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
27 #include "classfile/javaClasses.hpp"
28 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
29 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
30 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
31 #include "gc/shenandoah/shenandoahForwarding.hpp"
32 #include "gc/shenandoah/shenandoahHeap.hpp"
33 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
34 #include "gc/shenandoah/shenandoahRuntime.hpp"
35 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
36 #include "opto/arraycopynode.hpp"
37 #include "opto/block.hpp"
38 #include "opto/callnode.hpp"
39 #include "opto/castnode.hpp"
40 #include "opto/movenode.hpp"
41 #include "opto/phaseX.hpp"
42 #include "opto/rootnode.hpp"
43 #include "opto/runtime.hpp"
44 #include "opto/subnode.hpp"
45
46 bool ShenandoahBarrierC2Support::expand(Compile* C, PhaseIterGVN& igvn) {
47 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
48 if (state->load_reference_barriers_count() > 0) {
49 assert(C->post_loop_opts_phase(), "no loop opts allowed");
50 C->reset_post_loop_opts_phase(); // ... but we know what we are doing
51 C->clear_major_progress();
52 PhaseIdealLoop::optimize(igvn, LoopOptsShenandoahExpand);
53 if (C->failing()) return false;
54
55 C->set_major_progress();
56 if (!C->optimize_loops(igvn, LoopOptsShenandoahPostExpand)) {
57 return false;
58 }
59 C->clear_major_progress();
60 C->process_for_post_loop_opts_igvn(igvn);
61 if (C->failing()) return false;
62
63 C->set_post_loop_opts_phase(); // now for real!
64 }
65 return true;
66 }
67
68 bool ShenandoahBarrierC2Support::is_gc_state_test(Node* iff, int mask) {
69 if (!UseShenandoahGC) {
70 return false;
71 }
72 assert(iff->is_If(), "bad input");
73 if (iff->Opcode() != Op_If) {
74 return false;
75 }
76 Node* bol = iff->in(1);
77 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
78 return false;
79 }
80 Node* cmp = bol->in(1);
81 if (cmp->Opcode() != Op_CmpI) {
82 return false;
83 }
84 Node* in1 = cmp->in(1);
85 Node* in2 = cmp->in(2);
86 if (in2->find_int_con(-1) != 0) {
87 return false;
88 }
89 if (in1->Opcode() != Op_AndI) {
90 return false;
91 }
92 in2 = in1->in(2);
93 if (in2->find_int_con(-1) != mask) {
94 return false;
95 }
96 in1 = in1->in(1);
97
98 return is_gc_state_load(in1);
99 }
100
101 bool ShenandoahBarrierC2Support::is_heap_stable_test(Node* iff) {
102 return is_gc_state_test(iff, ShenandoahHeap::HAS_FORWARDED);
103 }
104
105 bool ShenandoahBarrierC2Support::is_gc_state_load(Node *n) {
106 if (!UseShenandoahGC) {
107 return false;
108 }
109 if (n->Opcode() != Op_LoadB && n->Opcode() != Op_LoadUB) {
110 return false;
111 }
112 Node* addp = n->in(MemNode::Address);
113 if (!addp->is_AddP()) {
114 return false;
115 }
116 Node* base = addp->in(AddPNode::Address);
117 Node* off = addp->in(AddPNode::Offset);
118 if (base->Opcode() != Op_ThreadLocal) {
119 return false;
120 }
121 if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) {
122 return false;
123 }
124 return true;
125 }
126
127 bool ShenandoahBarrierC2Support::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) {
128 assert(phase->is_dominator(stop, start), "bad inputs");
129 ResourceMark rm;
130 Unique_Node_List wq;
131 wq.push(start);
132 for (uint next = 0; next < wq.size(); next++) {
133 Node *m = wq.at(next);
134 if (m == stop) {
135 continue;
136 }
137 if (m->is_SafePoint() && !m->is_CallLeaf()) {
138 return true;
139 }
140 if (m->is_Region()) {
141 for (uint i = 1; i < m->req(); i++) {
142 wq.push(m->in(i));
143 }
144 } else {
145 wq.push(m->in(0));
146 }
147 }
148 return false;
149 }
150
151 #ifdef ASSERT
152 bool ShenandoahBarrierC2Support::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) {
153 assert(phis.size() == 0, "");
154
155 while (true) {
156 if (in->bottom_type() == TypePtr::NULL_PTR) {
157 if (trace) {tty->print_cr("null");}
158 } else if (!in->bottom_type()->make_ptr()->make_oopptr()) {
159 if (trace) {tty->print_cr("Non oop");}
160 } else {
161 if (in->is_ConstraintCast()) {
162 in = in->in(1);
163 continue;
164 } else if (in->is_AddP()) {
165 assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access");
166 in = in->in(AddPNode::Address);
167 continue;
168 } else if (in->is_Con()) {
169 if (trace) {
170 tty->print("Found constant");
171 in->dump();
172 }
173 } else if (in->Opcode() == Op_Parm) {
174 if (trace) {
175 tty->print("Found argument");
176 }
177 } else if (in->Opcode() == Op_CreateEx) {
178 if (trace) {
179 tty->print("Found create-exception");
180 }
181 } else if (in->Opcode() == Op_LoadP && in->adr_type() == TypeRawPtr::BOTTOM) {
182 if (trace) {
183 tty->print("Found raw LoadP (OSR argument?)");
184 }
185 } else if (in->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
186 if (t == ShenandoahOopStore) {
187 return false;
188 }
189 barriers_used.push(in);
190 if (trace) {tty->print("Found barrier"); in->dump();}
191 } else if (in->is_Proj() && in->in(0)->is_Allocate()) {
192 if (trace) {
193 tty->print("Found alloc");
194 in->in(0)->dump();
195 }
196 } else if (in->is_Proj() && (in->in(0)->Opcode() == Op_CallStaticJava || in->in(0)->Opcode() == Op_CallDynamicJava)) {
197 if (trace) {
198 tty->print("Found Java call");
199 }
200 } else if (in->is_Phi()) {
201 if (!visited.test_set(in->_idx)) {
202 if (trace) {tty->print("Pushed phi:"); in->dump();}
203 phis.push(in, 2);
204 in = in->in(1);
205 continue;
206 }
207 if (trace) {tty->print("Already seen phi:"); in->dump();}
208 } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) {
209 if (!visited.test_set(in->_idx)) {
210 if (trace) {tty->print("Pushed cmovep:"); in->dump();}
211 phis.push(in, CMoveNode::IfTrue);
212 in = in->in(CMoveNode::IfFalse);
213 continue;
214 }
215 if (trace) {tty->print("Already seen cmovep:"); in->dump();}
216 } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) {
217 in = in->in(1);
218 continue;
219 } else {
220 return false;
221 }
222 }
223 bool cont = false;
224 while (phis.is_nonempty()) {
225 uint idx = phis.index();
226 Node* phi = phis.node();
227 if (idx >= phi->req()) {
228 if (trace) {tty->print("Popped phi:"); phi->dump();}
229 phis.pop();
230 continue;
231 }
232 if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();}
233 in = phi->in(idx);
234 phis.set_index(idx+1);
235 cont = true;
236 break;
237 }
238 if (!cont) {
239 break;
240 }
241 }
242 return true;
243 }
244
245 void ShenandoahBarrierC2Support::report_verify_failure(const char* msg, Node* n1, Node* n2) {
246 if (n1 != nullptr) {
247 n1->dump(+10);
248 }
249 if (n2 != nullptr) {
250 n2->dump(+10);
251 }
252 fatal("%s", msg);
253 }
254
255 void ShenandoahBarrierC2Support::verify(RootNode* root) {
256 ResourceMark rm;
257 Unique_Node_List wq;
258 GrowableArray<Node*> barriers;
259 Unique_Node_List barriers_used;
260 Node_Stack phis(0);
261 VectorSet visited;
262 const bool trace = false;
263 const bool verify_no_useless_barrier = false;
264
265 wq.push(root);
266 for (uint next = 0; next < wq.size(); next++) {
267 Node *n = wq.at(next);
268 if (n->is_Load()) {
269 const bool trace = false;
270 if (trace) {tty->print("Verifying"); n->dump();}
271 if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) {
272 if (trace) {tty->print_cr("Load range/klass");}
273 } else {
274 const TypePtr* adr_type = n->as_Load()->adr_type();
275
276 if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) {
277 if (trace) {tty->print_cr("Mark load");}
278 } else if (adr_type->isa_instptr() &&
279 adr_type->is_instptr()->instance_klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
280 adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset()) {
281 if (trace) {tty->print_cr("Reference.get()");}
282 } else if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) {
283 report_verify_failure("Shenandoah verification: Load should have barriers", n);
284 }
285 }
286 } else if (n->is_Store()) {
287 const bool trace = false;
288
289 if (trace) {tty->print("Verifying"); n->dump();}
290 if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) {
291 Node* adr = n->in(MemNode::Address);
292 bool verify = true;
293
294 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) {
295 adr = adr->in(AddPNode::Address);
296 if (adr->is_AddP()) {
297 assert(adr->in(AddPNode::Base)->is_top(), "");
298 adr = adr->in(AddPNode::Address);
299 if (adr->Opcode() == Op_LoadP &&
300 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() &&
301 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
302 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) {
303 if (trace) {tty->print_cr("SATB prebarrier");}
304 verify = false;
305 }
306 }
307 }
308
309 if (verify && !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahValue, trace, barriers_used)) {
310 report_verify_failure("Shenandoah verification: Store should have barriers", n);
311 }
312 }
313 if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
314 report_verify_failure("Shenandoah verification: Store (address) should have barriers", n);
315 }
316 } else if (n->Opcode() == Op_CmpP) {
317 const bool trace = false;
318
319 Node* in1 = n->in(1);
320 Node* in2 = n->in(2);
321 if (in1->bottom_type()->isa_oopptr()) {
322 if (trace) {tty->print("Verifying"); n->dump();}
323
324 bool mark_inputs = false;
325 if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR ||
326 (in1->is_Con() || in2->is_Con())) {
327 if (trace) {tty->print_cr("Comparison against a constant");}
328 mark_inputs = true;
329 } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) ||
330 (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) {
331 if (trace) {tty->print_cr("Comparison with newly alloc'ed object");}
332 mark_inputs = true;
333 } else {
334 assert(in2->bottom_type()->isa_oopptr(), "");
335
336 if (!verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) ||
337 !verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) {
338 report_verify_failure("Shenandoah verification: Cmp should have barriers", n);
339 }
340 }
341 if (verify_no_useless_barrier &&
342 mark_inputs &&
343 (!verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) ||
344 !verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) {
345 phis.clear();
346 visited.reset();
347 }
348 }
349 } else if (n->is_LoadStore()) {
350 if (n->in(MemNode::ValueIn)->bottom_type()->make_ptr() &&
351 !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahValue, trace, barriers_used)) {
352 report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n);
353 }
354
355 if (n->in(MemNode::Address)->bottom_type()->make_oopptr() && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
356 report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n);
357 }
358 } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) {
359 CallNode* call = n->as_Call();
360
361 static struct {
362 const char* name;
363 struct {
364 int pos;
365 verify_type t;
366 } args[6];
367 } calls[] = {
368 "array_partition_stub",
369 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+4, ShenandoahStore }, { -1, ShenandoahNone },
370 { -1, ShenandoahNone }, { -1, ShenandoahNone }, { -1, ShenandoahNone } },
371 "arraysort_stub",
372 { { TypeFunc::Parms, ShenandoahStore }, { -1, ShenandoahNone }, { -1, ShenandoahNone },
373 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
374 "aescrypt_encryptBlock",
375 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
376 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
377 "aescrypt_decryptBlock",
378 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
379 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
380 "multiplyToLen",
381 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { TypeFunc::Parms+4, ShenandoahStore },
