1 /*
2 * Copyright (c) 2015, 2019, Red Hat, Inc. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26
27 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
28 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
29 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
30 #include "gc/shenandoah/shenandoahForwarding.hpp"
31 #include "gc/shenandoah/shenandoahHeap.hpp"
32 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
33 #include "gc/shenandoah/shenandoahRuntime.hpp"
34 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
35 #include "opto/arraycopynode.hpp"
36 #include "opto/block.hpp"
37 #include "opto/callnode.hpp"
38 #include "opto/castnode.hpp"
39 #include "opto/movenode.hpp"
40 #include "opto/phaseX.hpp"
41 #include "opto/rootnode.hpp"
42 #include "opto/runtime.hpp"
43 #include "opto/subnode.hpp"
44
45 bool ShenandoahBarrierC2Support::expand(Compile* C, PhaseIterGVN& igvn) {
46 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
47 if ((state->enqueue_barriers_count() +
48 state->load_reference_barriers_count()) > 0) {
49 bool attempt_more_loopopts = ShenandoahLoopOptsAfterExpansion;
50 C->clear_major_progress();
51 PhaseIdealLoop ideal_loop(igvn, LoopOptsShenandoahExpand);
52 if (C->failing()) return false;
53 PhaseIdealLoop::verify(igvn);
54 DEBUG_ONLY(verify_raw_mem(C->root());)
55 if (attempt_more_loopopts) {
56 C->set_major_progress();
57 if (!C->optimize_loops(igvn, LoopOptsShenandoahPostExpand)) {
58 return false;
59 }
60 C->clear_major_progress();
61 if (C->range_check_cast_count() > 0) {
62 // No more loop optimizations. Remove all range check dependent CastIINodes.
63 C->remove_range_check_casts(igvn);
64 igvn.optimize();
65 }
66 }
67 }
68 return true;
69 }
70
71 bool ShenandoahBarrierC2Support::is_gc_state_test(Node* iff, int mask) {
72 if (!UseShenandoahGC) {
73 return false;
74 }
75 assert(iff->is_If(), "bad input");
76 if (iff->Opcode() != Op_If) {
77 return false;
78 }
79 Node* bol = iff->in(1);
80 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
81 return false;
82 }
83 Node* cmp = bol->in(1);
84 if (cmp->Opcode() != Op_CmpI) {
85 return false;
86 }
87 Node* in1 = cmp->in(1);
88 Node* in2 = cmp->in(2);
89 if (in2->find_int_con(-1) != 0) {
90 return false;
91 }
92 if (in1->Opcode() != Op_AndI) {
93 return false;
94 }
95 in2 = in1->in(2);
96 if (in2->find_int_con(-1) != mask) {
97 return false;
98 }
99 in1 = in1->in(1);
100
101 return is_gc_state_load(in1);
102 }
103
104 bool ShenandoahBarrierC2Support::is_heap_stable_test(Node* iff) {
105 return is_gc_state_test(iff, ShenandoahHeap::HAS_FORWARDED);
106 }
107
108 bool ShenandoahBarrierC2Support::is_gc_state_load(Node *n) {
109 if (!UseShenandoahGC) {
110 return false;
111 }
112 if (n->Opcode() != Op_LoadB && n->Opcode() != Op_LoadUB) {
113 return false;
114 }
115 Node* addp = n->in(MemNode::Address);
116 if (!addp->is_AddP()) {
117 return false;
118 }
119 Node* base = addp->in(AddPNode::Address);
120 Node* off = addp->in(AddPNode::Offset);
121 if (base->Opcode() != Op_ThreadLocal) {
122 return false;
123 }
124 if (off->find_intptr_t_con(-1) != in_bytes(ShenandoahThreadLocalData::gc_state_offset())) {
125 return false;
126 }
127 return true;
128 }
129
130 bool ShenandoahBarrierC2Support::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) {
131 assert(phase->is_dominator(stop, start), "bad inputs");
132 ResourceMark rm;
133 Unique_Node_List wq;
134 wq.push(start);
135 for (uint next = 0; next < wq.size(); next++) {
136 Node *m = wq.at(next);
137 if (m == stop) {
138 continue;
139 }
140 if (m->is_SafePoint() && !m->is_CallLeaf()) {
141 return true;
142 }
143 if (m->is_Region()) {
144 for (uint i = 1; i < m->req(); i++) {
145 wq.push(m->in(i));
146 }
147 } else {
148 wq.push(m->in(0));
149 }
150 }
151 return false;
152 }
153
154 #ifdef ASSERT
155 bool ShenandoahBarrierC2Support::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) {
156 assert(phis.size() == 0, "");
157
158 while (true) {
159 if (in->bottom_type() == TypePtr::NULL_PTR) {
160 if (trace) {tty->print_cr("NULL");}
161 } else if (!in->bottom_type()->make_ptr()->make_oopptr()) {
162 if (trace) {tty->print_cr("Non oop");}
163 } else {
164 if (in->is_ConstraintCast()) {
165 in = in->in(1);
166 continue;
167 } else if (in->is_AddP()) {
168 assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access");
169 in = in->in(AddPNode::Address);
170 continue;
171 } else if (in->is_Con()) {
172 if (trace) {
173 tty->print("Found constant");
174 in->dump();
175 }
176 } else if (in->Opcode() == Op_Parm) {
177 if (trace) {
178 tty->print("Found argument");
179 }
180 } else if (in->Opcode() == Op_CreateEx) {
181 if (trace) {
182 tty->print("Found create-exception");
183 }
184 } else if (in->Opcode() == Op_LoadP && in->adr_type() == TypeRawPtr::BOTTOM) {
185 if (trace) {
186 tty->print("Found raw LoadP (OSR argument?)");
187 }
188 } else if (in->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
189 if (t == ShenandoahOopStore) {
190 uint i = 0;
191 for (; i < phis.size(); i++) {
192 Node* n = phis.node_at(i);
193 if (n->Opcode() == Op_ShenandoahEnqueueBarrier) {
194 break;
195 }
196 }
197 if (i == phis.size()) {
198 return false;
199 }
200 }
201 barriers_used.push(in);
202 if (trace) {tty->print("Found barrier"); in->dump();}
203 } else if (in->Opcode() == Op_ShenandoahEnqueueBarrier) {
204 if (t != ShenandoahOopStore) {
205 in = in->in(1);
206 continue;
207 }
208 if (trace) {tty->print("Found enqueue barrier"); in->dump();}
209 phis.push(in, in->req());
210 in = in->in(1);
211 continue;
212 } else if (in->is_Proj() && in->in(0)->is_Allocate()) {
213 if (trace) {
214 tty->print("Found alloc");
215 in->in(0)->dump();
216 }
217 } else if (in->is_Proj() && (in->in(0)->Opcode() == Op_CallStaticJava || in->in(0)->Opcode() == Op_CallDynamicJava)) {
218 if (trace) {
219 tty->print("Found Java call");
220 }
221 } else if (in->is_Phi()) {
222 if (!visited.test_set(in->_idx)) {
223 if (trace) {tty->print("Pushed phi:"); in->dump();}
224 phis.push(in, 2);
225 in = in->in(1);
226 continue;
227 }
228 if (trace) {tty->print("Already seen phi:"); in->dump();}
229 } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) {
230 if (!visited.test_set(in->_idx)) {
231 if (trace) {tty->print("Pushed cmovep:"); in->dump();}
232 phis.push(in, CMoveNode::IfTrue);
233 in = in->in(CMoveNode::IfFalse);
234 continue;
235 }
236 if (trace) {tty->print("Already seen cmovep:"); in->dump();}
237 } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) {
238 in = in->in(1);
239 continue;
240 } else {
241 return false;
242 }
243 }
244 bool cont = false;
245 while (phis.is_nonempty()) {
246 uint idx = phis.index();
247 Node* phi = phis.node();
248 if (idx >= phi->req()) {
249 if (trace) {tty->print("Popped phi:"); phi->dump();}
250 phis.pop();
251 continue;
252 }
253 if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();}
254 in = phi->in(idx);
255 phis.set_index(idx+1);
256 cont = true;
257 break;
258 }
259 if (!cont) {
260 break;
261 }
262 }
263 return true;
264 }
265
266 void ShenandoahBarrierC2Support::report_verify_failure(const char* msg, Node* n1, Node* n2) {
267 if (n1 != NULL) {
268 n1->dump(+10);
269 }
270 if (n2 != NULL) {
271 n2->dump(+10);
272 }
273 fatal("%s", msg);
274 }
275
276 void ShenandoahBarrierC2Support::verify(RootNode* root) {
277 ResourceMark rm;
278 Unique_Node_List wq;
279 GrowableArray<Node*> barriers;
280 Unique_Node_List barriers_used;
281 Node_Stack phis(0);
282 VectorSet visited(Thread::current()->resource_area());
283 const bool trace = false;
284 const bool verify_no_useless_barrier = false;
285
286 wq.push(root);
287 for (uint next = 0; next < wq.size(); next++) {
288 Node *n = wq.at(next);
289 if (n->is_Load()) {
290 const bool trace = false;
291 if (trace) {tty->print("Verifying"); n->dump();}
292 if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) {
293 if (trace) {tty->print_cr("Load range/klass");}
294 } else {
295 const TypePtr* adr_type = n->as_Load()->adr_type();
296
297 if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) {
298 if (trace) {tty->print_cr("Mark load");}
299 } else if (adr_type->isa_instptr() &&
300 adr_type->is_instptr()->klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
301 adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset) {
302 if (trace) {tty->print_cr("Reference.get()");}
303 } else if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) {
304 report_verify_failure("Shenandoah verification: Load should have barriers", n);
305 }
306 }
307 } else if (n->is_Store()) {
308 const bool trace = false;
309
310 if (trace) {tty->print("Verifying"); n->dump();}
311 if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) {
312 Node* adr = n->in(MemNode::Address);
313 bool verify = true;
314
315 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) {
316 adr = adr->in(AddPNode::Address);
317 if (adr->is_AddP()) {
318 assert(adr->in(AddPNode::Base)->is_top(), "");
319 adr = adr->in(AddPNode::Address);
320 if (adr->Opcode() == Op_LoadP &&
321 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() &&
322 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
323 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset())) {
324 if (trace) {tty->print_cr("SATB prebarrier");}
325 verify = false;
326 }
327 }
328 }
329
330 if (verify && !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahStoreValEnqueueBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
331 report_verify_failure("Shenandoah verification: Store should have barriers", n);
332 }
333 }
334 if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
335 report_verify_failure("Shenandoah verification: Store (address) should have barriers", n);
336 }
337 } else if (n->Opcode() == Op_CmpP) {
338 const bool trace = false;
339
340 Node* in1 = n->in(1);
341 Node* in2 = n->in(2);
342 if (in1->bottom_type()->isa_oopptr()) {
343 if (trace) {tty->print("Verifying"); n->dump();}
344
345 bool mark_inputs = false;
346 if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR ||
347 (in1->is_Con() || in2->is_Con())) {
348 if (trace) {tty->print_cr("Comparison against a constant");}
349 mark_inputs = true;
350 } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) ||
351 (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) {
352 if (trace) {tty->print_cr("Comparison with newly alloc'ed object");}
353 mark_inputs = true;
354 } else {
355 assert(in2->bottom_type()->isa_oopptr(), "");
356
357 if (!verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) ||
358 !verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) {
359 report_verify_failure("Shenandoah verification: Cmp should have barriers", n);
360 }
361 }
362 if (verify_no_useless_barrier &&
363 mark_inputs &&
364 (!verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) ||
365 !verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) {
366 phis.clear();
367 visited.reset();
368 }
369 }
370 } else if (n->is_LoadStore()) {
371 if (n->in(MemNode::ValueIn)->bottom_type()->make_ptr() &&
372 !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahStoreValEnqueueBarrier ? ShenandoahOopStore : ShenandoahValue, trace, barriers_used)) {
373 report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n);
374 }
375
376 if (n->in(MemNode::Address)->bottom_type()->make_oopptr() && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
377 report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n);
378 }
379 } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) {
380 CallNode* call = n->as_Call();
381
382 static struct {
383 const char* name;
384 struct {
385 int pos;
386 verify_type t;
387 } args[6];
388 } calls[] = {
389 "aescrypt_encryptBlock",
390 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
391 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
392 "aescrypt_decryptBlock",
393 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
394 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
395 "multiplyToLen",
396 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { TypeFunc::Parms+4, ShenandoahStore },
397 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
398 "squareToLen",
399 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { -1, ShenandoahNone},
400 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
401 "montgomery_multiply",
402 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
403 { TypeFunc::Parms+6, ShenandoahStore }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
404 "montgomery_square",
405 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+5, ShenandoahStore },
406 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
407 "mulAdd",
408 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
409 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
410 "vectorizedMismatch",
411 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
412 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
413 "updateBytesCRC32",
414 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
415 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
416 "updateBytesAdler32",
417 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
418 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
419 "updateBytesCRC32C",
420 { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad}, { -1, ShenandoahNone},
421 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
422 "counterMode_AESCrypt",
423 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
424 { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } },
425 "cipherBlockChaining_encryptAESCrypt",
426 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
427 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
428 "cipherBlockChaining_decryptAESCrypt",
429 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
430 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
431 "shenandoah_clone_barrier",
432 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
433 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
434 "ghash_processBlocks",
435 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
436 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
437 "sha1_implCompress",
438 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
439 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
440 "sha256_implCompress",
441 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
442 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
443 "sha512_implCompress",
444 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
445 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
446 "sha1_implCompressMB",
447 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
448 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
449 "sha256_implCompressMB",
450 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
451 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
452 "sha512_implCompressMB",
453 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
454 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
455 "encodeBlock",
456 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahStore }, { -1, ShenandoahNone },
457 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
458 };
459
460 if (call->is_call_to_arraycopystub()) {
461 Node* dest = NULL;
462 const TypeTuple* args = n->as_Call()->_tf->domain();
463 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
