1 /*
2 * Copyright (c) 2015, 2021, Red Hat, Inc. All rights reserved.
3 * Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "precompiled.hpp"
27
28 #include "classfile/javaClasses.hpp"
29 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
30 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
31 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
32 #include "gc/shenandoah/shenandoahForwarding.hpp"
33 #include "gc/shenandoah/shenandoahHeap.hpp"
34 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
35 #include "gc/shenandoah/shenandoahRuntime.hpp"
36 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
37 #include "opto/arraycopynode.hpp"
38 #include "opto/block.hpp"
39 #include "opto/callnode.hpp"
40 #include "opto/castnode.hpp"
41 #include "opto/movenode.hpp"
42 #include "opto/phaseX.hpp"
43 #include "opto/rootnode.hpp"
44 #include "opto/runtime.hpp"
45 #include "opto/subnode.hpp"
46
47 bool ShenandoahBarrierC2Support::expand(Compile* C, PhaseIterGVN& igvn) {
48 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
49 if ((state->iu_barriers_count() +
50 state->load_reference_barriers_count()) > 0) {
51 assert(C->post_loop_opts_phase(), "no loop opts allowed");
52 C->reset_post_loop_opts_phase(); // ... but we know what we are doing
53 C->clear_major_progress();
54 PhaseIdealLoop::optimize(igvn, LoopOptsShenandoahExpand);
55 if (C->failing()) return false;
56
57 C->set_major_progress();
58 if (!C->optimize_loops(igvn, LoopOptsShenandoahPostExpand)) {
59 return false;
60 }
61 C->clear_major_progress();
62 C->process_for_post_loop_opts_igvn(igvn);
63 if (C->failing()) return false;
64
65 C->set_post_loop_opts_phase(); // now for real!
66 }
67 return true;
68 }
69
70 bool ShenandoahBarrierC2Support::is_gc_state_test(Node* iff, int mask) {
71 if (!UseShenandoahGC) {
72 return false;
73 }
74 assert(iff->is_If(), "bad input");
75 if (iff->Opcode() != Op_If) {
76 return false;
77 }
78 Node* bol = iff->in(1);
79 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
80 return false;
81 }
82 Node* cmp = bol->in(1);
83 if (cmp->Opcode() != Op_CmpI) {
84 return false;
85 }
86 Node* in1 = cmp->in(1);
87 Node* in2 = cmp->in(2);
88 if (in2->find_int_con(-1) != 0) {
89 return false;
90 }
91 if (in1->Opcode() != Op_AndI) {
92 return false;
93 }
94 in2 = in1->in(2);
95 if (in2->find_int_con(-1) != mask) {
96 return false;
97 }
98 in1 = in1->in(1);
99
100 return is_gc_state_load(in1);
101 }
102
103 bool ShenandoahBarrierC2Support::is_heap_stable_test(Node* iff) {
104 return is_gc_state_test(iff, ShenandoahHeap::HAS_FORWARDED);
105 }
106
107 bool ShenandoahBarrierC2Support::is_gc_state_load(Node *n) {
108 if (!UseShenandoahGC) {
109 return false;
110 }
111 if (n->Opcode() != Op_LoadB && n->Opcode() != Op_LoadUB) {
112 return false;
113 }
114 Node* addp = n->in(MemNode::Address);
115 if (!addp->is_AddP()) {
116 return false;
117 }
118 Node* base = addp->in(AddPNode::Address);
119 Node* off = addp->in(AddPNode::Offset);
120 if (base->Opcode() != Op_ThreadLocal) {
121 return false;
122 }
123 if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) {
124 return false;
125 }
126 return true;
127 }
128
129 bool ShenandoahBarrierC2Support::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) {
130 assert(phase->is_dominator(stop, start), "bad inputs");
131 ResourceMark rm;
132 Unique_Node_List wq;
133 wq.push(start);
134 for (uint next = 0; next < wq.size(); next++) {
135 Node *m = wq.at(next);
136 if (m == stop) {
137 continue;
138 }
139 if (m->is_SafePoint() && !m->is_CallLeaf()) {
140 return true;
141 }
142 if (m->is_Region()) {
143 for (uint i = 1; i < m->req(); i++) {
144 wq.push(m->in(i));
145 }
146 } else {
147 wq.push(m->in(0));
148 }
149 }
150 return false;
151 }
152
153 #ifdef ASSERT
154 bool ShenandoahBarrierC2Support::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) {
155 assert(phis.size() == 0, "");
156
157 while (true) {
158 if (in->bottom_type() == TypePtr::NULL_PTR) {
159 if (trace) {tty->print_cr("null");}
160 } else if (!in->bottom_type()->make_ptr()->make_oopptr()) {
161 if (trace) {tty->print_cr("Non oop");}
162 } else {
163 if (in->is_ConstraintCast()) {
164 in = in->in(1);
165 continue;
166 } else if (in->is_AddP()) {
167 assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access");
168 in = in->in(AddPNode::Address);
169 continue;
170 } else if (in->is_Con()) {
171 if (trace) {
172 tty->print("Found constant");
173 in->dump();
174 }
175 } else if (in->Opcode() == Op_Parm) {
176 if (trace) {
177 tty->print("Found argument");
178 }
179 } else if (in->Opcode() == Op_CreateEx) {
180 if (trace) {
181 tty->print("Found create-exception");
182 }
183 } else if (in->Opcode() == Op_LoadP && in->adr_type() == TypeRawPtr::BOTTOM) {
184 if (trace) {
185 tty->print("Found raw LoadP (OSR argument?)");
186 }
187 } else if (in->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
188 if (t == ShenandoahOopStore) {
189 uint i = 0;
190 for (; i < phis.size(); i++) {
191 Node* n = phis.node_at(i);
192 if (n->Opcode() == Op_ShenandoahIUBarrier) {
193 break;
194 }
195 }
196 if (i == phis.size()) {
197 return false;
198 }
199 }
200 barriers_used.push(in);
201 if (trace) {tty->print("Found barrier"); in->dump();}
202 } else if (in->Opcode() == Op_ShenandoahIUBarrier) {
203 if (t != ShenandoahOopStore) {
204 in = in->in(1);
205 continue;
206 }
207 if (trace) {tty->print("Found enqueue barrier"); in->dump();}
208 phis.push(in, in->req());
209 in = in->in(1);
210 continue;
211 } else if (in->is_Proj() && in->in(0)->is_Allocate()) {
212 if (trace) {
213 tty->print("Found alloc");
214 in->in(0)->dump();
215 }
216 } else if (in->is_Proj() && (in->in(0)->Opcode() == Op_CallStaticJava || in->in(0)->Opcode() == Op_CallDynamicJava)) {
217 if (trace) {
218 tty->print("Found Java call");
219 }
220 } else if (in->is_Phi()) {
221 if (!visited.test_set(in->_idx)) {
222 if (trace) {tty->print("Pushed phi:"); in->dump();}
223 phis.push(in, 2);
224 in = in->in(1);
225 continue;
226 }
227 if (trace) {tty->print("Already seen phi:"); in->dump();}
228 } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) {
229 if (!visited.test_set(in->_idx)) {
230 if (trace) {tty->print("Pushed cmovep:"); in->dump();}
231 phis.push(in, CMoveNode::IfTrue);
232 in = in->in(CMoveNode::IfFalse);
233 continue;
234 }
235 if (trace) {tty->print("Already seen cmovep:"); in->dump();}
236 } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) {
237 in = in->in(1);
238 continue;
239 } else {
240 return false;
241 }
242 }
243 bool cont = false;
244 while (phis.is_nonempty()) {
245 uint idx = phis.index();
246 Node* phi = phis.node();
247 if (idx >= phi->req()) {
248 if (trace) {tty->print("Popped phi:"); phi->dump();}
249 phis.pop();
250 continue;
251 }
252 if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();}
253 in = phi->in(idx);
254 phis.set_index(idx+1);
255 cont = true;
256 break;
257 }
258 if (!cont) {
259 break;
260 }
261 }
262 return true;
263 }
264
265 void ShenandoahBarrierC2Support::report_verify_failure(const char* msg, Node* n1, Node* n2) {
266 if (n1 != nullptr) {
267 n1->dump(+10);
268 }
269 if (n2 != nullptr) {
270 n2->dump(+10);
271 }
272 fatal("%s", msg);
273 }
274
275 void ShenandoahBarrierC2Support::verify(RootNode* root) {
276 ResourceMark rm;
277 Unique_Node_List wq;
278 GrowableArray<Node*> barriers;
279 Unique_Node_List barriers_used;
280 Node_Stack phis(0);
281 VectorSet visited;
282 const bool trace = false;
283 const bool verify_no_useless_barrier = false;
284
285 wq.push(root);
286 for (uint next = 0; next < wq.size(); next++) {
287 Node *n = wq.at(next);
288 if (n->is_Load()) {
289 const bool trace = false;
290 if (trace) {tty->print("Verifying"); n->dump();}
291 if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) {
292 if (trace) {tty->print_cr("Load range/klass");}
293 } else {
294 const TypePtr* adr_type = n->as_Load()->adr_type();
295
296 if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) {
297 if (trace) {tty->print_cr("Mark load");}
298 } else if (adr_type->isa_instptr() &&
299 adr_type->is_instptr()->instance_klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
300 adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset()) {
301 if (trace) {tty->print_cr("Reference.get()");}
302 } else if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) {
303 report_verify_failure("Shenandoah verification: Load should have barriers", n);
304 }
305 }
306 } else if (n->is_Store()) {
307 const bool trace = false;
308
309 if (trace) {tty->print("Verifying"); n->dump();}
310 if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) {
311 Node* adr = n->in(MemNode::Address);
312 bool verify = true;
313
314 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) {
315 adr = adr->in(AddPNode::Address);
316 if (adr->is_AddP()) {
317 assert(adr->in(AddPNode::Base)->is_top(), "");
318 adr = adr->in(AddPNode::Address);
319 if (adr->Opcode() == Op_LoadP &&
320 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() &&
321 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
322 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) {
323 if (trace) {tty->print_cr("SATB prebarrier");}
324 verify = false;
325 }
326 }
327 }
328
329 if (verify && !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahIUBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
330 report_verify_failure("Shenandoah verification: Store should have barriers", n);
331 }
332 }
333 if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
334 report_verify_failure("Shenandoah verification: Store (address) should have barriers", n);
335 }
336 } else if (n->Opcode() == Op_CmpP) {
337 const bool trace = false;
338
339 Node* in1 = n->in(1);
340 Node* in2 = n->in(2);
341 if (in1->bottom_type()->isa_oopptr()) {
342 if (trace) {tty->print("Verifying"); n->dump();}
343
344 bool mark_inputs = false;
345 if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR ||
346 (in1->is_Con() || in2->is_Con())) {
347 if (trace) {tty->print_cr("Comparison against a constant");}
348 mark_inputs = true;
349 } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) ||
350 (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) {
351 if (trace) {tty->print_cr("Comparison with newly alloc'ed object");}
352 mark_inputs = true;
353 } else {
354 assert(in2->bottom_type()->isa_oopptr(), "");
355
356 if (!verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) ||
357 !verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) {
358 report_verify_failure("Shenandoah verification: Cmp should have barriers", n);
359 }
360 }
361 if (verify_no_useless_barrier &&
362 mark_inputs &&
363 (!verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) ||
364 !verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) {
365 phis.clear();
366 visited.reset();
367 }
368 }
369 } else if (n->is_LoadStore()) {
370 if (n->in(MemNode::ValueIn)->bottom_type()->make_ptr() &&
371 !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahIUBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
372 report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n);
373 }
374
375 if (n->in(MemNode::Address)->bottom_type()->make_oopptr() && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
376 report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n);
377 }
378 } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) {
379 CallNode* call = n->as_Call();
380
381 static struct {
382 const char* name;
383 struct {
384 int pos;
385 verify_type t;
386 } args[6];
387 } calls[] = {
388 "array_partition_stub",
389 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+4, ShenandoahStore }, { -1, ShenandoahNone },
390 { -1, ShenandoahNone }, { -1, ShenandoahNone }, { -1, ShenandoahNone } },
391 "arraysort_stub",
392 { { TypeFunc::Parms, ShenandoahStore }, { -1, ShenandoahNone }, { -1, ShenandoahNone },
393 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
394 "aescrypt_encryptBlock",
395 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
396 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
397 "aescrypt_decryptBlock",
398 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
399 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
400 "multiplyToLen",
401 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { TypeFunc::Parms+4, ShenandoahStore },
402 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
403 "squareToLen",
404 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { -1, ShenandoahNone},
405 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
406 "montgomery_multiply",
407 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
408 { TypeFunc::Parms+6, ShenandoahStore }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
409 "montgomery_square",
410 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+5, ShenandoahStore },
411 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
412 "mulAdd",
413 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
414 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
415 "vectorizedMismatch",
416 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
417 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
418 "updateBytesCRC32",
419 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
420 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
421 "updateBytesAdler32",
422 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
423 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
424 "updateBytesCRC32C",
425 { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad}, { -1, ShenandoahNone},
426 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
427 "counterMode_AESCrypt",
428 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
429 { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } },
430 "cipherBlockChaining_encryptAESCrypt",
431 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
432 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
433 "cipherBlockChaining_decryptAESCrypt",
434 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
435 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
436 "shenandoah_clone_barrier",
437 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
438 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
439 "ghash_processBlocks",
440 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
441 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
442 "sha1_implCompress",
443 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
444 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
445 "sha256_implCompress",
446 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
447 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
448 "sha512_implCompress",
449 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
450 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
451 "sha1_implCompressMB",
452 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
453 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
454 "sha256_implCompressMB",
455 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
456 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
457 "sha512_implCompressMB",
458 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
459 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
460 "encodeBlock",
461 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahStore }, { -1, ShenandoahNone },
462 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
463 "decodeBlock",
464 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahStore }, { -1, ShenandoahNone },
465 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