382 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
383 "squareToLen",
384 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { -1, ShenandoahNone},
385 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
386 "montgomery_multiply",
387 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
388 { TypeFunc::Parms+6, ShenandoahStore }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
389 "montgomery_square",
390 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+5, ShenandoahStore },
391 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
392 "mulAdd",
393 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
394 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
395 "vectorizedMismatch",
396 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
397 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
398 "updateBytesCRC32",
399 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
400 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
401 "updateBytesAdler32",
402 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
403 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
404 "updateBytesCRC32C",
405 { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad}, { -1, ShenandoahNone},
406 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
407 "counterMode_AESCrypt",
408 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
409 { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } },
410 "cipherBlockChaining_encryptAESCrypt",
411 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
412 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
413 "cipherBlockChaining_decryptAESCrypt",
414 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
415 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
416 "shenandoah_clone",
417 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
418 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
419 "ghash_processBlocks",
420 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
421 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
422 "sha1_implCompress",
423 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
424 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
425 "sha256_implCompress",
426 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
427 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
428 "sha512_implCompress",
429 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
430 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
431 "sha1_implCompressMB",
432 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
433 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
434 "sha256_implCompressMB",
435 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
436 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
437 "sha512_implCompressMB",
438 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
439 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
440 "encodeBlock",
441 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahStore }, { -1, ShenandoahNone },
442 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
443 "decodeBlock",
444 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahStore }, { -1, ShenandoahNone },
445 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
446 "intpoly_montgomeryMult_P256",
447 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahStore },
448 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
449 "intpoly_assign",
450 { { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad }, { -1, ShenandoahNone },
451 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
452 };
453
454 if (call->is_call_to_arraycopystub()) {
455 Node* dest = nullptr;
456 const TypeTuple* args = n->as_Call()->_tf->domain();
457 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
458 if (args->field_at(i)->isa_ptr()) {
459 j++;
460 if (j == 2) {
461 dest = n->in(i);
462 break;
463 }
464 }
465 }
466 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) ||
467 !verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) {
468 report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n);
469 }
470 } else if (strlen(call->_name) > 5 &&
471 !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) {
472 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) {
473 report_verify_failure("Shenandoah verification: _fill should have barriers", n);
474 }
475 } else if (!strcmp(call->_name, "shenandoah_wb_pre")) {
476 // skip
477 } else {
478 const int calls_len = sizeof(calls) / sizeof(calls[0]);
479 int i = 0;
480 for (; i < calls_len; i++) {
481 if (!strcmp(calls[i].name, call->_name)) {
482 break;
483 }
484 }
485 if (i != calls_len) {
486 const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]);
487 for (uint j = 0; j < args_len; j++) {
488 int pos = calls[i].args[j].pos;
489 if (pos == -1) {
490 break;
491 }
492 if (!verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) {
493 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
494 }
495 }
496 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
497 if (call->in(j)->bottom_type()->make_ptr() &&
498 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
499 uint k = 0;
500 for (; k < args_len && calls[i].args[k].pos != (int)j; k++);
501 if (k == args_len) {
502 fatal("arg %d for call %s not covered", j, call->_name);
503 }
504 }
505 }
506 } else {
507 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
508 if (call->in(j)->bottom_type()->make_ptr() &&
509 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
510 fatal("%s not covered", call->_name);
511 }
512 }
513 }
514 }
515 } else if (n->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
516 // skip
517 } else if (n->is_AddP()
518 || n->is_Phi()
519 || n->is_ConstraintCast()
520 || n->Opcode() == Op_Return
521 || n->Opcode() == Op_CMoveP
522 || n->Opcode() == Op_CMoveN
523 || n->Opcode() == Op_Rethrow
524 || n->is_MemBar()
525 || n->Opcode() == Op_Conv2B
526 || n->Opcode() == Op_SafePoint
527 || n->is_CallJava()
528 || n->Opcode() == Op_Unlock
529 || n->Opcode() == Op_EncodeP
530 || n->Opcode() == Op_DecodeN) {
531 // nothing to do
532 } else {
533 static struct {
534 int opcode;
535 struct {
536 int pos;
537 verify_type t;
538 } inputs[2];
539 } others[] = {
540 Op_FastLock,
541 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
542 Op_Lock,
543 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone} },
544 Op_ArrayCopy,
545 { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } },
546 Op_StrCompressedCopy,
547 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
548 Op_StrInflatedCopy,
549 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
550 Op_AryEq,
551 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
552 Op_StrIndexOf,
553 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
554 Op_StrComp,
555 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
556 Op_StrEquals,
557 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
558 Op_VectorizedHashCode,
559 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } },
560 Op_EncodeISOArray,
561 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
562 Op_CountPositives,
563 { { 2, ShenandoahLoad }, { -1, ShenandoahNone} },
564 Op_CastP2X,
565 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
566 Op_StrIndexOfChar,
567 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } },
568 };
569
570 const int others_len = sizeof(others) / sizeof(others[0]);
571 int i = 0;
572 for (; i < others_len; i++) {
573 if (others[i].opcode == n->Opcode()) {
574 break;
575 }
576 }
577 uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req();
578 if (i != others_len) {
579 const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]);
580 for (uint j = 0; j < inputs_len; j++) {
581 int pos = others[i].inputs[j].pos;
582 if (pos == -1) {
583 break;
584 }
585 if (!verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) {
586 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
587 }
588 }
589 for (uint j = 1; j < stop; j++) {
590 if (n->in(j) != nullptr && n->in(j)->bottom_type()->make_ptr() &&
591 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
592 uint k = 0;
593 for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++);
594 if (k == inputs_len) {
595 fatal("arg %d for node %s not covered", j, n->Name());
596 }
597 }
598 }
599 } else {
600 for (uint j = 1; j < stop; j++) {
601 if (n->in(j) != nullptr && n->in(j)->bottom_type()->make_ptr() &&
602 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
603 fatal("%s not covered", n->Name());
604 }
605 }
606 }
607 }
608
609 if (n->is_SafePoint()) {
610 SafePointNode* sfpt = n->as_SafePoint();
611 if (verify_no_useless_barrier && sfpt->jvms() != nullptr) {
612 for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) {
613 if (!verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) {
614 phis.clear();
615 visited.reset();
616 }
617 }
618 }
619 }
620 }
621
622 if (verify_no_useless_barrier) {
623 for (int i = 0; i < barriers.length(); i++) {
624 Node* n = barriers.at(i);
625 if (!barriers_used.member(n)) {
626 tty->print("XXX useless barrier"); n->dump(-2);
627 ShouldNotReachHere();
628 }
629 }
630 }
631 }
632 #endif
633
634 bool ShenandoahBarrierC2Support::is_dominator_same_ctrl(Node* c, Node* d, Node* n, PhaseIdealLoop* phase) {
635 // That both nodes have the same control is not sufficient to prove
636 // domination, verify that there's no path from d to n
637 ResourceMark rm;
638 Unique_Node_List wq;
639 wq.push(d);
640 for (uint next = 0; next < wq.size(); next++) {
641 Node *m = wq.at(next);
642 if (m == n) {
643 return false;
644 }
645 if (m->is_Phi() && m->in(0)->is_Loop()) {
646 assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control");
647 } else {
648 if (m->is_Store() || m->is_LoadStore()) {
649 // Take anti-dependencies into account
650 Node* mem = m->in(MemNode::Memory);
651 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
652 Node* u = mem->fast_out(i);
653 if (u->is_Load() && phase->C->can_alias(m->adr_type(), phase->C->get_alias_index(u->adr_type())) &&
654 phase->ctrl_or_self(u) == c) {
655 wq.push(u);
656 }
657 }
658 }
659 for (uint i = 0; i < m->req(); i++) {
660 if (m->in(i) != nullptr && phase->ctrl_or_self(m->in(i)) == c) {
661 wq.push(m->in(i));
662 }
663 }
664 }
665 }
666 return true;
667 }
668
669 bool ShenandoahBarrierC2Support::is_dominator(Node* d_c, Node* n_c, Node* d, Node* n, PhaseIdealLoop* phase) {
670 if (d_c != n_c) {
671 return phase->is_dominator(d_c, n_c);
672 }
673 return is_dominator_same_ctrl(d_c, d, n, phase);
674 }
675
676 Node* next_mem(Node* mem, int alias) {
677 Node* res = nullptr;
678 if (mem->is_Proj()) {
679 res = mem->in(0);
680 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
681 res = mem->in(TypeFunc::Memory);
682 } else if (mem->is_Phi()) {
683 res = mem->in(1);
684 } else if (mem->is_MergeMem()) {
685 res = mem->as_MergeMem()->memory_at(alias);
686 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
687 assert(alias == Compile::AliasIdxRaw, "following raw memory can't lead to a barrier");
688 res = mem->in(MemNode::Memory);
689 } else {
690 #ifdef ASSERT
691 mem->dump();
692 #endif
693 ShouldNotReachHere();
694 }
695 return res;
696 }
697
698 Node* ShenandoahBarrierC2Support::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) {
699 Node* iffproj = nullptr;
700 while (c != dom) {
701 Node* next = phase->idom(c);
702 assert(next->unique_ctrl_out_or_null() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?");
703 if (c->is_Region()) {
704 ResourceMark rm;
705 Unique_Node_List wq;
706 wq.push(c);
707 for (uint i = 0; i < wq.size(); i++) {
708 Node *n = wq.at(i);
709 if (n == next) {
710 continue;
711 }
712 if (n->is_Region()) {
713 for (uint j = 1; j < n->req(); j++) {
