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->iu_barriers_count() +
50 state->load_reference_barriers_count()) > 0) {
51 assert(C->post_loop_opts_phase(), "no loop opts allowed");
52 C->reset_post_loop_opts_phase(); // ... but we know what we are doing
53 C->clear_major_progress();
54 PhaseIdealLoop::optimize(igvn, LoopOptsShenandoahExpand);
55 if (C->failing()) return false;
56
57 C->set_major_progress();
58 if (!C->optimize_loops(igvn, LoopOptsShenandoahPostExpand)) {
59 return false;
60 }
61 C->clear_major_progress();
62 C->process_for_post_loop_opts_igvn(igvn);
63 if (C->failing()) return false;
64
65 C->set_post_loop_opts_phase(); // now for real!
66 }
67 return true;
68 }
69
70 bool ShenandoahBarrierC2Support::is_gc_state_test(Node* iff, int mask) {
71 if (!UseShenandoahGC) {
72 return false;
73 }
74 assert(iff->is_If(), "bad input");
75 if (iff->Opcode() != Op_If) {
76 return false;
77 }
78 Node* bol = iff->in(1);
79 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
80 return false;
81 }
82 Node* cmp = bol->in(1);
83 if (cmp->Opcode() != Op_CmpI) {
84 return false;
85 }
86 Node* in1 = cmp->in(1);
87 Node* in2 = cmp->in(2);
88 if (in2->find_int_con(-1) != 0) {
89 return false;
90 }
91 if (in1->Opcode() != Op_AndI) {
92 return false;
93 }
94 in2 = in1->in(2);
95 if (in2->find_int_con(-1) != mask) {
96 return false;
97 }
98 in1 = in1->in(1);
99
100 return is_gc_state_load(in1);
101 }
102
103 bool ShenandoahBarrierC2Support::is_heap_stable_test(Node* iff) {
104 return is_gc_state_test(iff, ShenandoahHeap::HAS_FORWARDED);
105 }
106
107 bool ShenandoahBarrierC2Support::is_gc_state_load(Node *n) {
108 if (!UseShenandoahGC) {
109 return false;
110 }
111 if (n->Opcode() != Op_LoadB && n->Opcode() != Op_LoadUB) {
112 return false;
113 }
114 Node* addp = n->in(MemNode::Address);
115 if (!addp->is_AddP()) {
116 return false;
117 }
118 Node* base = addp->in(AddPNode::Address);
119 Node* off = addp->in(AddPNode::Offset);
120 if (base->Opcode() != Op_ThreadLocal) {
121 return false;
122 }
123 if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) {
124 return false;
125 }
126 return true;
127 }
128
129 bool ShenandoahBarrierC2Support::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) {
130 assert(phase->is_dominator(stop, start), "bad inputs");
131 ResourceMark rm;
132 Unique_Node_List wq;
133 wq.push(start);
134 for (uint next = 0; next < wq.size(); next++) {
135 Node *m = wq.at(next);
136 if (m == stop) {
137 continue;
138 }
139 if (m->is_SafePoint() && !m->is_CallLeaf()) {
140 return true;
141 }
142 if (m->is_Region()) {
143 for (uint i = 1; i < m->req(); i++) {
144 wq.push(m->in(i));
145 }
146 } else {
147 wq.push(m->in(0));
148 }
149 }
150 return false;
151 }
152
153 #ifdef ASSERT
154 bool ShenandoahBarrierC2Support::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) {
155 assert(phis.size() == 0, "");
156
157 while (true) {
158 if (in->bottom_type() == TypePtr::NULL_PTR) {
159 if (trace) {tty->print_cr("null");}
160 } else if (!in->bottom_type()->make_ptr()->make_oopptr()) {
161 if (trace) {tty->print_cr("Non oop");}
162 } else {
163 if (in->is_ConstraintCast()) {
164 in = in->in(1);
165 continue;
166 } else if (in->is_AddP()) {
167 assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access");
168 in = in->in(AddPNode::Address);
169 continue;
170 } else if (in->is_Con()) {
171 if (trace) {
172 tty->print("Found constant");
173 in->dump();
174 }
175 } else if (in->Opcode() == Op_Parm) {
176 if (trace) {
177 tty->print("Found argument");
178 }
179 } else if (in->Opcode() == Op_CreateEx) {
180 if (trace) {
181 tty->print("Found create-exception");
182 }
183 } else if (in->Opcode() == Op_LoadP && in->adr_type() == TypeRawPtr::BOTTOM) {
184 if (trace) {
185 tty->print("Found raw LoadP (OSR argument?)");
186 }
187 } else if (in->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
188 if (t == ShenandoahOopStore) {
189 uint i = 0;
190 for (; i < phis.size(); i++) {
191 Node* n = phis.node_at(i);
192 if (n->Opcode() == Op_ShenandoahIUBarrier) {
193 break;
194 }
195 }
196 if (i == phis.size()) {
197 return false;
198 }
199 }
200 barriers_used.push(in);
201 if (trace) {tty->print("Found barrier"); in->dump();}
202 } else if (in->Opcode() == Op_ShenandoahIUBarrier) {
203 if (t != ShenandoahOopStore) {
204 in = in->in(1);
205 continue;
206 }
207 if (trace) {tty->print("Found enqueue barrier"); in->dump();}
208 phis.push(in, in->req());
209 in = in->in(1);
210 continue;
211 } else if (in->is_Proj() && in->in(0)->is_Allocate()) {
212 if (trace) {
213 tty->print("Found alloc");
214 in->in(0)->dump();
215 }
216 } else if (in->is_Proj() && (in->in(0)->Opcode() == Op_CallStaticJava || in->in(0)->Opcode() == Op_CallDynamicJava)) {
217 if (trace) {
218 tty->print("Found Java call");
219 }
220 } else if (in->is_Phi()) {
221 if (!visited.test_set(in->_idx)) {
222 if (trace) {tty->print("Pushed phi:"); in->dump();}
223 phis.push(in, 2);
224 in = in->in(1);
225 continue;
226 }
227 if (trace) {tty->print("Already seen phi:"); in->dump();}
228 } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) {
229 if (!visited.test_set(in->_idx)) {
230 if (trace) {tty->print("Pushed cmovep:"); in->dump();}
231 phis.push(in, CMoveNode::IfTrue);
232 in = in->in(CMoveNode::IfFalse);
233 continue;
234 }
235 if (trace) {tty->print("Already seen cmovep:"); in->dump();}
236 } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) {
237 in = in->in(1);
238 continue;
239 } else {
240 return false;
241 }
242 }
243 bool cont = false;
244 while (phis.is_nonempty()) {
245 uint idx = phis.index();
246 Node* phi = phis.node();
247 if (idx >= phi->req()) {
248 if (trace) {tty->print("Popped phi:"); phi->dump();}
249 phis.pop();
250 continue;
251 }
252 if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();}
253 in = phi->in(idx);
254 phis.set_index(idx+1);
255 cont = true;
256 break;
257 }
258 if (!cont) {
259 break;
260 }
261 }
262 return true;
263 }
264
265 void ShenandoahBarrierC2Support::report_verify_failure(const char* msg, Node* n1, Node* n2) {
266 if (n1 != nullptr) {
267 n1->dump(+10);
268 }
269 if (n2 != nullptr) {
270 n2->dump(+10);
271 }
272 fatal("%s", msg);
273 }
274
275 void ShenandoahBarrierC2Support::verify(RootNode* root) {
276 ResourceMark rm;
277 Unique_Node_List wq;
278 GrowableArray<Node*> barriers;
279 Unique_Node_List barriers_used;
280 Node_Stack phis(0);
281 VectorSet visited;
282 const bool trace = false;
283 const bool verify_no_useless_barrier = false;
284
285 wq.push(root);
286 for (uint next = 0; next < wq.size(); next++) {
287 Node *n = wq.at(next);
288 if (n->is_Load()) {
289 const bool trace = false;
290 if (trace) {tty->print("Verifying"); n->dump();}
291 if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) {
292 if (trace) {tty->print_cr("Load range/klass");}
293 } else {
294 const TypePtr* adr_type = n->as_Load()->adr_type();
295
296 if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) {
297 if (trace) {tty->print_cr("Mark load");}
298 } else if (adr_type->isa_instptr() &&
299 adr_type->is_instptr()->instance_klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
300 adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset()) {
301 if (trace) {tty->print_cr("Reference.get()");}
302 } else if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) {
303 report_verify_failure("Shenandoah verification: Load should have barriers", n);
304 }
305 }
306 } else if (n->is_Store()) {
307 const bool trace = false;
308
309 if (trace) {tty->print("Verifying"); n->dump();}
310 if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) {
311 Node* adr = n->in(MemNode::Address);
312 bool verify = true;
313
314 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) {
315 adr = adr->in(AddPNode::Address);
316 if (adr->is_AddP()) {
317 assert(adr->in(AddPNode::Base)->is_top(), "");
318 adr = adr->in(AddPNode::Address);
319 if (adr->Opcode() == Op_LoadP &&
320 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() &&
321 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
322 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) {
323 if (trace) {tty->print_cr("SATB prebarrier");}
324 verify = false;
325 }
326 }
327 }
328
329 if (verify && !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahIUBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
330 report_verify_failure("Shenandoah verification: Store should have barriers", n);
331 }
332 }
333 if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
334 report_verify_failure("Shenandoah verification: Store (address) should have barriers", n);
335 }
336 } else if (n->Opcode() == Op_CmpP) {
337 const bool trace = false;
338
339 Node* in1 = n->in(1);
340 Node* in2 = n->in(2);
341 if (in1->bottom_type()->isa_oopptr()) {
342 if (trace) {tty->print("Verifying"); n->dump();}
343
344 bool mark_inputs = false;
345 if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR ||
346 (in1->is_Con() || in2->is_Con())) {
347 if (trace) {tty->print_cr("Comparison against a constant");}
348 mark_inputs = true;
349 } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) ||
350 (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) {
351 if (trace) {tty->print_cr("Comparison with newly alloc'ed object");}
352 mark_inputs = true;
353 } else {
354 assert(in2->bottom_type()->isa_oopptr(), "");
355
356 if (!verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) ||
357 !verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) {
358 report_verify_failure("Shenandoah verification: Cmp should have barriers", n);
359 }
360 }
361 if (verify_no_useless_barrier &&
362 mark_inputs &&
363 (!verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) ||
364 !verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) {
365 phis.clear();
366 visited.reset();
367 }
368 }
369 } else if (n->is_LoadStore()) {
370 if (n->in(MemNode::ValueIn)->bottom_type()->make_ptr() &&
371 !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahIUBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
372 report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n);
373 }
374
375 if (n->in(MemNode::Address)->bottom_type()->make_oopptr() && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
376 report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n);
377 }
378 } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) {
379 CallNode* call = n->as_Call();
380
381 static struct {
382 const char* name;
383 struct {
384 int pos;
385 verify_type t;
386 } args[6];
387 } calls[] = {
388 "array_partition_stub",
389 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+4, ShenandoahStore }, { -1, ShenandoahNone },
390 { -1, ShenandoahNone }, { -1, ShenandoahNone }, { -1, ShenandoahNone } },
391 "arraysort_stub",
392 { { TypeFunc::Parms, ShenandoahStore }, { -1, ShenandoahNone }, { -1, ShenandoahNone },
393 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
394 "aescrypt_encryptBlock",
395 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
396 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
397 "aescrypt_decryptBlock",
398 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
399 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
400 "multiplyToLen",
401 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { TypeFunc::Parms+4, ShenandoahStore },
402 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
403 "squareToLen",
404 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { -1, ShenandoahNone},
405 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
406 "montgomery_multiply",
407 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
408 { TypeFunc::Parms+6, ShenandoahStore }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
409 "montgomery_square",
410 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+5, ShenandoahStore },
411 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
412 "mulAdd",
413 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
414 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
415 "vectorizedMismatch",
416 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
417 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
418 "updateBytesCRC32",
419 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
420 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
421 "updateBytesAdler32",
422 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
423 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
424 "updateBytesCRC32C",
425 { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad}, { -1, ShenandoahNone},
426 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
427 "counterMode_AESCrypt",
428 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
429 { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } },
430 "cipherBlockChaining_encryptAESCrypt",
431 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
432 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
433 "cipherBlockChaining_decryptAESCrypt",
434 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
435 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
436 "shenandoah_clone",
437 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
438 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
439 "ghash_processBlocks",
440 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
441 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
442 "sha1_implCompress",
443 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
444 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
445 "sha256_implCompress",
446 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
447 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
448 "sha512_implCompress",
449 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
450 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
451 "sha1_implCompressMB",
452 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
453 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
454 "sha256_implCompressMB",
455 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
456 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
457 "sha512_implCompressMB",
458 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
459 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
460 "encodeBlock",
461 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahStore }, { -1, ShenandoahNone },
462 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
463 "decodeBlock",
464 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahStore }, { -1, ShenandoahNone },
465 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
466 };
467
468 if (call->is_call_to_arraycopystub()) {
469 Node* dest = nullptr;
470 const TypeTuple* args = n->as_Call()->_tf->domain();
471 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
472 if (args->field_at(i)->isa_ptr()) {
473 j++;
474 if (j == 2) {
475 dest = n->in(i);
476 break;
477 }
478 }
479 }
480 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) ||
481 !verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) {