464 if (args->field_at(i)->isa_ptr()) {
465 j++;
466 if (j == 2) {
467 dest = n->in(i);
468 break;
469 }
470 }
471 }
472 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) ||
473 !verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) {
474 report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n);
475 }
476 } else if (strlen(call->_name) > 5 &&
477 !strcmp(call->_name + strlen(call->_name) - 5, "_fill")) {
478 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) {
479 report_verify_failure("Shenandoah verification: _fill should have barriers", n);
480 }
481 } else if (!strcmp(call->_name, "shenandoah_wb_pre")) {
482 // skip
483 } else {
484 const int calls_len = sizeof(calls) / sizeof(calls[0]);
485 int i = 0;
486 for (; i < calls_len; i++) {
487 if (!strcmp(calls[i].name, call->_name)) {
488 break;
489 }
490 }
491 if (i != calls_len) {
492 const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]);
493 for (uint j = 0; j < args_len; j++) {
494 int pos = calls[i].args[j].pos;
495 if (pos == -1) {
496 break;
497 }
498 if (!verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) {
499 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
500 }
501 }
502 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
503 if (call->in(j)->bottom_type()->make_ptr() &&
504 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
505 uint k = 0;
506 for (; k < args_len && calls[i].args[k].pos != (int)j; k++);
507 if (k == args_len) {
508 fatal("arg %d for call %s not covered", j, call->_name);
509 }
510 }
511 }
512 } else {
513 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
514 if (call->in(j)->bottom_type()->make_ptr() &&
515 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
516 fatal("%s not covered", call->_name);
517 }
518 }
519 }
520 }
521 } else if (n->Opcode() == Op_ShenandoahEnqueueBarrier || n->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
522 // skip
523 } else if (n->is_AddP()
524 || n->is_Phi()
525 || n->is_ConstraintCast()
526 || n->Opcode() == Op_Return
527 || n->Opcode() == Op_CMoveP
528 || n->Opcode() == Op_CMoveN
529 || n->Opcode() == Op_Rethrow
530 || n->is_MemBar()
531 || n->Opcode() == Op_Conv2B
532 || n->Opcode() == Op_SafePoint
533 || n->is_CallJava()
534 || n->Opcode() == Op_Unlock
535 || n->Opcode() == Op_EncodeP
536 || n->Opcode() == Op_DecodeN) {
537 // nothing to do
538 } else {
539 static struct {
540 int opcode;
541 struct {
542 int pos;
543 verify_type t;
544 } inputs[2];
545 } others[] = {
546 Op_FastLock,
547 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
548 Op_Lock,
549 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone} },
550 Op_ArrayCopy,
551 { { ArrayCopyNode::Src, ShenandoahLoad }, { ArrayCopyNode::Dest, ShenandoahStore } },
552 Op_StrCompressedCopy,
553 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
554 Op_StrInflatedCopy,
555 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
556 Op_AryEq,
557 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
558 Op_StrIndexOf,
559 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
560 Op_StrComp,
561 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
562 Op_StrEquals,
563 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
564 Op_EncodeISOArray,
565 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
566 Op_HasNegatives,
567 { { 2, ShenandoahLoad }, { -1, ShenandoahNone} },
568 Op_CastP2X,
569 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
570 Op_StrIndexOfChar,
571 { { 2, ShenandoahLoad }, { -1, ShenandoahNone } },
572 };
573
574 const int others_len = sizeof(others) / sizeof(others[0]);
575 int i = 0;
576 for (; i < others_len; i++) {
577 if (others[i].opcode == n->Opcode()) {
578 break;
579 }
580 }
581 uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req();
582 if (i != others_len) {
583 const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]);
584 for (uint j = 0; j < inputs_len; j++) {
585 int pos = others[i].inputs[j].pos;
586 if (pos == -1) {
587 break;
588 }
589 if (!verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) {
590 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
591 }
592 }
593 for (uint j = 1; j < stop; j++) {
594 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
595 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
596 uint k = 0;
597 for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++);
598 if (k == inputs_len) {
599 fatal("arg %d for node %s not covered", j, n->Name());
600 }
601 }
602 }
603 } else {
604 for (uint j = 1; j < stop; j++) {
605 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
606 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
607 fatal("%s not covered", n->Name());
608 }
609 }
610 }
611 }
612
613 if (n->is_SafePoint()) {
614 SafePointNode* sfpt = n->as_SafePoint();
615 if (verify_no_useless_barrier && sfpt->jvms() != NULL) {
616 for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) {
617 if (!verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) {
618 phis.clear();
619 visited.reset();
620 }
621 }
622 }
623 }
624 }
625
626 if (verify_no_useless_barrier) {
627 for (int i = 0; i < barriers.length(); i++) {
628 Node* n = barriers.at(i);
629 if (!barriers_used.member(n)) {
630 tty->print("XXX useless barrier"); n->dump(-2);
631 ShouldNotReachHere();
632 }
633 }
634 }
635 }
636 #endif
637
638 bool ShenandoahBarrierC2Support::is_dominator_same_ctrl(Node* c, Node* d, Node* n, PhaseIdealLoop* phase) {
639 // That both nodes have the same control is not sufficient to prove
640 // domination, verify that there's no path from d to n
641 ResourceMark rm;
642 Unique_Node_List wq;
643 wq.push(d);
644 for (uint next = 0; next < wq.size(); next++) {
645 Node *m = wq.at(next);
646 if (m == n) {
647 return false;
648 }
649 if (m->is_Phi() && m->in(0)->is_Loop()) {
650 assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control");
651 } else {
652 if (m->is_Store() || m->is_LoadStore()) {
653 // Take anti-dependencies into account
654 Node* mem = m->in(MemNode::Memory);
655 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
656 Node* u = mem->fast_out(i);
657 if (u->is_Load() && phase->C->can_alias(m->adr_type(), phase->C->get_alias_index(u->adr_type())) &&
658 phase->ctrl_or_self(u) == c) {
659 wq.push(u);
660 }
661 }
662 }
663 for (uint i = 0; i < m->req(); i++) {
664 if (m->in(i) != NULL && phase->ctrl_or_self(m->in(i)) == c) {
665 wq.push(m->in(i));
666 }
667 }
668 }
669 }
670 return true;
671 }
672
673 bool ShenandoahBarrierC2Support::is_dominator(Node* d_c, Node* n_c, Node* d, Node* n, PhaseIdealLoop* phase) {
674 if (d_c != n_c) {
675 return phase->is_dominator(d_c, n_c);
676 }
677 return is_dominator_same_ctrl(d_c, d, n, phase);
678 }
679
680 Node* next_mem(Node* mem, int alias) {
681 Node* res = NULL;
682 if (mem->is_Proj()) {
683 res = mem->in(0);
684 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
685 res = mem->in(TypeFunc::Memory);
686 } else if (mem->is_Phi()) {
687 res = mem->in(1);
688 } else if (mem->is_MergeMem()) {
689 res = mem->as_MergeMem()->memory_at(alias);
690 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
691 assert(alias = Compile::AliasIdxRaw, "following raw memory can't lead to a barrier");
692 res = mem->in(MemNode::Memory);
693 } else {
694 #ifdef ASSERT
695 mem->dump();
696 #endif
697 ShouldNotReachHere();
698 }
699 return res;
700 }
701
702 Node* ShenandoahBarrierC2Support::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) {
703 Node* iffproj = NULL;
704 while (c != dom) {
705 Node* next = phase->idom(c);
706 assert(next->unique_ctrl_out() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?");
707 if (c->is_Region()) {
708 ResourceMark rm;
709 Unique_Node_List wq;
710 wq.push(c);
711 for (uint i = 0; i < wq.size(); i++) {
712 Node *n = wq.at(i);
713 if (n == next) {
714 continue;
715 }
716 if (n->is_Region()) {
717 for (uint j = 1; j < n->req(); j++) {
718 wq.push(n->in(j));
719 }
720 } else {
721 wq.push(n->in(0));
722 }
723 }
724 for (uint i = 0; i < wq.size(); i++) {
725 Node *n = wq.at(i);
726 assert(n->is_CFG(), "");
727 if (n->is_Multi()) {
728 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
729 Node* u = n->fast_out(j);
730 if (u->is_CFG()) {
731 if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
732 return NodeSentinel;
733 }
734 }
735 }
736 }
737 }
738 } else if (c->is_Proj()) {
739 if (c->is_IfProj()) {
740 if (c->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) != NULL) {
741 // continue;
742 } else {
743 if (!allow_one_proj) {
744 return NodeSentinel;
745 }
746 if (iffproj == NULL) {
747 iffproj = c;
748 } else {
749 return NodeSentinel;
750 }
751 }
752 } else if (c->Opcode() == Op_JumpProj) {
753 return NodeSentinel; // unsupported
754 } else if (c->Opcode() == Op_CatchProj) {
755 return NodeSentinel; // unsupported
756 } else if (c->Opcode() == Op_CProj && next->Opcode() == Op_NeverBranch) {
757 return NodeSentinel; // unsupported
758 } else {
759 assert(next->unique_ctrl_out() == c, "unsupported branch pattern");
760 }
761 }
762 c = next;
763 }
764 return iffproj;
765 }
766
767 Node* ShenandoahBarrierC2Support::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) {
768 ResourceMark rm;
769 VectorSet wq(Thread::current()->resource_area());
770 wq.set(mem->_idx);
771 mem_ctrl = phase->ctrl_or_self(mem);
772 while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) {
773 mem = next_mem(mem, alias);
774 if (wq.test_set(mem->_idx)) {
775 return NULL;
776 }
777 mem_ctrl = phase->ctrl_or_self(mem);
778 }
779 if (mem->is_MergeMem()) {
780 mem = mem->as_MergeMem()->memory_at(alias);
781 mem_ctrl = phase->ctrl_or_self(mem);
782 }
783 return mem;
784 }
785
786 Node* ShenandoahBarrierC2Support::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) {
787 Node* mem = NULL;
788 Node* c = ctrl;
789 do {
790 if (c->is_Region()) {
791 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == NULL; i++) {
792 Node* u = c->fast_out(i);
793 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
794 if (u->adr_type() == TypePtr::BOTTOM) {
795 mem = u;
796 }
797 }
798 }
799 } else {
800 if (c->is_Call() && c->as_Call()->adr_type() != NULL) {
801 CallProjections projs;
802 c->as_Call()->extract_projections(&projs, true, false);
803 if (projs.fallthrough_memproj != NULL) {
804 if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
805 if (projs.catchall_memproj == NULL) {
806 mem = projs.fallthrough_memproj;
807 } else {
808 if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) {
809 mem = projs.fallthrough_memproj;
810 } else {
811 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier");
812 mem = projs.catchall_memproj;
813 }
814 }
815 }
816 } else {
817 Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
818 if (proj != NULL &&
819 proj->adr_type() == TypePtr::BOTTOM) {
820 mem = proj;
821 }
822 }
823 } else {
824 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
825 Node* u = c->fast_out(i);
826 if (u->is_Proj() &&
827 u->bottom_type() == Type::MEMORY &&
828 u->adr_type() == TypePtr::BOTTOM) {
829 assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
830 assert(mem == NULL, "only one proj");
831 mem = u;
832 }
833 }
834 assert(!c->is_Call() || c->as_Call()->adr_type() != NULL || mem == NULL, "no mem projection expected");
835 }
836 }
837 c = phase->idom(c);
838 } while (mem == NULL);
839 return mem;
840 }
841
842 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
843 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
844 Node* u = n->fast_out(i);
845 if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
846 uses.push(u);
847 }
848 }
849 }
850
851 static void hide_strip_mined_loop(OuterStripMinedLoopNode* outer, CountedLoopNode* inner, PhaseIdealLoop* phase) {
852 OuterStripMinedLoopEndNode* le = inner->outer_loop_end();
853 Node* new_outer = new LoopNode(outer->in(LoopNode::EntryControl), outer->in(LoopNode::LoopBackControl));
854 phase->register_control(new_outer, phase->get_loop(outer), outer->in(LoopNode::EntryControl));
855 Node* new_le = new IfNode(le->in(0), le->in(1), le->_prob, le->_fcnt);
856 phase->register_control(new_le, phase->get_loop(le), le->in(0));
857 phase->lazy_replace(outer, new_outer);
858 phase->lazy_replace(le, new_le);
859 inner->clear_strip_mined();
860 }
861
862 void ShenandoahBarrierC2Support::test_gc_state(Node*& ctrl, Node* raw_mem, Node*& test_fail_ctrl,
863 PhaseIdealLoop* phase, int flags) {
864 PhaseIterGVN& igvn = phase->igvn();
865 Node* old_ctrl = ctrl;
866
867 Node* thread = new ThreadLocalNode();
868 Node* gc_state_offset = igvn.MakeConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
869 Node* gc_state_addr = new AddPNode(phase->C->top(), thread, gc_state_offset);
870 Node* gc_state = new LoadBNode(old_ctrl, raw_mem, gc_state_addr,
871 DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(NULL),
872 TypeInt::BYTE, MemNode::unordered);
873 Node* gc_state_and = new AndINode(gc_state, igvn.intcon(flags));
874 Node* gc_state_cmp = new CmpINode(gc_state_and, igvn.zerocon(T_INT));
875 Node* gc_state_bool = new BoolNode(gc_state_cmp, BoolTest::ne);
876
877 IfNode* gc_state_iff = new IfNode(old_ctrl, gc_state_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
878 ctrl = new IfTrueNode(gc_state_iff);
879 test_fail_ctrl = new IfFalseNode(gc_state_iff);
880
881 IdealLoopTree* loop = phase->get_loop(old_ctrl);
882 phase->register_control(gc_state_iff, loop, old_ctrl);
883 phase->register_control(ctrl, loop, gc_state_iff);
884 phase->register_control(test_fail_ctrl, loop, gc_state_iff);
885
886 phase->register_new_node(thread, old_ctrl);
887 phase->register_new_node(gc_state_addr, old_ctrl);
888 phase->register_new_node(gc_state, old_ctrl);
889 phase->register_new_node(gc_state_and, old_ctrl);
890 phase->register_new_node(gc_state_cmp, old_ctrl);
891 phase->register_new_node(gc_state_bool, old_ctrl);
892
893 phase->set_ctrl(gc_state_offset, phase->C->root());
894
895 assert(is_gc_state_test(gc_state_iff, flags), "Should match the shape");
896 }
897
898 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) {
899 Node* old_ctrl = ctrl;
900 PhaseIterGVN& igvn = phase->igvn();
901
902 const Type* val_t = igvn.type(val);
903 if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
904 Node* null_cmp = new CmpPNode(val, igvn.zerocon(T_OBJECT));
905 Node* null_test = new BoolNode(null_cmp, BoolTest::ne);
906
907 IfNode* null_iff = new IfNode(old_ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
908 ctrl = new IfTrueNode(null_iff);
909 null_ctrl = new IfFalseNode(null_iff);
910
911 IdealLoopTree* loop = phase->get_loop(old_ctrl);
912 phase->register_control(null_iff, loop, old_ctrl);
913 phase->register_control(ctrl, loop, null_iff);
914 phase->register_control(null_ctrl, loop, null_iff);
915
916 phase->register_new_node(null_cmp, old_ctrl);
917 phase->register_new_node(null_test, old_ctrl);
918 }
919 }
920
921 void ShenandoahBarrierC2Support::test_in_cset(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) {
922 Node* old_ctrl = ctrl;
923 PhaseIterGVN& igvn = phase->igvn();
924
925 Node* raw_val = new CastP2XNode(old_ctrl, val);
926 Node* cset_idx = new URShiftXNode(raw_val, igvn.intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
927
928 // Figure out the target cset address with raw pointer math.