714 wq.push(n->in(j));
715 }
716 } else {
717 wq.push(n->in(0));
718 }
719 }
720 for (uint i = 0; i < wq.size(); i++) {
721 Node *n = wq.at(i);
722 assert(n->is_CFG(), "");
723 if (n->is_Multi()) {
724 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
725 Node* u = n->fast_out(j);
726 if (u->is_CFG()) {
727 if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj()) {
728 return NodeSentinel;
729 }
730 }
731 }
732 }
733 }
734 } else if (c->is_Proj()) {
735 if (c->is_IfProj()) {
736 if (c->as_Proj()->is_uncommon_trap_if_pattern() != nullptr) {
737 // continue;
738 } else {
739 if (!allow_one_proj) {
740 return NodeSentinel;
741 }
742 if (iffproj == nullptr) {
743 iffproj = c;
744 } else {
745 return NodeSentinel;
746 }
747 }
748 } else if (c->Opcode() == Op_JumpProj) {
749 return NodeSentinel; // unsupported
750 } else if (c->Opcode() == Op_CatchProj) {
751 return NodeSentinel; // unsupported
752 } else if (c->Opcode() == Op_CProj && next->is_NeverBranch()) {
753 return NodeSentinel; // unsupported
754 } else {
755 assert(next->unique_ctrl_out() == c, "unsupported branch pattern");
756 }
757 }
758 c = next;
759 }
760 return iffproj;
761 }
762
763 Node* ShenandoahBarrierC2Support::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) {
764 ResourceMark rm;
765 VectorSet wq;
766 wq.set(mem->_idx);
767 mem_ctrl = phase->ctrl_or_self(mem);
768 while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) {
769 mem = next_mem(mem, alias);
770 if (wq.test_set(mem->_idx)) {
771 return nullptr;
772 }
773 mem_ctrl = phase->ctrl_or_self(mem);
774 }
775 if (mem->is_MergeMem()) {
776 mem = mem->as_MergeMem()->memory_at(alias);
777 mem_ctrl = phase->ctrl_or_self(mem);
778 }
779 return mem;
780 }
781
782 Node* ShenandoahBarrierC2Support::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) {
783 Node* mem = nullptr;
784 Node* c = ctrl;
785 do {
786 if (c->is_Region()) {
787 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == nullptr; i++) {
788 Node* u = c->fast_out(i);
789 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
790 if (u->adr_type() == TypePtr::BOTTOM) {
791 mem = u;
792 }
793 }
794 }
795 } else {
796 if (c->is_Call() && c->as_Call()->adr_type() != nullptr) {
797 CallProjections projs;
798 c->as_Call()->extract_projections(&projs, true, false);
799 if (projs.fallthrough_memproj != nullptr) {
800 if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
801 if (projs.catchall_memproj == nullptr) {
802 mem = projs.fallthrough_memproj;
803 } else {
804 if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) {
805 mem = projs.fallthrough_memproj;
806 } else {
807 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier");
808 mem = projs.catchall_memproj;
809 }
810 }
811 }
812 } else {
813 Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
814 if (proj != nullptr &&
815 proj->adr_type() == TypePtr::BOTTOM) {
816 mem = proj;
817 }
818 }
819 } else {
820 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
821 Node* u = c->fast_out(i);
822 if (u->is_Proj() &&
823 u->bottom_type() == Type::MEMORY &&
824 u->adr_type() == TypePtr::BOTTOM) {
825 assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
826 assert(mem == nullptr, "only one proj");
827 mem = u;
828 }
829 }
830 assert(!c->is_Call() || c->as_Call()->adr_type() != nullptr || mem == nullptr, "no mem projection expected");
831 }
832 }
833 c = phase->idom(c);
834 } while (mem == nullptr);
835 return mem;
836 }
837
838 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
839 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
840 Node* u = n->fast_out(i);
841 if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
842 uses.push(u);
843 }
844 }
845 }
846
847 static void hide_strip_mined_loop(OuterStripMinedLoopNode* outer, CountedLoopNode* inner, PhaseIdealLoop* phase) {
848 OuterStripMinedLoopEndNode* le = inner->outer_loop_end();
849 Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl));
850 phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl));
851 Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt);
852 phase->register_control(new_le, phase->get_loop(le), le->in(0));
853 phase->lazy_replace(outer, new_outer);
854 phase->lazy_replace(le, new_le);
855 inner->clear_strip_mined();
856 }
857
858 void ShenandoahBarrierC2Support::test_gc_state(Node*& ctrl, Node* raw_mem, Node*& test_fail_ctrl,
859 PhaseIdealLoop* phase, int flags) {
860 PhaseIterGVN& igvn = phase->igvn();
861 Node* old_ctrl = ctrl;
862
863 Node* thread = new ThreadLocalNode();
864 Node* gc_state_offset = igvn.MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
865 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, gc_state_offset);
866 Node* gc_state = new LoadBNode(old_ctrl, raw_mem, gc_state_addr,
867 DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
868 TypeInt::BYTE, MemNode::unordered);
869 Node* gc_state_and = new AndINode(gc_state, igvn.intcon(flags));
870 Node* gc_state_cmp = new CmpINode(gc_state_and, igvn.zerocon(T_INT));
871 Node* gc_state_bool = new BoolNode(gc_state_cmp, BoolTest::ne);
872
873 IfNode* gc_state_iff = new IfNode(old_ctrl, gc_state_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
874 ctrl = new IfTrueNode(gc_state_iff);
875 test_fail_ctrl = new IfFalseNode(gc_state_iff);
876
877 IdealLoopTree* loop = phase->get_loop(old_ctrl);
878 phase->register_control(gc_state_iff, loop, old_ctrl);
879 phase->register_control(ctrl, loop, gc_state_iff);
880 phase->register_control(test_fail_ctrl, loop, gc_state_iff);
881
882 phase->register_new_node(thread, old_ctrl);
883 phase->register_new_node(gc_state_addr, old_ctrl);
884 phase->register_new_node(gc_state, old_ctrl);
885 phase->register_new_node(gc_state_and, old_ctrl);
886 phase->register_new_node(gc_state_cmp, old_ctrl);
887 phase->register_new_node(gc_state_bool, old_ctrl);
888
889 phase->set_root_as_ctrl(gc_state_offset);
890
891 assert(is_gc_state_test(gc_state_iff, flags), "Should match the shape");
892 }
893
894 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) {
895 Node* old_ctrl = ctrl;
896 PhaseIterGVN& igvn = phase->igvn();
897
898 const Type* val_t = igvn.type(val);
899 if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
900 Node* null_cmp = new CmpPNode(val, igvn.zerocon(T_OBJECT));
901 Node* null_test = new BoolNode(null_cmp, BoolTest::ne);
902
903 IfNode* null_iff = new IfNode(old_ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
904 ctrl = new IfTrueNode(null_iff);
905 null_ctrl = new IfFalseNode(null_iff);
906
907 IdealLoopTree* loop = phase->get_loop(old_ctrl);
908 phase->register_control(null_iff, loop, old_ctrl);
909 phase->register_control(ctrl, loop, null_iff);
910 phase->register_control(null_ctrl, loop, null_iff);
911
912 phase->register_new_node(null_cmp, old_ctrl);
913 phase->register_new_node(null_test, old_ctrl);
914 }
915 }
916
917 void ShenandoahBarrierC2Support::test_in_cset(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) {
918 Node* old_ctrl = ctrl;
919 PhaseIterGVN& igvn = phase->igvn();
920
921 Node* raw_val = new CastP2XNode(old_ctrl, val);
922 Node* cset_idx = new URShiftXNode(raw_val, igvn.intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
923
924 // Figure out the target cset address with raw pointer math.
925 // This avoids matching AddP+LoadB that would emit inefficient code.
926 // See JDK-8245465.
927 Node* cset_addr_ptr = igvn.makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
928 Node* cset_addr = new CastP2XNode(old_ctrl, cset_addr_ptr);
929 Node* cset_load_addr = new AddXNode(cset_addr, cset_idx);
930 Node* cset_load_ptr = new CastX2PNode(cset_load_addr);
931
932 Node* cset_load = new LoadBNode(old_ctrl, raw_mem, cset_load_ptr,
933 DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
934 TypeInt::BYTE, MemNode::unordered);
935 Node* cset_cmp = new CmpINode(cset_load, igvn.zerocon(T_INT));
936 Node* cset_bool = new BoolNode(cset_cmp, BoolTest::ne);
937
938 IfNode* cset_iff = new IfNode(old_ctrl, cset_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
939 ctrl = new IfTrueNode(cset_iff);
940 not_cset_ctrl = new IfFalseNode(cset_iff);
941
942 IdealLoopTree *loop = phase->get_loop(old_ctrl);
943 phase->register_control(cset_iff, loop, old_ctrl);
944 phase->register_control(ctrl, loop, cset_iff);
945 phase->register_control(not_cset_ctrl, loop, cset_iff);
946
947 phase->set_root_as_ctrl(cset_addr_ptr);
948
949 phase->register_new_node(raw_val, old_ctrl);
950 phase->register_new_node(cset_idx, old_ctrl);
951 phase->register_new_node(cset_addr, old_ctrl);
952 phase->register_new_node(cset_load_addr, old_ctrl);
953 phase->register_new_node(cset_load_ptr, old_ctrl);
954 phase->register_new_node(cset_load, old_ctrl);
955 phase->register_new_node(cset_cmp, old_ctrl);
956 phase->register_new_node(cset_bool, old_ctrl);
957 }
958
959 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node* load_addr,
960 DecoratorSet decorators, PhaseIdealLoop* phase) {
961 IdealLoopTree*loop = phase->get_loop(ctrl);
962 const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr();
963
964 address calladdr = nullptr;
965 const char* name = nullptr;
966 bool is_strong = ShenandoahBarrierSet::is_strong_access(decorators);
967 bool is_weak = ShenandoahBarrierSet::is_weak_access(decorators);
968 bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
969 bool is_native = ShenandoahBarrierSet::is_native_access(decorators);
970 bool is_narrow = UseCompressedOops && !is_native;
971 if (is_strong) {
972 if (is_narrow) {
973 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow);
974 name = "load_reference_barrier_strong_narrow";
975 } else {
976 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
977 name = "load_reference_barrier_strong";
978 }
979 } else if (is_weak) {
980 if (is_narrow) {
981 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow);
982 name = "load_reference_barrier_weak_narrow";
983 } else {
984 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak);
985 name = "load_reference_barrier_weak";
986 }
987 } else {
988 assert(is_phantom, "only remaining strength");
989 if (is_narrow) {
990 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom_narrow);
991 name = "load_reference_barrier_phantom_narrow";
992 } else {
993 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom);
994 name = "load_reference_barrier_phantom";
995 }
996 }
997 Node* call = new CallLeafNode(ShenandoahBarrierSetC2::load_reference_barrier_Type(), calladdr, name, TypeRawPtr::BOTTOM);
998
999 call->init_req(TypeFunc::Control, ctrl);
1000 call->init_req(TypeFunc::I_O, phase->C->top());
1001 call->init_req(TypeFunc::Memory, phase->C->top());
1002 call->init_req(TypeFunc::FramePtr, phase->C->top());
1003 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1004 call->init_req(TypeFunc::Parms, val);
1005 call->init_req(TypeFunc::Parms+1, load_addr);
1006 phase->register_control(call, loop, ctrl);
1007 ctrl = new ProjNode(call, TypeFunc::Control);
1008 phase->register_control(ctrl, loop, call);
1009 val = new ProjNode(call, TypeFunc::Parms);
1010 phase->register_new_node(val, call);
1011 val = new CheckCastPPNode(ctrl, val, obj_type);
1012 phase->register_new_node(val, ctrl);
1013 }
1014
1015 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) {
1016 Node* ctrl = phase->get_ctrl(barrier);
1017 Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
1018
1019 // Update the control of all nodes that should be after the
1020 // barrier control flow
1021 uses.clear();
1022 // Every node that is control dependent on the barrier's input
1023 // control will be after the expanded barrier. The raw memory (if
1024 // its memory is control dependent on the barrier's input control)
1025 // must stay above the barrier.
