(); next++ ) {
1740 Node *n = uses.at(next);
1741 assert(phase->get_ctrl(n) == init_ctrl, "bad control");
1742 assert(n != init_raw_mem, "should leave input raw mem above the barrier");
1743 phase->set_ctrl(n, region);
1744 follow_barrier_uses(n, init_ctrl, uses, phase);
1745 }
1746 fixer.fix_mem(init_ctrl, region, init_raw_mem, raw_mem_for_ctrl, phi, uses);
1747
1748 phase->igvn().replace_node(barrier, pre_val);
1749 }
1750 assert(state->enqueue_barriers_count() == 0, "all enqueue barrier nodes should have been replaced");
1751
1752 }
1753
1754 void ShenandoahBarrierC2Support::move_heap_stable_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
1755 IdealLoopTree *loop = phase->get_loop(iff);
1756 Node* loop_head = loop->_head;
1757 Node* entry_c = loop_head->in(LoopNode::EntryControl);
1758
1759 Node* bol = iff->in(1);
1760 Node* cmp = bol->in(1);
1761 Node* andi = cmp->in(1);
1762 Node* load = andi->in(1);
1763
1764 assert(is_gc_state_load(load), "broken");
1765 if (!phase->is_dominator(load->in(0), entry_c)) {
1766 Node* mem_ctrl = NULL;
1767 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
1768 load = load->clone();
1769 load->set_req(MemNode::Memory, mem);
1770 load->set_req(0, entry_c);
1771 phase->register_new_node(load, entry_c);
1772 andi = andi->clone();
1773 andi->set_req(1, load);
1774 phase->register_new_node(andi, entry_c);
1775 cmp = cmp->clone();
1776 cmp->set_req(1, andi);
1777 phase->register_new_node(cmp, entry_c);
1778 bol = bol->clone();
1779 bol->set_req(1, cmp);
1780 phase->register_new_node(bol, entry_c);
1781
1782 Node* old_bol =iff->in(1);
1783 phase->igvn().replace_input_of(iff, 1, bol);
1784 }
1785 }
1786
1787 bool ShenandoahBarrierC2Support::identical_backtoback_ifs(Node* n, PhaseIdealLoop* phase) {
1788 if (!n->is_If() || n->is_CountedLoopEnd()) {
1789 return false;
1790 }
1791 Node* region = n->in(0);
1792
1793 if (!region->is_Region()) {
1794 return false;
1795 }
1796 Node* dom = phase->idom(region);
1797 if (!dom->is_If()) {
1798 return false;
1799 }
1800
1801 if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) {
1802 return false;
1803 }
1804
1805 IfNode* dom_if = dom->as_If();
1806 Node* proj_true = dom_if->proj_out(1);
1807 Node* proj_false = dom_if->proj_out(0);
1808
1809 for (uint i = 1; i < region->req(); i++) {
1810 if (phase->is_dominator(proj_true, region->in(i))) {
1811 continue;
1812 }
1813 if (phase->is_dominator(proj_false, region->in(i))) {
1814 continue;
1815 }
1816 return false;
1817 }
1818
1819 return true;
1820 }
1821
1822 void ShenandoahBarrierC2Support::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) {
1823 assert(is_heap_stable_test(n), "no other tests");
1824 if (identical_backtoback_ifs(n, phase)) {
1825 Node* n_ctrl = n->in(0);
1826 if (phase->can_split_if(n_ctrl)) {
1827 IfNode* dom_if = phase->idom(n_ctrl)->as_If();
1828 if (is_heap_stable_test(n)) {
1829 Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1);
1830 assert(is_gc_state_load(gc_state_load), "broken");
1831 Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1);
1832 assert(is_gc_state_load(dom_gc_state_load), "broken");
1833 if (gc_state_load != dom_gc_state_load) {
1834 phase->igvn().replace_node(gc_state_load, dom_gc_state_load);
1835 }
1836 }
1837 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1838 Node* proj_true = dom_if->proj_out(1);
1839 Node* proj_false = dom_if->proj_out(0);
1840 Node* con_true = phase->igvn().makecon(TypeInt::ONE);
1841 Node* con_false = phase->igvn().makecon(TypeInt::ZERO);
1842
1843 for (uint i = 1; i < n_ctrl->req(); i++) {
1844 if (phase->is_dominator(proj_true, n_ctrl->in(i))) {
1845 bolphi->init_req(i, con_true);
1846 } else {
1847 assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1848 bolphi->init_req(i, con_false);
1849 }
1850 }
1851 phase->register_new_node(bolphi, n_ctrl);
1852 phase->igvn().replace_input_of(n, 1, bolphi);
1853 phase->do_split_if(n);
1854 }
1855 }
1856 }
1857
1858 IfNode* ShenandoahBarrierC2Support::find_unswitching_candidate(const IdealLoopTree* loop, PhaseIdealLoop* phase) {
1859 // Find first invariant test that doesn't exit the loop
1860 LoopNode *head = loop->_head->as_Loop();
1861 IfNode* unswitch_iff = NULL;
1862 Node* n = head->in(LoopNode::LoopBackControl);
1863 int loop_has_sfpts = -1;
1864 while (n != head) {
1865 Node* n_dom = phase->idom(n);
1866 if (n->is_Region()) {
1867 if (n_dom->is_If()) {
1868 IfNode* iff = n_dom->as_If();
1869 if (iff->in(1)->is_Bool()) {
1870 BoolNode* bol = iff->in(1)->as_Bool();
1871 if (bol->in(1)->is_Cmp()) {
1872 // If condition is invariant and not a loop exit,
1873 // then found reason to unswitch.
