1 /*
2 * Copyright (c) 2015, 2021, Red Hat, Inc. All rights reserved.
3 * Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "precompiled.hpp"
27
28 #include "classfile/javaClasses.hpp"
29 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
30 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
31 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
32 #include "gc/shenandoah/shenandoahForwarding.hpp"
33 #include "gc/shenandoah/shenandoahHeap.hpp"
34 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
35 #include "gc/shenandoah/shenandoahRuntime.hpp"
36 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
37 #include "opto/arraycopynode.hpp"
38 #include "opto/block.hpp"
39 #include "opto/callnode.hpp"
40 #include "opto/castnode.hpp"
41 #include "opto/movenode.hpp"
42 #include "opto/phaseX.hpp"
43 #include "opto/rootnode.hpp"
44 #include "opto/runtime.hpp"
45 #include "opto/subnode.hpp"
46
47 bool ShenandoahBarrierC2Support::expand(Compile* C, PhaseIterGVN& igvn) {
48 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
49 if (state->load_reference_barriers_count() > 0) {
50 assert(C->post_loop_opts_phase(), "no loop opts allowed");
51 C->reset_post_loop_opts_phase(); // ... but we know what we are doing
52 C->clear_major_progress();
53 PhaseIdealLoop::optimize(igvn, LoopOptsShenandoahExpand);
54 if (C->failing()) return false;
55
56 C->set_major_progress();
57 if (!C->optimize_loops(igvn, LoopOptsShenandoahPostExpand)) {
58 return false;
59 }
60 C->clear_major_progress();
61 C->process_for_post_loop_opts_igvn(igvn);
62 if (C->failing()) return false;
63
64 C->set_post_loop_opts_phase(); // now for real!
65 }
66 return true;
67 }
68
69 bool ShenandoahBarrierC2Support::is_gc_state_test(Node* iff, int mask) {
70 if (!UseShenandoahGC) {
71 return false;
72 }
73 assert(iff->is_If(), "bad input");
74 if (iff->Opcode() != Op_If) {
75 return false;
76 }
77 Node* bol = iff->in(1);
78 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
79 return false;
80 }
81 Node* cmp = bol->in(1);
82 if (cmp->Opcode() != Op_CmpI) {
83 return false;
84 }
85 Node* in1 = cmp->in(1);
86 Node* in2 = cmp->in(2);
87 if (in2->find_int_con(-1) != 0) {
88 return false;
89 }
90 if (in1->Opcode() != Op_AndI) {
91 return false;
92 }
93 in2 = in1->in(2);
94 if (in2->find_int_con(-1) != mask) {
95 return false;
96 }
97 in1 = in1->in(1);
98
99 return is_gc_state_load(in1);
100 }
101
102 bool ShenandoahBarrierC2Support::is_heap_stable_test(Node* iff) {
103 return is_gc_state_test(iff, ShenandoahHeap::HAS_FORWARDED);
104 }
105
106 bool ShenandoahBarrierC2Support::is_gc_state_load(Node *n) {
107 if (!UseShenandoahGC) {
108 return false;
109 }
110 if (n->Opcode() != Op_LoadB && n->Opcode() != Op_LoadUB) {
111 return false;
112 }
113 Node* addp = n->in(MemNode::Address);
114 if (!addp->is_AddP()) {
115 return false;
116 }
117 Node* base = addp->in(AddPNode::Address);
118 Node* off = addp->in(AddPNode::Offset);
119 if (base->Opcode() != Op_ThreadLocal) {
120 return false;
121 }
122 if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) {
123 return false;
124 }
125 return true;
126 }
127
128 bool ShenandoahBarrierC2Support::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) {
129 assert(phase->is_dominator(stop, start), "bad inputs");
130 ResourceMark rm;
131 Unique_Node_List wq;
132 wq.push(start);
133 for (uint next = 0; next < wq.size(); next++) {
134 Node *m = wq.at(next);
135 if (m == stop) {
136 continue;
137 }
138 if (m->is_SafePoint() && !m->is_CallLeaf()) {
139 return true;
140 }
141 if (m->is_Region()) {
142 for (uint i = 1; i < m->req(); i++) {
143 wq.push(m->in(i));
144 }
145 } else {
146 wq.push(m->in(0));
147 }
148 }
149 return false;
150 }
151
152 #ifdef ASSERT
153 bool ShenandoahBarrierC2Support::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) {
154 assert(phis.size() == 0, "");
155
156 while (true) {
157 if (in->bottom_type() == TypePtr::NULL_PTR) {
158 if (trace) {tty->print_cr("null");}
159 } else if (!in->bottom_type()->make_ptr()->make_oopptr()) {
160 if (trace) {tty->print_cr("Non oop");}
161 } else {
162 if (in->is_ConstraintCast()) {
163 in = in->in(1);
164 continue;
165 } else if (in->is_AddP()) {
166 assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access");
167 in = in->in(AddPNode::Address);
168 continue;
169 } else if (in->is_Con()) {
170 if (trace) {
171 tty->print("Found constant");
172 in->dump();
173 }
174 } else if (in->Opcode() == Op_Parm) {
175 if (trace) {
176 tty->print("Found argument");
177 }
178 } else if (in->Opcode() == Op_CreateEx) {
179 if (trace) {
180 tty->print("Found create-exception");
181 }
182 } else if (in->Opcode() == Op_LoadP && in->adr_type() == TypeRawPtr::BOTTOM) {
183 if (trace) {
184 tty->print("Found raw LoadP (OSR argument?)");
185 }
186 } else if (in->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
187 if (t == ShenandoahOopStore) {
188 return false;
189 }
190 barriers_used.push(in);
191 if (trace) {tty->print("Found barrier"); in->dump();}
192 } else if (in->is_Proj() && in->in(0)->is_Allocate()) {
193 if (trace) {
194 tty->print("Found alloc");
195 in->in(0)->dump();
196 }
197 } else if (in->is_Proj() && (in->in(0)->Opcode() == Op_CallStaticJava || in->in(0)->Opcode() == Op_CallDynamicJava)) {
198 if (trace) {
199 tty->print("Found Java call");
200 }
201 } else if (in->is_Phi()) {
202 if (!visited.test_set(in->_idx)) {
203 if (trace) {tty->print("Pushed phi:"); in->dump();}
204 phis.push(in, 2);
205 in = in->in(1);
206 continue;
207 }
208 if (trace) {tty->print("Already seen phi:"); in->dump();}
209 } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) {
210 if (!visited.test_set(in->_idx)) {
211 if (trace) {tty->print("Pushed cmovep:"); in->dump();}
212 phis.push(in, CMoveNode::IfTrue);
213 in = in->in(CMoveNode::IfFalse);
214 continue;
215 }
216 if (trace) {tty->print("Already seen cmovep:"); in->dump();}
217 } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) {
218 in = in->in(1);
219 continue;
220 } else {
221 return false;
222 }
223 }
224 bool cont = false;
225 while (phis.is_nonempty()) {
226 uint idx = phis.index();
227 Node* phi = phis.node();
228 if (idx >= phi->req()) {
229 if (trace) {tty->print("Popped phi:"); phi->dump();}
230 phis.pop();
231 continue;
232 }
233 if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();}
234 in = phi->in(idx);
235 phis.set_index(idx+1);
236 cont = true;
237 break;
238 }
239 if (!cont) {
240 break;
241 }
242 }
243 return true;
244 }
245
246 void ShenandoahBarrierC2Support::report_verify_failure(const char* msg, Node* n1, Node* n2) {
247 if (n1 != nullptr) {
248 n1->dump(+10);
249 }
250 if (n2 != nullptr) {
251 n2->dump(+10);
252 }
253 fatal("%s", msg);
254 }
255
256 void ShenandoahBarrierC2Support::verify(RootNode* root) {
257 ResourceMark rm;
258 Unique_Node_List wq;
259 GrowableArray<Node*> barriers;
260 Unique_Node_List barriers_used;
261 Node_Stack phis(0);
262 VectorSet visited;
263 const bool trace = false;
264 const bool verify_no_useless_barrier = false;
265
266 wq.push(root);
267 for (uint next = 0; next < wq.size(); next++) {
268 Node *n = wq.at(next);
269 if (n->is_Load()) {
270 const bool trace = false;
271 if (trace) {tty->print("Verifying"); n->dump();}
272 if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) {
273 if (trace) {tty->print_cr("Load range/klass");}
274 } else {
275 const TypePtr* adr_type = n->as_Load()->adr_type();
276
277 if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) {
278 if (trace) {tty->print_cr("Mark load");}
279 } else if (adr_type->isa_instptr() &&
280 adr_type->is_instptr()->instance_klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
281 adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset()) {
282 if (trace) {tty->print_cr("Reference.get()");}
283 } else if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) {
284 report_verify_failure("Shenandoah verification: Load should have barriers", n);
285 }
286 }
287 } else if (n->is_Store()) {
288 const bool trace = false;
289
290 if (trace) {tty->print("Verifying"); n->dump();}
291 if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) {
292 Node* adr = n->in(MemNode::Address);
293 bool verify = true;
294
295 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) {
296 adr = adr->in(AddPNode::Address);
297 if (adr->is_AddP()) {
298 assert(adr->in(AddPNode::Base)->is_top(), "");
299 adr = adr->in(AddPNode::Address);
300 if (adr->Opcode() == Op_LoadP &&
301 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() &&
302 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
303 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) {
304 if (trace) {tty->print_cr("SATB prebarrier");}
305 verify = false;
306 }
307 }
308 }
309
310 if (verify && !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahValue, trace, barriers_used)) {
311 report_verify_failure("Shenandoah verification: Store should have barriers", n);
312 }
313 }
314 if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
315 report_verify_failure("Shenandoah verification: Store (address) should have barriers", n);
316 }
317 } else if (n->Opcode() == Op_CmpP) {
318 const bool trace = false;
319
320 Node* in1 = n->in(1);
321 Node* in2 = n->in(2);
322 if (in1->bottom_type()->isa_oopptr()) {
323 if (trace) {tty->print("Verifying"); n->dump();}
324
325 bool mark_inputs = false;
326 if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR ||
327 (in1->is_Con() || in2->is_Con())) {
328 if (trace) {tty->print_cr("Comparison against a constant");}
329 mark_inputs = true;
330 } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) ||
331 (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) {
332 if (trace) {tty->print_cr("Comparison with newly alloc'ed object");}
333 mark_inputs = true;
334 } else {
335 assert(in2->bottom_type()->isa_oopptr(), "");
336
337 if (!verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) ||
338 !verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) {
339 report_verify_failure("Shenandoah verification: Cmp should have barriers", n);
340 }
341 }
342 if (verify_no_useless_barrier &&
343 mark_inputs &&
344 (!verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) ||
345 !verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) {
346 phis.clear();
347 visited.reset();
348 }
349 }
350 } else if (n->is_LoadStore()) {
351 if (n->in(MemNode::ValueIn)->bottom_type()->make_ptr() &&
352 !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahValue, trace, barriers_used)) {
353 report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n);
354 }
355
356 if (n->in(MemNode::Address)->bottom_type()->make_oopptr() && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
357 report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n);
358 }
359 } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) {
360 CallNode* call = n->as_Call();
361
362 static struct {
363 const char* name;
364 struct {
365 int pos;
366 verify_type t;
367 } args[6];
368 } calls[] = {
369 "array_partition_stub",
370 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+4, ShenandoahStore }, { -1, ShenandoahNone },
371 { -1, ShenandoahNone }, { -1, ShenandoahNone }, { -1, ShenandoahNone } },
372 "arraysort_stub",
373 { { TypeFunc::Parms, ShenandoahStore }, { -1, ShenandoahNone }, { -1, ShenandoahNone },
374 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
375 "aescrypt_encryptBlock",
376 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
377 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
378 "aescrypt_decryptBlock",
379 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
380 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
381 "multiplyToLen",
382 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { TypeFunc::Parms+4, ShenandoahStore },
