1 /*
2 * Copyright (c) 2015, 2021, Red Hat, Inc. All rights reserved.
3 * Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "precompiled.hpp"
27
28 #include "classfile/javaClasses.hpp"
29 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
30 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
31 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
32 #include "gc/shenandoah/shenandoahForwarding.hpp"
33 #include "gc/shenandoah/shenandoahHeap.hpp"
34 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
35 #include "gc/shenandoah/shenandoahRuntime.hpp"
36 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
37 #include "opto/arraycopynode.hpp"
38 #include "opto/block.hpp"
39 #include "opto/callnode.hpp"
40 #include "opto/castnode.hpp"
41 #include "opto/movenode.hpp"
42 #include "opto/phaseX.hpp"
43 #include "opto/rootnode.hpp"
44 #include "opto/runtime.hpp"
45 #include "opto/subnode.hpp"
46
47 bool ShenandoahBarrierC2Support::expand(Compile* C, PhaseIterGVN& igvn) {
48 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
49 if (state->load_reference_barriers_count() > 0) {
50 assert(C->post_loop_opts_phase(), "no loop opts allowed");
51 C->reset_post_loop_opts_phase(); // ... but we know what we are doing
52 C->clear_major_progress();
53 PhaseIdealLoop::optimize(igvn, LoopOptsShenandoahExpand);
54 if (C->failing()) return false;
55
56 C->set_major_progress();
57 if (!C->optimize_loops(igvn, LoopOptsShenandoahPostExpand)) {
58 return false;
59 }
60 C->clear_major_progress();
61 C->process_for_post_loop_opts_igvn(igvn);
62 if (C->failing()) return false;
63
64 C->set_post_loop_opts_phase(); // now for real!
65 }
66 return true;
67 }
68
69 bool ShenandoahBarrierC2Support::is_gc_state_test(Node* iff, int mask) {
70 if (!UseShenandoahGC) {
71 return false;
72 }
73 assert(iff->is_If(), "bad input");
74 if (iff->Opcode() != Op_If) {
75 return false;
76 }
77 Node* bol = iff->in(1);
78 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
79 return false;
80 }
81 Node* cmp = bol->in(1);
82 if (cmp->Opcode() != Op_CmpI) {
83 return false;
84 }
85 Node* in1 = cmp->in(1);
86 Node* in2 = cmp->in(2);
87 if (in2->find_int_con(-1) != 0) {
88 return false;
89 }
90 if (in1->Opcode() != Op_AndI) {
91 return false;
92 }
93 in2 = in1->in(2);
94 if (in2->find_int_con(-1) != mask) {
95 return false;
96 }
97 in1 = in1->in(1);
98
99 return is_gc_state_load(in1);
100 }
101
102 bool ShenandoahBarrierC2Support::is_heap_stable_test(Node* iff) {
103 return is_gc_state_test(iff, ShenandoahHeap::HAS_FORWARDED);
104 }
105
106 bool ShenandoahBarrierC2Support::is_gc_state_load(Node *n) {
107 if (!UseShenandoahGC) {
108 return false;
109 }
110 if (n->Opcode() != Op_LoadB && n->Opcode() != Op_LoadUB) {
111 return false;
112 }
113 Node* addp = n->in(MemNode::Address);
114 if (!addp->is_AddP()) {
115 return false;
116 }
117 Node* base = addp->in(AddPNode::Address);
118 Node* off = addp->in(AddPNode::Offset);
119 if (base->Opcode() != Op_ThreadLocal) {
120 return false;
121 }
122 if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) {
123 return false;
124 }
125 return true;
126 }
127
128 bool ShenandoahBarrierC2Support::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) {
129 assert(phase->is_dominator(stop, start), "bad inputs");
130 ResourceMark rm;
131 Unique_Node_List wq;
132 wq.push(start);
133 for (uint next = 0; next < wq.size(); next++) {
134 Node *m = wq.at(next);
135 if (m == stop) {
136 continue;
137 }
138 if (m->is_SafePoint() && !m->is_CallLeaf()) {
139 return true;
140 }
141 if (m->is_Region()) {
142 for (uint i = 1; i < m->req(); i++) {
143 wq.push(m->in(i));
144 }
145 } else {
146 wq.push(m->in(0));
147 }
148 }
149 return false;
150 }
151
152 #ifdef ASSERT
153 bool ShenandoahBarrierC2Support::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) {
154 assert(phis.size() == 0, "");
155
156 while (true) {
157 if (in->bottom_type() == TypePtr::NULL_PTR) {
158 if (trace) {tty->print_cr("null");}
159 } else if (!in->bottom_type()->make_ptr()->make_oopptr()) {
160 if (trace) {tty->print_cr("Non oop");}
161 } else {
162 if (in->is_ConstraintCast()) {
163 in = in->in(1);
164 continue;
165 } else if (in->is_AddP()) {
166 assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access");
167 in = in->in(AddPNode::Address);
168 continue;
169 } else if (in->is_Con()) {
170 if (trace) {
171 tty->print("Found constant");
172 in->dump();
173 }
174 } else if (in->Opcode() == Op_Parm) {
175 if (trace) {
176 tty->print("Found argument");
177 }
178 } else if (in->Opcode() == Op_CreateEx) {
179 if (trace) {
180 tty->print("Found create-exception");
181 }
182 } else if (in->Opcode() == Op_LoadP && in->adr_type() == TypeRawPtr::BOTTOM) {
183 if (trace) {
184 tty->print("Found raw LoadP (OSR argument?)");
185 }
186 } else if (in->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
187 if (t == ShenandoahOopStore) {
188 return false;
189 }
190 barriers_used.push(in);
191 if (trace) {tty->print("Found barrier"); in->dump();}
192 } else if (in->is_Proj() && in->in(0)->is_Allocate()) {
193 if (trace) {
194 tty->print("Found alloc");
195 in->in(0)->dump();
196 }
197 } else if (in->is_Proj() && (in->in(0)->Opcode() == Op_CallStaticJava || in->in(0)->Opcode() == Op_CallDynamicJava)) {
198 if (trace) {
199 tty->print("Found Java call");
200 }
201 } else if (in->is_Phi()) {
202 if (!visited.test_set(in->_idx)) {
203 if (trace) {tty->print("Pushed phi:"); in->dump();}
204 phis.push(in, 2);
205 in = in->in(1);
206 continue;
207 }
208 if (trace) {tty->print("Already seen phi:"); in->dump();}
209 } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) {
210 if (!visited.test_set(in->_idx)) {
211 if (trace) {tty->print("Pushed cmovep:"); in->dump();}
212 phis.push(in, CMoveNode::IfTrue);
213 in = in->in(CMoveNode::IfFalse);
214 continue;
215 }
216 if (trace) {tty->print("Already seen cmovep:"); in->dump();}
217 } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) {
218 in = in->in(1);
219 continue;
220 } else {
221 return false;
222 }
223 }
224 bool cont = false;
225 while (phis.is_nonempty()) {
226 uint idx = phis.index();
227 Node* phi = phis.node();
228 if (idx >= phi->req()) {
229 if (trace) {tty->print("Popped phi:"); phi->dump();}
230 phis.pop();
231 continue;
232 }
233 if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();}
234 in = phi->in(idx);
235 phis.set_index(idx+1);
236 cont = true;
237 break;
238 }
239 if (!cont) {
240 break;
241 }
242 }
243 return true;
244 }
245
246 void ShenandoahBarrierC2Support::report_verify_failure(const char* msg, Node* n1, Node* n2) {
247 if (n1 != nullptr) {
248 n1->dump(+10);
249 }
250 if (n2 != nullptr) {
251 n2->dump(+10);
252 }
253 fatal("%s", msg);
254 }
255
256 void ShenandoahBarrierC2Support::verify(RootNode* root) {
257 ResourceMark rm;
258 Unique_Node_List wq;
259 GrowableArray<Node*> barriers;
260 Unique_Node_List barriers_used;
261 Node_Stack phis(0);
262 VectorSet visited;
263 const bool trace = false;
264 const bool verify_no_useless_barrier = false;
265
266 wq.push(root);
267 for (uint next = 0; next < wq.size(); next++) {
268 Node *n = wq.at(next);
269 if (n->is_Load()) {
270 const bool trace = false;
271 if (trace) {tty->print("Verifying"); n->dump();}
272 if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) {
273 if (trace) {tty->print_cr("Load range/klass");}
274 } else {
275 const TypePtr* adr_type = n->as_Load()->adr_type();
276
277 if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) {
278 if (trace) {tty->print_cr("Mark load");}
279 } else if (adr_type->isa_instptr() &&
280 adr_type->is_instptr()->instance_klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
281 adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset()) {
282 if (trace) {tty->print_cr("Reference.get()");}
283 } else if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) {
284 report_verify_failure("Shenandoah verification: Load should have barriers", n);
285 }
286 }
287 } else if (n->is_Store()) {
288 const bool trace = false;
289
290 if (trace) {tty->print("Verifying"); n->dump();}
291 if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) {
292 Node* adr = n->in(MemNode::Address);
293 bool verify = true;
294
295 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) {
296 adr = adr->in(AddPNode::Address);
297 if (adr->is_AddP()) {
298 assert(adr->in(AddPNode::Base)->is_top(), "");
299 adr = adr->in(AddPNode::Address);
300 if (adr->Opcode() == Op_LoadP &&
301 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() &&
302 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
303 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) {
304 if (trace) {tty->print_cr("SATB prebarrier");}
305 verify = false;
306 }
307 }
308 }
309
310 if (verify && !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahValue, trace, barriers_used)) {
311 report_verify_failure("Shenandoah verification: Store should have barriers", n);
312 }
313 }
314 if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
315 report_verify_failure("Shenandoah verification: Store (address) should have barriers", n);
316 }
317 } else if (n->Opcode() == Op_CmpP) {
318 const bool trace = false;
319
320 Node* in1 = n->in(1);
321 Node* in2 = n->in(2);
322 if (in1->bottom_type()->isa_oopptr()) {
323 if (trace) {tty->print("Verifying"); n->dump();}
324
325 bool mark_inputs = false;
326 if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR ||
327 (in1->is_Con() || in2->is_Con())) {
328 if (trace) {tty->print_cr("Comparison against a constant");}
329 mark_inputs = true;
330 } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) ||
331 (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) {
332 if (trace) {tty->print_cr("Comparison with newly alloc'ed object");}
333 mark_inputs = true;
334 } else {
335 assert(in2->bottom_type()->isa_oopptr(), "");
336
337 if (!verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) ||
338 !verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) {
339 report_verify_failure("Shenandoah verification: Cmp should have barriers", n);
340 }
341 }
342 if (verify_no_useless_barrier &&
343 mark_inputs &&
344 (!verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) ||
345 !verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) {
346 phis.clear();
347 visited.reset();
348 }
349 }
350 } else if (n->is_LoadStore()) {
351 if (n->in(MemNode::ValueIn)->bottom_type()->make_ptr() &&
352 !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahValue, trace, barriers_used)) {
353 report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n);
354 }
355
356 if (n->in(MemNode::Address)->bottom_type()->make_oopptr() && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
357 report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n);
358 }
359 } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) {
360 CallNode* call = n->as_Call();
361
362 static struct {
363 const char* name;
364 struct {
365 int pos;
366 verify_type t;
367 } args[6];
368 } calls[] = {
369 "array_partition_stub",
370 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+4, ShenandoahStore }, { -1, ShenandoahNone },
371 { -1, ShenandoahNone }, { -1, ShenandoahNone }, { -1, ShenandoahNone } },
372 "arraysort_stub",
373 { { TypeFunc::Parms, ShenandoahStore }, { -1, ShenandoahNone }, { -1, ShenandoahNone },
374 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
375 "aescrypt_encryptBlock",
376 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
377 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
378 "aescrypt_decryptBlock",
379 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
