1
2 /*
3 * Copyright (c) 2015, 2019, Red Hat, Inc. All rights reserved.
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_implementation/shenandoah/shenandoahForwarding.hpp"
28 #include "gc_implementation/shenandoah/shenandoahHeap.hpp"
29 #include "gc_implementation/shenandoah/shenandoahHeapRegion.hpp"
30 #include "gc_implementation/shenandoah/shenandoahRuntime.hpp"
31 #include "gc_implementation/shenandoah/c2/shenandoahSupport.hpp"
32 #include "gc_implementation/shenandoah/c2/shenandoahBarrierSetC2.hpp"
33 #include "opto/block.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/phaseX.hpp"
36 #include "opto/rootnode.hpp"
37 #include "opto/runtime.hpp"
38 #include "opto/subnode.hpp"
39
40 #ifdef _LP64
41 #define LoadXNode LoadLNode
42 #else
43 #define LoadXNode LoadINode
44 #endif
45
46 bool ShenandoahBarrierC2Support::expand(Compile* C, PhaseIterGVN& igvn) {
47 if (C->shenandoah_barriers_count() > 0) {
48 C->clear_major_progress();
49 PhaseIdealLoop ideal_loop(igvn, false, true);
50 if (C->failing()) return false;
51 PhaseIdealLoop::verify(igvn);
52 DEBUG_ONLY(verify_raw_mem(C->root());)
53 }
54 return true;
55 }
56
57 bool ShenandoahBarrierC2Support::is_gc_state_test(Node* iff, int mask) {
58 if (!UseShenandoahGC) {
59 return false;
60 }
61 assert(iff->is_If(), "bad input");
62 if (iff->Opcode() != Op_If) {
63 return false;
64 }
65 Node* bol = iff->in(1);
66 if (!bol->is_Bool() || bol->as_Bool()->_test._test != BoolTest::ne) {
67 return false;
68 }
69 Node* cmp = bol->in(1);
70 if (cmp->Opcode() != Op_CmpI) {
71 return false;
72 }
73 Node* in1 = cmp->in(1);
74 Node* in2 = cmp->in(2);
75 if (in2->find_int_con(-1) != 0) {
76 return false;
77 }
78 if (in1->Opcode() != Op_AndI) {
79 return false;
80 }
81 in2 = in1->in(2);
82 if (in2->find_int_con(-1) != mask) {
83 return false;
84 }
85 in1 = in1->in(1);
86
87 return is_gc_state_load(in1);
88 }
89
90 bool ShenandoahBarrierC2Support::is_heap_stable_test(Node* iff) {
91 return is_gc_state_test(iff, ShenandoahHeap::HAS_FORWARDED);
92 }
93
94 bool ShenandoahBarrierC2Support::is_gc_state_load(Node *n) {
95 if (!UseShenandoahGC) {
96 return false;
97 }
98 if (n->Opcode() != Op_LoadB && n->Opcode() != Op_LoadUB) {
99 return false;
100 }
101 Node* addp = n->in(MemNode::Address);
102 if (!addp->is_AddP()) {
103 return false;
104 }
105 Node* base = addp->in(AddPNode::Address);
106 Node* off = addp->in(AddPNode::Offset);
107 if (base->Opcode() != Op_ThreadLocal) {
108 return false;
109 }
110 if (off->find_intptr_t_con(-1) != in_bytes(JavaThread::gc_state_offset())) {
111 return false;
112 }
113 return true;
114 }
115
116 bool ShenandoahBarrierC2Support::has_safepoint_between(Node* start, Node* stop, PhaseIdealLoop *phase) {
117 assert(phase->is_dominator(stop, start), "bad inputs");
118 ResourceMark rm;
119 Unique_Node_List wq;
120 wq.push(start);
121 for (uint next = 0; next < wq.size(); next++) {
122 Node *m = wq.at(next);
123 if (m == stop) {
124 continue;
125 }
126 if (m->is_SafePoint() && !m->is_CallLeaf()) {
127 return true;
128 }
129 if (m->is_Region()) {
130 for (uint i = 1; i < m->req(); i++) {
131 wq.push(m->in(i));
132 }
133 } else {
134 wq.push(m->in(0));
135 }
136 }
137 return false;
138 }
139
140 #ifdef ASSERT
141 bool ShenandoahBarrierC2Support::verify_helper(Node* in, Node_Stack& phis, VectorSet& visited, verify_type t, bool trace, Unique_Node_List& barriers_used) {
142 assert(phis.size() == 0, "");
143
144 while (true) {
145 if (in->bottom_type() == TypePtr::NULL_PTR) {
146 if (trace) {tty->print_cr("NULL");}
147 } else if (!in->bottom_type()->make_ptr()->make_oopptr()) {
148 if (trace) {tty->print_cr("Non oop");}
149 } else if (in->bottom_type()->make_ptr()->make_oopptr() == TypeInstPtr::MIRROR) {
150 if (trace) {tty->print_cr("Java mirror");}
151 } else {
152 if (in->is_ConstraintCast() || in->Opcode() == Op_CheckCastPP) {
153 in = in->in(1);
154 continue;
155 } else if (in->is_AddP()) {
156 assert(!in->in(AddPNode::Address)->is_top(), "no raw memory access");
157 in = in->in(AddPNode::Address);
158 continue;
159 } else if (in->is_Con()) {
160 if (trace) {
161 tty->print("Found constant");
162 in->dump();
163 }
164 } else if (in->Opcode() == Op_Parm) {
165 if (trace) {
166 tty->print("Found argument");
167 }
168 } else if (in->Opcode() == Op_CreateEx) {
169 if (trace) {
170 tty->print("Found create-exception");
171 }
172 } else if (in->Opcode() == Op_LoadP && in->adr_type() == TypeRawPtr::BOTTOM) {
173 if (trace) {
174 tty->print("Found raw LoadP (OSR argument?)");
175 }
176 } else if (in->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
177 if (t == ShenandoahOopStore) {
178 uint i = 0;
179 for (; i < phis.size(); i++) {
180 Node* n = phis.node_at(i);
181 }
182 if (i == phis.size()) {
183 return false;
184 }
185 }
186 barriers_used.push(in);
187 if (trace) {tty->print("Found barrier"); in->dump();}
188 } else if (in->is_Proj() && in->in(0)->is_Allocate()) {
189 if (trace) {
190 tty->print("Found alloc");
191 in->in(0)->dump();
192 }
193 } else if (in->is_Proj() && (in->in(0)->Opcode() == Op_CallStaticJava || in->in(0)->Opcode() == Op_CallDynamicJava)) {
194 if (trace) {
195 tty->print("Found Java call");
196 }
197 } else if (in->is_Phi()) {
198 if (!visited.test_set(in->_idx)) {
199 if (trace) {tty->print("Pushed phi:"); in->dump();}
200 phis.push(in, 2);
201 in = in->in(1);
202 continue;
203 }
204 if (trace) {tty->print("Already seen phi:"); in->dump();}
205 } else if (in->Opcode() == Op_CMoveP || in->Opcode() == Op_CMoveN) {
206 if (!visited.test_set(in->_idx)) {
207 if (trace) {tty->print("Pushed cmovep:"); in->dump();}
208 phis.push(in, CMoveNode::IfTrue);
209 in = in->in(CMoveNode::IfFalse);
210 continue;
211 }
212 if (trace) {tty->print("Already seen cmovep:"); in->dump();}
213 } else if (in->Opcode() == Op_EncodeP || in->Opcode() == Op_DecodeN) {
214 in = in->in(1);
215 continue;
216 } else {
217 return false;
218 }
219 }
220 bool cont = false;
221 while (phis.is_nonempty()) {
222 uint idx = phis.index();
223 Node* phi = phis.node();
224 if (idx >= phi->req()) {
225 if (trace) {tty->print("Popped phi:"); phi->dump();}
226 phis.pop();
227 continue;
228 }
229 if (trace) {tty->print("Next entry(%d) for phi:", idx); phi->dump();}
230 in = phi->in(idx);
231 phis.set_index(idx+1);
232 cont = true;
233 break;
234 }
235 if (!cont) {
236 break;
237 }
238 }
239 return true;
240 }
241
242 void ShenandoahBarrierC2Support::report_verify_failure(const char* msg, Node* n1, Node* n2) {
243 if (n1 != NULL) {
244 n1->dump(+10);
245 }
246 if (n2 != NULL) {
247 n2->dump(+10);
248 }
249 fatal(err_msg("%s", msg));
250 }
251
252 static const char* call_name(CallNode* call) {
253 if (call->is_CallRuntime()) {
254 return call->as_CallRuntime()->_name;
255 }
256 if (call->is_CallStaticJava()) {
257 return call->as_CallStaticJava()->_name;
258 }
259 return NULL;
260 }
261
262 void ShenandoahBarrierC2Support::verify(RootNode* root) {
263 ResourceMark rm;
264 Unique_Node_List wq;
265 GrowableArray<Node*> barriers;
266 Unique_Node_List barriers_used;
267 Node_Stack phis(0);
268 VectorSet visited(Thread::current()->resource_area());
269 const bool trace = true;
270 const bool verify_no_useless_barrier = false;
271
272 wq.push(root);
273 for (uint next = 0; next < wq.size(); next++) {
274 Node *n = wq.at(next);
275 if (n->is_Load()) {
276 const bool trace = false;
277 if (trace) {tty->print("Verifying"); n->dump();}
278 if (n->Opcode() == Op_LoadRange || n->Opcode() == Op_LoadKlass || n->Opcode() == Op_LoadNKlass) {
279 if (trace) {tty->print_cr("Load range/klass");}
280 } else {
281 const TypePtr* adr_type = n->as_Load()->adr_type();
282
283 if (adr_type->isa_oopptr() && adr_type->is_oopptr()->offset() == oopDesc::mark_offset_in_bytes()) {
284 if (trace) {tty->print_cr("Mark load");}
285 } else if (adr_type->isa_instptr() &&
286 adr_type->is_instptr()->klass()->is_subtype_of(Compile::current()->env()->Reference_klass()) &&
287 adr_type->is_instptr()->offset() == java_lang_ref_Reference::referent_offset) {
288 if (trace) {tty->print_cr("Reference.get()");}
289 } else if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahLoad, trace, barriers_used)) {
290 report_verify_failure("Shenandoah verification: Load should have barriers", n);
291 }
292 }
293 } else if (n->is_Store()) {
294 const bool trace = false;
295
296 if (trace) {tty->print("Verifying"); n->dump();}
297 if (n->in(MemNode::ValueIn)->bottom_type()->make_oopptr()) {
298 Node* adr = n->in(MemNode::Address);
299 bool verify = true;
300
301 if (adr->is_AddP() && adr->in(AddPNode::Base)->is_top()) {
302 adr = adr->in(AddPNode::Address);
303 if (adr->is_AddP()) {
304 assert(adr->in(AddPNode::Base)->is_top(), "");
305 adr = adr->in(AddPNode::Address);
306 if (adr->Opcode() == Op_LoadP &&
307 adr->in(MemNode::Address)->in(AddPNode::Base)->is_top() &&
308 adr->in(MemNode::Address)->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
309 adr->in(MemNode::Address)->in(AddPNode::Offset)->find_intptr_t_con(-1) == in_bytes(JavaThread::satb_mark_queue_offset()) + in_bytes(PtrQueue::byte_offset_of_buf())) {
310 if (trace) {tty->print_cr("SATB prebarrier");}
311 verify = false;
312 }
313 }
314 }
315
316 if (verify && !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahValue, trace, barriers_used)) {
317 report_verify_failure("Shenandoah verification: Store should have barriers", n);
318 }
319 }
320 if (!verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
321 report_verify_failure("Shenandoah verification: Store (address) should have barriers", n);
322 }
323 } else if (n->Opcode() == Op_CmpP) {
324 const bool trace = false;
325
326 Node* in1 = n->in(1);
327 Node* in2 = n->in(2);
328 if (in1->bottom_type()->isa_oopptr()) {
329 if (trace) {tty->print("Verifying"); n->dump();}
330
331 bool mark_inputs = false;
332 if (in1->bottom_type() == TypePtr::NULL_PTR || in2->bottom_type() == TypePtr::NULL_PTR ||
333 (in1->is_Con() || in2->is_Con())) {
334 if (trace) {tty->print_cr("Comparison against a constant");}
335 mark_inputs = true;
336 } else if ((in1->is_CheckCastPP() && in1->in(1)->is_Proj() && in1->in(1)->in(0)->is_Allocate()) ||
337 (in2->is_CheckCastPP() && in2->in(1)->is_Proj() && in2->in(1)->in(0)->is_Allocate())) {
338 if (trace) {tty->print_cr("Comparison with newly alloc'ed object");}
339 mark_inputs = true;
340 } else {
341 assert(in2->bottom_type()->isa_oopptr(), "");
342
343 if (!verify_helper(in1, phis, visited, ShenandoahStore, trace, barriers_used) ||
344 !verify_helper(in2, phis, visited, ShenandoahStore, trace, barriers_used)) {
345 report_verify_failure("Shenandoah verification: Cmp should have barriers", n);
346 }
347 }
348 if (verify_no_useless_barrier &&
349 mark_inputs &&
350 (!verify_helper(in1, phis, visited, ShenandoahValue, trace, barriers_used) ||
351 !verify_helper(in2, phis, visited, ShenandoahValue, trace, barriers_used))) {
352 phis.clear();
353 visited.Reset();
354 }
355 }
356 } else if (n->is_LoadStore()) {
357 if (n->in(MemNode::ValueIn)->bottom_type()->make_ptr() &&
358 !verify_helper(n->in(MemNode::ValueIn), phis, visited, ShenandoahValue, trace, barriers_used)) {
359 report_verify_failure("Shenandoah verification: LoadStore (value) should have barriers", n);