383 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
384 "squareToLen",
385 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { -1, ShenandoahNone},
386 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
387 "montgomery_multiply",
388 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
389 { TypeFunc::Parms+6, ShenandoahStore }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
390 "montgomery_square",
391 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+5, ShenandoahStore },
392 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
393 "mulAdd",
394 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
395 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
396 "vectorizedMismatch",
397 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
398 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
399 "updateBytesCRC32",
400 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
401 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
402 "updateBytesAdler32",
403 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
404 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
405 "updateBytesCRC32C",
406 { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad}, { -1, ShenandoahNone},
407 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
408 "counterMode_AESCrypt",
409 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
410 { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } },
411 "cipherBlockChaining_encryptAESCrypt",
412 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
413 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
414 "cipherBlockChaining_decryptAESCrypt",
415 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
416 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
417 "shenandoah_clone_barrier",
418 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
419 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
420 "ghash_processBlocks",
421 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
422 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
423 "sha1_implCompress",
424 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
425 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
426 "sha256_implCompress",
427 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
428 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
429 "sha512_implCompress",
430 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
431 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
432 "sha1_implCompressMB",
433 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
434 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
435 "sha256_implCompressMB",
436 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
437 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
438 "sha512_implCompressMB",
439 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
440 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
441 "encodeBlock",
442 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahStore }, { -1, ShenandoahNone },
443 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
444 "decodeBlock",
445 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahStore }, { -1, ShenandoahNone },
446 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
447 "intpoly_montgomeryMult_P256",
448 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahStore },
449 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
450 "intpoly_assign",
451 { { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad }, { -1, ShenandoahNone },
452 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
453 };
454
455 if (call->is_call_to_arraycopystub()) {
456 Node* dest = nullptr;
457 const TypeTuple* args = n->as_Call()->_tf->domain();
458 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
459 if (args->field_at(i)->isa_ptr()) {
460 j++;
461 if (j == 2) {
462 dest = n->in(i);
463 break;
464 }
465 }
466 }
467 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) ||
468 !verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) {
469 report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n);
470 }
471 } else if (strlen(call->_name) > 5 &&
472 !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) {
473 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) {
474 report_verify_failure("Shenandoah verification: _fill should have barriers", n);
475 }
476 } else if (!strcmp(call->_name, "shenandoah_wb_pre")) {
477 // skip
478 } else {
479 const int calls_len = sizeof(calls) / sizeof(calls[0]);
480 int i = 0;
481 for (; i < calls_len; i++) {
482 if (!strcmp(calls[i].name, call->_name)) {
483 break;
484 }
485 }
486 if (i != calls_len) {
487 const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]);
488 for (uint j = 0; j < args_len; j++) {
489 int pos = calls[i].args[j].pos;
490 if (pos == -1) {
491 break;
492 }
493 if (!verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) {
494 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
495 }
496 }
497 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
498 if (call->in(j)->bottom_type()->make_ptr() &&
499 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
500 uint k = 0;
501 for (; k < args_len && calls[i].args[k].pos != (int)j; k++);
502 if (k == args_len) {
503 fatal("arg %d for call %s not covered", j, call->_name);
504 }
505 }
506 }
507 } else {
508 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
509 if (call->in(j)->bottom_type()->make_ptr() &&
510 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
511 fatal("%s not covered", call->_name);
512 }
513 }
514 }
515 }
516 } else if (n->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
517 // skip
518 } else if (n->is_AddP()
519 || n->is_Phi()
520 || n->is_ConstraintCast()
521 || n->Opcode() == Op_Return
522 || n->Opcode() == Op_CMoveP
523 || n->Opcode() == Op_CMoveN
524 || n->Opcode() == Op_Rethrow
525 || n->is_MemBar()
526 || n->Opcode() == Op_Conv2B
527 || n->Opcode() == Op_SafePoint
528 || n->is_CallJava()
529 || n->Opcode() == Op_Unlock
530 || n->Opcode() == Op_EncodeP
531 || n->Opcode() == Op_DecodeN) {
532 // nothing to do
533 } else {
534 static struct {
535 int opcode;
536 struct {
537 int pos;
538 verify_type t;
539 } inputs[2];
540 } others[] = {
541 Op_FastLock,
542 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
543 Op_Lock,
544 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone} },
545 Op_ArrayCopy,
546 { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } },
547 Op_StrCompressedCopy,
548 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
549 Op_StrInflatedCopy,
550 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
551 Op_AryEq,
552 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
553 Op_StrIndexOf,
554 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
555 Op_StrComp,
556 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
557 Op_StrEquals,
558 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
559 Op_VectorizedHashCode,
560 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } },
561 Op_EncodeISOArray,
562 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
563 Op_CountPositives,
564 { { 2, ShenandoahLoad }, { -1, ShenandoahNone} },
565 Op_CastP2X,
566 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
567 Op_StrIndexOfChar,
568 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } },
569 };
570
571 const int others_len = sizeof(others) / sizeof(others[0]);
572 int i = 0;
573 for (; i < others_len; i++) {
574 if (others[i].opcode == n->Opcode()) {
575 break;
576 }
577 }
578 uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req();
579 if (i != others_len) {
580 const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]);
581 for (uint j = 0; j < inputs_len; j++) {
582 int pos = others[i].inputs[j].pos;
583 if (pos == -1) {
584 break;
585 }
586 if (!verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) {
587 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
588 }
589 }
590 for (uint j = 1; j < stop; j++) {
591 if (n->in(j) != nullptr && n->in(j)->bottom_type()->make_ptr() &&
592 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
593 uint k = 0;
594 for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++);
595 if (k == inputs_len) {
596 fatal("arg %d for node %s not covered", j, n->Name());
597 }
598 }
599 }
600 } else {
601 for (uint j = 1; j < stop; j++) {
602 if (n->in(j) != nullptr && n->in(j)->bottom_type()->make_ptr() &&
603 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
604 fatal("%s not covered", n->Name());
605 }
606 }
607 }
608 }
609
610 if (n->is_SafePoint()) {
611 SafePointNode* sfpt = n->as_SafePoint();
612 if (verify_no_useless_barrier && sfpt->jvms() != nullptr) {
613 for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) {
614 if (!verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) {
615 phis.clear();
616 visited.reset();
617 }
618 }
619 }
620 }
621 }
622
623 if (verify_no_useless_barrier) {
624 for (int i = 0; i < barriers.length(); i++) {
625 Node* n = barriers.at(i);
626 if (!barriers_used.member(n)) {
627 tty->print("XXX useless barrier"); n->dump(-2);
628 ShouldNotReachHere();
629 }
630 }
631 }
632 }
633 #endif
634
635 bool ShenandoahBarrierC2Support::is_dominator_same_ctrl(Node* c, Node* d, Node* n, PhaseIdealLoop* phase) {
636 // That both nodes have the same control is not sufficient to prove
637 // domination, verify that there's no path from d to n
638 ResourceMark rm;
639 Unique_Node_List wq;
640 wq.push(d);
641 for (uint next = 0; next < wq.size(); next++) {
642 Node *m = wq.at(next);
643 if (m == n) {
644 return false;
645 }
646 if (m->is_Phi() && m->in(0)->is_Loop()) {
647 assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control");
648 } else {
649 if (m->is_Store() || m->is_LoadStore()) {
650 // Take anti-dependencies into account
651 Node* mem = m->in(MemNode::Memory);
652 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
653 Node* u = mem->fast_out(i);
654 if (u->is_Load() && phase->C->can_alias(m->adr_type(), phase->C->get_alias_index(u->adr_type())) &&
655 phase->ctrl_or_self(u) == c) {
656 wq.push(u);
657 }
658 }
659 }
660 for (uint i = 0; i < m->req(); i++) {
661 if (m->in(i) != nullptr && phase->ctrl_or_self(m->in(i)) == c) {
662 wq.push(m->in(i));
663 }
664 }
665 }
666 }
667 return true;
668 }
669
670 bool ShenandoahBarrierC2Support::is_dominator(Node* d_c, Node* n_c, Node* d, Node* n, PhaseIdealLoop* phase) {
671 if (d_c != n_c) {
672 return phase->is_dominator(d_c, n_c);
673 }
674 return is_dominator_same_ctrl(d_c, d, n, phase);
675 }
676
677 Node* next_mem(Node* mem, int alias) {
678 Node* res = nullptr;
679 if (mem->is_Proj()) {
680 res = mem->in(0);
681 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
682 res = mem->in(TypeFunc::Memory);
683 } else if (mem->is_Phi()) {
684 res = mem->in(1);
685 } else if (mem->is_MergeMem()) {
686 res = mem->as_MergeMem()->memory_at(alias);
687 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
688 assert(alias == Compile::AliasIdxRaw, "following raw memory can't lead to a barrier");
689 res = mem->in(MemNode::Memory);
690 } else {
691 #ifdef ASSERT
692 mem->dump();
693 #endif
694 ShouldNotReachHere();
695 }
696 return res;
697 }
698
699 Node* ShenandoahBarrierC2Support::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) {
700 Node* iffproj = nullptr;
701 while (c != dom) {
702 Node* next = phase->idom(c);
703 assert(next->unique_ctrl_out_or_null() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?");
704 if (c->is_Region()) {
705 ResourceMark rm;
706 Unique_Node_List wq;
707 wq.push(c);
708 for (uint i = 0; i < wq.size(); i++) {
709 Node *n = wq.at(i);
710 if (n == next) {
711 continue;
712 }
713 if (n->is_Region()) {
714 for (uint j = 1; j < n->req(); j++) {