1874 if (is_heap_stable_test(iff) &&
1875 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) {
1876 assert(!loop->is_loop_exit(iff), "both branches should be in the loop");
1877 if (loop_has_sfpts == -1) {
1878 for(uint i = 0; i < loop->_body.size(); i++) {
1879 Node *m = loop->_body[i];
1880 if (m->is_SafePoint() && !m->is_CallLeaf()) {
1881 loop_has_sfpts = 1;
1882 break;
1883 }
1884 }
1885 if (loop_has_sfpts == -1) {
1886 loop_has_sfpts = 0;
1887 }
1888 }
1889 if (!loop_has_sfpts) {
1890 unswitch_iff = iff;
1891 }
1892 }
1893 }
1894 }
1895 }
1896 }
1897 n = n_dom;
1898 }
1899 return unswitch_iff;
1900 }
1901
1902
1903 void ShenandoahBarrierC2Support::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) {
1904 Node_List heap_stable_tests;
1905 Node_List gc_state_loads;
1906 stack.push(phase->C->start(), 0);
1907 do {
1908 Node* n = stack.node();
1909 uint i = stack.index();
1910
1911 if (i < n->outcnt()) {
1912 Node* u = n->raw_out(i);
1913 stack.set_index(i+1);
1914 if (!visited.test_set(u->_idx)) {
1915 stack.push(u, 0);
1916 }
1917 } else {
1918 stack.pop();
1919 if (ShenandoahCommonGCStateLoads && is_gc_state_load(n)) {
1920 gc_state_loads.push(n);
1921 }
1922 if (n->is_If() && is_heap_stable_test(n)) {
1923 heap_stable_tests.push(n);
1924 }
1925 }
1926 } while (stack.size() > 0);
1927
1928 bool progress;
1929 do {
1930 progress = false;
1931 for (uint i = 0; i < gc_state_loads.size(); i++) {
1932 Node* n = gc_state_loads.at(i);
1933 if (n->outcnt() != 0) {
1934 progress |= try_common_gc_state_load(n, phase);
1935 }
1936 }
1937 } while (progress);
1938
1939 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1940 Node* n = heap_stable_tests.at(i);
1941 assert(is_heap_stable_test(n), "only evacuation test");
1942 merge_back_to_back_tests(n, phase);
1943 }
1944
1945 if (!phase->C->major_progress()) {
1946 VectorSet seen(Thread::current()->resource_area());
1947 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1948 Node* n = heap_stable_tests.at(i);
1949 IdealLoopTree* loop = phase->get_loop(n);
1950 if (loop != phase->ltree_root() &&
1951 loop->_child == NULL &&
1952 !loop->_irreducible) {
1953 LoopNode* head = loop->_head->as_Loop();
1954 if ((!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) &&
1955 !seen.test_set(head->_idx)) {
1956 IfNode* iff = find_unswitching_candidate(loop, phase);
1957 if (iff != NULL) {
1958 Node* bol = iff->in(1);
1959 if (head->is_strip_mined()) {
1960 head->verify_strip_mined(0);
1961 }
1962 move_heap_stable_test_out_of_loop(iff, phase);
1963 if (loop->policy_unswitching(phase)) {
1964 if (head->is_strip_mined()) {
1965 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop();
1966 hide_strip_mined_loop(outer, head->as_CountedLoop(), phase);
1967 }
1968 phase->do_unswitching(loop, old_new);
1969 } else {
1970 // Not proceeding with unswitching. Move load back in
1971 // the loop.
1972 phase->igvn().replace_input_of(iff, 1, bol);
1973 }
1974 }
1975 }
1976 }
1977 }
1978 }
1979 }
1980
1981 #ifdef ASSERT
1982 void ShenandoahBarrierC2Support::verify_raw_mem(RootNode* root) {
1983 const bool trace = false;
1984 ResourceMark rm;
1985 Unique_Node_List nodes;
1986 Unique_Node_List controls;
1987 Unique_Node_List memories;
1988
1989 nodes.push(root);
1990 for (uint next = 0; next < nodes.size(); next++) {
1991 Node *n = nodes.at(next);
1992 if (ShenandoahBarrierSetC2::is_shenandoah_wb_call(n)) {
1993 controls.push(n);
1994 if (trace) { tty->print("XXXXXX verifying"); n->dump(); }
1995 for (uint next2 = 0; next2 < controls.size(); next2++) {
1996 Node *m = controls.at(next2);
1997 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
1998 Node* u = m->fast_out(i);
1999 if (u->is_CFG() && !u->is_Root() &&
2000 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1) &&
2001 !(u->is_Region() && u->unique_ctrl_out()->Opcode() == Op_Halt)) {
2002 if (trace) { tty->print("XXXXXX pushing control"); u->dump(); }
2003 controls.push(u);
2004 }
2005 }
2006 }
2007 memories.push(n->as_Call()->proj_out(TypeFunc::Memory));
2008 for (uint next2 = 0; next2 < memories.size(); next2++) {
2009 Node *m = memories.at(next2);
2010 assert(m->bottom_type() == Type::MEMORY, "");
2011 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
2012 Node* u = m->fast_out(i);
2013 if (u->bottom_type() == Type::MEMORY && (u->is_Mem() || u->is_ClearArray())) {
2014 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
2015 memories.push(u);
2016 } else if (u->is_LoadStore()) {
2017 if (trace) { tty->print("XXXXXX pushing memory"); u->find_out_with(Op_SCMemProj)->dump(); }
2018 memories.push(u->find_out_with(Op_SCMemProj));
2019 } else if (u->is_MergeMem() && u->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == m) {
2020 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
2021 memories.push(u);
2022 } else if (u->is_Phi()) {
2023 assert(u->bottom_type() == Type::MEMORY, "");
2024 if (u->adr_type() == TypeRawPtr::BOTTOM || u->adr_type() == TypePtr::BOTTOM) {
2025 assert(controls.member(u->in(0)), "");
2026 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
2027 memories.push(u);
2028 }
2029 } else if (u->is_SafePoint() || u->is_MemBar()) {
2030 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2031 Node* uu = u->fast_out(j);
2032 if (uu->bottom_type() == Type::MEMORY) {
2033 if (trace) { tty->print("XXXXXX pushing memory"); uu->dump(); }
2034 memories.push(uu);
2035 }
2036 }
2037 }
2038 }
2039 }
2040 for (uint next2 = 0; next2 < controls.size(); next2++) {
2041 Node *m = controls.at(next2);
2042 if (m->is_Region()) {
2043 bool all_in = true;
2044 for (uint i = 1; i < m->req(); i++) {
2045 if (!controls.member(m->in(i))) {
2046 all_in = false;
2047 break;
2048 }
2049 }
2050 if (trace) { tty->print("XXX verifying %s", all_in ? "all in" : ""); m->dump(); }
2051 bool found_phi = false;
2052 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax && !found_phi; j++) {
2053 Node* u = m->fast_out(j);
2054 if (u->is_Phi() && memories.member(u)) {
2055 found_phi = true;
2056 for (uint i = 1; i < u->req() && found_phi; i++) {
2057 Node* k = u->in(i);
2058 if (memories.member(k) != controls.member(m->in(i))) {
2059 found_phi = false;
2060 }
2061 }
2062 }
2063 }
2064 assert(found_phi || all_in, "");
2065 }
2066 }
2067 controls.clear();
2068 memories.clear();
2069 }
2070 for( uint i = 0; i < n->len(); ++i ) {
2071 Node *m = n->in(i);
2072 if (m != NULL) {
2073 nodes.push(m);
2074 }
2075 }
2076 }
2077 }
2078 #endif
2079