466 };
467
468 if (call->is_call_to_arraycopystub()) {
469 Node* dest = nullptr;
470 const TypeTuple* args = n->as_Call()->_tf->domain_sig();
471 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
472 if (args->field_at(i)->isa_ptr()) {
473 j++;
474 if (j == 2) {
475 dest = n->in(i);
476 break;
477 }
478 }
479 }
480 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) ||
481 !verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) {
482 report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n);
483 }
484 } else if (strlen(call->_name) > 5 &&
485 !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) {
486 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) {
487 report_verify_failure("Shenandoah verification: _fill should have barriers", n);
488 }
489 } else if (!strcmp(call->_name, "shenandoah_wb_pre")) {
490 // skip
491 } else {
492 const int calls_len = sizeof(calls) / sizeof(calls[0]);
493 int i = 0;
494 for (; i < calls_len; i++) {
495 if (!strcmp(calls[i].name, call->_name)) {
496 break;
497 }
498 }
499 if (i != calls_len) {
500 const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]);
501 for (uint j = 0; j < args_len; j++) {
502 int pos = calls[i].args[j].pos;
503 if (pos == -1) {
504 break;
505 }
506 if (!verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) {
507 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
508 }
509 }
510 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
511 if (call->in(j)->bottom_type()->make_ptr() &&
512 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
513 uint k = 0;
514 for (; k < args_len && calls[i].args[k].pos != (int)j; k++);
515 if (k == args_len) {
516 fatal("arg %d for call %s not covered", j, call->_name);
517 }
518 }
519 }
520 } else {
521 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
522 if (call->in(j)->bottom_type()->make_ptr() &&
523 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
524 fatal("%s not covered", call->_name);
525 }
526 }
527 }
528 }
529 } else if (n->Opcode() == Op_ShenandoahIUBarrier || n->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
530 // skip
531 } else if (n->is_AddP()
532 || n->is_Phi()
533 || n->is_ConstraintCast()
534 || n->Opcode() == Op_Return
535 || n->Opcode() == Op_CMoveP
536 || n->Opcode() == Op_CMoveN
537 || n->Opcode() == Op_Rethrow
538 || n->is_MemBar()
539 || n->Opcode() == Op_Conv2B
540 || n->Opcode() == Op_SafePoint
541 || n->is_CallJava()
542 || n->Opcode() == Op_Unlock
543 || n->Opcode() == Op_EncodeP
544 || n->Opcode() == Op_DecodeN) {
545 // nothing to do
546 } else {
547 static struct {
548 int opcode;
549 struct {
550 int pos;
551 verify_type t;
552 } inputs[2];
553 } others[] = {
554 Op_FastLock,
555 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
556 Op_Lock,
557 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone} },
558 Op_ArrayCopy,
559 { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } },
560 Op_StrCompressedCopy,
561 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
562 Op_StrInflatedCopy,
563 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
564 Op_AryEq,
565 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
566 Op_StrIndexOf,
567 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
568 Op_StrComp,
569 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
570 Op_StrEquals,
571 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
572 Op_VectorizedHashCode,
573 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } },
574 Op_EncodeISOArray,
575 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
576 Op_CountPositives,
577 { { 2, ShenandoahLoad }, { -1, ShenandoahNone} },
578 Op_CastP2X,
579 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
580 Op_StrIndexOfChar,
581 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } },
582 };
583
584 const int others_len = sizeof(others) / sizeof(others[0]);
585 int i = 0;
586 for (; i < others_len; i++) {
587 if (others[i].opcode == n->Opcode()) {
588 break;
589 }
590 }
591 uint stop = n->is_Call() ? n->as_Call()->tf()->domain_sig()->cnt() : n->req();
592 if (i != others_len) {
593 const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]);
594 for (uint j = 0; j < inputs_len; j++) {
595 int pos = others[i].inputs[j].pos;
596 if (pos == -1) {
597 break;
598 }
599 if (!verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) {
600 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
601 }
602 }
603 for (uint j = 1; j < stop; j++) {
604 if (n->in(j) != nullptr && n->in(j)->bottom_type()->make_ptr() &&
605 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
606 uint k = 0;
607 for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++);
608 if (k == inputs_len) {
609 fatal("arg %d for node %s not covered", j, n->Name());
610 }
611 }
612 }
613 } else {
614 for (uint j = 1; j < stop; j++) {
615 if (n->in(j) != nullptr && n->in(j)->bottom_type()->make_ptr() &&
616 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
617 fatal("%s not covered", n->Name());
618 }
619 }
620 }
621 }
622
623 if (n->is_SafePoint()) {
624 SafePointNode* sfpt = n->as_SafePoint();
625 if (verify_no_useless_barrier && sfpt->jvms() != nullptr) {
626 for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) {
627 if (!verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) {
628 phis.clear();
629 visited.reset();
630 }
631 }
632 }
633 }
634 }
635
636 if (verify_no_useless_barrier) {
637 for (int i = 0; i < barriers.length(); i++) {
638 Node* n = barriers.at(i);
639 if (!barriers_used.member(n)) {
640 tty->print("XXX useless barrier"); n->dump(-2);
641 ShouldNotReachHere();
642 }
643 }
644 }
645 }
646 #endif
647
648 bool ShenandoahBarrierC2Support::is_dominator_same_ctrl(Node* c, Node* d, Node* n, PhaseIdealLoop* phase) {
649 // That both nodes have the same control is not sufficient to prove
650 // domination, verify that there's no path from d to n
651 ResourceMark rm;
652 Unique_Node_List wq;
653 wq.push(d);
654 for (uint next = 0; next < wq.size(); next++) {
655 Node *m = wq.at(next);
656 if (m == n) {
657 return false;
658 }
659 if (m->is_Phi() && m->in(0)->is_Loop()) {
660 assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control");
661 } else {
662 if (m->is_Store() || m->is_LoadStore()) {
663 // Take anti-dependencies into account
664 Node* mem = m->in(MemNode::Memory);
665 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
666 Node* u = mem->fast_out(i);
667 if (u->is_Load() && phase->C->can_alias(m->adr_type(), phase->C->get_alias_index(u->adr_type())) &&
668 phase->ctrl_or_self(u) == c) {
669 wq.push(u);
670 }
671 }
672 }
673 for (uint i = 0; i < m->req(); i++) {
674 if (m->in(i) != nullptr && phase->ctrl_or_self(m->in(i)) == c) {
675 wq.push(m->in(i));
676 }
677 }
678 }
679 }
680 return true;
681 }
682
683 bool ShenandoahBarrierC2Support::is_dominator(Node* d_c, Node* n_c, Node* d, Node* n, PhaseIdealLoop* phase) {
684 if (d_c != n_c) {
685 return phase->is_dominator(d_c, n_c);
686 }
687 return is_dominator_same_ctrl(d_c, d, n, phase);
688 }
689
690 Node* next_mem(Node* mem, int alias) {
691 Node* res = nullptr;
692 if (mem->is_Proj()) {
693 res = mem->in(0);
694 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
695 res = mem->in(TypeFunc::Memory);
696 } else if (mem->is_Phi()) {
697 res = mem->in(1);
698 } else if (mem->is_MergeMem()) {
699 res = mem->as_MergeMem()->memory_at(alias);
700 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
701 assert(alias == Compile::AliasIdxRaw, "following raw memory can't lead to a barrier");
702 res = mem->in(MemNode::Memory);
703 } else {
704 #ifdef ASSERT
705 mem->dump();
706 #endif
707 ShouldNotReachHere();
708 }
709 return res;
710 }
711
712 Node* ShenandoahBarrierC2Support::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) {
713 Node* iffproj = nullptr;
714 while (c != dom) {
715 Node* next = phase->idom(c);
716 assert(next->unique_ctrl_out_or_null() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?");
717 if (c->is_Region()) {
718 ResourceMark rm;
719 Unique_Node_List wq;
720 wq.push(c);
721 for (uint i = 0; i < wq.size(); i++) {
722 Node *n = wq.at(i);
723 if (n == next) {
724 continue;
725 }
726 if (n->is_Region()) {
727 for (uint j = 1; j < n->req(); j++) {
728 wq.push(n->in(j));
729 }
730 } else {
731 wq.push(n->in(0));
732 }
733 }
734 for (uint i = 0; i < wq.size(); i++) {
735 Node *n = wq.at(i);
736 assert(n->is_CFG(), "");
737 if (n->is_Multi()) {
738 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
739 Node* u = n->fast_out(j);
740 if (u->is_CFG()) {
741 if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj()) {
742 return NodeSentinel;
743 }
744 }
745 }
746 }
747 }
748 } else if (c->is_Proj()) {
749 if (c->is_IfProj()) {
750 if (c->as_Proj()->is_uncommon_trap_if_pattern() != nullptr) {
751 // continue;
752 } else {
753 if (!allow_one_proj) {
754 return NodeSentinel;
755 }
756 if (iffproj == nullptr) {
757 iffproj = c;
758 } else {
759 return NodeSentinel;
760 }
761 }
762 } else if (c->Opcode() == Op_JumpProj) {
763 return NodeSentinel; // unsupported
764 } else if (c->Opcode() == Op_CatchProj) {
765 return NodeSentinel; // unsupported
766 } else if (c->Opcode() == Op_CProj && next->is_NeverBranch()) {
767 return NodeSentinel; // unsupported
768 } else {
769 assert(next->unique_ctrl_out() == c, "unsupported branch pattern");
770 }
771 }
772 c = next;
773 }
774 return iffproj;
775 }
776
777 Node* ShenandoahBarrierC2Support::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) {
778 ResourceMark rm;
779 VectorSet wq;
780 wq.set(mem->_idx);
781 mem_ctrl = phase->ctrl_or_self(mem);
782 while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) {
783 mem = next_mem(mem, alias);
784 if (wq.test_set(mem->_idx)) {
785 return nullptr;
786 }
787 mem_ctrl = phase->ctrl_or_self(mem);
788 }
789 if (mem->is_MergeMem()) {
790 mem = mem->as_MergeMem()->memory_at(alias);
791 mem_ctrl = phase->ctrl_or_self(mem);
792 }
793 return mem;
794 }
795
796 Node* ShenandoahBarrierC2Support::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) {
797 Node* mem = nullptr;
798 Node* c = ctrl;
799 do {
800 if (c->is_Region()) {
801 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == nullptr; i++) {
802 Node* u = c->fast_out(i);
803 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
804 if (u->adr_type() == TypePtr::BOTTOM) {
805 mem = u;
806 }
807 }
808 }
809 } else {
810 if (c->is_Call() && c->as_Call()->adr_type() != nullptr) {
811 CallProjections* projs = c->as_Call()->extract_projections(true, false);
812 if (projs->fallthrough_memproj != nullptr) {
813 if (projs->fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
814 if (projs->catchall_memproj == nullptr) {
815 mem = projs->fallthrough_memproj;
816 } else {
817 if (phase->is_dominator(projs->fallthrough_catchproj, ctrl)) {
818 mem = projs->fallthrough_memproj;
819 } else {
820 assert(phase->is_dominator(projs->catchall_catchproj, ctrl), "one proj must dominate barrier");
821 mem = projs->catchall_memproj;
822 }
823 }
824 }
825 } else {
826 Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
827 if (proj != nullptr &&
828 proj->adr_type() == TypePtr::BOTTOM) {
829 mem = proj;
830 }
831 }
832 } else {
833 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
834 Node* u = c->fast_out(i);
835 if (u->is_Proj() &&
836 u->bottom_type() == Type::MEMORY &&
837 u->adr_type() == TypePtr::BOTTOM) {
838 assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
839 assert(mem == nullptr, "only one proj");
840 mem = u;
841 }
842 }
843 assert(!c->is_Call() || c->as_Call()->adr_type() != nullptr || mem == nullptr, "no mem projection expected");
844 }
845 }
846 c = phase->idom(c);
847 } while (mem == nullptr);
848 return mem;
849 }
850
851 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
852 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
853 Node* u = n->fast_out(i);
854 if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
855 uses.push(u);
856 }
857 }
858 }
859
860 static void hide_strip_mined_loop(OuterStripMinedLoopNode* outer, CountedLoopNode* inner, PhaseIdealLoop* phase) {
861 OuterStripMinedLoopEndNode* le = inner->outer_loop_end();
862 Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl));
863 phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl));
864 Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt);
865 phase->register_control(new_le, phase->get_loop(le), le->in(0));
866 phase->lazy_replace(outer, new_outer);
867 phase->lazy_replace(le, new_le);
868 inner->clear_strip_mined();
869 }
870
871 void ShenandoahBarrierC2Support::test_gc_state(Node*& ctrl, Node* raw_mem, Node*& test_fail_ctrl,
872 PhaseIdealLoop* phase, int flags) {
873 PhaseIterGVN& igvn = phase->igvn();
874 Node* old_ctrl = ctrl;
875
876 Node* thread = new ThreadLocalNode();
877 Node* gc_state_offset = igvn.MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
878 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, gc_state_offset);
879 Node* gc_state = new LoadBNode(old_ctrl, raw_mem, gc_state_addr,
880 DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
881 TypeInt::BYTE, MemNode::unordered);
882 Node* gc_state_and = new AndINode(gc_state, igvn.intcon(flags));
883 Node* gc_state_cmp = new CmpINode(gc_state_and, igvn.zerocon(T_INT));
884 Node* gc_state_bool = new BoolNode(gc_state_cmp, BoolTest::ne);
885
886 IfNode* gc_state_iff = new IfNode(old_ctrl, gc_state_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
887 ctrl = new IfTrueNode(gc_state_iff);
888 test_fail_ctrl = new IfFalseNode(gc_state_iff);
889
890 IdealLoopTree* loop = phase->get_loop(old_ctrl);
891 phase->register_control(gc_state_iff, loop, old_ctrl);
892 phase->register_control(ctrl, loop, gc_state_iff);
893 phase->register_control(test_fail_ctrl, loop, gc_state_iff);
894
895 phase->register_new_node(thread, old_ctrl);
896 phase->register_new_node(gc_state_addr, old_ctrl);
897 phase->register_new_node(gc_state, old_ctrl);
898 phase->register_new_node(gc_state_and, old_ctrl);
899 phase->register_new_node(gc_state_cmp, old_ctrl);
900 phase->register_new_node(gc_state_bool, old_ctrl);
901
902 phase->set_ctrl(gc_state_offset, phase->C->root());
903
904 assert(is_gc_state_test(gc_state_iff, flags), "Should match the shape");
905 }
906
907 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) {
908 Node* old_ctrl = ctrl;
909 PhaseIterGVN& igvn = phase->igvn();
910
911 const Type* val_t = igvn.type(val);
912 if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
913 Node* null_cmp = new CmpPNode(val, igvn.zerocon(T_OBJECT));
914 Node* null_test = new BoolNode(null_cmp, BoolTest::ne);
915
916 IfNode* null_iff = new IfNode(old_ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
917 ctrl = new IfTrueNode(null_iff);
918 null_ctrl = new IfFalseNode(null_iff);
919
920 IdealLoopTree* loop = phase->get_loop(old_ctrl);
921 phase->register_control(null_iff, loop, old_ctrl);
922 phase->register_control(ctrl, loop, null_iff);
923 phase->register_control(null_ctrl, loop, null_iff);
924
925 phase->register_new_node(null_cmp, old_ctrl);
926 phase->register_new_node(null_test, old_ctrl);
927 }
928 }
929
930 void ShenandoahBarrierC2Support::test_in_cset(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) {
931 Node* old_ctrl = ctrl;
932 PhaseIterGVN& igvn = phase->igvn();
933
934 Node* raw_val = new CastP2XNode(old_ctrl, val);
935 Node* cset_idx = new URShiftXNode(raw_val, igvn.intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
936
937 // Figure out the target cset address with raw pointer math.