715 wq.push(n->in(j));
716 }
717 } else {
718 wq.push(n->in(0));
719 }
720 }
721 for (uint i = 0; i < wq.size(); i++) {
722 Node *n = wq.at(i);
723 assert(n->is_CFG(), "");
724 if (n->is_Multi()) {
725 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
726 Node* u = n->fast_out(j);
727 if (u->is_CFG()) {
728 if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj()) {
729 return NodeSentinel;
730 }
731 }
732 }
733 }
734 }
735 } else if (c->is_Proj()) {
736 if (c->is_IfProj()) {
737 if (c->as_Proj()->is_uncommon_trap_if_pattern() != nullptr) {
738 // continue;
739 } else {
740 if (!allow_one_proj) {
741 return NodeSentinel;
742 }
743 if (iffproj == nullptr) {
744 iffproj = c;
745 } else {
746 return NodeSentinel;
747 }
748 }
749 } else if (c->Opcode() == Op_JumpProj) {
750 return NodeSentinel; // unsupported
751 } else if (c->Opcode() == Op_CatchProj) {
752 return NodeSentinel; // unsupported
753 } else if (c->Opcode() == Op_CProj && next->is_NeverBranch()) {
754 return NodeSentinel; // unsupported
755 } else {
756 assert(next->unique_ctrl_out() == c, "unsupported branch pattern");
757 }
758 }
759 c = next;
760 }
761 return iffproj;
762 }
763
764 Node* ShenandoahBarrierC2Support::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) {
765 ResourceMark rm;
766 VectorSet wq;
767 wq.set(mem->_idx);
768 mem_ctrl = phase->ctrl_or_self(mem);
769 while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) {
770 mem = next_mem(mem, alias);
771 if (wq.test_set(mem->_idx)) {
772 return nullptr;
773 }
774 mem_ctrl = phase->ctrl_or_self(mem);
775 }
776 if (mem->is_MergeMem()) {
777 mem = mem->as_MergeMem()->memory_at(alias);
778 mem_ctrl = phase->ctrl_or_self(mem);
779 }
780 return mem;
781 }
782
783 Node* ShenandoahBarrierC2Support::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) {
784 Node* mem = nullptr;
785 Node* c = ctrl;
786 do {
787 if (c->is_Region()) {
788 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == nullptr; i++) {
789 Node* u = c->fast_out(i);
790 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
791 if (u->adr_type() == TypePtr::BOTTOM) {
792 mem = u;
793 }
794 }
795 }
796 } else {
797 if (c->is_Call() && c->as_Call()->adr_type() != nullptr) {
798 CallProjections projs;
799 c->as_Call()->extract_projections(&projs, true, false);
800 if (projs.fallthrough_memproj != nullptr) {
801 if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
802 if (projs.catchall_memproj == nullptr) {
803 mem = projs.fallthrough_memproj;
804 } else {
805 if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) {
806 mem = projs.fallthrough_memproj;
807 } else {
808 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier");
809 mem = projs.catchall_memproj;
810 }
811 }
812 }
813 } else {
814 Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
815 if (proj != nullptr &&
816 proj->adr_type() == TypePtr::BOTTOM) {
817 mem = proj;
818 }
819 }
820 } else {
821 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
822 Node* u = c->fast_out(i);
823 if (u->is_Proj() &&
824 u->bottom_type() == Type::MEMORY &&
825 u->adr_type() == TypePtr::BOTTOM) {
826 assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
827 assert(mem == nullptr, "only one proj");
828 mem = u;
829 }
830 }
831 assert(!c->is_Call() || c->as_Call()->adr_type() != nullptr || mem == nullptr, "no mem projection expected");
832 }
833 }
834 c = phase->idom(c);
835 } while (mem == nullptr);
836 return mem;
837 }
838
839 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
840 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
841 Node* u = n->fast_out(i);
842 if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
843 uses.push(u);
844 }
845 }
846 }
847
848 static void hide_strip_mined_loop(OuterStripMinedLoopNode* outer, CountedLoopNode* inner, PhaseIdealLoop* phase) {
849 OuterStripMinedLoopEndNode* le = inner->outer_loop_end();
850 Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl));
851 phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl));
852 Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt);
853 phase->register_control(new_le, phase->get_loop(le), le->in(0));
854 phase->lazy_replace(outer, new_outer);
855 phase->lazy_replace(le, new_le);
856 inner->clear_strip_mined();
857 }
858
859 void ShenandoahBarrierC2Support::test_gc_state(Node*& ctrl, Node* raw_mem, Node*& test_fail_ctrl,
860 PhaseIdealLoop* phase, int flags) {
861 PhaseIterGVN& igvn = phase->igvn();
862 Node* old_ctrl = ctrl;
863
864 Node* thread = new ThreadLocalNode();
865 Node* gc_state_offset = igvn.MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
866 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, gc_state_offset);
867 Node* gc_state = new LoadBNode(old_ctrl, raw_mem, gc_state_addr,
868 DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
869 TypeInt::BYTE, MemNode::unordered);
870 Node* gc_state_and = new AndINode(gc_state, igvn.intcon(flags));
871 Node* gc_state_cmp = new CmpINode(gc_state_and, igvn.zerocon(T_INT));
872 Node* gc_state_bool = new BoolNode(gc_state_cmp, BoolTest::ne);
873
874 IfNode* gc_state_iff = new IfNode(old_ctrl, gc_state_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
875 ctrl = new IfTrueNode(gc_state_iff);
876 test_fail_ctrl = new IfFalseNode(gc_state_iff);
877
878 IdealLoopTree* loop = phase->get_loop(old_ctrl);
879 phase->register_control(gc_state_iff, loop, old_ctrl);
880 phase->register_control(ctrl, loop, gc_state_iff);
881 phase->register_control(test_fail_ctrl, loop, gc_state_iff);
882
883 phase->register_new_node(thread, old_ctrl);
884 phase->register_new_node(gc_state_addr, old_ctrl);
885 phase->register_new_node(gc_state, old_ctrl);
886 phase->register_new_node(gc_state_and, old_ctrl);
887 phase->register_new_node(gc_state_cmp, old_ctrl);
888 phase->register_new_node(gc_state_bool, old_ctrl);
889
890 phase->set_ctrl(gc_state_offset, phase->C->root());
891
892 assert(is_gc_state_test(gc_state_iff, flags), "Should match the shape");
893 }
894
895 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) {
896 Node* old_ctrl = ctrl;
897 PhaseIterGVN& igvn = phase->igvn();
898
899 const Type* val_t = igvn.type(val);
900 if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
901 Node* null_cmp = new CmpPNode(val, igvn.zerocon(T_OBJECT));
902 Node* null_test = new BoolNode(null_cmp, BoolTest::ne);
903
904 IfNode* null_iff = new IfNode(old_ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
905 ctrl = new IfTrueNode(null_iff);
906 null_ctrl = new IfFalseNode(null_iff);
907
908 IdealLoopTree* loop = phase->get_loop(old_ctrl);
909 phase->register_control(null_iff, loop, old_ctrl);
910 phase->register_control(ctrl, loop, null_iff);
911 phase->register_control(null_ctrl, loop, null_iff);
912
913 phase->register_new_node(null_cmp, old_ctrl);
914 phase->register_new_node(null_test, old_ctrl);
915 }
916 }
917
918 void ShenandoahBarrierC2Support::test_in_cset(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) {
919 Node* old_ctrl = ctrl;
920 PhaseIterGVN& igvn = phase->igvn();
921
922 Node* raw_val = new CastP2XNode(old_ctrl, val);
923 Node* cset_idx = new URShiftXNode(raw_val, igvn.intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
924
925 // Figure out the target cset address with raw pointer math.
926 // This avoids matching AddP+LoadB that would emit inefficient code.
927 // See JDK-8245465.
928 Node* cset_addr_ptr = igvn.makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
929 Node* cset_addr = new CastP2XNode(old_ctrl, cset_addr_ptr);
930 Node* cset_load_addr = new AddXNode(cset_addr, cset_idx);
931 Node* cset_load_ptr = new CastX2PNode(cset_load_addr);
932
933 Node* cset_load = new LoadBNode(old_ctrl, raw_mem, cset_load_ptr,
934 DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
935 TypeInt::BYTE, MemNode::unordered);
936 Node* cset_cmp = new CmpINode(cset_load, igvn.zerocon(T_INT));
937 Node* cset_bool = new BoolNode(cset_cmp, BoolTest::ne);
938
939 IfNode* cset_iff = new IfNode(old_ctrl, cset_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
940 ctrl = new IfTrueNode(cset_iff);
941 not_cset_ctrl = new IfFalseNode(cset_iff);
942
943 IdealLoopTree *loop = phase->get_loop(old_ctrl);
944 phase->register_control(cset_iff, loop, old_ctrl);
945 phase->register_control(ctrl, loop, cset_iff);
946 phase->register_control(not_cset_ctrl, loop, cset_iff);
947
948 phase->set_ctrl(cset_addr_ptr, phase->C->root());
949
950 phase->register_new_node(raw_val, old_ctrl);
951 phase->register_new_node(cset_idx, old_ctrl);
952 phase->register_new_node(cset_addr, old_ctrl);
953 phase->register_new_node(cset_load_addr, old_ctrl);
954 phase->register_new_node(cset_load_ptr, old_ctrl);
955 phase->register_new_node(cset_load, old_ctrl);
956 phase->register_new_node(cset_cmp, old_ctrl);
957 phase->register_new_node(cset_bool, old_ctrl);
958 }
959
960 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node* load_addr,
961 DecoratorSet decorators, PhaseIdealLoop* phase) {
962 IdealLoopTree*loop = phase->get_loop(ctrl);
963 const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr();
964
965 address calladdr = nullptr;
966 const char* name = nullptr;
967 bool is_strong = ShenandoahBarrierSet::is_strong_access(decorators);
968 bool is_weak = ShenandoahBarrierSet::is_weak_access(decorators);
969 bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
970 bool is_native = ShenandoahBarrierSet::is_native_access(decorators);
971 bool is_narrow = UseCompressedOops && !is_native;
972 if (is_strong) {
973 if (is_narrow) {
974 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow);
975 name = "load_reference_barrier_strong_narrow";
976 } else {
977 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
978 name = "load_reference_barrier_strong";
979 }
980 } else if (is_weak) {
981 if (is_narrow) {
982 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow);
983 name = "load_reference_barrier_weak_narrow";
984 } else {
985 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak);
986 name = "load_reference_barrier_weak";
987 }
988 } else {
989 assert(is_phantom, "only remaining strength");
990 if (is_narrow) {
991 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom_narrow);
992 name = "load_reference_barrier_phantom_narrow";
993 } else {
994 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom);
995 name = "load_reference_barrier_phantom";
996 }
997 }
998 Node* call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type(), calladdr, name, TypeRawPtr::BOTTOM);
999
1000 call->init_req(TypeFunc::Control, ctrl);
1001 call->init_req(TypeFunc::I_O, phase->C->top());
1002 call->init_req(TypeFunc::Memory, phase->C->top());
1003 call->init_req(TypeFunc::FramePtr, phase->C->top());
1004 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1005 call->init_req(TypeFunc::Parms, val);
1006 call->init_req(TypeFunc::Parms+1, load_addr);
1007 phase->register_control(call, loop, ctrl);
1008 ctrl = new ProjNode(call, TypeFunc::Control);
1009 phase->register_control(ctrl, loop, call);
1010 val = new ProjNode(call, TypeFunc::Parms);
1011 phase->register_new_node(val, call);
1012 val = new CheckCastPPNode(ctrl, val, obj_type);
1013 phase->register_new_node(val, ctrl);
1014 }
1015
1016 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) {
1017 Node* ctrl = phase->get_ctrl(barrier);
1018 Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
1019
1020 // Update the control of all nodes that should be after the
1021 // barrier control flow
1022 uses.clear();
1023 // Every node that is control dependent on the barrier's input
1024 // control will be after the expanded barrier. The raw memory (if
1025 // its memory is control dependent on the barrier's input control)
1026 // must stay above the barrier.
