1 /*
2 * Copyright (c) 2015, 2019, Red Hat, Inc. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26
27 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
28 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
29 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
30 #include "gc/shenandoah/shenandoahForwarding.hpp"
31 #include "gc/shenandoah/shenandoahHeap.hpp"
32 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
33 #include "gc/shenandoah/shenandoahRuntime.hpp"
34 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
35 #include "opto/arraycopynode.hpp"
36 #include "opto/block.hpp"
37 #include "opto/callnode.hpp"
38 #include "opto/castnode.hpp"
39 #include "opto/movenode.hpp"
40 #include "opto/phaseX.hpp"
41 #include "opto/rootnode.hpp"
42 #include "opto/runtime.hpp"
43 #include "opto/subnode.hpp"
44
45 bool ShenandoahBarrierC2Support::expand(Compile* C, PhaseIterGVN& igvn) {
46 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
47 if ((state->enqueue_barriers_count() +
48 state->load_reference_barriers_count()) > 0) {
49 bool attempt_more_loopopts = ShenandoahLoopOptsAfterExpansion;
50 C->clear_major_progress();
51 PhaseIdealLoop ideal_loop(igvn, LoopOptsShenandoahExpand);
52 if (C->failing()) return false;
53 PhaseIdealLoop::verify(igvn);
54 DEBUG_ONLY(verify_raw_mem(C->root());)
55 if (attempt_more_loopopts) {
56 C->set_major_progress();
57 if (!C->optimize_loops(igvn, LoopOptsShenandoahPostExpand)) {
58 return false;
59 }
60 C->clear_major_progress();
61 if (C->range_check_cast_count() > 0) {
62 // No more loop optimizations. Remove all range check dependent CastIINodes.
63 C->remove_range_check_casts(igvn);
64 igvn.optimize();
65 }
66 }
67 }
68 return true;
69 }
70
71 bool ShenandoahBarrierC2Support::is_heap_state_test(Node* iff, int mask) {
72 if (!UseShenandoahGC) {
73 return false;
74 }
75 assert(iff->is_If(), "bad input");
76 if (iff->Opcode() != Op_If) {
77 return false;
78 }
79 Node* bol = iff->in(1);
80 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
81 return false;
82 }
83 Node* cmp = bol->in(1);
84 if (cmp->Opcode() != Op_CmpI) {
85 return false;
86 }
87 Node* in1 = cmp->in(1);
88 Node* in2 = cmp->in(2);
89 if (in2->find_int_con(-1) != 0) {
90 return false;
91 }
92 if (in1->Opcode() != Op_AndI) {
93 return false;
94 }
95 in2 = in1->in(2);
96 if (in2->find_int_con(-1) != mask) {
97 return false;
98 }
99 in1 = in1->in(1);
100
101 return is_gc_state_load(in1);
102 }
103
104 bool ShenandoahBarrierC2Support::is_heap_stable_test(Node* iff) {
105 return is_heap_state_test(iff, ShenandoahHeap::HAS_FORWARDED);
106 }
107
108 bool ShenandoahBarrierC2Support::is_gc_state_load(Node *n) {
109 if (!UseShenandoahGC) {
110 return false;
111 }
112 if (n->Opcode() != Op_LoadB && n->Opcode() != Op_LoadUB) {
113 return false;
114 }
115 Node* addp = n->in(MemNode::Address);
116 if (!addp->is_AddP()) {
117 return false;
118 }
119 Node* base = addp->in(AddPNode::Address);
120 Node* off = addp->in(AddPNode::Offset);
121 if (base->Opcode() != Op_ThreadLocal) {
122 return false;
123 }
124 if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) {
125 return false;
126 }
127 return true;
128 }
129
130 bool ShenandoahBarrierC2Support::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) {
131 assert(phase->is_dominator(stop, start), "bad inputs");
132 ResourceMark rm;
133 Unique_Node_List wq;
134 wq.push(start);
135 for (uint next = 0; next < wq.size(); next++) {
136 Node *m = wq.at(next);
137 if (m == stop) {
138 continue;
139 }
140 if (m->is_SafePoint() && !m->is_CallLeaf()) {
141 return true;
142 }
143 if (m->is_Region()) {
144 for (uint i = 1; i < m->req(); i++) {
145 wq.push(m->in(i));
146 }
147 } else {
148 wq.push(m->in(0));
149 }
150 }
151 return false;
152 }
153
154 #ifdef ASSERT
155 bool ShenandoahBarrierC2Support::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) {
156 assert(phis.size() == 0, "");
157
158 while (true) {
159 if (in->bottom_type() == TypePtr::NULL_PTR) {
160 if (trace) {tty->print_cr("NULL");}
161 } else if (!in->bottom_type()->make_ptr()->make_oopptr()) {
162 if (trace) {tty->print_cr("Non oop");}
163 } else {
164 if (in->is_ConstraintCast()) {
165 in = in->in(1);
166 continue;
167 } else if (in->is_AddP()) {
168 assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access");
169 in = in->in(AddPNode::Address);
170 continue;
171 } else if (in->is_Con()) {
172 if (trace) {
173 tty->print("Found constant");
174 in->dump();
175 }
176 } else if (in->Opcode() == Op_Parm) {
177 if (trace) {
178 tty->print("Found argument");
179 }
180 } else if (in->Opcode() == Op_CreateEx) {
181 if (trace) {
182 tty->print("Found create-exception");
183 }
184 } else if (in->Opcode() == Op_LoadP && in->adr_type() == TypeRawPtr::BOTTOM) {
185 if (trace) {
186 tty->print("Found raw LoadP (OSR argument?)");
187 }
188 } else if (in->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
189 if (t == ShenandoahOopStore) {
190 uint i = 0;
191 for (; i < phis.size(); i++) {
192 Node* n = phis.node_at(i);
193 if (n->Opcode() == Op_ShenandoahEnqueueBarrier) {
194 break;
195 }
196 }
197 if (i == phis.size()) {
198 return false;
199 }
200 }
201 barriers_used.push(in);
202 if (trace) {tty->print("Found barrier"); in->dump();}
203 } else if (in->Opcode() == Op_ShenandoahEnqueueBarrier) {
204 if (t != ShenandoahOopStore) {
205 in = in->in(1);
206 continue;
207 }
208 if (trace) {tty->print("Found enqueue barrier"); in->dump();}
209 phis.push(in, in->req());
210 in = in->in(1);
211 continue;
212 } else if (in->is_Proj() && in->in(0)->is_Allocate()) {
213 if (trace) {
214 tty->print("Found alloc");
215 in->in(0)->dump();
216 }
217 } else if (in->is_Proj() && (in->in(0)->Opcode() == Op_CallStaticJava || in->in(0)->Opcode() == Op_CallDynamicJava)) {
218 if (trace) {
219 tty->print("Found Java call");
220 }
221 } else if (in->is_Phi()) {
222 if (!visited.test_set(in->_idx)) {
223 if (trace) {tty->print("Pushed phi:"); in->dump();}
224 phis.push(in, 2);
225 in = in->in(1);
226 continue;
227 }
228 if (trace) {tty->print("Already seen phi:"); in->dump();}
229 } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) {
230 if (!visited.test_set(in->_idx)) {
231 if (trace) {tty->print("Pushed cmovep:"); in->dump();}
232 phis.push(in, CMoveNode::IfTrue);
233 in = in->in(CMoveNode::IfFalse);
234 continue;
235 }
236 if (trace) {tty->print("Already seen cmovep:"); in->dump();}
237 } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) {
238 in = in->in(1);
239 continue;
240 } else {
241 return false;
242 }
243 }
244 bool cont = false;
245 while (phis.is_nonempty()) {
246 uint idx = phis.index();
247 Node* phi = phis.node();
248 if (idx >= phi->req()) {
249 if (trace) {tty->print("Popped phi:"); phi->dump();}
250 phis.pop();
251 continue;
252 }
253 if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();}
254 in = phi->in(idx);
255 phis.set_index(idx+1);
256 cont = true;
257 break;
258 }
259 if (!cont) {
260 break;
261 }
262 }
263 return true;
264 }
265
266 void ShenandoahBarrierC2Support::report_verify_failure(const char* msg, Node* n1, Node* n2) {
267 if (n1 != NULL) {
268 n1->dump(+10);
269 }
270 if (n2 != NULL) {
271 n2->dump(+10);
272 }
273 fatal("%s", msg);
274 }
275
276 void ShenandoahBarrierC2Support::verify(RootNode* root) {
277 ResourceMark rm;
278 Unique_Node_List wq;
279 GrowableArray<Node*> barriers;
280 Unique_Node_List barriers_used;
281 Node_Stack phis(0);
282 VectorSet visited(Thread::current()->resource_area());
283 const bool trace = false;
284 const bool verify_no_useless_barrier = false;
285
286 wq.push(root);
287 for (uint next = 0; next < wq.size(); next++) {
288 Node *n = wq.at(next);
289 if (n->is_Load()) {
290 const bool trace = false;
291 if (trace) {tty->print("Verifying"); n->dump();}
292 if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) {
293 if (trace) {tty->print_cr("Load range/klass");}
294 } else {
295 const TypePtr* adr_type = n->as_Load()->adr_type();
296
297 if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) {
298 if (trace) {tty->print_cr("Mark load");}
299 } else if (adr_type->isa_instptr() &&
300 adr_type->is_instptr()->klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
301 adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset) {
302 if (trace) {tty->print_cr("Reference.get()");}
303 } else if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) {
304 report_verify_failure("Shenandoah verification: Load should have barriers", n);
305 }
306 }
307 } else if (n->is_Store()) {
308 const bool trace = false;
309
310 if (trace) {tty->print("Verifying"); n->dump();}
311 if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) {
312 Node* adr = n->in(MemNode::Address);
313 bool verify = true;
314
315 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) {
316 adr = adr->in(AddPNode::Address);
317 if (adr->is_AddP()) {
318 assert(adr->in(AddPNode::Base)->is_top(), "");
319 adr = adr->in(AddPNode::Address);
320 if (adr->Opcode() == Op_LoadP &&
321 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() &&
322 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
323 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) {
324 if (trace) {tty->print_cr("SATB prebarrier");}
325 verify = false;
326 }
327 }
328 }
329
330 if (verify && !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahStoreValEnqueueBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
331 report_verify_failure("Shenandoah verification: Store should have barriers", n);
332 }
333 }
334 if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
335 report_verify_failure("Shenandoah verification: Store (address) should have barriers", n);
336 }
337 } else if (n->Opcode() == Op_CmpP) {
338 const bool trace = false;
339
340 Node* in1 = n->in(1);
341 Node* in2 = n->in(2);
342 if (in1->bottom_type()->isa_oopptr()) {
343 if (trace) {tty->print("Verifying"); n->dump();}
344
345 bool mark_inputs = false;
346 if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR ||
347 (in1->is_Con() || in2->is_Con())) {
348 if (trace) {tty->print_cr("Comparison against a constant");}
349 mark_inputs = true;
350 } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) ||
351 (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) {
352 if (trace) {tty->print_cr("Comparison with newly alloc'ed object");}
353 mark_inputs = true;
354 } else {
355 assert(in2->bottom_type()->isa_oopptr(), "");
356
357 if (!verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) ||
358 !verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) {
359 report_verify_failure("Shenandoah verification: Cmp should have barriers", n);
360 }
361 }
362 if (verify_no_useless_barrier &&
363 mark_inputs &&
364 (!verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) ||
365 !verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) {
366 phis.clear();
367 visited.reset();
368 }
369 }
370 } else if (n->is_LoadStore()) {
371 if (n->in(MemNode::ValueIn)->bottom_type()->make_ptr() &&
372 !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahStoreValEnqueueBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
373 report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n);
374 }
375
376 if (n->in(MemNode::Address)->bottom_type()->make_oopptr() && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
377 report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n);
378 }
379 } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) {
380 CallNode* call = n->as_Call();
381
382 static struct {
383 const char* name;
384 struct {
385 int pos;
386 verify_type t;
387 } args[6];
388 } calls[] = {
389 "aescrypt_encryptBlock",
390 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
391 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
392 "aescrypt_decryptBlock",
393 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
394 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
395 "multiplyToLen",
396 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { TypeFunc::Parms+4, ShenandoahStore },
397 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
398 "squareToLen",
399 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { -1, ShenandoahNone},
400 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
401 "montgomery_multiply",
402 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
403 { TypeFunc::Parms+6, ShenandoahStore }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
404 "montgomery_square",
405 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+5, ShenandoahStore },
406 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
407 "mulAdd",
408 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
409 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
410 "vectorizedMismatch",
411 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
412 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
413 "updateBytesCRC32",
414 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
415 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
416 "updateBytesAdler32",
417 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
418 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
419 "updateBytesCRC32C",
420 { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad}, { -1, ShenandoahNone},
421 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
422 "counterMode_AESCrypt",
423 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
424 { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } },
425 "cipherBlockChaining_encryptAESCrypt",
426 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
427 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
428 "cipherBlockChaining_decryptAESCrypt",
429 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
430 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
431 "shenandoah_clone_barrier",
432 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
433 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
434 "ghash_processBlocks",
435 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
436 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
437 "sha1_implCompress",
438 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
439 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
440 "sha256_implCompress",
441 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
442 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
443 "sha512_implCompress",
444 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
445 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
446 "sha1_implCompressMB",
447 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
448 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
449 "sha256_implCompressMB",
450 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
451 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
452 "sha512_implCompressMB",
453 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
454 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
455 "encodeBlock",
456 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahStore }, { -1, ShenandoahNone },
457 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
458 };
459
460 if (call->is_call_to_arraycopystub()) {
461 Node* dest = NULL;
462 const TypeTuple* args = n->as_Call()->_tf->domain();
463 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
464 if (args->field_at(i)->isa_ptr()) {
465 j++;
466 if (j == 2) {
467 dest = n->in(i);
468 break;
469 }
470 }
471 }
472 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) ||
473 !verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) {
474 report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n);
475 }
476 } else if (strlen(call->_name) > 5 &&
477 !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) {
478 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) {
479 report_verify_failure("Shenandoah verification: _fill should have barriers", n);
480 }
481 } else if (!strcmp(call->_name, "shenandoah_wb_pre")) {
482 // skip
483 } else {
484 const int calls_len = sizeof(calls) / sizeof(calls[0]);
485 int i = 0;
486 for (; i < calls_len; i++) {
487 if (!strcmp(calls[i].name, call->_name)) {
488 break;
489 }
490 }
491 if (i != calls_len) {
492 const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]);
493 for (uint j = 0; j < args_len; j++) {
494 int pos = calls[i].args[j].pos;
495 if (pos == -1) {
496 break;
497 }
498 if (!verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) {
499 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
500 }
501 }
502 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
503 if (call->in(j)->bottom_type()->make_ptr() &&
504 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
505 uint k = 0;
506 for (; k < args_len && calls[i].args[k].pos != (int)j; k++);
507 if (k == args_len) {
508 fatal("arg %d for call %s not covered", j, call->_name);
509 }
510 }
511 }
512 } else {
513 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
514 if (call->in(j)->bottom_type()->make_ptr() &&
515 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
516 fatal("%s not covered", call->_name);
517 }
518 }
519 }
520 }
521 } else if (n->Opcode() == Op_ShenandoahEnqueueBarrier || n->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
522 // skip
523 } else if (n->is_AddP()
524 || n->is_Phi()
525 || n->is_ConstraintCast()
526 || n->Opcode() == Op_Return
527 || n->Opcode() == Op_CMoveP
528 || n->Opcode() == Op_CMoveN
529 || n->Opcode() == Op_Rethrow
530 || n->is_MemBar()
531 || n->Opcode() == Op_Conv2B
532 || n->Opcode() == Op_SafePoint
533 || n->is_CallJava()
