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 if (head->is_BaseCountedLoop() && ctrl->is_IfProj() && ctrl->in(0)->is_BaseCountedLoopEnd() &&
1326 head->as_BaseCountedLoop()->loopexit() == ctrl->in(0)) {
1327 Node* entry = head->in(LoopNode::EntryControl);
1328 Node* backedge = head->in(LoopNode::LoopBackControl);
1329 Node* new_head = new LoopNode(entry, backedge);
1330 phase->register_control(new_head, phase->get_loop(entry), entry);
1331 phase->lazy_replace(head, new_head);
1332 }
1333 }
1334
1335 // Expand load-reference-barriers
1336 MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
1337 Unique_Node_List uses_to_ignore;
1338 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1339 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1340 uint last = phase->C->unique();
1341 Node* ctrl = phase->get_ctrl(lrb);
1342 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1343
1344 Node* orig_ctrl = ctrl;
1345
1346 Node* raw_mem = fixer.find_mem(ctrl, lrb);
1347 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, nullptr);
1348
1349 IdealLoopTree *loop = phase->get_loop(ctrl);
1350
1351 Node* heap_stable_ctrl = nullptr;
1352 Node* null_ctrl = nullptr;
1353
1354 assert(val->bottom_type()->make_oopptr(), "need oop");
1355 assert(val->bottom_type()->make_oopptr()->const_oop() == nullptr, "expect non-constant");
1356
1357 enum { _heap_stable = 1, _evac_path, _not_cset, PATH_LIMIT };
1358 Node* region = new RegionNode(PATH_LIMIT);
1359 Node* val_phi = new PhiNode(region, val->bottom_type()->is_oopptr());
1360
1361 // Stable path.
1362 int flags = ShenandoahHeap::HAS_FORWARDED;
1363 if (!ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1364 flags |= ShenandoahHeap::WEAK_ROOTS;
1365 }
1366 test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, flags);
1367 IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
1368
1369 // Heap stable case
1370 region->init_req(_heap_stable, heap_stable_ctrl);
1371 val_phi->init_req(_heap_stable, val);
1372
1373 // Test for in-cset, unless it's a native-LRB. Native LRBs need to return null
1374 // even for non-cset objects to prevent resurrection of such objects.
1375 // Wires !in_cset(obj) to slot 2 of region and phis
1376 Node* not_cset_ctrl = nullptr;
1377 if (ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1378 test_in_cset(ctrl, not_cset_ctrl, val, raw_mem, phase);
1379 }
1380 if (not_cset_ctrl != nullptr) {
1381 region->init_req(_not_cset, not_cset_ctrl);
1382 val_phi->init_req(_not_cset, val);
1383 } else {
1384 region->del_req(_not_cset);
1385 val_phi->del_req(_not_cset);
1386 }
1387
1388 // Resolve object when orig-value is in cset.
1389 // Make the unconditional resolve for fwdptr.
1390
1391 // Call lrb-stub and wire up that path in slots 4
1392 Node* result_mem = nullptr;
1393
1394 Node* addr;
1395 {
1396 VectorSet visited;
1397 addr = get_load_addr(phase, visited, lrb);
1398 }
1399 if (addr->Opcode() == Op_AddP) {
1400 Node* orig_base = addr->in(AddPNode::Base);
1401 Node* base = new CheckCastPPNode(ctrl, orig_base, orig_base->bottom_type(), ConstraintCastNode::StrongDependency);
1402 phase->register_new_node(base, ctrl);
1403 if (addr->in(AddPNode::Base) == addr->in((AddPNode::Address))) {
1404 // Field access
1405 addr = addr->clone();
1406 addr->set_req(AddPNode::Base, base);
1407 addr->set_req(AddPNode::Address, base);
1408 phase->register_new_node(addr, ctrl);
1409 } else {
1410 Node* addr2 = addr->in(AddPNode::Address);
1411 if (addr2->Opcode() == Op_AddP && addr2->in(AddPNode::Base) == addr2->in(AddPNode::Address) &&
1412 addr2->in(AddPNode::Base) == orig_base) {
1413 addr2 = addr2->clone();
1414 addr2->set_req(AddPNode::Base, base);
1415 addr2->set_req(AddPNode::Address, base);
1416 phase->register_new_node(addr2, ctrl);
1417 addr = addr->clone();
1418 addr->set_req(AddPNode::Base, base);
1419 addr->set_req(AddPNode::Address, addr2);
1420 phase->register_new_node(addr, ctrl);
1421 }
1422 }
1423 }
1424 call_lrb_stub(ctrl, val, addr, lrb->decorators(), phase);
1425 region->init_req(_evac_path, ctrl);
1426 val_phi->init_req(_evac_path, val);
1427
1428 phase->register_control(region, loop, heap_stable_iff);
1429 Node* out_val = val_phi;
1430 phase->register_new_node(val_phi, region);
1431
1432 fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase);
1433
1434 ctrl = orig_ctrl;
1435
1436 phase->igvn().replace_node(lrb, out_val);
1437
1438 follow_barrier_uses(out_val, ctrl, uses, phase);
1439
1440 for(uint next = 0; next < uses.size(); next++ ) {
1441 Node *n = uses.at(next);
1442 assert(phase->get_ctrl(n) == ctrl, "bad control");
1443 assert(n != raw_mem, "should leave input raw mem above the barrier");
1444 phase->set_ctrl(n, region);
1445 follow_barrier_uses(n, ctrl, uses, phase);
1446 }
1447 fixer.record_new_ctrl(ctrl, region, raw_mem, raw_mem_for_ctrl);
1448 }
1449 // Done expanding load-reference-barriers.
1450 assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced");
1451
1452 for (int i = state->iu_barriers_count() - 1; i >= 0; i--) {
1453 Node* barrier = state->iu_barrier(i);
1454 Node* pre_val = barrier->in(1);
1455
1456 if (phase->igvn().type(pre_val)->higher_equal(TypePtr::NULL_PTR)) {
1457 ShouldNotReachHere();
1458 continue;
1459 }
1460
1461 Node* ctrl = phase->get_ctrl(barrier);
1462
1463 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) {
1464 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0)->in(0), pre_val, ctrl->in(0), phase), "can't move");
1465 ctrl = ctrl->in(0)->in(0);
1466 phase->set_ctrl(barrier, ctrl);
1467 } else if (ctrl->is_CallRuntime()) {
1468 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0), pre_val, ctrl, phase), "can't move");
1469 ctrl = ctrl->in(0);
1470 phase->set_ctrl(barrier, ctrl);
1471 }
1472
1473 Node* init_ctrl = ctrl;
1474 IdealLoopTree* loop = phase->get_loop(ctrl);
1475 Node* raw_mem = fixer.find_mem(ctrl, barrier);
1476 Node* init_raw_mem = raw_mem;
1477 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, nullptr);
1478 Node* heap_stable_ctrl = nullptr;
1479 Node* null_ctrl = nullptr;
1480 uint last = phase->C->unique();
1481
1482 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
1483 Node* region = new RegionNode(PATH_LIMIT);
1484 Node* phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1485
1486 enum { _fast_path = 1, _slow_path, _null_path, PATH_LIMIT2 };
1487 Node* region2 = new RegionNode(PATH_LIMIT2);
1488 Node* phi2 = PhiNode::make(region2, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1489
1490 // Stable path.