360 }
361
362 if (n->in(MemNode::Address)->bottom_type()->make_oopptr() && !verify_helper(n->in(MemNode::Address), phis, visited, ShenandoahStore, trace, barriers_used)) {
363 report_verify_failure("Shenandoah verification: LoadStore (address) should have barriers", n);
364 }
365 } else if (n->Opcode() == Op_CallLeafNoFP || n->Opcode() == Op_CallLeaf) {
366 CallNode* call = n->as_Call();
367
368 static struct {
369 const char* name;
370 struct {
371 int pos;
372 verify_type t;
373 } args[6];
374 } calls[] = {
375 "aescrypt_encryptBlock",
376 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
377 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
378 "aescrypt_decryptBlock",
379 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
380 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
381 "multiplyToLen",
382 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { TypeFunc::Parms+4, ShenandoahStore },
383 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
384 "squareToLen",
385 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad }, { -1, ShenandoahNone},
386 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
387 "montgomery_multiply",
388 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
389 { TypeFunc::Parms+6, ShenandoahStore }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
390 "montgomery_square",
391 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+5, ShenandoahStore },
392 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
393 "mulAdd",
394 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
395 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
396 "vectorizedMismatch",
397 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone},
398 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
399 "updateBytesCRC32",
400 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
401 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
402 "updateBytesAdler32",
403 { { TypeFunc::Parms+1, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
404 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
405 "updateBytesCRC32C",
406 { { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahLoad}, { -1, ShenandoahNone},
407 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
408 "counterMode_AESCrypt",
409 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
410 { TypeFunc::Parms+3, ShenandoahStore }, { TypeFunc::Parms+5, ShenandoahStore }, { TypeFunc::Parms+6, ShenandoahStore } },
411 "cipherBlockChaining_encryptAESCrypt",
412 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
413 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
414 "cipherBlockChaining_decryptAESCrypt",
415 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { TypeFunc::Parms+2, ShenandoahLoad },
416 { TypeFunc::Parms+3, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
417 "shenandoah_clone_barrier",
418 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone}, { -1, ShenandoahNone},
419 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
420 "ghash_processBlocks",
421 { { TypeFunc::Parms, ShenandoahStore }, { TypeFunc::Parms+1, ShenandoahLoad }, { TypeFunc::Parms+2, ShenandoahLoad },
422 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
423 "sha1_implCompress",
424 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
425 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
426 "sha256_implCompress",
427 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
428 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
429 "sha512_implCompress",
430 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
431 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
432 "sha1_implCompressMB",
433 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
434 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
435 "sha256_implCompressMB",
436 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
437 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
438 "sha512_implCompressMB",
439 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+1, ShenandoahStore }, { -1, ShenandoahNone },
440 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
441 "encodeBlock",
442 { { TypeFunc::Parms, ShenandoahLoad }, { TypeFunc::Parms+3, ShenandoahStore }, { -1, ShenandoahNone },
443 { -1, ShenandoahNone}, { -1, ShenandoahNone}, { -1, ShenandoahNone} },
444 };
445
446 if (call->is_CallRuntime() && call->as_CallRuntime()->is_call_to_arraycopystub()) {
447 Node* dest = NULL;
448 const TypeTuple* args = n->as_Call()->_tf->domain();
449 for (uint i = TypeFunc::Parms, j = 0; i < args->cnt(); i++) {
450 if (args->field_at(i)->isa_ptr()) {
451 j++;
452 if (j == 2) {
453 dest = n->in(i);
454 break;
455 }
456 }
457 }
458 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahLoad, trace, barriers_used) ||
459 !verify_helper(dest, phis, visited, ShenandoahStore, trace, barriers_used)) {
460 report_verify_failure("Shenandoah verification: ArrayCopy should have barriers", n);
461 }
462 } else if (strlen(call_name(call)) > 5 &&
463 !strcmp(call_name(call) + strlen(call_name(call)) - 5, "_fill")) {
464 if (!verify_helper(n->in(TypeFunc::Parms), phis, visited, ShenandoahStore, trace, barriers_used)) {
465 report_verify_failure("Shenandoah verification: _fill should have barriers", n);
466 }
467 } else if (!strcmp(call_name(call), "g1_wb_pre")) {
468 // skip
469 } else {
470 const int calls_len = sizeof(calls) / sizeof(calls[0]);
471 int i = 0;
472 for (; i < calls_len; i++) {
473 if (!strcmp(calls[i].name, call_name(call))) {
474 break;
475 }
476 }
477 if (i != calls_len) {
478 const uint args_len = sizeof(calls[0].args) / sizeof(calls[0].args[0]);
479 for (uint j = 0; j < args_len; j++) {
480 int pos = calls[i].args[j].pos;
481 if (pos == -1) {
482 break;
483 }
484 if (!verify_helper(call->in(pos), phis, visited, calls[i].args[j].t, trace, barriers_used)) {
485 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
486 }
487 }
488 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
489 if (call->in(j)->bottom_type()->make_ptr() &&
490 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
491 uint k = 0;
492 for (; k < args_len && calls[i].args[k].pos != (int)j; k++);
493 if (k == args_len) {
494 fatal(err_msg("arg %d for call %s not covered", j, call_name(call)));
495 }
496 }
497 }
498 } else {
499 for (uint j = TypeFunc::Parms; j < call->req(); j++) {
500 if (call->in(j)->bottom_type()->make_ptr() &&
501 call->in(j)->bottom_type()->make_ptr()->isa_oopptr()) {
502 fatal(err_msg("%s not covered", call_name(call)));
503 }
504 }
505 }
506 }
507 } else if (n->Opcode() == Op_ShenandoahLoadReferenceBarrier) {
508 // skip
509 } else if (n->is_AddP()
510 || n->is_Phi()
511 || n->is_ConstraintCast()
512 || n->Opcode() == Op_CheckCastPP
513 || n->Opcode() == Op_Return
514 || n->Opcode() == Op_CMoveP
515 || n->Opcode() == Op_CMoveN
516 || n->Opcode() == Op_Rethrow
517 || n->is_MemBar()
518 || n->Opcode() == Op_Conv2B
519 || n->Opcode() == Op_SafePoint
520 || n->is_CallJava()
521 || n->Opcode() == Op_Unlock
522 || n->Opcode() == Op_EncodeP
523 || n->Opcode() == Op_DecodeN) {
524 // nothing to do
525 } else {
526 static struct {
527 int opcode;
528 struct {
529 int pos;
530 verify_type t;
531 } inputs[2];
532 } others[] = {
533 Op_FastLock,
534 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
535 Op_Lock,
536 { { TypeFunc::Parms, ShenandoahLoad }, { -1, ShenandoahNone} },
537 Op_AryEq,
538 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
539 Op_StrIndexOf,
540 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
541 Op_StrComp,
542 { { 2, ShenandoahLoad }, { 4, ShenandoahLoad } },
543 Op_StrEquals,
544 { { 2, ShenandoahLoad }, { 3, ShenandoahLoad } },
545 Op_EncodeISOArray,
546 { { 2, ShenandoahLoad }, { 3, ShenandoahStore } },
547 Op_CastP2X,
548 { { 1, ShenandoahLoad }, { -1, ShenandoahNone} },
549 Op_ClearArray,
550 { { 3, ShenandoahStore }, { -1, ShenandoahNone} },
551 };
552
553 const int others_len = sizeof(others) / sizeof(others[0]);
554 int i = 0;
555 for (; i < others_len; i++) {
556 if (others[i].opcode == n->Opcode()) {
557 break;
558 }
559 }
560 uint stop = n->is_Call() ? n->as_Call()->tf()->domain()->cnt() : n->req();
561 if (i != others_len) {
562 const uint inputs_len = sizeof(others[0].inputs) / sizeof(others[0].inputs[0]);
563 for (uint j = 0; j < inputs_len; j++) {
564 int pos = others[i].inputs[j].pos;
565 if (pos == -1) {
566 break;
567 }
568 if (!verify_helper(n->in(pos), phis, visited, others[i].inputs[j].t, trace, barriers_used)) {
569 report_verify_failure("Shenandoah verification: intrinsic calls should have barriers", n);
570 }
571 }
572 for (uint j = 1; j < stop; j++) {
573 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
574 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
575 uint k = 0;
576 for (; k < inputs_len && others[i].inputs[k].pos != (int)j; k++);
577 if (k == inputs_len) {
578 fatal(err_msg("arg %d for node %s not covered", j, n->Name()));
579 }
580 }
581 }
582 } else {
583 for (uint j = 1; j < stop; j++) {
584 if (n->in(j) != NULL && n->in(j)->bottom_type()->make_ptr() &&
585 n->in(j)->bottom_type()->make_ptr()->make_oopptr()) {
586 fatal(err_msg("%s not covered", n->Name()));
587 }
588 }
589 }
590 }
591
592 if (n->is_SafePoint()) {
593 SafePointNode* sfpt = n->as_SafePoint();
594 if (verify_no_useless_barrier && sfpt->jvms() != NULL) {
595 for (uint i = sfpt->jvms()->scloff(); i < sfpt->jvms()->endoff(); i++) {
596 if (!verify_helper(sfpt->in(i), phis, visited, ShenandoahLoad, trace, barriers_used)) {
597 phis.clear();
598 visited.Reset();
599 }
600 }
601 }
602 }
603 }
604
605 if (verify_no_useless_barrier) {
606 for (int i = 0; i < barriers.length(); i++) {
607 Node* n = barriers.at(i);
608 if (!barriers_used.member(n)) {
609 tty->print("XXX useless barrier"); n->dump(-2);
610 ShouldNotReachHere();
611 }
612 }
613 }
614 }
615 #endif
616
617 bool ShenandoahBarrierC2Support::is_dominator_same_ctrl(Node* c, Node* d, Node* n, PhaseIdealLoop* phase) {
618 // That both nodes have the same control is not sufficient to prove
619 // domination, verify that there's no path from d to n
620 ResourceMark rm;
621 Unique_Node_List wq;
622 wq.push(d);
623 for (uint next = 0; next < wq.size(); next++) {
624 Node *m = wq.at(next);
625 if (m == n) {
626 return false;
627 }
628 if (m->is_Phi() && m->in(0)->is_Loop()) {
629 assert(phase->ctrl_or_self(m->in(LoopNode::EntryControl)) != c, "following loop entry should lead to new control");
630 } else {
631 if (m->is_Store() || m->is_LoadStore()) {
632 // Take anti-dependencies into account
633 Node* mem = m->in(MemNode::Memory);
634 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) {
635 Node* u = mem->fast_out(i);
636 if (u->is_Load() && phase->C->can_alias(m->adr_type(), phase->C->get_alias_index(u->adr_type())) &&
637 phase->ctrl_or_self(u) == c) {
638 wq.push(u);
639 }
640 }
641 }
642 for (uint i = 0; i < m->req(); i++) {
643 if (m->in(i) != NULL && phase->ctrl_or_self(m->in(i)) == c) {
644 wq.push(m->in(i));
645 }
646 }
647 }
648 }
649 return true;
650 }
651
652 bool ShenandoahBarrierC2Support::is_dominator(Node* d_c, Node* n_c, Node* d, Node* n, PhaseIdealLoop* phase) {