534 || n->Opcode() == Op_Unlock
535 || n->Opcode() == Op_EncodeP
536 || n->Opcode() == Op_DecodeN) {
537 // nothing to do
538 } else {
539 static struct {
540 int opcode;
541 struct {
542 int pos;
543 verify_type t;
544 } inputs[2];
545 } others[] = {
546 Op_FastLock,
547 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
548 Op_Lock,
549 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone} },
550 Op_ArrayCopy,
551 { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } },
552 Op_StrCompressedCopy,
553 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
554 Op_StrInflatedCopy,
555 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
556 Op_AryEq,
557 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
558 Op_StrIndexOf,
559 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
560 Op_StrComp,
561 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
562 Op_StrEquals,
563 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
564 Op_EncodeISOArray,
565 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
566 Op_HasNegatives,
567 { { 2, ShenandoahLoad }, { -1, ShenandoahNone} },
568 Op_CastP2X,
569 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
570 Op_StrIndexOfChar,
571 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } },
572 };
573
574 const int others_len = sizeof(others) / sizeof(others[0]);
575 int i = 0;
576 for (; i < others_len; i++) {
577 if (others[i].opcode == n->Opcode()) {
578 break;
579 }
580 }
581 uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req();
582 if (i != others_len) {
583 const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]);
584 for (uint j = 0; j < inputs_len; j++) {
585 int pos = others[i].inputs[j].pos;
586 if (pos == -1) {
587 break;
588 }
589 if (!verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) {
590 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
591 }
592 }
593 for (uint j = 1; j < stop; j++) {
594 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
595 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
596 uint k = 0;
597 for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++);
598 if (k == inputs_len) {
599 fatal("arg %d for node %s not covered", j, n->Name());
600 }
601 }
602 }
603 } else {
604 for (uint j = 1; j < stop; j++) {
605 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
606 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
607 fatal("%s not covered", n->Name());
608 }
609 }
610 }
611 }
612
613 if (n->is_SafePoint()) {
614 SafePointNode* sfpt = n->as_SafePoint();
615 if (verify_no_useless_barrier && sfpt->jvms() != NULL) {
616 for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) {
617 if (!verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) {
618 phis.clear();
619 visited.reset();
620 }
621 }
622 }
623 }
624 }
625
626 if (verify_no_useless_barrier) {
627 for (int i = 0; i < barriers.length(); i++) {
628 Node* n = barriers.at(i);
629 if (!barriers_used.member(n)) {
630 tty->print("XXX useless barrier"); n->dump(-2);
631 ShouldNotReachHere();
632 }
633 }
634 }
635 }
636 #endif
637
638 bool ShenandoahBarrierC2Support::is_dominator_same_ctrl(Node* c, Node* d, Node* n, PhaseIdealLoop* phase) {
639 // That both nodes have the same control is not sufficient to prove
640 // domination, verify that there's no path from d to n
641 ResourceMark rm;
642 Unique_Node_List wq;
643 wq.push(d);
644 for (uint next = 0; next < wq.size(); next++) {
645 Node *m = wq.at(next);
646 if (m == n) {
647 return false;
648 }
649 if (m->is_Phi() && m->in(0)->is_Loop()) {
650 assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control");
651 } else {
652 if (m->is_Store() || m->is_LoadStore()) {
653 // Take anti-dependencies into account
654 Node* mem = m->in(MemNode::Memory);
655 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
656 Node* u = mem->fast_out(i);
657 if (u->is_Load() && phase->C->can_alias(m->adr_type(), phase->C->get_alias_index(u->adr_type())) &&
658 phase->ctrl_or_self(u) == c) {
659 wq.push(u);
660 }
661 }
662 }
663 for (uint i = 0; i < m->req(); i++) {
664 if (m->in(i) != NULL && phase->ctrl_or_self(m->in(i)) == c) {
665 wq.push(m->in(i));
666 }
667 }
668 }
669 }
670 return true;
671 }
672
673 bool ShenandoahBarrierC2Support::is_dominator(Node* d_c, Node* n_c, Node* d, Node* n, PhaseIdealLoop* phase) {
674 if (d_c != n_c) {
675 return phase->is_dominator(d_c, n_c);
676 }
677 return is_dominator_same_ctrl(d_c, d, n, phase);
678 }
679
680 Node* next_mem(Node* mem, int alias) {
681 Node* res = NULL;
682 if (mem->is_Proj()) {
683 res = mem->in(0);
684 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
685 res = mem->in(TypeFunc::Memory);
686 } else if (mem->is_Phi()) {
687 res = mem->in(1);
688 } else if (mem->is_MergeMem()) {
689 res = mem->as_MergeMem()->memory_at(alias);
690 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
691 assert(alias = Compile::AliasIdxRaw, "following raw memory can't lead to a barrier");
692 res = mem->in(MemNode::Memory);
693 } else {
694 #ifdef ASSERT
695 mem->dump();
696 #endif
697 ShouldNotReachHere();
698 }
699 return res;
700 }
701
702 Node* ShenandoahBarrierC2Support::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) {
703 Node* iffproj = NULL;
704 while (c != dom) {
705 Node* next = phase->idom(c);
706 assert(next->unique_ctrl_out() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?");
707 if (c->is_Region()) {
708 ResourceMark rm;
709 Unique_Node_List wq;
710 wq.push(c);
711 for (uint i = 0; i < wq.size(); i++) {
712 Node *n = wq.at(i);
713 if (n == next) {
714 continue;
715 }
716 if (n->is_Region()) {
717 for (uint j = 1; j < n->req(); j++) {
718 wq.push(n->in(j));
719 }
720 } else {
721 wq.push(n->in(0));
722 }
723 }
724 for (uint i = 0; i < wq.size(); i++) {
725 Node *n = wq.at(i);
726 assert(n->is_CFG(), "");
727 if (n->is_Multi()) {
728 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
729 Node* u = n->fast_out(j);
730 if (u->is_CFG()) {
731 if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
732 return NodeSentinel;
733 }
734 }
735 }
736 }
737 }
738 } else if (c->is_Proj()) {
739 if (c->is_IfProj()) {
740 if (c->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) != NULL) {
741 // continue;
742 } else {
743 if (!allow_one_proj) {
744 return NodeSentinel;
745 }
746 if (iffproj == NULL) {
747 iffproj = c;
748 } else {
749 return NodeSentinel;
750 }
751 }
752 } else if (c->Opcode() == Op_JumpProj) {
753 return NodeSentinel; // unsupported
754 } else if (c->Opcode() == Op_CatchProj) {
755 return NodeSentinel; // unsupported
756 } else if (c->Opcode() == Op_CProj && next->Opcode() == Op_NeverBranch) {
757 return NodeSentinel; // unsupported
758 } else {
759 assert(next->unique_ctrl_out() == c, "unsupported branch pattern");
760 }
761 }
762 c = next;
763 }
764 return iffproj;
765 }
766
767 Node* ShenandoahBarrierC2Support::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) {
768 ResourceMark rm;
769 VectorSet wq(Thread::current()->resource_area());
770 wq.set(mem->_idx);
771 mem_ctrl = phase->ctrl_or_self(mem);
772 while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) {
773 mem = next_mem(mem, alias);
774 if (wq.test_set(mem->_idx)) {
775 return NULL;
776 }
777 mem_ctrl = phase->ctrl_or_self(mem);
778 }
779 if (mem->is_MergeMem()) {
780 mem = mem->as_MergeMem()->memory_at(alias);
781 mem_ctrl = phase->ctrl_or_self(mem);
782 }
783 return mem;
784 }
785
786 Node* ShenandoahBarrierC2Support::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) {
787 Node* mem = NULL;
788 Node* c = ctrl;
789 do {
790 if (c->is_Region()) {
791 Node* phi_bottom = NULL;
792 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == NULL; i++) {
793 Node* u = c->fast_out(i);
794 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
795 if (u->adr_type() == TypePtr::BOTTOM) {
796 mem = u;
797 }
798 }
799 }
800 } else {
801 if (c->is_Call() && c->as_Call()->adr_type() != NULL) {
802 CallProjections projs;
803 c->as_Call()->extract_projections(&projs, true, false);
804 if (projs.fallthrough_memproj != NULL) {
805 if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
806 if (projs.catchall_memproj == NULL) {
807 mem = projs.fallthrough_memproj;
808 } else {
809 if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) {
810 mem = projs.fallthrough_memproj;
811 } else {
812 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier");
813 mem = projs.catchall_memproj;
814 }
815 }
816 }
817 } else {
818 Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
819 if (proj != NULL &&
820 proj->adr_type() == TypePtr::BOTTOM) {
821 mem = proj;
822 }
823 }
824 } else {
825 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
826 Node* u = c->fast_out(i);
827 if (u->is_Proj() &&
828 u->bottom_type() == Type::MEMORY &&
829 u->adr_type() == TypePtr::BOTTOM) {
830 assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
831 assert(mem == NULL, "only one proj");
832 mem = u;
833 }
834 }
835 assert(!c->is_Call() || c->as_Call()->adr_type() != NULL || mem == NULL, "no mem projection expected");
836 }
837 }
838 c = phase->idom(c);
839 } while (mem == NULL);
840 return mem;
841 }
842
843 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
844 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
845 Node* u = n->fast_out(i);
846 if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
847 uses.push(u);
848 }
849 }
850 }
851
852 static void hide_strip_mined_loop(OuterStripMinedLoopNode* outer, CountedLoopNode* inner, PhaseIdealLoop* phase) {
853 OuterStripMinedLoopEndNode* le = inner->outer_loop_end();
854 Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl));
855 phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl));
856 Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt);
857 phase->register_control(new_le, phase->get_loop(le), le->in(0));
858 phase->lazy_replace(outer, new_outer);
859 phase->lazy_replace(le, new_le);
860 inner->clear_strip_mined();
861 }
862
863 void ShenandoahBarrierC2Support::test_heap_stable(Node*& ctrl, Node* raw_mem, Node*& heap_stable_ctrl,
864 PhaseIdealLoop* phase, int flags) {
865 IdealLoopTree* loop = phase->get_loop(ctrl);
866 Node* thread = new ThreadLocalNode();
867 phase->register_new_node(thread, ctrl);
868 Node* offset = phase->igvn().MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
869 phase->set_ctrl(offset, phase->C->root());
870 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, offset);
871 phase->register_new_node(gc_state_addr, ctrl);
872 uint gc_state_idx = Compile::AliasIdxRaw;
873 const TypePtr* gc_state_adr_type = NULL; // debug-mode-only argument
874 debug_only(gc_state_adr_type = phase->C->get_adr_type(gc_state_idx));
875
876 Node* gc_state = new LoadBNode(ctrl, raw_mem, gc_state_addr, gc_state_adr_type, TypeInt::BYTE, MemNode::unordered);
877 phase->register_new_node(gc_state, ctrl);
878 Node* heap_stable_and = new AndINode(gc_state, phase->igvn().intcon(flags));
879 phase->register_new_node(heap_stable_and, ctrl);
880 Node* heap_stable_cmp = new CmpINode(heap_stable_and, phase->igvn().zerocon(T_INT));
881 phase->register_new_node(heap_stable_cmp, ctrl);
882 Node* heap_stable_test = new BoolNode(heap_stable_cmp, BoolTest::ne);
883 phase->register_new_node(heap_stable_test, ctrl);
884 IfNode* heap_stable_iff = new IfNode(ctrl, heap_stable_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
885 phase->register_control(heap_stable_iff, loop, ctrl);
886
887 heap_stable_ctrl = new IfFalseNode(heap_stable_iff);
888 phase->register_control(heap_stable_ctrl, loop, heap_stable_iff);
889 ctrl = new IfTrueNode(heap_stable_iff);
890 phase->register_control(ctrl, loop, heap_stable_iff);
891
892 assert(is_heap_state_test(heap_stable_iff, flags), "Should match the shape");
893 }
894
895 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) {
896 const Type* val_t = phase->igvn().type(val);
897 if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
898 IdealLoopTree* loop = phase->get_loop(ctrl);
899 Node* null_cmp = new CmpPNode(val, phase->igvn().zerocon(T_OBJECT));
900 phase->register_new_node(null_cmp, ctrl);
901 Node* null_test = new BoolNode(null_cmp, BoolTest::ne);
902 phase->register_new_node(null_test, ctrl);
903 IfNode* null_iff = new IfNode(ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
904 phase->register_control(null_iff, loop, ctrl);
905 ctrl = new IfTrueNode(null_iff);
906 phase->register_control(ctrl, loop, null_iff);
907 null_ctrl = new IfFalseNode(null_iff);
908 phase->register_control(null_ctrl, loop, null_iff);
909 }
910 }
911
912 Node* ShenandoahBarrierC2Support::clone_null_check(Node*& c, Node* val, Node* unc_ctrl, PhaseIdealLoop* phase) {
913 IdealLoopTree *loop = phase->get_loop(c);
914 Node* iff = unc_ctrl->in(0);
915 assert(iff->is_If(), "broken");
916 Node* new_iff = iff->clone();
917 new_iff->set_req(0, c);
918 phase->register_control(new_iff, loop, c);
919 Node* iffalse = new IfFalseNode(new_iff->as_If());
920 phase->register_control(iffalse, loop, new_iff);
921 Node* iftrue = new IfTrueNode(new_iff->as_If());
922 phase->register_control(iftrue, loop, new_iff);
923 c = iftrue;
924 const Type *t = phase->igvn().type(val);
925 assert(val->Opcode() == Op_CastPP, "expect cast to non null here");
926 Node* uncasted_val = val->in(1);
927 val = new CastPPNode(uncasted_val, t);
928 val->init_req(0, c);
929 phase->register_new_node(val, c);
930 return val;
931 }
932
933 void ShenandoahBarrierC2Support::fix_null_check(Node* unc, Node* unc_ctrl, Node* new_unc_ctrl,
934 Unique_Node_List& uses, PhaseIdealLoop* phase) {
935 IfNode* iff = unc_ctrl->in(0)->as_If();
936 Node* proj = iff->proj_out(0);
937 assert(proj != unc_ctrl, "bad projection");
938 Node* use = proj->unique_ctrl_out();
939
940 assert(use == unc || use->is_Region(), "what else?");
941
942 uses.clear();
943 if (use == unc) {
944 phase->set_idom(use, new_unc_ctrl, phase->dom_depth(use));
945 for (uint i = 1; i < unc->req(); i++) {
946 Node* n = unc->in(i);
947 if (phase->has_ctrl(n) && phase->get_ctrl(n) == proj) {
948 uses.push(n);
949 }
950 }
951 } else {
952 assert(use->is_Region(), "what else?");
953 uint idx = 1;
954 for (; use->in(idx) != proj; idx++);
955 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
956 Node* u = use->fast_out(i);
957 if (u->is_Phi() && phase->get_ctrl(u->in(idx)) == proj) {
958 uses.push(u->in(idx));
959 }
960 }
961 }
962 for(uint next = 0; next < uses.size(); next++ ) {
963 Node *n = uses.at(next);
964 assert(phase->get_ctrl(n) == proj, "bad control");
965 phase->set_ctrl_and_loop(n, new_unc_ctrl);
966 if (n->in(0) == proj) {
967 phase->igvn().replace_input_of(n, 0, new_unc_ctrl);
968 }
969 for (uint i = 0; i < n->req(); i++) {
970 Node* m = n->in(i);
971 if (m != NULL && phase->has_ctrl(m) && phase->get_ctrl(m) == proj) {
972 uses.push(m);
973 }
974 }
975 }
976
977 phase->igvn().rehash_node_delayed(use);
978 int nb = use->replace_edge(proj, new_unc_ctrl);
979 assert(nb == 1, "only use expected");
980 }
981
982 void ShenandoahBarrierC2Support::in_cset_fast_test(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) {
983 IdealLoopTree *loop = phase->get_loop(ctrl);
984 Node* raw_rbtrue = new CastP2XNode(ctrl, val);
985 phase->register_new_node(raw_rbtrue, ctrl);
986 Node* cset_offset = new URShiftXNode(raw_rbtrue, phase->igvn().intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
987 phase->register_new_node(cset_offset, ctrl);
988 Node* in_cset_fast_test_base_addr = phase->igvn().makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
989 phase->set_ctrl(in_cset_fast_test_base_addr, phase->C->root());
990 Node* in_cset_fast_test_adr = new AddPNode(phase->C->top(), in_cset_fast_test_base_addr, cset_offset);
991 phase->register_new_node(in_cset_fast_test_adr, ctrl);
992 uint in_cset_fast_test_idx = Compile::AliasIdxRaw;
993 const TypePtr* in_cset_fast_test_adr_type = NULL; // debug-mode-only argument
994 debug_only(in_cset_fast_test_adr_type = phase->C->get_adr_type(in_cset_fast_test_idx));
995 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);
996 phase->register_new_node(in_cset_fast_test_load, ctrl);
997 Node* in_cset_fast_test_cmp = new CmpINode(in_cset_fast_test_load, phase->igvn().zerocon(T_INT));
998 phase->register_new_node(in_cset_fast_test_cmp, ctrl);
999 Node* in_cset_fast_test_test = new BoolNode(in_cset_fast_test_cmp, BoolTest::eq);
1000 phase->register_new_node(in_cset_fast_test_test, ctrl);
1001 IfNode* in_cset_fast_test_iff = new IfNode(ctrl, in_cset_fast_test_test, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
1002 phase->register_control(in_cset_fast_test_iff, loop, ctrl);
1003
1004 not_cset_ctrl = new IfTrueNode(in_cset_fast_test_iff);
1005 phase->register_control(not_cset_ctrl, loop, in_cset_fast_test_iff);
1006
1007 ctrl = new IfFalseNode(in_cset_fast_test_iff);
1008 phase->register_control(ctrl, loop, in_cset_fast_test_iff);
1009 }
1010
1011 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node* load_addr, Node*& result_mem, Node* raw_mem, bool is_native, PhaseIdealLoop* phase) {
1012 IdealLoopTree*loop = phase->get_loop(ctrl);
1013 const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr();
1014
1015 // The slow path stub consumes and produces raw memory in addition
1016 // to the existing memory edges
1017 Node* base = find_bottom_mem(ctrl, phase);
1018 MergeMemNode* mm = MergeMemNode::make(base);
1019 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
1020 phase->register_new_node(mm, ctrl);
1021
1022 address target = LP64_ONLY(UseCompressedOops) NOT_LP64(false) ?