929 // This avoids matching AddP+LoadB that would emit inefficient code.
930 // See JDK-8245465.
931 Node* cset_addr_ptr = igvn.makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
932 Node* cset_addr = new CastP2XNode(old_ctrl, cset_addr_ptr);
933 Node* cset_load_addr = new AddXNode(cset_addr, cset_idx);
934 Node* cset_load_ptr = new CastX2PNode(cset_load_addr);
935
936 Node* cset_load = new LoadBNode(old_ctrl, raw_mem, cset_load_ptr,
937 DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(NULL),
938 TypeInt::BYTE, MemNode::unordered);
939 Node* cset_cmp = new CmpINode(cset_load, igvn.zerocon(T_INT));
940 Node* cset_bool = new BoolNode(cset_cmp, BoolTest::ne);
941
942 IfNode* cset_iff = new IfNode(old_ctrl, cset_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
943 ctrl = new IfTrueNode(cset_iff);
944 not_cset_ctrl = new IfFalseNode(cset_iff);
945
946 IdealLoopTree *loop = phase->get_loop(old_ctrl);
947 phase->register_control(cset_iff, loop, old_ctrl);
948 phase->register_control(ctrl, loop, cset_iff);
949 phase->register_control(not_cset_ctrl, loop, cset_iff);
950
951 phase->set_ctrl(cset_addr_ptr, phase->C->root());
952
953 phase->register_new_node(raw_val, old_ctrl);
954 phase->register_new_node(cset_idx, old_ctrl);
955 phase->register_new_node(cset_addr, old_ctrl);
956 phase->register_new_node(cset_load_addr, old_ctrl);
957 phase->register_new_node(cset_load_ptr, old_ctrl);
958 phase->register_new_node(cset_load, old_ctrl);
959 phase->register_new_node(cset_cmp, old_ctrl);
960 phase->register_new_node(cset_bool, old_ctrl);
961 }
962
963 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node* load_addr, Node*& result_mem, Node* raw_mem, bool is_native, PhaseIdealLoop* phase) {
964 IdealLoopTree*loop = phase->get_loop(ctrl);
965 const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr();
966
967 // The slow path stub consumes and produces raw memory in addition
968 // to the existing memory edges
969 Node* base = find_bottom_mem(ctrl, phase);
970 MergeMemNode* mm = MergeMemNode::make(base);
971 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
972 phase->register_new_node(mm, ctrl);
973
974 address target = LP64_ONLY(UseCompressedOops) NOT_LP64(false) ?
975 CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_narrow) :
976 CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier);
977
978 address calladdr = is_native ? CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_native)
979 : target;
980 const char* name = is_native ? "load_reference_barrier_native" : "load_reference_barrier";
981 Node* call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type(), calladdr, name, TypeRawPtr::BOTTOM);
982
983 call->init_req(TypeFunc::Control, ctrl);
984 call->init_req(TypeFunc::I_O, phase->C->top());
985 call->init_req(TypeFunc::Memory, mm);
986 call->init_req(TypeFunc::FramePtr, phase->C->top());
987 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
988 call->init_req(TypeFunc::Parms, val);
989 call->init_req(TypeFunc::Parms+1, load_addr);
990 phase->register_control(call, loop, ctrl);
991 ctrl = new ProjNode(call, TypeFunc::Control);
992 phase->register_control(ctrl, loop, call);
993 result_mem = new ProjNode(call, TypeFunc::Memory);
994 phase->register_new_node(result_mem, call);
995 val = new ProjNode(call, TypeFunc::Parms);
996 phase->register_new_node(val, call);
997 val = new CheckCastPPNode(ctrl, val, obj_type);
998 phase->register_new_node(val, ctrl);
999 }
1000
1001 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) {
1002 Node* ctrl = phase->get_ctrl(barrier);
1003 Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
1004
1005 // Update the control of all nodes that should be after the
1006 // barrier control flow
1007 uses.clear();
1008 // Every node that is control dependent on the barrier's input
1009 // control will be after the expanded barrier. The raw memory (if
1010 // its memory is control dependent on the barrier's input control)
1011 // must stay above the barrier.
1012 uses_to_ignore.clear();
1013 if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
1014 uses_to_ignore.push(init_raw_mem);
1015 }
1016 for (uint next = 0; next < uses_to_ignore.size(); next++) {
1017 Node *n = uses_to_ignore.at(next);
1018 for (uint i = 0; i < n->req(); i++) {
1019 Node* in = n->in(i);
1020 if (in != NULL && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
1021 uses_to_ignore.push(in);
1022 }
1023 }
1024 }
1025 for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
1026 Node* u = ctrl->fast_out(i);
1027 if (u->_idx < last &&
1028 u != barrier &&
1029 !uses_to_ignore.member(u) &&
1030 (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) &&
1031 (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) {
1032 Node* old_c = phase->ctrl_or_self(u);
1033 Node* c = old_c;
1034 if (c != ctrl ||
1035 is_dominator_same_ctrl(old_c, barrier, u, phase) ||
1036 ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) {
1037 phase->igvn().rehash_node_delayed(u);
1038 int nb = u->replace_edge(ctrl, region);
1039 if (u->is_CFG()) {
1040 if (phase->idom(u) == ctrl) {
1041 phase->set_idom(u, region, phase->dom_depth(region));
1042 }
1043 } else if (phase->get_ctrl(u) == ctrl) {
1044 assert(u != init_raw_mem, "should leave input raw mem above the barrier");
1045 uses.push(u);
1046 }
1047 assert(nb == 1, "more than 1 ctrl input?");
1048 --i, imax -= nb;
1049 }
1050 }
1051 }
1052 }
1053
1054 static Node* create_phis_on_call_return(Node* ctrl, Node* c, Node* n, Node* n_clone, const CallProjections& projs, PhaseIdealLoop* phase) {
1055 Node* region = NULL;
1056 while (c != ctrl) {
1057 if (c->is_Region()) {
1058 region = c;
1059 }
1060 c = phase->idom(c);
1061 }
1062 assert(region != NULL, "");
1063 Node* phi = new PhiNode(region, n->bottom_type());
1064 for (uint j = 1; j < region->req(); j++) {
1065 Node* in = region->in(j);
1066 if (phase->is_dominator(projs.fallthrough_catchproj, in)) {
1067 phi->init_req(j, n);
1068 } else if (phase->is_dominator(projs.catchall_catchproj, in)) {
1069 phi->init_req(j, n_clone);
1070 } else {
1071 phi->init_req(j, create_phis_on_call_return(ctrl, in, n, n_clone, projs, phase));
1072 }
1073 }
1074 phase->register_new_node(phi, region);
1075 return phi;
1076 }
1077
1078 void ShenandoahBarrierC2Support::pin_and_expand(PhaseIdealLoop* phase) {
1079 ShenandoahBarrierSetC2State* state = ShenandoahBarrierSetC2::bsc2()->state();
1080
1081 Unique_Node_List uses;
1082 for (int i = 0; i < state->enqueue_barriers_count(); i++) {
1083 Node* barrier = state->enqueue_barrier(i);
1084 Node* ctrl = phase->get_ctrl(barrier);
1085 IdealLoopTree* loop = phase->get_loop(ctrl);
1086 if (loop->_head->is_OuterStripMinedLoop()) {
1087 // Expanding a barrier here will break loop strip mining
1088 // verification. Transform the loop so the loop nest doesn't
1089 // appear as strip mined.
1090 OuterStripMinedLoopNode* outer = loop->_head->as_OuterStripMinedLoop();
1091 hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1092 }
1093 }
1094
1095 Node_Stack stack(0);
1096 Node_List clones;
1097 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1098 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1099 if (lrb->is_redundant()) {
1100 continue;
1101 }
1102
1103 Node* ctrl = phase->get_ctrl(lrb);
1104 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1105
1106 CallStaticJavaNode* unc = NULL;
1107 Node* unc_ctrl = NULL;
1108 Node* uncasted_val = val;
1109
1110 for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1111 Node* u = lrb->fast_out(i);
1112 if (u->Opcode() == Op_CastPP &&
1113 u->in(0) != NULL &&
1114 phase->is_dominator(u->in(0), ctrl)) {
1115 const Type* u_t = phase->igvn().type(u);
1116
1117 if (u_t->meet(TypePtr::NULL_PTR) != u_t &&
1118 u->in(0)->Opcode() == Op_IfTrue &&
1119 u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
1120 u->in(0)->in(0)->is_If() &&
1121 u->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
1122 u->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
1123 u->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
1124 u->in(0)->in(0)->in(1)->in(1)->in(1) == val &&
1125 u->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
1126 IdealLoopTree* loop = phase->get_loop(ctrl);
1127 IdealLoopTree* unc_loop = phase->get_loop(u->in(0));
1128
1129 if (!unc_loop->is_member(loop)) {
1130 continue;
1131 }
1132
1133 Node* branch = no_branches(ctrl, u->in(0), false, phase);
1134 assert(branch == NULL || branch == NodeSentinel, "was not looking for a branch");
1135 if (branch == NodeSentinel) {
1136 continue;
1137 }
1138
1139 Node* iff = u->in(0)->in(0);
1140 Node* bol = iff->in(1)->clone();
1141 Node* cmp = bol->in(1)->clone();
1142 cmp->set_req(1, lrb);
1143 bol->set_req(1, cmp);
1144 phase->igvn().replace_input_of(iff, 1, bol);
1145 phase->set_ctrl(lrb, iff->in(0));
1146 phase->register_new_node(cmp, iff->in(0));
1147 phase->register_new_node(bol, iff->in(0));
1148 break;
1149 }
1150 }
1151 }
1152 if ((ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) || ctrl->is_CallJava()) {
1153 CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_CallJava() : ctrl->as_CallJava();
1154 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
1155 // The rethrow call may have too many projections to be
1156 // properly handled here. Given there's no reason for a
1157 // barrier to depend on the call, move it above the call
1158 stack.push(lrb, 0);
1159 do {
1160 Node* n = stack.node();
1161 uint idx = stack.index();
1162 if (idx < n->req()) {
1163 Node* in = n->in(idx);
1164 stack.set_index(idx+1);
1165 if (in != NULL) {
1166 if (phase->has_ctrl(in)) {
1167 if (phase->is_dominator(call, phase->get_ctrl(in))) {
1168 #ifdef ASSERT
1169 for (uint i = 0; i < stack.size(); i++) {
1170 assert(stack.node_at(i) != in, "node shouldn't have been seen yet");
1171 }
1172 #endif
1173 stack.push(in, 0);
1174 }
1175 } else {
1176 assert(phase->is_dominator(in, call->in(0)), "no dependency on the call");
1177 }
1178 }
1179 } else {
1180 phase->set_ctrl(n, call->in(0));
1181 stack.pop();
1182 }
1183 } while(stack.size() > 0);
1184 continue;
1185 }
1186 CallProjections projs;
1187 call->extract_projections(&projs, false, false);
1188
1189 #ifdef ASSERT
1190 VectorSet cloned(Thread::current()->resource_area());
1191 #endif
1192 Node* lrb_clone = lrb->clone();
1193 phase->register_new_node(lrb_clone, projs.catchall_catchproj);
1194 phase->set_ctrl(lrb, projs.fallthrough_catchproj);
1195
1196 stack.push(lrb, 0);
1197 clones.push(lrb_clone);
1198
1199 do {
1200 assert(stack.size() == clones.size(), "");
1201 Node* n = stack.node();
1202 #ifdef ASSERT
1203 if (n->is_Load()) {
1204 Node* mem = n->in(MemNode::Memory);
1205 for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
1206 Node* u = mem->fast_out(j);
1207 assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?");
1208 }
1209 }
1210 #endif
1211 uint idx = stack.index();
1212 Node* n_clone = clones.at(clones.size()-1);
1213 if (idx < n->outcnt()) {
1214 Node* u = n->raw_out(idx);
1215 Node* c = phase->ctrl_or_self(u);
1216 if (phase->is_dominator(call, c) && phase->is_dominator(c, projs.fallthrough_proj)) {
1217 stack.set_index(idx+1);
1218 assert(!u->is_CFG(), "");
1219 stack.push(u, 0);
1220 assert(!cloned.test_set(u->_idx), "only one clone");
1221 Node* u_clone = u->clone();
1222 int nb = u_clone->replace_edge(n, n_clone);
1223 assert(nb > 0, "should have replaced some uses");
1224 phase->register_new_node(u_clone, projs.catchall_catchproj);
1225 clones.push(u_clone);
1226 phase->set_ctrl(u, projs.fallthrough_catchproj);
1227 } else {
1228 bool replaced = false;
1229 if (u->is_Phi()) {
1230 for (uint k = 1; k < u->req(); k++) {
1231 if (u->in(k) == n) {
1232 if (phase->is_dominator(projs.catchall_catchproj, u->in(0)->in(k))) {
1233 phase->igvn().replace_input_of(u, k, n_clone);
1234 replaced = true;
1235 } else if (!phase->is_dominator(projs.fallthrough_catchproj, u->in(0)->in(k))) {
1236 phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase));
1237 replaced = true;
1238 }
1239 }
1240 }
1241 } else {
1242 if (phase->is_dominator(projs.catchall_catchproj, c)) {
1243 phase->igvn().rehash_node_delayed(u);
1244 int nb = u->replace_edge(n, n_clone);
1245 assert(nb > 0, "should have replaced some uses");
1246 replaced = true;
1247 } else if (!phase->is_dominator(projs.fallthrough_catchproj, c)) {
1248 if (u->is_If()) {
1249 // Can't break If/Bool/Cmp chain
1250 assert(n->is_Bool(), "unexpected If shape");
1251 assert(stack.node_at(stack.size()-2)->is_Cmp(), "unexpected If shape");
1252 assert(n_clone->is_Bool(), "unexpected clone");
1253 assert(clones.at(clones.size()-2)->is_Cmp(), "unexpected clone");
1254 Node* bol_clone = n->clone();
1255 Node* cmp_clone = stack.node_at(stack.size()-2)->clone();
1256 bol_clone->set_req(1, cmp_clone);
1257
1258 Node* nn = stack.node_at(stack.size()-3);
1259 Node* nn_clone = clones.at(clones.size()-3);
1260 assert(nn->Opcode() == nn_clone->Opcode(), "mismatch");
1261
1262 int nb = cmp_clone->replace_edge(nn, create_phis_on_call_return(ctrl, c, nn, nn_clone, projs, phase));
1263 assert(nb > 0, "should have replaced some uses");
1264
1265 phase->register_new_node(bol_clone, u->in(0));
1266 phase->register_new_node(cmp_clone, u->in(0));
1267
1268 phase->igvn().replace_input_of(u, 1, bol_clone);
1269
1270 } else {
1271 phase->igvn().rehash_node_delayed(u);
1272 int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase));
1273 assert(nb > 0, "should have replaced some uses");
1274 }
1275 replaced = true;
1276 }
1277 }
1278 if (!replaced) {
1279 stack.set_index(idx+1);
1280 }
1281 }
1282 } else {
1283 stack.pop();
1284 clones.pop();
1285 }
1286 } while (stack.size() > 0);
1287 assert(stack.size() == 0 && clones.size() == 0, "");
1288 }
1289 }
1290
1291 for (int i = 0; i < state->load_reference_barriers_count(); i++) {
1292 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1293 if (lrb->is_redundant()) {
1294 continue;
1295 }
1296 Node* ctrl = phase->get_ctrl(lrb);
1297 IdealLoopTree* loop = phase->get_loop(ctrl);
1298 if (loop->_head->is_OuterStripMinedLoop()) {
1299 // Expanding a barrier here will break loop strip mining
1300 // verification. Transform the loop so the loop nest doesn't
1301 // appear as strip mined.