380 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
381 "multiplyToLen",
382 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { TypeFunc::Parms+4, ShenandoahStore },
383 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
384 "squareToLen",
385 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { -1, ShenandoahNone},
386 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
387 "montgomery_multiply",
388 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
389 { TypeFunc::Parms+6, ShenandoahStore }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
390 "montgomery_square",
391 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+5, ShenandoahStore },
392 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
393 "mulAdd",
394 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
395 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
396 "vectorizedMismatch",
397 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
398 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
399 "updateBytesCRC32",
400 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
401 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
402 "updateBytesAdler32",
403 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
404 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
405 "updateBytesCRC32C",
406 { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad}, { -1, ShenandoahNone},
407 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
408 "counterMode_AESCrypt",
409 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
410 { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } },
411 "cipherBlockChaining_encryptAESCrypt",
412 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
413 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
414 "cipherBlockChaining_decryptAESCrypt",
415 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
416 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
417 "shenandoah_clone",
418 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
419 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
420 "ghash_processBlocks",
421 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
422 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
423 "sha1_implCompress",
424 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
425 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
426 "sha256_implCompress",
427 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
428 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
429 "sha512_implCompress",
430 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
431 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
432 "sha1_implCompressMB",
433 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
434 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
435 "sha256_implCompressMB",
436 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
437 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
438 "sha512_implCompressMB",
439 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
440 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
441 "encodeBlock",
442 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahStore }, { -1, ShenandoahNone },
443 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
444 "decodeBlock",
445 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahStore }, { -1, ShenandoahNone },
446 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
447 };
448
449 if (call->is_call_to_arraycopystub()) {
450 Node* dest = nullptr;
451 const TypeTuple* args = n->as_Call()->_tf->domain();
452 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
453 if (args->field_at(i)->isa_ptr()) {
454 j++;
455 if (j == 2) {
456 dest = n->in(i);
457 break;
458 }
459 }
460 }
461 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) ||
462 !verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) {
463 report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n);
464 }
465 } else if (strlen(call->_name) > 5 &&
466 !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) {
467 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) {
468 report_verify_failure("Shenandoah verification: _fill should have barriers", n);
469 }
470 } else if (!strcmp(call->_name, "shenandoah_wb_pre")) {
471 // skip
472 } else {
473 const int calls_len = sizeof(calls) / sizeof(calls[0]);
474 int i = 0;
475 for (; i < calls_len; i++) {
476 if (!strcmp(calls[i].name, call->_name)) {
477 break;
478 }
479 }
480 if (i != calls_len) {
481 const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]);
482 for (uint j = 0; j < args_len; j++) {
483 int pos = calls[i].args[j].pos;
484 if (pos == -1) {
485 break;
486 }
487 if (!verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) {
488 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
489 }
490 }
491 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
492 if (call->in(j)->bottom_type()->make_ptr() &&
493 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
494 uint k = 0;
495 for (; k < args_len && calls[i].args[k].pos != (int)j; k++);
496 if (k == args_len) {
497 fatal("arg %d for call %s not covered", j, call->_name);
498 }
499 }
500 }
501 } else {
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 fatal("%s not covered", call->_name);
506 }
507 }
508 }
509 }
510 } else if (n->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
511 // skip
512 } else if (n->is_AddP()
513 || n->is_Phi()
514 || n->is_ConstraintCast()
515 || n->Opcode() == Op_Return
516 || n->Opcode() == Op_CMoveP
517 || n->Opcode() == Op_CMoveN
518 || n->Opcode() == Op_Rethrow
519 || n->is_MemBar()
520 || n->Opcode() == Op_Conv2B
521 || n->Opcode() == Op_SafePoint
522 || n->is_CallJava()
523 || n->Opcode() == Op_Unlock
524 || n->Opcode() == Op_EncodeP
525 || n->Opcode() == Op_DecodeN) {
526 // nothing to do
527 } else {
528 static struct {
529 int opcode;
530 struct {
531 int pos;
532 verify_type t;
533 } inputs[2];
534 } others[] = {
535 Op_FastLock,
536 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
537 Op_Lock,
538 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone} },
539 Op_ArrayCopy,
540 { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } },
541 Op_StrCompressedCopy,
542 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
543 Op_StrInflatedCopy,
544 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
545 Op_AryEq,
546 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
547 Op_StrIndexOf,
548 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
549 Op_StrComp,
550 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
551 Op_StrEquals,
552 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
553 Op_VectorizedHashCode,
554 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } },
555 Op_EncodeISOArray,
556 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
557 Op_CountPositives,
558 { { 2, ShenandoahLoad }, { -1, ShenandoahNone} },
559 Op_CastP2X,
560 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
561 Op_StrIndexOfChar,
562 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } },
563 };
564
565 const int others_len = sizeof(others) / sizeof(others[0]);
566 int i = 0;
567 for (; i < others_len; i++) {
568 if (others[i].opcode == n->Opcode()) {
569 break;
570 }
571 }
572 uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req();
573 if (i != others_len) {
574 const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]);
575 for (uint j = 0; j < inputs_len; j++) {
576 int pos = others[i].inputs[j].pos;
577 if (pos == -1) {
578 break;
579 }
580 if (!verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) {
581 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
582 }
583 }
584 for (uint j = 1; j < stop; j++) {
585 if (n->in(j) != nullptr && n->in(j)->bottom_type()->make_ptr() &&
586 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
587 uint k = 0;
588 for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++);
589 if (k == inputs_len) {
590 fatal("arg %d for node %s not covered", j, n->Name());
591 }
592 }
593 }
594 } else {
595 for (uint j = 1; j < stop; j++) {
596 if (n->in(j) != nullptr && n->in(j)->bottom_type()->make_ptr() &&
597 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
598 fatal("%s not covered", n->Name());
599 }
600 }
601 }
602 }
603
604 if (n->is_SafePoint()) {
605 SafePointNode* sfpt = n->as_SafePoint();
606 if (verify_no_useless_barrier && sfpt->jvms() != nullptr) {
607 for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) {
608 if (!verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) {
609 phis.clear();
610 visited.reset();
611 }
612 }
613 }
614 }
615 }
616
617 if (verify_no_useless_barrier) {
618 for (int i = 0; i < barriers.length(); i++) {
619 Node* n = barriers.at(i);
620 if (!barriers_used.member(n)) {
621 tty->print("XXX useless barrier"); n->dump(-2);
622 ShouldNotReachHere();
623 }
624 }
625 }
626 }
627 #endif
628
629 bool ShenandoahBarrierC2Support::is_dominator_same_ctrl(Node* c, Node* d, Node* n, PhaseIdealLoop* phase) {
630 // That both nodes have the same control is not sufficient to prove
631 // domination, verify that there's no path from d to n
632 ResourceMark rm;
633 Unique_Node_List wq;
634 wq.push(d);
635 for (uint next = 0; next < wq.size(); next++) {
636 Node *m = wq.at(next);
637 if (m == n) {
638 return false;
639 }
640 if (m->is_Phi() && m->in(0)->is_Loop()) {
641 assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control");
642 } else {
643 if (m->is_Store() || m->is_LoadStore()) {
644 // Take anti-dependencies into account
645 Node* mem = m->in(MemNode::Memory);
646 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
647 Node* u = mem->fast_out(i);
648 if (u->is_Load() && phase->C->can_alias(m->adr_type(), phase->C->get_alias_index(u->adr_type())) &&
649 phase->ctrl_or_self(u) == c) {
650 wq.push(u);
651 }
652 }
653 }
654 for (uint i = 0; i < m->req(); i++) {
655 if (m->in(i) != nullptr && phase->ctrl_or_self(m->in(i)) == c) {
656 wq.push(m->in(i));
657 }
658 }
659 }
660 }
661 return true;
662 }
663
664 bool ShenandoahBarrierC2Support::is_dominator(Node* d_c, Node* n_c, Node* d, Node* n, PhaseIdealLoop* phase) {
665 if (d_c != n_c) {
666 return phase->is_dominator(d_c, n_c);
667 }
668 return is_dominator_same_ctrl(d_c, d, n, phase);
669 }
670
671 Node* next_mem(Node* mem, int alias) {
672 Node* res = nullptr;
673 if (mem->is_Proj()) {
674 res = mem->in(0);
675 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
676 res = mem->in(TypeFunc::Memory);
677 } else if (mem->is_Phi()) {
678 res = mem->in(1);
679 } else if (mem->is_MergeMem()) {
680 res = mem->as_MergeMem()->memory_at(alias);
681 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
682 assert(alias == Compile::AliasIdxRaw, "following raw memory can't lead to a barrier");
683 res = mem->in(MemNode::Memory);
684 } else {
685 #ifdef ASSERT
686 mem->dump();
687 #endif
688 ShouldNotReachHere();
689 }
690 return res;
691 }
692
693 Node* ShenandoahBarrierC2Support::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) {
694 Node* iffproj = nullptr;
695 while (c != dom) {
696 Node* next = phase->idom(c);
697 assert(next->unique_ctrl_out_or_null() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?");
698 if (c->is_Region()) {
699 ResourceMark rm;
700 Unique_Node_List wq;
701 wq.push(c);
702 for (uint i = 0; i < wq.size(); i++) {
703 Node *n = wq.at(i);
704 if (n == next) {
705 continue;
706 }
707 if (n->is_Region()) {
708 for (uint j = 1; j < n->req(); j++) {
709 wq.push(n->in(j));
710 }
711 } else {
712 wq.push(n->in(0));
713 }
714 }
715 for (uint i = 0; i < wq.size(); i++) {
716 Node *n = wq.at(i);