1023 CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_narrow) :
1024 CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier);
1025
1026 address calladdr = is_native ? CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_native)
1027 : target;
1028 const char* name = is_native ? "load_reference_barrier_native" : "load_reference_barrier";
1029 Node* call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type(), calladdr, name, TypeRawPtr::BOTTOM);
1030
1031 call->init_req(TypeFunc::Control, ctrl);
1032 call->init_req(TypeFunc::I_O, phase->C->top());
1033 call->init_req(TypeFunc::Memory, mm);
1034 call->init_req(TypeFunc::FramePtr, phase->C->top());
1035 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1036 call->init_req(TypeFunc::Parms, val);
1037 call->init_req(TypeFunc::Parms+1, load_addr);
1038 phase->register_control(call, loop, ctrl);
1039 ctrl = new ProjNode(call, TypeFunc::Control);
1040 phase->register_control(ctrl, loop, call);
1041 result_mem = new ProjNode(call, TypeFunc::Memory);
1042 phase->register_new_node(result_mem, call);
1043 val = new ProjNode(call, TypeFunc::Parms);
1044 phase->register_new_node(val, call);
1045 val = new CheckCastPPNode(ctrl, val, obj_type);
1046 phase->register_new_node(val, ctrl);
1047 }
1048
1049 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) {
1050 Node* ctrl = phase->get_ctrl(barrier);
1051 Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
1052
1053 // Update the control of all nodes that should be after the
1054 // barrier control flow
1055 uses.clear();
1056 // Every node that is control dependent on the barrier's input
1057 // control will be after the expanded barrier. The raw memory (if
1058 // its memory is control dependent on the barrier's input control)
1059 // must stay above the barrier.
1060 uses_to_ignore.clear();
1061 if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
1062 uses_to_ignore.push(init_raw_mem);
1063 }
1064 for (uint next = 0; next < uses_to_ignore.size(); next++) {
1065 Node *n = uses_to_ignore.at(next);
1066 for (uint i = 0; i < n->req(); i++) {
1067 Node* in = n->in(i);
1068 if (in != NULL && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
1069 uses_to_ignore.push(in);
1070 }
1071 }
1072 }
1073 for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
1074 Node* u = ctrl->fast_out(i);
1075 if (u->_idx < last &&
1076 u != barrier &&
1077 !uses_to_ignore.member(u) &&
1078 (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) &&
1079 (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) {
1080 Node* old_c = phase->ctrl_or_self(u);
1081 Node* c = old_c;
1082 if (c != ctrl ||
1083 is_dominator_same_ctrl(old_c, barrier, u, phase) ||
1084 ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) {
1085 phase->igvn().rehash_node_delayed(u);
1086 int nb = u->replace_edge(ctrl, region);
1087 if (u->is_CFG()) {
1088 if (phase->idom(u) == ctrl) {
1089 phase->set_idom(u, region, phase->dom_depth(region));
1090 }
1091 } else if (phase->get_ctrl(u) == ctrl) {
1092 assert(u != init_raw_mem, "should leave input raw mem above the barrier");
1093 uses.push(u);
1094 }
1095 assert(nb == 1, "more than 1 ctrl input?");
1096 --i, imax -= nb;
1097 }
1098 }
1099 }
1100 }
1101
1102 static Node* create_phis_on_call_return(Node* ctrl, Node* c, Node* n, Node* n_clone, const CallProjections& projs, PhaseIdealLoop* phase) {
1103 Node* region = NULL;
1104 while (c != ctrl) {
1105 if (c->is_Region()) {
1106 region = c;
1107 }
1108 c = phase->idom(c);
1109 }
1110 assert(region != NULL, "");
1111 Node* phi = new PhiNode(region, n->bottom_type());
1112 for (uint j = 1; j < region->req(); j++) {
1113 Node* in = region->in(j);
1114 if (phase->is_dominator(projs.fallthrough_catchproj, in)) {
1115 phi->init_req(j, n);
1116 } else if (phase->is_dominator(projs.catchall_catchproj, in)) {
1117 phi->init_req(j, n_clone);
1118 } else {
1119 phi->init_req(j, create_phis_on_call_return(ctrl, in, n, n_clone, projs, phase));
1120 }
1121 }
1122 phase->register_new_node(phi, region);
1123 return phi;
1124 }
1125
1126 void ShenandoahBarrierC2Support::pin_and_expand(PhaseIdealLoop* phase) {
1127 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
1128
1129 Unique_Node_List uses;
1130 for (int i = 0; i < state->enqueue_barriers_count(); i++) {
1131 Node* barrier = state->enqueue_barrier(i);
1132 Node* ctrl = phase->get_ctrl(barrier);
1133 IdealLoopTree* loop = phase->get_loop(ctrl);
1134 if (loop->_head->is_OuterStripMinedLoop()) {
1135 // Expanding a barrier here will break loop strip mining
1136 // verification. Transform the loop so the loop nest doesn't
1137 // appear as strip mined.
1138 OuterStripMinedLoopNode* outer = loop->_head->as_OuterStripMinedLoop();
1139 hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1140 }
1141 }
1142
1143 Node_Stack stack(0);
1144 Node_List clones;
1145 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1146 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1147 if (lrb->is_redundant()) {
1148 continue;
1149 }
1150
1151 Node* ctrl = phase->get_ctrl(lrb);
1152 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1153
1154 CallStaticJavaNode* unc = NULL;
1155 Node* unc_ctrl = NULL;
1156 Node* uncasted_val = val;
1157
1158 for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1159 Node* u = lrb->fast_out(i);
1160 if (u->Opcode() == Op_CastPP &&
1161 u->in(0) != NULL &&
1162 phase->is_dominator(u->in(0), ctrl)) {
1163 const Type* u_t = phase->igvn().type(u);
1164
1165 if (u_t->meet(TypePtr::NULL_PTR) != u_t &&
1166 u->in(0)->Opcode() == Op_IfTrue &&
1167 u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
1168 u->in(0)->in(0)->is_If() &&
1169 u->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
1170 u->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
1171 u->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
1172 u->in(0)->in(0)->in(1)->in(1)->in(1) == val &&
1173 u->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
1174 IdealLoopTree* loop = phase->get_loop(ctrl);
1175 IdealLoopTree* unc_loop = phase->get_loop(u->in(0));
1176
1177 if (!unc_loop->is_member(loop)) {
1178 continue;
1179 }
1180
1181 Node* branch = no_branches(ctrl, u->in(0), false, phase);
1182 assert(branch == NULL || branch == NodeSentinel, "was not looking for a branch");
1183 if (branch == NodeSentinel) {
1184 continue;
1185 }
1186
1187 phase->igvn().replace_input_of(u, 1, val);
1188 phase->igvn().replace_input_of(lrb, ShenandoahLoadReferenceBarrierNode::ValueIn, u);
1189 phase->set_ctrl(u, u->in(0));
1190 phase->set_ctrl(lrb, u->in(0));
1191 unc = u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
1192 unc_ctrl = u->in(0);
1193 val = u;
1194
1195 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
1196 Node* u = val->fast_out(j);
1197 if (u == lrb) continue;
1198 phase->igvn().rehash_node_delayed(u);
1199 int nb = u->replace_edge(val, lrb);
1200 --j; jmax -= nb;
1201 }
1202
1203 RegionNode* r = new RegionNode(3);
1204 IfNode* iff = unc_ctrl->in(0)->as_If();
1205
1206 Node* ctrl_use = unc_ctrl->unique_ctrl_out();
1207 Node* unc_ctrl_clone = unc_ctrl->clone();
1208 phase->register_control(unc_ctrl_clone, loop, iff);
1209 Node* c = unc_ctrl_clone;
1210 Node* new_cast = clone_null_check(c, val, unc_ctrl_clone, phase);
1211 r->init_req(1, new_cast->in(0)->in(0)->as_If()->proj_out(0));
1212
1213 phase->igvn().replace_input_of(unc_ctrl, 0, c->in(0));
1214 phase->set_idom(unc_ctrl, c->in(0), phase->dom_depth(unc_ctrl));
1215 phase->lazy_replace(c, unc_ctrl);
1216 c = NULL;;
1217 phase->igvn().replace_input_of(val, 0, unc_ctrl_clone);
1218 phase->set_ctrl(val, unc_ctrl_clone);
1219
1220 IfNode* new_iff = new_cast->in(0)->in(0)->as_If();
1221 fix_null_check(unc, unc_ctrl_clone, r, uses, phase);
1222 Node* iff_proj = iff->proj_out(0);
1223 r->init_req(2, iff_proj);
1224 phase->register_control(r, phase->ltree_root(), iff);
1225
1226 Node* new_bol = new_iff->in(1)->clone();
1227 Node* new_cmp = new_bol->in(1)->clone();
1228 assert(new_cmp->Opcode() == Op_CmpP, "broken");
1229 assert(new_cmp->in(1) == val->in(1), "broken");
1230 new_bol->set_req(1, new_cmp);
1231 new_cmp->set_req(1, lrb);
1232 phase->register_new_node(new_bol, new_iff->in(0));
1233 phase->register_new_node(new_cmp, new_iff->in(0));
1234 phase->igvn().replace_input_of(new_iff, 1, new_bol);
1235 phase->igvn().replace_input_of(new_cast, 1, lrb);
1236
1237 for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1238 Node* u = lrb->fast_out(i);
1239 if (u == new_cast || u == new_cmp) {
1240 continue;
1241 }
1242 phase->igvn().rehash_node_delayed(u);
1243 int nb = u->replace_edge(lrb, new_cast);
1244 assert(nb > 0, "no update?");
1245 --i; imax -= nb;
1246 }
1247
1248 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
1249 Node* u = val->fast_out(i);
1250 if (u == lrb) {
1251 continue;
1252 }
1253 phase->igvn().rehash_node_delayed(u);
1254 int nb = u->replace_edge(val, new_cast);
1255 assert(nb > 0, "no update?");
1256 --i; imax -= nb;
1257 }
1258
1259 ctrl = unc_ctrl_clone;
1260 phase->set_ctrl_and_loop(lrb, ctrl);
1261 break;
1262 }
1263 }
1264 }
1265 if ((ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) || ctrl->is_CallJava()) {
1266 CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_CallJava() : ctrl->as_CallJava();
1267 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
1268 // The rethrow call may have too many projections to be
1269 // properly handled here. Given there's no reason for a
1270 // barrier to depend on the call, move it above the call
1271 stack.push(lrb, 0);
1272 do {
1273 Node* n = stack.node();
1274 uint idx = stack.index();
1275 if (idx < n->req()) {
1276 Node* in = n->in(idx);
1277 stack.set_index(idx+1);
1278 if (in != NULL) {
1279 if (phase->has_ctrl(in)) {
1280 if (phase->is_dominator(call, phase->get_ctrl(in))) {
1281 #ifdef ASSERT
1282 for (uint i = 0; i < stack.size(); i++) {
1283 assert(stack.node_at(i) != in, "node shouldn't have been seen yet");
1284 }
1285 #endif
1286 stack.push(in, 0);
1287 }
1288 } else {
1289 assert(phase->is_dominator(in, call->in(0)), "no dependency on the call");
1290 }
1291 }
1292 } else {
1293 phase->set_ctrl(n, call->in(0));
1294 stack.pop();
1295 }
1296 } while(stack.size() > 0);
1297 continue;
1298 }
1299 CallProjections projs;
1300 call->extract_projections(&projs, false, false);
1301
1302 Node* lrb_clone = lrb->clone();
1303 phase->register_new_node(lrb_clone, projs.catchall_catchproj);
1304 phase->set_ctrl(lrb, projs.fallthrough_catchproj);
1305
1306 stack.push(lrb, 0);
1307 clones.push(lrb_clone);
1308
1309 do {
1310 assert(stack.size() == clones.size(), "");
1311 Node* n = stack.node();
1312 #ifdef ASSERT
1313 if (n->is_Load()) {
1314 Node* mem = n->in(MemNode::Memory);
1315 for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
1316 Node* u = mem->fast_out(j);
1317 assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?");
1318 }
1319 }
1320 #endif
1321 uint idx = stack.index();
1322 Node* n_clone = clones.at(clones.size()-1);
1323 if (idx < n->outcnt()) {
1324 Node* u = n->raw_out(idx);
1325 Node* c = phase->ctrl_or_self(u);
1326 if (phase->is_dominator(call, c) && phase->is_dominator(c, projs.fallthrough_proj)) {
1327 stack.set_index(idx+1);
1328 assert(!u->is_CFG(), "");
1329 stack.push(u, 0);
1330 Node* u_clone = u->clone();
1331 int nb = u_clone->replace_edge(n, n_clone);
1332 assert(nb > 0, "should have replaced some uses");
1333 phase->register_new_node(u_clone, projs.catchall_catchproj);
1334 clones.push(u_clone);
1335 phase->set_ctrl(u, projs.fallthrough_catchproj);
1336 } else {
1337 bool replaced = false;
1338 if (u->is_Phi()) {
1339 for (uint k = 1; k < u->req(); k++) {
1340 if (u->in(k) == n) {
1341 if (phase->is_dominator(projs.catchall_catchproj, u->in(0)->in(k))) {
1342 phase->igvn().replace_input_of(u, k, n_clone);
1343 replaced = true;
1344 } else if (!phase->is_dominator(projs.fallthrough_catchproj, u->in(0)->in(k))) {
1345 phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase));
1346 replaced = true;
1347 }
1348 }
1349 }
1350 } else {
1351 if (phase->is_dominator(projs.catchall_catchproj, c)) {
1352 phase->igvn().rehash_node_delayed(u);
1353 int nb = u->replace_edge(n, n_clone);
1354 assert(nb > 0, "should have replaced some uses");
1355 replaced = true;
1356 } else if (!phase->is_dominator(projs.fallthrough_catchproj, c)) {
1357 phase->igvn().rehash_node_delayed(u);
1358 int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase));
1359 assert(nb > 0, "should have replaced some uses");
1360 replaced = true;
1361 }
1362 }
1363 if (!replaced) {
1364 stack.set_index(idx+1);
1365 }
1366 }
1367 } else {
1368 stack.pop();
1369 clones.pop();
1370 }
1371 } while (stack.size() > 0);
1372 assert(stack.size() == 0 && clones.size() == 0, "");
1373 }
1374 }
1375
1376 for (int i = 0; i < state->load_reference_barriers_count(); i++) {
1377 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1378 if (lrb->is_redundant()) {
1379 continue;
1380 }
1381 Node* ctrl = phase->get_ctrl(lrb);
1382 IdealLoopTree* loop = phase->get_loop(ctrl);
1383 if (loop->_head->is_OuterStripMinedLoop()) {
1384 // Expanding a barrier here will break loop strip mining
1385 // verification. Transform the loop so the loop nest doesn't
1386 // appear as strip mined.