1026 uses_to_ignore.clear();
1027 if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
1028 uses_to_ignore.push(init_raw_mem);
1029 }
1030 for (uint next = 0; next < uses_to_ignore.size(); next++) {
1031 Node *n = uses_to_ignore.at(next);
1032 for (uint i = 0; i < n->req(); i++) {
1033 Node* in = n->in(i);
1034 if (in != nullptr && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
1035 uses_to_ignore.push(in);
1036 }
1037 }
1038 }
1039 for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
1040 Node* u = ctrl->fast_out(i);
1041 if (u->_idx < last &&
1042 u != barrier &&
1043 !u->depends_only_on_test() && // preserve dependency on test
1044 !uses_to_ignore.member(u) &&
1045 (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) &&
1046 (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) {
1047 Node* old_c = phase->ctrl_or_self(u);
1048 Node* c = old_c;
1049 if (c != ctrl ||
1050 is_dominator_same_ctrl(old_c, barrier, u, phase) ||
1051 ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) {
1052 phase->igvn().rehash_node_delayed(u);
1053 int nb = u->replace_edge(ctrl, region, &phase->igvn());
1054 if (u->is_CFG()) {
1055 if (phase->idom(u) == ctrl) {
1056 phase->set_idom(u, region, phase->dom_depth(region));
1057 }
1058 } else if (phase->get_ctrl(u) == ctrl) {
1059 assert(u != init_raw_mem, "should leave input raw mem above the barrier");
1060 uses.push(u);
1061 }
1062 assert(nb == 1, "more than 1 ctrl input?");
1063 --i, imax -= nb;
1064 }
1065 }
1066 }
1067 }
1068
1069 static Node* create_phis_on_call_return(Node* ctrl, Node* c, Node* n, Node* n_clone, const CallProjections& projs, PhaseIdealLoop* phase) {
1070 Node* region = nullptr;
1071 while (c != ctrl) {
1072 if (c->is_Region()) {
1073 region = c;
1074 }
1075 c = phase->idom(c);
1076 }
1077 assert(region != nullptr, "");
1078 Node* phi = new PhiNode(region, n->bottom_type());
1079 for (uint j = 1; j < region->req(); j++) {
1080 Node* in = region->in(j);
1081 if (phase->is_dominator(projs.fallthrough_catchproj, in)) {
1082 phi->init_req(j, n);
1083 } else if (phase->is_dominator(projs.catchall_catchproj, in)) {
1084 phi->init_req(j, n_clone);
1085 } else {
1086 phi->init_req(j, create_phis_on_call_return(ctrl, in, n, n_clone, projs, phase));
1087 }
1088 }
1089 phase->register_new_node(phi, region);
1090 return phi;
1091 }
1092
1093 void ShenandoahBarrierC2Support::pin_and_expand(PhaseIdealLoop* phase) {
1094 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
1095
1096 Unique_Node_List uses;
1097 Node_Stack stack(0);
1098 Node_List clones;
1099 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1100 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1101
1102 Node* ctrl = phase->get_ctrl(lrb);
1103 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1104
1105 CallStaticJavaNode* unc = nullptr;
1106 Node* unc_ctrl = nullptr;
1107 Node* uncasted_val = val;
1108
1109 for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1110 Node* u = lrb->fast_out(i);
1111 if (u->Opcode() == Op_CastPP &&
1112 u->in(0) != nullptr &&
1113 phase->is_dominator(u->in(0), ctrl)) {
1114 const Type* u_t = phase->igvn().type(u);
1115
1116 if (u_t->meet(TypePtr::NULL_PTR) != u_t &&
1117 u->in(0)->Opcode() == Op_IfTrue &&
1118 u->in(0)->as_Proj()->is_uncommon_trap_if_pattern() &&
1119 u->in(0)->in(0)->is_If() &&
1120 u->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
1121 u->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
1122 u->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
1123 u->in(0)->in(0)->in(1)->in(1)->in(1) == val &&
1124 u->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
1125 IdealLoopTree* loop = phase->get_loop(ctrl);
1126 IdealLoopTree* unc_loop = phase->get_loop(u->in(0));
1127
1128 if (!unc_loop->is_member(loop)) {
1129 continue;
1130 }
1131
1132 Node* branch = no_branches(ctrl, u->in(0), false, phase);
1133 assert(branch == nullptr || branch == NodeSentinel, "was not looking for a branch");
1134 if (branch == NodeSentinel) {
1135 continue;
1136 }
1137
1138 Node* iff = u->in(0)->in(0);
1139 Node* bol = iff->in(1)->clone();
1140 Node* cmp = bol->in(1)->clone();
1141 cmp->set_req(1, lrb);
1142 bol->set_req(1, cmp);
1143 phase->igvn().replace_input_of(iff, 1, bol);
1144 phase->set_ctrl(lrb, iff->in(0));
1145 phase->register_new_node(cmp, iff->in(0));
1146 phase->register_new_node(bol, iff->in(0));
1147 break;
1148 }
1149 }
1150 }
1151 // Load barrier on the control output of a call
1152 if ((ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) || ctrl->is_CallJava()) {
1153 CallJavaNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_CallJava() : ctrl->as_CallJava();
1154 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
1155 // The rethrow call may have too many projections to be
1156 // properly handled here. Given there's no reason for a
1157 // barrier to depend on the call, move it above the call
1158 stack.push(lrb, 0);
1159 do {
1160 Node* n = stack.node();
1161 uint idx = stack.index();
1162 if (idx < n->req()) {
1163 Node* in = n->in(idx);
1164 stack.set_index(idx+1);
1165 if (in != nullptr) {
1166 if (phase->has_ctrl(in)) {
1167 if (phase->is_dominator(call, phase->get_ctrl(in))) {
1168 #ifdef ASSERT
1169 for (uint i = 0; i < stack.size(); i++) {
1170 assert(stack.node_at(i) != in, "node shouldn't have been seen yet");
1171 }
1172 #endif
1173 stack.push(in, 0);
1174 }
1175 } else {
1176 assert(phase->is_dominator(in, call->in(0)), "no dependency on the call");
1177 }
1178 }
1179 } else {
1180 phase->set_ctrl(n, call->in(0));
1181 stack.pop();
1182 }
1183 } while(stack.size() > 0);
1184 continue;
1185 }
1186 CallProjections projs;
1187 call->extract_projections(&projs, false, false);
1188
1189 // If this is a runtime call, it doesn't have an exception handling path
1190 if (projs.fallthrough_catchproj == nullptr) {
1191 assert(call->method() == nullptr, "should be runtime call");
1192 assert(projs.catchall_catchproj == nullptr, "runtime call should not have catch all projection");
1193 continue;
1194 }
1195
1196 // Otherwise, clone the barrier so there's one for the fallthrough and one for the exception handling path
1197 #ifdef ASSERT
1198 VectorSet cloned;
1199 #endif
1200 Node* lrb_clone = lrb->clone();
1201 phase->register_new_node(lrb_clone, projs.catchall_catchproj);
1202 phase->set_ctrl(lrb, projs.fallthrough_catchproj);
1203
1204 stack.push(lrb, 0);
1205 clones.push(lrb_clone);
1206
1207 do {
1208 assert(stack.size() == clones.size(), "");
1209 Node* n = stack.node();
1210 #ifdef ASSERT
1211 if (n->is_Load()) {
1212 Node* mem = n->in(MemNode::Memory);
1213 for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
1214 Node* u = mem->fast_out(j);
1215 assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?");
1216 }
1217 }
1218 #endif
1219 uint idx = stack.index();
1220 Node* n_clone = clones.at(clones.size()-1);
1221 if (idx < n->outcnt()) {
1222 Node* u = n->raw_out(idx);
1223 Node* c = phase->ctrl_or_self(u);
1224 if (phase->is_dominator(call, c) && phase->is_dominator(c, projs.fallthrough_proj)) {
1225 stack.set_index(idx+1);
1226 assert(!u->is_CFG(), "");
1227 stack.push(u, 0);
1228 assert(!cloned.test_set(u->_idx), "only one clone");
1229 Node* u_clone = u->clone();
1230 int nb = u_clone->replace_edge(n, n_clone, &phase->igvn());
1231 assert(nb > 0, "should have replaced some uses");
1232 phase->register_new_node(u_clone, projs.catchall_catchproj);
1233 clones.push(u_clone);
1234 phase->set_ctrl(u, projs.fallthrough_catchproj);
1235 } else {
1236 bool replaced = false;
1237 if (u->is_Phi()) {
1238 for (uint k = 1; k < u->req(); k++) {
1239 if (u->in(k) == n) {
1240 if (phase->is_dominator(projs.catchall_catchproj, u->in(0)->in(k))) {
1241 phase->igvn().replace_input_of(u, k, n_clone);
1242 replaced = true;
1243 } else if (!phase->is_dominator(projs.fallthrough_catchproj, u->in(0)->in(k))) {
1244 phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase));
1245 replaced = true;
1246 }
1247 }
1248 }
1249 } else {
1250 if (phase->is_dominator(projs.catchall_catchproj, c)) {
1251 phase->igvn().rehash_node_delayed(u);
1252 int nb = u->replace_edge(n, n_clone, &phase->igvn());
1253 assert(nb > 0, "should have replaced some uses");
1254 replaced = true;
1255 } else if (!phase->is_dominator(projs.fallthrough_catchproj, c)) {
1256 if (u->is_If()) {
1257 // Can't break If/Bool/Cmp chain
1258 assert(n->is_Bool(), "unexpected If shape");
1259 assert(stack.node_at(stack.size()-2)->is_Cmp(), "unexpected If shape");
1260 assert(n_clone->is_Bool(), "unexpected clone");
1261 assert(clones.at(clones.size()-2)->is_Cmp(), "unexpected clone");
1262 Node* bol_clone = n->clone();
1263 Node* cmp_clone = stack.node_at(stack.size()-2)->clone();
1264 bol_clone->set_req(1, cmp_clone);
1265
1266 Node* nn = stack.node_at(stack.size()-3);
1267 Node* nn_clone = clones.at(clones.size()-3);
1268 assert(nn->Opcode() == nn_clone->Opcode(), "mismatch");
1269
1270 int nb = cmp_clone->replace_edge(nn, create_phis_on_call_return(ctrl, c, nn, nn_clone, projs, phase),
1271 &phase->igvn());
1272 assert(nb > 0, "should have replaced some uses");
1273
1274 phase->register_new_node(bol_clone, u->in(0));
1275 phase->register_new_node(cmp_clone, u->in(0));
1276
1277 phase->igvn().replace_input_of(u, 1, bol_clone);
1278
1279 } else {
1280 phase->igvn().rehash_node_delayed(u);
1281 int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase), &phase->igvn());
1282 assert(nb > 0, "should have replaced some uses");
1283 }
1284 replaced = true;
1285 }
1286 }
1287 if (!replaced) {
1288 stack.set_index(idx+1);
1289 }
1290 }
1291 } else {
1292 stack.pop();
1293 clones.pop();
1294 }
1295 } while (stack.size() > 0);
1296 assert(stack.size() == 0 && clones.size() == 0, "");
1297 }
1298 }
1299
1300 for (int i = 0; i < state->load_reference_barriers_count(); i++) {
1301 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1302 Node* ctrl = phase->get_ctrl(lrb);
1303 IdealLoopTree* loop = phase->get_loop(ctrl);
1304 Node* head = loop->head();
1305 if (head->is_OuterStripMinedLoop()) {
1306 // Expanding a barrier here will break loop strip mining
1307 // verification. Transform the loop so the loop nest doesn't
1308 // appear as strip mined.
1309 OuterStripMinedLoopNode* outer = head->as_OuterStripMinedLoop();
1310 hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1311 }
1312 if (head->is_BaseCountedLoop() && ctrl->is_IfProj() && ctrl->in(0)->is_BaseCountedLoopEnd() &&
1313 head->as_BaseCountedLoop()->loopexit() == ctrl->in(0)) {
1314 Node* entry = head->in(LoopNode::EntryControl);
1315 Node* backedge = head->in(LoopNode::LoopBackControl);
1316 Node* new_head = new LoopNode(entry, backedge);
1317 phase->register_control(new_head, phase->get_loop(entry), entry);
1318 phase->lazy_replace(head, new_head);
1319 }
1320 }
1321
1322 // Expand load-reference-barriers
1323 MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
1324 Unique_Node_List uses_to_ignore;
1325 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1326 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1327 uint last = phase->C->unique();
1328 Node* ctrl = phase->get_ctrl(lrb);
1329 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1330
1331 Node* orig_ctrl = ctrl;
1332
1333 Node* raw_mem = fixer.find_mem(ctrl, lrb);
1334 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, nullptr);
1335
1336 IdealLoopTree *loop = phase->get_loop(ctrl);
1337
1338 Node* heap_stable_ctrl = nullptr;
1339 Node* null_ctrl = nullptr;
1340
1341 assert(val->bottom_type()->make_oopptr(), "need oop");
1342 assert(val->bottom_type()->make_oopptr()->const_oop() == nullptr, "expect non-constant");
1343
1344 enum { _heap_stable = 1, _evac_path, _not_cset, PATH_LIMIT };
1345 Node* region = new RegionNode(PATH_LIMIT);
1346 Node* val_phi = new PhiNode(region, val->bottom_type()->is_oopptr());
1347
1348 // Stable path.
1349 int flags = ShenandoahHeap::HAS_FORWARDED;
1350 if (!ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1351 flags |= ShenandoahHeap::WEAK_ROOTS;
1352 }
1353 test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, flags);
1354 IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
1355
1356 // Heap stable case
1357 region->init_req(_heap_stable, heap_stable_ctrl);
1358 val_phi->init_req(_heap_stable, val);
1359
1360 // Test for in-cset, unless it's a native-LRB. Native LRBs need to return null
1361 // even for non-cset objects to prevent resurrection of such objects.