1491 test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, ShenandoahHeap::MARKING);
1492 region->init_req(_heap_stable, heap_stable_ctrl);
1493 phi->init_req(_heap_stable, raw_mem);
1494
1495 // Null path
1496 Node* reg2_ctrl = nullptr;
1497 test_null(ctrl, pre_val, null_ctrl, phase);
1498 if (null_ctrl != nullptr) {
1499 reg2_ctrl = null_ctrl->in(0);
1500 region2->init_req(_null_path, null_ctrl);
1501 phi2->init_req(_null_path, raw_mem);
1502 } else {
1503 region2->del_req(_null_path);
1504 phi2->del_req(_null_path);
1505 }
1506
1507 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
1508 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1509 Node* thread = new ThreadLocalNode();
1510 phase->register_new_node(thread, ctrl);
1511 Node* buffer_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(buffer_offset));
1512 phase->register_new_node(buffer_adr, ctrl);
1513 Node* index_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(index_offset));
1514 phase->register_new_node(index_adr, ctrl);
1515
1516 BasicType index_bt = TypeX_X->basic_type();
1517 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading Shenandoah SATBMarkQueue::_index with wrong size.");
1518 const TypePtr* adr_type = TypeRawPtr::BOTTOM;
1519 Node* index = new LoadXNode(ctrl, raw_mem, index_adr, adr_type, TypeX_X, MemNode::unordered);
1520 phase->register_new_node(index, ctrl);
1521 Node* index_cmp = new CmpXNode(index, phase->igvn().MakeConX(0));
1522 phase->register_new_node(index_cmp, ctrl);
1523 Node* index_test = new BoolNode(index_cmp, BoolTest::ne);
1524 phase->register_new_node(index_test, ctrl);
1525 IfNode* queue_full_iff = new IfNode(ctrl, index_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
1526 if (reg2_ctrl == nullptr) reg2_ctrl = queue_full_iff;
1527 phase->register_control(queue_full_iff, loop, ctrl);
1528 Node* not_full = new IfTrueNode(queue_full_iff);
1529 phase->register_control(not_full, loop, queue_full_iff);
1530 Node* full = new IfFalseNode(queue_full_iff);
1531 phase->register_control(full, loop, queue_full_iff);
1532
1533 ctrl = not_full;
1534
1535 Node* next_index = new SubXNode(index, phase->igvn().MakeConX(sizeof(intptr_t)));
1536 phase->register_new_node(next_index, ctrl);
1537
1538 Node* buffer = new LoadPNode(ctrl, raw_mem, buffer_adr, adr_type, TypeRawPtr::NOTNULL, MemNode::unordered);
1539 phase->register_new_node(buffer, ctrl);
1540 Node *log_addr = new AddPNode(phase->C->top(), buffer, next_index);
1541 phase->register_new_node(log_addr, ctrl);
1542 Node* log_store = new StorePNode(ctrl, raw_mem, log_addr, adr_type, pre_val, MemNode::unordered);
1543 phase->register_new_node(log_store, ctrl);
1544 // update the index
1545 Node* index_update = new StoreXNode(ctrl, log_store, index_adr, adr_type, next_index, MemNode::unordered);
1546 phase->register_new_node(index_update, ctrl);
1547
1548 // Fast-path case
1549 region2->init_req(_fast_path, ctrl);
1550 phi2->init_req(_fast_path, index_update);
1551
1552 ctrl = full;
1553
1554 Node* base = find_bottom_mem(ctrl, phase);
1555
1556 MergeMemNode* mm = MergeMemNode::make(base);
1557 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
1558 phase->register_new_node(mm, ctrl);
1559
1560 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);
1561 call->init_req(TypeFunc::Control, ctrl);
1562 call->init_req(TypeFunc::I_O, phase->C->top());
1563 call->init_req(TypeFunc::Memory, mm);
1564 call->init_req(TypeFunc::FramePtr, phase->C->top());
1565 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1566 call->init_req(TypeFunc::Parms, pre_val);
1567 call->init_req(TypeFunc::Parms+1, thread);
1568 phase->register_control(call, loop, ctrl);
1569
1570 Node* ctrl_proj = new ProjNode(call, TypeFunc::Control);
1571 phase->register_control(ctrl_proj, loop, call);
1572 Node* mem_proj = new ProjNode(call, TypeFunc::Memory);
1573 phase->register_new_node(mem_proj, call);
1574
1575 // Slow-path case
1576 region2->init_req(_slow_path, ctrl_proj);
1577 phi2->init_req(_slow_path, mem_proj);
1578
1579 phase->register_control(region2, loop, reg2_ctrl);
1580 phase->register_new_node(phi2, region2);
1581
1582 region->init_req(_heap_unstable, region2);
1583 phi->init_req(_heap_unstable, phi2);
1584
1585 phase->register_control(region, loop, heap_stable_ctrl->in(0));
1586 phase->register_new_node(phi, region);
1587
1588 fix_ctrl(barrier, region, fixer, uses, uses_to_ignore, last, phase);
1589 for(uint next = 0; next < uses.size(); next++ ) {
1590 Node *n = uses.at(next);
1591 assert(phase->get_ctrl(n) == init_ctrl, "bad control");
1592 assert(n != init_raw_mem, "should leave input raw mem above the barrier");
1593 phase->set_ctrl(n, region);
1594 follow_barrier_uses(n, init_ctrl, uses, phase);
1595 }
1596 fixer.fix_mem(init_ctrl, region, init_raw_mem, raw_mem_for_ctrl, phi, uses);
1597
1598 phase->igvn().replace_node(barrier, pre_val);
1599 }
1600 assert(state->iu_barriers_count() == 0, "all enqueue barrier nodes should have been replaced");
1601
1602 }
1603
1604 Node* ShenandoahBarrierC2Support::get_load_addr(PhaseIdealLoop* phase, VectorSet& visited, Node* in) {
1605 if (visited.test_set(in->_idx)) {
1606 return nullptr;
1607 }
1608 switch (in->Opcode()) {
1609 case Op_Proj:
1610 return get_load_addr(phase, visited, in->in(0));
1611 case Op_CastPP:
1612 case Op_CheckCastPP:
1613 case Op_DecodeN:
1614 case Op_EncodeP:
1615 return get_load_addr(phase, visited, in->in(1));
1616 case Op_LoadN:
1617 case Op_LoadP:
1618 return in->in(MemNode::Address);
1619 case Op_CompareAndExchangeN:
1620 case Op_CompareAndExchangeP:
1621 case Op_GetAndSetN:
1622 case Op_GetAndSetP:
1623 case Op_ShenandoahCompareAndExchangeP:
1624 case Op_ShenandoahCompareAndExchangeN:
1625 // Those instructions would just have stored a different
1626 // value into the field. No use to attempt to fix it at this point.
1627 return phase->igvn().zerocon(T_OBJECT);
1628 case Op_CMoveP:
1629 case Op_CMoveN: {
1630 Node* t = get_load_addr(phase, visited, in->in(CMoveNode::IfTrue));
1631 Node* f = get_load_addr(phase, visited, in->in(CMoveNode::IfFalse));
1632 // Handle unambiguous cases: single address reported on both branches.
1633 if (t != nullptr && f == nullptr) return t;
1634 if (t == nullptr && f != nullptr) return f;
1635 if (t != nullptr && t == f) return t;
1636 // Ambiguity.