1387 OuterStripMinedLoopNode* outer = loop->_head->as_OuterStripMinedLoop();
1388 hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1389 }
1390 }
1391
1392 // Expand load-reference-barriers
1393 MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
1394 Unique_Node_List uses_to_ignore;
1395 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1396 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1397 if (lrb->is_redundant()) {
1398 phase->igvn().replace_node(lrb, lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1399 continue;
1400 }
1401 uint last = phase->C->unique();
1402 Node* ctrl = phase->get_ctrl(lrb);
1403 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1404
1405
1406 Node* orig_ctrl = ctrl;
1407
1408 Node* raw_mem = fixer.find_mem(ctrl, lrb);
1409 Node* init_raw_mem = raw_mem;
1410 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL);
1411
1412 IdealLoopTree *loop = phase->get_loop(ctrl);
1413 CallStaticJavaNode* unc = lrb->pin_and_expand_null_check(phase->igvn());
1414 Node* unc_ctrl = NULL;
1415 if (unc != NULL) {
1416 if (val->in(ShenandoahLoadReferenceBarrierNode::Control) != ctrl) {
1417 unc = NULL;
1418 } else {
1419 unc_ctrl = val->in(ShenandoahLoadReferenceBarrierNode::Control);
1420 }
1421 }
1422
1423 Node* uncasted_val = val;
1424 if (unc != NULL) {
1425 uncasted_val = val->in(1);
1426 }
1427
1428 Node* heap_stable_ctrl = NULL;
1429 Node* null_ctrl = NULL;
1430
1431 assert(val->bottom_type()->make_oopptr(), "need oop");
1432 assert(val->bottom_type()->make_oopptr()->const_oop() == NULL, "expect non-constant");
1433
1434 enum { _heap_stable = 1, _not_cset, _evac_path, _null_path, PATH_LIMIT };
1435 Node* region = new RegionNode(PATH_LIMIT);
1436 Node* val_phi = new PhiNode(region, uncasted_val->bottom_type()->is_oopptr());
1437 Node* raw_mem_phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1438
1439 // Stable path.
1440 test_heap_stable(ctrl, raw_mem, heap_stable_ctrl, phase, ShenandoahHeap::HAS_FORWARDED);
1441 IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
1442
1443 // Heap stable case
1444 region->init_req(_heap_stable, heap_stable_ctrl);
1445 val_phi->init_req(_heap_stable, uncasted_val);
1446 raw_mem_phi->init_req(_heap_stable, raw_mem);
1447
1448 Node* reg2_ctrl = NULL;
1449 // Null case
1450 test_null(ctrl, val, null_ctrl, phase);
1451 if (null_ctrl != NULL) {
1452 reg2_ctrl = null_ctrl->in(0);
1453 region->init_req(_null_path, null_ctrl);
1454 val_phi->init_req(_null_path, uncasted_val);
1455 raw_mem_phi->init_req(_null_path, raw_mem);
1456 } else {
1457 region->del_req(_null_path);
1458 val_phi->del_req(_null_path);
1459 raw_mem_phi->del_req(_null_path);
1460 }
1461
1462 // Test for in-cset.
1463 // Wires !in_cset(obj) to slot 2 of region and phis
1464 Node* not_cset_ctrl = NULL;
1465 in_cset_fast_test(ctrl, not_cset_ctrl, uncasted_val, raw_mem, phase);
1466 if (not_cset_ctrl != NULL) {
1467 if (reg2_ctrl == NULL) reg2_ctrl = not_cset_ctrl->in(0);
1468 region->init_req(_not_cset, not_cset_ctrl);
1469 val_phi->init_req(_not_cset, uncasted_val);
1470 raw_mem_phi->init_req(_not_cset, raw_mem);
1471 }
1472
1473 // Resolve object when orig-value is in cset.
1474 // Make the unconditional resolve for fwdptr.
1475 Node* new_val = uncasted_val;
1476 if (unc_ctrl != NULL) {
1477 // Clone the null check in this branch to allow implicit null check
1478 new_val = clone_null_check(ctrl, val, unc_ctrl, phase);
1479 fix_null_check(unc, unc_ctrl, ctrl->in(0)->as_If()->proj_out(0), uses, phase);
1480
1481 IfNode* iff = unc_ctrl->in(0)->as_If();
1482 phase->igvn().replace_input_of(iff, 1, phase->igvn().intcon(1));
1483 }
1484
1485 // Call lrb-stub and wire up that path in slots 4
1486 Node* result_mem = NULL;
1487
1488 Node* fwd = new_val;
1489 Node* addr;
1490 if (ShenandoahSelfFixing) {
1491 VectorSet visited(Thread::current()->resource_area());
1492 addr = get_load_addr(phase, visited, lrb);
1493 } else {
1494 addr = phase->igvn().zerocon(T_OBJECT);
1495 }
1496 if (addr->Opcode() == Op_AddP) {
1497 Node* orig_base = addr->in(AddPNode::Base);
1498 Node* base = new CheckCastPPNode(ctrl, orig_base, orig_base->bottom_type(), true);
1499 phase->register_new_node(base, ctrl);
1500 if (addr->in(AddPNode::Base) == addr->in((AddPNode::Address))) {
1501 // Field access
1502 addr = addr->clone();
1503 addr->set_req(AddPNode::Base, base);
1504 addr->set_req(AddPNode::Address, base);
1505 phase->register_new_node(addr, ctrl);
1506 } else {
1507 Node* addr2 = addr->in(AddPNode::Address);
1508 if (addr2->Opcode() == Op_AddP && addr2->in(AddPNode::Base) == addr2->in(AddPNode::Address) &&
1509 addr2->in(AddPNode::Base) == orig_base) {
1510 addr2 = addr2->clone();
1511 addr2->set_req(AddPNode::Base, base);
1512 addr2->set_req(AddPNode::Address, base);
1513 phase->register_new_node(addr2, ctrl);
1514 addr = addr->clone();
1515 addr->set_req(AddPNode::Base, base);
1516 addr->set_req(AddPNode::Address, addr2);
1517 phase->register_new_node(addr, ctrl);
1518 }
1519 }
1520 }
1521 call_lrb_stub(ctrl, fwd, addr, result_mem, raw_mem, lrb->is_native(), phase);
1522 region->init_req(_evac_path, ctrl);
1523 val_phi->init_req(_evac_path, fwd);
1524 raw_mem_phi->init_req(_evac_path, result_mem);
1525
1526 phase->register_control(region, loop, heap_stable_iff);
1527 Node* out_val = val_phi;
1528 phase->register_new_node(val_phi, region);
1529 phase->register_new_node(raw_mem_phi, region);
1530
1531 fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase);
1532
1533 ctrl = orig_ctrl;
1534
1535 if (unc != NULL) {
1536 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
1537 Node* u = val->fast_out(i);
1538 Node* c = phase->ctrl_or_self(u);
1539 if (u != lrb && (c != ctrl || is_dominator_same_ctrl(c, lrb, u, phase))) {
1540 phase->igvn().rehash_node_delayed(u);
1541 int nb = u->replace_edge(val, out_val);
1542 --i, imax -= nb;
1543 }
1544 }
1545 if (val->outcnt() == 0) {
1546 phase->igvn()._worklist.push(val);
1547 }
1548 }
1549 phase->igvn().replace_node(lrb, out_val);
1550
1551 follow_barrier_uses(out_val, ctrl, uses, phase);
1552
1553 for(uint next = 0; next < uses.size(); next++ ) {
1554 Node *n = uses.at(next);
1555 assert(phase->get_ctrl(n) == ctrl, "bad control");
1556 assert(n != init_raw_mem, "should leave input raw mem above the barrier");
1557 phase->set_ctrl(n, region);
1558 follow_barrier_uses(n, ctrl, uses, phase);
1559 }
1560
1561 // The slow path call produces memory: hook the raw memory phi
1562 // from the expanded load reference barrier with the rest of the graph
1563 // which may require adding memory phis at every post dominated
1564 // region and at enclosing loop heads. Use the memory state
1565 // collected in memory_nodes to fix the memory graph. Update that
1566 // memory state as we go.
1567 fixer.fix_mem(ctrl, region, init_raw_mem, raw_mem_for_ctrl, raw_mem_phi, uses);
1568 }
1569 // Done expanding load-reference-barriers.
1570 assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced");
1571
1572 for (int i = state->enqueue_barriers_count() - 1; i >= 0; i--) {
1573 Node* barrier = state->enqueue_barrier(i);
1574 Node* pre_val = barrier->in(1);
1575
1576 assert(!phase->igvn().type(pre_val)->higher_equal(TypePtr::NULL_PTR), "no known-NULLs here");
1577
1578 if (((ShenandoahEnqueueBarrierNode*)barrier)->can_eliminate(phase)) {
1579 phase->igvn().replace_node(barrier, pre_val);
1580 continue;
1581 }
1582
1583 Node* ctrl = phase->get_ctrl(barrier);
1584
1585 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) {
1586 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0)->in(0), pre_val, ctrl->in(0), phase), "can't move");
1587 ctrl = ctrl->in(0)->in(0);
1588 phase->set_ctrl(barrier, ctrl);
1589 } else if (ctrl->is_CallRuntime()) {
1590 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0), pre_val, ctrl, phase), "can't move");
1591 ctrl = ctrl->in(0);
1592 phase->set_ctrl(barrier, ctrl);
1593 }
1594
1595 Node* init_ctrl = ctrl;
1596 IdealLoopTree* loop = phase->get_loop(ctrl);
1597 Node* raw_mem = fixer.find_mem(ctrl, barrier);
1598 Node* init_raw_mem = raw_mem;
1599 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL);
1600 Node* heap_stable_ctrl = NULL;
1601 Node* null_ctrl = NULL;
1602 uint last = phase->C->unique();
1603
1604 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
1605 Node* region = new RegionNode(PATH_LIMIT);
1606 Node* phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1607
1608 enum { _fast_path = 1, _slow_path, _null_path, PATH_LIMIT2 };
1609 Node* region2 = new RegionNode(PATH_LIMIT2);
1610 Node* phi2 = PhiNode::make(region2, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1611
1612 // Stable path.