482 report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n);
483 }
484 } else if (strlen(call->_name) > 5 &&
485 !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) {
486 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) {
487 report_verify_failure("Shenandoah verification: _fill should have barriers", n);
488 }
489 } else if (!strcmp(call->_name, "shenandoah_wb_pre")) {
490 // skip
491 } else {
492 const int calls_len = sizeof(calls) / sizeof(calls[0]);
493 int i = 0;
494 for (; i < calls_len; i++) {
495 if (!strcmp(calls[i].name, call->_name)) {
496 break;
497 }
498 }
499 if (i != calls_len) {
500 const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]);
501 for (uint j = 0; j < args_len; j++) {
502 int pos = calls[i].args[j].pos;
503 if (pos == -1) {
504 break;
505 }
506 if (!verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) {
507 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
508 }
509 }
510 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
511 if (call->in(j)->bottom_type()->make_ptr() &&
512 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
513 uint k = 0;
514 for (; k < args_len && calls[i].args[k].pos != (int)j; k++);
515 if (k == args_len) {
516 fatal("arg %d for call %s not covered", j, call->_name);
517 }
518 }
519 }
520 } else {
521 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
522 if (call->in(j)->bottom_type()->make_ptr() &&
523 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
524 fatal("%s not covered", call->_name);
525 }
526 }
527 }
528 }
529 } else if (n->Opcode() == Op_ShenandoahIUBarrier || n->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
530 // skip
531 } else if (n->is_AddP()
532 || n->is_Phi()
533 || n->is_ConstraintCast()
534 || n->Opcode() == Op_Return
535 || n->Opcode() == Op_CMoveP
536 || n->Opcode() == Op_CMoveN
537 || n->Opcode() == Op_Rethrow
538 || n->is_MemBar()
539 || n->Opcode() == Op_Conv2B
540 || n->Opcode() == Op_SafePoint
541 || n->is_CallJava()
542 || n->Opcode() == Op_Unlock
543 || n->Opcode() == Op_EncodeP
544 || n->Opcode() == Op_DecodeN) {
545 // nothing to do
546 } else {
547 static struct {
548 int opcode;
549 struct {
550 int pos;
551 verify_type t;
552 } inputs[2];
553 } others[] = {
554 Op_FastLock,
555 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
556 Op_Lock,
557 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone} },
558 Op_ArrayCopy,
559 { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } },
560 Op_StrCompressedCopy,
561 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
562 Op_StrInflatedCopy,
563 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
564 Op_AryEq,
565 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
566 Op_StrIndexOf,
567 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
568 Op_StrComp,
569 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
570 Op_StrEquals,
571 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
572 Op_VectorizedHashCode,
573 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } },
574 Op_EncodeISOArray,
575 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
576 Op_CountPositives,
577 { { 2, ShenandoahLoad }, { -1, ShenandoahNone} },
578 Op_CastP2X,
579 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
580 Op_StrIndexOfChar,
581 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } },
582 };
583
584 const int others_len = sizeof(others) / sizeof(others[0]);
585 int i = 0;
586 for (; i < others_len; i++) {
587 if (others[i].opcode == n->Opcode()) {
588 break;
589 }
590 }
591 uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req();
592 if (i != others_len) {
593 const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]);
594 for (uint j = 0; j < inputs_len; j++) {
595 int pos = others[i].inputs[j].pos;
596 if (pos == -1) {
597 break;
598 }
599 if (!verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) {
600 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
601 }
602 }
603 for (uint j = 1; j < stop; j++) {
604 if (n->in(j) != nullptr && n->in(j)->bottom_type()->make_ptr() &&
605 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
606 uint k = 0;
607 for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++);
608 if (k == inputs_len) {
609 fatal("arg %d for node %s not covered", j, n->Name());
610 }
611 }
612 }
613 } else {
614 for (uint j = 1; j < stop; j++) {
615 if (n->in(j) != nullptr && n->in(j)->bottom_type()->make_ptr() &&
616 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
617 fatal("%s not covered", n->Name());
618 }
619 }
620 }
621 }
622
623 if (n->is_SafePoint()) {
624 SafePointNode* sfpt = n->as_SafePoint();
625 if (verify_no_useless_barrier && sfpt->jvms() != nullptr) {
626 for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) {
627 if (!verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) {
628 phis.clear();
629 visited.reset();
630 }
631 }
632 }
633 }
634 }
635
636 if (verify_no_useless_barrier) {
637 for (int i = 0; i < barriers.length(); i++) {
638 Node* n = barriers.at(i);
639 if (!barriers_used.member(n)) {
640 tty->print("XXX useless barrier"); n->dump(-2);
641 ShouldNotReachHere();
642 }
643 }
644 }
645 }
646 #endif
647
648 bool ShenandoahBarrierC2Support::is_dominator_same_ctrl(Node* c, Node* d, Node* n, PhaseIdealLoop* phase) {
649 // That both nodes have the same control is not sufficient to prove
650 // domination, verify that there's no path from d to n
651 ResourceMark rm;
652 Unique_Node_List wq;
653 wq.push(d);
654 for (uint next = 0; next < wq.size(); next++) {
655 Node *m = wq.at(next);
656 if (m == n) {
657 return false;
658 }
659 if (m->is_Phi() && m->in(0)->is_Loop()) {
660 assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control");
661 } else {
662 if (m->is_Store() || m->is_LoadStore()) {
663 // Take anti-dependencies into account
664 Node* mem = m->in(MemNode::Memory);
665 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
666 Node* u = mem->fast_out(i);
667 if (u->is_Load() && phase->C->can_alias(m->adr_type(), phase->C->get_alias_index(u->adr_type())) &&
668 phase->ctrl_or_self(u) == c) {
669 wq.push(u);
670 }
671 }
672 }
673 for (uint i = 0; i < m->req(); i++) {
674 if (m->in(i) != nullptr && phase->ctrl_or_self(m->in(i)) == c) {
675 wq.push(m->in(i));
676 }
677 }
678 }
679 }
680 return true;
681 }
682
683 bool ShenandoahBarrierC2Support::is_dominator(Node* d_c, Node* n_c, Node* d, Node* n, PhaseIdealLoop* phase) {
684 if (d_c != n_c) {
685 return phase->is_dominator(d_c, n_c);
686 }
687 return is_dominator_same_ctrl(d_c, d, n, phase);
688 }
689
690 Node* next_mem(Node* mem, int alias) {
691 Node* res = nullptr;
692 if (mem->is_Proj()) {
693 res = mem->in(0);
694 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
695 res = mem->in(TypeFunc::Memory);
696 } else if (mem->is_Phi()) {
697 res = mem->in(1);
698 } else if (mem->is_MergeMem()) {
699 res = mem->as_MergeMem()->memory_at(alias);
700 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
701 assert(alias == Compile::AliasIdxRaw, "following raw memory can't lead to a barrier");
702 res = mem->in(MemNode::Memory);
703 } else {
704 #ifdef ASSERT
705 mem->dump();
706 #endif
707 ShouldNotReachHere();
708 }
709 return res;
710 }
711
712 Node* ShenandoahBarrierC2Support::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) {
713 Node* iffproj = nullptr;
714 while (c != dom) {
715 Node* next = phase->idom(c);
716 assert(next->unique_ctrl_out_or_null() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?");
717 if (c->is_Region()) {
718 ResourceMark rm;
719 Unique_Node_List wq;
720 wq.push(c);
721 for (uint i = 0; i < wq.size(); i++) {
722 Node *n = wq.at(i);
723 if (n == next) {
724 continue;
725 }
726 if (n->is_Region()) {
727 for (uint j = 1; j < n->req(); j++) {
728 wq.push(n->in(j));
729 }
730 } else {
731 wq.push(n->in(0));
732 }
733 }
734 for (uint i = 0; i < wq.size(); i++) {
735 Node *n = wq.at(i);
736 assert(n->is_CFG(), "");
737 if (n->is_Multi()) {
738 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
739 Node* u = n->fast_out(j);
740 if (u->is_CFG()) {
741 if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
742 return NodeSentinel;
743 }
744 }
745 }
746 }
747 }
748 } else if (c->is_Proj()) {
749 if (c->is_IfProj()) {
750 if (c->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) != nullptr) {
751 // continue;
752 } else {
753 if (!allow_one_proj) {
754 return NodeSentinel;
755 }
756 if (iffproj == nullptr) {
757 iffproj = c;
758 } else {
759 return NodeSentinel;
760 }
761 }
762 } else if (c->Opcode() == Op_JumpProj) {
763 return NodeSentinel; // unsupported
764 } else if (c->Opcode() == Op_CatchProj) {
765 return NodeSentinel; // unsupported
766 } else if (c->Opcode() == Op_CProj && next->is_NeverBranch()) {
767 return NodeSentinel; // unsupported
768 } else {
769 assert(next->unique_ctrl_out() == c, "unsupported branch pattern");
770 }
771 }
772 c = next;
773 }
774 return iffproj;
775 }
776
777 Node* ShenandoahBarrierC2Support::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) {
778 ResourceMark rm;
779 VectorSet wq;
780 wq.set(mem->_idx);
781 mem_ctrl = phase->ctrl_or_self(mem);
782 while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) {
783 mem = next_mem(mem, alias);
784 if (wq.test_set(mem->_idx)) {
785 return nullptr;
786 }
787 mem_ctrl = phase->ctrl_or_self(mem);
788 }
789 if (mem->is_MergeMem()) {
790 mem = mem->as_MergeMem()->memory_at(alias);
791 mem_ctrl = phase->ctrl_or_self(mem);
792 }
793 return mem;
794 }
795
796 Node* ShenandoahBarrierC2Support::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) {
797 Node* mem = nullptr;
798 Node* c = ctrl;
799 do {
800 if (c->is_Region()) {
801 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == nullptr; i++) {
802 Node* u = c->fast_out(i);
803 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
804 if (u->adr_type() == TypePtr::BOTTOM) {
805 mem = u;
806 }
807 }
808 }
809 } else {
810 if (c->is_Call() && c->as_Call()->adr_type() != nullptr) {
811 CallProjections projs;
812 c->as_Call()->extract_projections(&projs, true, false);
813 if (projs.fallthrough_memproj != nullptr) {
814 if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
815 if (projs.catchall_memproj == nullptr) {
816 mem = projs.fallthrough_memproj;
817 } else {
818 if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) {
819 mem = projs.fallthrough_memproj;
820 } else {
821 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier");
822 mem = projs.catchall_memproj;
823 }
824 }
825 }
826 } else {
827 Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
828 if (proj != nullptr &&
829 proj->adr_type() == TypePtr::BOTTOM) {
830 mem = proj;
831 }
832 }
833 } else {
834 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
835 Node* u = c->fast_out(i);
836 if (u->is_Proj() &&
837 u->bottom_type() == Type::MEMORY &&
838 u->adr_type() == TypePtr::BOTTOM) {
839 assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
840 assert(mem == nullptr, "only one proj");
841 mem = u;
842 }
843 }
844 assert(!c->is_Call() || c->as_Call()->adr_type() != nullptr || mem == nullptr, "no mem projection expected");
845 }
846 }
847 c = phase->idom(c);
848 } while (mem == nullptr);
849 return mem;
850 }
851
852 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
853 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
854 Node* u = n->fast_out(i);
855 if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
856 uses.push(u);
857 }
858 }
859 }
860
861 static void hide_strip_mined_loop(OuterStripMinedLoopNode* outer, CountedLoopNode* inner, PhaseIdealLoop* phase) {
862 OuterStripMinedLoopEndNode* le = inner->outer_loop_end();
863 Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl));
864 phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl));
865 Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt);
866 phase->register_control(new_le, phase->get_loop(le), le->in(0));
867 phase->lazy_replace(outer, new_outer);
868 phase->lazy_replace(le, new_le);
869 inner->clear_strip_mined();
870 }
871
872 void ShenandoahBarrierC2Support::test_gc_state(Node*& ctrl, Node* raw_mem, Node*& test_fail_ctrl,
873 PhaseIdealLoop* phase, int flags) {
874 PhaseIterGVN& igvn = phase->igvn();
875 Node* old_ctrl = ctrl;
876
877 Node* thread = new ThreadLocalNode();
878 Node* gc_state_offset = igvn.MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
879 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, gc_state_offset);
880 Node* gc_state = new LoadBNode(old_ctrl, raw_mem, gc_state_addr,
881 DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
882 TypeInt::BYTE, MemNode::unordered);
883 Node* gc_state_and = new AndINode(gc_state, igvn.intcon(flags));
884 Node* gc_state_cmp = new CmpINode(gc_state_and, igvn.zerocon(T_INT));
885 Node* gc_state_bool = new BoolNode(gc_state_cmp, BoolTest::ne);
886
887 IfNode* gc_state_iff = new IfNode(old_ctrl, gc_state_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
888 ctrl = new IfTrueNode(gc_state_iff);
889 test_fail_ctrl = new IfFalseNode(gc_state_iff);
890
891 IdealLoopTree* loop = phase->get_loop(old_ctrl);
892 phase->register_control(gc_state_iff, loop, old_ctrl);
893 phase->register_control(ctrl, loop, gc_state_iff);
894 phase->register_control(test_fail_ctrl, loop, gc_state_iff);
895
896 phase->register_new_node(thread, old_ctrl);
897 phase->register_new_node(gc_state_addr, old_ctrl);
898 phase->register_new_node(gc_state, old_ctrl);
899 phase->register_new_node(gc_state_and, old_ctrl);
900 phase->register_new_node(gc_state_cmp, old_ctrl);
901 phase->register_new_node(gc_state_bool, old_ctrl);
902
903 phase->set_ctrl(gc_state_offset, phase->C->root());
904
905 assert(is_gc_state_test(gc_state_iff, flags), "Should match the shape");
906 }
907
908 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) {
909 Node* old_ctrl = ctrl;
910 PhaseIterGVN& igvn = phase->igvn();
911
912 const Type* val_t = igvn.type(val);
913 if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
914 Node* null_cmp = new CmpPNode(val, igvn.zerocon(T_OBJECT));
915 Node* null_test = new BoolNode(null_cmp, BoolTest::ne);
916
917 IfNode* null_iff = new IfNode(old_ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
918 ctrl = new IfTrueNode(null_iff);
919 null_ctrl = new IfFalseNode(null_iff);
920
921 IdealLoopTree* loop = phase->get_loop(old_ctrl);
922 phase->register_control(null_iff, loop, old_ctrl);
923 phase->register_control(ctrl, loop, null_iff);
924 phase->register_control(null_ctrl, loop, null_iff);
925
926 phase->register_new_node(null_cmp, old_ctrl);
927 phase->register_new_node(null_test, old_ctrl);
928 }
929 }
930
931 void ShenandoahBarrierC2Support::test_in_cset(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) {
932 Node* old_ctrl = ctrl;
933 PhaseIterGVN& igvn = phase->igvn();
934
935 Node* raw_val = new CastP2XNode(old_ctrl, val);
936 Node* cset_idx = new URShiftXNode(raw_val, igvn.intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
937
938 // Figure out the target cset address with raw pointer math.