1637 return phase->igvn().zerocon(T_OBJECT);
1638 }
1639 case Op_Phi: {
1640 Node* addr = nullptr;
1641 for (uint i = 1; i < in->req(); i++) {
1642 Node* addr1 = get_load_addr(phase, visited, in->in(i));
1643 if (addr == nullptr) {
1644 addr = addr1;
1645 }
1646 if (addr != addr1) {
1647 return phase->igvn().zerocon(T_OBJECT);
1648 }
1649 }
1650 return addr;
1651 }
1652 case Op_ShenandoahLoadReferenceBarrier:
1653 return get_load_addr(phase, visited, in->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1654 case Op_ShenandoahIUBarrier:
1655 return get_load_addr(phase, visited, in->in(1));
1656 case Op_CallDynamicJava:
1657 case Op_CallLeaf:
1658 case Op_CallStaticJava:
1659 case Op_ConN:
1660 case Op_ConP:
1661 case Op_Parm:
1662 case Op_CreateEx:
1663 return phase->igvn().zerocon(T_OBJECT);
1664 default:
1665 #ifdef ASSERT
1666 fatal("Unknown node in get_load_addr: %s", NodeClassNames[in->Opcode()]);
1667 #endif
1668 return phase->igvn().zerocon(T_OBJECT);
1669 }
1670
1671 }
1672
1673 void ShenandoahBarrierC2Support::move_gc_state_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
1674 IdealLoopTree *loop = phase->get_loop(iff);
1675 Node* loop_head = loop->_head;
1676 Node* entry_c = loop_head->in(LoopNode::EntryControl);
1677
1678 Node* bol = iff->in(1);
1679 Node* cmp = bol->in(1);
1680 Node* andi = cmp->in(1);
1681 Node* load = andi->in(1);
1682
1683 assert(is_gc_state_load(load), "broken");
1684 if (!phase->is_dominator(load->in(0), entry_c)) {
1685 Node* mem_ctrl = nullptr;
1686 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
1687 load = load->clone();
1688 load->set_req(MemNode::Memory, mem);
1689 load->set_req(0, entry_c);
1690 phase->register_new_node(load, entry_c);
1691 andi = andi->clone();
1692 andi->set_req(1, load);
1693 phase->register_new_node(andi, entry_c);
1694 cmp = cmp->clone();
1695 cmp->set_req(1, andi);
1696 phase->register_new_node(cmp, entry_c);
1697 bol = bol->clone();
1698 bol->set_req(1, cmp);
1699 phase->register_new_node(bol, entry_c);
1700
1701 phase->igvn().replace_input_of(iff, 1, bol);
1702 }
1703 }
1704
1705 bool ShenandoahBarrierC2Support::identical_backtoback_ifs(Node* n, PhaseIdealLoop* phase) {
1706 if (!n->is_If() || n->is_CountedLoopEnd()) {
1707 return false;
1708 }
1709 Node* region = n->in(0);
1710
1711 if (!region->is_Region()) {
1712 return false;
1713 }
1714 Node* dom = phase->idom(region);
1715 if (!dom->is_If()) {
1716 return false;
1717 }
1718
1719 if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) {
1720 return false;
1721 }
1722
1723 IfNode* dom_if = dom->as_If();
1724 Node* proj_true = dom_if->proj_out(1);
1725 Node* proj_false = dom_if->proj_out(0);
1726
1727 for (uint i = 1; i < region->req(); i++) {
1728 if (phase->is_dominator(proj_true, region->in(i))) {
1729 continue;
1730 }
1731 if (phase->is_dominator(proj_false, region->in(i))) {
1732 continue;
1733 }
1734 return false;
1735 }
1736
1737 return true;
1738 }
1739
1740 bool ShenandoahBarrierC2Support::merge_point_safe(Node* region) {
1741 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1742 Node* n = region->fast_out(i);
1743 if (n->is_LoadStore()) {
1744 // Splitting a LoadStore node through phi, causes it to lose its SCMemProj: the split if code doesn't have support
1745 // for a LoadStore at the region the if is split through because that's not expected to happen (LoadStore nodes
1746 // should be between barrier nodes). It does however happen with Shenandoah though because barriers can get
1747 // expanded around a LoadStore node.
1748 return false;
1749 }
1750 }
1751 return true;
1752 }
1753
1754
1755 void ShenandoahBarrierC2Support::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) {
1756 assert(is_heap_stable_test(n), "no other tests");
1757 if (identical_backtoback_ifs(n, phase)) {
1758 Node* n_ctrl = n->in(0);
1759 if (phase->can_split_if(n_ctrl) && merge_point_safe(n_ctrl)) {
1760 IfNode* dom_if = phase->idom(n_ctrl)->as_If();
1761 if (is_heap_stable_test(n)) {
1762 Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1);
1763 assert(is_gc_state_load(gc_state_load), "broken");
1764 Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1);
1765 assert(is_gc_state_load(dom_gc_state_load), "broken");
1766 if (gc_state_load != dom_gc_state_load) {
1767 phase->igvn().replace_node(gc_state_load, dom_gc_state_load);
1768 }
1769 }
1770 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1771 Node* proj_true = dom_if->proj_out(1);
1772 Node* proj_false = dom_if->proj_out(0);
1773 Node* con_true = phase->igvn().makecon(TypeInt::ONE);
1774 Node* con_false = phase->igvn().makecon(TypeInt::ZERO);
1775
1776 for (uint i = 1; i < n_ctrl->req(); i++) {
1777 if (phase->is_dominator(proj_true, n_ctrl->in(i))) {
1778 bolphi->init_req(i, con_true);
1779 } else {
1780 assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1781 bolphi->init_req(i, con_false);
1782 }
1783 }
1784 phase->register_new_node(bolphi, n_ctrl);
1785 phase->igvn().replace_input_of(n, 1, bolphi);
1786 phase->do_split_if(n);
1787 }
1788 }
1789 }
1790
1791 IfNode* ShenandoahBarrierC2Support::find_unswitching_candidate(const IdealLoopTree* loop, PhaseIdealLoop* phase) {
1792 // Find first invariant test that doesn't exit the loop
1793 LoopNode *head = loop->_head->as_Loop();
1794 IfNode* unswitch_iff = nullptr;
1795 Node* n = head->in(LoopNode::LoopBackControl);
1796 int loop_has_sfpts = -1;
1797 while (n != head) {
1798 Node* n_dom = phase->idom(n);
1799 if (n->is_Region()) {
1800 if (n_dom->is_If()) {
1801 IfNode* iff = n_dom->as_If();
1802 if (iff->in(1)->is_Bool()) {
1803 BoolNode* bol = iff->in(1)->as_Bool();
1804 if (bol->in(1)->is_Cmp()) {
1805 // If condition is invariant and not a loop exit,
1806 // then found reason to unswitch.
1807 if (is_heap_stable_test(iff) &&
1808 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) {
1809 assert(!loop->is_loop_exit(iff), "both branches should be in the loop");
1810 if (loop_has_sfpts == -1) {
1811 for(uint i = 0; i < loop->_body.size(); i++) {
1812 Node *m = loop->_body[i];
1813 if (m->is_SafePoint() && !m->is_CallLeaf()) {
1814 loop_has_sfpts = 1;
1815 break;
1816 }
1817 }
1818 if (loop_has_sfpts == -1) {
1819 loop_has_sfpts = 0;
1820 }
1821 }
1822 if (!loop_has_sfpts) {
1823 unswitch_iff = iff;
1824 }
1825 }
1826 }
1827 }
1828 }
1829 }
1830 n = n_dom;
1831 }
1832 return unswitch_iff;
1833 }
1834
1835
1836 void ShenandoahBarrierC2Support::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) {
1837 Node_List heap_stable_tests;
1838 stack.push(phase->C->start(), 0);
1839 do {
1840 Node* n = stack.node();
1841 uint i = stack.index();
1842
1843 if (i < n->outcnt()) {
1844 Node* u = n->raw_out(i);
1845 stack.set_index(i+1);
1846 if (!visited.test_set(u->_idx)) {
1847 stack.push(u, 0);
1848 }
1849 } else {
1850 stack.pop();
1851 if (n->is_If() && is_heap_stable_test(n)) {
1852 heap_stable_tests.push(n);
1853 }
1854 }
1855 } while (stack.size() > 0);
1856
1857 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1858 Node* n = heap_stable_tests.at(i);
1859 assert(is_heap_stable_test(n), "only evacuation test");
1860 merge_back_to_back_tests(n, phase);
1861 }
1862
1863 if (!phase->C->major_progress()) {
1864 VectorSet seen;
1865 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1866 Node* n = heap_stable_tests.at(i);
1867 IdealLoopTree* loop = phase->get_loop(n);
1868 if (loop != phase->ltree_root() &&
1869 loop->_child == nullptr &&
1870 !loop->_irreducible) {
1871 Node* head = loop->_head;
1872 if (head->is_Loop() &&
1873 (!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) &&
1874 !seen.test_set(head->_idx)) {
1875 IfNode* iff = find_unswitching_candidate(loop, phase);
1876 if (iff != nullptr) {
1877 Node* bol = iff->in(1);
1878 if (head->as_Loop()->is_strip_mined()) {
1879 head->as_Loop()->verify_strip_mined(0);
1880 }
1881 move_gc_state_test_out_of_loop(iff, phase);
1882
1883 AutoNodeBudget node_budget(phase);
1884
1885 if (loop->policy_unswitching(phase)) {
1886 if (head->as_Loop()->is_strip_mined()) {
1887 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop();
1888 hide_strip_mined_loop(outer, head->as_CountedLoop(), phase);
1889 }
1890 phase->do_unswitching(loop, old_new);
1891 } else {
1892 // Not proceeding with unswitching. Move load back in
1893 // the loop.