1613 test_heap_stable(ctrl, raw_mem, heap_stable_ctrl, phase, ShenandoahHeap::TRAVERSAL | ShenandoahHeap::MARKING);
1614 region->init_req(_heap_stable, heap_stable_ctrl);
1615 phi->init_req(_heap_stable, raw_mem);
1616
1617 // Null path
1618 Node* reg2_ctrl = NULL;
1619 test_null(ctrl, pre_val, null_ctrl, phase);
1620 if (null_ctrl != NULL) {
1621 reg2_ctrl = null_ctrl->in(0);
1622 region2->init_req(_null_path, null_ctrl);
1623 phi2->init_req(_null_path, raw_mem);
1624 } else {
1625 region2->del_req(_null_path);
1626 phi2->del_req(_null_path);
1627 }
1628
1629 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
1630 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1631 Node* thread = new ThreadLocalNode();
1632 phase->register_new_node(thread, ctrl);
1633 Node* buffer_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(buffer_offset));
1634 phase->register_new_node(buffer_adr, ctrl);
1635 Node* index_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(index_offset));
1636 phase->register_new_node(index_adr, ctrl);
1637
1638 BasicType index_bt = TypeX_X->basic_type();
1639 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
1640 const TypePtr* adr_type = TypeRawPtr::BOTTOM;
1641 Node* index = new LoadXNode(ctrl, raw_mem, index_adr, adr_type, TypeX_X, MemNode::unordered);
1642 phase->register_new_node(index, ctrl);
1643 Node* index_cmp = new CmpXNode(index, phase->igvn().MakeConX(0));
1644 phase->register_new_node(index_cmp, ctrl);
1645 Node* index_test = new BoolNode(index_cmp, BoolTest::ne);
1646 phase->register_new_node(index_test, ctrl);
1647 IfNode* queue_full_iff = new IfNode(ctrl, index_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
1648 if (reg2_ctrl == NULL) reg2_ctrl = queue_full_iff;
1649 phase->register_control(queue_full_iff, loop, ctrl);
1650 Node* not_full = new IfTrueNode(queue_full_iff);
1651 phase->register_control(not_full, loop, queue_full_iff);
1652 Node* full = new IfFalseNode(queue_full_iff);
1653 phase->register_control(full, loop, queue_full_iff);
1654
1655 ctrl = not_full;
1656
1657 Node* next_index = new SubXNode(index, phase->igvn().MakeConX(sizeof(intptr_t)));
1658 phase->register_new_node(next_index, ctrl);
1659
1660 Node* buffer = new LoadPNode(ctrl, raw_mem, buffer_adr, adr_type, TypeRawPtr::NOTNULL, MemNode::unordered);
1661 phase->register_new_node(buffer, ctrl);
1662 Node *log_addr = new AddPNode(phase->C->top(), buffer, next_index);
1663 phase->register_new_node(log_addr, ctrl);
1664 Node* log_store = new StorePNode(ctrl, raw_mem, log_addr, adr_type, pre_val, MemNode::unordered);
1665 phase->register_new_node(log_store, ctrl);
1666 // update the index
1667 Node* index_update = new StoreXNode(ctrl, log_store, index_adr, adr_type, next_index, MemNode::unordered);
1668 phase->register_new_node(index_update, ctrl);
1669
1670 // Fast-path case
1671 region2->init_req(_fast_path, ctrl);
1672 phi2->init_req(_fast_path, index_update);
1673
1674 ctrl = full;
1675
1676 Node* base = find_bottom_mem(ctrl, phase);
1677
1678 MergeMemNode* mm = MergeMemNode::make(base);
1679 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
1680 phase->register_new_node(mm, ctrl);
1681
1682 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);
1683 call->init_req(TypeFunc::Control, ctrl);
1684 call->init_req(TypeFunc::I_O, phase->C->top());
1685 call->init_req(TypeFunc::Memory, mm);
1686 call->init_req(TypeFunc::FramePtr, phase->C->top());
1687 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1688 call->init_req(TypeFunc::Parms, pre_val);
1689 call->init_req(TypeFunc::Parms+1, thread);
1690 phase->register_control(call, loop, ctrl);
1691
1692 Node* ctrl_proj = new ProjNode(call, TypeFunc::Control);
1693 phase->register_control(ctrl_proj, loop, call);
1694 Node* mem_proj = new ProjNode(call, TypeFunc::Memory);
1695 phase->register_new_node(mem_proj, call);
1696
1697 // Slow-path case
1698 region2->init_req(_slow_path, ctrl_proj);
1699 phi2->init_req(_slow_path, mem_proj);
1700
1701 phase->register_control(region2, loop, reg2_ctrl);
1702 phase->register_new_node(phi2, region2);
1703
1704 region->init_req(_heap_unstable, region2);
1705 phi->init_req(_heap_unstable, phi2);
1706
1707 phase->register_control(region, loop, heap_stable_ctrl->in(0));
1708 phase->register_new_node(phi, region);
1709
1710 fix_ctrl(barrier, region, fixer, uses, uses_to_ignore, last, phase);
1711 for(uint next = 0; next < uses.size(); next++ ) {
1712 Node *n = uses.at(next);
1713 assert(phase->get_ctrl(n) == init_ctrl, "bad control");
1714 assert(n != init_raw_mem, "should leave input raw mem above the barrier");
1715 phase->set_ctrl(n, region);
1716 follow_barrier_uses(n, init_ctrl, uses, phase);
1717 }
1718 fixer.fix_mem(init_ctrl, region, init_raw_mem, raw_mem_for_ctrl, phi, uses);
1719
1720 phase->igvn().replace_node(barrier, pre_val);
1721 }
1722 assert(state->enqueue_barriers_count() == 0, "all enqueue barrier nodes should have been replaced");
1723
1724 }
1725
1726 Node* ShenandoahBarrierC2Support::get_load_addr(PhaseIdealLoop* phase, VectorSet& visited, Node* in) {
1727 if (visited.test_set(in->_idx)) {
1728 return NULL;
1729 }
1730 switch (in->Opcode()) {
1731 case Op_Proj:
1732 return get_load_addr(phase, visited, in->in(0));
1733 case Op_CastPP:
1734 case Op_CheckCastPP:
1735 case Op_DecodeN:
1736 case Op_EncodeP:
1737 return get_load_addr(phase, visited, in->in(1));
1738 case Op_LoadN:
1739 case Op_LoadP:
1740 return in->in(MemNode::Address);
1741 case Op_CompareAndExchangeN:
1742 case Op_CompareAndExchangeP:
1743 case Op_GetAndSetN:
1744 case Op_GetAndSetP:
1745 case Op_ShenandoahCompareAndExchangeP:
1746 case Op_ShenandoahCompareAndExchangeN:
1747 // Those instructions would just have stored a different
1748 // value into the field. No use to attempt to fix it at this point.
1749 return phase->igvn().zerocon(T_OBJECT);
1750 case Op_CMoveP:
1751 case Op_CMoveN: {
1752 Node* t = get_load_addr(phase, visited, in->in(CMoveNode::IfTrue));
1753 Node* f = get_load_addr(phase, visited, in->in(CMoveNode::IfFalse));
1754 // Handle unambiguous cases: single address reported on both branches.
1755 if (t != NULL && f == NULL) return t;
1756 if (t == NULL && f != NULL) return f;
1757 if (t != NULL && t == f) return t;
1758 // Ambiguity.
1759 return phase->igvn().zerocon(T_OBJECT);
1760 }
1761 case Op_Phi: {
1762 Node* addr = NULL;
1763 for (uint i = 1; i < in->req(); i++) {
1764 Node* addr1 = get_load_addr(phase, visited, in->in(i));
1765 if (addr == NULL) {
1766 addr = addr1;
1767 }
1768 if (addr != addr1) {
1769 return phase->igvn().zerocon(T_OBJECT);
1770 }
1771 }
1772 return addr;
1773 }
1774 case Op_ShenandoahLoadReferenceBarrier:
1775 return get_load_addr(phase, visited, in->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1776 case Op_ShenandoahEnqueueBarrier:
1777 return get_load_addr(phase, visited, in->in(1));
1778 case Op_CallDynamicJava:
1779 case Op_CallLeaf:
1780 case Op_CallStaticJava:
1781 case Op_ConN:
1782 case Op_ConP:
1783 case Op_Parm:
1784 case Op_CreateEx:
1785 return phase->igvn().zerocon(T_OBJECT);
1786 default:
1787 #ifdef ASSERT
1788 fatal("Unknown node in get_load_addr: %s", NodeClassNames[in->Opcode()]);
1789 #endif
1790 return phase->igvn().zerocon(T_OBJECT);
1791 }
1792
1793 }
1794
1795 void ShenandoahBarrierC2Support::move_heap_stable_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
1796 IdealLoopTree *loop = phase->get_loop(iff);
1797 Node* loop_head = loop->_head;
1798 Node* entry_c = loop_head->in(LoopNode::EntryControl);
1799
1800 Node* bol = iff->in(1);
1801 Node* cmp = bol->in(1);
1802 Node* andi = cmp->in(1);
1803 Node* load = andi->in(1);
1804
1805 assert(is_gc_state_load(load), "broken");
1806 if (!phase->is_dominator(load->in(0), entry_c)) {
1807 Node* mem_ctrl = NULL;
1808 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
1809 load = load->clone();
1810 load->set_req(MemNode::Memory, mem);
1811 load->set_req(0, entry_c);
1812 phase->register_new_node(load, entry_c);
1813 andi = andi->clone();
1814 andi->set_req(1, load);
1815 phase->register_new_node(andi, entry_c);
1816 cmp = cmp->clone();
1817 cmp->set_req(1, andi);
1818 phase->register_new_node(cmp, entry_c);
1819 bol = bol->clone();
1820 bol->set_req(1, cmp);
1821 phase->register_new_node(bol, entry_c);
1822
1823 Node* old_bol =iff->in(1);
1824 phase->igvn().replace_input_of(iff, 1, bol);
1825 }
1826 }
1827
1828 bool ShenandoahBarrierC2Support::identical_backtoback_ifs(Node* n, PhaseIdealLoop* phase) {
1829 if (!n->is_If() || n->is_CountedLoopEnd()) {
1830 return false;
1831 }
1832 Node* region = n->in(0);
1833
1834 if (!region->is_Region()) {
1835 return false;
1836 }
1837 Node* dom = phase->idom(region);
1838 if (!dom->is_If()) {
1839 return false;
1840 }
1841
1842 if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) {
1843 return false;
1844 }
1845
1846 IfNode* dom_if = dom->as_If();
1847 Node* proj_true = dom_if->proj_out(1);
1848 Node* proj_false = dom_if->proj_out(0);
1849
1850 for (uint i = 1; i < region->req(); i++) {
1851 if (phase->is_dominator(proj_true, region->in(i))) {
1852 continue;
1853 }
1854 if (phase->is_dominator(proj_false, region->in(i))) {
1855 continue;
1856 }
1857 return false;
1858 }
1859
1860 return true;
1861 }
1862
1863 void ShenandoahBarrierC2Support::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) {
1864 assert(is_heap_stable_test(n), "no other tests");
1865 if (identical_backtoback_ifs(n, phase)) {
1866 Node* n_ctrl = n->in(0);
1867 if (phase->can_split_if(n_ctrl)) {
1868 IfNode* dom_if = phase->idom(n_ctrl)->as_If();
1869 if (is_heap_stable_test(n)) {
1870 Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1);
1871 assert(is_gc_state_load(gc_state_load), "broken");
1872 Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1);
1873 assert(is_gc_state_load(dom_gc_state_load), "broken");
1874 if (gc_state_load != dom_gc_state_load) {
1875 phase->igvn().replace_node(gc_state_load, dom_gc_state_load);
1876 }
1877 }
1878 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1879 Node* proj_true = dom_if->proj_out(1);
1880 Node* proj_false = dom_if->proj_out(0);
1881 Node* con_true = phase->igvn().makecon(TypeInt::ONE);
1882 Node* con_false = phase->igvn().makecon(TypeInt::ZERO);
1883
1884 for (uint i = 1; i < n_ctrl->req(); i++) {
1885 if (phase->is_dominator(proj_true, n_ctrl->in(i))) {
1886 bolphi->init_req(i, con_true);
1887 } else {
1888 assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1889 bolphi->init_req(i, con_false);
1890 }
1891 }
1892 phase->register_new_node(bolphi, n_ctrl);
1893 phase->igvn().replace_input_of(n, 1, bolphi);
1894 phase->do_split_if(n);
1895 }
1896 }
1897 }
1898
1899 IfNode* ShenandoahBarrierC2Support::find_unswitching_candidate(const IdealLoopTree* loop, PhaseIdealLoop* phase) {
1900 // Find first invariant test that doesn't exit the loop
1901 LoopNode *head = loop->_head->as_Loop();
1902 IfNode* unswitch_iff = NULL;
1903 Node* n = head->in(LoopNode::LoopBackControl);
1904 int loop_has_sfpts = -1;
1905 while (n != head) {
1906 Node* n_dom = phase->idom(n);
1907 if (n->is_Region()) {
1908 if (n_dom->is_If()) {
1909 IfNode* iff = n_dom->as_If();
1910 if (iff->in(1)->is_Bool()) {
1911 BoolNode* bol = iff->in(1)->as_Bool();
1912 if (bol->in(1)->is_Cmp()) {
1913 // If condition is invariant and not a loop exit,
1914 // then found reason to unswitch.
1915 if (is_heap_stable_test(iff) &&
1916 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) {
1917 assert(!loop->is_loop_exit(iff), "both branches should be in the loop");
1918 if (loop_has_sfpts == -1) {
1919 for(uint i = 0; i < loop->_body.size(); i++) {
1920 Node *m = loop->_body[i];
1921 if (m->is_SafePoint() && !m->is_CallLeaf()) {
1922 loop_has_sfpts = 1;
1923 break;
1924 }
1925 }
1926 if (loop_has_sfpts == -1) {
1927 loop_has_sfpts = 0;
1928 }
1929 }
1930 if (!loop_has_sfpts) {
1931 unswitch_iff = iff;
1932 }
1933 }
1934 }
1935 }
1936 }
1937 }
1938 n = n_dom;
1939 }
1940 return unswitch_iff;
1941 }
1942
1943
1944 void ShenandoahBarrierC2Support::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) {
1945 Node_List heap_stable_tests;
1946 stack.push(phase->C->start(), 0);
1947 do {
1948 Node* n = stack.node();
1949 uint i = stack.index();
1950
1951 if (i < n->outcnt()) {
1952 Node* u = n->raw_out(i);
1953 stack.set_index(i+1);
1954 if (!visited.test_set(u->_idx)) {
1955 stack.push(u, 0);
1956 }
1957 } else {
1958 stack.pop();
1959 if (n->is_If() && is_heap_stable_test(n)) {
1960 heap_stable_tests.push(n);
1961 }
1962 }
1963 } while (stack.size() > 0);
1964
1965 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1966 Node* n = heap_stable_tests.at(i);
1967 assert(is_heap_stable_test(n), "only evacuation test");
1968 merge_back_to_back_tests(n, phase);
1969 }
1970
1971 if (!phase->C->major_progress()) {
1972 VectorSet seen(Thread::current()->resource_area());
1973 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1974 Node* n = heap_stable_tests.at(i);
1975 IdealLoopTree* loop = phase->get_loop(n);
1976 if (loop != phase->ltree_root() &&
1977 loop->_child == NULL &&
1978 !loop->_irreducible) {
1979 Node* head = loop->_head;
1980 if (head->is_Loop() &&
1981 (!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) &&
1982 !seen.test_set(head->_idx)) {
1983 IfNode* iff = find_unswitching_candidate(loop, phase);
1984 if (iff != NULL) {
1985 Node* bol = iff->in(1);
1986 if (head->as_Loop()->is_strip_mined()) {
1987 head->as_Loop()->verify_strip_mined(0);
1988 }
1989 move_heap_stable_test_out_of_loop(iff, phase);
1990
1991 AutoNodeBudget node_budget(phase);
1992
1993 if (loop->policy_unswitching(phase)) {
1994 if (head->as_Loop()->is_strip_mined()) {
1995 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop();
1996 hide_strip_mined_loop(outer, head->as_CountedLoop(), phase);
1997 }
1998 phase->do_unswitching(loop, old_new);
1999 } else {
2000 // Not proceeding with unswitching. Move load back in
2001 // the loop.