717 assert(n->is_CFG(), "");
718 if (n->is_Multi()) {
719 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
720 Node* u = n->fast_out(j);
721 if (u->is_CFG()) {
722 if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
723 return NodeSentinel;
724 }
725 }
726 }
727 }
728 }
729 } else if (c->is_Proj()) {
730 if (c->is_IfProj()) {
731 if (c->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) != nullptr) {
732 // continue;
733 } else {
734 if (!allow_one_proj) {
735 return NodeSentinel;
736 }
737 if (iffproj == nullptr) {
738 iffproj = c;
739 } else {
740 return NodeSentinel;
741 }
742 }
743 } else if (c->Opcode() == Op_JumpProj) {
744 return NodeSentinel; // unsupported
745 } else if (c->Opcode() == Op_CatchProj) {
746 return NodeSentinel; // unsupported
747 } else if (c->Opcode() == Op_CProj && next->is_NeverBranch()) {
748 return NodeSentinel; // unsupported
749 } else {
750 assert(next->unique_ctrl_out() == c, "unsupported branch pattern");
751 }
752 }
753 c = next;
754 }
755 return iffproj;
756 }
757
758 Node* ShenandoahBarrierC2Support::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) {
759 ResourceMark rm;
760 VectorSet wq;
761 wq.set(mem->_idx);
762 mem_ctrl = phase->ctrl_or_self(mem);
763 while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) {
764 mem = next_mem(mem, alias);
765 if (wq.test_set(mem->_idx)) {
766 return nullptr;
767 }
768 mem_ctrl = phase->ctrl_or_self(mem);
769 }
770 if (mem->is_MergeMem()) {
771 mem = mem->as_MergeMem()->memory_at(alias);
772 mem_ctrl = phase->ctrl_or_self(mem);
773 }
774 return mem;
775 }
776
777 Node* ShenandoahBarrierC2Support::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) {
778 Node* mem = nullptr;
779 Node* c = ctrl;
780 do {
781 if (c->is_Region()) {
782 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == nullptr; i++) {
783 Node* u = c->fast_out(i);
784 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
785 if (u->adr_type() == TypePtr::BOTTOM) {
786 mem = u;
787 }
788 }
789 }
790 } else {
791 if (c->is_Call() && c->as_Call()->adr_type() != nullptr) {
792 CallProjections projs;
793 c->as_Call()->extract_projections(&projs, true, false);
794 if (projs.fallthrough_memproj != nullptr) {
795 if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
796 if (projs.catchall_memproj == nullptr) {
797 mem = projs.fallthrough_memproj;
798 } else {
799 if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) {
800 mem = projs.fallthrough_memproj;
801 } else {
802 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier");
803 mem = projs.catchall_memproj;
804 }
805 }
806 }
807 } else {
808 Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
809 if (proj != nullptr &&
810 proj->adr_type() == TypePtr::BOTTOM) {
811 mem = proj;
812 }
813 }
814 } else {
815 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
816 Node* u = c->fast_out(i);
817 if (u->is_Proj() &&
818 u->bottom_type() == Type::MEMORY &&
819 u->adr_type() == TypePtr::BOTTOM) {
820 assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
821 assert(mem == nullptr, "only one proj");
822 mem = u;
823 }
824 }
825 assert(!c->is_Call() || c->as_Call()->adr_type() != nullptr || mem == nullptr, "no mem projection expected");
826 }
827 }
828 c = phase->idom(c);
829 } while (mem == nullptr);
830 return mem;
831 }
832
833 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
834 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
835 Node* u = n->fast_out(i);
836 if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
837 uses.push(u);
838 }
839 }
840 }
841
842 static void hide_strip_mined_loop(OuterStripMinedLoopNode* outer, CountedLoopNode* inner, PhaseIdealLoop* phase) {
843 OuterStripMinedLoopEndNode* le = inner->outer_loop_end();
844 Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl));
845 phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl));
846 Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt);
847 phase->register_control(new_le, phase->get_loop(le), le->in(0));
848 phase->lazy_replace(outer, new_outer);
849 phase->lazy_replace(le, new_le);
850 inner->clear_strip_mined();
851 }
852
853 void ShenandoahBarrierC2Support::test_gc_state(Node*& ctrl, Node* raw_mem, Node*& test_fail_ctrl,
854 PhaseIdealLoop* phase, int flags) {
855 PhaseIterGVN& igvn = phase->igvn();
856 Node* old_ctrl = ctrl;
857
858 Node* thread = new ThreadLocalNode();
859 Node* gc_state_offset = igvn.MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
860 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, gc_state_offset);
861 Node* gc_state = new LoadBNode(old_ctrl, raw_mem, gc_state_addr,
862 DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
863 TypeInt::BYTE, MemNode::unordered);
864 Node* gc_state_and = new AndINode(gc_state, igvn.intcon(flags));
865 Node* gc_state_cmp = new CmpINode(gc_state_and, igvn.zerocon(T_INT));
866 Node* gc_state_bool = new BoolNode(gc_state_cmp, BoolTest::ne);
867
868 IfNode* gc_state_iff = new IfNode(old_ctrl, gc_state_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
869 ctrl = new IfTrueNode(gc_state_iff);
870 test_fail_ctrl = new IfFalseNode(gc_state_iff);
871
872 IdealLoopTree* loop = phase->get_loop(old_ctrl);
873 phase->register_control(gc_state_iff, loop, old_ctrl);
874 phase->register_control(ctrl, loop, gc_state_iff);
875 phase->register_control(test_fail_ctrl, loop, gc_state_iff);
876
877 phase->register_new_node(thread, old_ctrl);
878 phase->register_new_node(gc_state_addr, old_ctrl);
879 phase->register_new_node(gc_state, old_ctrl);
880 phase->register_new_node(gc_state_and, old_ctrl);
881 phase->register_new_node(gc_state_cmp, old_ctrl);
882 phase->register_new_node(gc_state_bool, old_ctrl);
883
884 phase->set_ctrl(gc_state_offset, phase->C->root());
885
886 assert(is_gc_state_test(gc_state_iff, flags), "Should match the shape");
887 }
888
889 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) {
890 Node* old_ctrl = ctrl;
891 PhaseIterGVN& igvn = phase->igvn();
892
893 const Type* val_t = igvn.type(val);
894 if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
895 Node* null_cmp = new CmpPNode(val, igvn.zerocon(T_OBJECT));
896 Node* null_test = new BoolNode(null_cmp, BoolTest::ne);
897
898 IfNode* null_iff = new IfNode(old_ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
899 ctrl = new IfTrueNode(null_iff);
900 null_ctrl = new IfFalseNode(null_iff);
901
902 IdealLoopTree* loop = phase->get_loop(old_ctrl);
903 phase->register_control(null_iff, loop, old_ctrl);
904 phase->register_control(ctrl, loop, null_iff);
905 phase->register_control(null_ctrl, loop, null_iff);
906
907 phase->register_new_node(null_cmp, old_ctrl);
908 phase->register_new_node(null_test, old_ctrl);
909 }
910 }
911
912 void ShenandoahBarrierC2Support::test_in_cset(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) {
913 Node* old_ctrl = ctrl;
914 PhaseIterGVN& igvn = phase->igvn();
915
916 Node* raw_val = new CastP2XNode(old_ctrl, val);
917 Node* cset_idx = new URShiftXNode(raw_val, igvn.intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
918
919 // Figure out the target cset address with raw pointer math.
920 // This avoids matching AddP+LoadB that would emit inefficient code.
921 // See JDK-8245465.
922 Node* cset_addr_ptr = igvn.makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
923 Node* cset_addr = new CastP2XNode(old_ctrl, cset_addr_ptr);
924 Node* cset_load_addr = new AddXNode(cset_addr, cset_idx);
925 Node* cset_load_ptr = new CastX2PNode(cset_load_addr);
926
927 Node* cset_load = new LoadBNode(old_ctrl, raw_mem, cset_load_ptr,
928 DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
929 TypeInt::BYTE, MemNode::unordered);
930 Node* cset_cmp = new CmpINode(cset_load, igvn.zerocon(T_INT));
931 Node* cset_bool = new BoolNode(cset_cmp, BoolTest::ne);
932
933 IfNode* cset_iff = new IfNode(old_ctrl, cset_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
934 ctrl = new IfTrueNode(cset_iff);
935 not_cset_ctrl = new IfFalseNode(cset_iff);
936
937 IdealLoopTree *loop = phase->get_loop(old_ctrl);
938 phase->register_control(cset_iff, loop, old_ctrl);
939 phase->register_control(ctrl, loop, cset_iff);
940 phase->register_control(not_cset_ctrl, loop, cset_iff);
941
942 phase->set_ctrl(cset_addr_ptr, phase->C->root());
943
944 phase->register_new_node(raw_val, old_ctrl);
945 phase->register_new_node(cset_idx, old_ctrl);
946 phase->register_new_node(cset_addr, old_ctrl);
947 phase->register_new_node(cset_load_addr, old_ctrl);
948 phase->register_new_node(cset_load_ptr, old_ctrl);
949 phase->register_new_node(cset_load, old_ctrl);
950 phase->register_new_node(cset_cmp, old_ctrl);
951 phase->register_new_node(cset_bool, old_ctrl);
952 }
953
954 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node* load_addr,
955 DecoratorSet decorators, PhaseIdealLoop* phase) {
956 IdealLoopTree*loop = phase->get_loop(ctrl);
957 const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr();
958
959 address calladdr = nullptr;
960 const char* name = nullptr;
961 bool is_strong = ShenandoahBarrierSet::is_strong_access(decorators);
962 bool is_weak = ShenandoahBarrierSet::is_weak_access(decorators);
963 bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
964 bool is_native = ShenandoahBarrierSet::is_native_access(decorators);
965 bool is_narrow = UseCompressedOops && !is_native;
966 if (is_strong) {
967 if (is_narrow) {
968 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow);
969 name = "load_reference_barrier_strong_narrow";
970 } else {
971 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
972 name = "load_reference_barrier_strong";
973 }
974 } else if (is_weak) {
975 if (is_narrow) {
976 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow);
977 name = "load_reference_barrier_weak_narrow";
978 } else {
979 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak);
980 name = "load_reference_barrier_weak";
981 }
982 } else {
983 assert(is_phantom, "only remaining strength");
984 if (is_narrow) {
985 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom_narrow);
986 name = "load_reference_barrier_phantom_narrow";
987 } else {
988 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom);
989 name = "load_reference_barrier_phantom";
990 }
991 }
992 Node* call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type(), calladdr, name, TypeRawPtr::BOTTOM);
993
994 call->init_req(TypeFunc::Control, ctrl);
995 call->init_req(TypeFunc::I_O, phase->C->top());
996 call->init_req(TypeFunc::Memory, phase->C->top());
997 call->init_req(TypeFunc::FramePtr, phase->C->top());
998 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
999 call->init_req(TypeFunc::Parms, val);
1000 call->init_req(TypeFunc::Parms+1, load_addr);
1001 phase->register_control(call, loop, ctrl);
1002 ctrl = new ProjNode(call, TypeFunc::Control);
1003 phase->register_control(ctrl, loop, call);
1004 val = new ProjNode(call, TypeFunc::Parms);
1005 phase->register_new_node(val, call);
1006 val = new CheckCastPPNode(ctrl, val, obj_type);
1007 phase->register_new_node(val, ctrl);
1008 }
1009
1010 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) {
1011 Node* ctrl = phase->get_ctrl(barrier);
1012 Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
1013
1014 // Update the control of all nodes that should be after the
1015 // barrier control flow
1016 uses.clear();
1017 // Every node that is control dependent on the barrier's input
1018 // control will be after the expanded barrier. The raw memory (if
1019 // its memory is control dependent on the barrier's input control)
1020 // must stay above the barrier.