2080 ShenandoahEnqueueBarrierNode::ShenandoahEnqueueBarrierNode(Node* val) : Node(NULL, val) {
2081 ShenandoahBarrierSetC2::bsc2()->state()->add_enqueue_barrier(this);
2082 }
2083
2084 const Type* ShenandoahEnqueueBarrierNode::bottom_type() const {
2085 if (in(1) == NULL || in(1)->is_top()) {
2086 return Type::TOP;
2087 }
2088 const Type* t = in(1)->bottom_type();
2089 if (t == TypePtr::NULL_PTR) {
2090 return t;
2091 }
2092 return t->is_oopptr()->cast_to_nonconst();
2093 }
2094
2095 const Type* ShenandoahEnqueueBarrierNode::Value(PhaseGVN* phase) const {
2096 if (in(1) == NULL) {
2097 return Type::TOP;
2098 }
2099 const Type* t = phase->type(in(1));
2100 if (t == Type::TOP) {
2101 return Type::TOP;
2102 }
2103 if (t == TypePtr::NULL_PTR) {
2104 return t;
2105 }
2106 return t->is_oopptr()->cast_to_nonconst();
2107 }
2108
2109 int ShenandoahEnqueueBarrierNode::needed(Node* n) {
2110 if (n == NULL ||
2111 n->is_Allocate() ||
2112 n->Opcode() == Op_ShenandoahEnqueueBarrier ||
2113 n->bottom_type() == TypePtr::NULL_PTR ||
2114 (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL)) {
2115 return NotNeeded;
2116 }
2117 if (n->is_Phi() ||
2118 n->is_CMove()) {
2119 return MaybeNeeded;
2120 }
2121 return Needed;
2122 }
2123
2124 Node* ShenandoahEnqueueBarrierNode::next(Node* n) {
2125 for (;;) {
2126 if (n == NULL) {
2127 return n;
2128 } else if (n->bottom_type() == TypePtr::NULL_PTR) {
2129 return n;
2130 } else if (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL) {
2131 return n;
2132 } else if (n->is_ConstraintCast() ||
2133 n->Opcode() == Op_DecodeN ||
2134 n->Opcode() == Op_EncodeP) {
2135 n = n->in(1);
2136 } else if (n->is_Proj()) {
2137 n = n->in(0);
2138 } else {
2139 return n;
2140 }
2141 }
2142 ShouldNotReachHere();
2143 return NULL;
2144 }
2145
2146 Node* ShenandoahEnqueueBarrierNode::Identity(PhaseGVN* phase) {
2147 PhaseIterGVN* igvn = phase->is_IterGVN();
2148
2149 Node* n = next(in(1));
2150
2151 int cont = needed(n);
2152
2153 if (cont == NotNeeded) {
2154 return in(1);
2155 } else if (cont == MaybeNeeded) {
2156 if (igvn == NULL) {
2157 phase->record_for_igvn(this);
2158 return this;
2159 } else {
2160 ResourceMark rm;
2161 Unique_Node_List wq;
2162 uint wq_i = 0;
2163
2164 for (;;) {
2165 if (n->is_Phi()) {
2166 for (uint i = 1; i < n->req(); i++) {
2167 Node* m = n->in(i);
2168 if (m != NULL) {
2169 wq.push(m);
2170 }
2171 }
2172 } else {
2173 assert(n->is_CMove(), "nothing else here");
2174 Node* m = n->in(CMoveNode::IfFalse);
2175 wq.push(m);
2176 m = n->in(CMoveNode::IfTrue);
2177 wq.push(m);
2178 }
2179 Node* orig_n = NULL;
2180 do {
2181 if (wq_i >= wq.size()) {
2182 return in(1);
2183 }
2184 n = wq.at(wq_i);
2185 wq_i++;
2186 orig_n = n;
2187 n = next(n);
2188 cont = needed(n);
2189 if (cont == Needed) {
2190 return this;
2191 }
2192 } while (cont != MaybeNeeded || (orig_n != n && wq.member(n)));
2193 }
2194 }
2195 }
2196
2197 return this;
2198 }
2199
2200 #ifdef ASSERT
2201 static bool has_never_branch(Node* root) {
2202 for (uint i = 1; i < root->req(); i++) {
2203 Node* in = root->in(i);
2204 if (in != NULL && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->Opcode() == Op_NeverBranch) {
2205 return true;
2206 }
2207 }
2208 return false;
2209 }
2210 #endif
2211
2212 void MemoryGraphFixer::collect_memory_nodes() {
2213 Node_Stack stack(0);
2214 VectorSet visited(Thread::current()->resource_area());
2215 Node_List regions;
2216
2217 // Walk the raw memory graph and create a mapping from CFG node to
2218 // memory node. Exclude phis for now.
2219 stack.push(_phase->C->root(), 1);
2220 do {
2221 Node* n = stack.node();
2222 int opc = n->Opcode();
2223 uint i = stack.index();
2224 if (i < n->req()) {
2225 Node* mem = NULL;
2226 if (opc == Op_Root) {
2227 Node* in = n->in(i);
2228 int in_opc = in->Opcode();
2229 if (in_opc == Op_Return || in_opc == Op_Rethrow) {
2230 mem = in->in(TypeFunc::Memory);
2231 } else if (in_opc == Op_Halt) {
2232 if (!in->in(0)->is_Region()) {
2233 Node* proj = in->in(0);
2234 assert(proj->is_Proj(), "");
2235 Node* in = proj->in(0);
2236 assert(in->is_CallStaticJava() || in->Opcode() == Op_NeverBranch || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
2237 if (in->is_CallStaticJava()) {
2238 mem = in->in(TypeFunc::Memory);
2239 } else if (in->Opcode() == Op_Catch) {
2240 Node* call = in->in(0)->in(0);
2241 assert(call->is_Call(), "");
2242 mem = call->in(TypeFunc::Memory);
2243 } else if (in->Opcode() == Op_NeverBranch) {
2244 ResourceMark rm;
2245 Unique_Node_List wq;
2246 wq.push(in);
2247 wq.push(in->as_Multi()->proj_out(0));
2248 for (uint j = 1; j < wq.size(); j++) {
2249 Node* c = wq.at(j);
2250 assert(!c->is_Root(), "shouldn't leave loop");
2251 if (c->is_SafePoint()) {
2252 assert(mem == NULL, "only one safepoint");
2253 mem = c->in(TypeFunc::Memory);
2254 }
2255 for (DUIterator_Fast kmax, k = c->fast_outs(kmax); k < kmax; k++) {
2256 Node* u = c->fast_out(k);
2257 if (u->is_CFG()) {
2258 wq.push(u);
2259 }
2260 }
2261 }
2262 assert(mem != NULL, "should have found safepoint");
2263 }
2264 }
2265 } else {
2266 #ifdef ASSERT
2267 n->dump();
2268 in->dump();
2269 #endif
2270 ShouldNotReachHere();
2271 }
2272 } else {
2273 assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
2274 assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
2275 mem = n->in(i);
2276 }
2277 i++;
2278 stack.set_index(i);
2279 if (mem == NULL) {
2280 continue;
2281 }
2282 for (;;) {
2283 if (visited.test_set(mem->_idx) || mem->is_Start()) {
2284 break;
2285 }
2286 if (mem->is_Phi()) {
2287 stack.push(mem, 2);
2288 mem = mem->in(1);
2289 } else if (mem->is_Proj()) {
2290 stack.push(mem, mem->req());
2291 mem = mem->in(0);
2292 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
2293 mem = mem->in(TypeFunc::Memory);
2294 } else if (mem->is_MergeMem()) {
2295 MergeMemNode* mm = mem->as_MergeMem();
2296 mem = mm->memory_at(_alias);
2297 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
2298 assert(_alias == Compile::AliasIdxRaw, "");
2299 stack.push(mem, mem->req());
2300 mem = mem->in(MemNode::Memory);
2301 } else {
2302 #ifdef ASSERT
2303 mem->dump();
2304 #endif
2305 ShouldNotReachHere();
2306 }
2307 }
2308 } else {
2309 if (n->is_Phi()) {
2310 // Nothing
2311 } else if (!n->is_Root()) {
2312 Node* c = get_ctrl(n);
2313 _memory_nodes.map(c->_idx, n);
2314 }
2315 stack.pop();
2316 }
2317 } while(stack.is_nonempty());
2318
2319 // Iterate over CFG nodes in rpo and propagate memory state to
2320 // compute memory state at regions, creating new phis if needed.