1027 uses_to_ignore.clear();
1028 if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
1029 uses_to_ignore.push(init_raw_mem);
1030 }
1031 for (uint next = 0; next < uses_to_ignore.size(); next++) {
1032 Node *n = uses_to_ignore.at(next);
1033 for (uint i = 0; i < n->req(); i++) {
1034 Node* in = n->in(i);
1035 if (in != nullptr && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
1036 uses_to_ignore.push(in);
1037 }
1038 }
1039 }
1040 for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
1041 Node* u = ctrl->fast_out(i);
1042 if (u->_idx < last &&
1043 u != barrier &&
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 if ((ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) || ctrl->is_CallJava()) {
1152 CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_CallJava() : ctrl->as_CallJava();
1153 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
1154 // The rethrow call may have too many projections to be
1155 // properly handled here. Given there's no reason for a
1156 // barrier to depend on the call, move it above the call
1157 stack.push(lrb, 0);
1158 do {
1159 Node* n = stack.node();
1160 uint idx = stack.index();
1161 if (idx < n->req()) {
1162 Node* in = n->in(idx);
1163 stack.set_index(idx+1);
1164 if (in != nullptr) {
1165 if (phase->has_ctrl(in)) {
1166 if (phase->is_dominator(call, phase->get_ctrl(in))) {
1167 #ifdef ASSERT
1168 for (uint i = 0; i < stack.size(); i++) {
1169 assert(stack.node_at(i) != in, "node shouldn't have been seen yet");
1170 }
1171 #endif
1172 stack.push(in, 0);
1173 }
1174 } else {
1175 assert(phase->is_dominator(in, call->in(0)), "no dependency on the call");
1176 }
1177 }
1178 } else {
1179 phase->set_ctrl(n, call->in(0));
1180 stack.pop();
1181 }
1182 } while(stack.size() > 0);
1183 continue;
1184 }
1185 CallProjections projs;
1186 call->extract_projections(&projs, false, false);
1187
1188 #ifdef ASSERT
1189 VectorSet cloned;
1190 #endif
1191 Node* lrb_clone = lrb->clone();
1192 phase->register_new_node(lrb_clone, projs.catchall_catchproj);
1193 phase->set_ctrl(lrb, projs.fallthrough_catchproj);
1194
1195 stack.push(lrb, 0);
1196 clones.push(lrb_clone);
1197
1198 do {
1199 assert(stack.size() == clones.size(), "");
1200 Node* n = stack.node();
1201 #ifdef ASSERT
1202 if (n->is_Load()) {
1203 Node* mem = n->in(MemNode::Memory);
1204 for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
1205 Node* u = mem->fast_out(j);
1206 assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?");
1207 }
1208 }
1209 #endif
1210 uint idx = stack.index();
1211 Node* n_clone = clones.at(clones.size()-1);
1212 if (idx < n->outcnt()) {
1213 Node* u = n->raw_out(idx);
1214 Node* c = phase->ctrl_or_self(u);
1215 if (phase->is_dominator(call, c) && phase->is_dominator(c, projs.fallthrough_proj)) {
1216 stack.set_index(idx+1);
1217 assert(!u->is_CFG(), "");
1218 stack.push(u, 0);
1219 assert(!cloned.test_set(u->_idx), "only one clone");
1220 Node* u_clone = u->clone();
1221 int nb = u_clone->replace_edge(n, n_clone, &phase->igvn());
1222 assert(nb > 0, "should have replaced some uses");
1223 phase->register_new_node(u_clone, projs.catchall_catchproj);
1224 clones.push(u_clone);
1225 phase->set_ctrl(u, projs.fallthrough_catchproj);
1226 } else {
1227 bool replaced = false;
1228 if (u->is_Phi()) {
1229 for (uint k = 1; k < u->req(); k++) {
1230 if (u->in(k) == n) {
1231 if (phase->is_dominator(projs.catchall_catchproj, u->in(0)->in(k))) {
1232 phase->igvn().replace_input_of(u, k, n_clone);
1233 replaced = true;
1234 } else if (!phase->is_dominator(projs.fallthrough_catchproj, u->in(0)->in(k))) {
1235 phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase));
1236 replaced = true;
1237 }
1238 }
1239 }
1240 } else {
1241 if (phase->is_dominator(projs.catchall_catchproj, c)) {
1242 phase->igvn().rehash_node_delayed(u);
1243 int nb = u->replace_edge(n, n_clone, &phase->igvn());
1244 assert(nb > 0, "should have replaced some uses");
1245 replaced = true;
1246 } else if (!phase->is_dominator(projs.fallthrough_catchproj, c)) {
1247 if (u->is_If()) {
1248 // Can't break If/Bool/Cmp chain
1249 assert(n->is_Bool(), "unexpected If shape");
1250 assert(stack.node_at(stack.size()-2)->is_Cmp(), "unexpected If shape");
1251 assert(n_clone->is_Bool(), "unexpected clone");
1252 assert(clones.at(clones.size()-2)->is_Cmp(), "unexpected clone");
1253 Node* bol_clone = n->clone();
1254 Node* cmp_clone = stack.node_at(stack.size()-2)->clone();
1255 bol_clone->set_req(1, cmp_clone);
1256
1257 Node* nn = stack.node_at(stack.size()-3);
1258 Node* nn_clone = clones.at(clones.size()-3);
1259 assert(nn->Opcode() == nn_clone->Opcode(), "mismatch");
1260
1261 int nb = cmp_clone->replace_edge(nn, create_phis_on_call_return(ctrl, c, nn, nn_clone, projs, phase),
1262 &phase->igvn());
1263 assert(nb > 0, "should have replaced some uses");
1264
1265 phase->register_new_node(bol_clone, u->in(0));
1266 phase->register_new_node(cmp_clone, u->in(0));
1267
1268 phase->igvn().replace_input_of(u, 1, bol_clone);
1269
1270 } else {
1271 phase->igvn().rehash_node_delayed(u);
1272 int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase), &phase->igvn());
1273 assert(nb > 0, "should have replaced some uses");
1274 }
1275 replaced = true;
1276 }
1277 }
1278 if (!replaced) {
1279 stack.set_index(idx+1);
1280 }
1281 }
1282 } else {
1283 stack.pop();
1284 clones.pop();
1285 }
1286 } while (stack.size() > 0);
1287 assert(stack.size() == 0 && clones.size() == 0, "");
1288 }
1289 }
1290
1291 for (int i = 0; i < state->load_reference_barriers_count(); i++) {
1292 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1293 Node* ctrl = phase->get_ctrl(lrb);
1294 IdealLoopTree* loop = phase->get_loop(ctrl);
1295 Node* head = loop->head();
1296 if (head->is_OuterStripMinedLoop()) {
1297 // Expanding a barrier here will break loop strip mining
1298 // verification. Transform the loop so the loop nest doesn't
1299 // appear as strip mined.
1300 OuterStripMinedLoopNode* outer = head->as_OuterStripMinedLoop();
1301 hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1302 }
1303 if (head->is_BaseCountedLoop() && ctrl->is_IfProj() && ctrl->in(0)->is_BaseCountedLoopEnd() &&
1304 head->as_BaseCountedLoop()->loopexit() == ctrl->in(0)) {
1305 Node* entry = head->in(LoopNode::EntryControl);
1306 Node* backedge = head->in(LoopNode::LoopBackControl);
1307 Node* new_head = new LoopNode(entry, backedge);
1308 phase->register_control(new_head, phase->get_loop(entry), entry);
1309 phase->lazy_replace(head, new_head);
1310 }
1311 }
1312
1313 // Expand load-reference-barriers
1314 MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
1315 Unique_Node_List uses_to_ignore;
1316 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1317 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1318 uint last = phase->C->unique();
1319 Node* ctrl = phase->get_ctrl(lrb);
1320 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1321
1322 Node* orig_ctrl = ctrl;
1323
1324 Node* raw_mem = fixer.find_mem(ctrl, lrb);
1325 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, nullptr);
1326
1327 IdealLoopTree *loop = phase->get_loop(ctrl);
1328
1329 Node* heap_stable_ctrl = nullptr;
1330 Node* null_ctrl = nullptr;
1331
1332 assert(val->bottom_type()->make_oopptr(), "need oop");
1333 assert(val->bottom_type()->make_oopptr()->const_oop() == nullptr, "expect non-constant");
1334
1335 enum { _heap_stable = 1, _evac_path, _not_cset, PATH_LIMIT };
1336 Node* region = new RegionNode(PATH_LIMIT);
1337 Node* val_phi = new PhiNode(region, val->bottom_type()->is_oopptr());
1338
1339 // Stable path.
1340 int flags = ShenandoahHeap::HAS_FORWARDED;
1341 if (!ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1342 flags |= ShenandoahHeap::WEAK_ROOTS;
1343 }
1344 test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, flags);
1345 IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
1346
1347 // Heap stable case
1348 region->init_req(_heap_stable, heap_stable_ctrl);
1349 val_phi->init_req(_heap_stable, val);
1350
1351 // Test for in-cset, unless it's a native-LRB. Native LRBs need to return null
1352 // even for non-cset objects to prevent resurrection of such objects.
1353 // Wires !in_cset(obj) to slot 2 of region and phis
1354 Node* not_cset_ctrl = nullptr;
1355 if (ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1356 test_in_cset(ctrl, not_cset_ctrl, val, raw_mem, phase);
1357 }
1358 if (not_cset_ctrl != nullptr) {
1359 region->init_req(_not_cset, not_cset_ctrl);
1360 val_phi->init_req(_not_cset, val);
1361 } else {
1362 region->del_req(_not_cset);
1363 val_phi->del_req(_not_cset);
1364 }
1365
1366 // Resolve object when orig-value is in cset.
1367 // Make the unconditional resolve for fwdptr.
1368
1369 // Call lrb-stub and wire up that path in slots 4
1370 Node* result_mem = nullptr;
1371
1372 Node* addr;
1373 {
1374 VectorSet visited;
1375 addr = get_load_addr(phase, visited, lrb);
1376 }
1377 if (addr->Opcode() == Op_AddP) {
1378 Node* orig_base = addr->in(AddPNode::Base);
1379 Node* base = new CheckCastPPNode(ctrl, orig_base, orig_base->bottom_type(), ConstraintCastNode::StrongDependency);
1380 phase->register_new_node(base, ctrl);
1381 if (addr->in(AddPNode::Base) == addr->in((AddPNode::Address))) {
1382 // Field access
1383 addr = addr->clone();
1384 addr->set_req(AddPNode::Base, base);
1385 addr->set_req(AddPNode::Address, base);
1386 phase->register_new_node(addr, ctrl);
1387 } else {
1388 Node* addr2 = addr->in(AddPNode::Address);
1389 if (addr2->Opcode() == Op_AddP && addr2->in(AddPNode::Base) == addr2->in(AddPNode::Address) &&
1390 addr2->in(AddPNode::Base) == orig_base) {
1391 addr2 = addr2->clone();
1392 addr2->set_req(AddPNode::Base, base);
1393 addr2->set_req(AddPNode::Address, base);
1394 phase->register_new_node(addr2, ctrl);
1395 addr = addr->clone();
1396 addr->set_req(AddPNode::Base, base);
1397 addr->set_req(AddPNode::Address, addr2);
1398 phase->register_new_node(addr, ctrl);
1399 }
1400 }
1401 }
1402 call_lrb_stub(ctrl, val, addr, lrb->decorators(), phase);
1403 region->init_req(_evac_path, ctrl);
1404 val_phi->init_req(_evac_path, val);
1405
1406 phase->register_control(region, loop, heap_stable_iff);
1407 Node* out_val = val_phi;
1408 phase->register_new_node(val_phi, region);
1409
1410 fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase);
1411
1412 ctrl = orig_ctrl;
1413
1414 phase->igvn().replace_node(lrb, out_val);
1415
1416 follow_barrier_uses(out_val, ctrl, uses, phase);
1417
1418 for(uint next = 0; next < uses.size(); next++ ) {
1419 Node *n = uses.at(next);
1420 assert(phase->get_ctrl(n) == ctrl, "bad control");
1421 assert(n != raw_mem, "should leave input raw mem above the barrier");
1422 phase->set_ctrl(n, region);
1423 follow_barrier_uses(n, ctrl, uses, phase);
1424 }
1425 fixer.record_new_ctrl(ctrl, region, raw_mem, raw_mem_for_ctrl);
1426 }
1427 // Done expanding load-reference-barriers.