1302 OuterStripMinedLoopNode* outer = loop->_head->as_OuterStripMinedLoop();
1303 hide_strip_mined_loop(outer, outer->unique_ctrl_out()->as_CountedLoop(), phase);
1304 }
1305 }
1306
1307 // Expand load-reference-barriers
1308 MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
1309 Unique_Node_List uses_to_ignore;
1310 for (int i = state->load_reference_barriers_count() - 1; i >= 0; i--) {
1311 ShenandoahLoadReferenceBarrierNode* lrb = state->load_reference_barrier(i);
1312 if (lrb->is_redundant()) {
1313 phase->igvn().replace_node(lrb, lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1314 continue;
1315 }
1316 uint last = phase->C->unique();
1317 Node* ctrl = phase->get_ctrl(lrb);
1318 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1319
1320
1321 Node* orig_ctrl = ctrl;
1322
1323 Node* raw_mem = fixer.find_mem(ctrl, lrb);
1324 Node* init_raw_mem = raw_mem;
1325 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL);
1326
1327 IdealLoopTree *loop = phase->get_loop(ctrl);
1328
1329 Node* heap_stable_ctrl = NULL;
1330 Node* null_ctrl = NULL;
1331
1332 assert(val->bottom_type()->make_oopptr(), "need oop");
1333 assert(val->bottom_type()->make_oopptr()->const_oop() == NULL, "expect non-constant");
1334
1335 enum { _heap_stable = 1, _not_cset, _evac_path, PATH_LIMIT };
1336 Node* region = new RegionNode(PATH_LIMIT);
1337 Node* val_phi = new PhiNode(region, val->bottom_type()->is_oopptr());
1338 Node* raw_mem_phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1339
1340 // Stable path.
1341 test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, ShenandoahHeap::HAS_FORWARDED);
1342 IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
1343
1344 // Heap stable case
1345 region->init_req(_heap_stable, heap_stable_ctrl);
1346 val_phi->init_req(_heap_stable, val);
1347 raw_mem_phi->init_req(_heap_stable, raw_mem);
1348
1349 // Test for in-cset.
1350 // Wires !in_cset(obj) to slot 2 of region and phis
1351 Node* not_cset_ctrl = NULL;
1352 test_in_cset(ctrl, not_cset_ctrl, val, raw_mem, phase);
1353 if (not_cset_ctrl != NULL) {
1354 region->init_req(_not_cset, not_cset_ctrl);
1355 val_phi->init_req(_not_cset, val);
1356 raw_mem_phi->init_req(_not_cset, raw_mem);
1357 }
1358
1359 // Resolve object when orig-value is in cset.
1360 // Make the unconditional resolve for fwdptr.
1361
1362 // Call lrb-stub and wire up that path in slots 4
1363 Node* result_mem = NULL;
1364
1365 Node* addr;
1366 if (ShenandoahSelfFixing) {
1367 VectorSet visited(Thread::current()->resource_area());
1368 addr = get_load_addr(phase, visited, lrb);
1369 } else {
1370 addr = phase->igvn().zerocon(T_OBJECT);
1371 }
1372 if (addr->Opcode() == Op_AddP) {
1373 Node* orig_base = addr->in(AddPNode::Base);
1374 Node* base = new CheckCastPPNode(ctrl, orig_base, orig_base->bottom_type(), true);
1375 phase->register_new_node(base, ctrl);
1376 if (addr->in(AddPNode::Base) == addr->in((AddPNode::Address))) {
1377 // Field access
1378 addr = addr->clone();
1379 addr->set_req(AddPNode::Base, base);
1380 addr->set_req(AddPNode::Address, base);
1381 phase->register_new_node(addr, ctrl);
1382 } else {
1383 Node* addr2 = addr->in(AddPNode::Address);
1384 if (addr2->Opcode() == Op_AddP && addr2->in(AddPNode::Base) == addr2->in(AddPNode::Address) &&
1385 addr2->in(AddPNode::Base) == orig_base) {
1386 addr2 = addr2->clone();
1387 addr2->set_req(AddPNode::Base, base);
1388 addr2->set_req(AddPNode::Address, base);
1389 phase->register_new_node(addr2, ctrl);
1390 addr = addr->clone();
1391 addr->set_req(AddPNode::Base, base);
1392 addr->set_req(AddPNode::Address, addr2);
1393 phase->register_new_node(addr, ctrl);
1394 }
1395 }
1396 }
1397 call_lrb_stub(ctrl, val, addr, result_mem, raw_mem, lrb->is_native(), phase);
1398 region->init_req(_evac_path, ctrl);
1399 val_phi->init_req(_evac_path, val);
1400 raw_mem_phi->init_req(_evac_path, result_mem);
1401
1402 phase->register_control(region, loop, heap_stable_iff);
1403 Node* out_val = val_phi;
1404 phase->register_new_node(val_phi, region);
1405 phase->register_new_node(raw_mem_phi, region);
1406
1407 fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase);
1408
1409 ctrl = orig_ctrl;
1410
1411 phase->igvn().replace_node(lrb, out_val);
1412
1413 follow_barrier_uses(out_val, ctrl, uses, phase);
1414
1415 for(uint next = 0; next < uses.size(); next++ ) {
1416 Node *n = uses.at(next);
1417 assert(phase->get_ctrl(n) == ctrl, "bad control");
1418 assert(n != init_raw_mem, "should leave input raw mem above the barrier");
1419 phase->set_ctrl(n, region);
1420 follow_barrier_uses(n, ctrl, uses, phase);
1421 }
1422
1423 // The slow path call produces memory: hook the raw memory phi
1424 // from the expanded load reference barrier with the rest of the graph
1425 // which may require adding memory phis at every post dominated
1426 // region and at enclosing loop heads. Use the memory state
1427 // collected in memory_nodes to fix the memory graph. Update that
1428 // memory state as we go.
1429 fixer.fix_mem(ctrl, region, init_raw_mem, raw_mem_for_ctrl, raw_mem_phi, uses);
1430 }
1431 // Done expanding load-reference-barriers.
1432 assert(ShenandoahBarrierSetC2::bsc2()->state()->load_reference_barriers_count() == 0, "all load reference barrier nodes should have been replaced");
1433
1434 for (int i = state->enqueue_barriers_count() - 1; i >= 0; i--) {
1435 Node* barrier = state->enqueue_barrier(i);
1436 Node* pre_val = barrier->in(1);
1437
1438 if (phase->igvn().type(pre_val)->higher_equal(TypePtr::NULL_PTR)) {
1439 ShouldNotReachHere();
1440 continue;
1441 }
1442
1443 Node* ctrl = phase->get_ctrl(barrier);
1444
1445 if (ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) {
1446 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0)->in(0), pre_val, ctrl->in(0), phase), "can't move");
1447 ctrl = ctrl->in(0)->in(0);
1448 phase->set_ctrl(barrier, ctrl);
1449 } else if (ctrl->is_CallRuntime()) {
1450 assert(is_dominator(phase->get_ctrl(pre_val), ctrl->in(0), pre_val, ctrl, phase), "can't move");
1451 ctrl = ctrl->in(0);
1452 phase->set_ctrl(barrier, ctrl);
1453 }
1454
1455 Node* init_ctrl = ctrl;
1456 IdealLoopTree* loop = phase->get_loop(ctrl);
1457 Node* raw_mem = fixer.find_mem(ctrl, barrier);
1458 Node* init_raw_mem = raw_mem;
1459 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL);
1460 Node* heap_stable_ctrl = NULL;
1461 Node* null_ctrl = NULL;
1462 uint last = phase->C->unique();
1463
1464 enum { _heap_stable = 1, _heap_unstable, PATH_LIMIT };
1465 Node* region = new RegionNode(PATH_LIMIT);
1466 Node* phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1467
1468 enum { _fast_path = 1, _slow_path, _null_path, PATH_LIMIT2 };
1469 Node* region2 = new RegionNode(PATH_LIMIT2);
1470 Node* phi2 = PhiNode::make(region2, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1471
1472 // Stable path.
1473 test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, ShenandoahHeap::MARKING);
1474 region->init_req(_heap_stable, heap_stable_ctrl);
1475 phi->init_req(_heap_stable, raw_mem);
1476
1477 // Null path
1478 Node* reg2_ctrl = NULL;
1479 test_null(ctrl, pre_val, null_ctrl, phase);
1480 if (null_ctrl != NULL) {
1481 reg2_ctrl = null_ctrl->in(0);
1482 region2->init_req(_null_path, null_ctrl);
1483 phi2->init_req(_null_path, raw_mem);
1484 } else {
1485 region2->del_req(_null_path);
1486 phi2->del_req(_null_path);
1487 }
1488
1489 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
1490 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1491 Node* thread = new ThreadLocalNode();
1492 phase->register_new_node(thread, ctrl);
1493 Node* buffer_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(buffer_offset));
1494 phase->register_new_node(buffer_adr, ctrl);
1495 Node* index_adr = new AddPNode(phase->C->top(), thread, phase->igvn().MakeConX(index_offset));
1496 phase->register_new_node(index_adr, ctrl);
1497
1498 BasicType index_bt = TypeX_X->basic_type();
1499 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
1500 const TypePtr* adr_type = TypeRawPtr::BOTTOM;
1501 Node* index = new LoadXNode(ctrl, raw_mem, index_adr, adr_type, TypeX_X, MemNode::unordered);
1502 phase->register_new_node(index, ctrl);
1503 Node* index_cmp = new CmpXNode(index, phase->igvn().MakeConX(0));
1504 phase->register_new_node(index_cmp, ctrl);
1505 Node* index_test = new BoolNode(index_cmp, BoolTest::ne);
1506 phase->register_new_node(index_test, ctrl);
1507 IfNode* queue_full_iff = new IfNode(ctrl, index_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
1508 if (reg2_ctrl == NULL) reg2_ctrl = queue_full_iff;
1509 phase->register_control(queue_full_iff, loop, ctrl);
1510 Node* not_full = new IfTrueNode(queue_full_iff);
1511 phase->register_control(not_full, loop, queue_full_iff);
1512 Node* full = new IfFalseNode(queue_full_iff);
1513 phase->register_control(full, loop, queue_full_iff);
1514
1515 ctrl = not_full;
1516
1517 Node* next_index = new SubXNode(index, phase->igvn().MakeConX(sizeof(intptr_t)));
1518 phase->register_new_node(next_index, ctrl);
1519
1520 Node* buffer = new LoadPNode(ctrl, raw_mem, buffer_adr, adr_type, TypeRawPtr::NOTNULL, MemNode::unordered);
1521 phase->register_new_node(buffer, ctrl);
1522 Node *log_addr = new AddPNode(phase->C->top(), buffer, next_index);
1523 phase->register_new_node(log_addr, ctrl);
1524 Node* log_store = new StorePNode(ctrl, raw_mem, log_addr, adr_type, pre_val, MemNode::unordered);
1525 phase->register_new_node(log_store, ctrl);
1526 // update the index
1527 Node* index_update = new StoreXNode(ctrl, log_store, index_adr, adr_type, next_index, MemNode::unordered);
1528 phase->register_new_node(index_update, ctrl);
1529
1530 // Fast-path case
1531 region2->init_req(_fast_path, ctrl);
1532 phi2->init_req(_fast_path, index_update);
1533
1534 ctrl = full;
1535
1536 Node* base = find_bottom_mem(ctrl, phase);
1537
1538 MergeMemNode* mm = MergeMemNode::make(base);
1539 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
1540 phase->register_new_node(mm, ctrl);
1541
1542 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);
1543 call->init_req(TypeFunc::Control, ctrl);
1544 call->init_req(TypeFunc::I_O, phase->C->top());
1545 call->init_req(TypeFunc::Memory, mm);
1546 call->init_req(TypeFunc::FramePtr, phase->C->top());
1547 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1548 call->init_req(TypeFunc::Parms, pre_val);
1549 call->init_req(TypeFunc::Parms+1, thread);
1550 phase->register_control(call, loop, ctrl);
1551
1552 Node* ctrl_proj = new ProjNode(call, TypeFunc::Control);
1553 phase->register_control(ctrl_proj, loop, call);
1554 Node* mem_proj = new ProjNode(call, TypeFunc::Memory);
1555 phase->register_new_node(mem_proj, call);
1556
1557 // Slow-path case
1558 region2->init_req(_slow_path, ctrl_proj);
1559 phi2->init_req(_slow_path, mem_proj);
1560
1561 phase->register_control(region2, loop, reg2_ctrl);
1562 phase->register_new_node(phi2, region2);
1563
1564 region->init_req(_heap_unstable, region2);
1565 phi->init_req(_heap_unstable, phi2);
1566
1567 phase->register_control(region, loop, heap_stable_ctrl->in(0));
1568 phase->register_new_node(phi, region);
1569
1570 fix_ctrl(barrier, region, fixer, uses, uses_to_ignore, last, phase);
1571 for(uint next = 0; next < uses.size(); next++ ) {
1572 Node *n = uses.at(next);
1573 assert(phase->get_ctrl(n) == init_ctrl, "bad control");
1574 assert(n != init_raw_mem, "should leave input raw mem above the barrier");
1575 phase->set_ctrl(n, region);
1576 follow_barrier_uses(n, init_ctrl, uses, phase);
1577 }
1578 fixer.fix_mem(init_ctrl, region, init_raw_mem, raw_mem_for_ctrl, phi, uses);
1579
1580 phase->igvn().replace_node(barrier, pre_val);
1581 }
1582 assert(state->enqueue_barriers_count() == 0, "all enqueue barrier nodes should have been replaced");
1583
1584 }
1585
1586 Node* ShenandoahBarrierC2Support::get_load_addr(PhaseIdealLoop* phase, VectorSet& visited, Node* in) {
1587 if (visited.test_set(in->_idx)) {
1588 return NULL;
1589 }
1590 switch (in->Opcode()) {
1591 case Op_Proj:
1592 return get_load_addr(phase, visited, in->in(0));
1593 case Op_CastPP:
1594 case Op_CheckCastPP:
1595 case Op_DecodeN:
1596 case Op_EncodeP:
1597 return get_load_addr(phase, visited, in->in(1));
1598 case Op_LoadN:
1599 case Op_LoadP:
1600 return in->in(MemNode::Address);
1601 case Op_CompareAndExchangeN:
1602 case Op_CompareAndExchangeP:
1603 case Op_GetAndSetN:
1604 case Op_GetAndSetP:
1605 case Op_ShenandoahCompareAndExchangeP:
1606 case Op_ShenandoahCompareAndExchangeN:
1607 // Those instructions would just have stored a different
1608 // value into the field. No use to attempt to fix it at this point.