2321 Node_List rpo_list;
2322 visited.Clear();
2323 _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
2324 Node* root = rpo_list.pop();
2325 assert(root == _phase->C->root(), "");
2326
2327 const bool trace = false;
2328 #ifdef ASSERT
2329 if (trace) {
2330 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2331 Node* c = rpo_list.at(i);
2332 if (_memory_nodes[c->_idx] != NULL) {
2333 tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump();
2334 }
2335 }
2336 }
2337 #endif
2338 uint last = _phase->C->unique();
2339
2340 #ifdef ASSERT
2341 uint8_t max_depth = 0;
2342 for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
2343 IdealLoopTree* lpt = iter.current();
2344 max_depth = MAX2(max_depth, lpt->_nest);
2345 }
2346 #endif
2347
2348 bool progress = true;
2349 int iteration = 0;
2350 Node_List dead_phis;
2351 while (progress) {
2352 progress = false;
2353 iteration++;
2354 assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2355 if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
2356 IdealLoopTree* last_updated_ilt = NULL;
2357 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2358 Node* c = rpo_list.at(i);
2359
2360 Node* prev_mem = _memory_nodes[c->_idx];
2361 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2362 Node* prev_region = regions[c->_idx];
2363 Node* unique = NULL;
2364 for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
2365 Node* m = _memory_nodes[c->in(j)->_idx];
2366 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");
2367 if (m != NULL) {
2368 if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
2369 assert(c->is_Loop() && j == LoopNode::LoopBackControl || _phase->C->has_irreducible_loop(), "");
2370 // continue
2371 } else if (unique == NULL) {
2372 unique = m;
2373 } else if (m == unique) {
2374 // continue
2375 } else {
2376 unique = NodeSentinel;
2377 }
2378 }
2379 }
2380 assert(unique != NULL, "empty phi???");
2381 if (unique != NodeSentinel) {
2382 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c) {
2383 dead_phis.push(prev_region);
2384 }
2385 regions.map(c->_idx, unique);
2386 } else {
2387 Node* phi = NULL;
2388 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
2389 phi = prev_region;
2390 for (uint k = 1; k < c->req(); k++) {
2391 Node* m = _memory_nodes[c->in(k)->_idx];
2392 assert(m != NULL, "expect memory state");
2393 phi->set_req(k, m);
2394 }
2395 } else {
2396 for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == NULL; j++) {
2397 Node* u = c->fast_out(j);
2398 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2399 (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
2400 phi = u;
2401 for (uint k = 1; k < c->req() && phi != NULL; k++) {
2402 Node* m = _memory_nodes[c->in(k)->_idx];
2403 assert(m != NULL, "expect memory state");
2404 if (u->in(k) != m) {
2405 phi = NULL;
2406 }
2407 }
2408 }
2409 }
2410 if (phi == NULL) {
2411 phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
2412 for (uint k = 1; k < c->req(); k++) {
2413 Node* m = _memory_nodes[c->in(k)->_idx];
2414 assert(m != NULL, "expect memory state");
2415 phi->init_req(k, m);
2416 }
2417 }
2418 }
2419 assert(phi != NULL, "");
2420 regions.map(c->_idx, phi);
2421 }
2422 Node* current_region = regions[c->_idx];
2423 if (current_region != prev_region) {
2424 progress = true;
2425 if (prev_region == prev_mem) {
2426 _memory_nodes.map(c->_idx, current_region);
2427 }
2428 }
2429 } else if (prev_mem == NULL || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
2430 Node* m = _memory_nodes[_phase->idom(c)->_idx];
2431 assert(m != NULL, "expect memory state");
2432 if (m != prev_mem) {
2433 _memory_nodes.map(c->_idx, m);
2434 progress = true;
2435 }
2436 }
2437 #ifdef ASSERT
2438 if (trace) { tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); }
2439 #endif
2440 }
2441 }
2442
2443 // Replace existing phi with computed memory state for that region
2444 // if different (could be a new phi or a dominating memory node if
2445 // that phi was found to be useless).