938 // This avoids matching AddP+LoadB that would emit inefficient code.
939 // See JDK-8245465.
940 Node* cset_addr_ptr = igvn.makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
941 Node* cset_addr = new CastP2XNode(old_ctrl, cset_addr_ptr);
942 Node* cset_load_addr = new AddXNode(cset_addr, cset_idx);
943 Node* cset_load_ptr = new CastX2PNode(cset_load_addr);
944
945 Node* cset_load = new LoadBNode(old_ctrl, raw_mem, cset_load_ptr,
946 DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(nullptr),
947 TypeInt::BYTE, MemNode::unordered);
948 Node* cset_cmp = new CmpINode(cset_load, igvn.zerocon(T_INT));
949 Node* cset_bool = new BoolNode(cset_cmp, BoolTest::ne);
950
951 IfNode* cset_iff = new IfNode(old_ctrl, cset_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
952 ctrl = new IfTrueNode(cset_iff);
953 not_cset_ctrl = new IfFalseNode(cset_iff);
954
955 IdealLoopTree *loop = phase->get_loop(old_ctrl);
956 phase->register_control(cset_iff, loop, old_ctrl);
957 phase->register_control(ctrl, loop, cset_iff);
958 phase->register_control(not_cset_ctrl, loop, cset_iff);
959
960 phase->set_ctrl(cset_addr_ptr, phase->C->root());
961
962 phase->register_new_node(raw_val, old_ctrl);
963 phase->register_new_node(cset_idx, old_ctrl);
964 phase->register_new_node(cset_addr, old_ctrl);
965 phase->register_new_node(cset_load_addr, old_ctrl);
966 phase->register_new_node(cset_load_ptr, old_ctrl);
967 phase->register_new_node(cset_load, old_ctrl);
968 phase->register_new_node(cset_cmp, old_ctrl);
969 phase->register_new_node(cset_bool, old_ctrl);
970 }
971
972 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node* load_addr,
973 DecoratorSet decorators, PhaseIdealLoop* phase) {
974 IdealLoopTree*loop = phase->get_loop(ctrl);
975 const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr();
976
977 address calladdr = nullptr;
978 const char* name = nullptr;
979 bool is_strong = ShenandoahBarrierSet::is_strong_access(decorators);
980 bool is_weak = ShenandoahBarrierSet::is_weak_access(decorators);
981 bool is_phantom = ShenandoahBarrierSet::is_phantom_access(decorators);
982 bool is_native = ShenandoahBarrierSet::is_native_access(decorators);
983 bool is_narrow = UseCompressedOops && !is_native;
984 if (is_strong) {
985 if (is_narrow) {
986 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong_narrow);
987 name = "load_reference_barrier_strong_narrow";
988 } else {
989 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_strong);
990 name = "load_reference_barrier_strong";
991 }
992 } else if (is_weak) {
993 if (is_narrow) {
994 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak_narrow);
995 name = "load_reference_barrier_weak_narrow";
996 } else {
997 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_weak);
998 name = "load_reference_barrier_weak";
999 }
1000 } else {
1001 assert(is_phantom, "only remaining strength");
1002 if (is_narrow) {
1003 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom_narrow);
1004 name = "load_reference_barrier_phantom_narrow";
1005 } else {
1006 calladdr = CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_phantom);
1007 name = "load_reference_barrier_phantom";
1008 }
1009 }
1010 Node* call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type(), calladdr, name, TypeRawPtr::BOTTOM);
1011
1012 call->init_req(TypeFunc::Control, ctrl);
1013 call->init_req(TypeFunc::I_O, phase->C->top());
1014 call->init_req(TypeFunc::Memory, phase->C->top());
1015 call->init_req(TypeFunc::FramePtr, phase->C->top());
1016 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1017 call->init_req(TypeFunc::Parms, val);
1018 call->init_req(TypeFunc::Parms+1, load_addr);
1019 phase->register_control(call, loop, ctrl);
1020 ctrl = new ProjNode(call, TypeFunc::Control);
1021 phase->register_control(ctrl, loop, call);
1022 val = new ProjNode(call, TypeFunc::Parms);
1023 phase->register_new_node(val, call);
1024 val = new CheckCastPPNode(ctrl, val, obj_type);
1025 phase->register_new_node(val, ctrl);
1026 }
1027
1028 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) {
1029 Node* ctrl = phase->get_ctrl(barrier);
1030 Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
1031
1032 // Update the control of all nodes that should be after the
1033 // barrier control flow
1034 uses.clear();
1035 // Every node that is control dependent on the barrier's input
1036 // control will be after the expanded barrier. The raw memory (if
1037 // its memory is control dependent on the barrier's input control)
1038 // must stay above the barrier.
1039 uses_to_ignore.clear();
1040 if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
1041 uses_to_ignore.push(init_raw_mem);
1042 }
1043 for (uint next = 0; next < uses_to_ignore.size(); next++) {
1044 Node *n = uses_to_ignore.at(next);
1045 for (uint i = 0; i < n->req(); i++) {
1046 Node* in = n->in(i);
1047 if (in != nullptr && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
1048 uses_to_ignore.push(in);
1049 }
1050 }
1051 }
1052 for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
1053 Node* u = ctrl->fast_out(i);
1054 if (u->_idx < last &&
1055 u != barrier &&
1056 !uses_to_ignore.member(u) &&
1057 (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) &&
1058 (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) {
1059 Node* old_c = phase->ctrl_or_self(u);
1060 Node* c = old_c;
1061 if (c != ctrl ||
1062 is_dominator_same_ctrl(old_c, barrier, u, phase) ||
1063 ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) {
1064 phase->igvn().rehash_node_delayed(u);
1065 int nb = u->replace_edge(ctrl, region, &phase->igvn());
1066 if (u->is_CFG()) {
1067 if (phase->idom(u) == ctrl) {
1068 phase->set_idom(u, region, phase->dom_depth(region));
1069 }
1070 } else if (phase->get_ctrl(u) == ctrl) {
1071 assert(u != init_raw_mem, "should leave input raw mem above the barrier");
1072 uses.push(u);
1073 }
1074 assert(nb == 1, "more than 1 ctrl input?");
1075 --i, imax -= nb;
1076 }
1077 }
1078 }
1079 }
1080
1081 static Node* create_phis_on_call_return(Node* ctrl, Node* c, Node* n, Node* n_clone, const CallProjections* projs, PhaseIdealLoop* phase) {
1082 Node* region = nullptr;
1083 while (c != ctrl) {
1084 if (c->is_Region()) {
1085 region = c;
1086 }
1087 c = phase->idom(c);
1088 }
1089 assert(region != nullptr, "");
1090 Node* phi = new PhiNode(region, n->bottom_type());
1091 for (uint j = 1; j < region->req(); j++) {
1092 Node* in = region->in(j);
1093 if (phase->is_dominator(projs->fallthrough_catchproj, in)) {
1094 phi->init_req(j, n);
1095 } else if (phase->is_dominator(projs->catchall_catchproj, in)) {
1096 phi->init_req(j, n_clone);
1097 } else {
1098 phi->init_req(j, create_phis_on_call_return(ctrl, in, n, n_clone, projs, phase));
1099 }
1100 }
1101 phase->register_new_node(phi, region);
1102 return phi;
1103 }
1104
1105 void ShenandoahBarrierC2Support::pin_and_expand(PhaseIdealLoop* phase) {
1106 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
1107
1108 Unique_Node_List uses;
1109 for (int i = 0; i < state->iu_barriers_count(); i++) {
1110 Node* barrier = state->iu_barrier(i);
1111 Node* ctrl = phase->get_ctrl(barrier);
1112 IdealLoopTree* loop = phase->get_loop(ctrl);
1113 Node* head = loop->head();
1114 if (head->is_OuterStripMinedLoop()) {
1115 // Expanding a barrier here will break loop strip mining
1116 // verification. Transform the loop so the loop nest doesn't
1117 // appear as strip mined.