1021 uses_to_ignore.clear();
1022 if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
1023 uses_to_ignore.push(init_raw_mem);
1024 }
1025 for (uint next = 0; next < uses_to_ignore.size(); next++) {
1026 Node *n = uses_to_ignore.at(next);
1027 for (uint i = 0; i < n->req(); i++) {
1028 Node* in = n->in(i);
1029 if (in != nullptr && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
1030 uses_to_ignore.push(in);
1031 }
1032 }
1033 }
1034 for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
1035 Node* u = ctrl->fast_out(i);
1036 if (u->_idx < last &&
1037 u != barrier &&
1038 !u->depends_only_on_test() && // preserve dependency on test
1039 !uses_to_ignore.member(u) &&
1040 (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) &&
1041 (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) {
1042 Node* old_c = phase->ctrl_or_self(u);
1043 Node* c = old_c;
1044 if (c != ctrl ||
1045 is_dominator_same_ctrl(old_c, barrier, u, phase) ||
1046 ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) {
1047 phase->igvn().rehash_node_delayed(u);
1048 int nb = u->replace_edge(ctrl, region, &phase->igvn());
1049 if (u->is_CFG()) {
1050 if (phase->idom(u) == ctrl) {
1051 phase->set_idom(u, region, phase->dom_depth(region));
1052 }
1053 } else if (phase->get_ctrl(u) == ctrl) {
1054 assert(u != init_raw_mem, "should leave input raw mem above the barrier");
1055 uses.push(u);
1056 }
1057 assert(nb == 1, "more than 1 ctrl input?");
1058 --i, imax -= nb;
1059 }
1060 }
1061 }
1062 }
1063
1064 static Node* create_phis_on_call_return(Node* ctrl, Node* c, Node* n, Node* n_clone, const CallProjections& projs, PhaseIdealLoop* phase) {
1065 Node* region = nullptr;
1066 while (c != ctrl) {
1067 if (c->is_Region()) {
1068 region = c;
1069 }
1070 c = phase->idom(c);
1071 }
1072 assert(region != nullptr, "");
1073 Node* phi = new PhiNode(region, n->bottom_type());
1074 for (uint j = 1; j < region->req(); j++) {
1075 Node* in = region->in(j);
1076 if (phase->is_dominator(projs.fallthrough_catchproj, in)) {
1077 phi->init_req(j, n);
1078 } else if (phase->is_dominator(projs.catchall_catchproj, in)) {
1079 phi->init_req(j, n_clone);
1080 } else {
1081 phi->init_req(j, create_phis_on_call_return(ctrl, in, n, n_clone, projs, phase));
1082 }
1083 }
1084 phase->register_new_node(phi, region);
1085 return phi;
1086 }
1087
1088 void ShenandoahBarrierC2Support::pin_and_expand(PhaseIdealLoop* phase) {
1089 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
1090
1091 Unique_Node_List uses;
1092 Node_Stack stack(0);
1093 Node_List clones;
1094 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1095 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1096
1097 Node* ctrl = phase->get_ctrl(lrb);
1098 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1099
1100 CallStaticJavaNode* unc = nullptr;
1101 Node* unc_ctrl = nullptr;
1102 Node* uncasted_val = val;
1103
1104 for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1105 Node* u = lrb->fast_out(i);
1106 if (u->Opcode() == Op_CastPP &&
1107 u->in(0) != nullptr &&
1108 phase->is_dominator(u->in(0), ctrl)) {
1109 const Type* u_t = phase->igvn().type(u);
1110
1111 if (u_t->meet(TypePtr::NULL_PTR) != u_t &&
1112 u->in(0)->Opcode() == Op_IfTrue &&
1113 u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
1114 u->in(0)->in(0)->is_If() &&
1115 u->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
1116 u->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
1117 u->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
1118 u->in(0)->in(0)->in(1)->in(1)->in(1) == val &&
1119 u->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
1120 IdealLoopTree* loop = phase->get_loop(ctrl);
1121 IdealLoopTree* unc_loop = phase->get_loop(u->in(0));
1122
1123 if (!unc_loop->is_member(loop)) {
1124 continue;
1125 }
1126
1127 Node* branch = no_branches(ctrl, u->in(0), false, phase);
1128 assert(branch == nullptr || branch == NodeSentinel, "was not looking for a branch");
1129 if (branch == NodeSentinel) {
1130 continue;
1131 }
1132
1133 Node* iff = u->in(0)->in(0);
1134 Node* bol = iff->in(1)->clone();
1135 Node* cmp = bol->in(1)->clone();
1136 cmp->set_req(1, lrb);
1137 bol->set_req(1, cmp);
1138 phase->igvn().replace_input_of(iff, 1, bol);
1139 phase->set_ctrl(lrb, iff->in(0));
1140 phase->register_new_node(cmp, iff->in(0));
1141 phase->register_new_node(bol, iff->in(0));
1142 break;
1143 }
1144 }
1145 }
1146 if ((ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) || ctrl->is_CallJava()) {
1147 CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_CallJava() : ctrl->as_CallJava();
1148 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
1149 // The rethrow call may have too many projections to be
1150 // properly handled here. Given there's no reason for a
1151 // barrier to depend on the call, move it above the call
1152 stack.push(lrb, 0);
1153 do {
1154 Node* n = stack.node();
1155 uint idx = stack.index();
1156 if (idx < n->req()) {
1157 Node* in = n->in(idx);
1158 stack.set_index(idx+1);
1159 if (in != nullptr) {
1160 if (phase->has_ctrl(in)) {
1161 if (phase->is_dominator(call, phase->get_ctrl(in))) {
1162 #ifdef ASSERT
1163 for (uint i = 0; i < stack.size(); i++) {
1164 assert(stack.node_at(i) != in, "node shouldn't have been seen yet");
1165 }
1166 #endif
1167 stack.push(in, 0);
1168 }
1169 } else {
1170 assert(phase->is_dominator(in, call->in(0)), "no dependency on the call");
1171 }
1172 }
1173 } else {
1174 phase->set_ctrl(n, call->in(0));
1175 stack.pop();
1176 }
1177 } while(stack.size() > 0);
1178 continue;
1179 }
1180 CallProjections projs;
1181 call->extract_projections(&projs, false, false);
1182
1183 #ifdef ASSERT
1184 VectorSet cloned;
1185 #endif
1186 Node* lrb_clone = lrb->clone();
1187 phase->register_new_node(lrb_clone, projs.catchall_catchproj);
1188 phase->set_ctrl(lrb, projs.fallthrough_catchproj);
1189
1190 stack.push(lrb, 0);
1191 clones.push(lrb_clone);
1192
1193 do {
1194 assert(stack.size() == clones.size(), "");
1195 Node* n = stack.node();
1196 #ifdef ASSERT
1197 if (n->is_Load()) {
1198 Node* mem = n->in(MemNode::Memory);
1199 for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
1200 Node* u = mem->fast_out(j);
1201 assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?");
1202 }
1203 }
1204 #endif
1205 uint idx = stack.index();
1206 Node* n_clone = clones.at(clones.size()-1);
1207 if (idx < n->outcnt()) {
1208 Node* u = n->raw_out(idx);
1209 Node* c = phase->ctrl_or_self(u);
1210 if (phase->is_dominator(call, c) && phase->is_dominator(c, projs.fallthrough_proj)) {
1211 stack.set_index(idx+1);
1212 assert(!u->is_CFG(), "");
1213 stack.push(u, 0);
1214 assert(!cloned.test_set(u->_idx), "only one clone");
1215 Node* u_clone = u->clone();
1216 int nb = u_clone->replace_edge(n, n_clone, &phase->igvn());
1217 assert(nb > 0, "should have replaced some uses");
1218 phase->register_new_node(u_clone, projs.catchall_catchproj);
1219 clones.push(u_clone);
1220 phase->set_ctrl(u, projs.fallthrough_catchproj);
1221 } else {
1222 bool replaced = false;
1223 if (u->is_Phi()) {
1224 for (uint k = 1; k < u->req(); k++) {
1225 if (u->in(k) == n) {
1226 if (phase->is_dominator(projs.catchall_catchproj, u->in(0)->in(k))) {
1227 phase->igvn().replace_input_of(u, k, n_clone);
1228 replaced = true;
1229 } else if (!phase->is_dominator(projs.fallthrough_catchproj, u->in(0)->in(k))) {
1230 phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase));
1231 replaced = true;
1232 }
1233 }
1234 }
1235 } else {
1236 if (phase->is_dominator(projs.catchall_catchproj, c)) {
1237 phase->igvn().rehash_node_delayed(u);
1238 int nb = u->replace_edge(n, n_clone, &phase->igvn());
1239 assert(nb > 0, "should have replaced some uses");
1240 replaced = true;
1241 } else if (!phase->is_dominator(projs.fallthrough_catchproj, c)) {
1242 if (u->is_If()) {
1243 // Can't break If/Bool/Cmp chain
1244 assert(n->is_Bool(), "unexpected If shape");
1245 assert(stack.node_at(stack.size()-2)->is_Cmp(), "unexpected If shape");
1246 assert(n_clone->is_Bool(), "unexpected clone");
1247 assert(clones.at(clones.size()-2)->is_Cmp(), "unexpected clone");
1248 Node* bol_clone = n->clone();
1249 Node* cmp_clone = stack.node_at(stack.size()-2)->clone();
1250 bol_clone->set_req(1, cmp_clone);
1251
1252 Node* nn = stack.node_at(stack.size()-3);
1253 Node* nn_clone = clones.at(clones.size()-3);
1254 assert(nn->Opcode() == nn_clone->Opcode(), "mismatch");
1255
1256 int nb = cmp_clone->replace_edge(nn, create_phis_on_call_return(ctrl, c, nn, nn_clone, projs, phase),
1257 &phase->igvn());
1258 assert(nb > 0, "should have replaced some uses");
1259
1260 phase->register_new_node(bol_clone, u->in(0));
1261 phase->register_new_node(cmp_clone, u->in(0));
1262
1263 phase->igvn().replace_input_of(u, 1, bol_clone);
1264
1265 } else {
1266 phase->igvn().rehash_node_delayed(u);
1267 int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase), &phase->igvn());
1268 assert(nb > 0, "should have replaced some uses");
1269 }
1270 replaced = true;
1271 }
1272 }
1273 if (!replaced) {
1274 stack.set_index(idx+1);
1275 }
1276 }
1277 } else {
1278 stack.pop();
1279 clones.pop();
1280 }
1281 } while (stack.size() > 0);
1282 assert(stack.size() == 0 && clones.size() == 0, "");
1283 }
1284 }
1285
1286 for (int i = 0; i < state->load_reference_barriers_count(); i++) {
1287 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1288 Node* ctrl = phase->get_ctrl(lrb);
1289 IdealLoopTree* loop = phase->get_loop(ctrl);
1290 Node* head = loop->head();
1291 if (head->is_OuterStripMinedLoop()) {
1292 // Expanding a barrier here will break loop strip mining
1293 // verification. Transform the loop so the loop nest doesn't
1294 // appear as strip mined.
1295 OuterStripMinedLoopNode* outer = head->as_OuterStripMinedLoop();
1296 hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1297 }
1298 }
1299
1300 // Expand load-reference-barriers
1301 MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
1302 Unique_Node_List uses_to_ignore;
1303 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1304 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1305 uint last = phase->C->unique();
1306 Node* ctrl = phase->get_ctrl(lrb);
1307 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1308
1309 Node* orig_ctrl = ctrl;
1310
1311 Node* raw_mem = fixer.find_mem(ctrl, lrb);
1312 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, nullptr);
1313
1314 IdealLoopTree *loop = phase->get_loop(ctrl);
1315
1316 Node* heap_stable_ctrl = nullptr;
1317 Node* null_ctrl = nullptr;
1318
1319 assert(val->bottom_type()->make_oopptr(), "need oop");
1320 assert(val->bottom_type()->make_oopptr()->const_oop() == nullptr, "expect non-constant");
1321
1322 enum { _heap_stable = 1, _evac_path, _not_cset, PATH_LIMIT };
1323 Node* region = new RegionNode(PATH_LIMIT);
1324 Node* val_phi = new PhiNode(region, val->bottom_type()->is_oopptr());
1325
1326 // Stable path.
1327 int flags = ShenandoahHeap::HAS_FORWARDED;
1328 if (!ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1329 flags |= ShenandoahHeap::WEAK_ROOTS;
1330 }
1331 test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, flags);
1332 IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
1333
1334 // Heap stable case
1335 region->init_req(_heap_stable, heap_stable_ctrl);
1336 val_phi->init_req(_heap_stable, val);
1337
1338 // Test for in-cset, unless it's a native-LRB. Native LRBs need to return null
1339 // even for non-cset objects to prevent resurrection of such objects.
1340 // Wires !in_cset(obj) to slot 2 of region and phis
1341 Node* not_cset_ctrl = nullptr;
1342 if (ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1343 test_in_cset(ctrl, not_cset_ctrl, val, raw_mem, phase);
1344 }
1345 if (not_cset_ctrl != nullptr) {
1346 region->init_req(_not_cset, not_cset_ctrl);
1347 val_phi->init_req(_not_cset, val);
1348 } else {
1349 region->del_req(_not_cset);
1350 val_phi->del_req(_not_cset);
1351 }
1352
1353 // Resolve object when orig-value is in cset.
1354 // Make the unconditional resolve for fwdptr.