2446 while (dead_phis.size() > 0) {
2447 Node* n = dead_phis.pop();
2448 n->replace_by(_phase->C->top());
2449 n->destruct();
2450 }
2451 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2452 Node* c = rpo_list.at(i);
2453 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2454 Node* n = regions[c->_idx];
2455 if (n->is_Phi() && n->_idx >= last && n->in(0) == c) {
2456 _phase->register_new_node(n, c);
2457 }
2458 }
2459 }
2460 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2461 Node* c = rpo_list.at(i);
2462 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2463 Node* n = regions[c->_idx];
2464 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
2465 Node* u = c->fast_out(i);
2466 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2467 u != n) {
2468 if (u->adr_type() == TypePtr::BOTTOM) {
2469 fix_memory_uses(u, n, n, c);
2470 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2471 _phase->lazy_replace(u, n);
2472 --i; --imax;
2473 }
2474 }
2475 }
2476 }
2477 }
2478 }
2479
2480 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
2481 Node* c = _phase->get_ctrl(n);
2482 if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Call()) {
2483 assert(c == n->in(0), "");
2484 CallNode* call = c->as_Call();
2485 CallProjections projs;
2486 call->extract_projections(&projs, true, false);
2487 if (projs.catchall_memproj != NULL) {
2488 if (projs.fallthrough_memproj == n) {
2489 c = projs.fallthrough_catchproj;
2490 } else {
2491 assert(projs.catchall_memproj == n, "");
2492 c = projs.catchall_catchproj;
2493 }
2494 }
2495 }
2496 return c;
2497 }
2498
2499 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
2500 if (_phase->has_ctrl(n))
2501 return get_ctrl(n);
2502 else {
2503 assert (n->is_CFG(), "must be a CFG node");
2504 return n;
2505 }
2506 }
2507
2508 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
2509 return m != NULL && get_ctrl(m) == c;
2510 }
2511
2512 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
2513 assert(n == NULL || _phase->ctrl_or_self(n) == ctrl, "");
2514 Node* mem = _memory_nodes[ctrl->_idx];
2515 Node* c = ctrl;
2516 while (!mem_is_valid(mem, c) &&
2517 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
2518 c = _phase->idom(c);
2519 mem = _memory_nodes[c->_idx];
2520 }
2521 if (n != NULL && mem_is_valid(mem, c)) {
2522 while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
2523 mem = next_mem(mem, _alias);
2524 }
2525 if (mem->is_MergeMem()) {
2526 mem = mem->as_MergeMem()->memory_at(_alias);
2527 }
2528 if (!mem_is_valid(mem, c)) {
2529 do {
2530 c = _phase->idom(c);
2531 mem = _memory_nodes[c->_idx];
2532 } while (!mem_is_valid(mem, c) &&
2533 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
2534 }
2535 }
2536 assert(mem->bottom_type() == Type::MEMORY, "");
2537 return mem;
2538 }
2539
2540 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
2541 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2542 Node* use = region->fast_out(i);
2543 if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
2544 (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
2545 return true;
2546 }
2547 }
2548 return false;
2549 }
2550
2551 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
2552 assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
2553 const bool trace = false;
2554 DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
2555 DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
2556 GrowableArray<Node*> phis;
2557 if (mem_for_ctrl != mem) {
2558 Node* old = mem_for_ctrl;
2559 Node* prev = NULL;
2560 while (old != mem) {
2561 prev = old;
2562 if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
2563 assert(_alias == Compile::AliasIdxRaw, "");
2564 old = old->in(MemNode::Memory);
2565 } else if (old->Opcode() == Op_SCMemProj) {
2566 assert(_alias == Compile::AliasIdxRaw, "");
2567 old = old->in(0);
2568 } else {
2569 ShouldNotReachHere();
2570 }
2571 }
2572 assert(prev != NULL, "");
2573 if (new_ctrl != ctrl) {
2574 _memory_nodes.map(ctrl->_idx, mem);
2575 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2576 }
2577 uint input = (uint)MemNode::Memory;
2578 _phase->igvn().replace_input_of(prev, input, new_mem);
2579 } else {
2580 uses.clear();
2581 _memory_nodes.map(new_ctrl->_idx, new_mem);
2582 uses.push(new_ctrl);
2583 for(uint next = 0; next < uses.size(); next++ ) {
2584 Node *n = uses.at(next);
2585 assert(n->is_CFG(), "");
2586 DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
2587 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2588 Node* u = n->fast_out(i);
2589 if (!u->is_Root() && u->is_CFG() && u != n) {
2590 Node* m = _memory_nodes[u->_idx];
2591 if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
2592 !has_mem_phi(u) &&
2593 u->unique_ctrl_out()->Opcode() != Op_Halt) {
2594 DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
2595 DEBUG_ONLY(if (trace && m != NULL) { tty->print("ZZZ mem"); m->dump(); });
2596
2597 if (!mem_is_valid(m, u) || !m->is_Phi()) {
2598 bool push = true;
2599 bool create_phi = true;
2600 if (_phase->is_dominator(new_ctrl, u)) {
2601 create_phi = false;
2602 } else if (!_phase->C->has_irreducible_loop()) {
2603 IdealLoopTree* loop = _phase->get_loop(ctrl);
2604 bool do_check = true;
2605 IdealLoopTree* l = loop;
2606 create_phi = false;
2607 while (l != _phase->ltree_root()) {
2608 if (_phase->is_dominator(l->_head, u) && _phase->is_dominator(_phase->idom(u), l->_head)) {
2609 create_phi = true;
2610 do_check = false;
2611 break;
2612 }
2613 l = l->_parent;
2614 }
2615
2616 if (do_check) {
2617 assert(!create_phi, "");
2618 IdealLoopTree* u_loop = _phase->get_loop(u);
2619 if (u_loop != _phase->ltree_root() && u_loop->is_member(loop)) {
2620 Node* c = ctrl;
2621 while (!_phase->is_dominator(c, u_loop->tail())) {
2622 c = _phase->idom(c);
2623 }
2624 if (!_phase->is_dominator(c, u)) {
2625 do_check = false;
2626 }
2627 }
2628 }
2629
2630 if (do_check && _phase->is_dominator(_phase->idom(u), new_ctrl)) {
2631 create_phi = true;
2632 }
2633 }
2634 if (create_phi) {
2635 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
2636 _phase->register_new_node(phi, u);
2637 phis.push(phi);
2638 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
2639 if (!mem_is_valid(m, u)) {
2640 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
2641 _memory_nodes.map(u->_idx, phi);
2642 } else {
2643 DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
2644 for (;;) {
2645 assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), "");
2646 Node* next = NULL;
2647 if (m->is_Proj()) {
2648 next = m->in(0);
2649 } else {