1118 OuterStripMinedLoopNode* outer = head->as_OuterStripMinedLoop();
1119 hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1120 }
1121 }
1122
1123 Node_Stack stack(0);
1124 Node_List clones;
1125 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1126 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1127
1128 Node* ctrl = phase->get_ctrl(lrb);
1129 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1130
1131 CallStaticJavaNode* unc = nullptr;
1132 Node* unc_ctrl = nullptr;
1133 Node* uncasted_val = val;
1134
1135 for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1136 Node* u = lrb->fast_out(i);
1137 if (u->Opcode() == Op_CastPP &&
1138 u->in(0) != nullptr &&
1139 phase->is_dominator(u->in(0), ctrl)) {
1140 const Type* u_t = phase->igvn().type(u);
1141
1142 if (u_t->meet(TypePtr::NULL_PTR) != u_t &&
1143 u->in(0)->Opcode() == Op_IfTrue &&
1144 u->in(0)->as_Proj()->is_uncommon_trap_if_pattern() &&
1145 u->in(0)->in(0)->is_If() &&
1146 u->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
1147 u->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
1148 u->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
1149 u->in(0)->in(0)->in(1)->in(1)->in(1) == val &&
1150 u->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
1151 IdealLoopTree* loop = phase->get_loop(ctrl);
1152 IdealLoopTree* unc_loop = phase->get_loop(u->in(0));
1153
1154 if (!unc_loop->is_member(loop)) {
1155 continue;
1156 }
1157
1158 Node* branch = no_branches(ctrl, u->in(0), false, phase);
1159 assert(branch == nullptr || branch == NodeSentinel, "was not looking for a branch");
1160 if (branch == NodeSentinel) {
1161 continue;
1162 }
1163
1164 Node* iff = u->in(0)->in(0);
1165 Node* bol = iff->in(1)->clone();
1166 Node* cmp = bol->in(1)->clone();
1167 cmp->set_req(1, lrb);
1168 bol->set_req(1, cmp);
1169 phase->igvn().replace_input_of(iff, 1, bol);
1170 phase->set_ctrl(lrb, iff->in(0));
1171 phase->register_new_node(cmp, iff->in(0));
1172 phase->register_new_node(bol, iff->in(0));
1173 break;
1174 }
1175 }
1176 }
1177 if ((ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) || ctrl->is_CallJava()) {
1178 CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_CallJava() : ctrl->as_CallJava();
1179 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
1180 // The rethrow call may have too many projections to be
1181 // properly handled here. Given there's no reason for a
1182 // barrier to depend on the call, move it above the call
1183 stack.push(lrb, 0);
1184 do {
1185 Node* n = stack.node();
1186 uint idx = stack.index();
1187 if (idx < n->req()) {
1188 Node* in = n->in(idx);
1189 stack.set_index(idx+1);
1190 if (in != nullptr) {
1191 if (phase->has_ctrl(in)) {
1192 if (phase->is_dominator(call, phase->get_ctrl(in))) {
1193 #ifdef ASSERT
1194 for (uint i = 0; i < stack.size(); i++) {
1195 assert(stack.node_at(i) != in, "node shouldn't have been seen yet");
1196 }
1197 #endif
1198 stack.push(in, 0);
1199 }
1200 } else {
1201 assert(phase->is_dominator(in, call->in(0)), "no dependency on the call");
1202 }
1203 }
1204 } else {
1205 phase->set_ctrl(n, call->in(0));
1206 stack.pop();
1207 }
1208 } while(stack.size() > 0);
1209 continue;
1210 }
1211 CallProjections* projs = call->extract_projections(false, false);
1212 #ifdef ASSERT
1213 VectorSet cloned;
1214 #endif
1215 Node* lrb_clone = lrb->clone();
1216 phase->register_new_node(lrb_clone, projs->catchall_catchproj);
1217 phase->set_ctrl(lrb, projs->fallthrough_catchproj);
1218
1219 stack.push(lrb, 0);
1220 clones.push(lrb_clone);
1221
1222 do {
1223 assert(stack.size() == clones.size(), "");
1224 Node* n = stack.node();
1225 #ifdef ASSERT
1226 if (n->is_Load()) {
1227 Node* mem = n->in(MemNode::Memory);
1228 for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
1229 Node* u = mem->fast_out(j);
1230 assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?");
1231 }
1232 }
1233 #endif
1234 uint idx = stack.index();
1235 Node* n_clone = clones.at(clones.size()-1);
1236 if (idx < n->outcnt()) {
1237 Node* u = n->raw_out(idx);
1238 Node* c = phase->ctrl_or_self(u);
1239 if (phase->is_dominator(call, c) && phase->is_dominator(c, projs->fallthrough_proj)) {
1240 stack.set_index(idx+1);
1241 assert(!u->is_CFG(), "");
1242 stack.push(u, 0);
1243 assert(!cloned.test_set(u->_idx), "only one clone");
1244 Node* u_clone = u->clone();
1245 int nb = u_clone->replace_edge(n, n_clone, &phase->igvn());
1246 assert(nb > 0, "should have replaced some uses");
1247 phase->register_new_node(u_clone, projs->catchall_catchproj);
1248 clones.push(u_clone);
1249 phase->set_ctrl(u, projs->fallthrough_catchproj);
1250 } else {
1251 bool replaced = false;
1252 if (u->is_Phi()) {
1253 for (uint k = 1; k < u->req(); k++) {
1254 if (u->in(k) == n) {
1255 if (phase->is_dominator(projs->catchall_catchproj, u->in(0)->in(k))) {
1256 phase->igvn().replace_input_of(u, k, n_clone);
1257 replaced = true;
1258 } else if (!phase->is_dominator(projs->fallthrough_catchproj, u->in(0)->in(k))) {
1259 phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase));
1260 replaced = true;
1261 }
1262 }
1263 }
1264 } else {
1265 if (phase->is_dominator(projs->catchall_catchproj, c)) {
1266 phase->igvn().rehash_node_delayed(u);
1267 int nb = u->replace_edge(n, n_clone, &phase->igvn());
1268 assert(nb > 0, "should have replaced some uses");
1269 replaced = true;
1270 } else if (!phase->is_dominator(projs->fallthrough_catchproj, c)) {
1271 if (u->is_If()) {
1272 // Can't break If/Bool/Cmp chain
1273 assert(n->is_Bool(), "unexpected If shape");
1274 assert(stack.node_at(stack.size()-2)->is_Cmp(), "unexpected If shape");
1275 assert(n_clone->is_Bool(), "unexpected clone");
1276 assert(clones.at(clones.size()-2)->is_Cmp(), "unexpected clone");
1277 Node* bol_clone = n->clone();
1278 Node* cmp_clone = stack.node_at(stack.size()-2)->clone();
1279 bol_clone->set_req(1, cmp_clone);
1280
1281 Node* nn = stack.node_at(stack.size()-3);
1282 Node* nn_clone = clones.at(clones.size()-3);
1283 assert(nn->Opcode() == nn_clone->Opcode(), "mismatch");
1284
1285 int nb = cmp_clone->replace_edge(nn, create_phis_on_call_return(ctrl, c, nn, nn_clone, projs, phase),
1286 &phase->igvn());
1287 assert(nb > 0, "should have replaced some uses");
1288
1289 phase->register_new_node(bol_clone, u->in(0));
1290 phase->register_new_node(cmp_clone, u->in(0));
1291
1292 phase->igvn().replace_input_of(u, 1, bol_clone);
1293
1294 } else {
1295 phase->igvn().rehash_node_delayed(u);
1296 int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase), &phase->igvn());
1297 assert(nb > 0, "should have replaced some uses");
1298 }
1299 replaced = true;
1300 }
1301 }
1302 if (!replaced) {
1303 stack.set_index(idx+1);
1304 }
1305 }
1306 } else {
1307 stack.pop();
1308 clones.pop();
1309 }
1310 } while (stack.size() > 0);
1311 assert(stack.size() == 0 && clones.size() == 0, "");
1312 }
1313 }
1314
1315 for (int i = 0; i < state->load_reference_barriers_count(); i++) {
1316 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1317 Node* ctrl = phase->get_ctrl(lrb);
1318 IdealLoopTree* loop = phase->get_loop(ctrl);
1319 Node* head = loop->head();
1320 if (head->is_OuterStripMinedLoop()) {
1321 // Expanding a barrier here will break loop strip mining
1322 // verification. Transform the loop so the loop nest doesn't
1323 // appear as strip mined.
1324 OuterStripMinedLoopNode* outer = head->as_OuterStripMinedLoop();
1325 hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1326 }
1327 }
1328
1329 // Expand load-reference-barriers
1330 MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
1331 Unique_Node_List uses_to_ignore;
1332 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1333 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1334 uint last = phase->C->unique();
1335 Node* ctrl = phase->get_ctrl(lrb);
1336 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1337
1338 Node* orig_ctrl = ctrl;
1339
1340 Node* raw_mem = fixer.find_mem(ctrl, lrb);
1341 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, nullptr);
1342
1343 IdealLoopTree *loop = phase->get_loop(ctrl);
1344
1345 Node* heap_stable_ctrl = nullptr;
1346 Node* null_ctrl = nullptr;
1347
1348 assert(val->bottom_type()->make_oopptr(), "need oop");
1349 assert(val->bottom_type()->make_oopptr()->const_oop() == nullptr, "expect non-constant");
1350
1351 enum { _heap_stable = 1, _evac_path, _not_cset, PATH_LIMIT };
1352 Node* region = new RegionNode(PATH_LIMIT);
1353 Node* val_phi = new PhiNode(region, val->bottom_type()->is_oopptr());
1354
1355 // Stable path.
1356 int flags = ShenandoahHeap::HAS_FORWARDED;
1357 if (!ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1358 flags |= ShenandoahHeap::WEAK_ROOTS;
1359 }
1360 test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, flags);
1361 IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
1362
1363 // Heap stable case
1364 region->init_req(_heap_stable, heap_stable_ctrl);
1365 val_phi->init_req(_heap_stable, val);
1366
1367 // Test for in-cset, unless it's a native-LRB. Native LRBs need to return null
1368 // even for non-cset objects to prevent resurrection of such objects.
1369 // Wires !in_cset(obj) to slot 2 of region and phis
1370 Node* not_cset_ctrl = nullptr;
1371 if (ShenandoahBarrierSet::is_strong_access(lrb->decorators())) {
1372 test_in_cset(ctrl, not_cset_ctrl, val, raw_mem, phase);
1373 }
1374 if (not_cset_ctrl != nullptr) {
1375 region->init_req(_not_cset, not_cset_ctrl);
1376 val_phi->init_req(_not_cset, val);
1377 } else {
1378 region->del_req(_not_cset);
1379 val_phi->del_req(_not_cset);
1380 }
1381
1382 // Resolve object when orig-value is in cset.
1383 // Make the unconditional resolve for fwdptr.
1384
1385 // Call lrb-stub and wire up that path in slots 4
1386 Node* result_mem = nullptr;
1387
1388 Node* addr;
1389 {
1390 VectorSet visited;
1391 addr = get_load_addr(phase, visited, lrb);
1392 }
1393 if (addr->Opcode() == Op_AddP) {
1394 Node* orig_base = addr->in(AddPNode::Base);
1395 Node* base = new CheckCastPPNode(ctrl, orig_base, orig_base->bottom_type(), ConstraintCastNode::StrongDependency);
1396 phase->register_new_node(base, ctrl);
1397 if (addr->in(AddPNode::Base) == addr->in((AddPNode::Address))) {
1398 // Field access
1399 addr = addr->clone();
1400 addr->set_req(AddPNode::Base, base);
1401 addr->set_req(AddPNode::Address, base);
1402 phase->register_new_node(addr, ctrl);
1403 } else {
1404 Node* addr2 = addr->in(AddPNode::Address);
1405 if (addr2->Opcode() == Op_AddP && addr2->in(AddPNode::Base) == addr2->in(AddPNode::Address) &&
1406 addr2->in(AddPNode::Base) == orig_base) {
1407 addr2 = addr2->clone();
1408 addr2->set_req(AddPNode::Base, base);
1409 addr2->set_req(AddPNode::Address, base);
1410 phase->register_new_node(addr2, ctrl);
1411 addr = addr->clone();
1412 addr->set_req(AddPNode::Base, base);
1413 addr->set_req(AddPNode::Address, addr2);
1414 phase->register_new_node(addr, ctrl);
1415 }
1416 }
1417 }
1418 call_lrb_stub(ctrl, val, addr, lrb->decorators(), phase);
1419 region->init_req(_evac_path, ctrl);
1420 val_phi->init_req(_evac_path, val);
1421
1422 phase->register_control(region, loop, heap_stable_iff);
1423 Node* out_val = val_phi;
1424 phase->register_new_node(val_phi, region);
1425
1426 fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase);
1427
1428 ctrl = orig_ctrl;
1429
1430 phase->igvn().replace_node(lrb, out_val);
1431
1432 follow_barrier_uses(out_val, ctrl, uses, phase);
1433
1434 for(uint next = 0; next < uses.size(); next++ ) {
1435 Node *n = uses.at(next);
1436 assert(phase->get_ctrl(n) == ctrl, "bad control");
1437 assert(n != raw_mem, "should leave input raw mem above the barrier");
1438 phase->set_ctrl(n, region);
1439 follow_barrier_uses(n, ctrl, uses, phase);
1440 }
1441 fixer.record_new_ctrl(ctrl, region, raw_mem, raw_mem_for_ctrl);
1442 }
1443 // Done expanding load-reference-barriers.
1444 assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced");
1445
1446 for (int i = state->iu_barriers_count() - 1; i >= 0; i--) {
1447 Node* barrier = state->iu_barrier(i);
1448 Node* pre_val = barrier->in(1);
1449
1450 if (phase->igvn().type(pre_val)->higher_equal(TypePtr::NULL_PTR)) {
1451 ShouldNotReachHere();
1452 continue;
1453 }
1454
1455 Node* ctrl = phase->get_ctrl(barrier);
1456
1457 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) {
1458 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0)->in(0), pre_val, ctrl->in(0), phase), "can't move");
1459 ctrl = ctrl->in(0)->in(0);
1460 phase->set_ctrl(barrier, ctrl);
1461 } else if (ctrl->is_CallRuntime()) {
1462 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0), pre_val, ctrl, phase), "can't move");
1463 ctrl = ctrl->in(0);
1464 phase->set_ctrl(barrier, ctrl);
1465 }
1466
1467 Node* init_ctrl = ctrl;
1468 IdealLoopTree* loop = phase->get_loop(ctrl);
1469 Node* raw_mem = fixer.find_mem(ctrl, barrier);
1470 Node* init_raw_mem = raw_mem;
1471 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, nullptr);
1472 Node* heap_stable_ctrl = nullptr;
1473 Node* null_ctrl = nullptr;
1474 uint last = phase->C->unique();
1475
1476 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
1477 Node* region = new RegionNode(PATH_LIMIT);
1478 Node* phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1479
1480 enum { _fast_path = 1, _slow_path, _null_path, PATH_LIMIT2 };
1481 Node* region2 = new RegionNode(PATH_LIMIT2);
1482 Node* phi2 = PhiNode::make(region2, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1483
1484 // Stable path.