1355
1356 // Call lrb-stub and wire up that path in slots 4
1357 Node* result_mem = nullptr;
1358
1359 Node* addr;
1360 {
1361 VectorSet visited;
1362 addr = get_load_addr(phase, visited, lrb);
1363 }
1364 if (addr->Opcode() == Op_AddP) {
1365 Node* orig_base = addr->in(AddPNode::Base);
1366 Node* base = new CheckCastPPNode(ctrl, orig_base, orig_base->bottom_type(), ConstraintCastNode::StrongDependency);
1367 phase->register_new_node(base, ctrl);
1368 if (addr->in(AddPNode::Base) == addr->in((AddPNode::Address))) {
1369 // Field access
1370 addr = addr->clone();
1371 addr->set_req(AddPNode::Base, base);
1372 addr->set_req(AddPNode::Address, base);
1373 phase->register_new_node(addr, ctrl);
1374 } else {
1375 Node* addr2 = addr->in(AddPNode::Address);
1376 if (addr2->Opcode() == Op_AddP && addr2->in(AddPNode::Base) == addr2->in(AddPNode::Address) &&
1377 addr2->in(AddPNode::Base) == orig_base) {
1378 addr2 = addr2->clone();
1379 addr2->set_req(AddPNode::Base, base);
1380 addr2->set_req(AddPNode::Address, base);
1381 phase->register_new_node(addr2, ctrl);
1382 addr = addr->clone();
1383 addr->set_req(AddPNode::Base, base);
1384 addr->set_req(AddPNode::Address, addr2);
1385 phase->register_new_node(addr, ctrl);
1386 }
1387 }
1388 }
1389 call_lrb_stub(ctrl, val, addr, lrb->decorators(), phase);
1390 region->init_req(_evac_path, ctrl);
1391 val_phi->init_req(_evac_path, val);
1392
1393 phase->register_control(region, loop, heap_stable_iff);
1394 Node* out_val = val_phi;
1395 phase->register_new_node(val_phi, region);
1396
1397 fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase);
1398
1399 ctrl = orig_ctrl;
1400
1401 phase->igvn().replace_node(lrb, out_val);
1402
1403 follow_barrier_uses(out_val, ctrl, uses, phase);
1404
1405 for(uint next = 0; next < uses.size(); next++ ) {
1406 Node *n = uses.at(next);
1407 assert(phase->get_ctrl(n) == ctrl, "bad control");
1408 assert(n != raw_mem, "should leave input raw mem above the barrier");
1409 phase->set_ctrl(n, region);
1410 follow_barrier_uses(n, ctrl, uses, phase);
1411 }
1412 fixer.record_new_ctrl(ctrl, region, raw_mem, raw_mem_for_ctrl);
1413 }
1414 // Done expanding load-reference-barriers.
1415 assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced");
1416 }
1417
1418 Node* ShenandoahBarrierC2Support::get_load_addr(PhaseIdealLoop* phase, VectorSet& visited, Node* in) {
1419 if (visited.test_set(in->_idx)) {
1420 return nullptr;
1421 }
1422 switch (in->Opcode()) {
1423 case Op_Proj:
1424 return get_load_addr(phase, visited, in->in(0));
1425 case Op_CastPP:
1426 case Op_CheckCastPP:
1427 case Op_DecodeN:
1428 case Op_EncodeP:
1429 return get_load_addr(phase, visited, in->in(1));
1430 case Op_LoadN:
1431 case Op_LoadP:
1432 return in->in(MemNode::Address);
1433 case Op_CompareAndExchangeN:
1434 case Op_CompareAndExchangeP:
1435 case Op_GetAndSetN:
1436 case Op_GetAndSetP:
1437 case Op_ShenandoahCompareAndExchangeP:
1438 case Op_ShenandoahCompareAndExchangeN:
1439 // Those instructions would just have stored a different
1440 // value into the field. No use to attempt to fix it at this point.
1441 return phase->igvn().zerocon(T_OBJECT);
1442 case Op_CMoveP:
1443 case Op_CMoveN: {
1444 Node* t = get_load_addr(phase, visited, in->in(CMoveNode::IfTrue));
1445 Node* f = get_load_addr(phase, visited, in->in(CMoveNode::IfFalse));
1446 // Handle unambiguous cases: single address reported on both branches.
1447 if (t != nullptr && f == nullptr) return t;
1448 if (t == nullptr && f != nullptr) return f;
1449 if (t != nullptr && t == f) return t;
1450 // Ambiguity.
1451 return phase->igvn().zerocon(T_OBJECT);
1452 }
1453 case Op_Phi: {
1454 Node* addr = nullptr;
1455 for (uint i = 1; i < in->req(); i++) {
1456 Node* addr1 = get_load_addr(phase, visited, in->in(i));
1457 if (addr == nullptr) {
1458 addr = addr1;
1459 }
1460 if (addr != addr1) {
1461 return phase->igvn().zerocon(T_OBJECT);
1462 }
1463 }
1464 return addr;
1465 }
1466 case Op_ShenandoahLoadReferenceBarrier:
1467 return get_load_addr(phase, visited, in->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1468 case Op_CallDynamicJava:
1469 case Op_CallLeaf:
1470 case Op_CallStaticJava:
1471 case Op_ConN:
1472 case Op_ConP:
1473 case Op_Parm:
1474 case Op_CreateEx:
1475 return phase->igvn().zerocon(T_OBJECT);
1476 default:
1477 #ifdef ASSERT
1478 fatal("Unknown node in get_load_addr: %s", NodeClassNames[in->Opcode()]);
1479 #endif
1480 return phase->igvn().zerocon(T_OBJECT);
1481 }
1482
1483 }
1484
1485 void ShenandoahBarrierC2Support::move_gc_state_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
1486 IdealLoopTree *loop = phase->get_loop(iff);
1487 Node* loop_head = loop->_head;
1488 Node* entry_c = loop_head->in(LoopNode::EntryControl);
1489
1490 Node* bol = iff->in(1);
1491 Node* cmp = bol->in(1);
1492 Node* andi = cmp->in(1);
1493 Node* load = andi->in(1);
1494
1495 assert(is_gc_state_load(load), "broken");
1496 if (!phase->is_dominator(load->in(0), entry_c)) {
1497 Node* mem_ctrl = nullptr;
1498 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
1499 load = load->clone();
1500 load->set_req(MemNode::Memory, mem);
1501 load->set_req(0, entry_c);
1502 phase->register_new_node(load, entry_c);
1503 andi = andi->clone();
1504 andi->set_req(1, load);
1505 phase->register_new_node(andi, entry_c);
1506 cmp = cmp->clone();
1507 cmp->set_req(1, andi);
1508 phase->register_new_node(cmp, entry_c);
1509 bol = bol->clone();
1510 bol->set_req(1, cmp);
1511 phase->register_new_node(bol, entry_c);
1512
1513 phase->igvn().replace_input_of(iff, 1, bol);
1514 }
1515 }
1516
1517 bool ShenandoahBarrierC2Support::identical_backtoback_ifs(Node* n, PhaseIdealLoop* phase) {
1518 if (!n->is_If() || n->is_CountedLoopEnd()) {
1519 return false;
1520 }
1521 Node* region = n->in(0);
1522
1523 if (!region->is_Region()) {
1524 return false;
1525 }
1526 Node* dom = phase->idom(region);
1527 if (!dom->is_If()) {
1528 return false;
1529 }
1530
1531 if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) {
1532 return false;
1533 }
1534
1535 IfNode* dom_if = dom->as_If();
1536 Node* proj_true = dom_if->proj_out(1);
1537 Node* proj_false = dom_if->proj_out(0);
1538
1539 for (uint i = 1; i < region->req(); i++) {
1540 if (phase->is_dominator(proj_true, region->in(i))) {
1541 continue;
1542 }
1543 if (phase->is_dominator(proj_false, region->in(i))) {
1544 continue;
1545 }
1546 return false;
1547 }
1548
1549 return true;
1550 }
1551
1552 bool ShenandoahBarrierC2Support::merge_point_safe(Node* region) {
1553 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1554 Node* n = region->fast_out(i);
1555 if (n->is_LoadStore()) {
1556 // Splitting a LoadStore node through phi, causes it to lose its SCMemProj: the split if code doesn't have support
1557 // for a LoadStore at the region the if is split through because that's not expected to happen (LoadStore nodes
1558 // should be between barrier nodes). It does however happen with Shenandoah though because barriers can get
1559 // expanded around a LoadStore node.
1560 return false;
1561 }
1562 }
1563 return true;
1564 }
1565
1566
1567 void ShenandoahBarrierC2Support::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) {
1568 assert(is_heap_stable_test(n), "no other tests");
1569 if (identical_backtoback_ifs(n, phase)) {
1570 Node* n_ctrl = n->in(0);
1571 if (phase->can_split_if(n_ctrl) && merge_point_safe(n_ctrl)) {
1572 IfNode* dom_if = phase->idom(n_ctrl)->as_If();
1573 if (is_heap_stable_test(n)) {
1574 Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1);
1575 assert(is_gc_state_load(gc_state_load), "broken");
1576 Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1);
1577 assert(is_gc_state_load(dom_gc_state_load), "broken");
1578 if (gc_state_load != dom_gc_state_load) {
1579 phase->igvn().replace_node(gc_state_load, dom_gc_state_load);
1580 }
1581 }
1582 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1583 Node* proj_true = dom_if->proj_out(1);
1584 Node* proj_false = dom_if->proj_out(0);
1585 Node* con_true = phase->igvn().makecon(TypeInt::ONE);
1586 Node* con_false = phase->igvn().makecon(TypeInt::ZERO);
1587
1588 for (uint i = 1; i < n_ctrl->req(); i++) {
1589 if (phase->is_dominator(proj_true, n_ctrl->in(i))) {
1590 bolphi->init_req(i, con_true);
1591 } else {
1592 assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1593 bolphi->init_req(i, con_false);
1594 }
1595 }
1596 phase->register_new_node(bolphi, n_ctrl);
1597 phase->igvn().replace_input_of(n, 1, bolphi);
1598 phase->do_split_if(n);
1599 }
1600 }
1601 }
1602
1603 IfNode* ShenandoahBarrierC2Support::find_unswitching_candidate(const IdealLoopTree* loop, PhaseIdealLoop* phase) {
1604 // Find first invariant test that doesn't exit the loop
1605 LoopNode *head = loop->_head->as_Loop();
1606 IfNode* unswitch_iff = nullptr;
1607 Node* n = head->in(LoopNode::LoopBackControl);
1608 int loop_has_sfpts = -1;
1609 while (n != head) {
1610 Node* n_dom = phase->idom(n);
1611 if (n->is_Region()) {
1612 if (n_dom->is_If()) {
1613 IfNode* iff = n_dom->as_If();
1614 if (iff->in(1)->is_Bool()) {
1615 BoolNode* bol = iff->in(1)->as_Bool();
1616 if (bol->in(1)->is_Cmp()) {
1617 // If condition is invariant and not a loop exit,
1618 // then found reason to unswitch.