2650 assert(m->is_Mem() || m->is_LoadStore(), "");
2651 assert(_alias == Compile::AliasIdxRaw, "");
2652 next = m->in(MemNode::Memory);
2653 }
2654 if (_phase->get_ctrl(next) != u) {
2655 break;
2656 }
2657 if (next->is_MergeMem()) {
2658 assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
2659 break;
2660 }
2661 if (next->is_Phi()) {
2662 assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
2663 break;
2664 }
2665 m = next;
2666 }
2667
2668 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
2669 assert(m->is_Mem() || m->is_LoadStore(), "");
2670 uint input = (uint)MemNode::Memory;
2671 _phase->igvn().replace_input_of(m, input, phi);
2672 push = false;
2673 }
2674 } else {
2675 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
2676 }
2677 if (push) {
2678 uses.push(u);
2679 }
2680 }
2681 } else if (!mem_is_valid(m, u) &&
2682 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1)) {
2683 uses.push(u);
2684 }
2685 }
2686 }
2687 }
2688 for (int i = 0; i < phis.length(); i++) {
2689 Node* n = phis.at(i);
2690 Node* r = n->in(0);
2691 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
2692 for (uint j = 1; j < n->req(); j++) {
2693 Node* m = find_mem(r->in(j), NULL);
2694 _phase->igvn().replace_input_of(n, j, m);
2695 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
2696 }
2697 }
2698 }
2699 uint last = _phase->C->unique();
2700 MergeMemNode* mm = NULL;
2701 int alias = _alias;
2702 DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
2703 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2704 Node* u = mem->out(i);
2705 if (u->_idx < last) {
2706 if (u->is_Mem()) {
2707 if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2708 Node* m = find_mem(_phase->get_ctrl(u), u);
2709 if (m != mem) {
2710 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2711 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2712 --i;
2713 }
2714 }
2715 } else if (u->is_MergeMem()) {
2716 MergeMemNode* u_mm = u->as_MergeMem();
2717 if (u_mm->memory_at(alias) == mem) {
2718 MergeMemNode* newmm = NULL;
2719 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2720 Node* uu = u->fast_out(j);
2721 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2722 if (uu->is_Phi()) {
2723 assert(uu->adr_type() == TypePtr::BOTTOM, "");
2724 Node* region = uu->in(0);
2725 int nb = 0;
2726 for (uint k = 1; k < uu->req(); k++) {
2727 if (uu->in(k) == u) {
2728 Node* m = find_mem(region->in(k), NULL);
2729 if (m != mem) {
2730 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
2731 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2732 if (newmm != u) {
2733 _phase->igvn().replace_input_of(uu, k, newmm);
2734 nb++;
2735 --jmax;
2736 }
2737 }
2738 }
2739 }
2740 if (nb > 0) {
2741 --j;
2742 }
2743 } else {
2744 Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
2745 if (m != mem) {
2746 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
2747 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2748 if (newmm != u) {
2749 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2750 --j, --jmax;
2751 }
2752 }
2753 }
2754 }
2755 }
2756 } else if (u->is_Phi()) {
2757 assert(u->bottom_type() == Type::MEMORY, "what else?");
2758 if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
2759 Node* region = u->in(0);
2760 bool replaced = false;
2761 for (uint j = 1; j < u->req(); j++) {
2762 if (u->in(j) == mem) {
2763 Node* m = find_mem(region->in(j), NULL);
2764 Node* nnew = m;
2765 if (m != mem) {
2766 if (u->adr_type() == TypePtr::BOTTOM) {
2767 mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
2768 nnew = mm;
2769 }
2770 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
2771 _phase->igvn().replace_input_of(u, j, nnew);
2772 replaced = true;
2773 }
2774 }
2775 }
2776 if (replaced) {
2777 --i;
2778 }
2779 }
2780 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2781 u->adr_type() == NULL) {
2782 assert(u->adr_type() != NULL ||
2783 u->Opcode() == Op_Rethrow ||
2784 u->Opcode() == Op_Return ||
2785 u->Opcode() == Op_SafePoint ||
2786 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2787 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2788 u->Opcode() == Op_CallLeaf, "");
2789 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2790 if (m != mem) {
2791 mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
2792 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2793 --i;
2794 }
2795 } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2796 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2797 if (m != mem) {
2798 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2799 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2800 --i;
2801 }
2802 } else if (u->adr_type() != TypePtr::BOTTOM &&
2803 _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
2804 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2805 assert(m != mem, "");
2806 // u is on the wrong slice...
2807 assert(u->is_ClearArray(), "");
2808 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2809 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2810 --i;
2811 }
2812 }
2813 }
2814 #ifdef ASSERT
2815 assert(new_mem->outcnt() > 0, "");
2816 for (int i = 0; i < phis.length(); i++) {
2817 Node* n = phis.at(i);
2818 assert(n->outcnt() > 0, "new phi must have uses now");
2819 }
2820 #endif
2821 }
2822
2823 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
2824 MergeMemNode* mm = MergeMemNode::make(mem);
2825 mm->set_memory_at(_alias, rep_proj);
2826 _phase->register_new_node(mm, rep_ctrl);
2827 return mm;
2828 }
2829
2830 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
2831 MergeMemNode* newmm = NULL;
2832 MergeMemNode* u_mm = u->as_MergeMem();
2833 Node* c = _phase->get_ctrl(u);
2834 if (_phase->is_dominator(c, rep_ctrl)) {
2835 c = rep_ctrl;
2836 } else {
2837 assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
2838 }
2839 if (u->outcnt() == 1) {
2840 if (u->req() > (uint)_alias && u->in(_alias) == mem) {
2841 _phase->igvn().replace_input_of(u, _alias, rep_proj);
2842 --i;
2843 } else {
2844 _phase->igvn().rehash_node_delayed(u);
2845 u_mm->set_memory_at(_alias, rep_proj);
2846 }
2847 newmm = u_mm;
2848 _phase->set_ctrl_and_loop(u, c);
2849 } else {
2850 // can't simply clone u and then change one of its input because
2851 // it adds and then removes an edge which messes with the
2852 // DUIterator
2853 newmm = MergeMemNode::make(u_mm->base_memory());
2854 for (uint j = 0; j < u->req(); j++) {
2855 if (j < newmm->req()) {
2856 if (j == (uint)_alias) {
2857 newmm->set_req(j, rep_proj);
2858 } else if (newmm->in(j) != u->in(j)) {
2859 newmm->set_req(j, u->in(j));