1428 assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced");
1429 }
1430
1431 Node* ShenandoahBarrierC2Support::get_load_addr(PhaseIdealLoop* phase, VectorSet& visited, Node* in) {
1432 if (visited.test_set(in->_idx)) {
1433 return nullptr;
1434 }
1435 switch (in->Opcode()) {
1436 case Op_Proj:
1437 return get_load_addr(phase, visited, in->in(0));
1438 case Op_CastPP:
1439 case Op_CheckCastPP:
1440 case Op_DecodeN:
1441 case Op_EncodeP:
1442 return get_load_addr(phase, visited, in->in(1));
1443 case Op_LoadN:
1444 case Op_LoadP:
1445 return in->in(MemNode::Address);
1446 case Op_CompareAndExchangeN:
1447 case Op_CompareAndExchangeP:
1448 case Op_GetAndSetN:
1449 case Op_GetAndSetP:
1450 case Op_ShenandoahCompareAndExchangeP:
1451 case Op_ShenandoahCompareAndExchangeN:
1452 // Those instructions would just have stored a different
1453 // value into the field. No use to attempt to fix it at this point.
1454 return phase->igvn().zerocon(T_OBJECT);
1455 case Op_CMoveP:
1456 case Op_CMoveN: {
1457 Node* t = get_load_addr(phase, visited, in->in(CMoveNode::IfTrue));
1458 Node* f = get_load_addr(phase, visited, in->in(CMoveNode::IfFalse));
1459 // Handle unambiguous cases: single address reported on both branches.
1460 if (t != nullptr && f == nullptr) return t;
1461 if (t == nullptr && f != nullptr) return f;
1462 if (t != nullptr && t == f) return t;
1463 // Ambiguity.
1464 return phase->igvn().zerocon(T_OBJECT);
1465 }
1466 case Op_Phi: {
1467 Node* addr = nullptr;
1468 for (uint i = 1; i < in->req(); i++) {
1469 Node* addr1 = get_load_addr(phase, visited, in->in(i));
1470 if (addr == nullptr) {
1471 addr = addr1;
1472 }
1473 if (addr != addr1) {
1474 return phase->igvn().zerocon(T_OBJECT);
1475 }
1476 }
1477 return addr;
1478 }
1479 case Op_ShenandoahLoadReferenceBarrier:
1480 return get_load_addr(phase, visited, in->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1481 case Op_CallDynamicJava:
1482 case Op_CallLeaf:
1483 case Op_CallStaticJava:
1484 case Op_ConN:
1485 case Op_ConP:
1486 case Op_Parm:
1487 case Op_CreateEx:
1488 return phase->igvn().zerocon(T_OBJECT);
1489 default:
1490 #ifdef ASSERT
1491 fatal("Unknown node in get_load_addr: %s", NodeClassNames[in->Opcode()]);
1492 #endif
1493 return phase->igvn().zerocon(T_OBJECT);
1494 }
1495
1496 }
1497
1498 void ShenandoahBarrierC2Support::move_gc_state_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
1499 IdealLoopTree *loop = phase->get_loop(iff);
1500 Node* loop_head = loop->_head;
1501 Node* entry_c = loop_head->in(LoopNode::EntryControl);
1502
1503 Node* bol = iff->in(1);
1504 Node* cmp = bol->in(1);
1505 Node* andi = cmp->in(1);
1506 Node* load = andi->in(1);
1507
1508 assert(is_gc_state_load(load), "broken");
1509 if (!phase->is_dominator(load->in(0), entry_c)) {
1510 Node* mem_ctrl = nullptr;
1511 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
1512 load = load->clone();
1513 load->set_req(MemNode::Memory, mem);
1514 load->set_req(0, entry_c);
1515 phase->register_new_node(load, entry_c);
1516 andi = andi->clone();
1517 andi->set_req(1, load);
1518 phase->register_new_node(andi, entry_c);
1519 cmp = cmp->clone();
1520 cmp->set_req(1, andi);
1521 phase->register_new_node(cmp, entry_c);
1522 bol = bol->clone();
1523 bol->set_req(1, cmp);
1524 phase->register_new_node(bol, entry_c);
1525
1526 phase->igvn().replace_input_of(iff, 1, bol);
1527 }
1528 }
1529
1530 bool ShenandoahBarrierC2Support::identical_backtoback_ifs(Node* n, PhaseIdealLoop* phase) {
1531 if (!n->is_If() || n->is_CountedLoopEnd()) {
1532 return false;
1533 }
1534 Node* region = n->in(0);
1535
1536 if (!region->is_Region()) {
1537 return false;
1538 }
1539 Node* dom = phase->idom(region);
1540 if (!dom->is_If()) {
1541 return false;
1542 }
1543
1544 if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) {
1545 return false;
1546 }
1547
1548 IfNode* dom_if = dom->as_If();
1549 Node* proj_true = dom_if->proj_out(1);
1550 Node* proj_false = dom_if->proj_out(0);
1551
1552 for (uint i = 1; i < region->req(); i++) {
1553 if (phase->is_dominator(proj_true, region->in(i))) {
1554 continue;
1555 }
1556 if (phase->is_dominator(proj_false, region->in(i))) {
1557 continue;
1558 }
1559 return false;
1560 }
1561
1562 return true;
1563 }
1564
1565 bool ShenandoahBarrierC2Support::merge_point_safe(Node* region) {
1566 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1567 Node* n = region->fast_out(i);
1568 if (n->is_LoadStore()) {
1569 // Splitting a LoadStore node through phi, causes it to lose its SCMemProj: the split if code doesn't have support
1570 // for a LoadStore at the region the if is split through because that's not expected to happen (LoadStore nodes
1571 // should be between barrier nodes). It does however happen with Shenandoah though because barriers can get
1572 // expanded around a LoadStore node.
1573 return false;
1574 }
1575 }
1576 return true;
1577 }
1578
1579
1580 void ShenandoahBarrierC2Support::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) {
1581 assert(is_heap_stable_test(n), "no other tests");
1582 if (identical_backtoback_ifs(n, phase)) {
1583 Node* n_ctrl = n->in(0);
1584 if (phase->can_split_if(n_ctrl) && merge_point_safe(n_ctrl)) {
1585 IfNode* dom_if = phase->idom(n_ctrl)->as_If();
1586 if (is_heap_stable_test(n)) {
1587 Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1);
1588 assert(is_gc_state_load(gc_state_load), "broken");
1589 Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1);
1590 assert(is_gc_state_load(dom_gc_state_load), "broken");
1591 if (gc_state_load != dom_gc_state_load) {
1592 phase->igvn().replace_node(gc_state_load, dom_gc_state_load);
1593 }
1594 }
1595 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1596 Node* proj_true = dom_if->proj_out(1);
1597 Node* proj_false = dom_if->proj_out(0);
1598 Node* con_true = phase->igvn().makecon(TypeInt::ONE);
1599 Node* con_false = phase->igvn().makecon(TypeInt::ZERO);
1600
1601 for (uint i = 1; i < n_ctrl->req(); i++) {
1602 if (phase->is_dominator(proj_true, n_ctrl->in(i))) {
1603 bolphi->init_req(i, con_true);
1604 } else {
1605 assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1606 bolphi->init_req(i, con_false);
1607 }
1608 }
1609 phase->register_new_node(bolphi, n_ctrl);
1610 phase->igvn().replace_input_of(n, 1, bolphi);
1611 phase->do_split_if(n);
1612 }
1613 }
1614 }
1615
1616 IfNode* ShenandoahBarrierC2Support::find_unswitching_candidate(const IdealLoopTree* loop, PhaseIdealLoop* phase) {
1617 // Find first invariant test that doesn't exit the loop
1618 LoopNode *head = loop->_head->as_Loop();
1619 IfNode* unswitch_iff = nullptr;
1620 Node* n = head->in(LoopNode::LoopBackControl);
1621 int loop_has_sfpts = -1;
1622 while (n != head) {
1623 Node* n_dom = phase->idom(n);
1624 if (n->is_Region()) {
1625 if (n_dom->is_If()) {
1626 IfNode* iff = n_dom->as_If();
1627 if (iff->in(1)->is_Bool()) {
1628 BoolNode* bol = iff->in(1)->as_Bool();
1629 if (bol->in(1)->is_Cmp()) {
1630 // If condition is invariant and not a loop exit,
1631 // then found reason to unswitch.
1632 if (is_heap_stable_test(iff) &&
1633 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) {
1634 assert(!loop->is_loop_exit(iff), "both branches should be in the loop");
1635 if (loop_has_sfpts == -1) {
1636 for(uint i = 0; i < loop->_body.size(); i++) {
1637 Node *m = loop->_body[i];
1638 if (m->is_SafePoint() && !m->is_CallLeaf()) {
1639 loop_has_sfpts = 1;
1640 break;
1641 }
1642 }
1643 if (loop_has_sfpts == -1) {
1644 loop_has_sfpts = 0;
1645 }
1646 }
1647 if (!loop_has_sfpts) {
1648 unswitch_iff = iff;
1649 }
1650 }
1651 }
1652 }
1653 }
1654 }
1655 n = n_dom;
1656 }
1657 return unswitch_iff;
1658 }
1659
1660
1661 void ShenandoahBarrierC2Support::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) {
1662 Node_List heap_stable_tests;
1663 stack.push(phase->C->start(), 0);
1664 do {
1665 Node* n = stack.node();
1666 uint i = stack.index();
1667
1668 if (i < n->outcnt()) {
1669 Node* u = n->raw_out(i);
1670 stack.set_index(i+1);
1671 if (!visited.test_set(u->_idx)) {
1672 stack.push(u, 0);
1673 }
1674 } else {
1675 stack.pop();
1676 if (n->is_If() && is_heap_stable_test(n)) {
1677 heap_stable_tests.push(n);
1678 }
1679 }
1680 } while (stack.size() > 0);
1681
1682 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1683 Node* n = heap_stable_tests.at(i);
1684 assert(is_heap_stable_test(n), "only evacuation test");
1685 merge_back_to_back_tests(n, phase);
1686 }
1687
1688 if (!phase->C->major_progress()) {
1689 VectorSet seen;
1690 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1691 Node* n = heap_stable_tests.at(i);
1692 IdealLoopTree* loop = phase->get_loop(n);
1693 if (loop != phase->ltree_root() &&
1694 loop->_child == nullptr &&
1695 !loop->_irreducible) {
1696 Node* head = loop->_head;
1697 if (head->is_Loop() &&
1698 (!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) &&
1699 !seen.test_set(head->_idx)) {
1700 IfNode* iff = find_unswitching_candidate(loop, phase);
1701 if (iff != nullptr) {
1702 Node* bol = iff->in(1);
1703 if (head->as_Loop()->is_strip_mined()) {
1704 head->as_Loop()->verify_strip_mined(0);
1705 }
1706 move_gc_state_test_out_of_loop(iff, phase);
1707
1708 AutoNodeBudget node_budget(phase);
1709
1710 if (loop->policy_unswitching(phase)) {
1711 if (head->as_Loop()->is_strip_mined()) {
1712 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop();
1713 hide_strip_mined_loop(outer, head->as_CountedLoop(), phase);
1714 }
1715 phase->do_unswitching(loop, old_new);
1716 } else {
1717 // Not proceeding with unswitching. Move load back in
1718 // the loop.