939 // This avoids matching AddP+LoadB that would emit inefficient code.
940 // See JDK-8245465.
941 Node* cset_addr_ptr = igvn.makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
942 Node* cset_addr = new CastP2XNode(old_ctrl, cset_addr_ptr);
943 Node* cset_load_addr = new AddXNode(cset_addr, cset_idx);
944 Node* cset_load_ptr = new CastX2PNode(cset_load_addr);
945
946 Node* cset_load = new LoadBNode(old_ctrl, raw_mem, cset_load_ptr,
947 DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
948 TypeInt::BYTE, MemNode::unordered);
949 Node* cset_cmp = new CmpINode(cset_load, igvn.zerocon(T_INT));
950 Node* cset_bool = new BoolNode(cset_cmp, BoolTest::ne);
951
952 IfNode* cset_iff = new IfNode(old_ctrl, cset_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
953 ctrl = new IfTrueNode(cset_iff);
954 not_cset_ctrl = new IfFalseNode(cset_iff);
955
956 IdealLoopTree *loop = phase->get_loop(old_ctrl);
957 phase->register_control(cset_iff, loop, old_ctrl);
958 phase->register_control(ctrl, loop, cset_iff);
959 phase->register_control(not_cset_ctrl, loop, cset_iff);
960
961 phase->set_ctrl(cset_addr_ptr, phase->C->root());
962
963 phase->register_new_node(raw_val, old_ctrl);
964 phase->register_new_node(cset_idx, old_ctrl);
965 phase->register_new_node(cset_addr, old_ctrl);
966 phase->register_new_node(cset_load_addr, old_ctrl);
967 phase->register_new_node(cset_load_ptr, old_ctrl);
968 phase->register_new_node(cset_load, old_ctrl);
969 phase->register_new_node(cset_cmp, old_ctrl);
970 phase->register_new_node(cset_bool, old_ctrl);
971 }
972
973 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node* load_addr,
974 DecoratorSet decorators, PhaseIdealLoop* phase) {
975 IdealLoopTree*loop = phase->get_loop(ctrl);
976 const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr();
977
978 address calladdr = nullptr;
979 const char* name = nullptr;
980 bool is_strong = ShenandoahBarrierSet::is_strong_access(decorators);
981 bool is_weak = ShenandoahBarrierSet::is_weak_access(decorators);
982 bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
983 bool is_native = ShenandoahBarrierSet::is_native_access(decorators);
984 bool is_narrow = UseCompressedOops && !is_native;
985 if (is_strong) {
986 if (is_narrow) {
987 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow);
988 name = "load_reference_barrier_strong_narrow";
989 } else {
990 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
991 name = "load_reference_barrier_strong";
992 }
993 } else if (is_weak) {
994 if (is_narrow) {
995 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow);
996 name = "load_reference_barrier_weak_narrow";
997 } else {
998 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak);
999 name = "load_reference_barrier_weak";
1000 }
1001 } else {
1002 assert(is_phantom, "only remaining strength");
1003 if (is_narrow) {
1004 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom_narrow);
1005 name = "load_reference_barrier_phantom_narrow";
1006 } else {
1007 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom);
1008 name = "load_reference_barrier_phantom";
1009 }
1010 }
1011 Node* call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type(), calladdr, name, TypeRawPtr::BOTTOM);
1012
1013 call->init_req(TypeFunc::Control, ctrl);
1014 call->init_req(TypeFunc::I_O, phase->C->top());
1015 call->init_req(TypeFunc::Memory, phase->C->top());
1016 call->init_req(TypeFunc::FramePtr, phase->C->top());
1017 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1018 call->init_req(TypeFunc::Parms, val);
1019 call->init_req(TypeFunc::Parms+1, load_addr);
1020 phase->register_control(call, loop, ctrl);
1021 ctrl = new ProjNode(call, TypeFunc::Control);
1022 phase->register_control(ctrl, loop, call);
1023 val = new ProjNode(call, TypeFunc::Parms);
1024 phase->register_new_node(val, call);
1025 val = new CheckCastPPNode(ctrl, val, obj_type);
1026 phase->register_new_node(val, ctrl);
1027 }
1028
1029 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) {
1030 Node* ctrl = phase->get_ctrl(barrier);
1031 Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
1032
1033 // Update the control of all nodes that should be after the
1034 // barrier control flow
1035 uses.clear();
1036 // Every node that is control dependent on the barrier's input
1037 // control will be after the expanded barrier. The raw memory (if
1038 // its memory is control dependent on the barrier's input control)
1039 // must stay above the barrier.
1040 uses_to_ignore.clear();
1041 if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
1042 uses_to_ignore.push(init_raw_mem);
1043 }
1044 for (uint next = 0; next < uses_to_ignore.size(); next++) {
1045 Node *n = uses_to_ignore.at(next);
1046 for (uint i = 0; i < n->req(); i++) {
1047 Node* in = n->in(i);
1048 if (in != nullptr && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
1049 uses_to_ignore.push(in);
1050 }
1051 }
1052 }
1053 for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
1054 Node* u = ctrl->fast_out(i);
1055 if (u->_idx < last &&
1056 u != barrier &&
1057 !u->depends_only_on_test() && // preserve dependency on test
1058 !uses_to_ignore.member(u) &&
1059 (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) &&
1060 (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) {
1061 Node* old_c = phase->ctrl_or_self(u);
1062 Node* c = old_c;
1063 if (c != ctrl ||
1064 is_dominator_same_ctrl(old_c, barrier, u, phase) ||
1065 ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) {
1066 phase->igvn().rehash_node_delayed(u);
1067 int nb = u->replace_edge(ctrl, region, &phase->igvn());
1068 if (u->is_CFG()) {
1069 if (phase->idom(u) == ctrl) {
1070 phase->set_idom(u, region, phase->dom_depth(region));
1071 }
1072 } else if (phase->get_ctrl(u) == ctrl) {
1073 assert(u != init_raw_mem, "should leave input raw mem above the barrier");
1074 uses.push(u);
1075 }
1076 assert(nb == 1, "more than 1 ctrl input?");
1077 --i, imax -= nb;
1078 }
1079 }
1080 }
1081 }
1082
1083 static Node* create_phis_on_call_return(Node* ctrl, Node* c, Node* n, Node* n_clone, const CallProjections& projs, PhaseIdealLoop* phase) {
1084 Node* region = nullptr;
1085 while (c != ctrl) {
1086 if (c->is_Region()) {
1087 region = c;
1088 }
1089 c = phase->idom(c);
1090 }
1091 assert(region != nullptr, "");
1092 Node* phi = new PhiNode(region, n->bottom_type());
1093 for (uint j = 1; j < region->req(); j++) {
1094 Node* in = region->in(j);
1095 if (phase->is_dominator(projs.fallthrough_catchproj, in)) {
1096 phi->init_req(j, n);
1097 } else if (phase->is_dominator(projs.catchall_catchproj, in)) {
1098 phi->init_req(j, n_clone);
1099 } else {
1100 phi->init_req(j, create_phis_on_call_return(ctrl, in, n, n_clone, projs, phase));
1101 }
1102 }
1103 phase->register_new_node(phi, region);
1104 return phi;
1105 }
1106
1107 void ShenandoahBarrierC2Support::pin_and_expand(PhaseIdealLoop* phase) {
1108 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
1109
1110 Unique_Node_List uses;
1111 for (int i = 0; i < state->iu_barriers_count(); i++) {
1112 Node* barrier = state->iu_barrier(i);
1113 Node* ctrl = phase->get_ctrl(barrier);
1114 IdealLoopTree* loop = phase->get_loop(ctrl);
1115 Node* head = loop->head();
1116 if (head->is_OuterStripMinedLoop()) {
1117 // Expanding a barrier here will break loop strip mining
1118 // verification. Transform the loop so the loop nest doesn't
1119 // appear as strip mined.