1362 // Wires !in_cset(obj) to slot 2 of region and phis
1363 Node* not_cset_ctrl = nullptr;
1364 if (ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1365 test_in_cset(ctrl, not_cset_ctrl, val, raw_mem, phase);
1366 }
1367 if (not_cset_ctrl != nullptr) {
1368 region->init_req(_not_cset, not_cset_ctrl);
1369 val_phi->init_req(_not_cset, val);
1370 } else {
1371 region->del_req(_not_cset);
1372 val_phi->del_req(_not_cset);
1373 }
1374
1375 // Resolve object when orig-value is in cset.
1376 // Make the unconditional resolve for fwdptr.
1377
1378 // Call lrb-stub and wire up that path in slots 4
1379 Node* result_mem = nullptr;
1380
1381 Node* addr;
1382 {
1383 VectorSet visited;
1384 addr = get_load_addr(phase, visited, lrb);
1385 }
1386 if (addr->Opcode() == Op_AddP) {
1387 Node* orig_base = addr->in(AddPNode::Base);
1388 Node* base = new CheckCastPPNode(ctrl, orig_base, orig_base->bottom_type(), ConstraintCastNode::StrongDependency);
1389 phase->register_new_node(base, ctrl);
1390 if (addr->in(AddPNode::Base) == addr->in((AddPNode::Address))) {
1391 // Field access
1392 addr = addr->clone();
1393 addr->set_req(AddPNode::Base, base);
1394 addr->set_req(AddPNode::Address, base);
1395 phase->register_new_node(addr, ctrl);
1396 } else {
1397 Node* addr2 = addr->in(AddPNode::Address);
1398 if (addr2->Opcode() == Op_AddP && addr2->in(AddPNode::Base) == addr2->in(AddPNode::Address) &&
1399 addr2->in(AddPNode::Base) == orig_base) {
1400 addr2 = addr2->clone();
1401 addr2->set_req(AddPNode::Base, base);
1402 addr2->set_req(AddPNode::Address, base);
1403 phase->register_new_node(addr2, ctrl);
1404 addr = addr->clone();
1405 addr->set_req(AddPNode::Base, base);
1406 addr->set_req(AddPNode::Address, addr2);
1407 phase->register_new_node(addr, ctrl);
1408 }
1409 }
1410 }
1411 call_lrb_stub(ctrl, val, addr, lrb->decorators(), phase);
1412 region->init_req(_evac_path, ctrl);
1413 val_phi->init_req(_evac_path, val);
1414
1415 phase->register_control(region, loop, heap_stable_iff);
1416 Node* out_val = val_phi;
1417 phase->register_new_node(val_phi, region);
1418
1419 fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase);
1420
1421 ctrl = orig_ctrl;
1422
1423 phase->igvn().replace_node(lrb, out_val);
1424
1425 follow_barrier_uses(out_val, ctrl, uses, phase);
1426
1427 for(uint next = 0; next < uses.size(); next++ ) {
1428 Node *n = uses.at(next);
1429 assert(phase->get_ctrl(n) == ctrl, "bad control");
1430 assert(n != raw_mem, "should leave input raw mem above the barrier");
1431 phase->set_ctrl(n, region);
1432 follow_barrier_uses(n, ctrl, uses, phase);
1433 }
1434 fixer.record_new_ctrl(ctrl, region, raw_mem, raw_mem_for_ctrl);
1435 }
1436 // Done expanding load-reference-barriers.
1437 assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced");
1438 }
1439
1440 Node* ShenandoahBarrierC2Support::get_load_addr(PhaseIdealLoop* phase, VectorSet& visited, Node* in) {
1441 if (visited.test_set(in->_idx)) {
1442 return nullptr;
1443 }
1444 switch (in->Opcode()) {
1445 case Op_Proj:
1446 return get_load_addr(phase, visited, in->in(0));
1447 case Op_CastPP:
1448 case Op_CheckCastPP:
1449 case Op_DecodeN:
1450 case Op_EncodeP:
1451 return get_load_addr(phase, visited, in->in(1));
1452 case Op_LoadN:
1453 case Op_LoadP:
1454 return in->in(MemNode::Address);
1455 case Op_CompareAndExchangeN:
1456 case Op_CompareAndExchangeP:
1457 case Op_GetAndSetN:
1458 case Op_GetAndSetP:
1459 case Op_ShenandoahCompareAndExchangeP:
1460 case Op_ShenandoahCompareAndExchangeN:
1461 // Those instructions would just have stored a different
1462 // value into the field. No use to attempt to fix it at this point.
1463 return phase->igvn().zerocon(T_OBJECT);
1464 case Op_CMoveP:
1465 case Op_CMoveN: {
1466 Node* t = get_load_addr(phase, visited, in->in(CMoveNode::IfTrue));
1467 Node* f = get_load_addr(phase, visited, in->in(CMoveNode::IfFalse));
1468 // Handle unambiguous cases: single address reported on both branches.
1469 if (t != nullptr && f == nullptr) return t;
1470 if (t == nullptr && f != nullptr) return f;
1471 if (t != nullptr && t == f) return t;
1472 // Ambiguity.
1473 return phase->igvn().zerocon(T_OBJECT);
1474 }
1475 case Op_Phi: {
1476 Node* addr = nullptr;
1477 for (uint i = 1; i < in->req(); i++) {
1478 Node* addr1 = get_load_addr(phase, visited, in->in(i));
1479 if (addr == nullptr) {
1480 addr = addr1;
1481 }
1482 if (addr != addr1) {
1483 return phase->igvn().zerocon(T_OBJECT);
1484 }
1485 }
1486 return addr;
1487 }
1488 case Op_ShenandoahLoadReferenceBarrier:
1489 return get_load_addr(phase, visited, in->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1490 case Op_CallDynamicJava:
1491 case Op_CallLeaf:
1492 case Op_CallStaticJava:
1493 case Op_ConN:
1494 case Op_ConP:
1495 case Op_Parm:
1496 case Op_CreateEx:
1497 return phase->igvn().zerocon(T_OBJECT);
1498 default:
1499 #ifdef ASSERT
1500 fatal("Unknown node in get_load_addr: %s", NodeClassNames[in->Opcode()]);
1501 #endif
1502 return phase->igvn().zerocon(T_OBJECT);
1503 }
1504
1505 }
1506
1507 void ShenandoahBarrierC2Support::move_gc_state_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
1508 IdealLoopTree *loop = phase->get_loop(iff);
1509 Node* loop_head = loop->_head;
1510 Node* entry_c = loop_head->in(LoopNode::EntryControl);
1511
1512 Node* bol = iff->in(1);
1513 Node* cmp = bol->in(1);
1514 Node* andi = cmp->in(1);
1515 Node* load = andi->in(1);
1516
1517 assert(is_gc_state_load(load), "broken");
1518 if (!phase->is_dominator(load->in(0), entry_c)) {
1519 Node* mem_ctrl = nullptr;
1520 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
1521 load = load->clone();
1522 load->set_req(MemNode::Memory, mem);
1523 load->set_req(0, entry_c);
1524 phase->register_new_node(load, entry_c);
1525 andi = andi->clone();
1526 andi->set_req(1, load);
1527 phase->register_new_node(andi, entry_c);
1528 cmp = cmp->clone();
1529 cmp->set_req(1, andi);
1530 phase->register_new_node(cmp, entry_c);
1531 bol = bol->clone();
1532 bol->set_req(1, cmp);
1533 phase->register_new_node(bol, entry_c);
1534
1535 phase->igvn().replace_input_of(iff, 1, bol);
1536 }
1537 }
1538
1539 bool ShenandoahBarrierC2Support::identical_backtoback_ifs(Node* n, PhaseIdealLoop* phase) {
1540 if (!n->is_If() || n->is_CountedLoopEnd()) {
1541 return false;
1542 }
1543 Node* region = n->in(0);
1544
1545 if (!region->is_Region()) {
1546 return false;
1547 }
1548 Node* dom = phase->idom(region);
1549 if (!dom->is_If()) {
1550 return false;
1551 }
1552
1553 if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) {
1554 return false;
1555 }
1556
1557 IfNode* dom_if = dom->as_If();
1558 Node* proj_true = dom_if->proj_out(1);
1559 Node* proj_false = dom_if->proj_out(0);
1560
1561 for (uint i = 1; i < region->req(); i++) {
1562 if (phase->is_dominator(proj_true, region->in(i))) {
1563 continue;
1564 }
1565 if (phase->is_dominator(proj_false, region->in(i))) {
1566 continue;
1567 }
1568 return false;
1569 }
1570
1571 return true;
1572 }
1573
1574 bool ShenandoahBarrierC2Support::merge_point_safe(Node* region) {
1575 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1576 Node* n = region->fast_out(i);
1577 if (n->is_LoadStore()) {
1578 // Splitting a LoadStore node through phi, causes it to lose its SCMemProj: the split if code doesn't have support
1579 // for a LoadStore at the region the if is split through because that's not expected to happen (LoadStore nodes
1580 // should be between barrier nodes). It does however happen with Shenandoah though because barriers can get
1581 // expanded around a LoadStore node.
1582 return false;
1583 }
1584 }
1585 return true;
1586 }
1587
1588
1589 void ShenandoahBarrierC2Support::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) {
1590 assert(is_heap_stable_test(n), "no other tests");
1591 if (identical_backtoback_ifs(n, phase)) {
1592 Node* n_ctrl = n->in(0);
1593 if (phase->can_split_if(n_ctrl) && merge_point_safe(n_ctrl)) {
1594 IfNode* dom_if = phase->idom(n_ctrl)->as_If();
1595 if (is_heap_stable_test(n)) {
1596 Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1);
1597 assert(is_gc_state_load(gc_state_load), "broken");
1598 Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1);
1599 assert(is_gc_state_load(dom_gc_state_load), "broken");
1600 if (gc_state_load != dom_gc_state_load) {
1601 phase->igvn().replace_node(gc_state_load, dom_gc_state_load);
1602 }
1603 }
1604 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1605 Node* proj_true = dom_if->proj_out(1);
1606 Node* proj_false = dom_if->proj_out(0);
1607 Node* con_true = phase->igvn().makecon(TypeInt::ONE);
1608 Node* con_false = phase->igvn().makecon(TypeInt::ZERO);
1609
1610 for (uint i = 1; i < n_ctrl->req(); i++) {
1611 if (phase->is_dominator(proj_true, n_ctrl->in(i))) {
1612 bolphi->init_req(i, con_true);
1613 } else {
1614 assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1615 bolphi->init_req(i, con_false);
1616 }
1617 }
1618 phase->register_new_node(bolphi, n_ctrl);
1619 phase->igvn().replace_input_of(n, 1, bolphi);
1620 phase->do_split_if(n);
1621 }
1622 }
1623 }
1624
1625 IfNode* ShenandoahBarrierC2Support::find_unswitching_candidate(const IdealLoopTree* loop, PhaseIdealLoop* phase) {
1626 // Find first invariant test that doesn't exit the loop
1627 LoopNode *head = loop->_head->as_Loop();
1628 IfNode* unswitch_iff = nullptr;
1629 Node* n = head->in(LoopNode::LoopBackControl);
1630 int loop_has_sfpts = -1;
1631 while (n != head) {
1632 Node* n_dom = phase->idom(n);
1633 if (n->is_Region()) {
1634 if (n_dom->is_If()) {
1635 IfNode* iff = n_dom->as_If();
1636 if (iff->in(1)->is_Bool()) {
1637 BoolNode* bol = iff->in(1)->as_Bool();
1638 if (bol->in(1)->is_Cmp()) {
1639 // If condition is invariant and not a loop exit,
1640 // then found reason to unswitch.