2002 phase->igvn().replace_input_of(iff, 1, bol);
2003 }
2004 }
2005 }
2006 }
2007 }
2008 }
2009 }
2010
2011 #ifdef ASSERT
2012 void ShenandoahBarrierC2Support::verify_raw_mem(RootNode* root) {
2013 const bool trace = false;
2014 ResourceMark rm;
2015 Unique_Node_List nodes;
2016 Unique_Node_List controls;
2017 Unique_Node_List memories;
2018
2019 nodes.push(root);
2020 for (uint next = 0; next < nodes.size(); next++) {
2021 Node *n = nodes.at(next);
2022 if (ShenandoahBarrierSetC2::is_shenandoah_lrb_call(n)) {
2023 controls.push(n);
2024 if (trace) { tty->print("XXXXXX verifying"); n->dump(); }
2025 for (uint next2 = 0; next2 < controls.size(); next2++) {
2026 Node *m = controls.at(next2);
2027 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
2028 Node* u = m->fast_out(i);
2029 if (u->is_CFG() && !u->is_Root() &&
2030 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1) &&
2031 !(u->is_Region() && u->unique_ctrl_out()->Opcode() == Op_Halt)) {
2032 if (trace) { tty->print("XXXXXX pushing control"); u->dump(); }
2033 controls.push(u);
2034 }
2035 }
2036 }
2037 memories.push(n->as_Call()->proj_out(TypeFunc::Memory));
2038 for (uint next2 = 0; next2 < memories.size(); next2++) {
2039 Node *m = memories.at(next2);
2040 assert(m->bottom_type() == Type::MEMORY, "");
2041 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
2042 Node* u = m->fast_out(i);
2043 if (u->bottom_type() == Type::MEMORY && (u->is_Mem() || u->is_ClearArray())) {
2044 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
2045 memories.push(u);
2046 } else if (u->is_LoadStore()) {
2047 if (trace) { tty->print("XXXXXX pushing memory"); u->find_out_with(Op_SCMemProj)->dump(); }
2048 memories.push(u->find_out_with(Op_SCMemProj));
2049 } else if (u->is_MergeMem() && u->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == m) {
2050 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
2051 memories.push(u);
2052 } else if (u->is_Phi()) {
2053 assert(u->bottom_type() == Type::MEMORY, "");
2054 if (u->adr_type() == TypeRawPtr::BOTTOM || u->adr_type() == TypePtr::BOTTOM) {
2055 assert(controls.member(u->in(0)), "");
2056 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
2057 memories.push(u);
2058 }
2059 } else if (u->is_SafePoint() || u->is_MemBar()) {
2060 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2061 Node* uu = u->fast_out(j);
2062 if (uu->bottom_type() == Type::MEMORY) {
2063 if (trace) { tty->print("XXXXXX pushing memory"); uu->dump(); }
2064 memories.push(uu);
2065 }
2066 }
2067 }
2068 }
2069 }
2070 for (uint next2 = 0; next2 < controls.size(); next2++) {
2071 Node *m = controls.at(next2);
2072 if (m->is_Region()) {
2073 bool all_in = true;
2074 for (uint i = 1; i < m->req(); i++) {
2075 if (!controls.member(m->in(i))) {
2076 all_in = false;
2077 break;
2078 }
2079 }
2080 if (trace) { tty->print("XXX verifying %s", all_in ? "all in" : ""); m->dump(); }
2081 bool found_phi = false;
2082 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax && !found_phi; j++) {
2083 Node* u = m->fast_out(j);
2084 if (u->is_Phi() && memories.member(u)) {
2085 found_phi = true;
2086 for (uint i = 1; i < u->req() && found_phi; i++) {
2087 Node* k = u->in(i);
2088 if (memories.member(k) != controls.member(m->in(i))) {
2089 found_phi = false;
2090 }
2091 }
2092 }
2093 }
2094 assert(found_phi || all_in, "");
2095 }
2096 }
2097 controls.clear();
2098 memories.clear();
2099 }
2100 for( uint i = 0; i < n->len(); ++i ) {
2101 Node *m = n->in(i);
2102 if (m != NULL) {
2103 nodes.push(m);
2104 }
2105 }
2106 }
2107 }
2108 #endif
2109
2110 ShenandoahEnqueueBarrierNode::ShenandoahEnqueueBarrierNode(Node* val) : Node(NULL, val) {
2111 ShenandoahBarrierSetC2::bsc2()->state()->add_enqueue_barrier(this);
2112 }
2113
2114 const Type* ShenandoahEnqueueBarrierNode::bottom_type() const {
2115 if (in(1) == NULL || in(1)->is_top()) {
2116 return Type::TOP;
2117 }
2118 const Type* t = in(1)->bottom_type();
2119 if (t == TypePtr::NULL_PTR) {
2120 return t;
2121 }
2122 return t->is_oopptr();
2123 }
2124
2125 const Type* ShenandoahEnqueueBarrierNode::Value(PhaseGVN* phase) const {
2126 if (in(1) == NULL) {
2127 return Type::TOP;
2128 }
2129 const Type* t = phase->type(in(1));
2130 if (t == Type::TOP) {
2131 return Type::TOP;
2132 }
2133 if (t == TypePtr::NULL_PTR) {
2134 return t;
2135 }
2136 return t->is_oopptr();
2137 }
2138
2139 int ShenandoahEnqueueBarrierNode::needed(Node* n) {
2140 if (n == NULL ||
2141 n->is_Allocate() ||
2142 n->Opcode() == Op_ShenandoahEnqueueBarrier ||
2143 n->bottom_type() == TypePtr::NULL_PTR ||
2144 (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL)) {
2145 return NotNeeded;
2146 }
2147 if (n->is_Phi() ||
2148 n->is_CMove()) {
2149 return MaybeNeeded;
2150 }
2151 return Needed;
2152 }
2153
2154 Node* ShenandoahEnqueueBarrierNode::next(Node* n) {
2155 for (;;) {
2156 if (n == NULL) {
2157 return n;
2158 } else if (n->bottom_type() == TypePtr::NULL_PTR) {
2159 return n;
2160 } else if (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL) {
2161 return n;
2162 } else if (n->is_ConstraintCast() ||
2163 n->Opcode() == Op_DecodeN ||
2164 n->Opcode() == Op_EncodeP) {
2165 n = n->in(1);
2166 } else if (n->is_Proj()) {
2167 n = n->in(0);
2168 } else {
2169 return n;
2170 }
2171 }
2172 ShouldNotReachHere();
2173 return NULL;
2174 }
2175
2176 bool ShenandoahEnqueueBarrierNode::can_eliminate(PhaseIdealLoop* phase) {
2177 return ShenandoahHeap::heap()->traversal_gc() == NULL &&
2178 is_redundant() && ShenandoahAggressiveReferenceDiscovery;
2179 }
2180
2181 bool ShenandoahEnqueueBarrierNode::is_redundant() {
2182 Unique_Node_List visited;
2183 Node_Stack stack(0);
2184 stack.push(this, 0);
2185
2186 while (stack.size() > 0) {
2187 Node* n = stack.node();
2188 if (visited.member(n)) {
2189 stack.pop();
2190 continue;
2191 }
2192 visited.push(n);
2193 bool visit_users = false;
2194 switch (n->Opcode()) {
2195 case Op_CallStaticJava:
2196 if (n->as_CallStaticJava()->uncommon_trap_request() == 0) {
2197 return false;
2198 }
2199 break;
2200 case Op_CallDynamicJava:
2201 case Op_CompareAndExchangeN:
2202 case Op_CompareAndExchangeP:
2203 case Op_CompareAndSwapN:
2204 case Op_CompareAndSwapP:
2205 case Op_ShenandoahCompareAndSwapN:
2206 case Op_ShenandoahCompareAndSwapP:
2207 case Op_GetAndSetN:
2208 case Op_GetAndSetP:
2209 case Op_Return:
2210 case Op_StoreN:
2211 case Op_StoreP:
2212 return false;
2213 break;
2214 case Op_AddP:
2215 case Op_Allocate:
2216 case Op_AllocateArray:
2217 case Op_ArrayCopy:
2218 case Op_CmpP:
2219 case Op_LoadL:
2220 case Op_SafePoint:
2221 case Op_SubTypeCheck:
2222 case Op_StoreLConditional:
2223 case Op_StoreIConditional:
2224 case Op_FastUnlock:
2225 break;
2226 case Op_CastPP:
2227 case Op_CheckCastPP:
2228 case Op_CMoveN:
2229 case Op_CMoveP:
2230 case Op_EncodeP:
2231 case Op_Phi:
2232 case Op_ShenandoahEnqueueBarrier:
2233 visit_users = true;
2234 break;
2235 default: {
2236 #ifdef ASSERT
2237 fatal("Unknown node in is_redundant: %s", NodeClassNames[n->Opcode()]);
2238 #endif
2239 // Default to useful: better to have excess barriers, rather than miss some.
2240 return false;
2241 }
2242 }
2243
2244 stack.pop();
2245 if (visit_users) {
2246 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2247 Node* user = n->fast_out(i);
2248 if (user != NULL) {
2249 stack.push(user, 0);
2250 }
2251 }
2252 }
2253 }
2254 return true;
2255 }
2256
2257 Node* ShenandoahEnqueueBarrierNode::Identity(PhaseGVN* phase) {
2258 PhaseIterGVN* igvn = phase->is_IterGVN();
2259
2260 Node* n = next(in(1));
2261
2262 int cont = needed(n);
2263
2264 if (cont == NotNeeded) {
2265 return in(1);
2266 } else if (cont == MaybeNeeded) {
2267 if (igvn == NULL) {
2268 phase->record_for_igvn(this);
2269 return this;
2270 } else {
2271 ResourceMark rm;
2272 Unique_Node_List wq;
2273 uint wq_i = 0;
2274
2275 for (;;) {
2276 if (n->is_Phi()) {
2277 for (uint i = 1; i < n->req(); i++) {
2278 Node* m = n->in(i);
2279 if (m != NULL) {
2280 wq.push(m);
2281 }
2282 }
2283 } else {
2284 assert(n->is_CMove(), "nothing else here");
2285 Node* m = n->in(CMoveNode::IfFalse);
2286 wq.push(m);
2287 m = n->in(CMoveNode::IfTrue);
2288 wq.push(m);
2289 }
2290 Node* orig_n = NULL;
2291 do {
2292 if (wq_i >= wq.size()) {
2293 return in(1);
2294 }
2295 n = wq.at(wq_i);
2296 wq_i++;
2297 orig_n = n;
2298 n = next(n);
2299 cont = needed(n);
2300 if (cont == Needed) {
2301 return this;
2302 }
2303 } while (cont != MaybeNeeded || (orig_n != n && wq.member(n)));
2304 }
2305 }
2306 }
2307
2308 return this;
2309 }
2310
2311 #ifdef ASSERT
2312 static bool has_never_branch(Node* root) {
2313 for (uint i = 1; i < root->req(); i++) {
2314 Node* in = root->in(i);
2315 if (in != NULL && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->Opcode() == Op_NeverBranch) {
2316 return true;
2317 }
2318 }
2319 return false;
2320 }
2321 #endif
2322
2323 void MemoryGraphFixer::collect_memory_nodes() {
2324 Node_Stack stack(0);
2325 VectorSet visited(Thread::current()->resource_area());
2326 Node_List regions;
2327
2328 // Walk the raw memory graph and create a mapping from CFG node to
2329 // memory node. Exclude phis for now.
2330 stack.push(_phase->C->root(), 1);
2331 do {
2332 Node* n = stack.node();
2333 int opc = n->Opcode();
2334 uint i = stack.index();
2335 if (i < n->req()) {
2336 Node* mem = NULL;
2337 if (opc == Op_Root) {
2338 Node* in = n->in(i);
2339 int in_opc = in->Opcode();
2340 if (in_opc == Op_Return || in_opc == Op_Rethrow) {
2341 mem = in->in(TypeFunc::Memory);
2342 } else if (in_opc == Op_Halt) {
2343 if (in->in(0)->is_Region()) {
2344 Node* r = in->in(0);
2345 for (uint j = 1; j < r->req(); j++) {
2346 assert(r->in(j)->Opcode() != Op_NeverBranch, "");
2347 }
2348 } else {
2349 Node* proj = in->in(0);
2350 assert(proj->is_Proj(), "");
2351 Node* in = proj->in(0);
2352 assert(in->is_CallStaticJava() || in->Opcode() == Op_NeverBranch || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
2353 if (in->is_CallStaticJava()) {
2354 mem = in->in(TypeFunc::Memory);
2355 } else if (in->Opcode() == Op_Catch) {
2356 Node* call = in->in(0)->in(0);
2357 assert(call->is_Call(), "");
2358 mem = call->in(TypeFunc::Memory);
2359 } else if (in->Opcode() == Op_NeverBranch) {
2360 Node* head = in->in(0);
2361 assert(head->is_Region() && head->req() == 3, "unexpected infinite loop graph shape");
2362 assert(_phase->is_dominator(head, head->in(1)) || _phase->is_dominator(head, head->in(2)), "no back branch?");
2363 Node* tail = _phase->is_dominator(head, head->in(1)) ? head->in(1) : head->in(2);
2364 Node* c = tail;
2365 while (c != head) {
2366 if (c->is_SafePoint() && !c->is_CallLeaf()) {
2367 mem = c->in(TypeFunc::Memory);
2368 }
2369 c = _phase->idom(c);
2370 }
2371 assert(mem != NULL, "should have found safepoint");
2372
2373 Node* phi_mem = NULL;
2374 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2375 Node* u = head->fast_out(j);
2376 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
2377 if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2378 assert(phi_mem == NULL || phi_mem->adr_type() == TypePtr::BOTTOM, "");
2379 phi_mem = u;
2380 } else if (u->adr_type() == TypePtr::BOTTOM) {
2381 assert(phi_mem == NULL || _phase->C->get_alias_index(phi_mem->adr_type()) == _alias, "");
2382 if (phi_mem == NULL) {
2383 phi_mem = u;
2384 }
2385 }
2386 }
2387 }
2388 if (phi_mem != NULL) {
2389 mem = phi_mem;
2390 }
2391 }
2392 }
2393 } else {
2394 #ifdef ASSERT
2395 n->dump();
2396 in->dump();
2397 #endif
2398 ShouldNotReachHere();
2399 }
2400 } else {
2401 assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
2402 assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
2403 mem = n->in(i);
2404 }
2405 i++;
2406 stack.set_index(i);
2407 if (mem == NULL) {
2408 continue;
2409 }
2410 for (;;) {
2411 if (visited.test_set(mem->_idx) || mem->is_Start()) {
2412 break;
2413 }
2414 if (mem->is_Phi()) {
2415 stack.push(mem, 2);
2416 mem = mem->in(1);
2417 } else if (mem->is_Proj()) {
2418 stack.push(mem, mem->req());
2419 mem = mem->in(0);
2420 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
2421 mem = mem->in(TypeFunc::Memory);
2422 } else if (mem->is_MergeMem()) {
2423 MergeMemNode* mm = mem->as_MergeMem();
2424 mem = mm->memory_at(_alias);
2425 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
2426 assert(_alias == Compile::AliasIdxRaw, "");
2427 stack.push(mem, mem->req());
2428 mem = mem->in(MemNode::Memory);
2429 } else {
2430 #ifdef ASSERT
2431 mem->dump();
2432 #endif
2433 ShouldNotReachHere();
2434 }
2435 }
2436 } else {
2437 if (n->is_Phi()) {
2438 // Nothing
2439 } else if (!n->is_Root()) {
2440 Node* c = get_ctrl(n);
2441 _memory_nodes.map(c->_idx, n);
2442 }
2443 stack.pop();
2444 }
2445 } while(stack.is_nonempty());
2446
2447 // Iterate over CFG nodes in rpo and propagate memory state to
2448 // compute memory state at regions, creating new phis if needed.