1120 OuterStripMinedLoopNode* outer = head->as_OuterStripMinedLoop();
1121 hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1122 }
1123 }
1124
1125 Node_Stack stack(0);
1126 Node_List clones;
1127 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1128 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1129
1130 Node* ctrl = phase->get_ctrl(lrb);
1131 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1132
1133 CallStaticJavaNode* unc = nullptr;
1134 Node* unc_ctrl = nullptr;
1135 Node* uncasted_val = val;
1136
1137 for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1138 Node* u = lrb->fast_out(i);
1139 if (u->Opcode() == Op_CastPP &&
1140 u->in(0) != nullptr &&
1141 phase->is_dominator(u->in(0), ctrl)) {
1142 const Type* u_t = phase->igvn().type(u);
1143
1144 if (u_t->meet(TypePtr::NULL_PTR) != u_t &&
1145 u->in(0)->Opcode() == Op_IfTrue &&
1146 u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
1147 u->in(0)->in(0)->is_If() &&
1148 u->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
1149 u->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
1150 u->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
1151 u->in(0)->in(0)->in(1)->in(1)->in(1) == val &&
1152 u->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
1153 IdealLoopTree* loop = phase->get_loop(ctrl);
1154 IdealLoopTree* unc_loop = phase->get_loop(u->in(0));
1155
1156 if (!unc_loop->is_member(loop)) {
1157 continue;
1158 }
1159
1160 Node* branch = no_branches(ctrl, u->in(0), false, phase);
1161 assert(branch == nullptr || branch == NodeSentinel, "was not looking for a branch");
1162 if (branch == NodeSentinel) {
1163 continue;
1164 }
1165
1166 Node* iff = u->in(0)->in(0);
1167 Node* bol = iff->in(1)->clone();
1168 Node* cmp = bol->in(1)->clone();
1169 cmp->set_req(1, lrb);
1170 bol->set_req(1, cmp);
1171 phase->igvn().replace_input_of(iff, 1, bol);
1172 phase->set_ctrl(lrb, iff->in(0));
1173 phase->register_new_node(cmp, iff->in(0));
1174 phase->register_new_node(bol, iff->in(0));
1175 break;
1176 }
1177 }
1178 }
1179 if ((ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) || ctrl->is_CallJava()) {
1180 CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_CallJava() : ctrl->as_CallJava();
1181 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
1182 // The rethrow call may have too many projections to be
1183 // properly handled here. Given there's no reason for a
1184 // barrier to depend on the call, move it above the call
1185 stack.push(lrb, 0);
1186 do {
1187 Node* n = stack.node();
1188 uint idx = stack.index();
1189 if (idx < n->req()) {
1190 Node* in = n->in(idx);
1191 stack.set_index(idx+1);
1192 if (in != nullptr) {
1193 if (phase->has_ctrl(in)) {
1194 if (phase->is_dominator(call, phase->get_ctrl(in))) {
1195 #ifdef ASSERT
1196 for (uint i = 0; i < stack.size(); i++) {
1197 assert(stack.node_at(i) != in, "node shouldn't have been seen yet");
1198 }
1199 #endif
1200 stack.push(in, 0);
1201 }
1202 } else {
1203 assert(phase->is_dominator(in, call->in(0)), "no dependency on the call");
1204 }
1205 }
1206 } else {
1207 phase->set_ctrl(n, call->in(0));
1208 stack.pop();
1209 }
1210 } while(stack.size() > 0);
1211 continue;
1212 }
1213 CallProjections projs;
1214 call->extract_projections(&projs, false, false);
1215
1216 #ifdef ASSERT
1217 VectorSet cloned;
1218 #endif
1219 Node* lrb_clone = lrb->clone();
1220 phase->register_new_node(lrb_clone, projs.catchall_catchproj);
1221 phase->set_ctrl(lrb, projs.fallthrough_catchproj);
1222
1223 stack.push(lrb, 0);
1224 clones.push(lrb_clone);
1225
1226 do {
1227 assert(stack.size() == clones.size(), "");
1228 Node* n = stack.node();
1229 #ifdef ASSERT
1230 if (n->is_Load()) {
1231 Node* mem = n->in(MemNode::Memory);
1232 for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
1233 Node* u = mem->fast_out(j);
1234 assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?");
1235 }
1236 }
1237 #endif
1238 uint idx = stack.index();
1239 Node* n_clone = clones.at(clones.size()-1);
1240 if (idx < n->outcnt()) {
1241 Node* u = n->raw_out(idx);
1242 Node* c = phase->ctrl_or_self(u);
1243 if (phase->is_dominator(call, c) && phase->is_dominator(c, projs.fallthrough_proj)) {
1244 stack.set_index(idx+1);
1245 assert(!u->is_CFG(), "");
1246 stack.push(u, 0);
1247 assert(!cloned.test_set(u->_idx), "only one clone");
1248 Node* u_clone = u->clone();
1249 int nb = u_clone->replace_edge(n, n_clone, &phase->igvn());
1250 assert(nb > 0, "should have replaced some uses");
1251 phase->register_new_node(u_clone, projs.catchall_catchproj);
1252 clones.push(u_clone);
1253 phase->set_ctrl(u, projs.fallthrough_catchproj);
1254 } else {
1255 bool replaced = false;
1256 if (u->is_Phi()) {
1257 for (uint k = 1; k < u->req(); k++) {
1258 if (u->in(k) == n) {
1259 if (phase->is_dominator(projs.catchall_catchproj, u->in(0)->in(k))) {
1260 phase->igvn().replace_input_of(u, k, n_clone);
1261 replaced = true;
1262 } else if (!phase->is_dominator(projs.fallthrough_catchproj, u->in(0)->in(k))) {
1263 phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase));
1264 replaced = true;
1265 }
1266 }
1267 }
1268 } else {
1269 if (phase->is_dominator(projs.catchall_catchproj, c)) {
1270 phase->igvn().rehash_node_delayed(u);
1271 int nb = u->replace_edge(n, n_clone, &phase->igvn());
1272 assert(nb > 0, "should have replaced some uses");
1273 replaced = true;
1274 } else if (!phase->is_dominator(projs.fallthrough_catchproj, c)) {
1275 if (u->is_If()) {
1276 // Can't break If/Bool/Cmp chain
1277 assert(n->is_Bool(), "unexpected If shape");
1278 assert(stack.node_at(stack.size()-2)->is_Cmp(), "unexpected If shape");
1279 assert(n_clone->is_Bool(), "unexpected clone");
1280 assert(clones.at(clones.size()-2)->is_Cmp(), "unexpected clone");
1281 Node* bol_clone = n->clone();
1282 Node* cmp_clone = stack.node_at(stack.size()-2)->clone();
1283 bol_clone->set_req(1, cmp_clone);
1284
1285 Node* nn = stack.node_at(stack.size()-3);
1286 Node* nn_clone = clones.at(clones.size()-3);
1287 assert(nn->Opcode() == nn_clone->Opcode(), "mismatch");
1288
1289 int nb = cmp_clone->replace_edge(nn, create_phis_on_call_return(ctrl, c, nn, nn_clone, projs, phase),
1290 &phase->igvn());
1291 assert(nb > 0, "should have replaced some uses");
1292
1293 phase->register_new_node(bol_clone, u->in(0));
1294 phase->register_new_node(cmp_clone, u->in(0));
1295
1296 phase->igvn().replace_input_of(u, 1, bol_clone);
1297
1298 } else {
1299 phase->igvn().rehash_node_delayed(u);
1300 int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase), &phase->igvn());
1301 assert(nb > 0, "should have replaced some uses");
1302 }
1303 replaced = true;
1304 }
1305 }
1306 if (!replaced) {
1307 stack.set_index(idx+1);
1308 }
1309 }
1310 } else {
1311 stack.pop();
1312 clones.pop();
1313 }
1314 } while (stack.size() > 0);
1315 assert(stack.size() == 0 && clones.size() == 0, "");
1316 }
1317 }
1318
1319 for (int i = 0; i < state->load_reference_barriers_count(); i++) {
1320 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1321 Node* ctrl = phase->get_ctrl(lrb);
1322 IdealLoopTree* loop = phase->get_loop(ctrl);
1323 Node* head = loop->head();
1324 if (head->is_OuterStripMinedLoop()) {
1325 // Expanding a barrier here will break loop strip mining
1326 // verification. Transform the loop so the loop nest doesn't
1327 // appear as strip mined.
1328 OuterStripMinedLoopNode* outer = head->as_OuterStripMinedLoop();
1329 hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1330 }
1331 }
1332
1333 // Expand load-reference-barriers
1334 MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
1335 Unique_Node_List uses_to_ignore;
1336 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1337 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1338 uint last = phase->C->unique();
1339 Node* ctrl = phase->get_ctrl(lrb);
1340 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1341
1342 Node* orig_ctrl = ctrl;
1343
1344 Node* raw_mem = fixer.find_mem(ctrl, lrb);
1345 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, nullptr);
1346
1347 IdealLoopTree *loop = phase->get_loop(ctrl);
1348
1349 Node* heap_stable_ctrl = nullptr;
1350 Node* null_ctrl = nullptr;
1351
1352 assert(val->bottom_type()->make_oopptr(), "need oop");
1353 assert(val->bottom_type()->make_oopptr()->const_oop() == nullptr, "expect non-constant");
1354
1355 enum { _heap_stable = 1, _evac_path, _not_cset, PATH_LIMIT };
1356 Node* region = new RegionNode(PATH_LIMIT);
1357 Node* val_phi = new PhiNode(region, val->bottom_type()->is_oopptr());
1358
1359 // Stable path.
1360 int flags = ShenandoahHeap::HAS_FORWARDED;
1361 if (!ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1362 flags |= ShenandoahHeap::WEAK_ROOTS;
1363 }
1364 test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, flags);
1365 IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
1366
1367 // Heap stable case
1368 region->init_req(_heap_stable, heap_stable_ctrl);
1369 val_phi->init_req(_heap_stable, val);
1370
1371 // Test for in-cset, unless it's a native-LRB. Native LRBs need to return null
1372 // even for non-cset objects to prevent resurrection of such objects.
1373 // Wires !in_cset(obj) to slot 2 of region and phis
1374 Node* not_cset_ctrl = nullptr;
1375 if (ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1376 test_in_cset(ctrl, not_cset_ctrl, val, raw_mem, phase);
1377 }
1378 if (not_cset_ctrl != nullptr) {
1379 region->init_req(_not_cset, not_cset_ctrl);
1380 val_phi->init_req(_not_cset, val);
1381 } else {
1382 region->del_req(_not_cset);
1383 val_phi->del_req(_not_cset);
1384 }
1385
1386 // Resolve object when orig-value is in cset.
1387 // Make the unconditional resolve for fwdptr.
1388
1389 // Call lrb-stub and wire up that path in slots 4
1390 Node* result_mem = nullptr;
1391
1392 Node* addr;
1393 {
1394 VectorSet visited;
1395 addr = get_load_addr(phase, visited, lrb);
1396 }
1397 if (addr->Opcode() == Op_AddP) {
1398 Node* orig_base = addr->in(AddPNode::Base);
1399 Node* base = new CheckCastPPNode(ctrl, orig_base, orig_base->bottom_type(), ConstraintCastNode::StrongDependency);
1400 phase->register_new_node(base, ctrl);
1401 if (addr->in(AddPNode::Base) == addr->in((AddPNode::Address))) {
1402 // Field access
1403 addr = addr->clone();
1404 addr->set_req(AddPNode::Base, base);
1405 addr->set_req(AddPNode::Address, base);
1406 phase->register_new_node(addr, ctrl);
1407 } else {
1408 Node* addr2 = addr->in(AddPNode::Address);
1409 if (addr2->Opcode() == Op_AddP && addr2->in(AddPNode::Base) == addr2->in(AddPNode::Address) &&
1410 addr2->in(AddPNode::Base) == orig_base) {
1411 addr2 = addr2->clone();
1412 addr2->set_req(AddPNode::Base, base);
1413 addr2->set_req(AddPNode::Address, base);
1414 phase->register_new_node(addr2, ctrl);
1415 addr = addr->clone();
1416 addr->set_req(AddPNode::Base, base);
1417 addr->set_req(AddPNode::Address, addr2);
1418 phase->register_new_node(addr, ctrl);
1419 }
1420 }
1421 }
1422 call_lrb_stub(ctrl, val, addr, lrb->decorators(), phase);
1423 region->init_req(_evac_path, ctrl);
1424 val_phi->init_req(_evac_path, val);
1425
1426 phase->register_control(region, loop, heap_stable_iff);
1427 Node* out_val = val_phi;
1428 phase->register_new_node(val_phi, region);
1429
1430 fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase);
1431
1432 ctrl = orig_ctrl;
1433
1434 phase->igvn().replace_node(lrb, out_val);
1435
1436 follow_barrier_uses(out_val, ctrl, uses, phase);
1437
1438 for(uint next = 0; next < uses.size(); next++ ) {
1439 Node *n = uses.at(next);
1440 assert(phase->get_ctrl(n) == ctrl, "bad control");
1441 assert(n != raw_mem, "should leave input raw mem above the barrier");
1442 phase->set_ctrl(n, region);
1443 follow_barrier_uses(n, ctrl, uses, phase);
1444 }
1445 fixer.record_new_ctrl(ctrl, region, raw_mem, raw_mem_for_ctrl);
1446 }
1447 // Done expanding load-reference-barriers.
1448 assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced");
1449
1450 for (int i = state->iu_barriers_count() - 1; i >= 0; i--) {
1451 Node* barrier = state->iu_barrier(i);
1452 Node* pre_val = barrier->in(1);
1453
1454 if (phase->igvn().type(pre_val)->higher_equal(TypePtr::NULL_PTR)) {
1455 ShouldNotReachHere();
1456 continue;
1457 }
1458
1459 Node* ctrl = phase->get_ctrl(barrier);
1460
1461 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) {
1462 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0)->in(0), pre_val, ctrl->in(0), phase), "can't move");
1463 ctrl = ctrl->in(0)->in(0);
1464 phase->set_ctrl(barrier, ctrl);
1465 } else if (ctrl->is_CallRuntime()) {
1466 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0), pre_val, ctrl, phase), "can't move");
1467 ctrl = ctrl->in(0);
1468 phase->set_ctrl(barrier, ctrl);
1469 }
1470
1471 Node* init_ctrl = ctrl;
1472 IdealLoopTree* loop = phase->get_loop(ctrl);
1473 Node* raw_mem = fixer.find_mem(ctrl, barrier);
1474 Node* init_raw_mem = raw_mem;
1475 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, nullptr);
1476 Node* heap_stable_ctrl = nullptr;
1477 Node* null_ctrl = nullptr;
1478 uint last = phase->C->unique();
1479
1480 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
1481 Node* region = new RegionNode(PATH_LIMIT);
1482 Node* phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1483
1484 enum { _fast_path = 1, _slow_path, _null_path, PATH_LIMIT2 };
1485 Node* region2 = new RegionNode(PATH_LIMIT2);
1486 Node* phi2 = PhiNode::make(region2, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1487
1488 // Stable path.