1485 test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, ShenandoahHeap::MARKING);
1486 region->init_req(_heap_stable, heap_stable_ctrl);
1487 phi->init_req(_heap_stable, raw_mem);
1488
1489 // Null path
1490 Node* reg2_ctrl = nullptr;
1491 test_null(ctrl, pre_val, null_ctrl, phase);
1492 if (null_ctrl != nullptr) {
1493 reg2_ctrl = null_ctrl->in(0);
1494 region2->init_req(_null_path, null_ctrl);
1495 phi2->init_req(_null_path, raw_mem);
1496 } else {
1497 region2->del_req(_null_path);
1498 phi2->del_req(_null_path);
1499 }
1500
1501 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
1502 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1503 Node* thread = new ThreadLocalNode();
1504 phase->register_new_node(thread, ctrl);
1505 Node* buffer_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(buffer_offset));
1506 phase->register_new_node(buffer_adr, ctrl);
1507 Node* index_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(index_offset));
1508 phase->register_new_node(index_adr, ctrl);
1509
1510 BasicType index_bt = TypeX_X->basic_type();
1511 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading Shenandoah SATBMarkQueue::_index with wrong size.");
1512 const TypePtr* adr_type = TypeRawPtr::BOTTOM;
1513 Node* index = new LoadXNode(ctrl, raw_mem, index_adr, adr_type, TypeX_X, MemNode::unordered);
1514 phase->register_new_node(index, ctrl);
1515 Node* index_cmp = new CmpXNode(index, phase->igvn().MakeConX(0));
1516 phase->register_new_node(index_cmp, ctrl);
1517 Node* index_test = new BoolNode(index_cmp, BoolTest::ne);
1518 phase->register_new_node(index_test, ctrl);
1519 IfNode* queue_full_iff = new IfNode(ctrl, index_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
1520 if (reg2_ctrl == nullptr) reg2_ctrl = queue_full_iff;
1521 phase->register_control(queue_full_iff, loop, ctrl);
1522 Node* not_full = new IfTrueNode(queue_full_iff);
1523 phase->register_control(not_full, loop, queue_full_iff);
1524 Node* full = new IfFalseNode(queue_full_iff);
1525 phase->register_control(full, loop, queue_full_iff);
1526
1527 ctrl = not_full;
1528
1529 Node* next_index = new SubXNode(index, phase->igvn().MakeConX(sizeof(intptr_t)));
1530 phase->register_new_node(next_index, ctrl);
1531
1532 Node* buffer = new LoadPNode(ctrl, raw_mem, buffer_adr, adr_type, TypeRawPtr::NOTNULL, MemNode::unordered);
1533 phase->register_new_node(buffer, ctrl);
1534 Node *log_addr = new AddPNode(phase->C->top(), buffer, next_index);
1535 phase->register_new_node(log_addr, ctrl);
1536 Node* log_store = new StorePNode(ctrl, raw_mem, log_addr, adr_type, pre_val, MemNode::unordered);
1537 phase->register_new_node(log_store, ctrl);
1538 // update the index
1539 Node* index_update = new StoreXNode(ctrl, log_store, index_adr, adr_type, next_index, MemNode::unordered);
1540 phase->register_new_node(index_update, ctrl);
1541
1542 // Fast-path case
1543 region2->init_req(_fast_path, ctrl);
1544 phi2->init_req(_fast_path, index_update);
1545
1546 ctrl = full;
1547
1548 Node* base = find_bottom_mem(ctrl, phase);
1549
1550 MergeMemNode* mm = MergeMemNode::make(base);
1551 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
1552 phase->register_new_node(mm, ctrl);
1553
1554 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);
1555 call->init_req(TypeFunc::Control, ctrl);
1556 call->init_req(TypeFunc::I_O, phase->C->top());
1557 call->init_req(TypeFunc::Memory, mm);
1558 call->init_req(TypeFunc::FramePtr, phase->C->top());
1559 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1560 call->init_req(TypeFunc::Parms, pre_val);
1561 call->init_req(TypeFunc::Parms+1, thread);
1562 phase->register_control(call, loop, ctrl);
1563
1564 Node* ctrl_proj = new ProjNode(call, TypeFunc::Control);
1565 phase->register_control(ctrl_proj, loop, call);
1566 Node* mem_proj = new ProjNode(call, TypeFunc::Memory);
1567 phase->register_new_node(mem_proj, call);
1568
1569 // Slow-path case
1570 region2->init_req(_slow_path, ctrl_proj);
1571 phi2->init_req(_slow_path, mem_proj);
1572
1573 phase->register_control(region2, loop, reg2_ctrl);
1574 phase->register_new_node(phi2, region2);
1575
1576 region->init_req(_heap_unstable, region2);
1577 phi->init_req(_heap_unstable, phi2);
1578
1579 phase->register_control(region, loop, heap_stable_ctrl->in(0));
1580 phase->register_new_node(phi, region);
1581
1582 fix_ctrl(barrier, region, fixer, uses, uses_to_ignore, last, phase);
1583 for(uint next = 0; next < uses.size(); next++ ) {
1584 Node *n = uses.at(next);
1585 assert(phase->get_ctrl(n) == init_ctrl, "bad control");
1586 assert(n != init_raw_mem, "should leave input raw mem above the barrier");
1587 phase->set_ctrl(n, region);
1588 follow_barrier_uses(n, init_ctrl, uses, phase);
1589 }
1590 fixer.fix_mem(init_ctrl, region, init_raw_mem, raw_mem_for_ctrl, phi, uses);
1591
1592 phase->igvn().replace_node(barrier, pre_val);
1593 }
1594 assert(state->iu_barriers_count() == 0, "all enqueue barrier nodes should have been replaced");
1595
1596 }
1597
1598 Node* ShenandoahBarrierC2Support::get_load_addr(PhaseIdealLoop* phase, VectorSet& visited, Node* in) {
1599 if (visited.test_set(in->_idx)) {
1600 return nullptr;
1601 }
1602 switch (in->Opcode()) {
1603 case Op_Proj:
1604 return get_load_addr(phase, visited, in->in(0));
1605 case Op_CastPP:
1606 case Op_CheckCastPP:
1607 case Op_DecodeN:
1608 case Op_EncodeP:
1609 return get_load_addr(phase, visited, in->in(1));
1610 case Op_LoadN:
1611 case Op_LoadP:
1612 return in->in(MemNode::Address);
1613 case Op_CompareAndExchangeN:
1614 case Op_CompareAndExchangeP:
1615 case Op_GetAndSetN:
1616 case Op_GetAndSetP:
1617 case Op_ShenandoahCompareAndExchangeP:
1618 case Op_ShenandoahCompareAndExchangeN:
1619 // Those instructions would just have stored a different
1620 // value into the field. No use to attempt to fix it at this point.
1621 return phase->igvn().zerocon(T_OBJECT);
1622 case Op_CMoveP:
1623 case Op_CMoveN: {
1624 Node* t = get_load_addr(phase, visited, in->in(CMoveNode::IfTrue));
1625 Node* f = get_load_addr(phase, visited, in->in(CMoveNode::IfFalse));
1626 // Handle unambiguous cases: single address reported on both branches.
1627 if (t != nullptr && f == nullptr) return t;
1628 if (t == nullptr && f != nullptr) return f;
1629 if (t != nullptr && t == f) return t;
1630 // Ambiguity.
1631 return phase->igvn().zerocon(T_OBJECT);
1632 }
1633 case Op_Phi: {
1634 Node* addr = nullptr;
1635 for (uint i = 1; i < in->req(); i++) {
1636 Node* addr1 = get_load_addr(phase, visited, in->in(i));
1637 if (addr == nullptr) {
1638 addr = addr1;
1639 }
1640 if (addr != addr1) {
1641 return phase->igvn().zerocon(T_OBJECT);
1642 }
1643 }
1644 return addr;
1645 }
1646 case Op_ShenandoahLoadReferenceBarrier:
1647 return get_load_addr(phase, visited, in->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1648 case Op_ShenandoahIUBarrier:
1649 return get_load_addr(phase, visited, in->in(1));
1650 case Op_CallDynamicJava:
1651 case Op_CallLeaf:
1652 case Op_CallStaticJava:
1653 case Op_ConN:
1654 case Op_ConP:
1655 case Op_Parm:
1656 case Op_CreateEx:
1657 return phase->igvn().zerocon(T_OBJECT);
1658 default:
1659 #ifdef ASSERT
1660 fatal("Unknown node in get_load_addr: %s", NodeClassNames[in->Opcode()]);
1661 #endif
1662 return phase->igvn().zerocon(T_OBJECT);
1663 }
1664
1665 }
1666
1667 void ShenandoahBarrierC2Support::move_gc_state_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
1668 IdealLoopTree *loop = phase->get_loop(iff);
1669 Node* loop_head = loop->_head;
1670 Node* entry_c = loop_head->in(LoopNode::EntryControl);
1671
1672 Node* bol = iff->in(1);
1673 Node* cmp = bol->in(1);
1674 Node* andi = cmp->in(1);
1675 Node* load = andi->in(1);
1676
1677 assert(is_gc_state_load(load), "broken");
1678 if (!phase->is_dominator(load->in(0), entry_c)) {
1679 Node* mem_ctrl = nullptr;
1680 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
1681 load = load->clone();
1682 load->set_req(MemNode::Memory, mem);
1683 load->set_req(0, entry_c);
1684 phase->register_new_node(load, entry_c);
1685 andi = andi->clone();
1686 andi->set_req(1, load);
1687 phase->register_new_node(andi, entry_c);
1688 cmp = cmp->clone();
1689 cmp->set_req(1, andi);
1690 phase->register_new_node(cmp, entry_c);
1691 bol = bol->clone();
1692 bol->set_req(1, cmp);
1693 phase->register_new_node(bol, entry_c);
1694
1695 phase->igvn().replace_input_of(iff, 1, bol);
1696 }
1697 }
1698
1699 bool ShenandoahBarrierC2Support::identical_backtoback_ifs(Node* n, PhaseIdealLoop* phase) {
1700 if (!n->is_If() || n->is_CountedLoopEnd()) {
1701 return false;
1702 }
1703 Node* region = n->in(0);
1704
1705 if (!region->is_Region()) {
1706 return false;
1707 }
1708 Node* dom = phase->idom(region);
1709 if (!dom->is_If()) {
1710 return false;
1711 }
1712
1713 if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) {
1714 return false;
1715 }
1716
1717 IfNode* dom_if = dom->as_If();
1718 Node* proj_true = dom_if->proj_out(1);
1719 Node* proj_false = dom_if->proj_out(0);
1720
1721 for (uint i = 1; i < region->req(); i++) {
1722 if (phase->is_dominator(proj_true, region->in(i))) {
1723 continue;
1724 }
1725 if (phase->is_dominator(proj_false, region->in(i))) {
1726 continue;
1727 }
1728 return false;
1729 }
1730
1731 return true;
1732 }
1733
1734 void ShenandoahBarrierC2Support::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) {
1735 assert(is_heap_stable_test(n), "no other tests");
1736 if (identical_backtoback_ifs(n, phase)) {
1737 Node* n_ctrl = n->in(0);
1738 if (phase->can_split_if(n_ctrl)) {
1739 IfNode* dom_if = phase->idom(n_ctrl)->as_If();
1740 if (is_heap_stable_test(n)) {
1741 Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1);
1742 assert(is_gc_state_load(gc_state_load), "broken");
1743 Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1);
1744 assert(is_gc_state_load(dom_gc_state_load), "broken");
1745 if (gc_state_load != dom_gc_state_load) {
1746 phase->igvn().replace_node(gc_state_load, dom_gc_state_load);
1747 }
1748 }
1749 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1750 Node* proj_true = dom_if->proj_out(1);
1751 Node* proj_false = dom_if->proj_out(0);
1752 Node* con_true = phase->igvn().makecon(TypeInt::ONE);
1753 Node* con_false = phase->igvn().makecon(TypeInt::ZERO);
1754
1755 for (uint i = 1; i < n_ctrl->req(); i++) {
1756 if (phase->is_dominator(proj_true, n_ctrl->in(i))) {
1757 bolphi->init_req(i, con_true);
1758 } else {
1759 assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1760 bolphi->init_req(i, con_false);
1761 }
1762 }
1763 phase->register_new_node(bolphi, n_ctrl);
1764 phase->igvn().replace_input_of(n, 1, bolphi);
1765 phase->do_split_if(n);
1766 }
1767 }
1768 }
1769
1770 IfNode* ShenandoahBarrierC2Support::find_unswitching_candidate(const IdealLoopTree* loop, PhaseIdealLoop* phase) {
1771 // Find first invariant test that doesn't exit the loop
1772 LoopNode *head = loop->_head->as_Loop();
1773 IfNode* unswitch_iff = nullptr;
1774 Node* n = head->in(LoopNode::LoopBackControl);
1775 int loop_has_sfpts = -1;
1776 while (n != head) {
1777 Node* n_dom = phase->idom(n);
1778 if (n->is_Region()) {
1779 if (n_dom->is_If()) {
1780 IfNode* iff = n_dom->as_If();
1781 if (iff->in(1)->is_Bool()) {
1782 BoolNode* bol = iff->in(1)->as_Bool();
1783 if (bol->in(1)->is_Cmp()) {
1784 // If condition is invariant and not a loop exit,
1785 // then found reason to unswitch.