653 if (d_c != n_c) {
654 return phase->is_dominator(d_c, n_c);
655 }
656 return is_dominator_same_ctrl(d_c, d, n, phase);
657 }
658
659 Node* next_mem(Node* mem, int alias) {
660 Node* res = NULL;
661 if (mem->is_Proj()) {
662 res = mem->in(0);
663 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
664 res = mem->in(TypeFunc::Memory);
665 } else if (mem->is_Phi()) {
666 res = mem->in(1);
667 } else if (mem->is_MergeMem()) {
668 res = mem->as_MergeMem()->memory_at(alias);
669 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
670 assert(alias = Compile::AliasIdxRaw, "following raw memory can't lead to a barrier");
671 res = mem->in(MemNode::Memory);
672 } else {
673 #ifdef ASSERT
674 mem->dump();
675 #endif
676 ShouldNotReachHere();
677 }
678 return res;
679 }
680
681 Node* ShenandoahBarrierC2Support::no_branches(Node* c, Node* dom, bool allow_one_proj, PhaseIdealLoop* phase) {
682 Node* iffproj = NULL;
683 while (c != dom) {
684 Node* next = phase->idom(c);
685 assert(next->unique_ctrl_out() == c || c->is_Proj() || c->is_Region(), "multiple control flow out but no proj or region?");
686 if (c->is_Region()) {
687 ResourceMark rm;
688 Unique_Node_List wq;
689 wq.push(c);
690 for (uint i = 0; i < wq.size(); i++) {
691 Node *n = wq.at(i);
692 if (n == next) {
693 continue;
694 }
695 if (n->is_Region()) {
696 for (uint j = 1; j < n->req(); j++) {
697 wq.push(n->in(j));
698 }
699 } else {
700 wq.push(n->in(0));
701 }
702 }
703 for (uint i = 0; i < wq.size(); i++) {
704 Node *n = wq.at(i);
705 assert(n->is_CFG(), "");
706 if (n->is_Multi()) {
707 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
708 Node* u = n->fast_out(j);
709 if (u->is_CFG()) {
710 if (!wq.member(u) && !u->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
711 return NodeSentinel;
712 }
713 }
714 }
715 }
716 }
717 } else if (c->is_Proj()) {
718 if (c->is_IfProj()) {
719 if (c->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) != NULL) {
720 // continue;
721 } else {
722 if (!allow_one_proj) {
723 return NodeSentinel;
724 }
725 if (iffproj == NULL) {
726 iffproj = c;
727 } else {
728 return NodeSentinel;
729 }
730 }
731 } else if (c->Opcode() == Op_JumpProj) {
732 return NodeSentinel; // unsupported
733 } else if (c->Opcode() == Op_CatchProj) {
734 return NodeSentinel; // unsupported
735 } else if (c->Opcode() == Op_CProj && next->Opcode() == Op_NeverBranch) {
736 return NodeSentinel; // unsupported
737 } else {
738 assert(next->unique_ctrl_out() == c, "unsupported branch pattern");
739 }
740 }
741 c = next;
742 }
743 return iffproj;
744 }
745
746 Node* ShenandoahBarrierC2Support::dom_mem(Node* mem, Node* ctrl, int alias, Node*& mem_ctrl, PhaseIdealLoop* phase) {
747 ResourceMark rm;
748 VectorSet wq(Thread::current()->resource_area());
749 wq.set(mem->_idx);
750 mem_ctrl = phase->ctrl_or_self(mem);
751 while (!phase->is_dominator(mem_ctrl, ctrl) || mem_ctrl == ctrl) {
752 mem = next_mem(mem, alias);
753 if (wq.test_set(mem->_idx)) {
754 return NULL;
755 }
756 mem_ctrl = phase->ctrl_or_self(mem);
757 }
758 if (mem->is_MergeMem()) {
759 mem = mem->as_MergeMem()->memory_at(alias);
760 mem_ctrl = phase->ctrl_or_self(mem);
761 }
762 return mem;
763 }
764
765 Node* ShenandoahBarrierC2Support::find_bottom_mem(Node* ctrl, PhaseIdealLoop* phase) {
766 Node* mem = NULL;
767 Node* c = ctrl;
768 do {
769 if (c->is_Region()) {
770 Node* phi_bottom = NULL;
771 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax && mem == NULL; i++) {
772 Node* u = c->fast_out(i);
773 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
774 if (u->adr_type() == TypePtr::BOTTOM) {
775 mem = u;
776 }
777 }
778 }
779 } else {
780 if (c->is_Call() && c->as_Call()->adr_type() != NULL) {
781 CallProjections projs;
782 c->as_Call()->extract_projections(&projs, true, false);
783 if (projs.fallthrough_memproj != NULL) {
784 if (projs.fallthrough_memproj->adr_type() == TypePtr::BOTTOM) {
785 if (projs.catchall_memproj == NULL) {
786 mem = projs.fallthrough_memproj;
787 } else {
788 if (phase->is_dominator(projs.fallthrough_catchproj, ctrl)) {
789 mem = projs.fallthrough_memproj;
790 } else {
791 assert(phase->is_dominator(projs.catchall_catchproj, ctrl), "one proj must dominate barrier");
792 mem = projs.catchall_memproj;
793 }
794 }
795 }
796 } else {
797 Node* proj = c->as_Call()->proj_out(TypeFunc::Memory);
798 if (proj != NULL &&
799 proj->adr_type() == TypePtr::BOTTOM) {
800 mem = proj;
801 }
802 }
803 } else {
804 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
805 Node* u = c->fast_out(i);
806 if (u->is_Proj() &&
807 u->bottom_type() == Type::MEMORY &&
808 u->adr_type() == TypePtr::BOTTOM) {
809 assert(c->is_SafePoint() || c->is_MemBar() || c->is_Start(), "");
810 assert(mem == NULL, "only one proj");
811 mem = u;
812 }
813 }
814 assert(!c->is_Call() || c->as_Call()->adr_type() != NULL || mem == NULL, "no mem projection expected");
815 }
816 }
817 c = phase->idom(c);
818 } while (mem == NULL);
819 return mem;
820 }
821
822 void ShenandoahBarrierC2Support::follow_barrier_uses(Node* n, Node* ctrl, Unique_Node_List& uses, PhaseIdealLoop* phase) {
823 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
824 Node* u = n->fast_out(i);
825 if (!u->is_CFG() && phase->get_ctrl(u) == ctrl && (!u->is_Phi() || !u->in(0)->is_Loop() || u->in(LoopNode::LoopBackControl) != n)) {
826 uses.push(u);
827 }
828 }
829 }
830
831 void ShenandoahBarrierC2Support::test_gc_state(Node*& ctrl, Node* raw_mem, Node*& test_fail_ctrl,
832 PhaseIdealLoop* phase, int flags) {
833 PhaseIterGVN& igvn = phase->igvn();
834 Node* old_ctrl = ctrl;
835
836 Node* thread = new (phase->C) ThreadLocalNode();
837 Node* gc_state_offset = igvn.MakeConX(in_bytes(JavaThread::gc_state_offset()));
838 Node* gc_state_addr = new (phase->C) AddPNode(phase->C->top(), thread, gc_state_offset);
839 Node* gc_state = new (phase->C) LoadBNode(old_ctrl, raw_mem, gc_state_addr,
840 DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(NULL),
841 TypeInt::BYTE, MemNode::unordered);
842
843 Node* gc_state_and = new (phase->C) AndINode(gc_state, igvn.intcon(flags));
844 Node* gc_state_cmp = new (phase->C) CmpINode(gc_state_and, igvn.zerocon(T_INT));
845 Node* gc_state_bool = new (phase->C) BoolNode(gc_state_cmp, BoolTest::ne);
846
847 IfNode* gc_state_iff = new (phase->C) IfNode(old_ctrl, gc_state_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
848 ctrl = new (phase->C) IfTrueNode(gc_state_iff);
849 test_fail_ctrl = new (phase->C) IfFalseNode(gc_state_iff);
850
851 IdealLoopTree* loop = phase->get_loop(old_ctrl);
852 phase->register_control(gc_state_iff, loop, old_ctrl);
853 phase->register_control(ctrl, loop, gc_state_iff);
854 phase->register_control(test_fail_ctrl, loop, gc_state_iff);
855
856 phase->register_new_node(thread, old_ctrl);
857 phase->register_new_node(gc_state_addr, old_ctrl);
858 phase->register_new_node(gc_state, old_ctrl);
859 phase->register_new_node(gc_state_and, old_ctrl);
860 phase->register_new_node(gc_state_cmp, old_ctrl);
861 phase->register_new_node(gc_state_bool, old_ctrl);
862
863 phase->set_ctrl(gc_state_offset, phase->C->root());
864 assert(is_gc_state_test(gc_state_iff, flags), "Should match the shape");
865 }
866
867 void ShenandoahBarrierC2Support::test_null(Node*& ctrl, Node* val, Node*& null_ctrl, PhaseIdealLoop* phase) {
868 Node* old_ctrl = ctrl;
869 PhaseIterGVN& igvn = phase->igvn();
870
871 const Type* val_t = igvn.type(val);
872 if (val_t->meet(TypePtr::NULL_PTR) == val_t) {
873 Node* null_cmp = new (phase->C) CmpPNode(val, igvn.zerocon(T_OBJECT));
874 Node* null_test = new (phase->C) BoolNode(null_cmp, BoolTest::ne);
875
876 IfNode* null_iff = new (phase->C) IfNode(old_ctrl, null_test, PROB_LIKELY(0.999), COUNT_UNKNOWN);
877 ctrl = new (phase->C) IfTrueNode(null_iff);
878 null_ctrl = new (phase->C) IfFalseNode(null_iff);
879
880 IdealLoopTree* loop = phase->get_loop(old_ctrl);
881 phase->register_control(null_iff, loop, old_ctrl);
882 phase->register_control(ctrl, loop, null_iff);
883 phase->register_control(null_ctrl, loop, null_iff);
884
885 phase->register_new_node(null_cmp, old_ctrl);
886 phase->register_new_node(null_test, old_ctrl);
887 }
888 }
889
890 Node* ShenandoahBarrierC2Support::clone_null_check(Node*& c, Node* val, Node* unc_ctrl, PhaseIdealLoop* phase) {
891 IdealLoopTree *loop = phase->get_loop(c);
892 Node* iff = unc_ctrl->in(0);
893 assert(iff->is_If(), "broken");
894 Node* new_iff = iff->clone();
895 new_iff->set_req(0, c);
896 phase->register_control(new_iff, loop, c);
897 Node* iffalse = new (phase->C) IfFalseNode(new_iff->as_If());
898 phase->register_control(iffalse, loop, new_iff);
899 Node* iftrue = new (phase->C) IfTrueNode(new_iff->as_If());
900 phase->register_control(iftrue, loop, new_iff);
901 c = iftrue;
902 const Type *t = phase->igvn().type(val);
903 assert(val->Opcode() == Op_CastPP, "expect cast to non null here");
904 Node* uncasted_val = val->in(1);
905 val = new (phase->C) CastPPNode(uncasted_val, t);
906 val->init_req(0, c);
907 phase->register_new_node(val, c);
908 return val;
909 }
910
911 void ShenandoahBarrierC2Support::fix_null_check(Node* unc, Node* unc_ctrl, Node* new_unc_ctrl,
912 Unique_Node_List& uses, PhaseIdealLoop* phase) {
913 IfNode* iff = unc_ctrl->in(0)->as_If();
914 Node* proj = iff->proj_out(0);
915 assert(proj != unc_ctrl, "bad projection");
916 Node* use = proj->unique_ctrl_out();
917
918 assert(use == unc || use->is_Region(), "what else?");
919
920 uses.clear();
921 if (use == unc) {
922 phase->set_idom(use, new_unc_ctrl, phase->dom_depth(use));
923 for (uint i = 1; i < unc->req(); i++) {
924 Node* n = unc->in(i);
925 if (phase->has_ctrl(n) && phase->get_ctrl(n) == proj) {
926 uses.push(n);
927 }
928 }
929 } else {
930 assert(use->is_Region(), "what else?");
931 uint idx = 1;
932 for (; use->in(idx) != proj; idx++);
933 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) {
934 Node* u = use->fast_out(i);
935 if (u->is_Phi() && phase->get_ctrl(u->in(idx)) == proj) {
936 uses.push(u->in(idx));
937 }
938 }
939 }
940 for(uint next = 0; next < uses.size(); next++ ) {
941 Node *n = uses.at(next);
942 assert(phase->get_ctrl(n) == proj, "bad control");
943 phase->set_ctrl_and_loop(n, new_unc_ctrl);
944 if (n->in(0) == proj) {
945 phase->igvn().replace_input_of(n, 0, new_unc_ctrl);
946 }
947 for (uint i = 0; i < n->req(); i++) {
948 Node* m = n->in(i);
949 if (m != NULL && phase->has_ctrl(m) && phase->get_ctrl(m) == proj) {
950 uses.push(m);
951 }
952 }
953 }
954
955 phase->igvn().rehash_node_delayed(use);
956 int nb = use->replace_edge(proj, new_unc_ctrl);
957 assert(nb == 1, "only use expected");
958 }
959
960 void ShenandoahBarrierC2Support::test_in_cset(Node*& ctrl, Node*& not_cset_ctrl, Node* val, Node* raw_mem, PhaseIdealLoop* phase) {
961 Node* old_ctrl = ctrl;
962 PhaseIterGVN& igvn = phase->igvn();
963
964 Node* raw_val = new (phase->C) CastP2XNode(old_ctrl, val);
965 Node* cset_idx = new (phase->C) URShiftXNode(raw_val, igvn.intcon(ShenandoahHeapRegion::region_size_bytes_shift_jint()));
966
967 // Figure out the target cset address with raw pointer math.
968 // This avoids matching AddP+LoadB that would emit inefficient code.
969 // See JDK-8245465.