1489 test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, ShenandoahHeap::MARKING);
1490 region->init_req(_heap_stable, heap_stable_ctrl);
1491 phi->init_req(_heap_stable, raw_mem);
1492
1493 // Null path
1494 Node* reg2_ctrl = nullptr;
1495 test_null(ctrl, pre_val, null_ctrl, phase);
1496 if (null_ctrl != nullptr) {
1497 reg2_ctrl = null_ctrl->in(0);
1498 region2->init_req(_null_path, null_ctrl);
1499 phi2->init_req(_null_path, raw_mem);
1500 } else {
1501 region2->del_req(_null_path);
1502 phi2->del_req(_null_path);
1503 }
1504
1505 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
1506 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1507 Node* thread = new ThreadLocalNode();
1508 phase->register_new_node(thread, ctrl);
1509 Node* buffer_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(buffer_offset));
1510 phase->register_new_node(buffer_adr, ctrl);
1511 Node* index_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(index_offset));
1512 phase->register_new_node(index_adr, ctrl);
1513
1514 BasicType index_bt = TypeX_X->basic_type();
1515 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading Shenandoah SATBMarkQueue::_index with wrong size.");
1516 const TypePtr* adr_type = TypeRawPtr::BOTTOM;
1517 Node* index = new LoadXNode(ctrl, raw_mem, index_adr, adr_type, TypeX_X, MemNode::unordered);
1518 phase->register_new_node(index, ctrl);
1519 Node* index_cmp = new CmpXNode(index, phase->igvn().MakeConX(0));
1520 phase->register_new_node(index_cmp, ctrl);
1521 Node* index_test = new BoolNode(index_cmp, BoolTest::ne);
1522 phase->register_new_node(index_test, ctrl);
1523 IfNode* queue_full_iff = new IfNode(ctrl, index_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
1524 if (reg2_ctrl == nullptr) reg2_ctrl = queue_full_iff;
1525 phase->register_control(queue_full_iff, loop, ctrl);
1526 Node* not_full = new IfTrueNode(queue_full_iff);
1527 phase->register_control(not_full, loop, queue_full_iff);
1528 Node* full = new IfFalseNode(queue_full_iff);
1529 phase->register_control(full, loop, queue_full_iff);
1530
1531 ctrl = not_full;
1532
1533 Node* next_index = new SubXNode(index, phase->igvn().MakeConX(sizeof(intptr_t)));
1534 phase->register_new_node(next_index, ctrl);
1535
1536 Node* buffer = new LoadPNode(ctrl, raw_mem, buffer_adr, adr_type, TypeRawPtr::NOTNULL, MemNode::unordered);
1537 phase->register_new_node(buffer, ctrl);
1538 Node *log_addr = new AddPNode(phase->C->top(), buffer, next_index);
1539 phase->register_new_node(log_addr, ctrl);
1540 Node* log_store = new StorePNode(ctrl, raw_mem, log_addr, adr_type, pre_val, MemNode::unordered);
1541 phase->register_new_node(log_store, ctrl);
1542 // update the index
1543 Node* index_update = new StoreXNode(ctrl, log_store, index_adr, adr_type, next_index, MemNode::unordered);
1544 phase->register_new_node(index_update, ctrl);
1545
1546 // Fast-path case
1547 region2->init_req(_fast_path, ctrl);
1548 phi2->init_req(_fast_path, index_update);
1549
1550 ctrl = full;
1551
1552 Node* base = find_bottom_mem(ctrl, phase);
1553
1554 MergeMemNode* mm = MergeMemNode::make(base);
1555 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
1556 phase->register_new_node(mm, ctrl);
1557
1558 Node* call = new CallLeafNode(ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type(), CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", TypeRawPtr::BOTTOM);
1559 call->init_req(TypeFunc::Control, ctrl);
1560 call->init_req(TypeFunc::I_O, phase->C->top());
1561 call->init_req(TypeFunc::Memory, mm);
1562 call->init_req(TypeFunc::FramePtr, phase->C->top());
1563 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1564 call->init_req(TypeFunc::Parms, pre_val);
1565 call->init_req(TypeFunc::Parms+1, thread);
1566 phase->register_control(call, loop, ctrl);
1567
1568 Node* ctrl_proj = new ProjNode(call, TypeFunc::Control);
1569 phase->register_control(ctrl_proj, loop, call);
1570 Node* mem_proj = new ProjNode(call, TypeFunc::Memory);
1571 phase->register_new_node(mem_proj, call);
1572
1573 // Slow-path case
1574 region2->init_req(_slow_path, ctrl_proj);
1575 phi2->init_req(_slow_path, mem_proj);
1576
1577 phase->register_control(region2, loop, reg2_ctrl);
1578 phase->register_new_node(phi2, region2);
1579
1580 region->init_req(_heap_unstable, region2);
1581 phi->init_req(_heap_unstable, phi2);
1582
1583 phase->register_control(region, loop, heap_stable_ctrl->in(0));
1584 phase->register_new_node(phi, region);
1585
1586 fix_ctrl(barrier, region, fixer, uses, uses_to_ignore, last, phase);
1587 for(uint next = 0; next < uses.size(); next++ ) {
1588 Node *n = uses.at(next);
1589 assert(phase->get_ctrl(n) == init_ctrl, "bad control");
1590 assert(n != init_raw_mem, "should leave input raw mem above the barrier");
1591 phase->set_ctrl(n, region);
1592 follow_barrier_uses(n, init_ctrl, uses, phase);
1593 }
1594 fixer.fix_mem(init_ctrl, region, init_raw_mem, raw_mem_for_ctrl, phi, uses);
1595
1596 phase->igvn().replace_node(barrier, pre_val);
1597 }
1598 assert(state->iu_barriers_count() == 0, "all enqueue barrier nodes should have been replaced");
1599
1600 }
1601
1602 Node* ShenandoahBarrierC2Support::get_load_addr(PhaseIdealLoop* phase, VectorSet& visited, Node* in) {
1603 if (visited.test_set(in->_idx)) {
1604 return nullptr;
1605 }
1606 switch (in->Opcode()) {
1607 case Op_Proj:
1608 return get_load_addr(phase, visited, in->in(0));
1609 case Op_CastPP:
1610 case Op_CheckCastPP:
1611 case Op_DecodeN:
1612 case Op_EncodeP:
1613 return get_load_addr(phase, visited, in->in(1));
1614 case Op_LoadN:
1615 case Op_LoadP:
1616 return in->in(MemNode::Address);
1617 case Op_CompareAndExchangeN:
1618 case Op_CompareAndExchangeP:
1619 case Op_GetAndSetN:
1620 case Op_GetAndSetP:
1621 case Op_ShenandoahCompareAndExchangeP:
1622 case Op_ShenandoahCompareAndExchangeN:
1623 // Those instructions would just have stored a different
1624 // value into the field. No use to attempt to fix it at this point.
1625 return phase->igvn().zerocon(T_OBJECT);
1626 case Op_CMoveP:
1627 case Op_CMoveN: {
1628 Node* t = get_load_addr(phase, visited, in->in(CMoveNode::IfTrue));
1629 Node* f = get_load_addr(phase, visited, in->in(CMoveNode::IfFalse));
1630 // Handle unambiguous cases: single address reported on both branches.
1631 if (t != nullptr && f == nullptr) return t;
1632 if (t == nullptr && f != nullptr) return f;
1633 if (t != nullptr && t == f) return t;
1634 // Ambiguity.
1635 return phase->igvn().zerocon(T_OBJECT);
1636 }
1637 case Op_Phi: {
1638 Node* addr = nullptr;
1639 for (uint i = 1; i < in->req(); i++) {
1640 Node* addr1 = get_load_addr(phase, visited, in->in(i));
1641 if (addr == nullptr) {
1642 addr = addr1;
1643 }
1644 if (addr != addr1) {
1645 return phase->igvn().zerocon(T_OBJECT);
1646 }
1647 }
1648 return addr;
1649 }
1650 case Op_ShenandoahLoadReferenceBarrier:
1651 return get_load_addr(phase, visited, in->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1652 case Op_ShenandoahIUBarrier:
1653 return get_load_addr(phase, visited, in->in(1));
1654 case Op_CallDynamicJava:
1655 case Op_CallLeaf:
1656 case Op_CallStaticJava:
1657 case Op_ConN:
1658 case Op_ConP:
1659 case Op_Parm:
1660 case Op_CreateEx:
1661 return phase->igvn().zerocon(T_OBJECT);
1662 default:
1663 #ifdef ASSERT
1664 fatal("Unknown node in get_load_addr: %s", NodeClassNames[in->Opcode()]);
1665 #endif
1666 return phase->igvn().zerocon(T_OBJECT);
1667 }
1668
1669 }
1670
1671 void ShenandoahBarrierC2Support::move_gc_state_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
1672 IdealLoopTree *loop = phase->get_loop(iff);
1673 Node* loop_head = loop->_head;
1674 Node* entry_c = loop_head->in(LoopNode::EntryControl);
1675
1676 Node* bol = iff->in(1);
1677 Node* cmp = bol->in(1);
1678 Node* andi = cmp->in(1);
1679 Node* load = andi->in(1);
1680
1681 assert(is_gc_state_load(load), "broken");
1682 if (!phase->is_dominator(load->in(0), entry_c)) {
1683 Node* mem_ctrl = nullptr;
1684 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
1685 load = load->clone();
1686 load->set_req(MemNode::Memory, mem);
1687 load->set_req(0, entry_c);
1688 phase->register_new_node(load, entry_c);
1689 andi = andi->clone();
1690 andi->set_req(1, load);
1691 phase->register_new_node(andi, entry_c);
1692 cmp = cmp->clone();
1693 cmp->set_req(1, andi);
1694 phase->register_new_node(cmp, entry_c);
1695 bol = bol->clone();
1696 bol->set_req(1, cmp);
1697 phase->register_new_node(bol, entry_c);
1698
1699 phase->igvn().replace_input_of(iff, 1, bol);
1700 }
1701 }
1702
1703 bool ShenandoahBarrierC2Support::identical_backtoback_ifs(Node* n, PhaseIdealLoop* phase) {
1704 if (!n->is_If() || n->is_CountedLoopEnd()) {
1705 return false;
1706 }
1707 Node* region = n->in(0);
1708
1709 if (!region->is_Region()) {
1710 return false;
1711 }
1712 Node* dom = phase->idom(region);
1713 if (!dom->is_If()) {
1714 return false;
1715 }
1716
1717 if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) {
1718 return false;
1719 }
1720
1721 IfNode* dom_if = dom->as_If();
1722 Node* proj_true = dom_if->proj_out(1);
1723 Node* proj_false = dom_if->proj_out(0);
1724
1725 for (uint i = 1; i < region->req(); i++) {
1726 if (phase->is_dominator(proj_true, region->in(i))) {
1727 continue;
1728 }
1729 if (phase->is_dominator(proj_false, region->in(i))) {
1730 continue;
1731 }
1732 return false;
1733 }
1734
1735 return true;
1736 }
1737
1738 bool ShenandoahBarrierC2Support::merge_point_safe(Node* region) {
1739 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1740 Node* n = region->fast_out(i);
1741 if (n->is_LoadStore()) {
1742 // Splitting a LoadStore node through phi, causes it to lose its SCMemProj: the split if code doesn't have support
1743 // for a LoadStore at the region the if is split through because that's not expected to happen (LoadStore nodes
1744 // should be between barrier nodes). It does however happen with Shenandoah though because barriers can get
1745 // expanded around a LoadStore node.
1746 return false;
1747 }
1748 }
1749 return true;
1750 }
1751
1752
1753 void ShenandoahBarrierC2Support::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) {
1754 assert(is_heap_stable_test(n), "no other tests");
1755 if (identical_backtoback_ifs(n, phase)) {
1756 Node* n_ctrl = n->in(0);
1757 if (phase->can_split_if(n_ctrl) && merge_point_safe(n_ctrl)) {
1758 IfNode* dom_if = phase->idom(n_ctrl)->as_If();
1759 if (is_heap_stable_test(n)) {
1760 Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1);
1761 assert(is_gc_state_load(gc_state_load), "broken");
1762 Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1);
1763 assert(is_gc_state_load(dom_gc_state_load), "broken");
1764 if (gc_state_load != dom_gc_state_load) {
1765 phase->igvn().replace_node(gc_state_load, dom_gc_state_load);
1766 }
1767 }
1768 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1769 Node* proj_true = dom_if->proj_out(1);
1770 Node* proj_false = dom_if->proj_out(0);
1771 Node* con_true = phase->igvn().makecon(TypeInt::ONE);
1772 Node* con_false = phase->igvn().makecon(TypeInt::ZERO);
1773
1774 for (uint i = 1; i < n_ctrl->req(); i++) {
1775 if (phase->is_dominator(proj_true, n_ctrl->in(i))) {
1776 bolphi->init_req(i, con_true);
1777 } else {
1778 assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1779 bolphi->init_req(i, con_false);
1780 }
1781 }
1782 phase->register_new_node(bolphi, n_ctrl);
1783 phase->igvn().replace_input_of(n, 1, bolphi);
1784 phase->do_split_if(n);
1785 }
1786 }
1787 }
1788
1789 IfNode* ShenandoahBarrierC2Support::find_unswitching_candidate(const IdealLoopTree* loop, PhaseIdealLoop* phase) {
1790 // Find first invariant test that doesn't exit the loop
1791 LoopNode *head = loop->_head->as_Loop();
1792 IfNode* unswitch_iff = nullptr;
1793 Node* n = head->in(LoopNode::LoopBackControl);
1794 int loop_has_sfpts = -1;
1795 while (n != head) {
1796 Node* n_dom = phase->idom(n);
1797 if (n->is_Region()) {
1798 if (n_dom->is_If()) {
1799 IfNode* iff = n_dom->as_If();
1800 if (iff->in(1)->is_Bool()) {
1801 BoolNode* bol = iff->in(1)->as_Bool();
1802 if (bol->in(1)->is_Cmp()) {
1803 // If condition is invariant and not a loop exit,
1804 // then found reason to unswitch.