1609 return phase->igvn().zerocon(T_OBJECT);
1610 case Op_CMoveP:
1611 case Op_CMoveN: {
1612 Node* t = get_load_addr(phase, visited, in->in(CMoveNode::IfTrue));
1613 Node* f = get_load_addr(phase, visited, in->in(CMoveNode::IfFalse));
1614 // Handle unambiguous cases: single address reported on both branches.
1615 if (t != NULL && f == NULL) return t;
1616 if (t == NULL && f != NULL) return f;
1617 if (t != NULL && t == f) return t;
1618 // Ambiguity.
1619 return phase->igvn().zerocon(T_OBJECT);
1620 }
1621 case Op_Phi: {
1622 Node* addr = NULL;
1623 for (uint i = 1; i < in->req(); i++) {
1624 Node* addr1 = get_load_addr(phase, visited, in->in(i));
1625 if (addr == NULL) {
1626 addr = addr1;
1627 }
1628 if (addr != addr1) {
1629 return phase->igvn().zerocon(T_OBJECT);
1630 }
1631 }
1632 return addr;
1633 }
1634 case Op_ShenandoahLoadReferenceBarrier:
1635 return get_load_addr(phase, visited, in->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1636 case Op_ShenandoahEnqueueBarrier:
1637 return get_load_addr(phase, visited, in->in(1));
1638 case Op_CallDynamicJava:
1639 case Op_CallLeaf:
1640 case Op_CallStaticJava:
1641 case Op_ConN:
1642 case Op_ConP:
1643 case Op_Parm:
1644 case Op_CreateEx:
1645 return phase->igvn().zerocon(T_OBJECT);
1646 default:
1647 #ifdef ASSERT
1648 fatal("Unknown node in get_load_addr: %s", NodeClassNames[in->Opcode()]);
1649 #endif
1650 return phase->igvn().zerocon(T_OBJECT);
1651 }
1652
1653 }
1654
1655 void ShenandoahBarrierC2Support::move_gc_state_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
1656 IdealLoopTree *loop = phase->get_loop(iff);
1657 Node* loop_head = loop->_head;
1658 Node* entry_c = loop_head->in(LoopNode::EntryControl);
1659
1660 Node* bol = iff->in(1);
1661 Node* cmp = bol->in(1);
1662 Node* andi = cmp->in(1);
1663 Node* load = andi->in(1);
1664
1665 assert(is_gc_state_load(load), "broken");
1666 if (!phase->is_dominator(load->in(0), entry_c)) {
1667 Node* mem_ctrl = NULL;
1668 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
1669 load = load->clone();
1670 load->set_req(MemNode::Memory, mem);
1671 load->set_req(0, entry_c);
1672 phase->register_new_node(load, entry_c);
1673 andi = andi->clone();
1674 andi->set_req(1, load);
1675 phase->register_new_node(andi, entry_c);
1676 cmp = cmp->clone();
1677 cmp->set_req(1, andi);
1678 phase->register_new_node(cmp, entry_c);
1679 bol = bol->clone();
1680 bol->set_req(1, cmp);
1681 phase->register_new_node(bol, entry_c);
1682
1683 phase->igvn().replace_input_of(iff, 1, bol);
1684 }
1685 }
1686
1687 bool ShenandoahBarrierC2Support::identical_backtoback_ifs(Node* n, PhaseIdealLoop* phase) {
1688 if (!n->is_If() || n->is_CountedLoopEnd()) {
1689 return false;
1690 }
1691 Node* region = n->in(0);
1692
1693 if (!region->is_Region()) {
1694 return false;
1695 }
1696 Node* dom = phase->idom(region);
1697 if (!dom->is_If()) {
1698 return false;
1699 }
1700
1701 if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) {
1702 return false;
1703 }
1704
1705 IfNode* dom_if = dom->as_If();
1706 Node* proj_true = dom_if->proj_out(1);
1707 Node* proj_false = dom_if->proj_out(0);
1708
1709 for (uint i = 1; i < region->req(); i++) {
1710 if (phase->is_dominator(proj_true, region->in(i))) {
1711 continue;
1712 }
1713 if (phase->is_dominator(proj_false, region->in(i))) {
1714 continue;
1715 }
1716 return false;
1717 }
1718
1719 return true;
1720 }
1721
1722 void ShenandoahBarrierC2Support::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) {
1723 assert(is_heap_stable_test(n), "no other tests");
1724 if (identical_backtoback_ifs(n, phase)) {
1725 Node* n_ctrl = n->in(0);
1726 if (phase->can_split_if(n_ctrl)) {
1727 IfNode* dom_if = phase->idom(n_ctrl)->as_If();
1728 if (is_heap_stable_test(n)) {
1729 Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1);
1730 assert(is_gc_state_load(gc_state_load), "broken");
1731 Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1);
1732 assert(is_gc_state_load(dom_gc_state_load), "broken");
1733 if (gc_state_load != dom_gc_state_load) {
1734 phase->igvn().replace_node(gc_state_load, dom_gc_state_load);
1735 }
1736 }
1737 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1738 Node* proj_true = dom_if->proj_out(1);
1739 Node* proj_false = dom_if->proj_out(0);
1740 Node* con_true = phase->igvn().makecon(TypeInt::ONE);
1741 Node* con_false = phase->igvn().makecon(TypeInt::ZERO);
1742
1743 for (uint i = 1; i < n_ctrl->req(); i++) {
1744 if (phase->is_dominator(proj_true, n_ctrl->in(i))) {
1745 bolphi->init_req(i, con_true);
1746 } else {
1747 assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1748 bolphi->init_req(i, con_false);
1749 }
1750 }
1751 phase->register_new_node(bolphi, n_ctrl);
1752 phase->igvn().replace_input_of(n, 1, bolphi);
1753 phase->do_split_if(n);
1754 }
1755 }
1756 }
1757
1758 IfNode* ShenandoahBarrierC2Support::find_unswitching_candidate(const IdealLoopTree* loop, PhaseIdealLoop* phase) {
1759 // Find first invariant test that doesn't exit the loop
1760 LoopNode *head = loop->_head->as_Loop();
1761 IfNode* unswitch_iff = NULL;
1762 Node* n = head->in(LoopNode::LoopBackControl);
1763 int loop_has_sfpts = -1;
1764 while (n != head) {
1765 Node* n_dom = phase->idom(n);
1766 if (n->is_Region()) {
1767 if (n_dom->is_If()) {
1768 IfNode* iff = n_dom->as_If();
1769 if (iff->in(1)->is_Bool()) {
1770 BoolNode* bol = iff->in(1)->as_Bool();
1771 if (bol->in(1)->is_Cmp()) {
1772 // If condition is invariant and not a loop exit,
1773 // then found reason to unswitch.
1774 if (is_heap_stable_test(iff) &&
1775 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) {
1776 assert(!loop->is_loop_exit(iff), "both branches should be in the loop");
1777 if (loop_has_sfpts == -1) {
1778 for(uint i = 0; i < loop->_body.size(); i++) {
1779 Node *m = loop->_body[i];
1780 if (m->is_SafePoint() && !m->is_CallLeaf()) {
1781 loop_has_sfpts = 1;
1782 break;
1783 }
1784 }
1785 if (loop_has_sfpts == -1) {
1786 loop_has_sfpts = 0;
1787 }
1788 }
1789 if (!loop_has_sfpts) {
1790 unswitch_iff = iff;
1791 }
1792 }
1793 }
1794 }
1795 }
1796 }
1797 n = n_dom;
1798 }
1799 return unswitch_iff;
1800 }
1801
1802
1803 void ShenandoahBarrierC2Support::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) {
1804 Node_List heap_stable_tests;
1805 stack.push(phase->C->start(), 0);
1806 do {
1807 Node* n = stack.node();
1808 uint i = stack.index();
1809
1810 if (i < n->outcnt()) {
1811 Node* u = n->raw_out(i);
1812 stack.set_index(i+1);
1813 if (!visited.test_set(u->_idx)) {
1814 stack.push(u, 0);
1815 }
1816 } else {
1817 stack.pop();
1818 if (n->is_If() && is_heap_stable_test(n)) {
1819 heap_stable_tests.push(n);
1820 }
1821 }
1822 } while (stack.size() > 0);
1823
1824 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1825 Node* n = heap_stable_tests.at(i);
1826 assert(is_heap_stable_test(n), "only evacuation test");
1827 merge_back_to_back_tests(n, phase);
1828 }
1829
1830 if (!phase->C->major_progress()) {
1831 VectorSet seen(Thread::current()->resource_area());
1832 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1833 Node* n = heap_stable_tests.at(i);
1834 IdealLoopTree* loop = phase->get_loop(n);
1835 if (loop != phase->ltree_root() &&
1836 loop->_child == NULL &&
1837 !loop->_irreducible) {
1838 Node* head = loop->_head;
1839 if (head->is_Loop() &&
1840 (!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) &&
1841 !seen.test_set(head->_idx)) {
1842 IfNode* iff = find_unswitching_candidate(loop, phase);
1843 if (iff != NULL) {
1844 Node* bol = iff->in(1);
1845 if (head->as_Loop()->is_strip_mined()) {
1846 head->as_Loop()->verify_strip_mined(0);
1847 }
1848 move_gc_state_test_out_of_loop(iff, phase);
1849
1850 AutoNodeBudget node_budget(phase);
1851
1852 if (loop->policy_unswitching(phase)) {
1853 if (head->as_Loop()->is_strip_mined()) {
1854 OuterStripMinedLoopNode* outer = head->as_CountedLoop()->outer_loop();
1855 hide_strip_mined_loop(outer, head->as_CountedLoop(), phase);
1856 }
1857 phase->do_unswitching(loop, old_new);
1858 } else {
1859 // Not proceeding with unswitching. Move load back in
1860 // the loop.