2860 }
2861 } else if (j == (uint)_alias) {
2862 newmm->add_req(rep_proj);
2863 } else {
2864 newmm->add_req(u->in(j));
2865 }
2866 }
2867 if ((uint)_alias >= u->req()) {
2868 newmm->set_memory_at(_alias, rep_proj);
2869 }
2870 _phase->register_new_node(newmm, c);
2871 }
2872 return newmm;
2873 }
2874
2875 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
2876 if (phi->adr_type() == TypePtr::BOTTOM) {
2877 Node* region = phi->in(0);
2878 for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
2879 Node* uu = region->fast_out(j);
2880 if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
2881 return false;
2882 }
2883 }
2884 return true;
2885 }
2886 return _phase->C->get_alias_index(phi->adr_type()) == _alias;
2887 }
2888
2889 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
2890 uint last = _phase-> C->unique();
2891 MergeMemNode* mm = NULL;
2892 assert(mem->bottom_type() == Type::MEMORY, "");
2893 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2894 Node* u = mem->out(i);
2895 if (u != replacement && u->_idx < last) {
2896 if (u->is_MergeMem()) {
2897 MergeMemNode* u_mm = u->as_MergeMem();
2898 if (u_mm->memory_at(_alias) == mem) {
2899 MergeMemNode* newmm = NULL;
2900 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2901 Node* uu = u->fast_out(j);
2902 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2903 if (uu->is_Phi()) {
2904 if (should_process_phi(uu)) {
2905 Node* region = uu->in(0);
2906 int nb = 0;
2907 for (uint k = 1; k < uu->req(); k++) {
2908 if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
2909 if (newmm == NULL) {
2910 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2911 }
2912 if (newmm != u) {
2913 _phase->igvn().replace_input_of(uu, k, newmm);
2914 nb++;
2915 --jmax;
2916 }
2917 }
2918 }
2919 if (nb > 0) {
2920 --j;
2921 }
2922 }
2923 } else {
2924 if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
2925 if (newmm == NULL) {
2926 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2927 }
2928 if (newmm != u) {
2929 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2930 --j, --jmax;
2931 }
2932 }
2933 }
2934 }
2935 }
2936 } else if (u->is_Phi()) {
2937 assert(u->bottom_type() == Type::MEMORY, "what else?");
2938 Node* region = u->in(0);
2939 if (should_process_phi(u)) {
2940 bool replaced = false;
2941 for (uint j = 1; j < u->req(); j++) {
2942 if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
2943 Node* nnew = rep_proj;
2944 if (u->adr_type() == TypePtr::BOTTOM) {
2945 if (mm == NULL) {
2946 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2947 }
2948 nnew = mm;
2949 }
2950 _phase->igvn().replace_input_of(u, j, nnew);
2951 replaced = true;
2952 }
2953 }
2954 if (replaced) {
2955 --i;
2956 }
2957
2958 }
2959 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2960 u->adr_type() == NULL) {
2961 assert(u->adr_type() != NULL ||
2962 u->Opcode() == Op_Rethrow ||
2963 u->Opcode() == Op_Return ||
2964 u->Opcode() == Op_SafePoint ||
2965 u->Opcode() == Op_StoreLConditional ||
2966 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2967 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2968 u->Opcode() == Op_CallLeaf, "");
2969 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2970 if (mm == NULL) {
2971 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2972 }
2973 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2974 --i;
2975 }
2976 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2977 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2978 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
2979 --i;
2980 }
2981 }
2982 }
2983 }
2984 }
2985
2986 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj)
2987 : Node(ctrl, obj) {
2988 ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this);
2989 }
2990
2991 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const {
2992 if (in(ValueIn) == NULL || in(ValueIn)->is_top()) {
2993 return Type::TOP;
2994 }
2995 const Type* t = in(ValueIn)->bottom_type();
2996 if (t == TypePtr::NULL_PTR) {
2997 return t;
2998 }
2999 return t->is_oopptr();
3000 }
3001
3002 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const {
3003 // Either input is TOP ==> the result is TOP
3004 const Type *t2 = phase->type(in(ValueIn));
3005 if( t2 == Type::TOP ) return Type::TOP;
3006
3007 if (t2 == TypePtr::NULL_PTR) {
3008 return t2;
3009 }
3010
3011 const Type* type = t2->is_oopptr()/*->cast_to_nonconst()*/;
3012 return type;
3013 }
3014
3015 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) {
3016 Node* value = in(ValueIn);
3017 if (!needs_barrier(phase, value)) {
3018 return value;
3019 }
3020 return this;
3021 }
3022
3023 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) {
3024 Unique_Node_List visited;
3025 return needs_barrier_impl(phase, n, visited);
3026 }
3027
3028 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) {
3029 if (n == NULL) return false;
3030 if (visited.member(n)) {
3031 return false; // Been there.
3032 }
3033 visited.push(n);
3034
3035 if (n->is_Allocate()) {
3036 // tty->print_cr("optimize barrier on alloc");
3037 return false;
3038 }
3039 if (n->is_Call()) {
3040 // tty->print_cr("optimize barrier on call");
3041 return false;
3042 }
3043
3044 const Type* type = phase->type(n);
3045 if (type == Type::TOP) {
3046 return false;
3047 }
3048 if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) {
3049 // tty->print_cr("optimize barrier on null");
3050 return false;
3051 }
3052 if (type->make_oopptr() && type->make_oopptr()->const_oop() != NULL) {
3053 // tty->print_cr("optimize barrier on constant");
3054 return false;
3055 }
3056
3057 switch (n->Opcode()) {
3058 case Op_AddP:
3059 return true; // TODO: Can refine?
3060 case Op_LoadP:
3061 case Op_ShenandoahCompareAndExchangeN:
3062 case Op_ShenandoahCompareAndExchangeP:
3063 case Op_CompareAndExchangeN:
3064 case Op_CompareAndExchangeP:
3065 case Op_GetAndSetN:
3066 case Op_GetAndSetP:
3067 return true;
3068 case Op_Phi: {
3069 for (uint i = 1; i < n->req(); i++) {
3070 if (needs_barrier_impl(phase, n->in(i), visited)) return true;
3071 }
3072 return false;
3073 }
3074 case Op_CheckCastPP:
3075 case Op_CastPP:
3076 return needs_barrier_impl(phase, n->in(1), visited);
3077 case Op_Proj:
3078 return needs_barrier_impl(phase, n->in(0), visited);
3079 case Op_ShenandoahLoadReferenceBarrier:
3080 // tty->print_cr("optimize barrier on barrier");
3081 return false;
3082 case Op_Parm:
3083 // tty->print_cr("optimize barrier on input arg");
3084 return false;
3085 case Op_DecodeN:
3086 case Op_EncodeP:
3087 return needs_barrier_impl(phase, n->in(1), visited);