1894 phase->igvn().replace_input_of(iff, 1, bol);
1895 }
1896 }
1897 }
1898 }
1899 }
1900 }
1901 }
1902
1903 ShenandoahIUBarrierNode::ShenandoahIUBarrierNode(Node* val) : Node(nullptr, val) {
1904 ShenandoahBarrierSetC2::bsc2()->state()->add_iu_barrier(this);
1905 }
1906
1907 const Type* ShenandoahIUBarrierNode::bottom_type() const {
1908 if (in(1) == nullptr || in(1)->is_top()) {
1909 return Type::TOP;
1910 }
1911 const Type* t = in(1)->bottom_type();
1912 if (t == TypePtr::NULL_PTR) {
1913 return t;
1914 }
1915 return t->is_oopptr();
1916 }
1917
1918 const Type* ShenandoahIUBarrierNode::Value(PhaseGVN* phase) const {
1919 if (in(1) == nullptr) {
1920 return Type::TOP;
1921 }
1922 const Type* t = phase->type(in(1));
1923 if (t == Type::TOP) {
1924 return Type::TOP;
1925 }
1926 if (t == TypePtr::NULL_PTR) {
1927 return t;
1928 }
1929 return t->is_oopptr();
1930 }
1931
1932 int ShenandoahIUBarrierNode::needed(Node* n) {
1933 if (n == nullptr ||
1934 n->is_Allocate() ||
1935 n->Opcode() == Op_ShenandoahIUBarrier ||
1936 n->bottom_type() == TypePtr::NULL_PTR ||
1937 (n->bottom_type()->make_oopptr() != nullptr && n->bottom_type()->make_oopptr()->const_oop() != nullptr)) {
1938 return NotNeeded;
1939 }
1940 if (n->is_Phi() ||
1941 n->is_CMove()) {
1942 return MaybeNeeded;
1943 }
1944 return Needed;
1945 }
1946
1947 Node* ShenandoahIUBarrierNode::next(Node* n) {
1948 for (;;) {
1949 if (n == nullptr) {
1950 return n;
1951 } else if (n->bottom_type() == TypePtr::NULL_PTR) {
1952 return n;
1953 } else if (n->bottom_type()->make_oopptr() != nullptr && n->bottom_type()->make_oopptr()->const_oop() != nullptr) {
1954 return n;
1955 } else if (n->is_ConstraintCast() ||
1956 n->Opcode() == Op_DecodeN ||
1957 n->Opcode() == Op_EncodeP) {
1958 n = n->in(1);
1959 } else if (n->is_Proj()) {
1960 n = n->in(0);
1961 } else {
1962 return n;
1963 }
1964 }
1965 ShouldNotReachHere();
1966 return nullptr;
1967 }
1968
1969 Node* ShenandoahIUBarrierNode::Identity(PhaseGVN* phase) {
1970 PhaseIterGVN* igvn = phase->is_IterGVN();
1971
1972 Node* n = next(in(1));
1973
1974 int cont = needed(n);
1975
1976 if (cont == NotNeeded) {
1977 return in(1);
1978 } else if (cont == MaybeNeeded) {
1979 if (igvn == nullptr) {
1980 phase->record_for_igvn(this);
1981 return this;
1982 } else {
1983 ResourceMark rm;
1984 Unique_Node_List wq;
1985 uint wq_i = 0;
1986
1987 for (;;) {
1988 if (n->is_Phi()) {
1989 for (uint i = 1; i < n->req(); i++) {
1990 Node* m = n->in(i);
1991 if (m != nullptr) {
1992 wq.push(m);
1993 }
1994 }
1995 } else {
1996 assert(n->is_CMove(), "nothing else here");
1997 Node* m = n->in(CMoveNode::IfFalse);
1998 wq.push(m);
1999 m = n->in(CMoveNode::IfTrue);
2000 wq.push(m);
2001 }
2002 Node* orig_n = nullptr;
2003 do {
2004 if (wq_i >= wq.size()) {
2005 return in(1);
2006 }
2007 n = wq.at(wq_i);
2008 wq_i++;
2009 orig_n = n;
2010 n = next(n);
2011 cont = needed(n);
2012 if (cont == Needed) {
2013 return this;
2014 }
2015 } while (cont != MaybeNeeded || (orig_n != n && wq.member(n)));
2016 }
2017 }
2018 }
2019
2020 return this;
2021 }
2022
2023 #ifdef ASSERT
2024 static bool has_never_branch(Node* root) {
2025 for (uint i = 1; i < root->req(); i++) {
2026 Node* in = root->in(i);
2027 if (in != nullptr && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->is_NeverBranch()) {
2028 return true;
2029 }
2030 }
2031 return false;
2032 }
2033 #endif
2034
2035 void MemoryGraphFixer::collect_memory_nodes() {
2036 Node_Stack stack(0);
2037 VectorSet visited;
2038 Node_List regions;
2039
2040 // Walk the raw memory graph and create a mapping from CFG node to
2041 // memory node. Exclude phis for now.