1641 if (is_heap_stable_test(iff) &&
1642 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) {
1643 assert(!loop->is_loop_exit(iff), "both branches should be in the loop");
1644 if (loop_has_sfpts == -1) {
1645 for(uint i = 0; i < loop->_body.size(); i++) {
1646 Node *m = loop->_body[i];
1647 if (m->is_SafePoint() && !m->is_CallLeaf()) {
1648 loop_has_sfpts = 1;
1649 break;
1650 }
1651 }
1652 if (loop_has_sfpts == -1) {
1653 loop_has_sfpts = 0;
1654 }
1655 }
1656 if (!loop_has_sfpts) {
1657 unswitch_iff = iff;
1658 }
1659 }
1660 }
1661 }
1662 }
1663 }
1664 n = n_dom;
1665 }
1666 return unswitch_iff;
1667 }
1668
1669
1670 void ShenandoahBarrierC2Support::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) {
1671 Node_List heap_stable_tests;
1672 stack.push(phase->C->start(), 0);
1673 do {
1674 Node* n = stack.node();
1675 uint i = stack.index();
1676
1677 if (i < n->outcnt()) {
1678 Node* u = n->raw_out(i);
1679 stack.set_index(i+1);
1680 if (!visited.test_set(u->_idx)) {
1681 stack.push(u, 0);
1682 }
1683 } else {
1684 stack.pop();
1685 if (n->is_If() && is_heap_stable_test(n)) {
1686 heap_stable_tests.push(n);
1687 }
1688 }
1689 } while (stack.size() > 0);
1690
1691 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1692 Node* n = heap_stable_tests.at(i);
1693 assert(is_heap_stable_test(n), "only evacuation test");
1694 merge_back_to_back_tests(n, phase);
1695 }
1696
1697 if (!phase->C->major_progress()) {
1698 VectorSet seen;
1699 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1700 Node* n = heap_stable_tests.at(i);
1701 IdealLoopTree* loop = phase->get_loop(n);
1702 if (loop != phase->ltree_root() &&
1703 loop->_child == nullptr &&
1704 !loop->_irreducible) {
1705 Node* head = loop->_head;
1706 if (head->is_Loop() &&
1707 (!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) &&
1708 !seen.test_set(head->_idx)) {
1709 IfNode* iff = find_unswitching_candidate(loop, phase);
1710 if (iff != nullptr) {
1711 Node* bol = iff->in(1);
1712 if (head->as_Loop()->is_strip_mined()) {
1713 head->as_Loop()->verify_strip_mined(0);
1714 }
1715 move_gc_state_test_out_of_loop(iff, phase);
1716
1717 AutoNodeBudget node_budget(phase);
1718
1719 if (loop->policy_unswitching(phase)) {
1720 if (head->as_Loop()->is_strip_mined()) {
1721 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop();
1722 hide_strip_mined_loop(outer, head->as_CountedLoop(), phase);
1723 }
1724 phase->do_unswitching(loop, old_new);
1725 } else {
1726 // Not proceeding with unswitching. Move load back in
1727 // the loop.
1728 phase->igvn().replace_input_of(iff, 1, bol);
1729 }
1730 }
1731 }
1732 }
1733 }
1734 }
1735 }
1736
1737 #ifdef ASSERT
1738 static bool has_never_branch(Node* root) {
1739 for (uint i = 1; i < root->req(); i++) {
1740 Node* in = root->in(i);
1741 if (in != nullptr && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->is_NeverBranch()) {
1742 return true;
1743 }
1744 }
1745 return false;
1746 }
1747 #endif
1748
1749 void MemoryGraphFixer::collect_memory_nodes() {
1750 Node_Stack stack(0);
1751 VectorSet visited;
1752 Node_List regions;
1753
1754 // Walk the raw memory graph and create a mapping from CFG node to
1755 // memory node. Exclude phis for now.
1756 stack.push(_phase->C->root(), 1);
1757 do {
1758 Node* n = stack.node();
1759 int opc = n->Opcode();
1760 uint i = stack.index();
1761 if (i < n->req()) {
1762 Node* mem = nullptr;
1763 if (opc == Op_Root) {
1764 Node* in = n->in(i);
1765 int in_opc = in->Opcode();
1766 if (in_opc == Op_Return || in_opc == Op_Rethrow) {
1767 mem = in->in(TypeFunc::Memory);
1768 } else if (in_opc == Op_Halt) {
1769 if (in->in(0)->is_Region()) {
1770 Node* r = in->in(0);
1771 for (uint j = 1; j < r->req(); j++) {
1772 assert(!r->in(j)->is_NeverBranch(), "");
1773 }
1774 } else {
1775 Node* proj = in->in(0);
1776 assert(proj->is_Proj(), "");
1777 Node* in = proj->in(0);
1778 assert(in->is_CallStaticJava() || in->is_NeverBranch() || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
1779 if (in->is_CallStaticJava()) {
1780 mem = in->in(TypeFunc::Memory);
1781 } else if (in->Opcode() == Op_Catch) {
1782 Node* call = in->in(0)->in(0);
1783 assert(call->is_Call(), "");
1784 mem = call->in(TypeFunc::Memory);
1785 } else if (in->is_NeverBranch()) {
1786 mem = collect_memory_for_infinite_loop(in);
1787 }
1788 }
1789 } else {
1790 #ifdef ASSERT
1791 n->dump();
1792 in->dump();
1793 #endif
1794 ShouldNotReachHere();
1795 }
1796 } else {
1797 assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
1798 assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
1799 mem = n->in(i);
1800 }
1801 i++;
1802 stack.set_index(i);
1803 if (mem == nullptr) {
1804 continue;
1805 }
1806 for (;;) {
1807 if (visited.test_set(mem->_idx) || mem->is_Start()) {
1808 break;
1809 }
1810 if (mem->is_Phi()) {
1811 stack.push(mem, 2);
1812 mem = mem->in(1);
1813 } else if (mem->is_Proj()) {
1814 stack.push(mem, mem->req());
1815 mem = mem->in(0);
1816 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
1817 mem = mem->in(TypeFunc::Memory);
1818 } else if (mem->is_MergeMem()) {
1819 MergeMemNode* mm = mem->as_MergeMem();
1820 mem = mm->memory_at(_alias);
1821 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
1822 assert(_alias == Compile::AliasIdxRaw, "");
1823 stack.push(mem, mem->req());
1824 mem = mem->in(MemNode::Memory);
1825 } else {
1826 #ifdef ASSERT
1827 mem->dump();
1828 #endif
1829 ShouldNotReachHere();
1830 }
1831 }
1832 } else {
1833 if (n->is_Phi()) {
1834 // Nothing
1835 } else if (!n->is_Root()) {
1836 Node* c = get_ctrl(n);
1837 _memory_nodes.map(c->_idx, n);
1838 }
1839 stack.pop();
1840 }
1841 } while(stack.is_nonempty());
1842
1843 // Iterate over CFG nodes in rpo and propagate memory state to
1844 // compute memory state at regions, creating new phis if needed.
1845 Node_List rpo_list;
1846 visited.clear();
1847 _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
1848 Node* root = rpo_list.pop();
1849 assert(root == _phase->C->root(), "");
1850
1851 const bool trace = false;
1852 #ifdef ASSERT
1853 if (trace) {
1854 for (int i = rpo_list.size() - 1; i >= 0; i--) {
1855 Node* c = rpo_list.at(i);
1856 if (_memory_nodes[c->_idx] != nullptr) {
1857 tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump();
1858 }
1859 }
1860 }
1861 #endif
1862 uint last = _phase->C->unique();
1863
1864 #ifdef ASSERT
1865 uint16_t max_depth = 0;
1866 for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
1867 IdealLoopTree* lpt = iter.current();
1868 max_depth = MAX2(max_depth, lpt->_nest);
1869 }
1870 #endif
1871
1872 bool progress = true;
1873 int iteration = 0;
1874 Node_List dead_phis;
1875 while (progress) {
1876 progress = false;
1877 iteration++;
1878 assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
1879 if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
1880
1881 for (int i = rpo_list.size() - 1; i >= 0; i--) {
1882 Node* c = rpo_list.at(i);
1883
1884 Node* prev_mem = _memory_nodes[c->_idx];
1885 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
1886 Node* prev_region = regions[c->_idx];
1887 Node* unique = nullptr;
1888 for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
1889 Node* m = _memory_nodes[c->in(j)->_idx];
1890 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");
1891 if (m != nullptr) {
1892 if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
1893 assert((c->is_Loop() && j == LoopNode::LoopBackControl) || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
1894 // continue
1895 } else if (unique == nullptr) {
1896 unique = m;
1897 } else if (m == unique) {
1898 // continue
1899 } else {
1900 unique = NodeSentinel;
1901 }
1902 }
1903 }
1904 assert(unique != nullptr, "empty phi???");
1905 if (unique != NodeSentinel) {
1906 if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c) {
1907 dead_phis.push(prev_region);
1908 }
1909 regions.map(c->_idx, unique);
1910 } else {
1911 Node* phi = nullptr;
1912 if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
1913 phi = prev_region;
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->set_req(k, m);
1918 }
1919 } else {
1920 for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == nullptr; j++) {
1921 Node* u = c->fast_out(j);
1922 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
1923 (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
1924 phi = u;
1925 for (uint k = 1; k < c->req() && phi != nullptr; k++) {
1926 Node* m = _memory_nodes[c->in(k)->_idx];
1927 assert(m != nullptr, "expect memory state");
1928 if (u->in(k) != m) {
1929 phi = NodeSentinel;
1930 }
1931 }
1932 }
1933 }
1934 if (phi == NodeSentinel) {
1935 phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
1936 for (uint k = 1; k < c->req(); k++) {
1937 Node* m = _memory_nodes[c->in(k)->_idx];
1938 assert(m != nullptr, "expect memory state");
1939 phi->init_req(k, m);
1940 }
1941 }
1942 }
1943 if (phi != nullptr) {
1944 regions.map(c->_idx, phi);
1945 } else {
1946 assert(c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
1947 }
1948 }
1949 Node* current_region = regions[c->_idx];
1950 if (current_region != prev_region) {
1951 progress = true;
1952 if (prev_region == prev_mem) {
1953 _memory_nodes.map(c->_idx, current_region);
1954 }
1955 }
1956 } else if (prev_mem == nullptr || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
1957 Node* m = _memory_nodes[_phase->idom(c)->_idx];
1958 assert(m != nullptr || c->Opcode() == Op_Halt, "expect memory state");
1959 if (m != prev_mem) {
1960 _memory_nodes.map(c->_idx, m);
1961 progress = true;
1962 }
1963 }
1964 #ifdef ASSERT
1965 if (trace) { tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); }
1966 #endif
1967 }
1968 }
1969
1970 // Replace existing phi with computed memory state for that region
1971 // if different (could be a new phi or a dominating memory node if
1972 // that phi was found to be useless).