1619 if (is_heap_stable_test(iff) &&
1620 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) {
1621 assert(!loop->is_loop_exit(iff), "both branches should be in the loop");
1622 if (loop_has_sfpts == -1) {
1623 for(uint i = 0; i < loop->_body.size(); i++) {
1624 Node *m = loop->_body[i];
1625 if (m->is_SafePoint() && !m->is_CallLeaf()) {
1626 loop_has_sfpts = 1;
1627 break;
1628 }
1629 }
1630 if (loop_has_sfpts == -1) {
1631 loop_has_sfpts = 0;
1632 }
1633 }
1634 if (!loop_has_sfpts) {
1635 unswitch_iff = iff;
1636 }
1637 }
1638 }
1639 }
1640 }
1641 }
1642 n = n_dom;
1643 }
1644 return unswitch_iff;
1645 }
1646
1647
1648 void ShenandoahBarrierC2Support::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) {
1649 Node_List heap_stable_tests;
1650 stack.push(phase->C->start(), 0);
1651 do {
1652 Node* n = stack.node();
1653 uint i = stack.index();
1654
1655 if (i < n->outcnt()) {
1656 Node* u = n->raw_out(i);
1657 stack.set_index(i+1);
1658 if (!visited.test_set(u->_idx)) {
1659 stack.push(u, 0);
1660 }
1661 } else {
1662 stack.pop();
1663 if (n->is_If() && is_heap_stable_test(n)) {
1664 heap_stable_tests.push(n);
1665 }
1666 }
1667 } while (stack.size() > 0);
1668
1669 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1670 Node* n = heap_stable_tests.at(i);
1671 assert(is_heap_stable_test(n), "only evacuation test");
1672 merge_back_to_back_tests(n, phase);
1673 }
1674
1675 if (!phase->C->major_progress()) {
1676 VectorSet seen;
1677 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1678 Node* n = heap_stable_tests.at(i);
1679 IdealLoopTree* loop = phase->get_loop(n);
1680 if (loop != phase->ltree_root() &&
1681 loop->_child == nullptr &&
1682 !loop->_irreducible) {
1683 Node* head = loop->_head;
1684 if (head->is_Loop() &&
1685 (!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) &&
1686 !seen.test_set(head->_idx)) {
1687 IfNode* iff = find_unswitching_candidate(loop, phase);
1688 if (iff != nullptr) {
1689 Node* bol = iff->in(1);
1690 if (head->as_Loop()->is_strip_mined()) {
1691 head->as_Loop()->verify_strip_mined(0);
1692 }
1693 move_gc_state_test_out_of_loop(iff, phase);
1694
1695 AutoNodeBudget node_budget(phase);
1696
1697 if (loop->policy_unswitching(phase)) {
1698 if (head->as_Loop()->is_strip_mined()) {
1699 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop();
1700 hide_strip_mined_loop(outer, head->as_CountedLoop(), phase);
1701 }
1702 phase->do_unswitching(loop, old_new);
1703 } else {
1704 // Not proceeding with unswitching. Move load back in
1705 // the loop.
1706 phase->igvn().replace_input_of(iff, 1, bol);
1707 }
1708 }
1709 }
1710 }
1711 }
1712 }
1713 }
1714
1715 #ifdef ASSERT
1716 static bool has_never_branch(Node* root) {
1717 for (uint i = 1; i < root->req(); i++) {
1718 Node* in = root->in(i);
1719 if (in != nullptr && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->is_NeverBranch()) {
1720 return true;
1721 }
1722 }
1723 return false;
1724 }
1725 #endif
1726
1727 void MemoryGraphFixer::collect_memory_nodes() {
1728 Node_Stack stack(0);
1729 VectorSet visited;
1730 Node_List regions;
1731
1732 // Walk the raw memory graph and create a mapping from CFG node to
1733 // memory node. Exclude phis for now.
1734 stack.push(_phase->C->root(), 1);
1735 do {
1736 Node* n = stack.node();
1737 int opc = n->Opcode();
1738 uint i = stack.index();
1739 if (i < n->req()) {
1740 Node* mem = nullptr;
1741 if (opc == Op_Root) {
1742 Node* in = n->in(i);
1743 int in_opc = in->Opcode();
1744 if (in_opc == Op_Return || in_opc == Op_Rethrow) {
1745 mem = in->in(TypeFunc::Memory);
1746 } else if (in_opc == Op_Halt) {
1747 if (in->in(0)->is_Region()) {
1748 Node* r = in->in(0);
1749 for (uint j = 1; j < r->req(); j++) {
1750 assert(!r->in(j)->is_NeverBranch(), "");
1751 }
1752 } else {
1753 Node* proj = in->in(0);
1754 assert(proj->is_Proj(), "");
1755 Node* in = proj->in(0);
1756 assert(in->is_CallStaticJava() || in->is_NeverBranch() || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
1757 if (in->is_CallStaticJava()) {
1758 mem = in->in(TypeFunc::Memory);
1759 } else if (in->Opcode() == Op_Catch) {
1760 Node* call = in->in(0)->in(0);
1761 assert(call->is_Call(), "");
1762 mem = call->in(TypeFunc::Memory);
1763 } else if (in->is_NeverBranch()) {
1764 mem = collect_memory_for_infinite_loop(in);
1765 }
1766 }
1767 } else {
1768 #ifdef ASSERT
1769 n->dump();
1770 in->dump();
1771 #endif
1772 ShouldNotReachHere();
1773 }
1774 } else {
1775 assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
1776 assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
1777 mem = n->in(i);
1778 }
1779 i++;
1780 stack.set_index(i);
1781 if (mem == nullptr) {
1782 continue;
1783 }
1784 for (;;) {
1785 if (visited.test_set(mem->_idx) || mem->is_Start()) {
1786 break;
1787 }
1788 if (mem->is_Phi()) {
1789 stack.push(mem, 2);
1790 mem = mem->in(1);
1791 } else if (mem->is_Proj()) {
1792 stack.push(mem, mem->req());
1793 mem = mem->in(0);
1794 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
1795 mem = mem->in(TypeFunc::Memory);
1796 } else if (mem->is_MergeMem()) {
1797 MergeMemNode* mm = mem->as_MergeMem();
1798 mem = mm->memory_at(_alias);
1799 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
1800 assert(_alias == Compile::AliasIdxRaw, "");
1801 stack.push(mem, mem->req());
1802 mem = mem->in(MemNode::Memory);
1803 } else {
1804 #ifdef ASSERT
1805 mem->dump();
1806 #endif
1807 ShouldNotReachHere();
1808 }
1809 }
1810 } else {
1811 if (n->is_Phi()) {
1812 // Nothing
1813 } else if (!n->is_Root()) {
1814 Node* c = get_ctrl(n);
1815 _memory_nodes.map(c->_idx, n);
1816 }
1817 stack.pop();
1818 }
1819 } while(stack.is_nonempty());
1820
1821 // Iterate over CFG nodes in rpo and propagate memory state to
1822 // compute memory state at regions, creating new phis if needed.
1823 Node_List rpo_list;
1824 visited.clear();
1825 _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
1826 Node* root = rpo_list.pop();
1827 assert(root == _phase->C->root(), "");
1828
1829 const bool trace = false;
1830 #ifdef ASSERT
1831 if (trace) {
1832 for (int i = rpo_list.size() - 1; i >= 0; i--) {
1833 Node* c = rpo_list.at(i);
1834 if (_memory_nodes[c->_idx] != nullptr) {
1835 tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump();
1836 }
1837 }
1838 }
1839 #endif
1840 uint last = _phase->C->unique();
1841
1842 #ifdef ASSERT
1843 uint16_t max_depth = 0;
1844 for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
1845 IdealLoopTree* lpt = iter.current();
1846 max_depth = MAX2(max_depth, lpt->_nest);
1847 }
1848 #endif
1849
1850 bool progress = true;
1851 int iteration = 0;
1852 Node_List dead_phis;
1853 while (progress) {
1854 progress = false;
1855 iteration++;
1856 assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
1857 if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
1858
1859 for (int i = rpo_list.size() - 1; i >= 0; i--) {
1860 Node* c = rpo_list.at(i);
1861
1862 Node* prev_mem = _memory_nodes[c->_idx];
1863 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
1864 Node* prev_region = regions[c->_idx];
1865 Node* unique = nullptr;
1866 for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
1867 Node* m = _memory_nodes[c->in(j)->_idx];
1868 assert(m != nullptr || (c->is_Loop() && j == LoopNode::LoopBackControl && iteration == 1) || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "expect memory state");
1869 if (m != nullptr) {
1870 if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
1871 assert((c->is_Loop() && j == LoopNode::LoopBackControl) || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
1872 // continue
1873 } else if (unique == nullptr) {
1874 unique = m;
1875 } else if (m == unique) {
1876 // continue
1877 } else {
1878 unique = NodeSentinel;
1879 }
1880 }
1881 }
1882 assert(unique != nullptr, "empty phi???");
1883 if (unique != NodeSentinel) {
1884 if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c) {
1885 dead_phis.push(prev_region);
1886 }
1887 regions.map(c->_idx, unique);
1888 } else {
1889 Node* phi = nullptr;
1890 if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
1891 phi = prev_region;
1892 for (uint k = 1; k < c->req(); k++) {
1893 Node* m = _memory_nodes[c->in(k)->_idx];
1894 assert(m != nullptr, "expect memory state");
1895 phi->set_req(k, m);
1896 }
1897 } else {
1898 for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == nullptr; j++) {
1899 Node* u = c->fast_out(j);
1900 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
1901 (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
1902 phi = u;
1903 for (uint k = 1; k < c->req() && phi != nullptr; k++) {
1904 Node* m = _memory_nodes[c->in(k)->_idx];
1905 assert(m != nullptr, "expect memory state");
1906 if (u->in(k) != m) {
1907 phi = NodeSentinel;
1908 }
1909 }
1910 }
1911 }
1912 if (phi == NodeSentinel) {
1913 phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
1914 for (uint k = 1; k < c->req(); k++) {
1915 Node* m = _memory_nodes[c->in(k)->_idx];
1916 assert(m != nullptr, "expect memory state");
1917 phi->init_req(k, m);
1918 }
1919 }
1920 }
1921 if (phi != nullptr) {
1922 regions.map(c->_idx, phi);
1923 } else {
1924 assert(c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
1925 }
1926 }
1927 Node* current_region = regions[c->_idx];
1928 if (current_region != prev_region) {
1929 progress = true;
1930 if (prev_region == prev_mem) {
1931 _memory_nodes.map(c->_idx, current_region);
1932 }
1933 }
1934 } else if (prev_mem == nullptr || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
1935 Node* m = _memory_nodes[_phase->idom(c)->_idx];
1936 assert(m != nullptr || c->Opcode() == Op_Halt, "expect memory state");
1937 if (m != prev_mem) {
1938 _memory_nodes.map(c->_idx, m);
1939 progress = true;
1940 }
1941 }
1942 #ifdef ASSERT
1943 if (trace) { tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); }
1944 #endif
1945 }
1946 }
1947
1948 // Replace existing phi with computed memory state for that region
1949 // if different (could be a new phi or a dominating memory node if
1950 // that phi was found to be useless).