2042 stack.push(_phase->C->root(), 1);
2043 do {
2044 Node* n = stack.node();
2045 int opc = n->Opcode();
2046 uint i = stack.index();
2047 if (i < n->req()) {
2048 Node* mem = nullptr;
2049 if (opc == Op_Root) {
2050 Node* in = n->in(i);
2051 int in_opc = in->Opcode();
2052 if (in_opc == Op_Return || in_opc == Op_Rethrow) {
2053 mem = in->in(TypeFunc::Memory);
2054 } else if (in_opc == Op_Halt) {
2055 if (in->in(0)->is_Region()) {
2056 Node* r = in->in(0);
2057 for (uint j = 1; j < r->req(); j++) {
2058 assert(!r->in(j)->is_NeverBranch(), "");
2059 }
2060 } else {
2061 Node* proj = in->in(0);
2062 assert(proj->is_Proj(), "");
2063 Node* in = proj->in(0);
2064 assert(in->is_CallStaticJava() || in->is_NeverBranch() || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
2065 if (in->is_CallStaticJava()) {
2066 mem = in->in(TypeFunc::Memory);
2067 } else if (in->Opcode() == Op_Catch) {
2068 Node* call = in->in(0)->in(0);
2069 assert(call->is_Call(), "");
2070 mem = call->in(TypeFunc::Memory);
2071 } else if (in->is_NeverBranch()) {
2072 mem = collect_memory_for_infinite_loop(in);
2073 }
2074 }
2075 } else {
2076 #ifdef ASSERT
2077 n->dump();
2078 in->dump();
2079 #endif
2080 ShouldNotReachHere();
2081 }
2082 } else {
2083 assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
2084 assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
2085 mem = n->in(i);
2086 }
2087 i++;
2088 stack.set_index(i);
2089 if (mem == nullptr) {
2090 continue;
2091 }
2092 for (;;) {
2093 if (visited.test_set(mem->_idx) || mem->is_Start()) {
2094 break;
2095 }
2096 if (mem->is_Phi()) {
2097 stack.push(mem, 2);
2098 mem = mem->in(1);
2099 } else if (mem->is_Proj()) {
2100 stack.push(mem, mem->req());
2101 mem = mem->in(0);
2102 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
2103 mem = mem->in(TypeFunc::Memory);
2104 } else if (mem->is_MergeMem()) {
2105 MergeMemNode* mm = mem->as_MergeMem();
2106 mem = mm->memory_at(_alias);
2107 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
2108 assert(_alias == Compile::AliasIdxRaw, "");
2109 stack.push(mem, mem->req());
2110 mem = mem->in(MemNode::Memory);
2111 } else {
2112 #ifdef ASSERT
2113 mem->dump();
2114 #endif
2115 ShouldNotReachHere();
2116 }
2117 }
2118 } else {
2119 if (n->is_Phi()) {
2120 // Nothing
2121 } else if (!n->is_Root()) {
2122 Node* c = get_ctrl(n);
2123 _memory_nodes.map(c->_idx, n);
2124 }
2125 stack.pop();
2126 }
2127 } while(stack.is_nonempty());
2128
2129 // Iterate over CFG nodes in rpo and propagate memory state to
2130 // compute memory state at regions, creating new phis if needed.
2131 Node_List rpo_list;
2132 visited.clear();
2133 _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
2134 Node* root = rpo_list.pop();
2135 assert(root == _phase->C->root(), "");
2136
2137 const bool trace = false;
2138 #ifdef ASSERT
2139 if (trace) {
2140 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2141 Node* c = rpo_list.at(i);
2142 if (_memory_nodes[c->_idx] != nullptr) {
2143 tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump();
2144 }
2145 }
2146 }
2147 #endif
2148 uint last = _phase->C->unique();
2149
2150 #ifdef ASSERT
2151 uint16_t max_depth = 0;
2152 for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
2153 IdealLoopTree* lpt = iter.current();
2154 max_depth = MAX2(max_depth, lpt->_nest);
2155 }
2156 #endif
2157
2158 bool progress = true;
2159 int iteration = 0;
2160 Node_List dead_phis;
2161 while (progress) {
2162 progress = false;
2163 iteration++;
2164 assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2165 if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
2166
2167 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2168 Node* c = rpo_list.at(i);
2169
2170 Node* prev_mem = _memory_nodes[c->_idx];
2171 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2172 Node* prev_region = regions[c->_idx];
2173 Node* unique = nullptr;
2174 for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
2175 Node* m = _memory_nodes[c->in(j)->_idx];
2176 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");
2177 if (m != nullptr) {
2178 if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
2179 assert(c->is_Loop() && j == LoopNode::LoopBackControl || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2180 // continue
2181 } else if (unique == nullptr) {
2182 unique = m;
2183 } else if (m == unique) {
2184 // continue
2185 } else {
2186 unique = NodeSentinel;
2187 }
2188 }
2189 }
2190 assert(unique != nullptr, "empty phi???");
2191 if (unique != NodeSentinel) {
2192 if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c) {
2193 dead_phis.push(prev_region);
2194 }
2195 regions.map(c->_idx, unique);
2196 } else {
2197 Node* phi = nullptr;
2198 if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
2199 phi = prev_region;
2200 for (uint k = 1; k < c->req(); k++) {
2201 Node* m = _memory_nodes[c->in(k)->_idx];
2202 assert(m != nullptr, "expect memory state");
2203 phi->set_req(k, m);
2204 }
2205 } else {
2206 for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == nullptr; j++) {
2207 Node* u = c->fast_out(j);
2208 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2209 (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
2210 phi = u;
2211 for (uint k = 1; k < c->req() && phi != nullptr; k++) {
2212 Node* m = _memory_nodes[c->in(k)->_idx];
2213 assert(m != nullptr, "expect memory state");
2214 if (u->in(k) != m) {
2215 phi = NodeSentinel;
2216 }
2217 }
2218 }
2219 }
2220 if (phi == NodeSentinel) {
2221 phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
2222 for (uint k = 1; k < c->req(); k++) {
2223 Node* m = _memory_nodes[c->in(k)->_idx];
2224 assert(m != nullptr, "expect memory state");
2225 phi->init_req(k, m);
2226 }
2227 }
2228 }
2229 if (phi != nullptr) {
2230 regions.map(c->_idx, phi);
2231 } else {
2232 assert(c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2233 }
2234 }
2235 Node* current_region = regions[c->_idx];
2236 if (current_region != prev_region) {
2237 progress = true;
2238 if (prev_region == prev_mem) {
2239 _memory_nodes.map(c->_idx, current_region);
2240 }
2241 }
2242 } else if (prev_mem == nullptr || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
2243 Node* m = _memory_nodes[_phase->idom(c)->_idx];
2244 assert(m != nullptr || c->Opcode() == Op_Halt, "expect memory state");
2245 if (m != prev_mem) {
2246 _memory_nodes.map(c->_idx, m);
2247 progress = true;
2248 }
2249 }
2250 #ifdef ASSERT
2251 if (trace) { tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); }
2252 #endif
2253 }
2254 }
2255
2256 // Replace existing phi with computed memory state for that region
2257 // if different (could be a new phi or a dominating memory node if
2258 // that phi was found to be useless).