2449 Node_List rpo_list;
2450 visited.clear();
2451 _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
2452 Node* root = rpo_list.pop();
2453 assert(root == _phase->C->root(), "");
2454
2455 const bool trace = false;
2456 #ifdef ASSERT
2457 if (trace) {
2458 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2459 Node* c = rpo_list.at(i);
2460 if (_memory_nodes[c->_idx] != NULL) {
2461 tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump();
2462 }
2463 }
2464 }
2465 #endif
2466 uint last = _phase->C->unique();
2467
2468 #ifdef ASSERT
2469 uint8_t max_depth = 0;
2470 for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
2471 IdealLoopTree* lpt = iter.current();
2472 max_depth = MAX2(max_depth, lpt->_nest);
2473 }
2474 #endif
2475
2476 bool progress = true;
2477 int iteration = 0;
2478 Node_List dead_phis;
2479 while (progress) {
2480 progress = false;
2481 iteration++;
2482 assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2483 if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
2484 IdealLoopTree* last_updated_ilt = NULL;
2485 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2486 Node* c = rpo_list.at(i);
2487
2488 Node* prev_mem = _memory_nodes[c->_idx];
2489 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2490 Node* prev_region = regions[c->_idx];
2491 Node* unique = NULL;
2492 for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
2493 Node* m = _memory_nodes[c->in(j)->_idx];
2494 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");
2495 if (m != NULL) {
2496 if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
2497 assert(c->is_Loop() && j == LoopNode::LoopBackControl || _phase->C->has_irreducible_loop(), "");
2498 // continue
2499 } else if (unique == NULL) {
2500 unique = m;
2501 } else if (m == unique) {
2502 // continue
2503 } else {
2504 unique = NodeSentinel;
2505 }
2506 }
2507 }
2508 assert(unique != NULL, "empty phi???");
2509 if (unique != NodeSentinel) {
2510 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c) {
2511 dead_phis.push(prev_region);
2512 }
2513 regions.map(c->_idx, unique);
2514 } else {
2515 Node* phi = NULL;
2516 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
2517 phi = prev_region;
2518 for (uint k = 1; k < c->req(); k++) {
2519 Node* m = _memory_nodes[c->in(k)->_idx];
2520 assert(m != NULL, "expect memory state");
2521 phi->set_req(k, m);
2522 }
2523 } else {
2524 for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == NULL; j++) {
2525 Node* u = c->fast_out(j);
2526 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2527 (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
2528 phi = u;
2529 for (uint k = 1; k < c->req() && phi != NULL; k++) {
2530 Node* m = _memory_nodes[c->in(k)->_idx];
2531 assert(m != NULL, "expect memory state");
2532 if (u->in(k) != m) {
2533 phi = NULL;
2534 }
2535 }
2536 }
2537 }
2538 if (phi == NULL) {
2539 phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
2540 for (uint k = 1; k < c->req(); k++) {
2541 Node* m = _memory_nodes[c->in(k)->_idx];
2542 assert(m != NULL, "expect memory state");
2543 phi->init_req(k, m);
2544 }
2545 }
2546 }
2547 assert(phi != NULL, "");
2548 regions.map(c->_idx, phi);
2549 }
2550 Node* current_region = regions[c->_idx];
2551 if (current_region != prev_region) {
2552 progress = true;
2553 if (prev_region == prev_mem) {
2554 _memory_nodes.map(c->_idx, current_region);
2555 }
2556 }
2557 } else if (prev_mem == NULL || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
2558 Node* m = _memory_nodes[_phase->idom(c)->_idx];
2559 assert(m != NULL, "expect memory state");
2560 if (m != prev_mem) {
2561 _memory_nodes.map(c->_idx, m);
2562 progress = true;
2563 }
2564 }
2565 #ifdef ASSERT
2566 if (trace) { tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); }
2567 #endif
2568 }
2569 }
2570
2571 // Replace existing phi with computed memory state for that region
2572 // if different (could be a new phi or a dominating memory node if
2573 // that phi was found to be useless).
2574 while (dead_phis.size() > 0) {
2575 Node* n = dead_phis.pop();
2576 n->replace_by(_phase->C->top());
2577 n->destruct();
2578 }
2579 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2580 Node* c = rpo_list.at(i);
2581 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2582 Node* n = regions[c->_idx];
2583 if (n->is_Phi() && n->_idx >= last && n->in(0) == c) {
2584 _phase->register_new_node(n, c);
2585 }
2586 }
2587 }
2588 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2589 Node* c = rpo_list.at(i);
2590 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2591 Node* n = regions[c->_idx];
2592 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
2593 Node* u = c->fast_out(i);
2594 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2595 u != n) {
2596 if (u->adr_type() == TypePtr::BOTTOM) {
2597 fix_memory_uses(u, n, n, c);
2598 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2599 _phase->lazy_replace(u, n);
2600 --i; --imax;
2601 }
2602 }
2603 }
2604 }
2605 }
2606 }
2607
2608 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
2609 Node* c = _phase->get_ctrl(n);
2610 if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Call()) {
2611 assert(c == n->in(0), "");
2612 CallNode* call = c->as_Call();
2613 CallProjections projs;
2614 call->extract_projections(&projs, true, false);
2615 if (projs.catchall_memproj != NULL) {
2616 if (projs.fallthrough_memproj == n) {
2617 c = projs.fallthrough_catchproj;
2618 } else {
2619 assert(projs.catchall_memproj == n, "");
2620 c = projs.catchall_catchproj;
2621 }
2622 }
2623 }
2624 return c;
2625 }
2626
2627 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
2628 if (_phase->has_ctrl(n))
2629 return get_ctrl(n);
2630 else {
2631 assert (n->is_CFG(), "must be a CFG node");
2632 return n;
2633 }
2634 }
2635
2636 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
2637 return m != NULL && get_ctrl(m) == c;
2638 }
2639
2640 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
2641 assert(n == NULL || _phase->ctrl_or_self(n) == ctrl, "");
2642 Node* mem = _memory_nodes[ctrl->_idx];
2643 Node* c = ctrl;
2644 while (!mem_is_valid(mem, c) &&
2645 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
2646 c = _phase->idom(c);
2647 mem = _memory_nodes[c->_idx];
2648 }
2649 if (n != NULL && mem_is_valid(mem, c)) {
2650 while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
2651 mem = next_mem(mem, _alias);
2652 }
2653 if (mem->is_MergeMem()) {
2654 mem = mem->as_MergeMem()->memory_at(_alias);
2655 }
2656 if (!mem_is_valid(mem, c)) {
2657 do {
2658 c = _phase->idom(c);
2659 mem = _memory_nodes[c->_idx];
2660 } while (!mem_is_valid(mem, c) &&
2661 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
2662 }
2663 }
2664 assert(mem->bottom_type() == Type::MEMORY, "");
2665 return mem;
2666 }
2667
2668 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
2669 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2670 Node* use = region->fast_out(i);
2671 if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
2672 (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
2673 return true;
2674 }
2675 }
2676 return false;
2677 }
2678
2679 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
2680 assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
2681 const bool trace = false;
2682 DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
2683 DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
2684 GrowableArray<Node*> phis;
2685 if (mem_for_ctrl != mem) {
2686 Node* old = mem_for_ctrl;
2687 Node* prev = NULL;
2688 while (old != mem) {
2689 prev = old;
2690 if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
2691 assert(_alias == Compile::AliasIdxRaw, "");
2692 old = old->in(MemNode::Memory);
2693 } else if (old->Opcode() == Op_SCMemProj) {
2694 assert(_alias == Compile::AliasIdxRaw, "");
2695 old = old->in(0);
2696 } else {
2697 ShouldNotReachHere();
2698 }
2699 }
2700 assert(prev != NULL, "");
2701 if (new_ctrl != ctrl) {
2702 _memory_nodes.map(ctrl->_idx, mem);
2703 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2704 }
2705 uint input = (uint)MemNode::Memory;
2706 _phase->igvn().replace_input_of(prev, input, new_mem);
2707 } else {
2708 uses.clear();
2709 _memory_nodes.map(new_ctrl->_idx, new_mem);
2710 uses.push(new_ctrl);
2711 for(uint next = 0; next < uses.size(); next++ ) {
2712 Node *n = uses.at(next);
2713 assert(n->is_CFG(), "");
2714 DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
2715 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2716 Node* u = n->fast_out(i);
2717 if (!u->is_Root() && u->is_CFG() && u != n) {
2718 Node* m = _memory_nodes[u->_idx];
2719 if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
2720 !has_mem_phi(u) &&
2721 u->unique_ctrl_out()->Opcode() != Op_Halt) {
2722 DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
2723 DEBUG_ONLY(if (trace && m != NULL) { tty->print("ZZZ mem"); m->dump(); });
2724
2725 if (!mem_is_valid(m, u) || !m->is_Phi()) {
2726 bool push = true;
2727 bool create_phi = true;
2728 if (_phase->is_dominator(new_ctrl, u)) {
2729 create_phi = false;
2730 } else if (!_phase->C->has_irreducible_loop()) {
2731 IdealLoopTree* loop = _phase->get_loop(ctrl);
2732 bool do_check = true;
2733 IdealLoopTree* l = loop;
2734 create_phi = false;
2735 while (l != _phase->ltree_root()) {
2736 Node* head = l->_head;
2737 if (head->in(0) == NULL) {
2738 head = _phase->get_ctrl(head);
2739 }
2740 if (_phase->is_dominator(head, u) && _phase->is_dominator(_phase->idom(u), head)) {
2741 create_phi = true;
2742 do_check = false;
2743 break;
2744 }
2745 l = l->_parent;
2746 }
2747
2748 if (do_check) {
2749 assert(!create_phi, "");
2750 IdealLoopTree* u_loop = _phase->get_loop(u);
2751 if (u_loop != _phase->ltree_root() && u_loop->is_member(loop)) {
2752 Node* c = ctrl;
2753 while (!_phase->is_dominator(c, u_loop->tail())) {
2754 c = _phase->idom(c);
2755 }
2756 if (!_phase->is_dominator(c, u)) {
2757 do_check = false;
2758 }
2759 }
2760 }
2761
2762 if (do_check && _phase->is_dominator(_phase->idom(u), new_ctrl)) {
2763 create_phi = true;
2764 }
2765 }
2766 if (create_phi) {
2767 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
2768 _phase->register_new_node(phi, u);
2769 phis.push(phi);
2770 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
2771 if (!mem_is_valid(m, u)) {
2772 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
2773 _memory_nodes.map(u->_idx, phi);
2774 } else {
2775 DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
2776 for (;;) {
2777 assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), "");
2778 Node* next = NULL;
2779 if (m->is_Proj()) {
2780 next = m->in(0);
2781 } else {
2782 assert(m->is_Mem() || m->is_LoadStore(), "");
2783 assert(_alias == Compile::AliasIdxRaw, "");
2784 next = m->in(MemNode::Memory);
2785 }
2786 if (_phase->get_ctrl(next) != u) {
2787 break;
2788 }
2789 if (next->is_MergeMem()) {
2790 assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
2791 break;
2792 }
2793 if (next->is_Phi()) {
2794 assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
2795 break;
2796 }
2797 m = next;
2798 }
2799
2800 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
2801 assert(m->is_Mem() || m->is_LoadStore(), "");
2802 uint input = (uint)MemNode::Memory;
2803 _phase->igvn().replace_input_of(m, input, phi);
2804 push = false;
2805 }
2806 } else {
2807 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
2808 }
2809 if (push) {
2810 uses.push(u);
2811 }
2812 }
2813 } else if (!mem_is_valid(m, u) &&
2814 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1)) {
2815 uses.push(u);
2816 }
2817 }
2818 }
2819 }
2820 for (int i = 0; i < phis.length(); i++) {
2821 Node* n = phis.at(i);
2822 Node* r = n->in(0);
2823 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
2824 for (uint j = 1; j < n->req(); j++) {
2825 Node* m = find_mem(r->in(j), NULL);
2826 _phase->igvn().replace_input_of(n, j, m);
2827 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
2828 }
2829 }
2830 }
2831 uint last = _phase->C->unique();
2832 MergeMemNode* mm = NULL;
2833 int alias = _alias;
2834 DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
2835 // Process loads first to not miss an anti-dependency: if the memory
2836 // edge of a store is updated before a load is processed then an
2837 // anti-dependency may be missed.
2838 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2839 Node* u = mem->out(i);
2840 if (u->_idx < last && u->is_Load() && _phase->C->get_alias_index(u->adr_type()) == alias) {
2841 Node* m = find_mem(_phase->get_ctrl(u), u);
2842 if (m != mem) {
2843 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2844 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2845 --i;
2846 }
2847 }
2848 }
2849 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2850 Node* u = mem->out(i);
2851 if (u->_idx < last) {
2852 if (u->is_Mem()) {
2853 if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2854 Node* m = find_mem(_phase->get_ctrl(u), u);
2855 if (m != mem) {
2856 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2857 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2858 --i;
2859 }
2860 }
2861 } else if (u->is_MergeMem()) {
2862 MergeMemNode* u_mm = u->as_MergeMem();
2863 if (u_mm->memory_at(alias) == mem) {
2864 MergeMemNode* newmm = NULL;
2865 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2866 Node* uu = u->fast_out(j);
2867 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2868 if (uu->is_Phi()) {
2869 assert(uu->adr_type() == TypePtr::BOTTOM, "");
2870 Node* region = uu->in(0);
2871 int nb = 0;
2872 for (uint k = 1; k < uu->req(); k++) {
2873 if (uu->in(k) == u) {
2874 Node* m = find_mem(region->in(k), NULL);
2875 if (m != mem) {
2876 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
2877 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2878 if (newmm != u) {
2879 _phase->igvn().replace_input_of(uu, k, newmm);
2880 nb++;
2881 --jmax;
2882 }
2883 }
2884 }
2885 }
2886 if (nb > 0) {
2887 --j;
2888 }
2889 } else {
2890 Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
2891 if (m != mem) {
2892 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
2893 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2894 if (newmm != u) {
2895 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2896 --j, --jmax;
2897 }
2898 }
2899 }
2900 }
2901 }
2902 } else if (u->is_Phi()) {
2903 assert(u->bottom_type() == Type::MEMORY, "what else?");
2904 if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
2905 Node* region = u->in(0);
2906 bool replaced = false;
2907 for (uint j = 1; j < u->req(); j++) {
2908 if (u->in(j) == mem) {
2909 Node* m = find_mem(region->in(j), NULL);
2910 Node* nnew = m;
2911 if (m != mem) {
2912 if (u->adr_type() == TypePtr::BOTTOM) {
2913 mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
2914 nnew = mm;
2915 }
2916 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
2917 _phase->igvn().replace_input_of(u, j, nnew);
2918 replaced = true;
2919 }
2920 }
2921 }
2922 if (replaced) {
2923 --i;
2924 }
2925 }
2926 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2927 u->adr_type() == NULL) {
2928 assert(u->adr_type() != NULL ||
2929 u->Opcode() == Op_Rethrow ||
2930 u->Opcode() == Op_Return ||
2931 u->Opcode() == Op_SafePoint ||
2932 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2933 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2934 u->Opcode() == Op_CallLeaf, "");
2935 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2936 if (m != mem) {
2937 mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
2938 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2939 --i;
2940 }
2941 } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2942 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2943 if (m != mem) {
2944 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2945 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2946 --i;
2947 }
2948 } else if (u->adr_type() != TypePtr::BOTTOM &&
2949 _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
2950 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2951 assert(m != mem, "");
2952 // u is on the wrong slice...