3088 case Op_LoadN:
3089 return true;
3090 case Op_CMoveP:
3091 return needs_barrier_impl(phase, n->in(2), visited) ||
3092 needs_barrier_impl(phase, n->in(3), visited);
3093 case Op_ShenandoahEnqueueBarrier:
3094 return needs_barrier_impl(phase, n->in(1), visited);
3095 default:
3096 break;
3097 }
3098 #ifdef ASSERT
3099 tty->print("need barrier on?: ");
3100 tty->print_cr("ins:");
3101 n->dump(2);
3102 tty->print_cr("outs:");
3103 n->dump(-2);
3104 ShouldNotReachHere();
3105 #endif
3106 return true;
3107 }
3108
3109 ShenandoahLoadReferenceBarrierNode::Strength ShenandoahLoadReferenceBarrierNode::get_barrier_strength() {
3110 Unique_Node_List visited;
3111 Node_Stack stack(0);
3112 stack.push(this, 0);
3113 Strength strength = NONE;
3114 while (strength != STRONG && stack.size() > 0) {
3115 Node* n = stack.node();
3116 if (visited.member(n)) {
3117 stack.pop();
3118 continue;
3119 }
3120 visited.push(n);
3121 bool visit_users = false;
3122 switch (n->Opcode()) {
3123 case Op_StoreN:
3124 case Op_StoreP: {
3125 strength = STRONG;
3126 break;
3127 }
3128 case Op_CmpP: {
3129 if (!n->in(1)->bottom_type()->higher_equal(TypePtr::NULL_PTR) &&
3130 !n->in(2)->bottom_type()->higher_equal(TypePtr::NULL_PTR)) {
3131 strength = STRONG;
3132 }
3133 break;
3134 }
3135 case Op_CallStaticJava: {
3136 strength = STRONG;
3137 break;
3138 }
3139 case Op_CallDynamicJava:
3140 case Op_CallLeaf:
3141 case Op_CallLeafNoFP:
3142 case Op_CompareAndSwapL:
3143 case Op_CompareAndSwapI:
3144 case Op_CompareAndSwapB:
3145 case Op_CompareAndSwapS:
3146 case Op_CompareAndSwapN:
3147 case Op_CompareAndSwapP:
3148 case Op_CompareAndExchangeL:
3149 case Op_CompareAndExchangeI:
3150 case Op_CompareAndExchangeB:
3151 case Op_CompareAndExchangeS:
3152 case Op_CompareAndExchangeN:
3153 case Op_CompareAndExchangeP:
3154 case Op_WeakCompareAndSwapL:
3155 case Op_WeakCompareAndSwapI:
3156 case Op_WeakCompareAndSwapB:
3157 case Op_WeakCompareAndSwapS:
3158 case Op_WeakCompareAndSwapN:
3159 case Op_WeakCompareAndSwapP:
3160 case Op_ShenandoahCompareAndSwapN:
3161 case Op_ShenandoahCompareAndSwapP:
3162 case Op_ShenandoahWeakCompareAndSwapN:
3163 case Op_ShenandoahWeakCompareAndSwapP:
3164 case Op_ShenandoahCompareAndExchangeN:
3165 case Op_ShenandoahCompareAndExchangeP:
3166 case Op_GetAndSetL:
3167 case Op_GetAndSetI:
3168 case Op_GetAndSetB:
3169 case Op_GetAndSetS:
3170 case Op_GetAndSetP:
3171 case Op_GetAndSetN:
3172 case Op_GetAndAddL:
3173 case Op_GetAndAddI:
3174 case Op_GetAndAddB:
3175 case Op_GetAndAddS:
3176 case Op_ShenandoahEnqueueBarrier:
3177 case Op_FastLock:
3178 case Op_FastUnlock:
3179 case Op_Rethrow:
3180 case Op_Return:
3181 case Op_StoreB:
3182 case Op_StoreC:
3183 case Op_StoreD:
3184 case Op_StoreF:
3185 case Op_StoreL:
3186 case Op_StoreLConditional:
3187 case Op_StoreI:
3188 case Op_StoreVector:
3189 case Op_StrInflatedCopy:
3190 case Op_StrCompressedCopy:
3191 case Op_EncodeP:
3192 case Op_CastP2X:
3193 case Op_SafePoint:
3194 case Op_EncodeISOArray:
3195 case Op_LoadKlass:
3196 case Op_LoadNKlass:
3197 strength = STRONG;
3198 break;
3199 case Op_LoadB:
3200 case Op_LoadUB:
3201 case Op_LoadUS:
3202 case Op_LoadD:
3203 case Op_LoadF:
3204 case Op_LoadL:
3205 case Op_LoadI:
3206 case Op_LoadS:
3207 case Op_LoadN:
3208 case Op_LoadP:
3209 case Op_LoadVector: {
3210 const TypePtr* adr_type = n->adr_type();
3211 int alias_idx = Compile::current()->get_alias_index(adr_type);
3212 Compile::AliasType* alias_type = Compile::current()->alias_type(alias_idx);
3213 ciField* field = alias_type->field();
3214 bool is_static = field != NULL && field->is_static();
3215 bool is_final = field != NULL && field->is_final();
3216 bool is_stable = field != NULL && field->is_stable();
3217 if (ShenandoahOptimizeStaticFinals && is_static && is_final) {
3218 // Leave strength as is.
3219 } else if (ShenandoahOptimizeInstanceFinals && !is_static && is_final) {
3220 // Leave strength as is.
3221 } else if (ShenandoahOptimizeStableFinals && (is_stable || (adr_type->isa_aryptr() && adr_type->isa_aryptr()->is_stable()))) {
3222 // Leave strength as is.
3223 } else {
3224 strength = WEAK;
3225 }
3226 break;
3227 }
3228 case Op_AryEq: {
3229 Node* n1 = n->in(2);
3230 Node* n2 = n->in(3);
3231 if (!ShenandoahOptimizeStableFinals ||
3232 !n1->bottom_type()->isa_aryptr() || !n1->bottom_type()->isa_aryptr()->is_stable() ||
3233 !n2->bottom_type()->isa_aryptr() || !n2->bottom_type()->isa_aryptr()->is_stable()) {
3234 strength = WEAK;
3235 }
3236 break;
3237 }
3238 case Op_StrEquals:
3239 case Op_StrComp:
3240 case Op_StrIndexOf:
3241 case Op_StrIndexOfChar:
3242 if (!ShenandoahOptimizeStableFinals) {
3243 strength = WEAK;
3244 }
3245 break;
3246 case Op_Conv2B:
3247 case Op_LoadRange:
3248 // NONE, i.e. leave current strength as is
3249 break;
3250 case Op_AddP:
3251 case Op_CheckCastPP:
3252 case Op_CastPP:
3253 case Op_CMoveP:
3254 case Op_Phi:
3255 case Op_ShenandoahLoadReferenceBarrier:
3256 visit_users = true;
3257 break;
3258 default: {
3259 #ifdef ASSERT
3260 tty->print_cr("Unknown node in get_barrier_strength:");
3261 n->dump(1);
3262 ShouldNotReachHere();
3263 #else
3264 strength = STRONG;
3265 #endif
3266 }
3267 }
3268 #ifdef ASSERT
3269 /*
3270 if (strength == STRONG) {
3271 tty->print("strengthening node: ");
3272 n->dump();
3273 }
3274 */
3275 #endif
3276 stack.pop();
3277 if (visit_users) {
3278 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3279 Node* user = n->fast_out(i);
3280 if (user != NULL) {
3281 stack.push(user, 0);
3282 }
3283 }
3284 }
3285 }
3286 return strength;
3287 }
3288
3289 CallStaticJavaNode* ShenandoahLoadReferenceBarrierNode::pin_and_expand_null_check(PhaseIterGVN& igvn) {
3290 Node* val = in(ValueIn);
3291
3292 const Type* val_t = igvn.type(val);
3293
3294 if (val_t->meet(TypePtr::NULL_PTR) != val_t &&
3295 val->Opcode() == Op_CastPP &&
3296 val->in(0) != NULL &&
3297 val->in(0)->Opcode() == Op_IfTrue &&
3298 val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
3299 val->in(0)->in(0)->is_If() &&
3300 val->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
3301 val->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
3302 val->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
3303 val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1) &&
3304 val->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
3305 assert(val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1), "");
3306 CallStaticJavaNode* unc = val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
3307 return unc;
3308 }
3309 return NULL;
3310 }