2259 while (dead_phis.size() > 0) {
2260 Node* n = dead_phis.pop();
2261 n->replace_by(_phase->C->top());
2262 n->destruct(&_phase->igvn());
2263 }
2264 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2265 Node* c = rpo_list.at(i);
2266 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2267 Node* n = regions[c->_idx];
2268 assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2269 if (n != nullptr && n->is_Phi() && n->_idx >= last && n->in(0) == c) {
2270 _phase->register_new_node(n, c);
2271 }
2272 }
2273 }
2274 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2275 Node* c = rpo_list.at(i);
2276 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2277 Node* n = regions[c->_idx];
2278 assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2279 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
2280 Node* u = c->fast_out(i);
2281 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2282 u != n) {
2283 assert(c->unique_ctrl_out()->Opcode() != Op_Halt, "expected memory state");
2284 if (u->adr_type() == TypePtr::BOTTOM) {
2285 fix_memory_uses(u, n, n, c);
2286 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2287 _phase->lazy_replace(u, n);
2288 --i; --imax;
2289 }
2290 }
2291 }
2292 }
2293 }
2294 }
2295
2296 Node* MemoryGraphFixer::collect_memory_for_infinite_loop(const Node* in) {
2297 Node* mem = nullptr;
2298 Node* head = in->in(0);
2299 assert(head->is_Region(), "unexpected infinite loop graph shape");
2300
2301 Node* phi_mem = nullptr;
2302 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2303 Node* u = head->fast_out(j);
2304 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
2305 if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2306 assert(phi_mem == nullptr || phi_mem->adr_type() == TypePtr::BOTTOM, "");
2307 phi_mem = u;
2308 } else if (u->adr_type() == TypePtr::BOTTOM) {
2309 assert(phi_mem == nullptr || _phase->C->get_alias_index(phi_mem->adr_type()) == _alias, "");
2310 if (phi_mem == nullptr) {
2311 phi_mem = u;
2312 }
2313 }
2314 }
2315 }
2316 if (phi_mem == nullptr) {
2317 ResourceMark rm;
2318 Node_Stack stack(0);
2319 stack.push(head, 1);
2320 do {
2321 Node* n = stack.node();
2322 uint i = stack.index();
2323 if (i >= n->req()) {
2324 stack.pop();
2325 } else {
2326 stack.set_index(i + 1);
2327 Node* c = n->in(i);
2328 assert(c != head, "should have found a safepoint on the way");
2329 if (stack.size() != 1 || _phase->is_dominator(head, c)) {
2330 for (;;) {
2331 if (c->is_Region()) {
2332 stack.push(c, 1);
2333 break;
2334 } else if (c->is_SafePoint() && !c->is_CallLeaf()) {
2335 Node* m = c->in(TypeFunc::Memory);
2336 if (m->is_MergeMem()) {
2337 m = m->as_MergeMem()->memory_at(_alias);
2338 }
2339 assert(mem == nullptr || mem == m, "several memory states");
2340 mem = m;
2341 break;
2342 } else {
2343 assert(c != c->in(0), "");
2344 c = c->in(0);
2345 }
2346 }
2347 }
2348 }
2349 } while (stack.size() > 0);
2350 assert(mem != nullptr, "should have found safepoint");
2351 } else {
2352 mem = phi_mem;
2353 }
2354 return mem;
2355 }
2356
2357 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
2358 Node* c = _phase->get_ctrl(n);
2359 if (n->is_Proj() && n->in(0) != nullptr && n->in(0)->is_Call()) {
2360 assert(c == n->in(0), "");
2361 CallNode* call = c->as_Call();
2362 CallProjections projs;
2363 call->extract_projections(&projs, true, false);
2364 if (projs.catchall_memproj != nullptr) {
2365 if (projs.fallthrough_memproj == n) {
2366 c = projs.fallthrough_catchproj;
2367 } else {
2368 assert(projs.catchall_memproj == n, "");
2369 c = projs.catchall_catchproj;
2370 }
2371 }
2372 }
2373 return c;
2374 }
2375
2376 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
2377 if (_phase->has_ctrl(n))
2378 return get_ctrl(n);
2379 else {
2380 assert (n->is_CFG(), "must be a CFG node");
2381 return n;
2382 }
2383 }
2384
2385 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
2386 return m != nullptr && get_ctrl(m) == c;
2387 }
2388
2389 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
2390 assert(n == nullptr || _phase->ctrl_or_self(n) == ctrl, "");
2391 assert(!ctrl->is_Call() || ctrl == n, "projection expected");
2392 #ifdef ASSERT
2393 if ((ctrl->is_Proj() && ctrl->in(0)->is_Call()) ||
2394 (ctrl->is_Catch() && ctrl->in(0)->in(0)->is_Call())) {
2395 CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_Call() : ctrl->in(0)->in(0)->as_Call();
2396 int mems = 0;
2397 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
2398 Node* u = call->fast_out(i);
2399 if (u->bottom_type() == Type::MEMORY) {
2400 mems++;
2401 }
2402 }
2403 assert(mems <= 1, "No node right after call if multiple mem projections");
2404 }
2405 #endif
2406 Node* mem = _memory_nodes[ctrl->_idx];
2407 Node* c = ctrl;
2408 while (!mem_is_valid(mem, c) &&
2409 (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
2410 c = _phase->idom(c);
2411 mem = _memory_nodes[c->_idx];
2412 }
2413 if (n != nullptr && mem_is_valid(mem, c)) {
2414 while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
2415 mem = next_mem(mem, _alias);
2416 }
2417 if (mem->is_MergeMem()) {
2418 mem = mem->as_MergeMem()->memory_at(_alias);
2419 }
2420 if (!mem_is_valid(mem, c)) {
2421 do {
2422 c = _phase->idom(c);
2423 mem = _memory_nodes[c->_idx];
2424 } while (!mem_is_valid(mem, c) &&
2425 (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
2426 }
2427 }
2428 assert(mem->bottom_type() == Type::MEMORY, "");
2429 return mem;
2430 }
2431
2432 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
2433 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2434 Node* use = region->fast_out(i);
2435 if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
2436 (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
2437 return true;
2438 }
2439 }
2440 return false;
2441 }
2442
2443 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
2444 assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
2445 const bool trace = false;
2446 DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
2447 DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
2448 GrowableArray<Node*> phis;
2449 if (mem_for_ctrl != mem) {
2450 Node* old = mem_for_ctrl;
2451 Node* prev = nullptr;
2452 while (old != mem) {
2453 prev = old;
2454 if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
2455 assert(_alias == Compile::AliasIdxRaw, "");
2456 old = old->in(MemNode::Memory);
2457 } else if (old->Opcode() == Op_SCMemProj) {
2458 assert(_alias == Compile::AliasIdxRaw, "");
2459 old = old->in(0);
2460 } else {
2461 ShouldNotReachHere();
2462 }
2463 }
2464 assert(prev != nullptr, "");
2465 if (new_ctrl != ctrl) {
2466 _memory_nodes.map(ctrl->_idx, mem);
2467 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2468 }
2469 uint input = (uint)MemNode::Memory;
2470 _phase->igvn().replace_input_of(prev, input, new_mem);
2471 } else {
2472 uses.clear();
2473 _memory_nodes.map(new_ctrl->_idx, new_mem);
2474 uses.push(new_ctrl);
2475 for(uint next = 0; next < uses.size(); next++ ) {
2476 Node *n = uses.at(next);
2477 assert(n->is_CFG(), "");
2478 DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
2479 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2480 Node* u = n->fast_out(i);
2481 if (!u->is_Root() && u->is_CFG() && u != n) {
2482 Node* m = _memory_nodes[u->_idx];
2483 if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
2484 !has_mem_phi(u) &&
2485 u->unique_ctrl_out()->Opcode() != Op_Halt) {
2486 DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
2487 DEBUG_ONLY(if (trace && m != nullptr) { tty->print("ZZZ mem"); m->dump(); });
2488
2489 if (!mem_is_valid(m, u) || !m->is_Phi()) {
2490 bool push = true;
2491 bool create_phi = true;
2492 if (_phase->is_dominator(new_ctrl, u)) {
2493 create_phi = false;
2494 }
2495 if (create_phi) {
2496 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
2497 _phase->register_new_node(phi, u);
2498 phis.push(phi);
2499 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
2500 if (!mem_is_valid(m, u)) {
2501 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
2502 _memory_nodes.map(u->_idx, phi);
2503 } else {
2504 DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
2505 for (;;) {
2506 assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), "");
2507 Node* next = nullptr;
2508 if (m->is_Proj()) {
2509 next = m->in(0);
2510 } else {
2511 assert(m->is_Mem() || m->is_LoadStore(), "");
2512 assert(_alias == Compile::AliasIdxRaw, "");
2513 next = m->in(MemNode::Memory);
2514 }
2515 if (_phase->get_ctrl(next) != u) {
2516 break;
2517 }
2518 if (next->is_MergeMem()) {
2519 assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
2520 break;
2521 }
2522 if (next->is_Phi()) {
2523 assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
2524 break;
2525 }
2526 m = next;
2527 }
2528
2529 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
2530 assert(m->is_Mem() || m->is_LoadStore(), "");
2531 uint input = (uint)MemNode::Memory;
2532 _phase->igvn().replace_input_of(m, input, phi);
2533 push = false;
2534 }
2535 } else {
2536 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
2537 }
2538 if (push) {
2539 uses.push(u);
2540 }
2541 }
2542 } else if (!mem_is_valid(m, u) &&
2543 !(u->Opcode() == Op_CProj && u->in(0)->is_NeverBranch() && u->as_Proj()->_con == 1)) {
2544 uses.push(u);
2545 }
2546 }
2547 }
2548 }
2549 for (int i = 0; i < phis.length(); i++) {
2550 Node* n = phis.at(i);
2551 Node* r = n->in(0);
2552 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
2553 for (uint j = 1; j < n->req(); j++) {
2554 Node* m = find_mem(r->in(j), nullptr);
2555 _phase->igvn().replace_input_of(n, j, m);
2556 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
2557 }
2558 }
2559 }
2560 uint last = _phase->C->unique();
2561 MergeMemNode* mm = nullptr;
2562 int alias = _alias;
2563 DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
2564 // Process loads first to not miss an anti-dependency: if the memory
2565 // edge of a store is updated before a load is processed then an
2566 // anti-dependency may be missed.