1786 if (is_heap_stable_test(iff) &&
1787 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) {
1788 assert(!loop->is_loop_exit(iff), "both branches should be in the loop");
1789 if (loop_has_sfpts == -1) {
1790 for(uint i = 0; i < loop->_body.size(); i++) {
1791 Node *m = loop->_body[i];
1792 if (m->is_SafePoint() && !m->is_CallLeaf()) {
1793 loop_has_sfpts = 1;
1794 break;
1795 }
1796 }
1797 if (loop_has_sfpts == -1) {
1798 loop_has_sfpts = 0;
1799 }
1800 }
1801 if (!loop_has_sfpts) {
1802 unswitch_iff = iff;
1803 }
1804 }
1805 }
1806 }
1807 }
1808 }
1809 n = n_dom;
1810 }
1811 return unswitch_iff;
1812 }
1813
1814
1815 void ShenandoahBarrierC2Support::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) {
1816 Node_List heap_stable_tests;
1817 stack.push(phase->C->start(), 0);
1818 do {
1819 Node* n = stack.node();
1820 uint i = stack.index();
1821
1822 if (i < n->outcnt()) {
1823 Node* u = n->raw_out(i);
1824 stack.set_index(i+1);
1825 if (!visited.test_set(u->_idx)) {
1826 stack.push(u, 0);
1827 }
1828 } else {
1829 stack.pop();
1830 if (n->is_If() && is_heap_stable_test(n)) {
1831 heap_stable_tests.push(n);
1832 }
1833 }
1834 } while (stack.size() > 0);
1835
1836 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1837 Node* n = heap_stable_tests.at(i);
1838 assert(is_heap_stable_test(n), "only evacuation test");
1839 merge_back_to_back_tests(n, phase);
1840 }
1841
1842 if (!phase->C->major_progress()) {
1843 VectorSet seen;
1844 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1845 Node* n = heap_stable_tests.at(i);
1846 IdealLoopTree* loop = phase->get_loop(n);
1847 if (loop != phase->ltree_root() &&
1848 loop->_child == nullptr &&
1849 !loop->_irreducible) {
1850 Node* head = loop->_head;
1851 if (head->is_Loop() &&
1852 (!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) &&
1853 !seen.test_set(head->_idx)) {
1854 IfNode* iff = find_unswitching_candidate(loop, phase);
1855 if (iff != nullptr) {
1856 Node* bol = iff->in(1);
1857 if (head->as_Loop()->is_strip_mined()) {
1858 head->as_Loop()->verify_strip_mined(0);
1859 }
1860 move_gc_state_test_out_of_loop(iff, phase);
1861
1862 AutoNodeBudget node_budget(phase);
1863
1864 if (loop->policy_unswitching(phase)) {
1865 if (head->as_Loop()->is_strip_mined()) {
1866 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop();
1867 hide_strip_mined_loop(outer, head->as_CountedLoop(), phase);
1868 }
1869 phase->do_unswitching(loop, old_new);
1870 } else {
1871 // Not proceeding with unswitching. Move load back in
1872 // the loop.
1873 phase->igvn().replace_input_of(iff, 1, bol);
1874 }
1875 }
1876 }
1877 }
1878 }
1879 }
1880 }
1881
1882 ShenandoahIUBarrierNode::ShenandoahIUBarrierNode(Node* val) : Node(nullptr, val) {
1883 ShenandoahBarrierSetC2::bsc2()->state()->add_iu_barrier(this);
1884 }
1885
1886 const Type* ShenandoahIUBarrierNode::bottom_type() const {
1887 if (in(1) == nullptr || in(1)->is_top()) {
1888 return Type::TOP;
1889 }
1890 const Type* t = in(1)->bottom_type();
1891 if (t == TypePtr::NULL_PTR) {
1892 return t;
1893 }
1894 return t->is_oopptr();
1895 }
1896
1897 const Type* ShenandoahIUBarrierNode::Value(PhaseGVN* phase) const {
1898 if (in(1) == nullptr) {
1899 return Type::TOP;
1900 }
1901 const Type* t = phase->type(in(1));
1902 if (t == Type::TOP) {
1903 return Type::TOP;
1904 }
1905 if (t == TypePtr::NULL_PTR) {
1906 return t;
1907 }
1908 return t->is_oopptr();
1909 }
1910
1911 int ShenandoahIUBarrierNode::needed(Node* n) {
1912 if (n == nullptr ||
1913 n->is_Allocate() ||
1914 n->Opcode() == Op_ShenandoahIUBarrier ||
1915 n->bottom_type() == TypePtr::NULL_PTR ||
1916 (n->bottom_type()->make_oopptr() != nullptr && n->bottom_type()->make_oopptr()->const_oop() != nullptr)) {
1917 return NotNeeded;
1918 }
1919 if (n->is_Phi() ||
1920 n->is_CMove()) {
1921 return MaybeNeeded;
1922 }
1923 return Needed;
1924 }
1925
1926 Node* ShenandoahIUBarrierNode::next(Node* n) {
1927 for (;;) {
1928 if (n == nullptr) {
1929 return n;
1930 } else if (n->bottom_type() == TypePtr::NULL_PTR) {
1931 return n;
1932 } else if (n->bottom_type()->make_oopptr() != nullptr && n->bottom_type()->make_oopptr()->const_oop() != nullptr) {
1933 return n;
1934 } else if (n->is_ConstraintCast() ||
1935 n->Opcode() == Op_DecodeN ||
1936 n->Opcode() == Op_EncodeP) {
1937 n = n->in(1);
1938 } else if (n->is_Proj()) {
1939 n = n->in(0);
1940 } else {
1941 return n;
1942 }
1943 }
1944 ShouldNotReachHere();
1945 return nullptr;
1946 }
1947
1948 Node* ShenandoahIUBarrierNode::Identity(PhaseGVN* phase) {
1949 PhaseIterGVN* igvn = phase->is_IterGVN();
1950
1951 Node* n = next(in(1));
1952
1953 int cont = needed(n);
1954
1955 if (cont == NotNeeded) {
1956 return in(1);
1957 } else if (cont == MaybeNeeded) {
1958 if (igvn == nullptr) {
1959 phase->record_for_igvn(this);
1960 return this;
1961 } else {
1962 ResourceMark rm;
1963 Unique_Node_List wq;
1964 uint wq_i = 0;
1965
1966 for (;;) {
1967 if (n->is_Phi()) {
1968 for (uint i = 1; i < n->req(); i++) {
1969 Node* m = n->in(i);
1970 if (m != nullptr) {
1971 wq.push(m);
1972 }
1973 }
1974 } else {
1975 assert(n->is_CMove(), "nothing else here");
1976 Node* m = n->in(CMoveNode::IfFalse);
1977 wq.push(m);
1978 m = n->in(CMoveNode::IfTrue);
1979 wq.push(m);
1980 }
1981 Node* orig_n = nullptr;
1982 do {
1983 if (wq_i >= wq.size()) {
1984 return in(1);
1985 }
1986 n = wq.at(wq_i);
1987 wq_i++;
1988 orig_n = n;
1989 n = next(n);
1990 cont = needed(n);
1991 if (cont == Needed) {
1992 return this;
1993 }
1994 } while (cont != MaybeNeeded || (orig_n != n && wq.member(n)));
1995 }
1996 }
1997 }
1998
1999 return this;
2000 }
2001
2002 #ifdef ASSERT
2003 static bool has_never_branch(Node* root) {
2004 for (uint i = 1; i < root->req(); i++) {
2005 Node* in = root->in(i);
2006 if (in != nullptr && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->is_NeverBranch()) {
2007 return true;
2008 }
2009 }
2010 return false;
2011 }
2012 #endif
2013
2014 void MemoryGraphFixer::collect_memory_nodes() {
2015 Node_Stack stack(0);
2016 VectorSet visited;
2017 Node_List regions;
2018
2019 // Walk the raw memory graph and create a mapping from CFG node to
2020 // memory node. Exclude phis for now.
2021 stack.push(_phase->C->root(), 1);
2022 do {
2023 Node* n = stack.node();
2024 int opc = n->Opcode();
2025 uint i = stack.index();
2026 if (i < n->req()) {
2027 Node* mem = nullptr;
2028 if (opc == Op_Root) {
2029 Node* in = n->in(i);
2030 int in_opc = in->Opcode();
2031 if (in_opc == Op_Return || in_opc == Op_Rethrow) {
2032 mem = in->in(TypeFunc::Memory);
2033 } else if (in_opc == Op_Halt) {
2034 if (in->in(0)->is_Region()) {
2035 Node* r = in->in(0);
2036 for (uint j = 1; j < r->req(); j++) {
2037 assert(!r->in(j)->is_NeverBranch(), "");
2038 }
2039 } else {
2040 Node* proj = in->in(0);
2041 assert(proj->is_Proj(), "");
2042 Node* in = proj->in(0);
2043 assert(in->is_CallStaticJava() || in->is_NeverBranch() || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
2044 if (in->is_CallStaticJava()) {
2045 mem = in->in(TypeFunc::Memory);
2046 } else if (in->Opcode() == Op_Catch) {
2047 Node* call = in->in(0)->in(0);
2048 assert(call->is_Call(), "");
2049 mem = call->in(TypeFunc::Memory);
2050 } else if (in->is_NeverBranch()) {
2051 mem = collect_memory_for_infinite_loop(in);
2052 }
2053 }
2054 } else {
2055 #ifdef ASSERT
2056 n->dump();
2057 in->dump();
2058 #endif
2059 ShouldNotReachHere();
2060 }
2061 } else {
2062 assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
2063 assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
2064 mem = n->in(i);
2065 }
2066 i++;
2067 stack.set_index(i);
2068 if (mem == nullptr) {
2069 continue;
2070 }
2071 for (;;) {
2072 if (visited.test_set(mem->_idx) || mem->is_Start()) {
2073 break;
2074 }
2075 if (mem->is_Phi()) {
2076 stack.push(mem, 2);
2077 mem = mem->in(1);
2078 } else if (mem->is_Proj()) {
2079 stack.push(mem, mem->req());
2080 mem = mem->in(0);
2081 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
2082 mem = mem->in(TypeFunc::Memory);
2083 } else if (mem->is_MergeMem()) {
2084 MergeMemNode* mm = mem->as_MergeMem();
2085 mem = mm->memory_at(_alias);
2086 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
2087 assert(_alias == Compile::AliasIdxRaw, "");
2088 stack.push(mem, mem->req());
2089 mem = mem->in(MemNode::Memory);
2090 } else {
2091 #ifdef ASSERT
2092 mem->dump();
2093 #endif
2094 ShouldNotReachHere();
2095 }
2096 }
2097 } else {
2098 if (n->is_Phi()) {
2099 // Nothing
2100 } else if (!n->is_Root()) {
2101 Node* c = get_ctrl(n);
2102 _memory_nodes.map(c->_idx, n);
2103 }
2104 stack.pop();
2105 }
2106 } while(stack.is_nonempty());
2107
2108 // Iterate over CFG nodes in rpo and propagate memory state to
2109 // compute memory state at regions, creating new phis if needed.
2110 Node_List rpo_list;
2111 visited.clear();
2112 _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
2113 Node* root = rpo_list.pop();
2114 assert(root == _phase->C->root(), "");
2115
2116 const bool trace = false;
2117 #ifdef ASSERT
2118 if (trace) {
2119 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2120 Node* c = rpo_list.at(i);
2121 if (_memory_nodes[c->_idx] != nullptr) {
2122 tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump();
2123 }
2124 }
2125 }
2126 #endif
2127 uint last = _phase->C->unique();
2128
2129 #ifdef ASSERT
2130 uint16_t max_depth = 0;
2131 for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
2132 IdealLoopTree* lpt = iter.current();
2133 max_depth = MAX2(max_depth, lpt->_nest);
2134 }
2135 #endif
2136
2137 bool progress = true;
2138 int iteration = 0;
2139 Node_List dead_phis;
2140 while (progress) {
2141 progress = false;
2142 iteration++;
2143 assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2144 if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
2145
2146 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2147 Node* c = rpo_list.at(i);
2148
2149 Node* prev_mem = _memory_nodes[c->_idx];
2150 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2151 Node* prev_region = regions[c->_idx];
2152 Node* unique = nullptr;
2153 for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
2154 Node* m = _memory_nodes[c->in(j)->_idx];
2155 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");
2156 if (m != nullptr) {
2157 if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
2158 assert(c->is_Loop() && j == LoopNode::LoopBackControl || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2159 // continue
2160 } else if (unique == nullptr) {
2161 unique = m;
2162 } else if (m == unique) {
2163 // continue
2164 } else {
2165 unique = NodeSentinel;
2166 }
2167 }
2168 }
2169 assert(unique != nullptr, "empty phi???");
2170 if (unique != NodeSentinel) {
2171 if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c) {
2172 dead_phis.push(prev_region);
2173 }
2174 regions.map(c->_idx, unique);
2175 } else {
2176 Node* phi = nullptr;
2177 if (prev_region != nullptr && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
2178 phi = prev_region;
2179 for (uint k = 1; k < c->req(); k++) {
2180 Node* m = _memory_nodes[c->in(k)->_idx];
2181 assert(m != nullptr, "expect memory state");
2182 phi->set_req(k, m);
2183 }
2184 } else {
2185 for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == nullptr; j++) {
2186 Node* u = c->fast_out(j);
2187 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2188 (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
2189 phi = u;
2190 for (uint k = 1; k < c->req() && phi != nullptr; k++) {
2191 Node* m = _memory_nodes[c->in(k)->_idx];
2192 assert(m != nullptr, "expect memory state");
2193 if (u->in(k) != m) {
2194 phi = NodeSentinel;
2195 }
2196 }
2197 }
2198 }
2199 if (phi == NodeSentinel) {
2200 phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
2201 for (uint k = 1; k < c->req(); k++) {
2202 Node* m = _memory_nodes[c->in(k)->_idx];
2203 assert(m != nullptr, "expect memory state");
2204 phi->init_req(k, m);
2205 }
2206 }
2207 }
2208 if (phi != nullptr) {
2209 regions.map(c->_idx, phi);
2210 } else {
2211 assert(c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2212 }
2213 }
2214 Node* current_region = regions[c->_idx];
2215 if (current_region != prev_region) {
2216 progress = true;
2217 if (prev_region == prev_mem) {
2218 _memory_nodes.map(c->_idx, current_region);
2219 }
2220 }
2221 } else if (prev_mem == nullptr || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
2222 Node* m = _memory_nodes[_phase->idom(c)->_idx];
2223 assert(m != nullptr || c->Opcode() == Op_Halt, "expect memory state");
2224 if (m != prev_mem) {
2225 _memory_nodes.map(c->_idx, m);
2226 progress = true;
2227 }
2228 }
2229 #ifdef ASSERT
2230 if (trace) { tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); }
2231 #endif
2232 }
2233 }
2234
2235 // Replace existing phi with computed memory state for that region
2236 // if different (could be a new phi or a dominating memory node if
2237 // that phi was found to be useless).