970 Node* cset_addr_ptr = igvn.makecon(TypeRawPtr::make(ShenandoahHeap::in_cset_fast_test_addr()));
971 Node* cset_addr = new (phase->C) CastP2XNode(old_ctrl, cset_addr_ptr);
972 Node* cset_load_addr = new (phase->C) AddXNode(cset_addr, cset_idx);
973 Node* cset_load_ptr = new (phase->C) CastX2PNode(cset_load_addr);
974
975 Node* cset_load = new (phase->C) LoadBNode(old_ctrl, raw_mem, cset_load_ptr,
976 DEBUG_ONLY(phase->C->get_adr_type(Compile::AliasIdxRaw)) NOT_DEBUG(NULL),
977 TypeInt::BYTE, MemNode::unordered);
978 Node* cset_cmp = new (phase->C) CmpINode(cset_load, igvn.zerocon(T_INT));
979 Node* cset_bool = new (phase->C) BoolNode(cset_cmp, BoolTest::ne);
980
981 IfNode* cset_iff = new (phase->C) IfNode(old_ctrl, cset_bool, PROB_UNLIKELY(0.999), COUNT_UNKNOWN);
982 ctrl = new (phase->C) IfTrueNode(cset_iff);
983 not_cset_ctrl = new (phase->C) IfFalseNode(cset_iff);
984
985 IdealLoopTree* loop = phase->get_loop(old_ctrl);
986 phase->register_control(cset_iff, loop, old_ctrl);
987 phase->register_control(ctrl, loop, cset_iff);
988 phase->register_control(not_cset_ctrl, loop, cset_iff);
989
990 phase->set_ctrl(cset_addr_ptr, phase->C->root());
991
992 phase->register_new_node(raw_val, old_ctrl);
993 phase->register_new_node(cset_idx, old_ctrl);
994 phase->register_new_node(cset_addr, old_ctrl);
995 phase->register_new_node(cset_load_addr, old_ctrl);
996 phase->register_new_node(cset_load_ptr, old_ctrl);
997 phase->register_new_node(cset_load, old_ctrl);
998 phase->register_new_node(cset_cmp, old_ctrl);
999 phase->register_new_node(cset_bool, old_ctrl);
1000 }
1001
1002 void ShenandoahBarrierC2Support::call_lrb_stub(Node*& ctrl, Node*& val, Node* load_addr, Node*& result_mem, Node* raw_mem, bool is_native, PhaseIdealLoop* phase) {
1003 IdealLoopTree*loop = phase->get_loop(ctrl);
1004 const TypePtr* obj_type = phase->igvn().type(val)->is_oopptr();
1005
1006 // The slow path stub consumes and produces raw memory in addition
1007 // to the existing memory edges
1008 Node* base = find_bottom_mem(ctrl, phase);
1009 MergeMemNode* mm = MergeMemNode::make(phase->C, base);
1010 mm->set_memory_at(Compile::AliasIdxRaw, raw_mem);
1011 phase->register_new_node(mm, ctrl);
1012
1013 address target = LP64_ONLY(UseCompressedOops) NOT_LP64(false) ?
1014 CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_narrow) :
1015 CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier);
1016
1017 Node* call = new (phase->C) CallLeafNode(ShenandoahBarrierSetC2::shenandoah_load_reference_barrier_Type(),
1018 target,
1019 "shenandoah_load_reference_barrier", TypeRawPtr::BOTTOM);
1020 call->init_req(TypeFunc::Control, ctrl);
1021 call->init_req(TypeFunc::I_O, phase->C->top());
1022 call->init_req(TypeFunc::Memory, mm);
1023 call->init_req(TypeFunc::FramePtr, phase->C->top());
1024 call->init_req(TypeFunc::ReturnAdr, phase->C->top());
1025 call->init_req(TypeFunc::Parms, val);
1026 call->init_req(TypeFunc::Parms+1, load_addr);
1027 phase->register_control(call, loop, ctrl);
1028 ctrl = new (phase->C) ProjNode(call, TypeFunc::Control);
1029 phase->register_control(ctrl, loop, call);
1030 result_mem = new (phase->C) ProjNode(call, TypeFunc::Memory);
1031 phase->register_new_node(result_mem, call);
1032 val = new (phase->C) ProjNode(call, TypeFunc::Parms);
1033 phase->register_new_node(val, call);
1034 val = new (phase->C) CheckCastPPNode(ctrl, val, obj_type);
1035 phase->register_new_node(val, ctrl);
1036 }
1037
1038 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) {
1039 Node* ctrl = phase->get_ctrl(barrier);
1040 Node* init_raw_mem = fixer.find_mem(ctrl, barrier);
1041
1042 // Update the control of all nodes that should be after the
1043 // barrier control flow
1044 uses.clear();
1045 // Every node that is control dependent on the barrier's input
1046 // control will be after the expanded barrier. The raw memory (if
1047 // its memory is control dependent on the barrier's input control)
1048 // must stay above the barrier.
1049 uses_to_ignore.clear();
1050 if (phase->has_ctrl(init_raw_mem) && phase->get_ctrl(init_raw_mem) == ctrl && !init_raw_mem->is_Phi()) {
1051 uses_to_ignore.push(init_raw_mem);
1052 }
1053 for (uint next = 0; next < uses_to_ignore.size(); next++) {
1054 Node *n = uses_to_ignore.at(next);
1055 for (uint i = 0; i < n->req(); i++) {
1056 Node* in = n->in(i);
1057 if (in != NULL && phase->has_ctrl(in) && phase->get_ctrl(in) == ctrl) {
1058 uses_to_ignore.push(in);
1059 }
1060 }
1061 }
1062 for (DUIterator_Fast imax, i = ctrl->fast_outs(imax); i < imax; i++) {
1063 Node* u = ctrl->fast_out(i);
1064 if (u->_idx < last &&
1065 u != barrier &&
1066 !uses_to_ignore.member(u) &&
1067 (u->in(0) != ctrl || (!u->is_Region() && !u->is_Phi())) &&
1068 (ctrl->Opcode() != Op_CatchProj || u->Opcode() != Op_CreateEx)) {
1069 Node* old_c = phase->ctrl_or_self(u);
1070 Node* c = old_c;
1071 if (c != ctrl ||
1072 is_dominator_same_ctrl(old_c, barrier, u, phase) ||
1073 ShenandoahBarrierSetC2::is_shenandoah_state_load(u)) {
1074 phase->igvn().rehash_node_delayed(u);
1075 int nb = u->replace_edge(ctrl, region);
1076 if (u->is_CFG()) {
1077 if (phase->idom(u) == ctrl) {
1078 phase->set_idom(u, region, phase->dom_depth(region));
1079 }
1080 } else if (phase->get_ctrl(u) == ctrl) {
1081 assert(u != init_raw_mem, "should leave input raw mem above the barrier");
1082 uses.push(u);
1083 }
1084 assert(nb == 1, "more than 1 ctrl input?");
1085 --i, imax -= nb;
1086 }
1087 }
1088 }
1089 }
1090
1091 static Node* create_phis_on_call_return(Node* ctrl, Node* c, Node* n, Node* n_clone, const CallProjections& projs, PhaseIdealLoop* phase) {
1092 Node* region = NULL;
1093 while (c != ctrl) {
1094 if (c->is_Region()) {
1095 region = c;
1096 }
1097 c = phase->idom(c);
1098 }
1099 assert(region != NULL, "");
1100 if (n->is_Bool()) {
1101 Node* bol_clone = n->clone();
1102 n = n->in(1);
1103 n_clone = n_clone->in(1);
1104 assert(n->is_Cmp() && n_clone->is_Cmp(), "should be cmp");
1105 Node* cmp_clone = n->clone();
1106 bol_clone->set_req(1, cmp_clone);
1107 if (n->in(1) != n_clone->in(1)) {
1108 cmp_clone->set_req(1, create_phis_on_call_return(ctrl, region, n->in(1), n_clone->in(1), projs, phase));
1109 }
1110 if (n->in(2) != n_clone->in(2)) {
1111 cmp_clone->set_req(2, create_phis_on_call_return(ctrl, region, n->in(2), n_clone->in(2), projs, phase));
1112 }
1113 phase->register_new_node(cmp_clone, region);
1114 phase->register_new_node(bol_clone, region);
1115 return bol_clone;
1116 }
1117 Node* phi = new (phase->C) PhiNode(region, n->bottom_type());
1118 for (uint j = 1; j < region->req(); j++) {
1119 Node* in = region->in(j);
1120 if (phase->is_dominator(projs.fallthrough_catchproj, in)) {
1121 phi->init_req(j, n);
1122 } else if (phase->is_dominator(projs.catchall_catchproj, in)) {
1123 phi->init_req(j, n_clone);
1124 } else {
1125 phi->init_req(j, create_phis_on_call_return(ctrl, in, n, n_clone, projs, phase));
1126 }
1127 }
1128 phase->register_new_node(phi, region);
1129 return phi;
1130 }
1131
1132 void ShenandoahBarrierC2Support::pin_and_expand(PhaseIdealLoop* phase) {
1133
1134 Unique_Node_List uses;
1135 Node_Stack stack(0);
1136 Node_List clones;
1137 for (int i = phase->C->shenandoah_barriers_count() - 1; i >= 0; i--) {
1138 ShenandoahLoadReferenceBarrierNode* lrb = phase->C->shenandoah_barrier(i);
1139 if (lrb->is_redundant()) {
1140 continue;
1141 }
1142
1143 Node* ctrl = phase->get_ctrl(lrb);
1144 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1145
1146 CallStaticJavaNode* unc = NULL;
1147 Node* unc_ctrl = NULL;
1148 Node* uncasted_val = val;
1149
1150 for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1151 Node* u = lrb->fast_out(i);
1152 if (u->Opcode() == Op_CastPP &&
1153 u->in(0) != NULL &&
1154 phase->is_dominator(u->in(0), ctrl)) {
1155 const Type* u_t = phase->igvn().type(u);
1156
1157 if (u_t->meet(TypePtr::NULL_PTR) != u_t &&
1158 u->in(0)->Opcode() == Op_IfTrue &&
1159 u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
1160 u->in(0)->in(0)->is_If() &&
1161 u->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
1162 u->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
1163 u->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
1164 u->in(0)->in(0)->in(1)->in(1)->in(1) == val &&
1165 u->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
1166 IdealLoopTree* loop = phase->get_loop(ctrl);
1167 IdealLoopTree* unc_loop = phase->get_loop(u->in(0));
1168
1169 if (!unc_loop->is_member(loop)) {
1170 continue;
1171 }
1172
1173 Node* branch = no_branches(ctrl, u->in(0), false, phase);
1174 assert(branch == NULL || branch == NodeSentinel, "was not looking for a branch");
1175 if (branch == NodeSentinel) {
1176 continue;
1177 }
1178
1179 phase->igvn().replace_input_of(u, 1, val);
1180 phase->igvn().replace_input_of(lrb, ShenandoahLoadReferenceBarrierNode::ValueIn, u);
1181 phase->set_ctrl(u, u->in(0));
1182 phase->set_ctrl(lrb, u->in(0));
1183 unc = u->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
1184 unc_ctrl = u->in(0);
1185 val = u;
1186
1187 for (DUIterator_Fast jmax, j = val->fast_outs(jmax); j < jmax; j++) {
1188 Node* u = val->fast_out(j);
1189 if (u == lrb) continue;
1190 phase->igvn().rehash_node_delayed(u);
1191 int nb = u->replace_edge(val, lrb);
1192 --j; jmax -= nb;
1193 }
1194
1195 RegionNode* r = new (phase->C) RegionNode(3);
1196 IfNode* iff = unc_ctrl->in(0)->as_If();
1197
1198 Node* ctrl_use = unc_ctrl->unique_ctrl_out();
1199 Node* unc_ctrl_clone = unc_ctrl->clone();
1200 phase->register_control(unc_ctrl_clone, loop, iff);
1201 Node* c = unc_ctrl_clone;