1805 if (is_heap_stable_test(iff) &&
1806 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) {
1807 assert(!loop->is_loop_exit(iff), "both branches should be in the loop");
1808 if (loop_has_sfpts == -1) {
1809 for(uint i = 0; i < loop->_body.size(); i++) {
1810 Node *m = loop->_body[i];
1811 if (m->is_SafePoint() && !m->is_CallLeaf()) {
1812 loop_has_sfpts = 1;
1813 break;
1814 }
1815 }
1816 if (loop_has_sfpts == -1) {
1817 loop_has_sfpts = 0;
1818 }
1819 }
1820 if (!loop_has_sfpts) {
1821 unswitch_iff = iff;
1822 }
1823 }
1824 }
1825 }
1826 }
1827 }
1828 n = n_dom;
1829 }
1830 return unswitch_iff;
1831 }
1832
1833
1834 void ShenandoahBarrierC2Support::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) {
1835 Node_List heap_stable_tests;
1836 stack.push(phase->C->start(), 0);
1837 do {
1838 Node* n = stack.node();
1839 uint i = stack.index();
1840
1841 if (i < n->outcnt()) {
1842 Node* u = n->raw_out(i);
1843 stack.set_index(i+1);
1844 if (!visited.test_set(u->_idx)) {
1845 stack.push(u, 0);
1846 }
1847 } else {
1848 stack.pop();
1849 if (n->is_If() && is_heap_stable_test(n)) {
1850 heap_stable_tests.push(n);
1851 }
1852 }
1853 } while (stack.size() > 0);
1854
1855 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1856 Node* n = heap_stable_tests.at(i);
1857 assert(is_heap_stable_test(n), "only evacuation test");
1858 merge_back_to_back_tests(n, phase);
1859 }
1860
1861 if (!phase->C->major_progress()) {
1862 VectorSet seen;
1863 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1864 Node* n = heap_stable_tests.at(i);
1865 IdealLoopTree* loop = phase->get_loop(n);
1866 if (loop != phase->ltree_root() &&
1867 loop->_child == nullptr &&
1868 !loop->_irreducible) {
1869 Node* head = loop->_head;
1870 if (head->is_Loop() &&
1871 (!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) &&
1872 !seen.test_set(head->_idx)) {
1873 IfNode* iff = find_unswitching_candidate(loop, phase);
1874 if (iff != nullptr) {
1875 Node* bol = iff->in(1);
1876 if (head->as_Loop()->is_strip_mined()) {
1877 head->as_Loop()->verify_strip_mined(0);
1878 }
1879 move_gc_state_test_out_of_loop(iff, phase);
1880
1881 AutoNodeBudget node_budget(phase);
1882
1883 if (loop->policy_unswitching(phase)) {
1884 if (head->as_Loop()->is_strip_mined()) {
1885 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop();
1886 hide_strip_mined_loop(outer, head->as_CountedLoop(), phase);
1887 }
1888 phase->do_unswitching(loop, old_new);
1889 } else {
1890 // Not proceeding with unswitching. Move load back in
1891 // the loop.
1892 phase->igvn().replace_input_of(iff, 1, bol);
1893 }
1894 }
1895 }
1896 }
1897 }
1898 }
1899 }
1900
1901 ShenandoahIUBarrierNode::ShenandoahIUBarrierNode(Node* val) : Node(nullptr, val) {
1902 ShenandoahBarrierSetC2::bsc2()->state()->add_iu_barrier(this);
1903 }
1904
1905 const Type* ShenandoahIUBarrierNode::bottom_type() const {
1906 if (in(1) == nullptr || in(1)->is_top()) {
1907 return Type::TOP;
1908 }
1909 const Type* t = in(1)->bottom_type();
1910 if (t == TypePtr::NULL_PTR) {
1911 return t;
1912 }
1913 return t->is_oopptr();
1914 }
1915
1916 const Type* ShenandoahIUBarrierNode::Value(PhaseGVN* phase) const {
1917 if (in(1) == nullptr) {
1918 return Type::TOP;
1919 }
1920 const Type* t = phase->type(in(1));
1921 if (t == Type::TOP) {
1922 return Type::TOP;
1923 }
1924 if (t == TypePtr::NULL_PTR) {
1925 return t;
1926 }
1927 return t->is_oopptr();
1928 }
1929
1930 int ShenandoahIUBarrierNode::needed(Node* n) {
1931 if (n == nullptr ||
1932 n->is_Allocate() ||
1933 n->Opcode() == Op_ShenandoahIUBarrier ||
1934 n->bottom_type() == TypePtr::NULL_PTR ||
1935 (n->bottom_type()->make_oopptr() != nullptr && n->bottom_type()->make_oopptr()->const_oop() != nullptr)) {
1936 return NotNeeded;
1937 }
1938 if (n->is_Phi() ||
1939 n->is_CMove()) {
1940 return MaybeNeeded;
1941 }
1942 return Needed;
1943 }
1944
1945 Node* ShenandoahIUBarrierNode::next(Node* n) {
1946 for (;;) {
1947 if (n == nullptr) {
1948 return n;
1949 } else if (n->bottom_type() == TypePtr::NULL_PTR) {
1950 return n;
1951 } else if (n->bottom_type()->make_oopptr() != nullptr && n->bottom_type()->make_oopptr()->const_oop() != nullptr) {
1952 return n;
1953 } else if (n->is_ConstraintCast() ||
1954 n->Opcode() == Op_DecodeN ||
1955 n->Opcode() == Op_EncodeP) {
1956 n = n->in(1);
1957 } else if (n->is_Proj()) {
1958 n = n->in(0);
1959 } else {
1960 return n;
1961 }
1962 }
1963 ShouldNotReachHere();
1964 return nullptr;
1965 }
1966
1967 Node* ShenandoahIUBarrierNode::Identity(PhaseGVN* phase) {
1968 PhaseIterGVN* igvn = phase->is_IterGVN();
1969
1970 Node* n = next(in(1));
1971
1972 int cont = needed(n);
1973
1974 if (cont == NotNeeded) {
1975 return in(1);
1976 } else if (cont == MaybeNeeded) {
1977 if (igvn == nullptr) {
1978 phase->record_for_igvn(this);
1979 return this;
1980 } else {
1981 ResourceMark rm;
1982 Unique_Node_List wq;
1983 uint wq_i = 0;
1984
1985 for (;;) {
1986 if (n->is_Phi()) {
1987 for (uint i = 1; i < n->req(); i++) {
1988 Node* m = n->in(i);
1989 if (m != nullptr) {
1990 wq.push(m);
1991 }
1992 }
1993 } else {
1994 assert(n->is_CMove(), "nothing else here");
1995 Node* m = n->in(CMoveNode::IfFalse);
1996 wq.push(m);
1997 m = n->in(CMoveNode::IfTrue);
1998 wq.push(m);
1999 }
2000 Node* orig_n = nullptr;
2001 do {
2002 if (wq_i >= wq.size()) {
2003 return in(1);
2004 }
2005 n = wq.at(wq_i);
2006 wq_i++;
2007 orig_n = n;
2008 n = next(n);
2009 cont = needed(n);
2010 if (cont == Needed) {
2011 return this;
2012 }
2013 } while (cont != MaybeNeeded || (orig_n != n && wq.member(n)));
2014 }
2015 }
2016 }
2017
2018 return this;
2019 }
2020
2021 #ifdef ASSERT
2022 static bool has_never_branch(Node* root) {
2023 for (uint i = 1; i < root->req(); i++) {
2024 Node* in = root->in(i);
2025 if (in != nullptr && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->is_NeverBranch()) {
2026 return true;
2027 }
2028 }
2029 return false;
2030 }
2031 #endif
2032
2033 void MemoryGraphFixer::collect_memory_nodes() {
2034 Node_Stack stack(0);
2035 VectorSet visited;
2036 Node_List regions;
2037
2038 // Walk the raw memory graph and create a mapping from CFG node to
2039 // memory node. Exclude phis for now.
2040 stack.push(_phase->C->root(), 1);
2041 do {
2042 Node* n = stack.node();
2043 int opc = n->Opcode();
2044 uint i = stack.index();
2045 if (i < n->req()) {
2046 Node* mem = nullptr;
2047 if (opc == Op_Root) {
2048 Node* in = n->in(i);
2049 int in_opc = in->Opcode();
2050 if (in_opc == Op_Return || in_opc == Op_Rethrow) {
2051 mem = in->in(TypeFunc::Memory);
2052 } else if (in_opc == Op_Halt) {
2053 if (in->in(0)->is_Region()) {
2054 Node* r = in->in(0);
2055 for (uint j = 1; j < r->req(); j++) {
2056 assert(!r->in(j)->is_NeverBranch(), "");
2057 }
2058 } else {
2059 Node* proj = in->in(0);
2060 assert(proj->is_Proj(), "");
2061 Node* in = proj->in(0);
2062 assert(in->is_CallStaticJava() || in->is_NeverBranch() || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
2063 if (in->is_CallStaticJava()) {
2064 mem = in->in(TypeFunc::Memory);
2065 } else if (in->Opcode() == Op_Catch) {
2066 Node* call = in->in(0)->in(0);
2067 assert(call->is_Call(), "");
2068 mem = call->in(TypeFunc::Memory);
2069 } else if (in->is_NeverBranch()) {
2070 mem = collect_memory_for_infinite_loop(in);
2071 }
2072 }
2073 } else {
2074 #ifdef ASSERT
2075 n->dump();
2076 in->dump();
2077 #endif
2078 ShouldNotReachHere();
2079 }
2080 } else {
2081 assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
2082 assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
2083 mem = n->in(i);
2084 }
2085 i++;
2086 stack.set_index(i);
2087 if (mem == nullptr) {
2088 continue;
2089 }
2090 for (;;) {
2091 if (visited.test_set(mem->_idx) || mem->is_Start()) {
2092 break;
2093 }
2094 if (mem->is_Phi()) {
2095 stack.push(mem, 2);
2096 mem = mem->in(1);
2097 } else if (mem->is_Proj()) {
2098 stack.push(mem, mem->req());
2099 mem = mem->in(0);
2100 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
2101 mem = mem->in(TypeFunc::Memory);
2102 } else if (mem->is_MergeMem()) {
2103 MergeMemNode* mm = mem->as_MergeMem();
2104 mem = mm->memory_at(_alias);
2105 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
2106 assert(_alias == Compile::AliasIdxRaw, "");
2107 stack.push(mem, mem->req());
2108 mem = mem->in(MemNode::Memory);
2109 } else {
2110 #ifdef ASSERT
2111 mem->dump();
2112 #endif
2113 ShouldNotReachHere();
2114 }
2115 }
2116 } else {
2117 if (n->is_Phi()) {
2118 // Nothing
2119 } else if (!n->is_Root()) {
2120 Node* c = get_ctrl(n);
2121 _memory_nodes.map(c->_idx, n);
2122 }
2123 stack.pop();
2124 }
2125 } while(stack.is_nonempty());
2126
2127 // Iterate over CFG nodes in rpo and propagate memory state to
2128 // compute memory state at regions, creating new phis if needed.
2129 Node_List rpo_list;
2130 visited.clear();
2131 _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
2132 Node* root = rpo_list.pop();
2133 assert(root == _phase->C->root(), "");
2134
2135 const bool trace = false;
2136 #ifdef ASSERT
2137 if (trace) {
2138 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2139 Node* c = rpo_list.at(i);
2140 if (_memory_nodes[c->_idx] != nullptr) {
2141 tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump();
2142 }
2143 }
2144 }
2145 #endif
2146 uint last = _phase->C->unique();
2147
2148 #ifdef ASSERT
2149 uint16_t max_depth = 0;
2150 for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
2151 IdealLoopTree* lpt = iter.current();
2152 max_depth = MAX2(max_depth, lpt->_nest);
2153 }
2154 #endif
2155
2156 bool progress = true;
2157 int iteration = 0;
2158 Node_List dead_phis;
2159 while (progress) {
2160 progress = false;
2161 iteration++;
2162 assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2163 if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
2164
2165 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2166 Node* c = rpo_list.at(i);
2167
2168 Node* prev_mem = _memory_nodes[c->_idx];
2169 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2170 Node* prev_region = regions[c->_idx];
2171 Node* unique = nullptr;
2172 for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
2173 Node* m = _memory_nodes[c->in(j)->_idx];
2174 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");
2175 if (m != nullptr) {
2176 if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
2177 assert((c->is_Loop() && j == LoopNode::LoopBackControl) || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2178 // continue
2179 } else if (unique == nullptr) {
2180 unique = m;
2181 } else if (m == unique) {
2182 // continue
2183 } else {
2184 unique = NodeSentinel;
2185 }
2186 }
2187 }
2188 assert(unique != nullptr, "empty phi???");
2189 if (unique != NodeSentinel) {
2190 if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c) {
2191 dead_phis.push(prev_region);
2192 }
2193 regions.map(c->_idx, unique);
2194 } else {
2195 Node* phi = nullptr;
2196 if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
2197 phi = prev_region;
2198 for (uint k = 1; k < c->req(); k++) {
2199 Node* m = _memory_nodes[c->in(k)->_idx];
2200 assert(m != nullptr, "expect memory state");
2201 phi->set_req(k, m);
2202 }
2203 } else {
2204 for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == nullptr; j++) {
2205 Node* u = c->fast_out(j);
2206 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2207 (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
2208 phi = u;
2209 for (uint k = 1; k < c->req() && phi != nullptr; k++) {
2210 Node* m = _memory_nodes[c->in(k)->_idx];
2211 assert(m != nullptr, "expect memory state");
2212 if (u->in(k) != m) {
2213 phi = NodeSentinel;
2214 }
2215 }
2216 }
2217 }
2218 if (phi == NodeSentinel) {
2219 phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
2220 for (uint k = 1; k < c->req(); k++) {
2221 Node* m = _memory_nodes[c->in(k)->_idx];
2222 assert(m != nullptr, "expect memory state");
2223 phi->init_req(k, m);
2224 }
2225 }
2226 }
2227 if (phi != nullptr) {
2228 regions.map(c->_idx, phi);
2229 } else {
2230 assert(c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2231 }
2232 }
2233 Node* current_region = regions[c->_idx];
2234 if (current_region != prev_region) {
2235 progress = true;
2236 if (prev_region == prev_mem) {
2237 _memory_nodes.map(c->_idx, current_region);
2238 }
2239 }
2240 } else if (prev_mem == nullptr || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
2241 Node* m = _memory_nodes[_phase->idom(c)->_idx];
2242 assert(m != nullptr || c->Opcode() == Op_Halt, "expect memory state");
2243 if (m != prev_mem) {
2244 _memory_nodes.map(c->_idx, m);
2245 progress = true;
2246 }
2247 }
2248 #ifdef ASSERT
2249 if (trace) { tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); }
2250 #endif
2251 }
2252 }
2253
2254 // Replace existing phi with computed memory state for that region
2255 // if different (could be a new phi or a dominating memory node if
2256 // that phi was found to be useless).