1861 phase->igvn().replace_input_of(iff, 1, bol);
1862 }
1863 }
1864 }
1865 }
1866 }
1867 }
1868 }
1869
1870 #ifdef ASSERT
1871 void ShenandoahBarrierC2Support::verify_raw_mem(RootNode* root) {
1872 const bool trace = false;
1873 ResourceMark rm;
1874 Unique_Node_List nodes;
1875 Unique_Node_List controls;
1876 Unique_Node_List memories;
1877
1878 nodes.push(root);
1879 for (uint next = 0; next < nodes.size(); next++) {
1880 Node *n = nodes.at(next);
1881 if (ShenandoahBarrierSetC2::is_shenandoah_lrb_call(n)) {
1882 controls.push(n);
1883 if (trace) { tty->print("XXXXXX verifying"); n->dump(); }
1884 for (uint next2 = 0; next2 < controls.size(); next2++) {
1885 Node *m = controls.at(next2);
1886 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
1887 Node* u = m->fast_out(i);
1888 if (u->is_CFG() && !u->is_Root() &&
1889 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1) &&
1890 !(u->is_Region() && u->unique_ctrl_out()->Opcode() == Op_Halt)) {
1891 if (trace) { tty->print("XXXXXX pushing control"); u->dump(); }
1892 controls.push(u);
1893 }
1894 }
1895 }
1896 memories.push(n->as_Call()->proj_out(TypeFunc::Memory));
1897 for (uint next2 = 0; next2 < memories.size(); next2++) {
1898 Node *m = memories.at(next2);
1899 assert(m->bottom_type() == Type::MEMORY, "");
1900 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
1901 Node* u = m->fast_out(i);
1902 if (u->bottom_type() == Type::MEMORY && (u->is_Mem() || u->is_ClearArray())) {
1903 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
1904 memories.push(u);
1905 } else if (u->is_LoadStore()) {
1906 if (trace) { tty->print("XXXXXX pushing memory"); u->find_out_with(Op_SCMemProj)->dump(); }
1907 memories.push(u->find_out_with(Op_SCMemProj));
1908 } else if (u->is_MergeMem() && u->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == m) {
1909 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
1910 memories.push(u);
1911 } else if (u->is_Phi()) {
1912 assert(u->bottom_type() == Type::MEMORY, "");
1913 if (u->adr_type() == TypeRawPtr::BOTTOM || u->adr_type() == TypePtr::BOTTOM) {
1914 assert(controls.member(u->in(0)), "");
1915 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
1916 memories.push(u);
1917 }
1918 } else if (u->is_SafePoint() || u->is_MemBar()) {
1919 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
1920 Node* uu = u->fast_out(j);
1921 if (uu->bottom_type() == Type::MEMORY) {
1922 if (trace) { tty->print("XXXXXX pushing memory"); uu->dump(); }
1923 memories.push(uu);
1924 }
1925 }
1926 }
1927 }
1928 }
1929 for (uint next2 = 0; next2 < controls.size(); next2++) {
1930 Node *m = controls.at(next2);
1931 if (m->is_Region()) {
1932 bool all_in = true;
1933 for (uint i = 1; i < m->req(); i++) {
1934 if (!controls.member(m->in(i))) {
1935 all_in = false;
1936 break;
1937 }
1938 }
1939 if (trace) { tty->print("XXX verifying %s", all_in ? "all in" : ""); m->dump(); }
1940 bool found_phi = false;
1941 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax && !found_phi; j++) {
1942 Node* u = m->fast_out(j);
1943 if (u->is_Phi() && memories.member(u)) {
1944 found_phi = true;
1945 for (uint i = 1; i < u->req() && found_phi; i++) {
1946 Node* k = u->in(i);
1947 if (memories.member(k) != controls.member(m->in(i))) {
1948 found_phi = false;
1949 }
1950 }
1951 }
1952 }
1953 assert(found_phi || all_in, "");
1954 }
1955 }
1956 controls.clear();
1957 memories.clear();
1958 }
1959 for( uint i = 0; i < n->len(); ++i ) {
1960 Node *m = n->in(i);
1961 if (m != NULL) {
1962 nodes.push(m);
1963 }
1964 }
1965 }
1966 }
1967 #endif
1968
1969 ShenandoahEnqueueBarrierNode::ShenandoahEnqueueBarrierNode(Node* val) : Node(NULL, val) {
1970 ShenandoahBarrierSetC2::bsc2()->state()->add_enqueue_barrier(this);
1971 }
1972
1973 const Type* ShenandoahEnqueueBarrierNode::bottom_type() const {
1974 if (in(1) == NULL || in(1)->is_top()) {
1975 return Type::TOP;
1976 }
1977 const Type* t = in(1)->bottom_type();
1978 if (t == TypePtr::NULL_PTR) {
1979 return t;
1980 }
1981 return t->is_oopptr();
1982 }
1983
1984 const Type* ShenandoahEnqueueBarrierNode::Value(PhaseGVN* phase) const {
1985 if (in(1) == NULL) {
1986 return Type::TOP;
1987 }
1988 const Type* t = phase->type(in(1));
1989 if (t == Type::TOP) {
1990 return Type::TOP;
1991 }
1992 if (t == TypePtr::NULL_PTR) {
1993 return t;
1994 }
1995 return t->is_oopptr();
1996 }
1997
1998 int ShenandoahEnqueueBarrierNode::needed(Node* n) {
1999 if (n == NULL ||
2000 n->is_Allocate() ||
2001 n->Opcode() == Op_ShenandoahEnqueueBarrier ||
2002 n->bottom_type() == TypePtr::NULL_PTR ||
2003 (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL)) {
2004 return NotNeeded;
2005 }
2006 if (n->is_Phi() ||
2007 n->is_CMove()) {
2008 return MaybeNeeded;
2009 }
2010 return Needed;
2011 }
2012
2013 Node* ShenandoahEnqueueBarrierNode::next(Node* n) {
2014 for (;;) {
2015 if (n == NULL) {
2016 return n;
2017 } else if (n->bottom_type() == TypePtr::NULL_PTR) {
2018 return n;
2019 } else if (n->bottom_type()->make_oopptr() != NULL && n->bottom_type()->make_oopptr()->const_oop() != NULL) {
2020 return n;
2021 } else if (n->is_ConstraintCast() ||
2022 n->Opcode() == Op_DecodeN ||
2023 n->Opcode() == Op_EncodeP) {
2024 n = n->in(1);
2025 } else if (n->is_Proj()) {
2026 n = n->in(0);
2027 } else {
2028 return n;
2029 }
2030 }
2031 ShouldNotReachHere();
2032 return NULL;
2033 }
2034
2035 Node* ShenandoahEnqueueBarrierNode::Identity(PhaseGVN* phase) {
2036 PhaseIterGVN* igvn = phase->is_IterGVN();
2037
2038 Node* n = next(in(1));
2039
2040 int cont = needed(n);
2041
2042 if (cont == NotNeeded) {
2043 return in(1);
2044 } else if (cont == MaybeNeeded) {
2045 if (igvn == NULL) {
2046 phase->record_for_igvn(this);
2047 return this;
2048 } else {
2049 ResourceMark rm;
2050 Unique_Node_List wq;
2051 uint wq_i = 0;
2052
2053 for (;;) {
2054 if (n->is_Phi()) {
2055 for (uint i = 1; i < n->req(); i++) {
2056 Node* m = n->in(i);
2057 if (m != NULL) {
2058 wq.push(m);
2059 }
2060 }
2061 } else {
2062 assert(n->is_CMove(), "nothing else here");
2063 Node* m = n->in(CMoveNode::IfFalse);
2064 wq.push(m);
2065 m = n->in(CMoveNode::IfTrue);
2066 wq.push(m);
2067 }
2068 Node* orig_n = NULL;
2069 do {
2070 if (wq_i >= wq.size()) {
2071 return in(1);
2072 }
2073 n = wq.at(wq_i);
2074 wq_i++;
2075 orig_n = n;
2076 n = next(n);
2077 cont = needed(n);
2078 if (cont == Needed) {
2079 return this;
2080 }
2081 } while (cont != MaybeNeeded || (orig_n != n && wq.member(n)));
2082 }
2083 }
2084 }
2085
2086 return this;
2087 }
2088
2089 #ifdef ASSERT
2090 static bool has_never_branch(Node* root) {
2091 for (uint i = 1; i < root->req(); i++) {
2092 Node* in = root->in(i);
2093 if (in != NULL && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->Opcode() == Op_NeverBranch) {
2094 return true;
2095 }
2096 }
2097 return false;
2098 }
2099 #endif
2100
2101 void MemoryGraphFixer::collect_memory_nodes() {
2102 Node_Stack stack(0);
2103 VectorSet visited(Thread::current()->resource_area());
2104 Node_List regions;
2105
2106 // Walk the raw memory graph and create a mapping from CFG node to
2107 // memory node. Exclude phis for now.
2108 stack.push(_phase->C->root(), 1);
2109 do {
2110 Node* n = stack.node();
2111 int opc = n->Opcode();
2112 uint i = stack.index();
2113 if (i < n->req()) {
2114 Node* mem = NULL;
2115 if (opc == Op_Root) {
2116 Node* in = n->in(i);
2117 int in_opc = in->Opcode();
2118 if (in_opc == Op_Return || in_opc == Op_Rethrow) {
2119 mem = in->in(TypeFunc::Memory);
2120 } else if (in_opc == Op_Halt) {
2121 if (in->in(0)->is_Region()) {
2122 Node* r = in->in(0);
2123 for (uint j = 1; j < r->req(); j++) {
2124 assert(r->in(j)->Opcode() != Op_NeverBranch, "");
2125 }
2126 } else {
2127 Node* proj = in->in(0);
2128 assert(proj->is_Proj(), "");
2129 Node* in = proj->in(0);
2130 assert(in->is_CallStaticJava() || in->Opcode() == Op_NeverBranch || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
2131 if (in->is_CallStaticJava()) {
2132 mem = in->in(TypeFunc::Memory);
2133 } else if (in->Opcode() == Op_Catch) {
2134 Node* call = in->in(0)->in(0);
2135 assert(call->is_Call(), "");
2136 mem = call->in(TypeFunc::Memory);
2137 } else if (in->Opcode() == Op_NeverBranch) {
2138 Node* head = in->in(0);
2139 assert(head->is_Region(), "unexpected infinite loop graph shape");
2140
2141 Node* phi_mem = NULL;
2142 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2143 Node* u = head->fast_out(j);
2144 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
2145 if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2146 assert(phi_mem == NULL || phi_mem->adr_type() == TypePtr::BOTTOM, "");
2147 phi_mem = u;
2148 } else if (u->adr_type() == TypePtr::BOTTOM) {
2149 assert(phi_mem == NULL || _phase->C->get_alias_index(phi_mem->adr_type()) == _alias, "");
2150 if (phi_mem == NULL) {
2151 phi_mem = u;
2152 }
2153 }
2154 }
2155 }
2156 if (phi_mem == NULL) {
2157 for (uint j = 1; j < head->req(); j++) {
2158 Node* tail = head->in(j);
2159 if (!_phase->is_dominator(head, tail)) {
2160 continue;
2161 }
2162 Node* c = tail;
2163 while (c != head) {
2164 if (c->is_SafePoint() && !c->is_CallLeaf()) {
2165 Node* m =c->in(TypeFunc::Memory);
2166 if (m->is_MergeMem()) {
2167 m = m->as_MergeMem()->memory_at(_alias);
2168 }
2169 assert(mem == NULL || mem == m, "several memory states");
2170 mem = m;
2171 }
2172 c = _phase->idom(c);
2173 }
2174 assert(mem != NULL, "should have found safepoint");
2175 }
2176 assert(mem != NULL, "should have found safepoint");
2177 } else {
2178 mem = phi_mem;
2179 }
2180 }
2181 }
2182 } else {
2183 #ifdef ASSERT
2184 n->dump();
2185 in->dump();
2186 #endif
2187 ShouldNotReachHere();
2188 }
2189 } else {
2190 assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
2191 assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
2192 mem = n->in(i);
2193 }
2194 i++;
2195 stack.set_index(i);
2196 if (mem == NULL) {
2197 continue;
2198 }
2199 for (;;) {
2200 if (visited.test_set(mem->_idx) || mem->is_Start()) {
2201 break;
2202 }
2203 if (mem->is_Phi()) {
2204 stack.push(mem, 2);
2205 mem = mem->in(1);
2206 } else if (mem->is_Proj()) {
2207 stack.push(mem, mem->req());
2208 mem = mem->in(0);
2209 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
2210 mem = mem->in(TypeFunc::Memory);
2211 } else if (mem->is_MergeMem()) {
2212 MergeMemNode* mm = mem->as_MergeMem();
2213 mem = mm->memory_at(_alias);
2214 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
2215 assert(_alias == Compile::AliasIdxRaw, "");
2216 stack.push(mem, mem->req());
2217 mem = mem->in(MemNode::Memory);
2218 } else {
2219 #ifdef ASSERT
2220 mem->dump();
2221 #endif
2222 ShouldNotReachHere();
2223 }
2224 }
2225 } else {
2226 if (n->is_Phi()) {
2227 // Nothing
2228 } else if (!n->is_Root()) {
2229 Node* c = get_ctrl(n);
2230 _memory_nodes.map(c->_idx, n);
2231 }
2232 stack.pop();
2233 }
2234 } while(stack.is_nonempty());
2235
2236 // Iterate over CFG nodes in rpo and propagate memory state to
2237 // compute memory state at regions, creating new phis if needed.
2238 Node_List rpo_list;
2239 visited.clear();
2240 _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
2241 Node* root = rpo_list.pop();
2242 assert(root == _phase->C->root(), "");
2243
2244 const bool trace = false;
2245 #ifdef ASSERT
2246 if (trace) {
2247 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2248 Node* c = rpo_list.at(i);
2249 if (_memory_nodes[c->_idx] != NULL) {
2250 tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump();
2251 }
2252 }
2253 }
2254 #endif
2255 uint last = _phase->C->unique();
2256
2257 #ifdef ASSERT
2258 uint8_t max_depth = 0;
2259 for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
2260 IdealLoopTree* lpt = iter.current();
2261 max_depth = MAX2(max_depth, lpt->_nest);
2262 }
2263 #endif
2264
2265 bool progress = true;
2266 int iteration = 0;
2267 Node_List dead_phis;
2268 while (progress) {
2269 progress = false;
2270 iteration++;
2271 assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2272 if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
2273
2274 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2275 Node* c = rpo_list.at(i);
2276
2277 Node* prev_mem = _memory_nodes[c->_idx];
2278 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2279 Node* prev_region = regions[c->_idx];
2280 Node* unique = NULL;
2281 for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
2282 Node* m = _memory_nodes[c->in(j)->_idx];
2283 assert(m != NULL || (c->is_Loop() && j == LoopNode::LoopBackControl && iteration == 1) || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "expect memory state");
2284 if (m != NULL) {
2285 if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
2286 assert(c->is_Loop() && j == LoopNode::LoopBackControl || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2287 // continue
2288 } else if (unique == NULL) {
2289 unique = m;
2290 } else if (m == unique) {
2291 // continue
2292 } else {
2293 unique = NodeSentinel;
2294 }
2295 }
2296 }
2297 assert(unique != NULL, "empty phi???");
2298 if (unique != NodeSentinel) {
2299 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c) {
2300 dead_phis.push(prev_region);
2301 }
2302 regions.map(c->_idx, unique);
2303 } else {
2304 Node* phi = NULL;
2305 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
2306 phi = prev_region;
2307 for (uint k = 1; k < c->req(); k++) {
2308 Node* m = _memory_nodes[c->in(k)->_idx];
2309 assert(m != NULL, "expect memory state");
2310 phi->set_req(k, m);
2311 }
2312 } else {
2313 for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == NULL; j++) {
2314 Node* u = c->fast_out(j);
2315 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2316 (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
2317 phi = u;
2318 for (uint k = 1; k < c->req() && phi != NULL; k++) {
2319 Node* m = _memory_nodes[c->in(k)->_idx];
2320 assert(m != NULL, "expect memory state");
2321 if (u->in(k) != m) {
2322 phi = NULL;
2323 }
2324 }
2325 }
2326 }
2327 if (phi == NULL) {
2328 phi = new PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
2329 for (uint k = 1; k < c->req(); k++) {
2330 Node* m = _memory_nodes[c->in(k)->_idx];
2331 assert(m != NULL, "expect memory state");
2332 phi->init_req(k, m);
2333 }
2334 }
2335 }
2336 assert(phi != NULL, "");
2337 regions.map(c->_idx, phi);
2338 }
2339 Node* current_region = regions[c->_idx];
2340 if (current_region != prev_region) {
2341 progress = true;
2342 if (prev_region == prev_mem) {
2343 _memory_nodes.map(c->_idx, current_region);
2344 }
2345 }
2346 } else if (prev_mem == NULL || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
2347 Node* m = _memory_nodes[_phase->idom(c)->_idx];
2348 assert(m != NULL, "expect memory state");
2349 if (m != prev_mem) {
2350 _memory_nodes.map(c->_idx, m);
2351 progress = true;
2352 }
2353 }
2354 #ifdef ASSERT
2355 if (trace) { tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); }
2356 #endif
2357 }
2358 }
2359
2360 // Replace existing phi with computed memory state for that region
2361 // if different (could be a new phi or a dominating memory node if
2362 // that phi was found to be useless).