1202 Node* new_cast = clone_null_check(c, val, unc_ctrl_clone, phase);
1203 r->init_req(1, new_cast->in(0)->in(0)->as_If()->proj_out(0));
1204
1205 phase->igvn().replace_input_of(unc_ctrl, 0, c->in(0));
1206 phase->set_idom(unc_ctrl, c->in(0), phase->dom_depth(unc_ctrl));
1207 phase->lazy_replace(c, unc_ctrl);
1208 c = NULL;;
1209 phase->igvn().replace_input_of(val, 0, unc_ctrl_clone);
1210 phase->set_ctrl(val, unc_ctrl_clone);
1211
1212 IfNode* new_iff = new_cast->in(0)->in(0)->as_If();
1213 fix_null_check(unc, unc_ctrl_clone, r, uses, phase);
1214 Node* iff_proj = iff->proj_out(0);
1215 r->init_req(2, iff_proj);
1216 phase->register_control(r, phase->ltree_root(), iff);
1217
1218 Node* new_bol = new_iff->in(1)->clone();
1219 Node* new_cmp = new_bol->in(1)->clone();
1220 assert(new_cmp->Opcode() == Op_CmpP, "broken");
1221 assert(new_cmp->in(1) == val->in(1), "broken");
1222 new_bol->set_req(1, new_cmp);
1223 new_cmp->set_req(1, lrb);
1224 phase->register_new_node(new_bol, new_iff->in(0));
1225 phase->register_new_node(new_cmp, new_iff->in(0));
1226 phase->igvn().replace_input_of(new_iff, 1, new_bol);
1227 phase->igvn().replace_input_of(new_cast, 1, lrb);
1228
1229 for (DUIterator_Fast imax, i = lrb->fast_outs(imax); i < imax; i++) {
1230 Node* u = lrb->fast_out(i);
1231 if (u == new_cast || u == new_cmp) {
1232 continue;
1233 }
1234 phase->igvn().rehash_node_delayed(u);
1235 int nb = u->replace_edge(lrb, new_cast);
1236 assert(nb > 0, "no update?");
1237 --i; imax -= nb;
1238 }
1239
1240 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
1241 Node* u = val->fast_out(i);
1242 if (u == lrb) {
1243 continue;
1244 }
1245 phase->igvn().rehash_node_delayed(u);
1246 int nb = u->replace_edge(val, new_cast);
1247 assert(nb > 0, "no update?");
1248 --i; imax -= nb;
1249 }
1250
1251 ctrl = unc_ctrl_clone;
1252 phase->set_ctrl_and_loop(lrb, ctrl);
1253 break;
1254 }
1255 }
1256 }
1257 if ((ctrl->is_Proj() && ctrl->in(0)->is_CallJava()) || ctrl->is_CallJava()) {
1258 CallNode* call = ctrl->is_Proj() ? ctrl->in(0)->as_CallJava() : ctrl->as_CallJava();
1259 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
1260 // The rethrow call may have too many projections to be
1261 // properly handled here. Given there's no reason for a
1262 // barrier to depend on the call, move it above the call
1263 if (phase->get_ctrl(val) == ctrl) {
1264 assert(val->Opcode() == Op_DecodeN, "unexpected node");
1265 assert(phase->is_dominator(phase->get_ctrl(val->in(1)), call->in(0)), "Load is too low");
1266 phase->set_ctrl(val, call->in(0));
1267 }
1268 phase->set_ctrl(lrb, call->in(0));
1269 continue;
1270 }
1271 CallProjections projs;
1272 call->extract_projections(&projs, false, false);
1273
1274 #ifdef ASSERT
1275 VectorSet cloned(Thread::current()->resource_area());
1276 #endif
1277 Node* lrb_clone = lrb->clone();
1278 phase->register_new_node(lrb_clone, projs.catchall_catchproj);
1279 phase->set_ctrl(lrb, projs.fallthrough_catchproj);
1280
1281 stack.push(lrb, 0);
1282 clones.push(lrb_clone);
1283
1284 do {
1285 assert(stack.size() == clones.size(), "");
1286 Node* n = stack.node();
1287 #ifdef ASSERT
1288 if (n->is_Load()) {
1289 Node* mem = n->in(MemNode::Memory);
1290 for (DUIterator_Fast jmax, j = mem->fast_outs(jmax); j < jmax; j++) {
1291 Node* u = mem->fast_out(j);
1292 assert(!u->is_Store() || !u->is_LoadStore() || phase->get_ctrl(u) != ctrl, "anti dependent store?");
1293 }
1294 }
1295 #endif
1296 uint idx = stack.index();
1297 Node* n_clone = clones.at(clones.size()-1);
1298 if (idx < n->outcnt()) {
1299 Node* u = n->raw_out(idx);
1300 Node* c = phase->ctrl_or_self(u);
1301 if (phase->is_dominator(call, c) && phase->is_dominator(c, projs.fallthrough_proj)) {
1302 stack.set_index(idx+1);
1303 assert(!u->is_CFG(), "");
1304 stack.push(u, 0);
1305 assert(!cloned.test_set(u->_idx), "only one clone");
1306 Node* u_clone = u->clone();
1307 int nb = u_clone->replace_edge(n, n_clone);
1308 assert(nb > 0, "should have replaced some uses");
1309 phase->register_new_node(u_clone, projs.catchall_catchproj);
1310 clones.push(u_clone);
1311 phase->set_ctrl(u, projs.fallthrough_catchproj);
1312 } else {
1313 bool replaced = false;
1314 if (u->is_Phi()) {
1315 for (uint k = 1; k < u->req(); k++) {
1316 if (u->in(k) == n) {
1317 if (phase->is_dominator(projs.catchall_catchproj, u->in(0)->in(k))) {
1318 phase->igvn().replace_input_of(u, k, n_clone);
1319 replaced = true;
1320 } else if (!phase->is_dominator(projs.fallthrough_catchproj, u->in(0)->in(k))) {
1321 phase->igvn().replace_input_of(u, k, create_phis_on_call_return(ctrl, u->in(0)->in(k), n, n_clone, projs, phase));
1322 replaced = true;
1323 }
1324 }
1325 }
1326 } else {
1327 if (phase->is_dominator(projs.catchall_catchproj, c)) {
1328 phase->igvn().rehash_node_delayed(u);
1329 int nb = u->replace_edge(n, n_clone);
1330 assert(nb > 0, "should have replaced some uses");
1331 replaced = true;
1332 } else if (!phase->is_dominator(projs.fallthrough_catchproj, c)) {
1333 if (u->is_If()) {
1334 // Can't break If/Bool/Cmp chain
1335 assert(n->is_Bool(), "unexpected If shape");
1336 assert(stack.node_at(stack.size()-2)->is_Cmp(), "unexpected If shape");
1337 assert(n_clone->is_Bool(), "unexpected clone");
1338 assert(clones.at(clones.size()-2)->is_Cmp(), "unexpected clone");
1339 Node* bol_clone = n->clone();
1340 Node* cmp_clone = stack.node_at(stack.size()-2)->clone();
1341 bol_clone->set_req(1, cmp_clone);
1342
1343 Node* nn = stack.node_at(stack.size()-3);
1344 Node* nn_clone = clones.at(clones.size()-3);
1345 assert(nn->Opcode() == nn_clone->Opcode(), "mismatch");
1346
1347 int nb = cmp_clone->replace_edge(nn, create_phis_on_call_return(ctrl, c, nn, nn_clone, projs, phase));
1348 assert(nb > 0, "should have replaced some uses");
1349
1350 phase->register_new_node(bol_clone, u->in(0));
1351 phase->register_new_node(cmp_clone, u->in(0));
1352
1353 phase->igvn().replace_input_of(u, 1, bol_clone);
1354
1355 } else {
1356 phase->igvn().rehash_node_delayed(u);
1357 int nb = u->replace_edge(n, create_phis_on_call_return(ctrl, c, n, n_clone, projs, phase));
1358 assert(nb > 0, "should have replaced some uses");
1359 }
1360 replaced = true;
1361 }
1362 }
1363 if (!replaced) {
1364 stack.set_index(idx+1);
1365 }
1366 }
1367 } else {
1368 stack.pop();
1369 clones.pop();
1370 }
1371 } while (stack.size() > 0);
1372 assert(stack.size() == 0 && clones.size() == 0, "");
1373 }
1374 }
1375
1376 // Expand load-reference-barriers
1377 MemoryGraphFixer fixer(Compile::AliasIdxRaw, true, phase);
1378 Unique_Node_List uses_to_ignore;
1379 for (int i = phase->C->shenandoah_barriers_count() - 1; i >= 0; i--) {
1380 ShenandoahLoadReferenceBarrierNode* lrb = phase->C->shenandoah_barrier(i);
1381 if (lrb->is_redundant()) {
1382 phase->igvn().replace_node(lrb, lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1383 continue;
1384 }
1385 uint last = phase->C->unique();
1386 Node* ctrl = phase->get_ctrl(lrb);
1387 Node* val = lrb->in(ShenandoahLoadReferenceBarrierNode::ValueIn);
1388
1389
1390 Node* orig_ctrl = ctrl;
1391
1392 Node* raw_mem = fixer.find_mem(ctrl, lrb);
1393 Node* init_raw_mem = raw_mem;
1394 Node* raw_mem_for_ctrl = fixer.find_mem(ctrl, NULL);
1395
1396 IdealLoopTree *loop = phase->get_loop(ctrl);
1397 CallStaticJavaNode* unc = lrb->pin_and_expand_null_check(phase->igvn());
1398 Node* unc_ctrl = NULL;
1399 if (unc != NULL) {
1400 if (val->in(ShenandoahLoadReferenceBarrierNode::Control) != ctrl) {
1401 unc = NULL;
1402 } else {
1403 unc_ctrl = val->in(ShenandoahLoadReferenceBarrierNode::Control);
1404 }
1405 }
1406
1407 Node* uncasted_val = val;
1408 if (unc != NULL) {
1409 uncasted_val = val->in(1);
1410 }
1411
1412 Node* heap_stable_ctrl = NULL;
1413 Node* null_ctrl = NULL;
1414
1415 assert(val->bottom_type()->make_oopptr(), "need oop");
1416 assert(val->bottom_type()->make_oopptr()->const_oop() == NULL, "expect non-constant");
1417
1418 enum { _heap_stable = 1, _not_cset, _evac_path, _null_path, PATH_LIMIT };
1419 Node* region = new (phase->C) RegionNode(PATH_LIMIT);
1420 Node* val_phi = new (phase->C) PhiNode(region, uncasted_val->bottom_type()->is_oopptr());
1421 Node* raw_mem_phi = PhiNode::make(region, raw_mem, Type::MEMORY, TypeRawPtr::BOTTOM);
1422
1423 // Stable path.
1424 test_gc_state(ctrl, raw_mem, heap_stable_ctrl, phase, ShenandoahHeap::HAS_FORWARDED);
1425 IfNode* heap_stable_iff = heap_stable_ctrl->in(0)->as_If();
1426
1427 // Heap stable case
1428 region->init_req(_heap_stable, heap_stable_ctrl);
1429 val_phi->init_req(_heap_stable, uncasted_val);
1430 raw_mem_phi->init_req(_heap_stable, raw_mem);
1431
1432 Node* reg2_ctrl = NULL;
1433 // Null case
1434 test_null(ctrl, val, null_ctrl, phase);
1435 if (null_ctrl != NULL) {
1436 reg2_ctrl = null_ctrl->in(0);
1437 region->init_req(_null_path, null_ctrl);
1438 val_phi->init_req(_null_path, uncasted_val);
1439 raw_mem_phi->init_req(_null_path, raw_mem);
1440 } else {
1441 region->del_req(_null_path);
1442 val_phi->del_req(_null_path);
1443 raw_mem_phi->del_req(_null_path);
1444 }
1445
1446 // Test for in-cset.
1447 // Wires !in_cset(obj) to slot 2 of region and phis
1448 Node* not_cset_ctrl = NULL;
1449 test_in_cset(ctrl, not_cset_ctrl, uncasted_val, raw_mem, phase);
1450 if (not_cset_ctrl != NULL) {
1451 if (reg2_ctrl == NULL) reg2_ctrl = not_cset_ctrl->in(0);
1452 region->init_req(_not_cset, not_cset_ctrl);
1453 val_phi->init_req(_not_cset, uncasted_val);
1454 raw_mem_phi->init_req(_not_cset, raw_mem);
1455 }
1456
1457 // Resolve object when orig-value is in cset.
1458 // Make the unconditional resolve for fwdptr.
