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