1973 while (dead_phis.size() > 0) {
1974 Node* n = dead_phis.pop();
1975 n->replace_by(_phase->C->top());
1976 n->destruct(&_phase->igvn());
1977 }
1978 for (int i = rpo_list.size() - 1; i >= 0; i--) {
1979 Node* c = rpo_list.at(i);
1980 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
1981 Node* n = regions[c->_idx];
1982 assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
1983 if (n != nullptr && n->is_Phi() && n->_idx >= last && n->in(0) == c) {
1984 _phase->register_new_node(n, c);
1985 }
1986 }
1987 }
1988 for (int i = rpo_list.size() - 1; i >= 0; i--) {
1989 Node* c = rpo_list.at(i);
1990 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
1991 Node* n = regions[c->_idx];
1992 assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
1993 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
1994 Node* u = c->fast_out(i);
1995 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
1996 u != n) {
1997 assert(c->unique_ctrl_out()->Opcode() != Op_Halt, "expected memory state");
1998 if (u->adr_type() == TypePtr::BOTTOM) {
1999 fix_memory_uses(u, n, n, c);
2000 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2001 _phase->lazy_replace(u, n);
2002 --i; --imax;
2003 }
2004 }
2005 }
2006 }
2007 }
2008 }
2009
2010 Node* MemoryGraphFixer::collect_memory_for_infinite_loop(const Node* in) {
2011 Node* mem = nullptr;
2012 Node* head = in->in(0);
2013 assert(head->is_Region(), "unexpected infinite loop graph shape");
2014
2015 Node* phi_mem = nullptr;
2016 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2017 Node* u = head->fast_out(j);
2018 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
2019 if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2020 assert(phi_mem == nullptr || phi_mem->adr_type() == TypePtr::BOTTOM, "");
2021 phi_mem = u;
2022 } else if (u->adr_type() == TypePtr::BOTTOM) {
2023 assert(phi_mem == nullptr || _phase->C->get_alias_index(phi_mem->adr_type()) == _alias, "");
2024 if (phi_mem == nullptr) {
2025 phi_mem = u;
2026 }
2027 }
2028 }
2029 }
2030 if (phi_mem == nullptr) {
2031 ResourceMark rm;
2032 Node_Stack stack(0);
2033 stack.push(head, 1);
2034 do {
2035 Node* n = stack.node();
2036 uint i = stack.index();
2037 if (i >= n->req()) {
2038 stack.pop();
2039 } else {
2040 stack.set_index(i + 1);
2041 Node* c = n->in(i);
2042 assert(c != head, "should have found a safepoint on the way");
2043 if (stack.size() != 1 || _phase->is_dominator(head, c)) {
2044 for (;;) {
2045 if (c->is_Region()) {
2046 stack.push(c, 1);
2047 break;
2048 } else if (c->is_SafePoint() && !c->is_CallLeaf()) {
2049 Node* m = c->in(TypeFunc::Memory);
2050 if (m->is_MergeMem()) {
2051 m = m->as_MergeMem()->memory_at(_alias);
2052 }
2053 assert(mem == nullptr || mem == m, "several memory states");
2054 mem = m;
2055 break;
2056 } else {
2057 assert(c != c->in(0), "");
2058 c = c->in(0);
2059 }
2060 }
2061 }
2062 }
2063 } while (stack.size() > 0);
2064 assert(mem != nullptr, "should have found safepoint");
2065 } else {
2066 mem = phi_mem;
2067 }
2068 return mem;
2069 }
2070
2071 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
2072 Node* c = _phase->get_ctrl(n);
2073 if (n->is_Proj() && n->in(0) != nullptr && n->in(0)->is_Call()) {
2074 assert(c == n->in(0), "");
2075 CallNode* call = c->as_Call();
2076 CallProjections projs;
2077 call->extract_projections(&projs, true, false);
2078 if (projs.catchall_memproj != nullptr) {
2079 if (projs.fallthrough_memproj == n) {
2080 c = projs.fallthrough_catchproj;
2081 } else {
2082 assert(projs.catchall_memproj == n, "");
2083 c = projs.catchall_catchproj;
2084 }
2085 }
2086 }
2087 return c;
2088 }
2089
2090 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
2091 if (_phase->has_ctrl(n))
2092 return get_ctrl(n);
2093 else {
2094 assert (n->is_CFG(), "must be a CFG node");
2095 return n;
2096 }
2097 }
2098
2099 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
2100 return m != nullptr && get_ctrl(m) == c;
2101 }
2102
2103 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
2104 assert(n == nullptr || _phase->ctrl_or_self(n) == ctrl, "");
2105 assert(!ctrl->is_Call() || ctrl == n, "projection expected");
2106 #ifdef ASSERT
2107 if ((ctrl->is_Proj() && ctrl->in(0)->is_Call()) ||
2108 (ctrl->is_Catch() && ctrl->in(0)->in(0)->is_Call())) {
2109 CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_Call() : ctrl->in(0)->in(0)->as_Call();
2110 int mems = 0;
2111 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
2112 Node* u = call->fast_out(i);
2113 if (u->bottom_type() == Type::MEMORY) {
2114 mems++;
2115 }
2116 }
2117 assert(mems <= 1, "No node right after call if multiple mem projections");
2118 }
2119 #endif
2120 Node* mem = _memory_nodes[ctrl->_idx];
2121 Node* c = ctrl;
2122 while (!mem_is_valid(mem, c) &&
2123 (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
2124 c = _phase->idom(c);
2125 mem = _memory_nodes[c->_idx];
2126 }
2127 if (n != nullptr && mem_is_valid(mem, c)) {
2128 while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
2129 mem = next_mem(mem, _alias);
2130 }
2131 if (mem->is_MergeMem()) {
2132 mem = mem->as_MergeMem()->memory_at(_alias);
2133 }
2134 if (!mem_is_valid(mem, c)) {
2135 do {
2136 c = _phase->idom(c);
2137 mem = _memory_nodes[c->_idx];
2138 } while (!mem_is_valid(mem, c) &&
2139 (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
2140 }
2141 }
2142 assert(mem->bottom_type() == Type::MEMORY, "");
2143 return mem;
2144 }
2145
2146 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
2147 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2148 Node* use = region->fast_out(i);
2149 if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
2150 (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
2151 return true;
2152 }
2153 }
2154 return false;
2155 }
2156
2157 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
2158 assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
2159 const bool trace = false;
2160 DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
2161 DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
2162 GrowableArray<Node*> phis;
2163 if (mem_for_ctrl != mem) {
2164 Node* old = mem_for_ctrl;
2165 Node* prev = nullptr;
2166 while (old != mem) {
2167 prev = old;
2168 if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
2169 assert(_alias == Compile::AliasIdxRaw, "");
2170 old = old->in(MemNode::Memory);
2171 } else if (old->Opcode() == Op_SCMemProj) {
2172 assert(_alias == Compile::AliasIdxRaw, "");
2173 old = old->in(0);
2174 } else {
2175 ShouldNotReachHere();
2176 }
2177 }
2178 assert(prev != nullptr, "");
2179 if (new_ctrl != ctrl) {
2180 _memory_nodes.map(ctrl->_idx, mem);
2181 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2182 }
2183 uint input = (uint)MemNode::Memory;
2184 _phase->igvn().replace_input_of(prev, input, new_mem);
2185 } else {
2186 uses.clear();
2187 _memory_nodes.map(new_ctrl->_idx, new_mem);
2188 uses.push(new_ctrl);
2189 for(uint next = 0; next < uses.size(); next++ ) {
2190 Node *n = uses.at(next);
2191 assert(n->is_CFG(), "");
2192 DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
2193 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2194 Node* u = n->fast_out(i);
2195 if (!u->is_Root() && u->is_CFG() && u != n) {
2196 Node* m = _memory_nodes[u->_idx];
2197 if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
2198 !has_mem_phi(u) &&
2199 u->unique_ctrl_out()->Opcode() != Op_Halt) {
2200 DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
2201 DEBUG_ONLY(if (trace && m != nullptr) { tty->print("ZZZ mem"); m->dump(); });
2202
2203 if (!mem_is_valid(m, u) || !m->is_Phi()) {
2204 bool push = true;
2205 bool create_phi = true;
2206 if (_phase->is_dominator(new_ctrl, u)) {
2207 create_phi = false;
2208 }
2209 if (create_phi) {
2210 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
2211 _phase->register_new_node(phi, u);
2212 phis.push(phi);
2213 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
2214 if (!mem_is_valid(m, u)) {
2215 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
2216 _memory_nodes.map(u->_idx, phi);
2217 } else {
2218 DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
2219 for (;;) {
2220 assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), "");
2221 Node* next = nullptr;
2222 if (m->is_Proj()) {
2223 next = m->in(0);
2224 } else {
2225 assert(m->is_Mem() || m->is_LoadStore(), "");
2226 assert(_alias == Compile::AliasIdxRaw, "");
2227 next = m->in(MemNode::Memory);
2228 }
2229 if (_phase->get_ctrl(next) != u) {
2230 break;
2231 }
2232 if (next->is_MergeMem()) {
2233 assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
2234 break;
2235 }
2236 if (next->is_Phi()) {
2237 assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
2238 break;
2239 }
2240 m = next;
2241 }
2242
2243 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
2244 assert(m->is_Mem() || m->is_LoadStore(), "");
2245 uint input = (uint)MemNode::Memory;
2246 _phase->igvn().replace_input_of(m, input, phi);
2247 push = false;
2248 }
2249 } else {
2250 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
2251 }
2252 if (push) {
2253 uses.push(u);
2254 }
2255 }
2256 } else if (!mem_is_valid(m, u) &&
2257 !(u->Opcode() == Op_CProj && u->in(0)->is_NeverBranch() && u->as_Proj()->_con == 1)) {
2258 uses.push(u);
2259 }
2260 }
2261 }
2262 }
2263 for (int i = 0; i < phis.length(); i++) {
2264 Node* n = phis.at(i);
2265 Node* r = n->in(0);
2266 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
2267 for (uint j = 1; j < n->req(); j++) {
2268 Node* m = find_mem(r->in(j), nullptr);
2269 _phase->igvn().replace_input_of(n, j, m);
2270 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
2271 }
2272 }
2273 }
2274 uint last = _phase->C->unique();
2275 MergeMemNode* mm = nullptr;
2276 int alias = _alias;
2277 DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
2278 // Process loads first to not miss an anti-dependency: if the memory
2279 // edge of a store is updated before a load is processed then an
2280 // anti-dependency may be missed.
2281 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2282 Node* u = mem->out(i);
2283 if (u->_idx < last && u->is_Load() && _phase->C->get_alias_index(u->adr_type()) == alias) {
2284 Node* m = find_mem(_phase->get_ctrl(u), u);
2285 if (m != mem) {
2286 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2287 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2288 --i;
2289 }
2290 }
2291 }
2292 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2293 Node* u = mem->out(i);
2294 if (u->_idx < last) {
2295 if (u->is_Mem()) {
2296 if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2297 Node* m = find_mem(_phase->get_ctrl(u), u);
2298 if (m != mem) {
2299 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2300 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2301 --i;
2302 }
2303 }
2304 } else if (u->is_MergeMem()) {
2305 MergeMemNode* u_mm = u->as_MergeMem();
2306 if (u_mm->memory_at(alias) == mem) {
2307 MergeMemNode* newmm = nullptr;
2308 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2309 Node* uu = u->fast_out(j);
2310 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2311 if (uu->is_Phi()) {
2312 assert(uu->adr_type() == TypePtr::BOTTOM, "");
2313 Node* region = uu->in(0);
2314 int nb = 0;
2315 for (uint k = 1; k < uu->req(); k++) {
2316 if (uu->in(k) == u) {
2317 Node* m = find_mem(region->in(k), nullptr);
2318 if (m != mem) {
2319 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
2320 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2321 if (newmm != u) {
2322 _phase->igvn().replace_input_of(uu, k, newmm);
2323 nb++;
2324 --jmax;
2325 }
2326 }
2327 }
2328 }
2329 if (nb > 0) {
2330 --j;
2331 }
2332 } else {
2333 Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
2334 if (m != mem) {
2335 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
2336 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2337 if (newmm != u) {
2338 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2339 --j, --jmax;
2340 }
2341 }
2342 }
2343 }
2344 }
2345 } else if (u->is_Phi()) {
2346 assert(u->bottom_type() == Type::MEMORY, "what else?");
2347 if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
2348 Node* region = u->in(0);
2349 bool replaced = false;
2350 for (uint j = 1; j < u->req(); j++) {
2351 if (u->in(j) == mem) {
2352 Node* m = find_mem(region->in(j), nullptr);
2353 Node* nnew = m;
2354 if (m != mem) {
2355 if (u->adr_type() == TypePtr::BOTTOM) {
2356 mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
2357 nnew = mm;
2358 }
2359 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
2360 _phase->igvn().replace_input_of(u, j, nnew);
2361 replaced = true;
2362 }
2363 }
2364 }
2365 if (replaced) {
2366 --i;
2367 }
2368 }
2369 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2370 u->adr_type() == nullptr) {
2371 assert(u->adr_type() != nullptr ||
2372 u->Opcode() == Op_Rethrow ||
2373 u->Opcode() == Op_Return ||
2374 u->Opcode() == Op_SafePoint ||
2375 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2376 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2377 u->Opcode() == Op_CallLeaf, "");
2378 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2379 if (m != mem) {
2380 mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
2381 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2382 --i;
2383 }
2384 } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2385 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2386 if (m != mem) {
2387 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2388 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2389 --i;
2390 }
2391 } else if (u->adr_type() != TypePtr::BOTTOM &&
2392 _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
2393 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2394 assert(m != mem, "");
2395 // u is on the wrong slice...