1459 Node* new_val = uncasted_val;
1460 if (unc_ctrl != NULL) {
1461 // Clone the null check in this branch to allow implicit null check
1462 new_val = clone_null_check(ctrl, val, unc_ctrl, phase);
1463 fix_null_check(unc, unc_ctrl, ctrl->in(0)->as_If()->proj_out(0), uses, phase);
1464
1465 IfNode* iff = unc_ctrl->in(0)->as_If();
1466 phase->igvn().replace_input_of(iff, 1, phase->igvn().intcon(1));
1467 }
1468
1469 // Call lrb-stub and wire up that path in slots 4
1470 Node* result_mem = NULL;
1471
1472 Node* fwd = new_val;
1473 Node* addr;
1474 if (ShenandoahSelfFixing) {
1475 VectorSet visited(Thread::current()->resource_area());
1476 addr = get_load_addr(phase, visited, lrb);
1477 } else {
1478 addr = phase->igvn().zerocon(T_OBJECT);
1479 }
1480 if (addr->Opcode() == Op_AddP) {
1481 Node* orig_base = addr->in(AddPNode::Base);
1482 Node* base = new (phase->C) CheckCastPPNode(ctrl, orig_base, orig_base->bottom_type());
1483 phase->register_new_node(base, ctrl);
1484 if (addr->in(AddPNode::Base) == addr->in((AddPNode::Address))) {
1485 // Field access
1486 addr = addr->clone();
1487 addr->set_req(AddPNode::Base, base);
1488 addr->set_req(AddPNode::Address, base);
1489 phase->register_new_node(addr, ctrl);
1490 } else {
1491 Node* addr2 = addr->in(AddPNode::Address);
1492 if (addr2->Opcode() == Op_AddP && addr2->in(AddPNode::Base) == addr2->in(AddPNode::Address) &&
1493 addr2->in(AddPNode::Base) == orig_base) {
1494 addr2 = addr2->clone();
1495 addr2->set_req(AddPNode::Base, base);
1496 addr2->set_req(AddPNode::Address, base);
1497 phase->register_new_node(addr2, ctrl);
1498 addr = addr->clone();
1499 addr->set_req(AddPNode::Base, base);
1500 addr->set_req(AddPNode::Address, addr2);
1501 phase->register_new_node(addr, ctrl);
1502 }
1503 }
1504 }
1505 call_lrb_stub(ctrl, fwd, addr, result_mem, raw_mem, false, phase);
1506 region->init_req(_evac_path, ctrl);
1507 val_phi->init_req(_evac_path, fwd);
1508 raw_mem_phi->init_req(_evac_path, result_mem);
1509
1510 phase->register_control(region, loop, heap_stable_iff);
1511 Node* out_val = val_phi;
1512 phase->register_new_node(val_phi, region);
1513 phase->register_new_node(raw_mem_phi, region);
1514
1515 fix_ctrl(lrb, region, fixer, uses, uses_to_ignore, last, phase);
1516
1517 ctrl = orig_ctrl;
1518
1519 if (unc != NULL) {
1520 for (DUIterator_Fast imax, i = val->fast_outs(imax); i < imax; i++) {
1521 Node* u = val->fast_out(i);
1522 Node* c = phase->ctrl_or_self(u);
1523 if (u != lrb && (c != ctrl || is_dominator_same_ctrl(c, lrb, u, phase))) {
1524 phase->igvn().rehash_node_delayed(u);
1525 int nb = u->replace_edge(val, out_val);
1526 --i, imax -= nb;
1527 }
1528 }
1529 if (val->outcnt() == 0) {
1530 phase->igvn()._worklist.push(val);
1531 }
1532 }
1533 phase->igvn().replace_node(lrb, out_val);
1534
1535 follow_barrier_uses(out_val, ctrl, uses, phase);
1536
1537 for(uint next = 0; next < uses.size(); next++ ) {
1538 Node *n = uses.at(next);
1539 assert(phase->get_ctrl(n) == ctrl, "bad control");
1540 assert(n != init_raw_mem, "should leave input raw mem above the barrier");
1541 phase->set_ctrl(n, region);
1542 follow_barrier_uses(n, ctrl, uses, phase);
1543 }
1544
1545 // The slow path call produces memory: hook the raw memory phi
1546 // from the expanded load reference barrier with the rest of the graph
1547 // which may require adding memory phis at every post dominated
1548 // region and at enclosing loop heads. Use the memory state
1549 // collected in memory_nodes to fix the memory graph. Update that
1550 // memory state as we go.
1551 fixer.fix_mem(ctrl, region, init_raw_mem, raw_mem_for_ctrl, raw_mem_phi, uses);
1552 }
1553 // Done expanding load-reference-barriers.
1554 assert(phase->C->shenandoah_barriers_count() == 0, "all load reference barrier nodes should have been replaced");
1555
1556 }
1557
1558 Node* ShenandoahBarrierC2Support::get_load_addr(PhaseIdealLoop* phase, VectorSet& visited, Node* in) {
1559 if (visited.test_set(in->_idx)) {
1560 return NULL;
1561 }
1562 switch (in->Opcode()) {
1563 case Op_Proj:
1564 return get_load_addr(phase, visited, in->in(0));
1565 case Op_CastPP:
1566 case Op_CheckCastPP:
1567 case Op_DecodeN:
1568 case Op_EncodeP:
1569 return get_load_addr(phase, visited, in->in(1));
1570 case Op_LoadN:
1571 case Op_LoadP:
1572 return in->in(MemNode::Address);
1573 case Op_GetAndSetN:
1574 case Op_GetAndSetP:
1575 // Those instructions would just have stored a different
1576 // value into the field. No use to attempt to fix it at this point.
1577 return phase->igvn().zerocon(T_OBJECT);
1578 case Op_CMoveP:
1579 case Op_CMoveN: {
1580 Node* t = get_load_addr(phase, visited, in->in(CMoveNode::IfTrue));
1581 Node* f = get_load_addr(phase, visited, in->in(CMoveNode::IfFalse));
1582 // Handle unambiguous cases: single address reported on both branches.
1583 if (t != NULL && f == NULL) return t;
1584 if (t == NULL && f != NULL) return f;
1585 if (t != NULL && t == f) return t;
1586 // Ambiguity.
1587 return phase->igvn().zerocon(T_OBJECT);
1588 }
1589 case Op_Phi: {
1590 Node* addr = NULL;
1591 for (uint i = 1; i < in->req(); i++) {
1592 Node* addr1 = get_load_addr(phase, visited, in->in(i));
1593 if (addr == NULL) {
1594 addr = addr1;
1595 }
1596 if (addr != addr1) {
1597 return phase->igvn().zerocon(T_OBJECT);
1598 }
1599 }
1600 return addr;
1601 }
1602 case Op_ShenandoahLoadReferenceBarrier:
1603 return get_load_addr(phase, visited, in->in(ShenandoahLoadReferenceBarrierNode::ValueIn));
1604 case Op_CallDynamicJava:
1605 case Op_CallLeaf:
1606 case Op_CallStaticJava:
1607 case Op_ConN:
1608 case Op_ConP:
1609 case Op_Parm:
1610 case Op_CreateEx:
1611 return phase->igvn().zerocon(T_OBJECT);
1612 default:
1613 #ifdef ASSERT
1614 fatal(err_msg("Unknown node in get_load_addr: %s", NodeClassNames[in->Opcode()]));
1615 #endif
1616 return phase->igvn().zerocon(T_OBJECT);
1617 }
1618
1619 }
1620
1621 void ShenandoahBarrierC2Support::move_gc_state_test_out_of_loop(IfNode* iff, PhaseIdealLoop* phase) {
1622 IdealLoopTree *loop = phase->get_loop(iff);
1623 Node* loop_head = loop->_head;
1624 Node* entry_c = loop_head->in(LoopNode::EntryControl);
1625
1626 Node* bol = iff->in(1);
1627 Node* cmp = bol->in(1);
1628 Node* andi = cmp->in(1);
1629 Node* load = andi->in(1);
1630
1631 assert(is_gc_state_load(load), "broken");
1632 if (!phase->is_dominator(load->in(0), entry_c)) {
1633 Node* mem_ctrl = NULL;
1634 Node* mem = dom_mem(load->in(MemNode::Memory), loop_head, Compile::AliasIdxRaw, mem_ctrl, phase);
1635 load = load->clone();
1636 load->set_req(MemNode::Memory, mem);
1637 load->set_req(0, entry_c);
1638 phase->register_new_node(load, entry_c);
1639 andi = andi->clone();
1640 andi->set_req(1, load);
1641 phase->register_new_node(andi, entry_c);
1642 cmp = cmp->clone();
1643 cmp->set_req(1, andi);
1644 phase->register_new_node(cmp, entry_c);
1645 bol = bol->clone();
1646 bol->set_req(1, cmp);
1647 phase->register_new_node(bol, entry_c);
1648
1649 Node* old_bol =iff->in(1);
1650 phase->igvn().replace_input_of(iff, 1, bol);
1651 }
1652 }
1653
1654 bool ShenandoahBarrierC2Support::identical_backtoback_ifs(Node* n, PhaseIdealLoop* phase) {
1655 if (!n->is_If() || n->is_CountedLoopEnd()) {
1656 return false;
1657 }
1658 Node* region = n->in(0);
1659
1660 if (!region->is_Region()) {
1661 return false;
1662 }
1663 Node* dom = phase->idom(region);
1664 if (!dom->is_If()) {
1665 return false;
1666 }
1667
1668 if (!is_heap_stable_test(n) || !is_heap_stable_test(dom)) {
1669 return false;
1670 }
1671
1672 IfNode* dom_if = dom->as_If();
1673 Node* proj_true = dom_if->proj_out(1);
1674 Node* proj_false = dom_if->proj_out(0);
1675
1676 for (uint i = 1; i < region->req(); i++) {
1677 if (phase->is_dominator(proj_true, region->in(i))) {
1678 continue;
1679 }
1680 if (phase->is_dominator(proj_false, region->in(i))) {
1681 continue;
1682 }
1683 return false;
1684 }
1685
1686 return true;
1687 }
1688
1689 static bool merge_point_too_heavy(Compile* C, Node* region) {
1690 // Bail out if the region and its phis have too many users.
1691 int weight = 0;
1692 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1693 weight += region->fast_out(i)->outcnt();
1694 }
1695 int nodes_left = C->max_node_limit() - C->live_nodes();
1696 if (weight * 8 > nodes_left) {
1697 #ifndef PRODUCT
1698 if (PrintOpto) {
1699 tty->print_cr("*** Split-if bails out: %d nodes, region weight %d", C->unique(), weight);
1700 }
1701 #endif
1702 return true;
1703 } else {
1704 return false;
1705 }
1706 }
1707
1708 static bool merge_point_safe(Node* region) {
1709 // 4799512: Stop split_if_with_blocks from splitting a block with a ConvI2LNode
1710 // having a PhiNode input. This sidesteps the dangerous case where the split
1711 // ConvI2LNode may become TOP if the input Value() does not
1712 // overlap the ConvI2L range, leaving a node which may not dominate its
1713 // uses.
1714 // A better fix for this problem can be found in the BugTraq entry, but
1715 // expediency for Mantis demands this hack.
1716 // 6855164: If the merge point has a FastLockNode with a PhiNode input, we stop
1717 // split_if_with_blocks from splitting a block because we could not move around
1718 // the FastLockNode.
1719 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
1720 Node* n = region->fast_out(i);
1721 if (n->is_Phi()) {
1722 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
1723 Node* m = n->fast_out(j);
1724 if (m->is_FastLock())
1725 return false;
1726 #ifdef _LP64
1727 if (m->Opcode() == Op_ConvI2L)
1728 return false;
1729 if (m->is_CastII() && m->isa_CastII()->has_range_check()) {
1730 return false;
1731 }
1732 #endif
1733 }
1734 }
1735 }
1736 return true;
1737 }
1738
1739 static bool can_split_if(PhaseIdealLoop* phase, Node* n_ctrl) {
1740 if (phase->C->live_nodes() > 35000) {
1741 return false; // Method too big
1742 }
1743
1744 // Do not do 'split-if' if irreducible loops are present.
1745 if (phase->_has_irreducible_loops) {
1746 return false;
1747 }
1748
1749 if (merge_point_too_heavy(phase->C, n_ctrl)) {
1750 return false;
1751 }
1752
1753 // Do not do 'split-if' if some paths are dead. First do dead code
1754 // elimination and then see if its still profitable.
1755 for (uint i = 1; i < n_ctrl->req(); i++) {
1756 if (n_ctrl->in(i) == phase->C->top()) {
1757 return false;
1758 }
1759 }
1760
1761 // If trying to do a 'Split-If' at the loop head, it is only
1762 // profitable if the cmp folds up on BOTH paths. Otherwise we
1763 // risk peeling a loop forever.
1764
1765 // CNC - Disabled for now. Requires careful handling of loop
1766 // body selection for the cloned code. Also, make sure we check
1767 // for any input path not being in the same loop as n_ctrl. For
1768 // irreducible loops we cannot check for 'n_ctrl->is_Loop()'
1769 // because the alternative loop entry points won't be converted
1770 // into LoopNodes.
1771 IdealLoopTree *n_loop = phase->get_loop(n_ctrl);
1772 for (uint j = 1; j < n_ctrl->req(); j++) {
1773 if (phase->get_loop(n_ctrl->in(j)) != n_loop) {
1774 return false;
1775 }
1776 }
1777
1778 // Check for safety of the merge point.