1951 while (dead_phis.size() > 0) {
1952 Node* n = dead_phis.pop();
1953 n->replace_by(_phase->C->top());
1954 n->destruct(&_phase->igvn());
1955 }
1956 for (int i = rpo_list.size() - 1; i >= 0; i--) {
1957 Node* c = rpo_list.at(i);
1958 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
1959 Node* n = regions[c->_idx];
1960 assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
1961 if (n != nullptr && n->is_Phi() && n->_idx >= last && n->in(0) == c) {
1962 _phase->register_new_node(n, c);
1963 }
1964 }
1965 }
1966 for (int i = rpo_list.size() - 1; i >= 0; i--) {
1967 Node* c = rpo_list.at(i);
1968 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
1969 Node* n = regions[c->_idx];
1970 assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
1971 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
1972 Node* u = c->fast_out(i);
1973 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
1974 u != n) {
1975 assert(c->unique_ctrl_out()->Opcode() != Op_Halt, "expected memory state");
1976 if (u->adr_type() == TypePtr::BOTTOM) {
1977 fix_memory_uses(u, n, n, c);
1978 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
1979 _phase->lazy_replace(u, n);
1980 --i; --imax;
1981 }
1982 }
1983 }
1984 }
1985 }
1986 }
1987
1988 Node* MemoryGraphFixer::collect_memory_for_infinite_loop(const Node* in) {
1989 Node* mem = nullptr;
1990 Node* head = in->in(0);
1991 assert(head->is_Region(), "unexpected infinite loop graph shape");
1992
1993 Node* phi_mem = nullptr;
1994 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
1995 Node* u = head->fast_out(j);
1996 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
1997 if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
1998 assert(phi_mem == nullptr || phi_mem->adr_type() == TypePtr::BOTTOM, "");
1999 phi_mem = u;
2000 } else if (u->adr_type() == TypePtr::BOTTOM) {
2001 assert(phi_mem == nullptr || _phase->C->get_alias_index(phi_mem->adr_type()) == _alias, "");
2002 if (phi_mem == nullptr) {
2003 phi_mem = u;
2004 }
2005 }
2006 }
2007 }
2008 if (phi_mem == nullptr) {
2009 ResourceMark rm;
2010 Node_Stack stack(0);
2011 stack.push(head, 1);
2012 do {
2013 Node* n = stack.node();
2014 uint i = stack.index();
2015 if (i >= n->req()) {
2016 stack.pop();
2017 } else {
2018 stack.set_index(i + 1);
2019 Node* c = n->in(i);
2020 assert(c != head, "should have found a safepoint on the way");
2021 if (stack.size() != 1 || _phase->is_dominator(head, c)) {
2022 for (;;) {
2023 if (c->is_Region()) {
2024 stack.push(c, 1);
2025 break;
2026 } else if (c->is_SafePoint() && !c->is_CallLeaf()) {
2027 Node* m = c->in(TypeFunc::Memory);
2028 if (m->is_MergeMem()) {
2029 m = m->as_MergeMem()->memory_at(_alias);
2030 }
2031 assert(mem == nullptr || mem == m, "several memory states");
2032 mem = m;
2033 break;
2034 } else {
2035 assert(c != c->in(0), "");
2036 c = c->in(0);
2037 }
2038 }
2039 }
2040 }
2041 } while (stack.size() > 0);
2042 assert(mem != nullptr, "should have found safepoint");
2043 } else {
2044 mem = phi_mem;
2045 }
2046 return mem;
2047 }
2048
2049 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
2050 Node* c = _phase->get_ctrl(n);
2051 if (n->is_Proj() && n->in(0) != nullptr && n->in(0)->is_Call()) {
2052 assert(c == n->in(0), "");
2053 CallNode* call = c->as_Call();
2054 CallProjections projs;
2055 call->extract_projections(&projs, true, false);
2056 if (projs.catchall_memproj != nullptr) {
2057 if (projs.fallthrough_memproj == n) {
2058 c = projs.fallthrough_catchproj;
2059 } else {
2060 assert(projs.catchall_memproj == n, "");
2061 c = projs.catchall_catchproj;
2062 }
2063 }
2064 }
2065 return c;
2066 }
2067
2068 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
2069 if (_phase->has_ctrl(n))
2070 return get_ctrl(n);
2071 else {
2072 assert (n->is_CFG(), "must be a CFG node");
2073 return n;
2074 }
2075 }
2076
2077 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
2078 return m != nullptr && get_ctrl(m) == c;
2079 }
2080
2081 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
2082 assert(n == nullptr || _phase->ctrl_or_self(n) == ctrl, "");
2083 assert(!ctrl->is_Call() || ctrl == n, "projection expected");
2084 #ifdef ASSERT
2085 if ((ctrl->is_Proj() && ctrl->in(0)->is_Call()) ||
2086 (ctrl->is_Catch() && ctrl->in(0)->in(0)->is_Call())) {
2087 CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_Call() : ctrl->in(0)->in(0)->as_Call();
2088 int mems = 0;
2089 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
2090 Node* u = call->fast_out(i);
2091 if (u->bottom_type() == Type::MEMORY) {
2092 mems++;
2093 }
2094 }
2095 assert(mems <= 1, "No node right after call if multiple mem projections");
2096 }
2097 #endif
2098 Node* mem = _memory_nodes[ctrl->_idx];
2099 Node* c = ctrl;
2100 while (!mem_is_valid(mem, c) &&
2101 (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
2102 c = _phase->idom(c);
2103 mem = _memory_nodes[c->_idx];
2104 }
2105 if (n != nullptr && mem_is_valid(mem, c)) {
2106 while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
2107 mem = next_mem(mem, _alias);
2108 }
2109 if (mem->is_MergeMem()) {
2110 mem = mem->as_MergeMem()->memory_at(_alias);
2111 }
2112 if (!mem_is_valid(mem, c)) {
2113 do {
2114 c = _phase->idom(c);
2115 mem = _memory_nodes[c->_idx];
2116 } while (!mem_is_valid(mem, c) &&
2117 (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
2118 }
2119 }
2120 assert(mem->bottom_type() == Type::MEMORY, "");
2121 return mem;
2122 }
2123
2124 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
2125 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2126 Node* use = region->fast_out(i);
2127 if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
2128 (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
2129 return true;
2130 }
2131 }
2132 return false;
2133 }
2134
2135 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
2136 assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
2137 const bool trace = false;
2138 DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
2139 DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
2140 GrowableArray<Node*> phis;
2141 if (mem_for_ctrl != mem) {
2142 Node* old = mem_for_ctrl;
2143 Node* prev = nullptr;
2144 while (old != mem) {
2145 prev = old;
2146 if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
2147 assert(_alias == Compile::AliasIdxRaw, "");
2148 old = old->in(MemNode::Memory);
2149 } else if (old->Opcode() == Op_SCMemProj) {
2150 assert(_alias == Compile::AliasIdxRaw, "");
2151 old = old->in(0);
2152 } else {
2153 ShouldNotReachHere();
2154 }
2155 }
2156 assert(prev != nullptr, "");
2157 if (new_ctrl != ctrl) {
2158 _memory_nodes.map(ctrl->_idx, mem);
2159 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2160 }
2161 uint input = (uint)MemNode::Memory;
2162 _phase->igvn().replace_input_of(prev, input, new_mem);
2163 } else {
2164 uses.clear();
2165 _memory_nodes.map(new_ctrl->_idx, new_mem);
2166 uses.push(new_ctrl);
2167 for(uint next = 0; next < uses.size(); next++ ) {
2168 Node *n = uses.at(next);
2169 assert(n->is_CFG(), "");
2170 DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
2171 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2172 Node* u = n->fast_out(i);
2173 if (!u->is_Root() && u->is_CFG() && u != n) {
2174 Node* m = _memory_nodes[u->_idx];
2175 if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
2176 !has_mem_phi(u) &&
2177 u->unique_ctrl_out()->Opcode() != Op_Halt) {
2178 DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
2179 DEBUG_ONLY(if (trace && m != nullptr) { tty->print("ZZZ mem"); m->dump(); });
2180
2181 if (!mem_is_valid(m, u) || !m->is_Phi()) {
2182 bool push = true;
2183 bool create_phi = true;
2184 if (_phase->is_dominator(new_ctrl, u)) {
2185 create_phi = false;
2186 }
2187 if (create_phi) {
2188 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
2189 _phase->register_new_node(phi, u);
2190 phis.push(phi);
2191 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
2192 if (!mem_is_valid(m, u)) {
2193 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
2194 _memory_nodes.map(u->_idx, phi);
2195 } else {
2196 DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
2197 for (;;) {
2198 assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), "");
2199 Node* next = nullptr;
2200 if (m->is_Proj()) {
2201 next = m->in(0);
2202 } else {
2203 assert(m->is_Mem() || m->is_LoadStore(), "");
2204 assert(_alias == Compile::AliasIdxRaw, "");
2205 next = m->in(MemNode::Memory);
2206 }
2207 if (_phase->get_ctrl(next) != u) {
2208 break;
2209 }
2210 if (next->is_MergeMem()) {
2211 assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
2212 break;
2213 }
2214 if (next->is_Phi()) {
2215 assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
2216 break;
2217 }
2218 m = next;
2219 }
2220
2221 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
2222 assert(m->is_Mem() || m->is_LoadStore(), "");
2223 uint input = (uint)MemNode::Memory;
2224 _phase->igvn().replace_input_of(m, input, phi);
2225 push = false;
2226 }
2227 } else {
2228 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
2229 }
2230 if (push) {
2231 uses.push(u);
2232 }
2233 }
2234 } else if (!mem_is_valid(m, u) &&
2235 !(u->Opcode() == Op_CProj && u->in(0)->is_NeverBranch() && u->as_Proj()->_con == 1)) {
2236 uses.push(u);
2237 }
2238 }
2239 }
2240 }
2241 for (int i = 0; i < phis.length(); i++) {
2242 Node* n = phis.at(i);
2243 Node* r = n->in(0);
2244 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
2245 for (uint j = 1; j < n->req(); j++) {
2246 Node* m = find_mem(r->in(j), nullptr);
2247 _phase->igvn().replace_input_of(n, j, m);
2248 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
2249 }
2250 }
2251 }
2252 uint last = _phase->C->unique();
2253 MergeMemNode* mm = nullptr;
2254 int alias = _alias;
2255 DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
2256 // Process loads first to not miss an anti-dependency: if the memory
2257 // edge of a store is updated before a load is processed then an
2258 // anti-dependency may be missed.
2259 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2260 Node* u = mem->out(i);
2261 if (u->_idx < last && u->is_Load() && _phase->C->get_alias_index(u->adr_type()) == alias) {
2262 Node* m = find_mem(_phase->get_ctrl(u), u);
2263 if (m != mem) {
2264 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2265 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2266 --i;
2267 }
2268 }
2269 }
2270 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2271 Node* u = mem->out(i);
2272 if (u->_idx < last) {
2273 if (u->is_Mem()) {
2274 if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2275 Node* m = find_mem(_phase->get_ctrl(u), u);
2276 if (m != mem) {
2277 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2278 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2279 --i;
2280 }
2281 }
2282 } else if (u->is_MergeMem()) {
2283 MergeMemNode* u_mm = u->as_MergeMem();
2284 if (u_mm->memory_at(alias) == mem) {
2285 MergeMemNode* newmm = nullptr;
2286 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2287 Node* uu = u->fast_out(j);
2288 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2289 if (uu->is_Phi()) {
2290 assert(uu->adr_type() == TypePtr::BOTTOM, "");
2291 Node* region = uu->in(0);
2292 int nb = 0;
2293 for (uint k = 1; k < uu->req(); k++) {
2294 if (uu->in(k) == u) {
2295 Node* m = find_mem(region->in(k), nullptr);
2296 if (m != mem) {
2297 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
2298 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2299 if (newmm != u) {
2300 _phase->igvn().replace_input_of(uu, k, newmm);
2301 nb++;
2302 --jmax;
2303 }
2304 }
2305 }
2306 }
2307 if (nb > 0) {
2308 --j;
2309 }
2310 } else {
2311 Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
2312 if (m != mem) {
2313 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
2314 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2315 if (newmm != u) {
2316 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2317 --j, --jmax;
2318 }
2319 }
2320 }
2321 }
2322 }
2323 } else if (u->is_Phi()) {
2324 assert(u->bottom_type() == Type::MEMORY, "what else?");
2325 if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
2326 Node* region = u->in(0);
2327 bool replaced = false;
2328 for (uint j = 1; j < u->req(); j++) {
2329 if (u->in(j) == mem) {
2330 Node* m = find_mem(region->in(j), nullptr);
2331 Node* nnew = m;
2332 if (m != mem) {
2333 if (u->adr_type() == TypePtr::BOTTOM) {
2334 mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
2335 nnew = mm;
2336 }
2337 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
2338 _phase->igvn().replace_input_of(u, j, nnew);
2339 replaced = true;
2340 }
2341 }
2342 }
2343 if (replaced) {
2344 --i;
2345 }
2346 }
2347 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2348 u->adr_type() == nullptr) {
2349 assert(u->adr_type() != nullptr ||
2350 u->Opcode() == Op_Rethrow ||
2351 u->Opcode() == Op_Return ||
2352 u->Opcode() == Op_SafePoint ||
2353 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2354 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2355 u->Opcode() == Op_CallLeaf, "");
2356 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2357 if (m != mem) {
2358 mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
2359 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2360 --i;
2361 }
2362 } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2363 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2364 if (m != mem) {
2365 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2366 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2367 --i;
2368 }
2369 } else if (u->adr_type() != TypePtr::BOTTOM &&
2370 _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
2371 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2372 assert(m != mem, "");
2373 // u is on the wrong slice...