1719 phase->igvn().replace_input_of(iff, 1, bol);
1720 }
1721 }
1722 }
1723 }
1724 }
1725 }
1726 }
1727
1728 #ifdef ASSERT
1729 static bool has_never_branch(Node* root) {
1730 for (uint i = 1; i < root->req(); i++) {
1731 Node* in = root->in(i);
1732 if (in != nullptr && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->is_NeverBranch()) {
1733 return true;
1734 }
1735 }
1736 return false;
1737 }
1738 #endif
1739
1740 void MemoryGraphFixer::collect_memory_nodes() {
1741 Node_Stack stack(0);
1742 VectorSet visited;
1743 Node_List regions;
1744
1745 // Walk the raw memory graph and create a mapping from CFG node to
1746 // memory node. Exclude phis for now.
1747 stack.push(_phase->C->root(), 1);
1748 do {
1749 Node* n = stack.node();
1750 int opc = n->Opcode();
1751 uint i = stack.index();
1752 if (i < n->req()) {
1753 Node* mem = nullptr;
1754 if (opc == Op_Root) {
1755 Node* in = n->in(i);
1756 int in_opc = in->Opcode();
1757 if (in_opc == Op_Return || in_opc == Op_Rethrow) {
1758 mem = in->in(TypeFunc::Memory);
1759 } else if (in_opc == Op_Halt) {
1760 if (in->in(0)->is_Region()) {
1761 Node* r = in->in(0);
1762 for (uint j = 1; j < r->req(); j++) {
1763 assert(!r->in(j)->is_NeverBranch(), "");
1764 }
1765 } else {
1766 Node* proj = in->in(0);
1767 assert(proj->is_Proj(), "");
1768 Node* in = proj->in(0);
1769 assert(in->is_CallStaticJava() || in->is_NeverBranch() || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
1770 if (in->is_CallStaticJava()) {
1771 mem = in->in(TypeFunc::Memory);
1772 } else if (in->Opcode() == Op_Catch) {
1773 Node* call = in->in(0)->in(0);
1774 assert(call->is_Call(), "");
1775 mem = call->in(TypeFunc::Memory);
1776 } else if (in->is_NeverBranch()) {
1777 mem = collect_memory_for_infinite_loop(in);
1778 }
1779 }
1780 } else {
1781 #ifdef ASSERT
1782 n->dump();
1783 in->dump();
1784 #endif
1785 ShouldNotReachHere();
1786 }
1787 } else {
1788 assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
1789 assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
1790 mem = n->in(i);
1791 }
1792 i++;
1793 stack.set_index(i);
1794 if (mem == nullptr) {
1795 continue;
1796 }
1797 for (;;) {
1798 if (visited.test_set(mem->_idx) || mem->is_Start()) {
1799 break;
1800 }
1801 if (mem->is_Phi()) {
1802 stack.push(mem, 2);
1803 mem = mem->in(1);
1804 } else if (mem->is_Proj()) {
1805 stack.push(mem, mem->req());
1806 mem = mem->in(0);
1807 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
1808 mem = mem->in(TypeFunc::Memory);
1809 } else if (mem->is_MergeMem()) {
1810 MergeMemNode* mm = mem->as_MergeMem();
1811 mem = mm->memory_at(_alias);
1812 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
1813 assert(_alias == Compile::AliasIdxRaw, "");
1814 stack.push(mem, mem->req());
1815 mem = mem->in(MemNode::Memory);
1816 } else {
1817 #ifdef ASSERT
1818 mem->dump();
1819 #endif
1820 ShouldNotReachHere();
1821 }
1822 }
1823 } else {
1824 if (n->is_Phi()) {
1825 // Nothing
1826 } else if (!n->is_Root()) {
1827 Node* c = get_ctrl(n);
1828 _memory_nodes.map(c->_idx, n);
1829 }
1830 stack.pop();
1831 }
1832 } while(stack.is_nonempty());
1833
1834 // Iterate over CFG nodes in rpo and propagate memory state to
1835 // compute memory state at regions, creating new phis if needed.
1836 Node_List rpo_list;
1837 visited.clear();
1838 _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
1839 Node* root = rpo_list.pop();
1840 assert(root == _phase->C->root(), "");
1841
1842 const bool trace = false;
1843 #ifdef ASSERT
1844 if (trace) {
1845 for (int i = rpo_list.size() - 1; i >= 0; i--) {
1846 Node* c = rpo_list.at(i);
1847 if (_memory_nodes[c->_idx] != nullptr) {
1848 tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump();
1849 }
1850 }
1851 }
1852 #endif
1853 uint last = _phase->C->unique();
1854
1855 #ifdef ASSERT
1856 uint16_t max_depth = 0;
1857 for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
1858 IdealLoopTree* lpt = iter.current();
1859 max_depth = MAX2(max_depth, lpt->_nest);
1860 }
1861 #endif
1862
1863 bool progress = true;
1864 int iteration = 0;
1865 Node_List dead_phis;
1866 while (progress) {
1867 progress = false;
1868 iteration++;
1869 assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
1870 if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
1871
1872 for (int i = rpo_list.size() - 1; i >= 0; i--) {
1873 Node* c = rpo_list.at(i);
1874
1875 Node* prev_mem = _memory_nodes[c->_idx];
1876 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
1877 Node* prev_region = regions[c->_idx];
1878 Node* unique = nullptr;
1879 for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
1880 Node* m = _memory_nodes[c->in(j)->_idx];
1881 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");
1882 if (m != nullptr) {
1883 if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
1884 assert((c->is_Loop() && j == LoopNode::LoopBackControl) || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
1885 // continue
1886 } else if (unique == nullptr) {
1887 unique = m;
1888 } else if (m == unique) {
1889 // continue
1890 } else {
1891 unique = NodeSentinel;
1892 }
1893 }
1894 }
1895 assert(unique != nullptr, "empty phi???");
1896 if (unique != NodeSentinel) {
1897 if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c) {
1898 dead_phis.push(prev_region);
1899 }
1900 regions.map(c->_idx, unique);
1901 } else {
1902 Node* phi = nullptr;
1903 if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
1904 phi = prev_region;
1905 for (uint k = 1; k < c->req(); k++) {
1906 Node* m = _memory_nodes[c->in(k)->_idx];
1907 assert(m != nullptr, "expect memory state");
1908 phi->set_req(k, m);
1909 }
1910 } else {
1911 for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == nullptr; j++) {
1912 Node* u = c->fast_out(j);
1913 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
1914 (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
1915 phi = u;
1916 for (uint k = 1; k < c->req() && phi != nullptr; k++) {
1917 Node* m = _memory_nodes[c->in(k)->_idx];
1918 assert(m != nullptr, "expect memory state");
1919 if (u->in(k) != m) {
1920 phi = NodeSentinel;
1921 }
1922 }
1923 }
1924 }
1925 if (phi == NodeSentinel) {
1926 phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
1927 for (uint k = 1; k < c->req(); k++) {
1928 Node* m = _memory_nodes[c->in(k)->_idx];
1929 assert(m != nullptr, "expect memory state");
1930 phi->init_req(k, m);
1931 }
1932 }
1933 }
1934 if (phi != nullptr) {
1935 regions.map(c->_idx, phi);
1936 } else {
1937 assert(c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
1938 }
1939 }
1940 Node* current_region = regions[c->_idx];
1941 if (current_region != prev_region) {
1942 progress = true;
1943 if (prev_region == prev_mem) {
1944 _memory_nodes.map(c->_idx, current_region);
1945 }
1946 }
1947 } else if (prev_mem == nullptr || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
1948 Node* m = _memory_nodes[_phase->idom(c)->_idx];
1949 assert(m != nullptr || c->Opcode() == Op_Halt, "expect memory state");
1950 if (m != prev_mem) {
1951 _memory_nodes.map(c->_idx, m);
1952 progress = true;
1953 }
1954 }
1955 #ifdef ASSERT
1956 if (trace) { tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); }
1957 #endif
1958 }
1959 }
1960
1961 // Replace existing phi with computed memory state for that region
1962 // if different (could be a new phi or a dominating memory node if
1963 // that phi was found to be useless).