2363 while (dead_phis.size() > 0) {
2364 Node* n = dead_phis.pop();
2365 n->replace_by(_phase->C->top());
2366 n->destruct();
2367 }
2368 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2369 Node* c = rpo_list.at(i);
2370 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2371 Node* n = regions[c->_idx];
2372 if (n->is_Phi() && n->_idx >= last && n->in(0) == c) {
2373 _phase->register_new_node(n, c);
2374 }
2375 }
2376 }
2377 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2378 Node* c = rpo_list.at(i);
2379 if (c->is_Region() && (_include_lsm || !c->is_OuterStripMinedLoop())) {
2380 Node* n = regions[c->_idx];
2381 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
2382 Node* u = c->fast_out(i);
2383 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2384 u != n) {
2385 if (u->adr_type() == TypePtr::BOTTOM) {
2386 fix_memory_uses(u, n, n, c);
2387 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2388 _phase->lazy_replace(u, n);
2389 --i; --imax;
2390 }
2391 }
2392 }
2393 }
2394 }
2395 }
2396
2397 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
2398 Node* c = _phase->get_ctrl(n);
2399 if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Call()) {
2400 assert(c == n->in(0), "");
2401 CallNode* call = c->as_Call();
2402 CallProjections projs;
2403 call->extract_projections(&projs, true, false);
2404 if (projs.catchall_memproj != NULL) {
2405 if (projs.fallthrough_memproj == n) {
2406 c = projs.fallthrough_catchproj;
2407 } else {
2408 assert(projs.catchall_memproj == n, "");
2409 c = projs.catchall_catchproj;
2410 }
2411 }
2412 }
2413 return c;
2414 }
2415
2416 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
2417 if (_phase->has_ctrl(n))
2418 return get_ctrl(n);
2419 else {
2420 assert (n->is_CFG(), "must be a CFG node");
2421 return n;
2422 }
2423 }
2424
2425 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
2426 return m != NULL && get_ctrl(m) == c;
2427 }
2428
2429 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
2430 assert(n == NULL || _phase->ctrl_or_self(n) == ctrl, "");
2431 Node* mem = _memory_nodes[ctrl->_idx];
2432 Node* c = ctrl;
2433 while (!mem_is_valid(mem, c) &&
2434 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
2435 c = _phase->idom(c);
2436 mem = _memory_nodes[c->_idx];
2437 }
2438 if (n != NULL && mem_is_valid(mem, c)) {
2439 while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
2440 mem = next_mem(mem, _alias);
2441 }
2442 if (mem->is_MergeMem()) {
2443 mem = mem->as_MergeMem()->memory_at(_alias);
2444 }
2445 if (!mem_is_valid(mem, c)) {
2446 do {
2447 c = _phase->idom(c);
2448 mem = _memory_nodes[c->_idx];
2449 } while (!mem_is_valid(mem, c) &&
2450 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
2451 }
2452 }
2453 assert(mem->bottom_type() == Type::MEMORY, "");
2454 return mem;
2455 }
2456
2457 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
2458 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2459 Node* use = region->fast_out(i);
2460 if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
2461 (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
2462 return true;
2463 }
2464 }
2465 return false;
2466 }
2467
2468 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
2469 assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
2470 const bool trace = false;
2471 DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
2472 DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
2473 GrowableArray<Node*> phis;
2474 if (mem_for_ctrl != mem) {
2475 Node* old = mem_for_ctrl;
2476 Node* prev = NULL;
2477 while (old != mem) {
2478 prev = old;
2479 if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
2480 assert(_alias == Compile::AliasIdxRaw, "");
2481 old = old->in(MemNode::Memory);
2482 } else if (old->Opcode() == Op_SCMemProj) {
2483 assert(_alias == Compile::AliasIdxRaw, "");
2484 old = old->in(0);
2485 } else {
2486 ShouldNotReachHere();
2487 }
2488 }
2489 assert(prev != NULL, "");
2490 if (new_ctrl != ctrl) {
2491 _memory_nodes.map(ctrl->_idx, mem);
2492 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2493 }
2494 uint input = (uint)MemNode::Memory;
2495 _phase->igvn().replace_input_of(prev, input, new_mem);
2496 } else {
2497 uses.clear();
2498 _memory_nodes.map(new_ctrl->_idx, new_mem);
2499 uses.push(new_ctrl);
2500 for(uint next = 0; next < uses.size(); next++ ) {
2501 Node *n = uses.at(next);
2502 assert(n->is_CFG(), "");
2503 DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
2504 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2505 Node* u = n->fast_out(i);
2506 if (!u->is_Root() && u->is_CFG() && u != n) {
2507 Node* m = _memory_nodes[u->_idx];
2508 if (u->is_Region() && (!u->is_OuterStripMinedLoop() || _include_lsm) &&
2509 !has_mem_phi(u) &&
2510 u->unique_ctrl_out()->Opcode() != Op_Halt) {
2511 DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
2512 DEBUG_ONLY(if (trace && m != NULL) { tty->print("ZZZ mem"); m->dump(); });
2513
2514 if (!mem_is_valid(m, u) || !m->is_Phi()) {
2515 bool push = true;
2516 bool create_phi = true;
2517 if (_phase->is_dominator(new_ctrl, u)) {
2518 create_phi = false;
2519 }
2520 if (create_phi) {
2521 Node* phi = new PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
2522 _phase->register_new_node(phi, u);
2523 phis.push(phi);
2524 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
2525 if (!mem_is_valid(m, u)) {
2526 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
2527 _memory_nodes.map(u->_idx, phi);
2528 } else {
2529 DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
2530 for (;;) {
2531 assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), "");
2532 Node* next = NULL;
2533 if (m->is_Proj()) {
2534 next = m->in(0);
2535 } else {
2536 assert(m->is_Mem() || m->is_LoadStore(), "");
2537 assert(_alias == Compile::AliasIdxRaw, "");
2538 next = m->in(MemNode::Memory);
2539 }
2540 if (_phase->get_ctrl(next) != u) {
2541 break;
2542 }
2543 if (next->is_MergeMem()) {
2544 assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
2545 break;
2546 }
2547 if (next->is_Phi()) {
2548 assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
2549 break;
2550 }
2551 m = next;
2552 }
2553
2554 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
2555 assert(m->is_Mem() || m->is_LoadStore(), "");
2556 uint input = (uint)MemNode::Memory;
2557 _phase->igvn().replace_input_of(m, input, phi);
2558 push = false;
2559 }
2560 } else {
2561 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
2562 }
2563 if (push) {
2564 uses.push(u);
2565 }
2566 }
2567 } else if (!mem_is_valid(m, u) &&
2568 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1)) {
2569 uses.push(u);
2570 }
2571 }
2572 }
2573 }
2574 for (int i = 0; i < phis.length(); i++) {
2575 Node* n = phis.at(i);
2576 Node* r = n->in(0);
2577 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
2578 for (uint j = 1; j < n->req(); j++) {
2579 Node* m = find_mem(r->in(j), NULL);
2580 _phase->igvn().replace_input_of(n, j, m);
2581 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
2582 }
2583 }
2584 }
2585 uint last = _phase->C->unique();
2586 MergeMemNode* mm = NULL;
2587 int alias = _alias;
2588 DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
2589 // Process loads first to not miss an anti-dependency: if the memory
2590 // edge of a store is updated before a load is processed then an
2591 // anti-dependency may be missed.
2592 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2593 Node* u = mem->out(i);
2594 if (u->_idx < last && u->is_Load() && _phase->C->get_alias_index(u->adr_type()) == alias) {
2595 Node* m = find_mem(_phase->get_ctrl(u), u);
2596 if (m != mem) {
2597 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2598 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2599 --i;
2600 }
2601 }
2602 }
2603 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2604 Node* u = mem->out(i);
2605 if (u->_idx < last) {
2606 if (u->is_Mem()) {
2607 if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2608 Node* m = find_mem(_phase->get_ctrl(u), u);
2609 if (m != mem) {
2610 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2611 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2612 --i;
2613 }
2614 }
2615 } else if (u->is_MergeMem()) {
2616 MergeMemNode* u_mm = u->as_MergeMem();
2617 if (u_mm->memory_at(alias) == mem) {
2618 MergeMemNode* newmm = NULL;
2619 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2620 Node* uu = u->fast_out(j);
2621 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2622 if (uu->is_Phi()) {
2623 assert(uu->adr_type() == TypePtr::BOTTOM, "");
2624 Node* region = uu->in(0);
2625 int nb = 0;
2626 for (uint k = 1; k < uu->req(); k++) {
2627 if (uu->in(k) == u) {
2628 Node* m = find_mem(region->in(k), NULL);
2629 if (m != mem) {
2630 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
2631 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2632 if (newmm != u) {
2633 _phase->igvn().replace_input_of(uu, k, newmm);
2634 nb++;
2635 --jmax;
2636 }
2637 }
2638 }
2639 }
2640 if (nb > 0) {
2641 --j;
2642 }
2643 } else {
2644 Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
2645 if (m != mem) {
2646 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
2647 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2648 if (newmm != u) {
2649 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2650 --j, --jmax;
2651 }
2652 }
2653 }
2654 }
2655 }
2656 } else if (u->is_Phi()) {
2657 assert(u->bottom_type() == Type::MEMORY, "what else?");
2658 if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
2659 Node* region = u->in(0);
2660 bool replaced = false;
2661 for (uint j = 1; j < u->req(); j++) {
2662 if (u->in(j) == mem) {
2663 Node* m = find_mem(region->in(j), NULL);
2664 Node* nnew = m;
2665 if (m != mem) {
2666 if (u->adr_type() == TypePtr::BOTTOM) {
2667 mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
2668 nnew = mm;
2669 }
2670 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
2671 _phase->igvn().replace_input_of(u, j, nnew);
2672 replaced = true;
2673 }
2674 }
2675 }
2676 if (replaced) {
2677 --i;
2678 }
2679 }
2680 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2681 u->adr_type() == NULL) {
2682 assert(u->adr_type() != NULL ||
2683 u->Opcode() == Op_Rethrow ||
2684 u->Opcode() == Op_Return ||
2685 u->Opcode() == Op_SafePoint ||
2686 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2687 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2688 u->Opcode() == Op_CallLeaf, "");
2689 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2690 if (m != mem) {
2691 mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
2692 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2693 --i;
2694 }
2695 } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2696 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2697 if (m != mem) {
2698 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2699 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2700 --i;
2701 }
2702 } else if (u->adr_type() != TypePtr::BOTTOM &&
2703 _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
2704 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2705 assert(m != mem, "");
2706 // u is on the wrong slice...