2374 assert(u->is_ClearArray(), "");
2375 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2376 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2377 --i;
2378 }
2379 }
2380 }
2381 #ifdef ASSERT
2382 assert(new_mem->outcnt() > 0, "");
2383 for (int i = 0; i < phis.length(); i++) {
2384 Node* n = phis.at(i);
2385 assert(n->outcnt() > 0, "new phi must have uses now");
2386 }
2387 #endif
2388 }
2389
2390 void MemoryGraphFixer::record_new_ctrl(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl) {
2391 if (mem_for_ctrl != mem && new_ctrl != ctrl) {
2392 _memory_nodes.map(ctrl->_idx, mem);
2393 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2394 }
2395 }
2396
2397 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
2398 MergeMemNode* mm = MergeMemNode::make(mem);
2399 mm->set_memory_at(_alias, rep_proj);
2400 _phase->register_new_node(mm, rep_ctrl);
2401 return mm;
2402 }
2403
2404 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
2405 MergeMemNode* newmm = nullptr;
2406 MergeMemNode* u_mm = u->as_MergeMem();
2407 Node* c = _phase->get_ctrl(u);
2408 if (_phase->is_dominator(c, rep_ctrl)) {
2409 c = rep_ctrl;
2410 } else {
2411 assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
2412 }
2413 if (u->outcnt() == 1) {
2414 if (u->req() > (uint)_alias && u->in(_alias) == mem) {
2415 _phase->igvn().replace_input_of(u, _alias, rep_proj);
2416 --i;
2417 } else {
2418 _phase->igvn().rehash_node_delayed(u);
2419 u_mm->set_memory_at(_alias, rep_proj);
2420 }
2421 newmm = u_mm;
2422 _phase->set_ctrl_and_loop(u, c);
2423 } else {
2424 // can't simply clone u and then change one of its input because
2425 // it adds and then removes an edge which messes with the
2426 // DUIterator
2427 newmm = MergeMemNode::make(u_mm->base_memory());
2428 for (uint j = 0; j < u->req(); j++) {
2429 if (j < newmm->req()) {
2430 if (j == (uint)_alias) {
2431 newmm->set_req(j, rep_proj);
2432 } else if (newmm->in(j) != u->in(j)) {
2433 newmm->set_req(j, u->in(j));
2434 }
2435 } else if (j == (uint)_alias) {
2436 newmm->add_req(rep_proj);
2437 } else {
2438 newmm->add_req(u->in(j));
2439 }
2440 }
2441 if ((uint)_alias >= u->req()) {
2442 newmm->set_memory_at(_alias, rep_proj);
2443 }
2444 _phase->register_new_node(newmm, c);
2445 }
2446 return newmm;
2447 }
2448
2449 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
2450 if (phi->adr_type() == TypePtr::BOTTOM) {
2451 Node* region = phi->in(0);
2452 for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
2453 Node* uu = region->fast_out(j);
2454 if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
2455 return false;
2456 }
2457 }
2458 return true;
2459 }
2460 return _phase->C->get_alias_index(phi->adr_type()) == _alias;
2461 }
2462
2463 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
2464 uint last = _phase-> C->unique();
2465 MergeMemNode* mm = nullptr;
2466 assert(mem->bottom_type() == Type::MEMORY, "");
2467 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2468 Node* u = mem->out(i);
2469 if (u != replacement && u->_idx < last) {
2470 if (u->is_MergeMem()) {
2471 MergeMemNode* u_mm = u->as_MergeMem();
2472 if (u_mm->memory_at(_alias) == mem) {
2473 MergeMemNode* newmm = nullptr;
2474 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2475 Node* uu = u->fast_out(j);
2476 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2477 if (uu->is_Phi()) {
2478 if (should_process_phi(uu)) {
2479 Node* region = uu->in(0);
2480 int nb = 0;
2481 for (uint k = 1; k < uu->req(); k++) {
2482 if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
2483 if (newmm == nullptr) {
2484 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2485 }
2486 if (newmm != u) {
2487 _phase->igvn().replace_input_of(uu, k, newmm);
2488 nb++;
2489 --jmax;
2490 }
2491 }
2492 }
2493 if (nb > 0) {
2494 --j;
2495 }
2496 }
2497 } else {
2498 if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
2499 if (newmm == nullptr) {
2500 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2501 }
2502 if (newmm != u) {
2503 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2504 --j, --jmax;
2505 }
2506 }
2507 }
2508 }
2509 }
2510 } else if (u->is_Phi()) {
2511 assert(u->bottom_type() == Type::MEMORY, "what else?");
2512 Node* region = u->in(0);
2513 if (should_process_phi(u)) {
2514 bool replaced = false;
2515 for (uint j = 1; j < u->req(); j++) {
2516 if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
2517 Node* nnew = rep_proj;
2518 if (u->adr_type() == TypePtr::BOTTOM) {
2519 if (mm == nullptr) {
2520 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2521 }
2522 nnew = mm;
2523 }
2524 _phase->igvn().replace_input_of(u, j, nnew);
2525 replaced = true;
2526 }
2527 }
2528 if (replaced) {
2529 --i;
2530 }
2531
2532 }
2533 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2534 u->adr_type() == nullptr) {
2535 assert(u->adr_type() != nullptr ||
2536 u->Opcode() == Op_Rethrow ||
2537 u->Opcode() == Op_Return ||
2538 u->Opcode() == Op_SafePoint ||
2539 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2540 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2541 u->Opcode() == Op_CallLeaf, "%s", u->Name());
2542 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2543 if (mm == nullptr) {
2544 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2545 }
2546 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2547 --i;
2548 }
2549 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2550 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2551 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
2552 --i;
2553 }
2554 }
2555 }
2556 }
2557 }
2558
2559 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj, DecoratorSet decorators)
2560 : Node(ctrl, obj), _decorators(decorators) {
2561 ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this);
2562 }
2563
2564 DecoratorSet ShenandoahLoadReferenceBarrierNode::decorators() const {
2565 return _decorators;
2566 }
2567
2568 uint ShenandoahLoadReferenceBarrierNode::size_of() const {
2569 return sizeof(*this);
2570 }
2571
2572 static DecoratorSet mask_decorators(DecoratorSet decorators) {
2573 return decorators & (ON_STRONG_OOP_REF | ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF | ON_UNKNOWN_OOP_REF | IN_NATIVE);
2574 }
2575
2576 uint ShenandoahLoadReferenceBarrierNode::hash() const {
2577 uint hash = Node::hash();
2578 hash += mask_decorators(_decorators);
2579 return hash;
2580 }
2581
2582 bool ShenandoahLoadReferenceBarrierNode::cmp( const Node &n ) const {
2583 return Node::cmp(n) && n.Opcode() == Op_ShenandoahLoadReferenceBarrier &&
2584 mask_decorators(_decorators) == mask_decorators(((const ShenandoahLoadReferenceBarrierNode&)n)._decorators);
2585 }
2586
2587 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const {
2588 if (in(ValueIn) == nullptr || in(ValueIn)->is_top()) {
2589 return Type::TOP;
2590 }
2591 const Type* t = in(ValueIn)->bottom_type();
2592 if (t == TypePtr::NULL_PTR) {
2593 return t;
2594 }
2595
2596 if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2597 return t;
2598 }
2599
2600 return t->meet(TypePtr::NULL_PTR);
2601 }
2602
2603 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const {
2604 // Either input is TOP ==> the result is TOP
2605 const Type *t2 = phase->type(in(ValueIn));
2606 if( t2 == Type::TOP ) return Type::TOP;
2607
2608 if (t2 == TypePtr::NULL_PTR) {
2609 return t2;
2610 }
2611
2612 if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2613 return t2;
2614 }
2615
2616 return t2->meet(TypePtr::NULL_PTR);
2617 }
2618
2619 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) {
2620 Node* value = in(ValueIn);
2621 if (!needs_barrier(phase, value)) {
2622 return value;
2623 }
2624 return this;
2625 }
2626
2627 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) {
2628 Unique_Node_List visited;
2629 return needs_barrier_impl(phase, n, visited);
2630 }
2631
2632 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) {
2633 if (n == nullptr) return false;
2634 if (visited.member(n)) {
2635 return false; // Been there.
2636 }
2637 visited.push(n);
2638
2639 if (n->is_Allocate()) {
2640 // tty->print_cr("optimize barrier on alloc");
2641 return false;
2642 }
2643 if (n->is_Call()) {
2644 // tty->print_cr("optimize barrier on call");
2645 return false;
2646 }
2647
2648 const Type* type = phase->type(n);
2649 if (type == Type::TOP) {
2650 return false;
2651 }
2652 if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) {
2653 // tty->print_cr("optimize barrier on null");
2654 return false;
2655 }
2656 if (type->make_oopptr() && type->make_oopptr()->const_oop() != nullptr) {
2657 // tty->print_cr("optimize barrier on constant");
2658 return false;
2659 }
2660
2661 switch (n->Opcode()) {
2662 case Op_AddP:
2663 return true; // TODO: Can refine?
2664 case Op_LoadP:
2665 case Op_ShenandoahCompareAndExchangeN:
2666 case Op_ShenandoahCompareAndExchangeP:
2667 case Op_CompareAndExchangeN:
2668 case Op_CompareAndExchangeP:
2669 case Op_GetAndSetN:
2670 case Op_GetAndSetP:
2671 return true;
2672 case Op_Phi: {
2673 for (uint i = 1; i < n->req(); i++) {
2674 if (needs_barrier_impl(phase, n->in(i), visited)) return true;
2675 }
2676 return false;
2677 }
2678 case Op_CheckCastPP:
2679 case Op_CastPP:
2680 return needs_barrier_impl(phase, n->in(1), visited);
2681 case Op_Proj:
2682 return needs_barrier_impl(phase, n->in(0), visited);
2683 case Op_ShenandoahLoadReferenceBarrier:
2684 // tty->print_cr("optimize barrier on barrier");
2685 return false;
2686 case Op_Parm:
2687 // tty->print_cr("optimize barrier on input arg");
2688 return false;
2689 case Op_DecodeN:
2690 case Op_EncodeP:
2691 return needs_barrier_impl(phase, n->in(1), visited);
2692 case Op_LoadN:
2693 return true;
2694 case Op_CMoveN:
2695 case Op_CMoveP:
2696 return needs_barrier_impl(phase, n->in(2), visited) ||
2697 needs_barrier_impl(phase, n->in(3), visited);
2698 case Op_CreateEx:
2699 return false;
2700 default:
2701 break;
2702 }
2703 #ifdef ASSERT
2704 tty->print("need barrier on?: ");
2705 tty->print_cr("ins:");
2706 n->dump(2);
2707 tty->print_cr("outs:");
2708 n->dump(-2);
2709 ShouldNotReachHere();
2710 #endif
2711 return true;
2712 }