1779 if (!merge_point_safe(n_ctrl)) {
1780 return false;
1781 }
1782
1783 return true;
1784 }
1785
1786 void ShenandoahBarrierC2Support::merge_back_to_back_tests(Node* n, PhaseIdealLoop* phase) {
1787 assert(is_heap_stable_test(n), "no other tests");
1788 if (identical_backtoback_ifs(n, phase)) {
1789 Node* n_ctrl = n->in(0);
1790 if (can_split_if(phase, n_ctrl)) {
1791 IfNode* dom_if = phase->idom(n_ctrl)->as_If();
1792 if (is_heap_stable_test(n)) {
1793 Node* gc_state_load = n->in(1)->in(1)->in(1)->in(1);
1794 assert(is_gc_state_load(gc_state_load), "broken");
1795 Node* dom_gc_state_load = dom_if->in(1)->in(1)->in(1)->in(1);
1796 assert(is_gc_state_load(dom_gc_state_load), "broken");
1797 if (gc_state_load != dom_gc_state_load) {
1798 phase->igvn().replace_node(gc_state_load, dom_gc_state_load);
1799 }
1800 }
1801 PhiNode* bolphi = PhiNode::make_blank(n_ctrl, n->in(1));
1802 Node* proj_true = dom_if->proj_out(1);
1803 Node* proj_false = dom_if->proj_out(0);
1804 Node* con_true = phase->igvn().makecon(TypeInt::ONE);
1805 Node* con_false = phase->igvn().makecon(TypeInt::ZERO);
1806
1807 for (uint i = 1; i < n_ctrl->req(); i++) {
1808 if (phase->is_dominator(proj_true, n_ctrl->in(i))) {
1809 bolphi->init_req(i, con_true);
1810 } else {
1811 assert(phase->is_dominator(proj_false, n_ctrl->in(i)), "bad if");
1812 bolphi->init_req(i, con_false);
1813 }
1814 }
1815 phase->register_new_node(bolphi, n_ctrl);
1816 phase->igvn().replace_input_of(n, 1, bolphi);
1817 phase->do_split_if(n);
1818 }
1819 }
1820 }
1821
1822 IfNode* ShenandoahBarrierC2Support::find_unswitching_candidate(const IdealLoopTree* loop, PhaseIdealLoop* phase) {
1823 // Find first invariant test that doesn't exit the loop
1824 LoopNode *head = loop->_head->as_Loop();
1825 IfNode* unswitch_iff = NULL;
1826 Node* n = head->in(LoopNode::LoopBackControl);
1827 int loop_has_sfpts = -1;
1828 while (n != head) {
1829 Node* n_dom = phase->idom(n);
1830 if (n->is_Region()) {
1831 if (n_dom->is_If()) {
1832 IfNode* iff = n_dom->as_If();
1833 if (iff->in(1)->is_Bool()) {
1834 BoolNode* bol = iff->in(1)->as_Bool();
1835 if (bol->in(1)->is_Cmp()) {
1836 // If condition is invariant and not a loop exit,
1837 // then found reason to unswitch.
1838 if (is_heap_stable_test(iff) &&
1839 (loop_has_sfpts == -1 || loop_has_sfpts == 0)) {
1840 assert(!loop->is_loop_exit(iff), "both branches should be in the loop");
1841 if (loop_has_sfpts == -1) {
1842 for(uint i = 0; i < loop->_body.size(); i++) {
1843 Node *m = loop->_body[i];
1844 if (m->is_SafePoint() && !m->is_CallLeaf()) {
1845 loop_has_sfpts = 1;
1846 break;
1847 }
1848 }
1849 if (loop_has_sfpts == -1) {
1850 loop_has_sfpts = 0;
1851 }
1852 }
1853 if (!loop_has_sfpts) {
1854 unswitch_iff = iff;
1855 }
1856 }
1857 }
1858 }
1859 }
1860 }
1861 n = n_dom;
1862 }
1863 return unswitch_iff;
1864 }
1865
1866
1867 void ShenandoahBarrierC2Support::optimize_after_expansion(VectorSet &visited, Node_Stack &stack, Node_List &old_new, PhaseIdealLoop* phase) {
1868 Node_List heap_stable_tests;
1869 stack.push(phase->C->start(), 0);
1870 do {
1871 Node* n = stack.node();
1872 uint i = stack.index();
1873
1874 if (i < n->outcnt()) {
1875 Node* u = n->raw_out(i);
1876 stack.set_index(i+1);
1877 if (!visited.test_set(u->_idx)) {
1878 stack.push(u, 0);
1879 }
1880 } else {
1881 stack.pop();
1882 if (n->is_If() && is_heap_stable_test(n)) {
1883 heap_stable_tests.push(n);
1884 }
1885 }
1886 } while (stack.size() > 0);
1887
1888 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1889 Node* n = heap_stable_tests.at(i);
1890 assert(is_heap_stable_test(n), "only evacuation test");
1891 merge_back_to_back_tests(n, phase);
1892 }
1893
1894 if (!phase->C->major_progress()) {
1895 VectorSet seen(Thread::current()->resource_area());
1896 for (uint i = 0; i < heap_stable_tests.size(); i++) {
1897 Node* n = heap_stable_tests.at(i);
1898 IdealLoopTree* loop = phase->get_loop(n);
1899 if (loop != phase->ltree_root() &&
1900 loop->_child == NULL &&
1901 !loop->_irreducible) {
1902 Node* head = loop->_head;
1903 if (head->is_Loop() &&
1904 (!head->is_CountedLoop() || head->as_CountedLoop()->is_main_loop() || head->as_CountedLoop()->is_normal_loop()) &&
1905 !seen.test_set(head->_idx)) {
1906 IfNode* iff = find_unswitching_candidate(loop, phase);
1907 if (iff != NULL) {
1908 Node* bol = iff->in(1);
1909 move_gc_state_test_out_of_loop(iff, phase);
1910 if (loop->policy_unswitching(phase)) {
1911 phase->do_unswitching(loop, old_new);
1912 } else {
1913 // Not proceeding with unswitching. Move load back in
1914 // the loop.
1915 phase->igvn().replace_input_of(iff, 1, bol);
1916 }
1917 }
1918 }
1919 }
1920 }
1921 }
1922 }
1923
1924 #ifdef ASSERT
1925 void ShenandoahBarrierC2Support::verify_raw_mem(RootNode* root) {
1926 const bool trace = false;
1927 ResourceMark rm;
1928 Unique_Node_List nodes;
1929 Unique_Node_List controls;
1930 Unique_Node_List memories;
1931
1932 nodes.push(root);
1933 for (uint next = 0; next < nodes.size(); next++) {
1934 Node *n = nodes.at(next);
1935 if (ShenandoahBarrierSetC2::is_shenandoah_lrb_call(n)) {
1936 controls.push(n);
1937 if (trace) { tty->print("XXXXXX verifying"); n->dump(); }
1938 for (uint next2 = 0; next2 < controls.size(); next2++) {
1939 Node *m = controls.at(next2);
1940 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
1941 Node* u = m->fast_out(i);
1942 if (u->is_CFG() && !u->is_Root() &&
1943 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1) &&
1944 !(u->is_Region() && u->unique_ctrl_out()->Opcode() == Op_Halt)) {
1945 if (trace) { tty->print("XXXXXX pushing control"); u->dump(); }
1946 controls.push(u);
1947 }
1948 }
1949 }
1950 memories.push(n->as_Call()->proj_out(TypeFunc::Memory));
1951 for (uint next2 = 0; next2 < memories.size(); next2++) {
1952 Node *m = memories.at(next2);
1953 assert(m->bottom_type() == Type::MEMORY, "");
1954 for (DUIterator_Fast imax, i = m->fast_outs(imax); i < imax; i++) {
1955 Node* u = m->fast_out(i);
1956 if (u->bottom_type() == Type::MEMORY && (u->is_Mem() || u->is_ClearArray())) {
1957 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
1958 memories.push(u);
1959 } else if (u->is_LoadStore()) {
1960 if (trace) { tty->print("XXXXXX pushing memory"); u->find_out_with(Op_SCMemProj)->dump(); }
1961 memories.push(u->find_out_with(Op_SCMemProj));
1962 } else if (u->is_MergeMem() && u->as_MergeMem()->memory_at(Compile::AliasIdxRaw) == m) {
1963 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
1964 memories.push(u);
1965 } else if (u->is_Phi()) {
1966 assert(u->bottom_type() == Type::MEMORY, "");
1967 if (u->adr_type() == TypeRawPtr::BOTTOM || u->adr_type() == TypePtr::BOTTOM) {
1968 assert(controls.member(u->in(0)), "");
1969 if (trace) { tty->print("XXXXXX pushing memory"); u->dump(); }
1970 memories.push(u);
1971 }
1972 } else if (u->is_SafePoint() || u->is_MemBar()) {
1973 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
1974 Node* uu = u->fast_out(j);
1975 if (uu->bottom_type() == Type::MEMORY) {
1976 if (trace) { tty->print("XXXXXX pushing memory"); uu->dump(); }
1977 memories.push(uu);
1978 }
1979 }
1980 }
1981 }
1982 }
1983 for (uint next2 = 0; next2 < controls.size(); next2++) {
1984 Node *m = controls.at(next2);
1985 if (m->is_Region()) {
1986 bool all_in = true;
1987 for (uint i = 1; i < m->req(); i++) {
1988 if (!controls.member(m->in(i))) {
1989 all_in = false;
1990 break;
1991 }
1992 }
1993 if (trace) { tty->print("XXX verifying %s", all_in ? "all in" : ""); m->dump(); }
1994 bool found_phi = false;
1995 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax && !found_phi; j++) {
1996 Node* u = m->fast_out(j);
1997 if (u->is_Phi() && memories.member(u)) {
1998 found_phi = true;
1999 for (uint i = 1; i < u->req() && found_phi; i++) {
2000 Node* k = u->in(i);
2001 if (memories.member(k) != controls.member(m->in(i))) {
2002 found_phi = false;
2003 }
2004 }
2005 }
2006 }
2007 assert(found_phi || all_in, "");
2008 }
2009 }
2010 controls.clear();
2011 memories.clear();
2012 }
2013 for( uint i = 0; i < n->len(); ++i ) {
2014 Node *m = n->in(i);
2015 if (m != NULL) {
2016 nodes.push(m);
2017 }
2018 }
2019 }
2020 }
2021 #endif
2022
2023 #ifdef ASSERT
2024 static bool has_never_branch(Node* root) {
2025 for (uint i = 1; i < root->req(); i++) {
2026 Node* in = root->in(i);
2027 if (in != NULL && in->Opcode() == Op_Halt && in->in(0)->is_Proj() && in->in(0)->in(0)->Opcode() == Op_NeverBranch) {
2028 return true;
2029 }
2030 }
2031 return false;
2032 }
2033 #endif
2034
2035 void MemoryGraphFixer::collect_memory_nodes() {
2036 Node_Stack stack(0);
2037 VectorSet visited(Thread::current()->resource_area());
2038 Node_List regions;
2039
2040 // Walk the raw memory graph and create a mapping from CFG node to
2041 // memory node. Exclude phis for now.
2042 stack.push(_phase->C->root(), 1);
2043 do {
2044 Node* n = stack.node();
2045 int opc = n->Opcode();
2046 uint i = stack.index();
2047 if (i < n->req()) {
2048 Node* mem = NULL;
2049 if (opc == Op_Root) {
2050 Node* in = n->in(i);
2051 int in_opc = in->Opcode();
2052 if (in_opc == Op_Return || in_opc == Op_Rethrow) {
2053 mem = in->in(TypeFunc::Memory);
2054 } else if (in_opc == Op_Halt) {
2055 if (in->in(0)->is_Region()) {
2056 Node* r = in->in(0);
2057 for (uint j = 1; j < r->req(); j++) {
2058 assert(r->in(j)->Opcode() != Op_NeverBranch, "");
2059 }
2060 } else {
2061 Node* proj = in->in(0);
2062 assert(proj->is_Proj(), "");
2063 Node* in = proj->in(0);
2064 assert(in->is_CallStaticJava() || in->Opcode() == Op_NeverBranch || in->Opcode() == Op_Catch || proj->is_IfProj(), "");
2065 if (in->is_CallStaticJava()) {
2066 mem = in->in(TypeFunc::Memory);
2067 } else if (in->Opcode() == Op_Catch) {
2068 Node* call = in->in(0)->in(0);
2069 assert(call->is_Call(), "");
2070 mem = call->in(TypeFunc::Memory);
2071 } else if (in->Opcode() == Op_NeverBranch) {
2072 Node* head = in->in(0);
2073 assert(head->is_Region(), "unexpected infinite loop graph shape");
2074
2075 Node* phi_mem = NULL;
2076 for (DUIterator_Fast jmax, j = head->fast_outs(jmax); j < jmax; j++) {
2077 Node* u = head->fast_out(j);
2078 if (u->is_Phi() && u->bottom_type() == Type::MEMORY) {
2079 if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2080 assert(phi_mem == NULL || phi_mem->adr_type() == TypePtr::BOTTOM, "");
2081 phi_mem = u;
2082 } else if (u->adr_type() == TypePtr::BOTTOM) {
2083 assert(phi_mem == NULL || _phase->C->get_alias_index(phi_mem->adr_type()) == _alias, "");
2084 if (phi_mem == NULL) {
2085 phi_mem = u;
2086 }
2087 }
2088 }
2089 }
2090 if (phi_mem == NULL) {
2091 for (uint j = 1; j < head->req(); j++) {
2092 Node* tail = head->in(j);
2093 if (!_phase->is_dominator(head, tail)) {
2094 continue;
2095 }
2096 Node* c = tail;
2097 while (c != head) {
2098 if (c->is_SafePoint() && !c->is_CallLeaf()) {
2099 Node* m =c->in(TypeFunc::Memory);
2100 if (m->is_MergeMem()) {
2101 m = m->as_MergeMem()->memory_at(_alias);
2102 }
2103 assert(mem == NULL || mem == m, "several memory states");
2104 mem = m;
2105 }
2106 c = _phase->idom(c);
2107 }
2108 assert(mem != NULL, "should have found safepoint");
2109 }
2110 assert(mem != NULL, "should have found safepoint");
2111 } else {
2112 mem = phi_mem;
2113 }
2114 }
2115 }
2116 } else {
2117 #ifdef ASSERT
2118 n->dump();
2119 in->dump();
2120 #endif
2121 ShouldNotReachHere();
2122 }
2123 } else {
2124 assert(n->is_Phi() && n->bottom_type() == Type::MEMORY, "");
2125 assert(n->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(n->adr_type()) == _alias, "");
2126 mem = n->in(i);
2127 }
2128 i++;
2129 stack.set_index(i);
2130 if (mem == NULL) {
2131 continue;
2132 }
2133 for (;;) {
2134 if (visited.test_set(mem->_idx) || mem->is_Start()) {
2135 break;
2136 }
2137 if (mem->is_Phi()) {
2138 stack.push(mem, 2);
2139 mem = mem->in(1);
2140 } else if (mem->is_Proj()) {
2141 stack.push(mem, mem->req());
2142 mem = mem->in(0);
2143 } else if (mem->is_SafePoint() || mem->is_MemBar()) {
2144 mem = mem->in(TypeFunc::Memory);
2145 } else if (mem->is_MergeMem()) {
2146 MergeMemNode* mm = mem->as_MergeMem();
2147 mem = mm->memory_at(_alias);
2148 } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
2149 assert(_alias == Compile::AliasIdxRaw, "");
2150 stack.push(mem, mem->req());
2151 mem = mem->in(MemNode::Memory);
2152 } else {
2153 #ifdef ASSERT
2154 mem->dump();
2155 #endif
2156 ShouldNotReachHere();
2157 }
2158 }
2159 } else {
2160 if (n->is_Phi()) {
2161 // Nothing
2162 } else if (!n->is_Root()) {
2163 Node* c = get_ctrl(n);
2164 _memory_nodes.map(c->_idx, n);
2165 }
2166 stack.pop();
2167 }
2168 } while(stack.is_nonempty());
2169
2170 // Iterate over CFG nodes in rpo and propagate memory state to
2171 // compute memory state at regions, creating new phis if needed.