2953 assert(u->is_ClearArray(), "");
2954 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2955 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2956 --i;
2957 }
2958 }
2959 }
2960 #ifdef ASSERT
2961 assert(new_mem->outcnt() > 0, "");
2962 for (int i = 0; i < phis.length(); i++) {
2963 Node* n = phis.at(i);
2964 assert(n->outcnt() > 0, "new phi must have uses now");
2965 }
2966 #endif
2967 }
2968
2969 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
2970 MergeMemNode* mm = MergeMemNode::make(mem);
2971 mm->set_memory_at(_alias, rep_proj);
2972 _phase->register_new_node(mm, rep_ctrl);
2973 return mm;
2974 }
2975
2976 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
2977 MergeMemNode* newmm = NULL;
2978 MergeMemNode* u_mm = u->as_MergeMem();
2979 Node* c = _phase->get_ctrl(u);
2980 if (_phase->is_dominator(c, rep_ctrl)) {
2981 c = rep_ctrl;
2982 } else {
2983 assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
2984 }
2985 if (u->outcnt() == 1) {
2986 if (u->req() > (uint)_alias && u->in(_alias) == mem) {
2987 _phase->igvn().replace_input_of(u, _alias, rep_proj);
2988 --i;
2989 } else {
2990 _phase->igvn().rehash_node_delayed(u);
2991 u_mm->set_memory_at(_alias, rep_proj);
2992 }
2993 newmm = u_mm;
2994 _phase->set_ctrl_and_loop(u, c);
2995 } else {
2996 // can't simply clone u and then change one of its input because
2997 // it adds and then removes an edge which messes with the
2998 // DUIterator
2999 newmm = MergeMemNode::make(u_mm->base_memory());
3000 for (uint j = 0; j < u->req(); j++) {
3001 if (j < newmm->req()) {
3002 if (j == (uint)_alias) {
3003 newmm->set_req(j, rep_proj);
3004 } else if (newmm->in(j) != u->in(j)) {
3005 newmm->set_req(j, u->in(j));
3006 }
3007 } else if (j == (uint)_alias) {
3008 newmm->add_req(rep_proj);
3009 } else {
3010 newmm->add_req(u->in(j));
3011 }
3012 }
3013 if ((uint)_alias >= u->req()) {
3014 newmm->set_memory_at(_alias, rep_proj);
3015 }
3016 _phase->register_new_node(newmm, c);
3017 }
3018 return newmm;
3019 }
3020
3021 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
3022 if (phi->adr_type() == TypePtr::BOTTOM) {
3023 Node* region = phi->in(0);
3024 for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
3025 Node* uu = region->fast_out(j);
3026 if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
3027 return false;
3028 }
3029 }
3030 return true;
3031 }
3032 return _phase->C->get_alias_index(phi->adr_type()) == _alias;
3033 }
3034
3035 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
3036 uint last = _phase-> C->unique();
3037 MergeMemNode* mm = NULL;
3038 assert(mem->bottom_type() == Type::MEMORY, "");
3039 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
3040 Node* u = mem->out(i);
3041 if (u != replacement && u->_idx < last) {
3042 if (u->is_MergeMem()) {
3043 MergeMemNode* u_mm = u->as_MergeMem();
3044 if (u_mm->memory_at(_alias) == mem) {
3045 MergeMemNode* newmm = NULL;
3046 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
3047 Node* uu = u->fast_out(j);
3048 assert(!uu->is_MergeMem(), "chain of MergeMems?");
3049 if (uu->is_Phi()) {
3050 if (should_process_phi(uu)) {
3051 Node* region = uu->in(0);
3052 int nb = 0;
3053 for (uint k = 1; k < uu->req(); k++) {
3054 if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
3055 if (newmm == NULL) {
3056 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
3057 }
3058 if (newmm != u) {
3059 _phase->igvn().replace_input_of(uu, k, newmm);
3060 nb++;
3061 --jmax;
3062 }
3063 }
3064 }
3065 if (nb > 0) {
3066 --j;
3067 }
3068 }
3069 } else {
3070 if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
3071 if (newmm == NULL) {
3072 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
3073 }
3074 if (newmm != u) {
3075 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
3076 --j, --jmax;
3077 }
3078 }
3079 }
3080 }
3081 }
3082 } else if (u->is_Phi()) {
3083 assert(u->bottom_type() == Type::MEMORY, "what else?");
3084 Node* region = u->in(0);
3085 if (should_process_phi(u)) {
3086 bool replaced = false;
3087 for (uint j = 1; j < u->req(); j++) {
3088 if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
3089 Node* nnew = rep_proj;
3090 if (u->adr_type() == TypePtr::BOTTOM) {
3091 if (mm == NULL) {
3092 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
3093 }
3094 nnew = mm;
3095 }
3096 _phase->igvn().replace_input_of(u, j, nnew);
3097 replaced = true;
3098 }
3099 }
3100 if (replaced) {
3101 --i;
3102 }
3103
3104 }
3105 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
3106 u->adr_type() == NULL) {
3107 assert(u->adr_type() != NULL ||
3108 u->Opcode() == Op_Rethrow ||
3109 u->Opcode() == Op_Return ||
3110 u->Opcode() == Op_SafePoint ||
3111 u->Opcode() == Op_StoreIConditional ||
3112 u->Opcode() == Op_StoreLConditional ||
3113 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
3114 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
3115 u->Opcode() == Op_CallLeaf, "%s", u->Name());
3116 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
3117 if (mm == NULL) {
3118 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
3119 }
3120 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
3121 --i;
3122 }
3123 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
3124 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
3125 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
3126 --i;
3127 }
3128 }
3129 }
3130 }
3131 }
3132
3133 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj, bool native)
3134 : Node(ctrl, obj), _native(native) {
3135 ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this);
3136 }
3137
3138 bool ShenandoahLoadReferenceBarrierNode::is_native() const {
3139 return _native;
3140 }
3141
3142 uint ShenandoahLoadReferenceBarrierNode::size_of() const {
3143 return sizeof(*this);
3144 }
3145
3146 uint ShenandoahLoadReferenceBarrierNode::hash() const {
3147 return Node::hash() + (_native ? 1 : 0);
3148 }
3149
3150 bool ShenandoahLoadReferenceBarrierNode::cmp( const Node &n ) const {
3151 return Node::cmp(n) && n.Opcode() == Op_ShenandoahLoadReferenceBarrier &&
3152 _native == ((const ShenandoahLoadReferenceBarrierNode&)n)._native;
3153 }
3154
3155 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const {
3156 if (in(ValueIn) == NULL || in(ValueIn)->is_top()) {
3157 return Type::TOP;
3158 }
3159 const Type* t = in(ValueIn)->bottom_type();
3160 if (t == TypePtr::NULL_PTR) {
3161 return t;
3162 }
3163 return t->is_oopptr();
3164 }
3165
3166 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const {
3167 // Either input is TOP ==> the result is TOP
3168 const Type *t2 = phase->type(in(ValueIn));
3169 if( t2 == Type::TOP ) return Type::TOP;
3170
3171 if (t2 == TypePtr::NULL_PTR) {
3172 return t2;
3173 }
3174
3175 const Type* type = t2->is_oopptr();
3176 return type;
3177 }
3178
3179 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) {
3180 Node* value = in(ValueIn);
3181 if (!needs_barrier(phase, value)) {
3182 return value;
3183 }
3184 return this;
3185 }
3186
3187 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) {
3188 Unique_Node_List visited;
3189 return needs_barrier_impl(phase, n, visited);
3190 }
3191
3192 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) {
3193 if (n == NULL) return false;
3194 if (visited.member(n)) {
3195 return false; // Been there.
3196 }
3197 visited.push(n);
3198
3199 if (n->is_Allocate()) {
3200 // tty->print_cr("optimize barrier on alloc");
3201 return false;
3202 }
3203 if (n->is_Call()) {
3204 // tty->print_cr("optimize barrier on call");
3205 return false;
3206 }
3207
3208 const Type* type = phase->type(n);
3209 if (type == Type::TOP) {
3210 return false;
3211 }
3212 if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) {
3213 // tty->print_cr("optimize barrier on null");
3214 return false;
3215 }
3216 if (type->make_oopptr() && type->make_oopptr()->const_oop() != NULL) {
3217 // tty->print_cr("optimize barrier on constant");
3218 return false;
3219 }
3220
3221 switch (n->Opcode()) {
3222 case Op_AddP:
3223 return true; // TODO: Can refine?
3224 case Op_LoadP:
3225 case Op_ShenandoahCompareAndExchangeN:
3226 case Op_ShenandoahCompareAndExchangeP:
3227 case Op_CompareAndExchangeN:
3228 case Op_CompareAndExchangeP:
3229 case Op_GetAndSetN:
3230 case Op_GetAndSetP:
3231 return true;
3232 case Op_Phi: {
3233 for (uint i = 1; i < n->req(); i++) {
3234 if (needs_barrier_impl(phase, n->in(i), visited)) return true;
3235 }
3236 return false;
3237 }
3238 case Op_CheckCastPP:
3239 case Op_CastPP:
3240 return needs_barrier_impl(phase, n->in(1), visited);
3241 case Op_Proj:
3242 return needs_barrier_impl(phase, n->in(0), visited);
3243 case Op_ShenandoahLoadReferenceBarrier:
3244 // tty->print_cr("optimize barrier on barrier");
3245 return false;
3246 case Op_Parm:
3247 // tty->print_cr("optimize barrier on input arg");
3248 return false;
3249 case Op_DecodeN:
3250 case Op_EncodeP:
3251 return needs_barrier_impl(phase, n->in(1), visited);
3252 case Op_LoadN:
3253 return true;
3254 case Op_CMoveN:
3255 case Op_CMoveP:
3256 return needs_barrier_impl(phase, n->in(2), visited) ||
3257 needs_barrier_impl(phase, n->in(3), visited);
3258 case Op_ShenandoahEnqueueBarrier:
3259 return needs_barrier_impl(phase, n->in(1), visited);
3260 case Op_CreateEx:
3261 return false;
3262 default:
3263 break;
3264 }
3265 #ifdef ASSERT
3266 tty->print("need barrier on?: ");
3267 tty->print_cr("ins:");
3268 n->dump(2);
3269 tty->print_cr("outs:");
3270 n->dump(-2);
3271 ShouldNotReachHere();
3272 #endif
3273 return true;
3274 }
3275
3276 bool ShenandoahLoadReferenceBarrierNode::is_redundant() {
3277 Unique_Node_List visited;
3278 Node_Stack stack(0);
3279 stack.push(this, 0);
3280
3281 // Check if the barrier is actually useful: go over nodes looking for useful uses
3282 // (e.g. memory accesses). Stop once we detected a required use. Otherwise, walk
3283 // until we ran out of nodes, and then declare the barrier redundant.
3284 while (stack.size() > 0) {
3285 Node* n = stack.node();
3286 if (visited.member(n)) {
3287 stack.pop();
3288 continue;
3289 }
3290 visited.push(n);
3291 bool visit_users = false;
3292 switch (n->Opcode()) {
3293 case Op_CallStaticJava:
3294 case Op_CallDynamicJava:
3295 case Op_CallLeaf:
3296 case Op_CallLeafNoFP:
3297 case Op_CompareAndSwapL:
3298 case Op_CompareAndSwapI:
3299 case Op_CompareAndSwapB:
3300 case Op_CompareAndSwapS:
3301 case Op_CompareAndSwapN:
3302 case Op_CompareAndSwapP:
3303 case Op_CompareAndExchangeL:
3304 case Op_CompareAndExchangeI:
3305 case Op_CompareAndExchangeB:
3306 case Op_CompareAndExchangeS:
3307 case Op_CompareAndExchangeN:
3308 case Op_CompareAndExchangeP:
3309 case Op_WeakCompareAndSwapL:
3310 case Op_WeakCompareAndSwapI:
3311 case Op_WeakCompareAndSwapB:
3312 case Op_WeakCompareAndSwapS:
3313 case Op_WeakCompareAndSwapN:
3314 case Op_WeakCompareAndSwapP:
3315 case Op_ShenandoahCompareAndSwapN:
3316 case Op_ShenandoahCompareAndSwapP:
3317 case Op_ShenandoahWeakCompareAndSwapN:
3318 case Op_ShenandoahWeakCompareAndSwapP:
3319 case Op_ShenandoahCompareAndExchangeN:
3320 case Op_ShenandoahCompareAndExchangeP:
3321 case Op_GetAndSetL:
3322 case Op_GetAndSetI:
3323 case Op_GetAndSetB:
3324 case Op_GetAndSetS:
3325 case Op_GetAndSetP:
3326 case Op_GetAndSetN:
3327 case Op_GetAndAddL:
3328 case Op_GetAndAddI:
3329 case Op_GetAndAddB:
3330 case Op_GetAndAddS:
3331 case Op_FastLock:
3332 case Op_FastUnlock:
3333 case Op_Rethrow:
3334 case Op_Return:
3335 case Op_StoreB:
3336 case Op_StoreC:
3337 case Op_StoreD:
3338 case Op_StoreF:
3339 case Op_StoreL:
3340 case Op_StoreLConditional:
3341 case Op_StoreI:
3342 case Op_StoreIConditional:
3343 case Op_StoreN:
3344 case Op_StoreP:
3345 case Op_StoreVector:
3346 case Op_StrInflatedCopy:
3347 case Op_StrCompressedCopy:
3348 case Op_EncodeP:
3349 case Op_CastP2X:
3350 case Op_SafePoint:
3351 case Op_EncodeISOArray:
3352 case Op_AryEq:
3353 case Op_StrEquals:
3354 case Op_StrComp:
3355 case Op_StrIndexOf:
3356 case Op_StrIndexOfChar:
3357 case Op_HasNegatives:
3358 // Known to require barriers
3359 return false;
3360 case Op_CmpP: {
3361 if (n->in(1)->bottom_type()->higher_equal(TypePtr::NULL_PTR) ||
3362 n->in(2)->bottom_type()->higher_equal(TypePtr::NULL_PTR)) {
3363 // One of the sides is known null, no need for barrier.
3364 } else {
3365 return false;
3366 }
3367 break;
3368 }
3369 case Op_LoadB:
3370 case Op_LoadUB:
3371 case Op_LoadUS:
3372 case Op_LoadD:
3373 case Op_LoadF:
3374 case Op_LoadL:
3375 case Op_LoadI:
3376 case Op_LoadS:
3377 case Op_LoadN:
3378 case Op_LoadP:
3379 case Op_LoadVector: {
3380 const TypePtr* adr_type = n->adr_type();
3381 int alias_idx = Compile::current()->get_alias_index(adr_type);
3382 Compile::AliasType* alias_type = Compile::current()->alias_type(alias_idx);
3383 ciField* field = alias_type->field();
3384 bool is_static = field != NULL && field->is_static();
3385 bool is_final = field != NULL && field->is_final();
3386
3387 if (ShenandoahOptimizeStaticFinals && is_static && is_final) {
3388 // Loading the constant does not require barriers: it should be handled
3389 // as part of GC roots already.
3390 } else {
3391 return false;
3392 }
3393 break;
3394 }
3395 case Op_Conv2B:
3396 case Op_LoadRange:
3397 case Op_LoadKlass:
3398 case Op_LoadNKlass:
3399 // Do not require barriers
3400 break;
3401 case Op_AddP:
3402 case Op_CheckCastPP:
3403 case Op_CastPP:
3404 case Op_CMoveP:
3405 case Op_Phi:
3406 case Op_ShenandoahLoadReferenceBarrier:
3407 case Op_ShenandoahEnqueueBarrier:
3408 // Whether or not these need the barriers depends on their users
3409 visit_users = true;
3410 break;
3411 default: {
3412 #ifdef ASSERT
3413 fatal("Unknown node in is_redundant: %s", NodeClassNames[n->Opcode()]);
3414 #else
3415 // Default to have excess barriers, rather than miss some.
3416 return false;
3417 #endif
3418 }
3419 }
3420
3421 stack.pop();
3422 if (visit_users) {
3423 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3424 Node* user = n->fast_out(i);
3425 if (user != NULL) {
3426 stack.push(user, 0);
3427 }
3428 }
3429 }
3430 }
3431
3432 // No need for barrier found.
3433 return true;
3434 }
3435
3436 CallStaticJavaNode* ShenandoahLoadReferenceBarrierNode::pin_and_expand_null_check(PhaseIterGVN& igvn) {
3437 Node* val = in(ValueIn);
3438
3439 const Type* val_t = igvn.type(val);
3440
3441 if (val_t->meet(TypePtr::NULL_PTR) != val_t &&
3442 val->Opcode() == Op_CastPP &&
3443 val->in(0) != NULL &&
3444 val->in(0)->Opcode() == Op_IfTrue &&
3445 val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
3446 val->in(0)->in(0)->is_If() &&
3447 val->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
3448 val->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
3449 val->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
3450 val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1) &&
3451 val->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
3452 assert(val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1), "");
3453 CallStaticJavaNode* unc = val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
3454 return unc;
3455 }
3456 return NULL;
3457 }