2396 assert(u->is_ClearArray(), "");
2397 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2398 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2399 --i;
2400 }
2401 }
2402 }
2403 #ifdef ASSERT
2404 assert(new_mem->outcnt() > 0, "");
2405 for (int i = 0; i < phis.length(); i++) {
2406 Node* n = phis.at(i);
2407 assert(n->outcnt() > 0, "new phi must have uses now");
2408 }
2409 #endif
2410 }
2411
2412 void MemoryGraphFixer::record_new_ctrl(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl) {
2413 if (mem_for_ctrl != mem && new_ctrl != ctrl) {
2414 _memory_nodes.map(ctrl->_idx, mem);
2415 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2416 }
2417 }
2418
2419 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
2420 MergeMemNode* mm = MergeMemNode::make(mem);
2421 mm->set_memory_at(_alias, rep_proj);
2422 _phase->register_new_node(mm, rep_ctrl);
2423 return mm;
2424 }
2425
2426 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
2427 MergeMemNode* newmm = nullptr;
2428 MergeMemNode* u_mm = u->as_MergeMem();
2429 Node* c = _phase->get_ctrl(u);
2430 if (_phase->is_dominator(c, rep_ctrl)) {
2431 c = rep_ctrl;
2432 } else {
2433 assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
2434 }
2435 if (u->outcnt() == 1) {
2436 if (u->req() > (uint)_alias && u->in(_alias) == mem) {
2437 _phase->igvn().replace_input_of(u, _alias, rep_proj);
2438 --i;
2439 } else {
2440 _phase->igvn().rehash_node_delayed(u);
2441 u_mm->set_memory_at(_alias, rep_proj);
2442 }
2443 newmm = u_mm;
2444 _phase->set_ctrl_and_loop(u, c);
2445 } else {
2446 // can't simply clone u and then change one of its input because
2447 // it adds and then removes an edge which messes with the
2448 // DUIterator
2449 newmm = MergeMemNode::make(u_mm->base_memory());
2450 for (uint j = 0; j < u->req(); j++) {
2451 if (j < newmm->req()) {
2452 if (j == (uint)_alias) {
2453 newmm->set_req(j, rep_proj);
2454 } else if (newmm->in(j) != u->in(j)) {
2455 newmm->set_req(j, u->in(j));
2456 }
2457 } else if (j == (uint)_alias) {
2458 newmm->add_req(rep_proj);
2459 } else {
2460 newmm->add_req(u->in(j));
2461 }
2462 }
2463 if ((uint)_alias >= u->req()) {
2464 newmm->set_memory_at(_alias, rep_proj);
2465 }
2466 _phase->register_new_node(newmm, c);
2467 }
2468 return newmm;
2469 }
2470
2471 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
2472 if (phi->adr_type() == TypePtr::BOTTOM) {
2473 Node* region = phi->in(0);
2474 for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
2475 Node* uu = region->fast_out(j);
2476 if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
2477 return false;
2478 }
2479 }
2480 return true;
2481 }
2482 return _phase->C->get_alias_index(phi->adr_type()) == _alias;
2483 }
2484
2485 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
2486 uint last = _phase-> C->unique();
2487 MergeMemNode* mm = nullptr;
2488 assert(mem->bottom_type() == Type::MEMORY, "");
2489 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2490 Node* u = mem->out(i);
2491 if (u != replacement && u->_idx < last) {
2492 if (u->is_MergeMem()) {
2493 MergeMemNode* u_mm = u->as_MergeMem();
2494 if (u_mm->memory_at(_alias) == mem) {
2495 MergeMemNode* newmm = nullptr;
2496 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2497 Node* uu = u->fast_out(j);
2498 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2499 if (uu->is_Phi()) {
2500 if (should_process_phi(uu)) {
2501 Node* region = uu->in(0);
2502 int nb = 0;
2503 for (uint k = 1; k < uu->req(); k++) {
2504 if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
2505 if (newmm == nullptr) {
2506 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2507 }
2508 if (newmm != u) {
2509 _phase->igvn().replace_input_of(uu, k, newmm);
2510 nb++;
2511 --jmax;
2512 }
2513 }
2514 }
2515 if (nb > 0) {
2516 --j;
2517 }
2518 }
2519 } else {
2520 if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
2521 if (newmm == nullptr) {
2522 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2523 }
2524 if (newmm != u) {
2525 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2526 --j, --jmax;
2527 }
2528 }
2529 }
2530 }
2531 }
2532 } else if (u->is_Phi()) {
2533 assert(u->bottom_type() == Type::MEMORY, "what else?");
2534 Node* region = u->in(0);
2535 if (should_process_phi(u)) {
2536 bool replaced = false;
2537 for (uint j = 1; j < u->req(); j++) {
2538 if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
2539 Node* nnew = rep_proj;
2540 if (u->adr_type() == TypePtr::BOTTOM) {
2541 if (mm == nullptr) {
2542 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2543 }
2544 nnew = mm;
2545 }
2546 _phase->igvn().replace_input_of(u, j, nnew);
2547 replaced = true;
2548 }
2549 }
2550 if (replaced) {
2551 --i;
2552 }
2553
2554 }
2555 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2556 u->adr_type() == nullptr) {
2557 assert(u->adr_type() != nullptr ||
2558 u->Opcode() == Op_Rethrow ||
2559 u->Opcode() == Op_Return ||
2560 u->Opcode() == Op_SafePoint ||
2561 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2562 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2563 u->Opcode() == Op_CallLeaf, "%s", u->Name());
2564 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2565 if (mm == nullptr) {
2566 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2567 }
2568 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2569 --i;
2570 }
2571 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2572 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2573 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
2574 --i;
2575 }
2576 }
2577 }
2578 }
2579 }
2580
2581 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj, DecoratorSet decorators)
2582 : Node(ctrl, obj), _decorators(decorators) {
2583 ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this);
2584 }
2585
2586 DecoratorSet ShenandoahLoadReferenceBarrierNode::decorators() const {
2587 return _decorators;
2588 }
2589
2590 uint ShenandoahLoadReferenceBarrierNode::size_of() const {
2591 return sizeof(*this);
2592 }
2593
2594 static DecoratorSet mask_decorators(DecoratorSet decorators) {
2595 return decorators & (ON_STRONG_OOP_REF | ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF | ON_UNKNOWN_OOP_REF | IN_NATIVE);
2596 }
2597
2598 uint ShenandoahLoadReferenceBarrierNode::hash() const {
2599 uint hash = Node::hash();
2600 hash += mask_decorators(_decorators);
2601 return hash;
2602 }
2603
2604 bool ShenandoahLoadReferenceBarrierNode::cmp( const Node &n ) const {
2605 return Node::cmp(n) && n.Opcode() == Op_ShenandoahLoadReferenceBarrier &&
2606 mask_decorators(_decorators) == mask_decorators(((const ShenandoahLoadReferenceBarrierNode&)n)._decorators);
2607 }
2608
2609 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const {
2610 if (in(ValueIn) == nullptr || in(ValueIn)->is_top()) {
2611 return Type::TOP;
2612 }
2613 const Type* t = in(ValueIn)->bottom_type();
2614 if (t == TypePtr::NULL_PTR) {
2615 return t;
2616 }
2617
2618 if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2619 return t;
2620 }
2621
2622 return t->meet(TypePtr::NULL_PTR);
2623 }
2624
2625 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const {
2626 // Either input is TOP ==> the result is TOP
2627 const Type *t2 = phase->type(in(ValueIn));
2628 if( t2 == Type::TOP ) return Type::TOP;
2629
2630 if (t2 == TypePtr::NULL_PTR) {
2631 return t2;
2632 }
2633
2634 if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2635 return t2;
2636 }
2637
2638 return t2->meet(TypePtr::NULL_PTR);
2639 }
2640
2641 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) {
2642 Node* value = in(ValueIn);
2643 if (!needs_barrier(phase, value)) {
2644 return value;
2645 }
2646 return this;
2647 }
2648
2649 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) {
2650 Unique_Node_List visited;
2651 return needs_barrier_impl(phase, n, visited);
2652 }
2653
2654 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) {
2655 if (n == nullptr) return false;
2656 if (visited.member(n)) {
2657 return false; // Been there.
2658 }
2659 visited.push(n);
2660
2661 if (n->is_Allocate()) {
2662 // tty->print_cr("optimize barrier on alloc");
2663 return false;
2664 }
2665 if (n->is_Call()) {
2666 // tty->print_cr("optimize barrier on call");
2667 return false;
2668 }
2669
2670 const Type* type = phase->type(n);
2671 if (type == Type::TOP) {
2672 return false;
2673 }
2674 if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) {
2675 // tty->print_cr("optimize barrier on null");
2676 return false;
2677 }
2678 if (type->make_oopptr() && type->make_oopptr()->const_oop() != nullptr) {
2679 // tty->print_cr("optimize barrier on constant");
2680 return false;
2681 }
2682
2683 switch (n->Opcode()) {
2684 case Op_AddP:
2685 return true; // TODO: Can refine?
2686 case Op_LoadP:
2687 case Op_ShenandoahCompareAndExchangeN:
2688 case Op_ShenandoahCompareAndExchangeP:
2689 case Op_CompareAndExchangeN:
2690 case Op_CompareAndExchangeP:
2691 case Op_GetAndSetN:
2692 case Op_GetAndSetP:
2693 return true;
2694 case Op_Phi: {
2695 for (uint i = 1; i < n->req(); i++) {
2696 if (needs_barrier_impl(phase, n->in(i), visited)) return true;
2697 }
2698 return false;
2699 }
2700 case Op_CheckCastPP:
2701 case Op_CastPP:
2702 return needs_barrier_impl(phase, n->in(1), visited);
2703 case Op_Proj:
2704 return needs_barrier_impl(phase, n->in(0), visited);
2705 case Op_ShenandoahLoadReferenceBarrier:
2706 // tty->print_cr("optimize barrier on barrier");
2707 return false;
2708 case Op_Parm:
2709 // tty->print_cr("optimize barrier on input arg");
2710 return false;
2711 case Op_DecodeN:
2712 case Op_EncodeP:
2713 return needs_barrier_impl(phase, n->in(1), visited);
2714 case Op_LoadN:
2715 return true;
2716 case Op_CMoveN:
2717 case Op_CMoveP:
2718 return needs_barrier_impl(phase, n->in(2), visited) ||
2719 needs_barrier_impl(phase, n->in(3), visited);
2720 case Op_CreateEx:
2721 return false;
2722 default:
2723 break;
2724 }
2725 #ifdef ASSERT
2726 tty->print("need barrier on?: ");
2727 tty->print_cr("ins:");
2728 n->dump(2);
2729 tty->print_cr("outs:");
2730 n->dump(-2);
2731 ShouldNotReachHere();
2732 #endif
2733 return true;
2734 }