2238 while (dead_phis.size() > 0) {
2239 Node* n = dead_phis.pop();
2240 n->replace_by(_phase->C->top());
2241 n->destruct(&_phase->igvn());
2242 }
2243 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2244 Node* c = rpo_list.at(i);
2245 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2246 Node* n = regions[c->_idx];
2247 assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2248 if (n != nullptr && n->is_Phi() && n->_idx >= last && n->in(0) == c) {
2249 _phase->register_new_node(n, c);
2250 }
2251 }
2252 }
2253 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2254 Node* c = rpo_list.at(i);
2255 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2256 Node* n = regions[c->_idx];
2257 assert(n != nullptr || c->unique_ctrl_out()->Opcode() == Op_Halt, "expected memory state");
2258 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
2259 Node* u = c->fast_out(i);
2260 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2261 u != n) {
2262 assert(c->unique_ctrl_out()->Opcode() != Op_Halt, "expected memory state");
2263 if (u->adr_type() == TypePtr::BOTTOM) {
2264 fix_memory_uses(u, n, n, c);
2265 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2266 _phase->lazy_replace(u, n);
2267 --i; --imax;
2268 }
2269 }
2270 }
2271 }
2272 }
2273 }
2274
2275 Node* MemoryGraphFixer::collect_memory_for_infinite_loop(const Node* in) {
2276 Node* mem = nullptr;
2277 Node* head = in->in(0);
2278 assert(head->is_Region(), "unexpected infinite loop graph shape");
2279
2280 Node* phi_mem = nullptr;
2281 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2282 Node* u = head->fast_out(j);
2283 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
2284 if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2285 assert(phi_mem == nullptr || phi_mem->adr_type() == TypePtr::BOTTOM, "");
2286 phi_mem = u;
2287 } else if (u->adr_type() == TypePtr::BOTTOM) {
2288 assert(phi_mem == nullptr || _phase->C->get_alias_index(phi_mem->adr_type()) == _alias, "");
2289 if (phi_mem == nullptr) {
2290 phi_mem = u;
2291 }
2292 }
2293 }
2294 }
2295 if (phi_mem == nullptr) {
2296 ResourceMark rm;
2297 Node_Stack stack(0);
2298 stack.push(head, 1);
2299 do {
2300 Node* n = stack.node();
2301 uint i = stack.index();
2302 if (i >= n->req()) {
2303 stack.pop();
2304 } else {
2305 stack.set_index(i + 1);
2306 Node* c = n->in(i);
2307 assert(c != head, "should have found a safepoint on the way");
2308 if (stack.size() != 1 || _phase->is_dominator(head, c)) {
2309 for (;;) {
2310 if (c->is_Region()) {
2311 stack.push(c, 1);
2312 break;
2313 } else if (c->is_SafePoint() && !c->is_CallLeaf()) {
2314 Node* m = c->in(TypeFunc::Memory);
2315 if (m->is_MergeMem()) {
2316 m = m->as_MergeMem()->memory_at(_alias);
2317 }
2318 assert(mem == nullptr || mem == m, "several memory states");
2319 mem = m;
2320 break;
2321 } else {
2322 assert(c != c->in(0), "");
2323 c = c->in(0);
2324 }
2325 }
2326 }
2327 }
2328 } while (stack.size() > 0);
2329 assert(mem != nullptr, "should have found safepoint");
2330 } else {
2331 mem = phi_mem;
2332 }
2333 return mem;
2334 }
2335
2336 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
2337 Node* c = _phase->get_ctrl(n);
2338 if (n->is_Proj() && n->in(0) != nullptr && n->in(0)->is_Call()) {
2339 assert(c == n->in(0), "");
2340 CallNode* call = c->as_Call();
2341 CallProjections* projs = call->extract_projections(true, false);
2342 if (projs->catchall_memproj != nullptr) {
2343 if (projs->fallthrough_memproj == n) {
2344 c = projs->fallthrough_catchproj;
2345 } else {
2346 assert(projs->catchall_memproj == n, "");
2347 c = projs->catchall_catchproj;
2348 }
2349 }
2350 }
2351 return c;
2352 }
2353
2354 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
2355 if (_phase->has_ctrl(n))
2356 return get_ctrl(n);
2357 else {
2358 assert (n->is_CFG(), "must be a CFG node");
2359 return n;
2360 }
2361 }
2362
2363 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
2364 return m != nullptr && get_ctrl(m) == c;
2365 }
2366
2367 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
2368 assert(n == nullptr || _phase->ctrl_or_self(n) == ctrl, "");
2369 assert(!ctrl->is_Call() || ctrl == n, "projection expected");
2370 #ifdef ASSERT
2371 if ((ctrl->is_Proj() && ctrl->in(0)->is_Call()) ||
2372 (ctrl->is_Catch() && ctrl->in(0)->in(0)->is_Call())) {
2373 CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_Call() : ctrl->in(0)->in(0)->as_Call();
2374 int mems = 0;
2375 for (DUIterator_Fast imax, i = call->fast_outs(imax); i < imax; i++) {
2376 Node* u = call->fast_out(i);
2377 if (u->bottom_type() == Type::MEMORY) {
2378 mems++;
2379 }
2380 }
2381 assert(mems <= 1, "No node right after call if multiple mem projections");
2382 }
2383 #endif
2384 Node* mem = _memory_nodes[ctrl->_idx];
2385 Node* c = ctrl;
2386 while (!mem_is_valid(mem, c) &&
2387 (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
2388 c = _phase->idom(c);
2389 mem = _memory_nodes[c->_idx];
2390 }
2391 if (n != nullptr && mem_is_valid(mem, c)) {
2392 while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
2393 mem = next_mem(mem, _alias);
2394 }
2395 if (mem->is_MergeMem()) {
2396 mem = mem->as_MergeMem()->memory_at(_alias);
2397 }
2398 if (!mem_is_valid(mem, c)) {
2399 do {
2400 c = _phase->idom(c);
2401 mem = _memory_nodes[c->_idx];
2402 } while (!mem_is_valid(mem, c) &&
2403 (!c->is_CatchProj() || mem == nullptr || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
2404 }
2405 }
2406 assert(mem->bottom_type() == Type::MEMORY, "");
2407 return mem;
2408 }
2409
2410 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
2411 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2412 Node* use = region->fast_out(i);
2413 if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
2414 (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
2415 return true;
2416 }
2417 }
2418 return false;
2419 }
2420
2421 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
2422 assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
2423 const bool trace = false;
2424 DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
2425 DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
2426 GrowableArray<Node*> phis;
2427 if (mem_for_ctrl != mem) {
2428 Node* old = mem_for_ctrl;
2429 Node* prev = nullptr;
2430 while (old != mem) {
2431 prev = old;
2432 if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
2433 assert(_alias == Compile::AliasIdxRaw, "");
2434 old = old->in(MemNode::Memory);
2435 } else if (old->Opcode() == Op_SCMemProj) {
2436 assert(_alias == Compile::AliasIdxRaw, "");
2437 old = old->in(0);
2438 } else {
2439 ShouldNotReachHere();
2440 }
2441 }
2442 assert(prev != nullptr, "");
2443 if (new_ctrl != ctrl) {
2444 _memory_nodes.map(ctrl->_idx, mem);
2445 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2446 }
2447 uint input = (uint)MemNode::Memory;
2448 _phase->igvn().replace_input_of(prev, input, new_mem);
2449 } else {
2450 uses.clear();
2451 _memory_nodes.map(new_ctrl->_idx, new_mem);
2452 uses.push(new_ctrl);
2453 for(uint next = 0; next < uses.size(); next++ ) {
2454 Node *n = uses.at(next);
2455 assert(n->is_CFG(), "");
2456 DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
2457 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2458 Node* u = n->fast_out(i);
2459 if (!u->is_Root() && u->is_CFG() && u != n) {
2460 Node* m = _memory_nodes[u->_idx];
2461 if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
2462 !has_mem_phi(u) &&
2463 u->unique_ctrl_out()->Opcode() != Op_Halt) {
2464 DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
2465 DEBUG_ONLY(if (trace && m != nullptr) { tty->print("ZZZ mem"); m->dump(); });
2466
2467 if (!mem_is_valid(m, u) || !m->is_Phi()) {
2468 bool push = true;
2469 bool create_phi = true;
2470 if (_phase->is_dominator(new_ctrl, u)) {
2471 create_phi = false;
2472 }
2473 if (create_phi) {
2474 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
2475 _phase->register_new_node(phi, u);
2476 phis.push(phi);
2477 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
2478 if (!mem_is_valid(m, u)) {
2479 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
2480 _memory_nodes.map(u->_idx, phi);
2481 } else {
2482 DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
2483 for (;;) {
2484 assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), "");
2485 Node* next = nullptr;
2486 if (m->is_Proj()) {
2487 next = m->in(0);
2488 } else {
2489 assert(m->is_Mem() || m->is_LoadStore(), "");
2490 assert(_alias == Compile::AliasIdxRaw, "");
2491 next = m->in(MemNode::Memory);
2492 }
2493 if (_phase->get_ctrl(next) != u) {
2494 break;
2495 }
2496 if (next->is_MergeMem()) {
2497 assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
2498 break;
2499 }
2500 if (next->is_Phi()) {
2501 assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
2502 break;
2503 }
2504 m = next;
2505 }
2506
2507 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
2508 assert(m->is_Mem() || m->is_LoadStore(), "");
2509 uint input = (uint)MemNode::Memory;
2510 _phase->igvn().replace_input_of(m, input, phi);
2511 push = false;
2512 }
2513 } else {
2514 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
2515 }
2516 if (push) {
2517 uses.push(u);
2518 }
2519 }
2520 } else if (!mem_is_valid(m, u) &&
2521 !(u->Opcode() == Op_CProj && u->in(0)->is_NeverBranch() && u->as_Proj()->_con == 1)) {
2522 uses.push(u);
2523 }
2524 }
2525 }
2526 }
2527 for (int i = 0; i < phis.length(); i++) {
2528 Node* n = phis.at(i);
2529 Node* r = n->in(0);
2530 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
2531 for (uint j = 1; j < n->req(); j++) {
2532 Node* m = find_mem(r->in(j), nullptr);
2533 _phase->igvn().replace_input_of(n, j, m);
2534 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
2535 }
2536 }
2537 }
2538 uint last = _phase->C->unique();
2539 MergeMemNode* mm = nullptr;
2540 int alias = _alias;
2541 DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
2542 // Process loads first to not miss an anti-dependency: if the memory
2543 // edge of a store is updated before a load is processed then an
2544 // anti-dependency may be missed.