2567 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2568 Node* u = mem->out(i);
2569 if (u->_idx < last && u->is_Load() && _phase->C->get_alias_index(u->adr_type()) == alias) {
2570 Node* m = find_mem(_phase->get_ctrl(u), u);
2571 if (m != mem) {
2572 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2573 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2574 --i;
2575 }
2576 }
2577 }
2578 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2579 Node* u = mem->out(i);
2580 if (u->_idx < last) {
2581 if (u->is_Mem()) {
2582 if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2583 Node* m = find_mem(_phase->get_ctrl(u), u);
2584 if (m != mem) {
2585 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2586 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2587 --i;
2588 }
2589 }
2590 } else if (u->is_MergeMem()) {
2591 MergeMemNode* u_mm = u->as_MergeMem();
2592 if (u_mm->memory_at(alias) == mem) {
2593 MergeMemNode* newmm = nullptr;
2594 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2595 Node* uu = u->fast_out(j);
2596 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2597 if (uu->is_Phi()) {
2598 assert(uu->adr_type() == TypePtr::BOTTOM, "");
2599 Node* region = uu->in(0);
2600 int nb = 0;
2601 for (uint k = 1; k < uu->req(); k++) {
2602 if (uu->in(k) == u) {
2603 Node* m = find_mem(region->in(k), nullptr);
2604 if (m != mem) {
2605 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
2606 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2607 if (newmm != u) {
2608 _phase->igvn().replace_input_of(uu, k, newmm);
2609 nb++;
2610 --jmax;
2611 }
2612 }
2613 }
2614 }
2615 if (nb > 0) {
2616 --j;
2617 }
2618 } else {
2619 Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
2620 if (m != mem) {
2621 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
2622 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2623 if (newmm != u) {
2624 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2625 --j, --jmax;
2626 }
2627 }
2628 }
2629 }
2630 }
2631 } else if (u->is_Phi()) {
2632 assert(u->bottom_type() == Type::MEMORY, "what else?");
2633 if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
2634 Node* region = u->in(0);
2635 bool replaced = false;
2636 for (uint j = 1; j < u->req(); j++) {
2637 if (u->in(j) == mem) {
2638 Node* m = find_mem(region->in(j), nullptr);
2639 Node* nnew = m;
2640 if (m != mem) {
2641 if (u->adr_type() == TypePtr::BOTTOM) {
2642 mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
2643 nnew = mm;
2644 }
2645 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
2646 _phase->igvn().replace_input_of(u, j, nnew);
2647 replaced = true;
2648 }
2649 }
2650 }
2651 if (replaced) {
2652 --i;
2653 }
2654 }
2655 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2656 u->adr_type() == nullptr) {
2657 assert(u->adr_type() != nullptr ||
2658 u->Opcode() == Op_Rethrow ||
2659 u->Opcode() == Op_Return ||
2660 u->Opcode() == Op_SafePoint ||
2661 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2662 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2663 u->Opcode() == Op_CallLeaf, "");
2664 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2665 if (m != mem) {
2666 mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
2667 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2668 --i;
2669 }
2670 } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2671 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2672 if (m != mem) {
2673 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2674 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2675 --i;
2676 }
2677 } else if (u->adr_type() != TypePtr::BOTTOM &&
2678 _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
2679 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2680 assert(m != mem, "");
2681 // u is on the wrong slice...
2682 assert(u->is_ClearArray(), "");
2683 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2684 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2685 --i;
2686 }
2687 }
2688 }
2689 #ifdef ASSERT
2690 assert(new_mem->outcnt() > 0, "");
2691 for (int i = 0; i < phis.length(); i++) {
2692 Node* n = phis.at(i);
2693 assert(n->outcnt() > 0, "new phi must have uses now");
2694 }
2695 #endif
2696 }
2697
2698 void MemoryGraphFixer::record_new_ctrl(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl) {
2699 if (mem_for_ctrl != mem && new_ctrl != ctrl) {
2700 _memory_nodes.map(ctrl->_idx, mem);
2701 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2702 }
2703 }
2704
2705 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
2706 MergeMemNode* mm = MergeMemNode::make(mem);
2707 mm->set_memory_at(_alias, rep_proj);
2708 _phase->register_new_node(mm, rep_ctrl);
2709 return mm;
2710 }
2711
2712 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
2713 MergeMemNode* newmm = nullptr;
2714 MergeMemNode* u_mm = u->as_MergeMem();
2715 Node* c = _phase->get_ctrl(u);
2716 if (_phase->is_dominator(c, rep_ctrl)) {
2717 c = rep_ctrl;
2718 } else {
2719 assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
2720 }
2721 if (u->outcnt() == 1) {
2722 if (u->req() > (uint)_alias && u->in(_alias) == mem) {
2723 _phase->igvn().replace_input_of(u, _alias, rep_proj);
2724 --i;
2725 } else {
2726 _phase->igvn().rehash_node_delayed(u);
2727 u_mm->set_memory_at(_alias, rep_proj);
2728 }
2729 newmm = u_mm;
2730 _phase->set_ctrl_and_loop(u, c);
2731 } else {
2732 // can't simply clone u and then change one of its input because
2733 // it adds and then removes an edge which messes with the
2734 // DUIterator
2735 newmm = MergeMemNode::make(u_mm->base_memory());
2736 for (uint j = 0; j < u->req(); j++) {
2737 if (j < newmm->req()) {
2738 if (j == (uint)_alias) {
2739 newmm->set_req(j, rep_proj);
2740 } else if (newmm->in(j) != u->in(j)) {
2741 newmm->set_req(j, u->in(j));
2742 }
2743 } else if (j == (uint)_alias) {
2744 newmm->add_req(rep_proj);
2745 } else {
2746 newmm->add_req(u->in(j));
2747 }
2748 }
2749 if ((uint)_alias >= u->req()) {
2750 newmm->set_memory_at(_alias, rep_proj);
2751 }
2752 _phase->register_new_node(newmm, c);
2753 }
2754 return newmm;
2755 }
2756
2757 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
2758 if (phi->adr_type() == TypePtr::BOTTOM) {
2759 Node* region = phi->in(0);
2760 for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
2761 Node* uu = region->fast_out(j);
2762 if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
2763 return false;
2764 }
2765 }
2766 return true;
2767 }
2768 return _phase->C->get_alias_index(phi->adr_type()) == _alias;
2769 }
2770
2771 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
2772 uint last = _phase-> C->unique();
2773 MergeMemNode* mm = nullptr;
2774 assert(mem->bottom_type() == Type::MEMORY, "");
2775 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2776 Node* u = mem->out(i);
2777 if (u != replacement && u->_idx < last) {
2778 if (u->is_MergeMem()) {
2779 MergeMemNode* u_mm = u->as_MergeMem();
2780 if (u_mm->memory_at(_alias) == mem) {
2781 MergeMemNode* newmm = nullptr;