1964 while (dead_phis.size() > 0) {
1965 Node* n = dead_phis.pop();
1966 n->replace_by(_phase->C->top());
1967 n->destruct(&_phase->igvn());
1968 }
1969 for (int i = rpo_list.size() - 1; i >= 0; i--) {
1970 Node* c = rpo_list.at(i);
1971 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
1972 Node* n = regions[c->_idx];
1973 assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
1974 if (n != nullptr && n->is_Phi() && n->_idx >= last && n->in(0) == c) {
1975 _phase->register_new_node(n, c);
1976 }
1977 }
1978 }
1979 for (int i = rpo_list.size() - 1; i >= 0; i--) {
1980 Node* c = rpo_list.at(i);
1981 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
1982 Node* n = regions[c->_idx];
1983 assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
1984 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
1985 Node* u = c->fast_out(i);
1986 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
1987 u != n) {
1988 assert(c->unique_ctrl_out()->Opcode() != Op_Halt, "expected memory state");
1989 if (u->adr_type() == TypePtr::BOTTOM) {
1990 fix_memory_uses(u, n, n, c);
1991 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
1992 _phase->lazy_replace(u, n);
1993 --i; --imax;
1994 }
1995 }
1996 }
1997 }
1998 }
1999 }
2000
2001 Node* MemoryGraphFixer::collect_memory_for_infinite_loop(const Node* in) {
2002 Node* mem = nullptr;
2003 Node* head = in->in(0);
2004 assert(head->is_Region(), "unexpected infinite loop graph shape");
2005
2006 Node* phi_mem = nullptr;
2007 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2008 Node* u = head->fast_out(j);
2009 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
2010 if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2011 assert(phi_mem == nullptr || phi_mem->adr_type() == TypePtr::BOTTOM, "");
2012 phi_mem = u;
2013 } else if (u->adr_type() == TypePtr::BOTTOM) {
2014 assert(phi_mem == nullptr || _phase->C->get_alias_index(phi_mem->adr_type()) == _alias, "");
2015 if (phi_mem == nullptr) {
2016 phi_mem = u;
2017 }
2018 }
2019 }
2020 }
2021 if (phi_mem == nullptr) {
2022 ResourceMark rm;
2023 Node_Stack stack(0);
2024 stack.push(head, 1);
2025 do {
2026 Node* n = stack.node();
2027 uint i = stack.index();
2028 if (i >= n->req()) {
2029 stack.pop();
2030 } else {
2031 stack.set_index(i + 1);
2032 Node* c = n->in(i);
2033 assert(c != head, "should have found a safepoint on the way");
2034 if (stack.size() != 1 || _phase->is_dominator(head, c)) {
2035 for (;;) {
2036 if (c->is_Region()) {
2037 stack.push(c, 1);
2038 break;
2039 } else if (c->is_SafePoint() && !c->is_CallLeaf()) {
2040 Node* m = c->in(TypeFunc::Memory);
2041 if (m->is_MergeMem()) {
2042 m = m->as_MergeMem()->memory_at(_alias);
2043 }
2044 assert(mem == nullptr || mem == m, "several memory states");
2045 mem = m;
2046 break;
2047 } else {
2048 assert(c != c->in(0), "");
2049 c = c->in(0);
2050 }
2051 }
2052 }
2053 }
2054 } while (stack.size() > 0);
2055 assert(mem != nullptr, "should have found safepoint");
2056 } else {
2057 mem = phi_mem;
2058 }
2059 return mem;
2060 }
2061
2062 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
2063 Node* c = _phase->get_ctrl(n);
2064 if (n->is_Proj() && n->in(0) != nullptr && n->in(0)->is_Call()) {
2065 assert(c == n->in(0), "");
2066 CallNode* call = c->as_Call();
2067 CallProjections projs;
2068 call->extract_projections(&projs, true, false);
2069 if (projs.catchall_memproj != nullptr) {
2070 if (projs.fallthrough_memproj == n) {
2071 c = projs.fallthrough_catchproj;
2072 } else {
2073 assert(projs.catchall_memproj == n, "");
2074 c = projs.catchall_catchproj;
2075 }
2076 }
2077 }
2078 return c;
2079 }
2080
2081 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
2082 if (_phase->has_ctrl(n))
2083 return get_ctrl(n);
2084 else {
2085 assert (n->is_CFG(), "must be a CFG node");
2086 return n;
2087 }
2088 }
2089
2090 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
2091 return m != nullptr && get_ctrl(m) == c;
2092 }
2093
2094 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
2095 assert(n == nullptr || _phase->ctrl_or_self(n) == ctrl, "");
2096 assert(!ctrl->is_Call() || ctrl == n, "projection expected");
2097 #ifdef ASSERT
2098 if ((ctrl->is_Proj() && ctrl->in(0)->is_Call()) ||
2099 (ctrl->is_Catch() && ctrl->in(0)->in(0)->is_Call())) {
2100 CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_Call() : ctrl->in(0)->in(0)->as_Call();
2101 int mems = 0;
2102 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
2103 Node* u = call->fast_out(i);
2104 if (u->bottom_type() == Type::MEMORY) {
2105 mems++;
2106 }
2107 }
2108 assert(mems <= 1, "No node right after call if multiple mem projections");
2109 }
2110 #endif
2111 Node* mem = _memory_nodes[ctrl->_idx];
2112 Node* c = ctrl;
2113 while (!mem_is_valid(mem, c) &&
2114 (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
2115 c = _phase->idom(c);
2116 mem = _memory_nodes[c->_idx];
2117 }
2118 if (n != nullptr && mem_is_valid(mem, c)) {
2119 while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
2120 mem = next_mem(mem, _alias);
2121 }
2122 if (mem->is_MergeMem()) {
2123 mem = mem->as_MergeMem()->memory_at(_alias);
2124 }
2125 if (!mem_is_valid(mem, c)) {
2126 do {
2127 c = _phase->idom(c);
2128 mem = _memory_nodes[c->_idx];
2129 } while (!mem_is_valid(mem, c) &&
2130 (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
2131 }
2132 }
2133 assert(mem->bottom_type() == Type::MEMORY, "");
2134 return mem;
2135 }
2136
2137 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
2138 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2139 Node* use = region->fast_out(i);
2140 if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
2141 (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
2142 return true;
2143 }
2144 }
2145 return false;
2146 }
2147
2148 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
2149 assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
2150 const bool trace = false;
2151 DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
2152 DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
2153 GrowableArray<Node*> phis;
2154 if (mem_for_ctrl != mem) {
2155 Node* old = mem_for_ctrl;
2156 Node* prev = nullptr;
2157 while (old != mem) {
2158 prev = old;
2159 if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
2160 assert(_alias == Compile::AliasIdxRaw, "");
2161 old = old->in(MemNode::Memory);
2162 } else if (old->Opcode() == Op_SCMemProj) {
2163 assert(_alias == Compile::AliasIdxRaw, "");
2164 old = old->in(0);
2165 } else {
2166 ShouldNotReachHere();
2167 }
2168 }
2169 assert(prev != nullptr, "");
2170 if (new_ctrl != ctrl) {
2171 _memory_nodes.map(ctrl->_idx, mem);
2172 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2173 }
2174 uint input = (uint)MemNode::Memory;
2175 _phase->igvn().replace_input_of(prev, input, new_mem);
2176 } else {
2177 uses.clear();
2178 _memory_nodes.map(new_ctrl->_idx, new_mem);
2179 uses.push(new_ctrl);
2180 for(uint next = 0; next < uses.size(); next++ ) {
2181 Node *n = uses.at(next);
2182 assert(n->is_CFG(), "");
2183 DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
2184 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2185 Node* u = n->fast_out(i);
2186 if (!u->is_Root() && u->is_CFG() && u != n) {
2187 Node* m = _memory_nodes[u->_idx];
2188 if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
2189 !has_mem_phi(u) &&
2190 u->unique_ctrl_out()->Opcode() != Op_Halt) {
2191 DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
2192 DEBUG_ONLY(if (trace && m != nullptr) { tty->print("ZZZ mem"); m->dump(); });
2193
2194 if (!mem_is_valid(m, u) || !m->is_Phi()) {
2195 bool push = true;
2196 bool create_phi = true;
2197 if (_phase->is_dominator(new_ctrl, u)) {
2198 create_phi = false;
2199 }
2200 if (create_phi) {
2201 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
2202 _phase->register_new_node(phi, u);
2203 phis.push(phi);
2204 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
2205 if (!mem_is_valid(m, u)) {
2206 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
2207 _memory_nodes.map(u->_idx, phi);
2208 } else {
2209 DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
2210 for (;;) {
2211 assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), "");
2212 Node* next = nullptr;
2213 if (m->is_Proj()) {
2214 next = m->in(0);
2215 } else {
2216 assert(m->is_Mem() || m->is_LoadStore(), "");
2217 assert(_alias == Compile::AliasIdxRaw, "");
2218 next = m->in(MemNode::Memory);
2219 }
2220 if (_phase->get_ctrl(next) != u) {
2221 break;
2222 }
2223 if (next->is_MergeMem()) {
2224 assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
2225 break;
2226 }
2227 if (next->is_Phi()) {
2228 assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
2229 break;
2230 }
2231 m = next;
2232 }
2233
2234 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
2235 assert(m->is_Mem() || m->is_LoadStore(), "");
2236 uint input = (uint)MemNode::Memory;
2237 _phase->igvn().replace_input_of(m, input, phi);
2238 push = false;
2239 }
2240 } else {
2241 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
2242 }
2243 if (push) {
2244 uses.push(u);
2245 }
2246 }
2247 } else if (!mem_is_valid(m, u) &&
2248 !(u->Opcode() == Op_CProj && u->in(0)->is_NeverBranch() && u->as_Proj()->_con == 1)) {
2249 uses.push(u);
2250 }
2251 }
2252 }
2253 }
2254 for (int i = 0; i < phis.length(); i++) {
2255 Node* n = phis.at(i);
2256 Node* r = n->in(0);
2257 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
2258 for (uint j = 1; j < n->req(); j++) {
2259 Node* m = find_mem(r->in(j), nullptr);
2260 _phase->igvn().replace_input_of(n, j, m);
2261 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
2262 }
2263 }
2264 }
2265 uint last = _phase->C->unique();
2266 MergeMemNode* mm = nullptr;
2267 int alias = _alias;
2268 DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
2269 // Process loads first to not miss an anti-dependency: if the memory
2270 // edge of a store is updated before a load is processed then an
2271 // anti-dependency may be missed.
2272 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2273 Node* u = mem->out(i);
2274 if (u->_idx < last && u->is_Load() && _phase->C->get_alias_index(u->adr_type()) == alias) {
2275 Node* m = find_mem(_phase->get_ctrl(u), u);
2276 if (m != mem) {
2277 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2278 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2279 --i;
2280 }
2281 }
2282 }
2283 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2284 Node* u = mem->out(i);
2285 if (u->_idx < last) {
2286 if (u->is_Mem()) {
2287 if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2288 Node* m = find_mem(_phase->get_ctrl(u), u);
2289 if (m != mem) {
2290 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2291 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2292 --i;
2293 }
2294 }
2295 } else if (u->is_MergeMem()) {
2296 MergeMemNode* u_mm = u->as_MergeMem();
2297 if (u_mm->memory_at(alias) == mem) {
2298 MergeMemNode* newmm = nullptr;
2299 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2300 Node* uu = u->fast_out(j);
2301 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2302 if (uu->is_Phi()) {
2303 assert(uu->adr_type() == TypePtr::BOTTOM, "");
2304 Node* region = uu->in(0);
2305 int nb = 0;
2306 for (uint k = 1; k < uu->req(); k++) {
2307 if (uu->in(k) == u) {
2308 Node* m = find_mem(region->in(k), nullptr);
2309 if (m != mem) {
2310 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
2311 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2312 if (newmm != u) {
2313 _phase->igvn().replace_input_of(uu, k, newmm);
2314 nb++;
2315 --jmax;
2316 }
2317 }
2318 }
2319 }
2320 if (nb > 0) {
2321 --j;
2322 }
2323 } else {
2324 Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
2325 if (m != mem) {
2326 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
2327 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2328 if (newmm != u) {
2329 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2330 --j, --jmax;
2331 }
2332 }
2333 }
2334 }
2335 }
2336 } else if (u->is_Phi()) {
2337 assert(u->bottom_type() == Type::MEMORY, "what else?");
2338 if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
2339 Node* region = u->in(0);
2340 bool replaced = false;
2341 for (uint j = 1; j < u->req(); j++) {
2342 if (u->in(j) == mem) {
2343 Node* m = find_mem(region->in(j), nullptr);
2344 Node* nnew = m;
2345 if (m != mem) {
2346 if (u->adr_type() == TypePtr::BOTTOM) {
2347 mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
2348 nnew = mm;
2349 }
2350 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
2351 _phase->igvn().replace_input_of(u, j, nnew);
2352 replaced = true;
2353 }
2354 }
2355 }
2356 if (replaced) {
2357 --i;
2358 }
2359 }
2360 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2361 u->adr_type() == nullptr) {
2362 assert(u->adr_type() != nullptr ||
2363 u->Opcode() == Op_Rethrow ||
2364 u->Opcode() == Op_Return ||
2365 u->Opcode() == Op_SafePoint ||
2366 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2367 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2368 u->Opcode() == Op_CallLeaf, "");
2369 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2370 if (m != mem) {
2371 mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
2372 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2373 --i;
2374 }
2375 } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2376 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2377 if (m != mem) {
2378 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2379 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2380 --i;
2381 }
2382 } else if (u->adr_type() != TypePtr::BOTTOM &&
2383 _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
2384 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2385 assert(m != mem, "");
2386 // u is on the wrong slice...