2257 while (dead_phis.size() > 0) {
2258 Node* n = dead_phis.pop();
2259 n->replace_by(_phase->C->top());
2260 n->destruct(&_phase->igvn());
2261 }
2262 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2263 Node* c = rpo_list.at(i);
2264 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2265 Node* n = regions[c->_idx];
2266 assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2267 if (n != nullptr && n->is_Phi() && n->_idx >= last && n->in(0) == c) {
2268 _phase->register_new_node(n, c);
2269 }
2270 }
2271 }
2272 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2273 Node* c = rpo_list.at(i);
2274 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2275 Node* n = regions[c->_idx];
2276 assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2277 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
2278 Node* u = c->fast_out(i);
2279 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2280 u != n) {
2281 assert(c->unique_ctrl_out()->Opcode() != Op_Halt, "expected memory state");
2282 if (u->adr_type() == TypePtr::BOTTOM) {
2283 fix_memory_uses(u, n, n, c);
2284 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2285 _phase->lazy_replace(u, n);
2286 --i; --imax;
2287 }
2288 }
2289 }
2290 }
2291 }
2292 }
2293
2294 Node* MemoryGraphFixer::collect_memory_for_infinite_loop(const Node* in) {
2295 Node* mem = nullptr;
2296 Node* head = in->in(0);
2297 assert(head->is_Region(), "unexpected infinite loop graph shape");
2298
2299 Node* phi_mem = nullptr;
2300 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2301 Node* u = head->fast_out(j);
2302 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
2303 if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2304 assert(phi_mem == nullptr || phi_mem->adr_type() == TypePtr::BOTTOM, "");
2305 phi_mem = u;
2306 } else if (u->adr_type() == TypePtr::BOTTOM) {
2307 assert(phi_mem == nullptr || _phase->C->get_alias_index(phi_mem->adr_type()) == _alias, "");
2308 if (phi_mem == nullptr) {
2309 phi_mem = u;
2310 }
2311 }
2312 }
2313 }
2314 if (phi_mem == nullptr) {
2315 ResourceMark rm;
2316 Node_Stack stack(0);
2317 stack.push(head, 1);
2318 do {
2319 Node* n = stack.node();
2320 uint i = stack.index();
2321 if (i >= n->req()) {
2322 stack.pop();
2323 } else {
2324 stack.set_index(i + 1);
2325 Node* c = n->in(i);
2326 assert(c != head, "should have found a safepoint on the way");
2327 if (stack.size() != 1 || _phase->is_dominator(head, c)) {
2328 for (;;) {
2329 if (c->is_Region()) {
2330 stack.push(c, 1);
2331 break;
2332 } else if (c->is_SafePoint() && !c->is_CallLeaf()) {
2333 Node* m = c->in(TypeFunc::Memory);
2334 if (m->is_MergeMem()) {
2335 m = m->as_MergeMem()->memory_at(_alias);
2336 }
2337 assert(mem == nullptr || mem == m, "several memory states");
2338 mem = m;
2339 break;
2340 } else {
2341 assert(c != c->in(0), "");
2342 c = c->in(0);
2343 }
2344 }
2345 }
2346 }
2347 } while (stack.size() > 0);
2348 assert(mem != nullptr, "should have found safepoint");
2349 } else {
2350 mem = phi_mem;
2351 }
2352 return mem;
2353 }
2354
2355 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
2356 Node* c = _phase->get_ctrl(n);
2357 if (n->is_Proj() && n->in(0) != nullptr && n->in(0)->is_Call()) {
2358 assert(c == n->in(0), "");
2359 CallNode* call = c->as_Call();
2360 CallProjections projs;
2361 call->extract_projections(&projs, true, false);
2362 if (projs.catchall_memproj != nullptr) {
2363 if (projs.fallthrough_memproj == n) {
2364 c = projs.fallthrough_catchproj;
2365 } else {
2366 assert(projs.catchall_memproj == n, "");
2367 c = projs.catchall_catchproj;
2368 }
2369 }
2370 }
2371 return c;
2372 }
2373
2374 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
2375 if (_phase->has_ctrl(n))
2376 return get_ctrl(n);
2377 else {
2378 assert (n->is_CFG(), "must be a CFG node");
2379 return n;
2380 }
2381 }
2382
2383 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
2384 return m != nullptr && get_ctrl(m) == c;
2385 }
2386
2387 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
2388 assert(n == nullptr || _phase->ctrl_or_self(n) == ctrl, "");
2389 assert(!ctrl->is_Call() || ctrl == n, "projection expected");
2390 #ifdef ASSERT
2391 if ((ctrl->is_Proj() && ctrl->in(0)->is_Call()) ||
2392 (ctrl->is_Catch() && ctrl->in(0)->in(0)->is_Call())) {
2393 CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_Call() : ctrl->in(0)->in(0)->as_Call();
2394 int mems = 0;
2395 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
2396 Node* u = call->fast_out(i);
2397 if (u->bottom_type() == Type::MEMORY) {
2398 mems++;
2399 }
2400 }
2401 assert(mems <= 1, "No node right after call if multiple mem projections");
2402 }
2403 #endif
2404 Node* mem = _memory_nodes[ctrl->_idx];
2405 Node* c = ctrl;
2406 while (!mem_is_valid(mem, c) &&
2407 (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
2408 c = _phase->idom(c);
2409 mem = _memory_nodes[c->_idx];
2410 }
2411 if (n != nullptr && mem_is_valid(mem, c)) {
2412 while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
2413 mem = next_mem(mem, _alias);
2414 }
2415 if (mem->is_MergeMem()) {
2416 mem = mem->as_MergeMem()->memory_at(_alias);
2417 }
2418 if (!mem_is_valid(mem, c)) {
2419 do {
2420 c = _phase->idom(c);
2421 mem = _memory_nodes[c->_idx];
2422 } while (!mem_is_valid(mem, c) &&
2423 (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
2424 }
2425 }
2426 assert(mem->bottom_type() == Type::MEMORY, "");
2427 return mem;
2428 }
2429
2430 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
2431 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2432 Node* use = region->fast_out(i);
2433 if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
2434 (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
2435 return true;
2436 }
2437 }
2438 return false;
2439 }
2440
2441 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
2442 assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
2443 const bool trace = false;
2444 DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
2445 DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
2446 GrowableArray<Node*> phis;
2447 if (mem_for_ctrl != mem) {
2448 Node* old = mem_for_ctrl;
2449 Node* prev = nullptr;
2450 while (old != mem) {
2451 prev = old;
2452 if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
2453 assert(_alias == Compile::AliasIdxRaw, "");
2454 old = old->in(MemNode::Memory);
2455 } else if (old->Opcode() == Op_SCMemProj) {
2456 assert(_alias == Compile::AliasIdxRaw, "");
2457 old = old->in(0);
2458 } else {
2459 ShouldNotReachHere();
2460 }
2461 }
2462 assert(prev != nullptr, "");
2463 if (new_ctrl != ctrl) {
2464 _memory_nodes.map(ctrl->_idx, mem);
2465 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2466 }
2467 uint input = (uint)MemNode::Memory;
2468 _phase->igvn().replace_input_of(prev, input, new_mem);
2469 } else {
2470 uses.clear();
2471 _memory_nodes.map(new_ctrl->_idx, new_mem);
2472 uses.push(new_ctrl);
2473 for(uint next = 0; next < uses.size(); next++ ) {
2474 Node *n = uses.at(next);
2475 assert(n->is_CFG(), "");
2476 DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
2477 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2478 Node* u = n->fast_out(i);
2479 if (!u->is_Root() && u->is_CFG() && u != n) {
2480 Node* m = _memory_nodes[u->_idx];
2481 if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
2482 !has_mem_phi(u) &&
2483 u->unique_ctrl_out()->Opcode() != Op_Halt) {
2484 DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
2485 DEBUG_ONLY(if (trace && m != nullptr) { tty->print("ZZZ mem"); m->dump(); });
2486
2487 if (!mem_is_valid(m, u) || !m->is_Phi()) {
2488 bool push = true;
2489 bool create_phi = true;
2490 if (_phase->is_dominator(new_ctrl, u)) {
2491 create_phi = false;
2492 }
2493 if (create_phi) {
2494 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
2495 _phase->register_new_node(phi, u);
2496 phis.push(phi);
2497 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
2498 if (!mem_is_valid(m, u)) {
2499 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
2500 _memory_nodes.map(u->_idx, phi);
2501 } else {
2502 DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
2503 for (;;) {
2504 assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), "");
2505 Node* next = nullptr;
2506 if (m->is_Proj()) {
2507 next = m->in(0);
2508 } else {
2509 assert(m->is_Mem() || m->is_LoadStore(), "");
2510 assert(_alias == Compile::AliasIdxRaw, "");
2511 next = m->in(MemNode::Memory);
2512 }
2513 if (_phase->get_ctrl(next) != u) {
2514 break;
2515 }
2516 if (next->is_MergeMem()) {
2517 assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
2518 break;
2519 }
2520 if (next->is_Phi()) {
2521 assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
2522 break;
2523 }
2524 m = next;
2525 }
2526
2527 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
2528 assert(m->is_Mem() || m->is_LoadStore(), "");
2529 uint input = (uint)MemNode::Memory;
2530 _phase->igvn().replace_input_of(m, input, phi);
2531 push = false;
2532 }
2533 } else {
2534 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
2535 }
2536 if (push) {
2537 uses.push(u);
2538 }
2539 }
2540 } else if (!mem_is_valid(m, u) &&
2541 !(u->Opcode() == Op_CProj && u->in(0)->is_NeverBranch() && u->as_Proj()->_con == 1)) {
2542 uses.push(u);
2543 }
2544 }
2545 }
2546 }
2547 for (int i = 0; i < phis.length(); i++) {
2548 Node* n = phis.at(i);
2549 Node* r = n->in(0);
2550 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
2551 for (uint j = 1; j < n->req(); j++) {
2552 Node* m = find_mem(r->in(j), nullptr);
2553 _phase->igvn().replace_input_of(n, j, m);
2554 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
2555 }
2556 }
2557 }
2558 uint last = _phase->C->unique();
2559 MergeMemNode* mm = nullptr;
2560 int alias = _alias;
2561 DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
2562 // Process loads first to not miss an anti-dependency: if the memory
2563 // edge of a store is updated before a load is processed then an
2564 // anti-dependency may be missed.