2545 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2546 Node* u = mem->out(i);
2547 if (u->_idx < last && u->is_Load() && _phase->C->get_alias_index(u->adr_type()) == alias) {
2548 Node* m = find_mem(_phase->get_ctrl(u), u);
2549 if (m != mem) {
2550 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2551 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2552 --i;
2553 }
2554 }
2555 }
2556 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2557 Node* u = mem->out(i);
2558 if (u->_idx < last) {
2559 if (u->is_Mem()) {
2560 if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2561 Node* m = find_mem(_phase->get_ctrl(u), u);
2562 if (m != mem) {
2563 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2564 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2565 --i;
2566 }
2567 }
2568 } else if (u->is_MergeMem()) {
2569 MergeMemNode* u_mm = u->as_MergeMem();
2570 if (u_mm->memory_at(alias) == mem) {
2571 MergeMemNode* newmm = nullptr;
2572 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2573 Node* uu = u->fast_out(j);
2574 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2575 if (uu->is_Phi()) {
2576 assert(uu->adr_type() == TypePtr::BOTTOM, "");
2577 Node* region = uu->in(0);
2578 int nb = 0;
2579 for (uint k = 1; k < uu->req(); k++) {
2580 if (uu->in(k) == u) {
2581 Node* m = find_mem(region->in(k), nullptr);
2582 if (m != mem) {
2583 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
2584 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2585 if (newmm != u) {
2586 _phase->igvn().replace_input_of(uu, k, newmm);
2587 nb++;
2588 --jmax;
2589 }
2590 }
2591 }
2592 }
2593 if (nb > 0) {
2594 --j;
2595 }
2596 } else {
2597 Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
2598 if (m != mem) {
2599 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
2600 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2601 if (newmm != u) {
2602 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2603 --j, --jmax;
2604 }
2605 }
2606 }
2607 }
2608 }
2609 } else if (u->is_Phi()) {
2610 assert(u->bottom_type() == Type::MEMORY, "what else?");
2611 if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
2612 Node* region = u->in(0);
2613 bool replaced = false;
2614 for (uint j = 1; j < u->req(); j++) {
2615 if (u->in(j) == mem) {
2616 Node* m = find_mem(region->in(j), nullptr);
2617 Node* nnew = m;
2618 if (m != mem) {
2619 if (u->adr_type() == TypePtr::BOTTOM) {
2620 mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
2621 nnew = mm;
2622 }
2623 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
2624 _phase->igvn().replace_input_of(u, j, nnew);
2625 replaced = true;
2626 }
2627 }
2628 }
2629 if (replaced) {
2630 --i;
2631 }
2632 }
2633 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2634 u->adr_type() == nullptr) {
2635 assert(u->adr_type() != nullptr ||
2636 u->Opcode() == Op_Rethrow ||
2637 u->Opcode() == Op_Return ||
2638 u->Opcode() == Op_SafePoint ||
2639 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2640 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2641 u->Opcode() == Op_CallLeaf, "");
2642 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2643 if (m != mem) {
2644 mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
2645 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2646 --i;
2647 }
2648 } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2649 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2650 if (m != mem) {
2651 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2652 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2653 --i;
2654 }
2655 } else if (u->adr_type() != TypePtr::BOTTOM &&
2656 _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
2657 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2658 assert(m != mem, "");
2659 // u is on the wrong slice...
2660 assert(u->is_ClearArray(), "");
2661 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2662 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2663 --i;
2664 }
2665 }
2666 }
2667 #ifdef ASSERT
2668 assert(new_mem->outcnt() > 0, "");
2669 for (int i = 0; i < phis.length(); i++) {
2670 Node* n = phis.at(i);
2671 assert(n->outcnt() > 0, "new phi must have uses now");
2672 }
2673 #endif
2674 }
2675
2676 void MemoryGraphFixer::record_new_ctrl(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl) {
2677 if (mem_for_ctrl != mem && new_ctrl != ctrl) {
2678 _memory_nodes.map(ctrl->_idx, mem);
2679 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2680 }
2681 }
2682
2683 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
2684 MergeMemNode* mm = MergeMemNode::make(mem);
2685 mm->set_memory_at(_alias, rep_proj);
2686 _phase->register_new_node(mm, rep_ctrl);
2687 return mm;
2688 }
2689
2690 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
2691 MergeMemNode* newmm = nullptr;
2692 MergeMemNode* u_mm = u->as_MergeMem();
2693 Node* c = _phase->get_ctrl(u);
2694 if (_phase->is_dominator(c, rep_ctrl)) {
2695 c = rep_ctrl;
2696 } else {
2697 assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
2698 }
2699 if (u->outcnt() == 1) {
2700 if (u->req() > (uint)_alias && u->in(_alias) == mem) {
2701 _phase->igvn().replace_input_of(u, _alias, rep_proj);
2702 --i;
2703 } else {
2704 _phase->igvn().rehash_node_delayed(u);
2705 u_mm->set_memory_at(_alias, rep_proj);
2706 }
2707 newmm = u_mm;
2708 _phase->set_ctrl_and_loop(u, c);
2709 } else {
2710 // can't simply clone u and then change one of its input because
2711 // it adds and then removes an edge which messes with the
2712 // DUIterator
2713 newmm = MergeMemNode::make(u_mm->base_memory());
2714 for (uint j = 0; j < u->req(); j++) {
2715 if (j < newmm->req()) {
2716 if (j == (uint)_alias) {
2717 newmm->set_req(j, rep_proj);
2718 } else if (newmm->in(j) != u->in(j)) {
2719 newmm->set_req(j, u->in(j));
2720 }
2721 } else if (j == (uint)_alias) {
2722 newmm->add_req(rep_proj);
2723 } else {
2724 newmm->add_req(u->in(j));
2725 }
2726 }
2727 if ((uint)_alias >= u->req()) {
2728 newmm->set_memory_at(_alias, rep_proj);
2729 }
2730 _phase->register_new_node(newmm, c);
2731 }
2732 return newmm;
2733 }
2734
2735 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
2736 if (phi->adr_type() == TypePtr::BOTTOM) {
2737 Node* region = phi->in(0);
2738 for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
2739 Node* uu = region->fast_out(j);
2740 if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
2741 return false;
2742 }
2743 }
2744 return true;
2745 }
2746 return _phase->C->get_alias_index(phi->adr_type()) == _alias;
2747 }
2748
2749 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
2750 uint last = _phase-> C->unique();
2751 MergeMemNode* mm = nullptr;
2752 assert(mem->bottom_type() == Type::MEMORY, "");
2753 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2754 Node* u = mem->out(i);
2755 if (u != replacement && u->_idx < last) {
2756 if (u->is_MergeMem()) {
2757 MergeMemNode* u_mm = u->as_MergeMem();
2758 if (u_mm->memory_at(_alias) == mem) {
2759 MergeMemNode* newmm = nullptr;
2760 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2761 Node* uu = u->fast_out(j);
2762 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2763 if (uu->is_Phi()) {
2764 if (should_process_phi(uu)) {
2765 Node* region = uu->in(0);
2766 int nb = 0;
2767 for (uint k = 1; k < uu->req(); k++) {
2768 if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
2769 if (newmm == nullptr) {
2770 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2771 }
2772 if (newmm != u) {
2773 _phase->igvn().replace_input_of(uu, k, newmm);
2774 nb++;
2775 --jmax;
2776 }
2777 }
2778 }
2779 if (nb > 0) {
2780 --j;
2781 }
2782 }
2783 } else {
2784 if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
2785 if (newmm == nullptr) {
2786 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2787 }
2788 if (newmm != u) {
2789 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2790 --j, --jmax;
2791 }
2792 }
2793 }
2794 }
2795 }
2796 } else if (u->is_Phi()) {
2797 assert(u->bottom_type() == Type::MEMORY, "what else?");
2798 Node* region = u->in(0);
2799 if (should_process_phi(u)) {
2800 bool replaced = false;
2801 for (uint j = 1; j < u->req(); j++) {
2802 if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
2803 Node* nnew = rep_proj;
2804 if (u->adr_type() == TypePtr::BOTTOM) {
2805 if (mm == nullptr) {
2806 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2807 }
2808 nnew = mm;
2809 }
2810 _phase->igvn().replace_input_of(u, j, nnew);
2811 replaced = true;
2812 }
2813 }
2814 if (replaced) {
2815 --i;
2816 }
2817
2818 }
2819 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2820 u->adr_type() == nullptr) {
2821 assert(u->adr_type() != nullptr ||
2822 u->Opcode() == Op_Rethrow ||
2823 u->Opcode() == Op_Return ||
2824 u->Opcode() == Op_SafePoint ||
2825 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2826 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2827 u->Opcode() == Op_CallLeaf, "%s", u->Name());
2828 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2829 if (mm == nullptr) {
2830 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2831 }
2832 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2833 --i;
2834 }
2835 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2836 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2837 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
2838 --i;
2839 }
2840 }
2841 }
2842 }
2843 }
2844
2845 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj, DecoratorSet decorators)
2846 : Node(ctrl, obj), _decorators(decorators) {
2847 ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this);
2848 }
2849
2850 DecoratorSet ShenandoahLoadReferenceBarrierNode::decorators() const {
2851 return _decorators;
2852 }
2853
2854 uint ShenandoahLoadReferenceBarrierNode::size_of() const {
2855 return sizeof(*this);
2856 }
2857
2858 static DecoratorSet mask_decorators(DecoratorSet decorators) {
2859 return decorators & (ON_STRONG_OOP_REF | ON_WEAK_OOP_REF | ON_PHANTOM_OOP_REF | ON_UNKNOWN_OOP_REF | IN_NATIVE);
2860 }
2861
2862 uint ShenandoahLoadReferenceBarrierNode::hash() const {
2863 uint hash = Node::hash();
2864 hash += mask_decorators(_decorators);
2865 return hash;
2866 }
2867
2868 bool ShenandoahLoadReferenceBarrierNode::cmp( const Node &n ) const {
2869 return Node::cmp(n) && n.Opcode() == Op_ShenandoahLoadReferenceBarrier &&
2870 mask_decorators(_decorators) == mask_decorators(((const ShenandoahLoadReferenceBarrierNode&)n)._decorators);
2871 }
2872
2873 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const {
2874 if (in(ValueIn) == nullptr || in(ValueIn)->is_top()) {
2875 return Type::TOP;
2876 }
2877 const Type* t = in(ValueIn)->bottom_type();
2878 if (t == TypePtr::NULL_PTR) {
2879 return t;
2880 }
2881
2882 if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2883 return t;
2884 }
2885
2886 return t->meet(TypePtr::NULL_PTR);
2887 }
2888
2889 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const {
2890 // Either input is TOP ==> the result is TOP
2891 const Type *t2 = phase->type(in(ValueIn));
2892 if( t2 == Type::TOP ) return Type::TOP;
2893
2894 if (t2 == TypePtr::NULL_PTR) {
2895 return t2;
2896 }
2897
2898 if (ShenandoahBarrierSet::is_strong_access(decorators())) {
2899 return t2;
2900 }
2901
2902 return t2->meet(TypePtr::NULL_PTR);
2903 }
2904
2905 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) {
2906 Node* value = in(ValueIn);
2907 if (!needs_barrier(phase, value)) {
2908 return value;
2909 }
2910 return this;
2911 }
2912
2913 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) {
2914 Unique_Node_List visited;
2915 return needs_barrier_impl(phase, n, visited);
2916 }
2917
2918 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) {
2919 if (n == nullptr) return false;
2920 if (visited.member(n)) {
2921 return false; // Been there.
2922 }
2923 visited.push(n);
2924
2925 if (n->is_Allocate()) {
2926 // tty->print_cr("optimize barrier on alloc");
2927 return false;
2928 }
2929 if (n->is_Call()) {
2930 // tty->print_cr("optimize barrier on call");
2931 return false;
2932 }
2933
2934 const Type* type = phase->type(n);
2935 if (type == Type::TOP) {
2936 return false;
2937 }
2938 if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) {
2939 // tty->print_cr("optimize barrier on null");
2940 return false;
2941 }
2942 if (type->make_oopptr() && type->make_oopptr()->const_oop() != nullptr) {
2943 // tty->print_cr("optimize barrier on constant");
2944 return false;
2945 }
2946
2947 switch (n->Opcode()) {
2948 case Op_AddP:
2949 return true; // TODO: Can refine?
2950 case Op_LoadP:
2951 case Op_ShenandoahCompareAndExchangeN:
2952 case Op_ShenandoahCompareAndExchangeP:
2953 case Op_CompareAndExchangeN:
2954 case Op_CompareAndExchangeP:
2955 case Op_GetAndSetN:
2956 case Op_GetAndSetP:
2957 return true;
2958 case Op_Phi: {
2959 for (uint i = 1; i < n->req(); i++) {
2960 if (needs_barrier_impl(phase, n->in(i), visited)) return true;
2961 }
2962 return false;
2963 }
2964 case Op_CheckCastPP:
2965 case Op_CastPP:
2966 return needs_barrier_impl(phase, n->in(1), visited);
2967 case Op_Proj:
2968 return needs_barrier_impl(phase, n->in(0), visited);
2969 case Op_ShenandoahLoadReferenceBarrier:
2970 // tty->print_cr("optimize barrier on barrier");
2971 return false;
2972 case Op_Parm:
2973 // tty->print_cr("optimize barrier on input arg");
2974 return false;
2975 case Op_DecodeN:
2976 case Op_EncodeP:
2977 return needs_barrier_impl(phase, n->in(1), visited);
2978 case Op_LoadN:
2979 return true;
2980 case Op_CMoveN:
2981 case Op_CMoveP:
2982 return needs_barrier_impl(phase, n->in(2), visited) ||
2983 needs_barrier_impl(phase, n->in(3), visited);
2984 case Op_ShenandoahIUBarrier:
2985 return needs_barrier_impl(phase, n->in(1), visited);
2986 case Op_CreateEx:
2987 return false;
2988 default:
2989 break;
2990 }
2991 #ifdef ASSERT
2992 tty->print("need barrier on?: ");
2993 tty->print_cr("ins:");
2994 n->dump(2);
2995 tty->print_cr("outs:");
2996 n->dump(-2);
2997 ShouldNotReachHere();
2998 #endif
2999 return true;
3000 }