2387 assert(u->is_ClearArray(), "");
2388 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2389 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2390 --i;
2391 }
2392 }
2393 }
2394 #ifdef ASSERT
2395 assert(new_mem->outcnt() > 0, "");
2396 for (int i = 0; i < phis.length(); i++) {
2397 Node* n = phis.at(i);
2398 assert(n->outcnt() > 0, "new phi must have uses now");
2399 }
2400 #endif
2401 }
2402
2403 void MemoryGraphFixer::record_new_ctrl(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl) {
2404 if (mem_for_ctrl != mem && new_ctrl != ctrl) {
2405 _memory_nodes.map(ctrl->_idx, mem);
2406 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2407 }
2408 }
2409
2410 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
2411 MergeMemNode* mm = MergeMemNode::make(mem);
2412 mm->set_memory_at(_alias, rep_proj);
2413 _phase->register_new_node(mm, rep_ctrl);
2414 return mm;
2415 }
2416
2417 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
2418 MergeMemNode* newmm = nullptr;
2419 MergeMemNode* u_mm = u->as_MergeMem();
2420 Node* c = _phase->get_ctrl(u);
2421 if (_phase->is_dominator(c, rep_ctrl)) {
2422 c = rep_ctrl;
2423 } else {
2424 assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
2425 }
2426 if (u->outcnt() == 1) {
2427 if (u->req() > (uint)_alias && u->in(_alias) == mem) {
2428 _phase->igvn().replace_input_of(u, _alias, rep_proj);
2429 --i;
2430 } else {
2431 _phase->igvn().rehash_node_delayed(u);
2432 u_mm->set_memory_at(_alias, rep_proj);
2433 }
2434 newmm = u_mm;
2435 _phase->set_ctrl_and_loop(u, c);
2436 } else {
2437 // can't simply clone u and then change one of its input because
2438 // it adds and then removes an edge which messes with the
2439 // DUIterator
2440 newmm = MergeMemNode::make(u_mm->base_memory());
2441 for (uint j = 0; j < u->req(); j++) {
2442 if (j < newmm->req()) {
2443 if (j == (uint)_alias) {
2444 newmm->set_req(j, rep_proj);
2445 } else if (newmm->in(j) != u->in(j)) {
2446 newmm->set_req(j, u->in(j));
2447 }
2448 } else if (j == (uint)_alias) {
2449 newmm->add_req(rep_proj);
2450 } else {
2451 newmm->add_req(u->in(j));
2452 }
2453 }
2454 if ((uint)_alias >= u->req()) {
2455 newmm->set_memory_at(_alias, rep_proj);
2456 }
2457 _phase->register_new_node(newmm, c);
2458 }
2459 return newmm;
2460 }
2461
2462 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
2463 if (phi->adr_type() == TypePtr::BOTTOM) {
2464 Node* region = phi->in(0);
2465 for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
2466 Node* uu = region->fast_out(j);
2467 if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
2468 return false;
2469 }
2470 }
2471 return true;
2472 }
2473 return _phase->C->get_alias_index(phi->adr_type()) == _alias;
2474 }
2475
2476 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
2477 uint last = _phase-> C->unique();
2478 MergeMemNode* mm = nullptr;
2479 assert(mem->bottom_type() == Type::MEMORY, "");
2480 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2481 Node* u = mem->out(i);
2482 if (u != replacement && u->_idx < last) {
2483 if (u->is_MergeMem()) {
2484 MergeMemNode* u_mm = u->as_MergeMem();
2485 if (u_mm->memory_at(_alias) == mem) {
2486 MergeMemNode* newmm = nullptr;
2487 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2488 Node* uu = u->fast_out(j);
2489 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2490 if (uu->is_Phi()) {
2491 if (should_process_phi(uu)) {
2492 Node* region = uu->in(0);
2493 int nb = 0;
2494 for (uint k = 1; k < uu->req(); k++) {
2495 if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
2496 if (newmm == nullptr) {
2497 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2498 }
2499 if (newmm != u) {
2500 _phase->igvn().replace_input_of(uu, k, newmm);
2501 nb++;
2502 --jmax;
2503 }
2504 }
2505 }
2506 if (nb > 0) {
2507 --j;
2508 }
2509 }
2510 } else {
2511 if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
2512 if (newmm == nullptr) {
2513 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2514 }
2515 if (newmm != u) {
2516 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2517 --j, --jmax;
2518 }
2519 }
2520 }
2521 }
2522 }
2523 } else if (u->is_Phi()) {
2524 assert(u->bottom_type() == Type::MEMORY, "what else?");
2525 Node* region = u->in(0);
2526 if (should_process_phi(u)) {
2527 bool replaced = false;
2528 for (uint j = 1; j < u->req(); j++) {
2529 if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
2530 Node* nnew = rep_proj;
2531 if (u->adr_type() == TypePtr::BOTTOM) {
2532 if (mm == nullptr) {
2533 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2534 }
2535 nnew = mm;
2536 }
2537 _phase->igvn().replace_input_of(u, j, nnew);
2538 replaced = true;
2539 }
2540 }
2541 if (replaced) {
2542 --i;
2543 }
2544
2545 }
2546 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2547 u->adr_type() == nullptr) {
2548 assert(u->adr_type() != nullptr ||
2549 u->Opcode() == Op_Rethrow ||
2550 u->Opcode() == Op_Return ||
2551 u->Opcode() == Op_SafePoint ||
2552 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2553 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2554 u->Opcode() == Op_CallLeaf, "%s", u->Name());
2555 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2556 if (mm == nullptr) {
2557 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2558 }
2559 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2560 --i;
2561 }
2562 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2563 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2564 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
2565 --i;
2566 }
2567 }
2568 }
2569 }
2570 }
2571
2572 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj, DecoratorSet decorators)
2573 : Node(ctrl, obj), _decorators(decorators) {
2574 ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this);
2575 }
2576
2577 DecoratorSet ShenandoahLoadReferenceBarrierNode::decorators() const {
2578 return _decorators;
2579 }
2580
2581 uint ShenandoahLoadReferenceBarrierNode::size_of() const {
2582 return sizeof(*this);
2583 }
2584
2585 static DecoratorSet mask_decorators(DecoratorSet decorators) {
2586 return decorators & (ON_STRONG_OOP_REF | ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF | ON_UNKNOWN_OOP_REF | IN_NATIVE);
2587 }
2588
2589 uint ShenandoahLoadReferenceBarrierNode::hash() const {
2590 uint hash = Node::hash();
2591 hash += mask_decorators(_decorators);
2592 return hash;
2593 }
2594
2595 bool ShenandoahLoadReferenceBarrierNode::cmp( const Node &n ) const {
2596 return Node::cmp(n) && n.Opcode() == Op_ShenandoahLoadReferenceBarrier &&
2597 mask_decorators(_decorators) == mask_decorators(((const ShenandoahLoadReferenceBarrierNode&)n)._decorators);
2598 }
2599
2600 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const {
2601 if (in(ValueIn) == nullptr || in(ValueIn)->is_top()) {
2602 return Type::TOP;
2603 }
2604 const Type* t = in(ValueIn)->bottom_type();
2605 if (t == TypePtr::NULL_PTR) {
2606 return t;
2607 }
2608
2609 if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2610 return t;
2611 }
2612
2613 return t->meet(TypePtr::NULL_PTR);
2614 }
2615
2616 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const {
2617 // Either input is TOP ==> the result is TOP
2618 const Type *t2 = phase->type(in(ValueIn));
2619 if( t2 == Type::TOP ) return Type::TOP;
2620
2621 if (t2 == TypePtr::NULL_PTR) {
2622 return t2;
2623 }
2624
2625 if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2626 return t2;
2627 }
2628
2629 return t2->meet(TypePtr::NULL_PTR);
2630 }
2631
2632 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) {
2633 Node* value = in(ValueIn);
2634 if (!needs_barrier(phase, value)) {
2635 return value;
2636 }
2637 return this;
2638 }
2639
2640 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) {
2641 Unique_Node_List visited;
2642 return needs_barrier_impl(phase, n, visited);
2643 }
2644
2645 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) {
2646 if (n == nullptr) return false;
2647 if (visited.member(n)) {
2648 return false; // Been there.
2649 }
2650 visited.push(n);
2651
2652 if (n->is_Allocate()) {
2653 // tty->print_cr("optimize barrier on alloc");
2654 return false;
2655 }
2656 if (n->is_Call()) {
2657 // tty->print_cr("optimize barrier on call");
2658 return false;
2659 }
2660
2661 const Type* type = phase->type(n);
2662 if (type == Type::TOP) {
2663 return false;
2664 }
2665 if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) {
2666 // tty->print_cr("optimize barrier on null");
2667 return false;
2668 }
2669 if (type->make_oopptr() && type->make_oopptr()->const_oop() != nullptr) {
2670 // tty->print_cr("optimize barrier on constant");
2671 return false;
2672 }
2673
2674 switch (n->Opcode()) {
2675 case Op_AddP:
2676 return true; // TODO: Can refine?
2677 case Op_LoadP:
2678 case Op_ShenandoahCompareAndExchangeN:
2679 case Op_ShenandoahCompareAndExchangeP:
2680 case Op_CompareAndExchangeN:
2681 case Op_CompareAndExchangeP:
2682 case Op_GetAndSetN:
2683 case Op_GetAndSetP:
2684 return true;
2685 case Op_Phi: {
2686 for (uint i = 1; i < n->req(); i++) {
2687 if (needs_barrier_impl(phase, n->in(i), visited)) return true;
2688 }
2689 return false;
2690 }
2691 case Op_CheckCastPP:
2692 case Op_CastPP:
2693 return needs_barrier_impl(phase, n->in(1), visited);
2694 case Op_Proj:
2695 return needs_barrier_impl(phase, n->in(0), visited);
2696 case Op_ShenandoahLoadReferenceBarrier:
2697 // tty->print_cr("optimize barrier on barrier");
2698 return false;
2699 case Op_Parm:
2700 // tty->print_cr("optimize barrier on input arg");
2701 return false;
2702 case Op_DecodeN:
2703 case Op_EncodeP:
2704 return needs_barrier_impl(phase, n->in(1), visited);
2705 case Op_LoadN:
2706 return true;
2707 case Op_CMoveN:
2708 case Op_CMoveP:
2709 return needs_barrier_impl(phase, n->in(2), visited) ||
2710 needs_barrier_impl(phase, n->in(3), visited);
2711 case Op_CreateEx:
2712 return false;
2713 default:
2714 break;
2715 }
2716 #ifdef ASSERT
2717 tty->print("need barrier on?: ");
2718 tty->print_cr("ins:");
2719 n->dump(2);
2720 tty->print_cr("outs:");
2721 n->dump(-2);
2722 ShouldNotReachHere();
2723 #endif
2724 return true;
2725 }