2565 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2566 Node* u = mem->out(i);
2567 if (u->_idx < last && u->is_Load() && _phase->C->get_alias_index(u->adr_type()) == alias) {
2568 Node* m = find_mem(_phase->get_ctrl(u), u);
2569 if (m != mem) {
2570 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2571 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2572 --i;
2573 }
2574 }
2575 }
2576 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2577 Node* u = mem->out(i);
2578 if (u->_idx < last) {
2579 if (u->is_Mem()) {
2580 if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2581 Node* m = find_mem(_phase->get_ctrl(u), u);
2582 if (m != mem) {
2583 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2584 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2585 --i;
2586 }
2587 }
2588 } else if (u->is_MergeMem()) {
2589 MergeMemNode* u_mm = u->as_MergeMem();
2590 if (u_mm->memory_at(alias) == mem) {
2591 MergeMemNode* newmm = nullptr;
2592 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2593 Node* uu = u->fast_out(j);
2594 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2595 if (uu->is_Phi()) {
2596 assert(uu->adr_type() == TypePtr::BOTTOM, "");
2597 Node* region = uu->in(0);
2598 int nb = 0;
2599 for (uint k = 1; k < uu->req(); k++) {
2600 if (uu->in(k) == u) {
2601 Node* m = find_mem(region->in(k), nullptr);
2602 if (m != mem) {
2603 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
2604 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2605 if (newmm != u) {
2606 _phase->igvn().replace_input_of(uu, k, newmm);
2607 nb++;
2608 --jmax;
2609 }
2610 }
2611 }
2612 }
2613 if (nb > 0) {
2614 --j;
2615 }
2616 } else {
2617 Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
2618 if (m != mem) {
2619 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
2620 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2621 if (newmm != u) {
2622 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2623 --j, --jmax;
2624 }
2625 }
2626 }
2627 }
2628 }
2629 } else if (u->is_Phi()) {
2630 assert(u->bottom_type() == Type::MEMORY, "what else?");
2631 if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
2632 Node* region = u->in(0);
2633 bool replaced = false;
2634 for (uint j = 1; j < u->req(); j++) {
2635 if (u->in(j) == mem) {
2636 Node* m = find_mem(region->in(j), nullptr);
2637 Node* nnew = m;
2638 if (m != mem) {
2639 if (u->adr_type() == TypePtr::BOTTOM) {
2640 mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
2641 nnew = mm;
2642 }
2643 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
2644 _phase->igvn().replace_input_of(u, j, nnew);
2645 replaced = true;
2646 }
2647 }
2648 }
2649 if (replaced) {
2650 --i;
2651 }
2652 }
2653 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2654 u->adr_type() == nullptr) {
2655 assert(u->adr_type() != nullptr ||
2656 u->Opcode() == Op_Rethrow ||
2657 u->Opcode() == Op_Return ||
2658 u->Opcode() == Op_SafePoint ||
2659 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2660 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2661 u->Opcode() == Op_CallLeaf, "");
2662 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2663 if (m != mem) {
2664 mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
2665 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2666 --i;
2667 }
2668 } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2669 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2670 if (m != mem) {
2671 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2672 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2673 --i;
2674 }
2675 } else if (u->adr_type() != TypePtr::BOTTOM &&
2676 _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
2677 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2678 assert(m != mem, "");
2679 // u is on the wrong slice...
2680 assert(u->is_ClearArray(), "");
2681 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2682 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2683 --i;
2684 }
2685 }
2686 }
2687 #ifdef ASSERT
2688 assert(new_mem->outcnt() > 0, "");
2689 for (int i = 0; i < phis.length(); i++) {
2690 Node* n = phis.at(i);
2691 assert(n->outcnt() > 0, "new phi must have uses now");
2692 }
2693 #endif
2694 }
2695
2696 void MemoryGraphFixer::record_new_ctrl(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl) {
2697 if (mem_for_ctrl != mem && new_ctrl != ctrl) {
2698 _memory_nodes.map(ctrl->_idx, mem);
2699 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2700 }
2701 }
2702
2703 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
2704 MergeMemNode* mm = MergeMemNode::make(mem);
2705 mm->set_memory_at(_alias, rep_proj);
2706 _phase->register_new_node(mm, rep_ctrl);
2707 return mm;
2708 }
2709
2710 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
2711 MergeMemNode* newmm = nullptr;
2712 MergeMemNode* u_mm = u->as_MergeMem();
2713 Node* c = _phase->get_ctrl(u);
2714 if (_phase->is_dominator(c, rep_ctrl)) {
2715 c = rep_ctrl;
2716 } else {
2717 assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
2718 }
2719 if (u->outcnt() == 1) {
2720 if (u->req() > (uint)_alias && u->in(_alias) == mem) {
2721 _phase->igvn().replace_input_of(u, _alias, rep_proj);
2722 --i;
2723 } else {
2724 _phase->igvn().rehash_node_delayed(u);
2725 u_mm->set_memory_at(_alias, rep_proj);
2726 }
2727 newmm = u_mm;
2728 _phase->set_ctrl_and_loop(u, c);
2729 } else {
2730 // can't simply clone u and then change one of its input because
2731 // it adds and then removes an edge which messes with the
2732 // DUIterator
2733 newmm = MergeMemNode::make(u_mm->base_memory());
2734 for (uint j = 0; j < u->req(); j++) {
2735 if (j < newmm->req()) {
2736 if (j == (uint)_alias) {
2737 newmm->set_req(j, rep_proj);
2738 } else if (newmm->in(j) != u->in(j)) {
2739 newmm->set_req(j, u->in(j));
2740 }
2741 } else if (j == (uint)_alias) {
2742 newmm->add_req(rep_proj);
2743 } else {
2744 newmm->add_req(u->in(j));
2745 }
2746 }
2747 if ((uint)_alias >= u->req()) {
2748 newmm->set_memory_at(_alias, rep_proj);
2749 }
2750 _phase->register_new_node(newmm, c);
2751 }
2752 return newmm;
2753 }
2754
2755 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
2756 if (phi->adr_type() == TypePtr::BOTTOM) {
2757 Node* region = phi->in(0);
2758 for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
2759 Node* uu = region->fast_out(j);
2760 if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
2761 return false;
2762 }
2763 }
2764 return true;
2765 }
2766 return _phase->C->get_alias_index(phi->adr_type()) == _alias;
2767 }
2768
2769 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
2770 uint last = _phase-> C->unique();
2771 MergeMemNode* mm = nullptr;
2772 assert(mem->bottom_type() == Type::MEMORY, "");
2773 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2774 Node* u = mem->out(i);
2775 if (u != replacement && u->_idx < last) {
2776 if (u->is_MergeMem()) {
2777 MergeMemNode* u_mm = u->as_MergeMem();
2778 if (u_mm->memory_at(_alias) == mem) {
2779 MergeMemNode* newmm = nullptr;
2780 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2781 Node* uu = u->fast_out(j);
2782 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2783 if (uu->is_Phi()) {
2784 if (should_process_phi(uu)) {
2785 Node* region = uu->in(0);
2786 int nb = 0;
2787 for (uint k = 1; k < uu->req(); k++) {
2788 if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
2789 if (newmm == nullptr) {
2790 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2791 }
2792 if (newmm != u) {
2793 _phase->igvn().replace_input_of(uu, k, newmm);
2794 nb++;
2795 --jmax;
2796 }
2797 }
2798 }
2799 if (nb > 0) {
2800 --j;
2801 }
2802 }
2803 } else {
2804 if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
2805 if (newmm == nullptr) {
2806 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2807 }
2808 if (newmm != u) {
2809 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2810 --j, --jmax;
2811 }
2812 }
2813 }
2814 }
2815 }
2816 } else if (u->is_Phi()) {
2817 assert(u->bottom_type() == Type::MEMORY, "what else?");
2818 Node* region = u->in(0);
2819 if (should_process_phi(u)) {
2820 bool replaced = false;
2821 for (uint j = 1; j < u->req(); j++) {
2822 if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
2823 Node* nnew = rep_proj;
2824 if (u->adr_type() == TypePtr::BOTTOM) {
2825 if (mm == nullptr) {
2826 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2827 }
2828 nnew = mm;
2829 }
2830 _phase->igvn().replace_input_of(u, j, nnew);
2831 replaced = true;
2832 }
2833 }
2834 if (replaced) {
2835 --i;
2836 }
2837
2838 }
2839 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2840 u->adr_type() == nullptr) {
2841 assert(u->adr_type() != nullptr ||
2842 u->Opcode() == Op_Rethrow ||
2843 u->Opcode() == Op_Return ||
2844 u->Opcode() == Op_SafePoint ||
2845 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2846 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2847 u->Opcode() == Op_CallLeaf, "%s", u->Name());
2848 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2849 if (mm == nullptr) {
2850 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2851 }
2852 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2853 --i;
2854 }
2855 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2856 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2857 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
2858 --i;
2859 }
2860 }
2861 }
2862 }
2863 }
2864
2865 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj, DecoratorSet decorators)
2866 : Node(ctrl, obj), _decorators(decorators) {
2867 ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this);
2868 }
2869
2870 DecoratorSet ShenandoahLoadReferenceBarrierNode::decorators() const {
2871 return _decorators;
2872 }
2873
2874 uint ShenandoahLoadReferenceBarrierNode::size_of() const {
2875 return sizeof(*this);
2876 }
2877
2878 static DecoratorSet mask_decorators(DecoratorSet decorators) {
2879 return decorators & (ON_STRONG_OOP_REF | ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF | ON_UNKNOWN_OOP_REF | IN_NATIVE);
2880 }
2881
2882 uint ShenandoahLoadReferenceBarrierNode::hash() const {
2883 uint hash = Node::hash();
2884 hash += mask_decorators(_decorators);
2885 return hash;
2886 }
2887
2888 bool ShenandoahLoadReferenceBarrierNode::cmp( const Node &n ) const {
2889 return Node::cmp(n) && n.Opcode() == Op_ShenandoahLoadReferenceBarrier &&
2890 mask_decorators(_decorators) == mask_decorators(((const ShenandoahLoadReferenceBarrierNode&)n)._decorators);
2891 }
2892
2893 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const {
2894 if (in(ValueIn) == nullptr || in(ValueIn)->is_top()) {
2895 return Type::TOP;
2896 }
2897 const Type* t = in(ValueIn)->bottom_type();
2898 if (t == TypePtr::NULL_PTR) {
2899 return t;
2900 }
2901
2902 if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2903 return t;
2904 }
2905
2906 return t->meet(TypePtr::NULL_PTR);
2907 }
2908
2909 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const {
2910 // Either input is TOP ==> the result is TOP
2911 const Type *t2 = phase->type(in(ValueIn));
2912 if( t2 == Type::TOP ) return Type::TOP;
2913
2914 if (t2 == TypePtr::NULL_PTR) {
2915 return t2;
2916 }
2917
2918 if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2919 return t2;
2920 }
2921
2922 return t2->meet(TypePtr::NULL_PTR);
2923 }
2924
2925 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) {
2926 Node* value = in(ValueIn);
2927 if (!needs_barrier(phase, value)) {
2928 return value;
2929 }
2930 return this;
2931 }
2932
2933 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) {
2934 Unique_Node_List visited;
2935 return needs_barrier_impl(phase, n, visited);
2936 }
2937
2938 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) {
2939 if (n == nullptr) return false;
2940 if (visited.member(n)) {
2941 return false; // Been there.
2942 }
2943 visited.push(n);
2944
2945 if (n->is_Allocate()) {
2946 // tty->print_cr("optimize barrier on alloc");
2947 return false;
2948 }
2949 if (n->is_Call()) {
2950 // tty->print_cr("optimize barrier on call");
2951 return false;
2952 }
2953
2954 const Type* type = phase->type(n);
2955 if (type == Type::TOP) {
2956 return false;
2957 }
2958 if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) {
2959 // tty->print_cr("optimize barrier on null");
2960 return false;
2961 }
2962 if (type->make_oopptr() && type->make_oopptr()->const_oop() != nullptr) {
2963 // tty->print_cr("optimize barrier on constant");
2964 return false;
2965 }
2966
2967 switch (n->Opcode()) {
2968 case Op_AddP:
2969 return true; // TODO: Can refine?
2970 case Op_LoadP:
2971 case Op_ShenandoahCompareAndExchangeN:
2972 case Op_ShenandoahCompareAndExchangeP:
2973 case Op_CompareAndExchangeN:
2974 case Op_CompareAndExchangeP:
2975 case Op_GetAndSetN:
2976 case Op_GetAndSetP:
2977 return true;
2978 case Op_Phi: {
2979 for (uint i = 1; i < n->req(); i++) {
2980 if (needs_barrier_impl(phase, n->in(i), visited)) return true;
2981 }
2982 return false;
2983 }
2984 case Op_CheckCastPP:
2985 case Op_CastPP:
2986 return needs_barrier_impl(phase, n->in(1), visited);
2987 case Op_Proj:
2988 return needs_barrier_impl(phase, n->in(0), visited);
2989 case Op_ShenandoahLoadReferenceBarrier:
2990 // tty->print_cr("optimize barrier on barrier");
2991 return false;
2992 case Op_Parm:
2993 // tty->print_cr("optimize barrier on input arg");
2994 return false;
2995 case Op_DecodeN:
2996 case Op_EncodeP:
2997 return needs_barrier_impl(phase, n->in(1), visited);
2998 case Op_LoadN:
2999 return true;
3000 case Op_CMoveN:
3001 case Op_CMoveP:
3002 return needs_barrier_impl(phase, n->in(2), visited) ||
3003 needs_barrier_impl(phase, n->in(3), visited);
3004 case Op_ShenandoahIUBarrier:
3005 return needs_barrier_impl(phase, n->in(1), visited);
3006 case Op_CreateEx:
3007 return false;
3008 default:
3009 break;
3010 }
3011 #ifdef ASSERT
3012 tty->print("need barrier on?: ");
3013 tty->print_cr("ins:");
3014 n->dump(2);
3015 tty->print_cr("outs:");
3016 n->dump(-2);
3017 ShouldNotReachHere();
3018 #endif
3019 return true;
3020 }