2707 assert(u->is_ClearArray(), "");
2708 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2709 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2710 --i;
2711 }
2712 }
2713 }
2714 #ifdef ASSERT
2715 assert(new_mem->outcnt() > 0, "");
2716 for (int i = 0; i < phis.length(); i++) {
2717 Node* n = phis.at(i);
2718 assert(n->outcnt() > 0, "new phi must have uses now");
2719 }
2720 #endif
2721 }
2722
2723 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
2724 MergeMemNode* mm = MergeMemNode::make(mem);
2725 mm->set_memory_at(_alias, rep_proj);
2726 _phase->register_new_node(mm, rep_ctrl);
2727 return mm;
2728 }
2729
2730 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
2731 MergeMemNode* newmm = NULL;
2732 MergeMemNode* u_mm = u->as_MergeMem();
2733 Node* c = _phase->get_ctrl(u);
2734 if (_phase->is_dominator(c, rep_ctrl)) {
2735 c = rep_ctrl;
2736 } else {
2737 assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
2738 }
2739 if (u->outcnt() == 1) {
2740 if (u->req() > (uint)_alias && u->in(_alias) == mem) {
2741 _phase->igvn().replace_input_of(u, _alias, rep_proj);
2742 --i;
2743 } else {
2744 _phase->igvn().rehash_node_delayed(u);
2745 u_mm->set_memory_at(_alias, rep_proj);
2746 }
2747 newmm = u_mm;
2748 _phase->set_ctrl_and_loop(u, c);
2749 } else {
2750 // can't simply clone u and then change one of its input because
2751 // it adds and then removes an edge which messes with the
2752 // DUIterator
2753 newmm = MergeMemNode::make(u_mm->base_memory());
2754 for (uint j = 0; j < u->req(); j++) {
2755 if (j < newmm->req()) {
2756 if (j == (uint)_alias) {
2757 newmm->set_req(j, rep_proj);
2758 } else if (newmm->in(j) != u->in(j)) {
2759 newmm->set_req(j, u->in(j));
2760 }
2761 } else if (j == (uint)_alias) {
2762 newmm->add_req(rep_proj);
2763 } else {
2764 newmm->add_req(u->in(j));
2765 }
2766 }
2767 if ((uint)_alias >= u->req()) {
2768 newmm->set_memory_at(_alias, rep_proj);
2769 }
2770 _phase->register_new_node(newmm, c);
2771 }
2772 return newmm;
2773 }
2774
2775 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
2776 if (phi->adr_type() == TypePtr::BOTTOM) {
2777 Node* region = phi->in(0);
2778 for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
2779 Node* uu = region->fast_out(j);
2780 if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
2781 return false;
2782 }
2783 }
2784 return true;
2785 }
2786 return _phase->C->get_alias_index(phi->adr_type()) == _alias;
2787 }
2788
2789 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
2790 uint last = _phase-> C->unique();
2791 MergeMemNode* mm = NULL;
2792 assert(mem->bottom_type() == Type::MEMORY, "");
2793 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2794 Node* u = mem->out(i);
2795 if (u != replacement && u->_idx < last) {
2796 if (u->is_MergeMem()) {
2797 MergeMemNode* u_mm = u->as_MergeMem();
2798 if (u_mm->memory_at(_alias) == mem) {
2799 MergeMemNode* newmm = NULL;
2800 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2801 Node* uu = u->fast_out(j);
2802 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2803 if (uu->is_Phi()) {
2804 if (should_process_phi(uu)) {
2805 Node* region = uu->in(0);
2806 int nb = 0;
2807 for (uint k = 1; k < uu->req(); k++) {
2808 if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
2809 if (newmm == NULL) {
2810 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2811 }
2812 if (newmm != u) {
2813 _phase->igvn().replace_input_of(uu, k, newmm);
2814 nb++;
2815 --jmax;
2816 }
2817 }
2818 }
2819 if (nb > 0) {
2820 --j;
2821 }
2822 }
2823 } else {
2824 if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
2825 if (newmm == NULL) {
2826 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2827 }
2828 if (newmm != u) {
2829 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2830 --j, --jmax;
2831 }
2832 }
2833 }
2834 }
2835 }
2836 } else if (u->is_Phi()) {
2837 assert(u->bottom_type() == Type::MEMORY, "what else?");
2838 Node* region = u->in(0);
2839 if (should_process_phi(u)) {
2840 bool replaced = false;
2841 for (uint j = 1; j < u->req(); j++) {
2842 if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
2843 Node* nnew = rep_proj;
2844 if (u->adr_type() == TypePtr::BOTTOM) {
2845 if (mm == NULL) {
2846 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2847 }
2848 nnew = mm;
2849 }
2850 _phase->igvn().replace_input_of(u, j, nnew);
2851 replaced = true;
2852 }
2853 }
2854 if (replaced) {
2855 --i;
2856 }
2857
2858 }
2859 } else if ((u->adr_type() == TypePtr::BOTTOM && u->Opcode() != Op_StrInflatedCopy) ||
2860 u->adr_type() == NULL) {
2861 assert(u->adr_type() != NULL ||
2862 u->Opcode() == Op_Rethrow ||
2863 u->Opcode() == Op_Return ||
2864 u->Opcode() == Op_SafePoint ||
2865 u->Opcode() == Op_StoreIConditional ||
2866 u->Opcode() == Op_StoreLConditional ||
2867 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2868 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2869 u->Opcode() == Op_CallLeaf, "%s", u->Name());
2870 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2871 if (mm == NULL) {
2872 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2873 }
2874 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2875 --i;
2876 }
2877 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2878 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2879 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
2880 --i;
2881 }
2882 }
2883 }
2884 }
2885 }
2886
2887 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj, bool native)
2888 : Node(ctrl, obj), _native(native) {
2889 ShenandoahBarrierSetC2::bsc2()->state()->add_load_reference_barrier(this);
2890 }
2891
2892 bool ShenandoahLoadReferenceBarrierNode::is_native() const {
2893 return _native;
2894 }
2895
2896 uint ShenandoahLoadReferenceBarrierNode::size_of() const {
2897 return sizeof(*this);
2898 }
2899
2900 uint ShenandoahLoadReferenceBarrierNode::hash() const {
2901 return Node::hash() + (_native ? 1 : 0);
2902 }
2903
2904 bool ShenandoahLoadReferenceBarrierNode::cmp( const Node &n ) const {
2905 return Node::cmp(n) && n.Opcode() == Op_ShenandoahLoadReferenceBarrier &&
2906 _native == ((const ShenandoahLoadReferenceBarrierNode&)n)._native;
2907 }
2908
2909 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const {
2910 if (in(ValueIn) == NULL || in(ValueIn)->is_top()) {
2911 return Type::TOP;
2912 }
2913 const Type* t = in(ValueIn)->bottom_type();
2914 if (t == TypePtr::NULL_PTR) {
2915 return t;
2916 }
2917 return t->is_oopptr();
2918 }
2919
2920 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseGVN* phase) const {
2921 // Either input is TOP ==> the result is TOP
2922 const Type *t2 = phase->type(in(ValueIn));
2923 if( t2 == Type::TOP ) return Type::TOP;
2924
2925 if (t2 == TypePtr::NULL_PTR) {
2926 return t2;
2927 }
2928
2929 const Type* type = t2->is_oopptr();
2930 return type;
2931 }
2932
2933 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseGVN* phase) {
2934 Node* value = in(ValueIn);
2935 if (!needs_barrier(phase, value)) {
2936 return value;
2937 }
2938 return this;
2939 }
2940
2941 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseGVN* phase, Node* n) {
2942 Unique_Node_List visited;
2943 return needs_barrier_impl(phase, n, visited);
2944 }
2945
2946 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseGVN* phase, Node* n, Unique_Node_List &visited) {
2947 if (n == NULL) return false;
2948 if (visited.member(n)) {
2949 return false; // Been there.
2950 }
2951 visited.push(n);
2952
2953 if (n->is_Allocate()) {
2954 // tty->print_cr("optimize barrier on alloc");
2955 return false;
2956 }
2957 if (n->is_Call()) {
2958 // tty->print_cr("optimize barrier on call");
2959 return false;
2960 }
2961
2962 const Type* type = phase->type(n);
2963 if (type == Type::TOP) {
2964 return false;
2965 }
2966 if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) {
2967 // tty->print_cr("optimize barrier on null");
2968 return false;
2969 }
2970 if (type->make_oopptr() && type->make_oopptr()->const_oop() != NULL) {
2971 // tty->print_cr("optimize barrier on constant");
2972 return false;
2973 }
2974
2975 switch (n->Opcode()) {
2976 case Op_AddP:
2977 return true; // TODO: Can refine?
2978 case Op_LoadP:
2979 case Op_ShenandoahCompareAndExchangeN:
2980 case Op_ShenandoahCompareAndExchangeP:
2981 case Op_CompareAndExchangeN:
2982 case Op_CompareAndExchangeP:
2983 case Op_GetAndSetN:
2984 case Op_GetAndSetP:
2985 return true;
2986 case Op_Phi: {
2987 for (uint i = 1; i < n->req(); i++) {
2988 if (needs_barrier_impl(phase, n->in(i), visited)) return true;
2989 }
2990 return false;
2991 }
2992 case Op_CheckCastPP:
2993 case Op_CastPP:
2994 return needs_barrier_impl(phase, n->in(1), visited);
2995 case Op_Proj:
2996 return needs_barrier_impl(phase, n->in(0), visited);
2997 case Op_ShenandoahLoadReferenceBarrier:
2998 // tty->print_cr("optimize barrier on barrier");
2999 return false;
3000 case Op_Parm:
3001 // tty->print_cr("optimize barrier on input arg");
3002 return false;
3003 case Op_DecodeN:
3004 case Op_EncodeP:
3005 return needs_barrier_impl(phase, n->in(1), visited);
3006 case Op_LoadN:
3007 return true;
3008 case Op_CMoveN:
3009 case Op_CMoveP:
3010 return needs_barrier_impl(phase, n->in(2), visited) ||
3011 needs_barrier_impl(phase, n->in(3), visited);
3012 case Op_ShenandoahEnqueueBarrier:
3013 return needs_barrier_impl(phase, n->in(1), visited);
3014 case Op_CreateEx:
3015 return false;
3016 default:
3017 break;
3018 }
3019 #ifdef ASSERT
3020 tty->print("need barrier on?: ");
3021 tty->print_cr("ins:");
3022 n->dump(2);
3023 tty->print_cr("outs:");
3024 n->dump(-2);
3025 ShouldNotReachHere();
3026 #endif
3027 return true;
3028 }
3029
3030 bool ShenandoahLoadReferenceBarrierNode::is_redundant() {
3031 Unique_Node_List visited;
3032 Node_Stack stack(0);
3033 stack.push(this, 0);
3034
3035 // Check if the barrier is actually useful: go over nodes looking for useful uses
3036 // (e.g. memory accesses). Stop once we detected a required use. Otherwise, walk
3037 // until we ran out of nodes, and then declare the barrier redundant.
3038 while (stack.size() > 0) {
3039 Node* n = stack.node();
3040 if (visited.member(n)) {
3041 stack.pop();
3042 continue;
3043 }
3044 visited.push(n);
3045 bool visit_users = false;
3046 switch (n->Opcode()) {
3047 case Op_CallStaticJava:
3048 case Op_CallDynamicJava:
3049 case Op_CallLeaf:
3050 case Op_CallLeafNoFP:
3051 case Op_CompareAndSwapL:
3052 case Op_CompareAndSwapI:
3053 case Op_CompareAndSwapB:
3054 case Op_CompareAndSwapS:
3055 case Op_CompareAndSwapN:
3056 case Op_CompareAndSwapP:
3057 case Op_CompareAndExchangeL:
3058 case Op_CompareAndExchangeI:
3059 case Op_CompareAndExchangeB:
3060 case Op_CompareAndExchangeS:
3061 case Op_CompareAndExchangeN:
3062 case Op_CompareAndExchangeP:
3063 case Op_WeakCompareAndSwapL:
3064 case Op_WeakCompareAndSwapI:
3065 case Op_WeakCompareAndSwapB:
3066 case Op_WeakCompareAndSwapS:
3067 case Op_WeakCompareAndSwapN:
3068 case Op_WeakCompareAndSwapP:
3069 case Op_ShenandoahCompareAndSwapN:
3070 case Op_ShenandoahCompareAndSwapP:
3071 case Op_ShenandoahWeakCompareAndSwapN:
3072 case Op_ShenandoahWeakCompareAndSwapP:
3073 case Op_ShenandoahCompareAndExchangeN:
3074 case Op_ShenandoahCompareAndExchangeP:
3075 case Op_GetAndSetL:
3076 case Op_GetAndSetI:
3077 case Op_GetAndSetB:
3078 case Op_GetAndSetS:
3079 case Op_GetAndSetP:
3080 case Op_GetAndSetN:
3081 case Op_GetAndAddL:
3082 case Op_GetAndAddI:
3083 case Op_GetAndAddB:
3084 case Op_GetAndAddS:
3085 case Op_ShenandoahEnqueueBarrier:
3086 case Op_FastLock:
3087 case Op_FastUnlock:
3088 case Op_Rethrow:
3089 case Op_Return:
3090 case Op_StoreB:
3091 case Op_StoreC:
3092 case Op_StoreD:
3093 case Op_StoreF:
3094 case Op_StoreL:
3095 case Op_StoreLConditional:
3096 case Op_StoreI:
3097 case Op_StoreIConditional:
3098 case Op_StoreN:
3099 case Op_StoreP:
3100 case Op_StoreVector:
3101 case Op_StrInflatedCopy:
3102 case Op_StrCompressedCopy:
3103 case Op_EncodeP:
3104 case Op_CastP2X:
3105 case Op_SafePoint:
3106 case Op_EncodeISOArray:
3107 case Op_AryEq:
3108 case Op_StrEquals:
3109 case Op_StrComp:
3110 case Op_StrIndexOf:
3111 case Op_StrIndexOfChar:
3112 case Op_HasNegatives:
3113 // Known to require barriers
3114 return false;
3115 case Op_CmpP: {
3116 if (n->in(1)->bottom_type()->higher_equal(TypePtr::NULL_PTR) ||
3117 n->in(2)->bottom_type()->higher_equal(TypePtr::NULL_PTR)) {
3118 // One of the sides is known null, no need for barrier.
3119 } else {
3120 return false;
3121 }
3122 break;
3123 }
3124 case Op_LoadB:
3125 case Op_LoadUB:
3126 case Op_LoadUS:
3127 case Op_LoadD:
3128 case Op_LoadF:
3129 case Op_LoadL:
3130 case Op_LoadI:
3131 case Op_LoadS:
3132 case Op_LoadN:
3133 case Op_LoadP:
3134 case Op_LoadVector: {
3135 const TypePtr* adr_type = n->adr_type();
3136 int alias_idx = Compile::current()->get_alias_index(adr_type);
3137 Compile::AliasType* alias_type = Compile::current()->alias_type(alias_idx);
3138 ciField* field = alias_type->field();
3139 bool is_static = field != NULL && field->is_static();
3140 bool is_final = field != NULL && field->is_final();
3141
3142 if (ShenandoahOptimizeStaticFinals && is_static && is_final) {
3143 // Loading the constant does not require barriers: it should be handled
3144 // as part of GC roots already.
3145 } else {
3146 return false;
3147 }
3148 break;
3149 }
3150 case Op_Conv2B:
3151 case Op_LoadRange:
3152 case Op_LoadKlass:
3153 case Op_LoadNKlass:
3154 // Do not require barriers
3155 break;
3156 case Op_AddP:
3157 case Op_CheckCastPP:
3158 case Op_CastPP:
3159 case Op_CMoveP:
3160 case Op_Phi:
3161 case Op_ShenandoahLoadReferenceBarrier:
3162 // Whether or not these need the barriers depends on their users
3163 visit_users = true;
3164 break;
3165 default: {
3166 #ifdef ASSERT
3167 fatal("Unknown node in is_redundant: %s", NodeClassNames[n->Opcode()]);
3168 #else
3169 // Default to have excess barriers, rather than miss some.
3170 return false;
3171 #endif
3172 }
3173 }
3174
3175 stack.pop();
3176 if (visit_users) {
3177 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3178 Node* user = n->fast_out(i);
3179 if (user != NULL) {
3180 stack.push(user, 0);
3181 }
3182 }
3183 }
3184 }
3185
3186 // No need for barrier found.
3187 return true;
3188 }