2782 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2783 Node* uu = u->fast_out(j);
2784 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2785 if (uu->is_Phi()) {
2786 if (should_process_phi(uu)) {
2787 Node* region = uu->in(0);
2788 int nb = 0;
2789 for (uint k = 1; k < uu->req(); k++) {
2790 if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
2791 if (newmm == nullptr) {
2792 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2793 }
2794 if (newmm != u) {
2795 _phase->igvn().replace_input_of(uu, k, newmm);
2796 nb++;
2797 --jmax;
2798 }
2799 }
2800 }
2801 if (nb > 0) {
2802 --j;
2803 }
2804 }
2805 } else {
2806 if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
2807 if (newmm == nullptr) {
2808 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2809 }
2810 if (newmm != u) {
2811 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2812 --j, --jmax;
2813 }
2814 }
2815 }
2816 }
2817 }
2818 } else if (u->is_Phi()) {
2819 assert(u->bottom_type() == Type::MEMORY, "what else?");
2820 Node* region = u->in(0);
2821 if (should_process_phi(u)) {
2822 bool replaced = false;
2823 for (uint j = 1; j < u->req(); j++) {
2824 if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
2825 Node* nnew = rep_proj;
2826 if (u->adr_type() == TypePtr::BOTTOM) {
2827 if (mm == nullptr) {
2828 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2829 }
2830 nnew = mm;
2831 }
2832 _phase->igvn().replace_input_of(u, j, nnew);
2833 replaced = true;
2834 }
2835 }
2836 if (replaced) {
2837 --i;
2838 }
2839
2840 }
2841 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2842 u->adr_type() == nullptr) {
2843 assert(u->adr_type() != nullptr ||
2844 u->Opcode() == Op_Rethrow ||
2845 u->Opcode() == Op_Return ||
2846 u->Opcode() == Op_SafePoint ||
2847 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2848 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2849 u->Opcode() == Op_CallLeaf, "%s", u->Name());
2850 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2851 if (mm == nullptr) {
2852 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2853 }
2854 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2855 --i;
2856 }
2857 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2858 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2859 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
2860 --i;
2861 }
2862 }
2863 }
2864 }
2865 }
2866
2867 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj, DecoratorSet decorators)
2868 : Node(ctrl, obj), _decorators(decorators) {
2869 ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this);
2870 }
2871
2872 DecoratorSet ShenandoahLoadReferenceBarrierNode::decorators() const {
2873 return _decorators;
2874 }
2875
2876 uint ShenandoahLoadReferenceBarrierNode::size_of() const {
2877 return sizeof(*this);
2878 }
2879
2880 static DecoratorSet mask_decorators(DecoratorSet decorators) {
2881 return decorators & (ON_STRONG_OOP_REF | ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF | ON_UNKNOWN_OOP_REF | IN_NATIVE);
2882 }
2883
2884 uint ShenandoahLoadReferenceBarrierNode::hash() const {
2885 uint hash = Node::hash();
2886 hash += mask_decorators(_decorators);
2887 return hash;
2888 }
2889
2890 bool ShenandoahLoadReferenceBarrierNode::cmp( const Node &n ) const {
2891 return Node::cmp(n) && n.Opcode() == Op_ShenandoahLoadReferenceBarrier &&
2892 mask_decorators(_decorators) == mask_decorators(((const ShenandoahLoadReferenceBarrierNode&)n)._decorators);
2893 }
2894
2895 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const {
2896 if (in(ValueIn) == nullptr || in(ValueIn)->is_top()) {
2897 return Type::TOP;
2898 }
2899 const Type* t = in(ValueIn)->bottom_type();
2900 if (t == TypePtr::NULL_PTR) {
2901 return t;
2902 }
2903
2904 if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2905 return t;
2906 }
2907
2908 return t->meet(TypePtr::NULL_PTR);
2909 }
2910
2911 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const {
2912 // Either input is TOP ==> the result is TOP
2913 const Type *t2 = phase->type(in(ValueIn));
2914 if( t2 == Type::TOP ) return Type::TOP;
2915
2916 if (t2 == TypePtr::NULL_PTR) {
2917 return t2;
2918 }
2919
2920 if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2921 return t2;
2922 }
2923
2924 return t2->meet(TypePtr::NULL_PTR);
2925 }
2926
2927 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) {
2928 Node* value = in(ValueIn);
2929 if (!needs_barrier(phase, value)) {
2930 return value;
2931 }
2932 return this;
2933 }
2934
2935 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) {
2936 Unique_Node_List visited;
2937 return needs_barrier_impl(phase, n, visited);
2938 }
2939
2940 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) {
2941 if (n == nullptr) return false;
2942 if (visited.member(n)) {
2943 return false; // Been there.
2944 }
2945 visited.push(n);
2946
2947 if (n->is_Allocate()) {
2948 // tty->print_cr("optimize barrier on alloc");
2949 return false;
2950 }
2951 if (n->is_Call()) {
2952 // tty->print_cr("optimize barrier on call");
2953 return false;
2954 }
2955
2956 const Type* type = phase->type(n);
2957 if (type == Type::TOP) {
2958 return false;
2959 }
2960 if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) {
2961 // tty->print_cr("optimize barrier on null");
2962 return false;
2963 }
2964 if (type->make_oopptr() && type->make_oopptr()->const_oop() != nullptr) {
2965 // tty->print_cr("optimize barrier on constant");
2966 return false;
2967 }
2968
2969 switch (n->Opcode()) {
2970 case Op_AddP:
2971 return true; // TODO: Can refine?
2972 case Op_LoadP:
2973 case Op_ShenandoahCompareAndExchangeN:
2974 case Op_ShenandoahCompareAndExchangeP:
2975 case Op_CompareAndExchangeN:
2976 case Op_CompareAndExchangeP:
2977 case Op_GetAndSetN:
2978 case Op_GetAndSetP:
2979 return true;
2980 case Op_Phi: {
2981 for (uint i = 1; i < n->req(); i++) {
2982 if (needs_barrier_impl(phase, n->in(i), visited)) return true;
2983 }
2984 return false;
2985 }
2986 case Op_CheckCastPP:
2987 case Op_CastPP:
2988 return needs_barrier_impl(phase, n->in(1), visited);
2989 case Op_Proj:
2990 return needs_barrier_impl(phase, n->in(0), visited);
2991 case Op_ShenandoahLoadReferenceBarrier:
2992 // tty->print_cr("optimize barrier on barrier");
2993 return false;
2994 case Op_Parm:
2995 // tty->print_cr("optimize barrier on input arg");
2996 return false;
2997 case Op_DecodeN:
2998 case Op_EncodeP:
2999 return needs_barrier_impl(phase, n->in(1), visited);
3000 case Op_LoadN:
3001 return true;
3002 case Op_CMoveN:
3003 case Op_CMoveP:
3004 return needs_barrier_impl(phase, n->in(2), visited) ||
3005 needs_barrier_impl(phase, n->in(3), visited);
3006 case Op_ShenandoahIUBarrier:
3007 return needs_barrier_impl(phase, n->in(1), visited);
3008 case Op_CreateEx:
3009 return false;
3010 default:
3011 break;
3012 }
3013 #ifdef ASSERT
3014 tty->print("need barrier on?: ");
3015 tty->print_cr("ins:");
3016 n->dump(2);
3017 tty->print_cr("outs:");
3018 n->dump(-2);
3019 ShouldNotReachHere();
3020 #endif
3021 return true;
3022 }