2172 Node_List rpo_list;
2173 visited.Clear();
2174 _phase->rpo(_phase->C->root(), stack, visited, rpo_list);
2175 Node* root = rpo_list.pop();
2176 assert(root == _phase->C->root(), "");
2177
2178 const bool trace = false;
2179 #ifdef ASSERT
2180 if (trace) {
2181 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2182 Node* c = rpo_list.at(i);
2183 if (_memory_nodes[c->_idx] != NULL) {
2184 tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump();
2185 }
2186 }
2187 }
2188 #endif
2189 uint last = _phase->C->unique();
2190
2191 #ifdef ASSERT
2192 uint8_t max_depth = 0;
2193 for (LoopTreeIterator iter(_phase->ltree_root()); !iter.done(); iter.next()) {
2194 IdealLoopTree* lpt = iter.current();
2195 max_depth = MAX2(max_depth, lpt->_nest);
2196 }
2197 #endif
2198
2199 bool progress = true;
2200 int iteration = 0;
2201 Node_List dead_phis;
2202 while (progress) {
2203 progress = false;
2204 iteration++;
2205 assert(iteration <= 2+max_depth || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2206 if (trace) { tty->print_cr("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX"); }
2207 IdealLoopTree* last_updated_ilt = NULL;
2208 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2209 Node* c = rpo_list.at(i);
2210
2211 Node* prev_mem = _memory_nodes[c->_idx];
2212 if (c->is_Region()) {
2213 Node* prev_region = regions[c->_idx];
2214 Node* unique = NULL;
2215 for (uint j = 1; j < c->req() && unique != NodeSentinel; j++) {
2216 Node* m = _memory_nodes[c->in(j)->_idx];
2217 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");
2218 if (m != NULL) {
2219 if (m == prev_region && ((c->is_Loop() && j == LoopNode::LoopBackControl) || (prev_region->is_Phi() && prev_region->in(0) == c))) {
2220 assert(c->is_Loop() && j == LoopNode::LoopBackControl || _phase->C->has_irreducible_loop() || has_never_branch(_phase->C->root()), "");
2221 // continue
2222 } else if (unique == NULL) {
2223 unique = m;
2224 } else if (m == unique) {
2225 // continue
2226 } else {
2227 unique = NodeSentinel;
2228 }
2229 }
2230 }
2231 assert(unique != NULL, "empty phi???");
2232 if (unique != NodeSentinel) {
2233 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c) {
2234 dead_phis.push(prev_region);
2235 }
2236 regions.map(c->_idx, unique);
2237 } else {
2238 Node* phi = NULL;
2239 if (prev_region != NULL && prev_region->is_Phi() && prev_region->in(0) == c && prev_region->_idx >= last) {
2240 phi = prev_region;
2241 for (uint k = 1; k < c->req(); k++) {
2242 Node* m = _memory_nodes[c->in(k)->_idx];
2243 assert(m != NULL, "expect memory state");
2244 phi->set_req(k, m);
2245 }
2246 } else {
2247 for (DUIterator_Fast jmax, j = c->fast_outs(jmax); j < jmax && phi == NULL; j++) {
2248 Node* u = c->fast_out(j);
2249 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2250 (u->adr_type() == TypePtr::BOTTOM || _phase->C->get_alias_index(u->adr_type()) == _alias)) {
2251 phi = u;
2252 for (uint k = 1; k < c->req() && phi != NULL; k++) {
2253 Node* m = _memory_nodes[c->in(k)->_idx];
2254 assert(m != NULL, "expect memory state");
2255 if (u->in(k) != m) {
2256 phi = NULL;
2257 }
2258 }
2259 }
2260 }
2261 if (phi == NULL) {
2262 phi = new (_phase->C) PhiNode(c, Type::MEMORY, _phase->C->get_adr_type(_alias));
2263 for (uint k = 1; k < c->req(); k++) {
2264 Node* m = _memory_nodes[c->in(k)->_idx];
2265 assert(m != NULL, "expect memory state");
2266 phi->init_req(k, m);
2267 }
2268 }
2269 }
2270 assert(phi != NULL, "");
2271 regions.map(c->_idx, phi);
2272 }
2273 Node* current_region = regions[c->_idx];
2274 if (current_region != prev_region) {
2275 progress = true;
2276 if (prev_region == prev_mem) {
2277 _memory_nodes.map(c->_idx, current_region);
2278 }
2279 }
2280 } else if (prev_mem == NULL || prev_mem->is_Phi() || ctrl_or_self(prev_mem) != c) {
2281 Node* m = _memory_nodes[_phase->idom(c)->_idx];
2282 assert(m != NULL, "expect memory state");
2283 if (m != prev_mem) {
2284 _memory_nodes.map(c->_idx, m);
2285 progress = true;
2286 }
2287 }
2288 #ifdef ASSERT
2289 if (trace) { tty->print("X %d", c->_idx); _memory_nodes[c->_idx]->dump(); }
2290 #endif
2291 }
2292 }
2293
2294 // Replace existing phi with computed memory state for that region
2295 // if different (could be a new phi or a dominating memory node if
2296 // that phi was found to be useless).
2297 while (dead_phis.size() > 0) {
2298 Node* n = dead_phis.pop();
2299 n->replace_by(_phase->C->top());
2300 n->destruct();
2301 }
2302 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2303 Node* c = rpo_list.at(i);
2304 if (c->is_Region()) {
2305 Node* n = regions[c->_idx];
2306 if (n->is_Phi() && n->_idx >= last && n->in(0) == c) {
2307 _phase->register_new_node(n, c);
2308 }
2309 }
2310 }
2311 for (int i = rpo_list.size() - 1; i >= 0; i--) {
2312 Node* c = rpo_list.at(i);
2313 if (c->is_Region()) {
2314 Node* n = regions[c->_idx];
2315 for (DUIterator_Fast imax, i = c->fast_outs(imax); i < imax; i++) {
2316 Node* u = c->fast_out(i);
2317 if (u->is_Phi() && u->bottom_type() == Type::MEMORY &&
2318 u != n) {
2319 if (u->adr_type() == TypePtr::BOTTOM) {
2320 fix_memory_uses(u, n, n, c);
2321 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2322 _phase->lazy_replace(u, n);
2323 --i; --imax;
2324 }
2325 }
2326 }
2327 }
2328 }
2329 }
2330
2331 Node* MemoryGraphFixer::get_ctrl(Node* n) const {
2332 Node* c = _phase->get_ctrl(n);
2333 if (n->is_Proj() && n->in(0) != NULL && n->in(0)->is_Call()) {
2334 assert(c == n->in(0), "");
2335 CallNode* call = c->as_Call();
2336 CallProjections projs;
2337 call->extract_projections(&projs, true, false);
2338 if (projs.catchall_memproj != NULL) {
2339 if (projs.fallthrough_memproj == n) {
2340 c = projs.fallthrough_catchproj;
2341 } else {
2342 assert(projs.catchall_memproj == n, "");
2343 c = projs.catchall_catchproj;
2344 }
2345 }
2346 }
2347 return c;
2348 }
2349
2350 Node* MemoryGraphFixer::ctrl_or_self(Node* n) const {
2351 if (_phase->has_ctrl(n))
2352 return get_ctrl(n);
2353 else {
2354 assert (n->is_CFG(), "must be a CFG node");
2355 return n;
2356 }
2357 }
2358
2359 bool MemoryGraphFixer::mem_is_valid(Node* m, Node* c) const {
2360 return m != NULL && get_ctrl(m) == c;
2361 }
2362
2363 Node* MemoryGraphFixer::find_mem(Node* ctrl, Node* n) const {
2364 assert(n == NULL || _phase->ctrl_or_self(n) == ctrl, "");
2365 Node* mem = _memory_nodes[ctrl->_idx];
2366 Node* c = ctrl;
2367 while (!mem_is_valid(mem, c) &&
2368 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem))) {
2369 c = _phase->idom(c);
2370 mem = _memory_nodes[c->_idx];
2371 }
2372 if (n != NULL && mem_is_valid(mem, c)) {
2373 while (!ShenandoahBarrierC2Support::is_dominator_same_ctrl(c, mem, n, _phase) && _phase->ctrl_or_self(mem) == ctrl) {
2374 mem = next_mem(mem, _alias);
2375 }
2376 if (mem->is_MergeMem()) {
2377 mem = mem->as_MergeMem()->memory_at(_alias);
2378 }
2379 if (!mem_is_valid(mem, c)) {
2380 do {
2381 c = _phase->idom(c);
2382 mem = _memory_nodes[c->_idx];
2383 } while (!mem_is_valid(mem, c) &&
2384 (!c->is_CatchProj() || mem == NULL || c->in(0)->in(0)->in(0) != get_ctrl(mem)));
2385 }
2386 }
2387 assert(mem->bottom_type() == Type::MEMORY, "");
2388 return mem;
2389 }
2390
2391 bool MemoryGraphFixer::has_mem_phi(Node* region) const {
2392 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) {
2393 Node* use = region->fast_out(i);
2394 if (use->is_Phi() && use->bottom_type() == Type::MEMORY &&
2395 (_phase->C->get_alias_index(use->adr_type()) == _alias)) {
2396 return true;
2397 }
2398 }
2399 return false;
2400 }
2401
2402 void MemoryGraphFixer::fix_mem(Node* ctrl, Node* new_ctrl, Node* mem, Node* mem_for_ctrl, Node* new_mem, Unique_Node_List& uses) {
2403 assert(_phase->ctrl_or_self(new_mem) == new_ctrl, "");
2404 const bool trace = false;
2405 DEBUG_ONLY(if (trace) { tty->print("ZZZ control is"); ctrl->dump(); });
2406 DEBUG_ONLY(if (trace) { tty->print("ZZZ mem is"); mem->dump(); });
2407 GrowableArray<Node*> phis;
2408 if (mem_for_ctrl != mem) {
2409 Node* old = mem_for_ctrl;
2410 Node* prev = NULL;
2411 while (old != mem) {
2412 prev = old;
2413 if (old->is_Store() || old->is_ClearArray() || old->is_LoadStore()) {
2414 assert(_alias == Compile::AliasIdxRaw, "");
2415 old = old->in(MemNode::Memory);
2416 } else if (old->Opcode() == Op_SCMemProj) {
2417 assert(_alias == Compile::AliasIdxRaw, "");
2418 old = old->in(0);
2419 } else {
2420 ShouldNotReachHere();
2421 }
2422 }
2423 assert(prev != NULL, "");
2424 if (new_ctrl != ctrl) {
2425 _memory_nodes.map(ctrl->_idx, mem);
2426 _memory_nodes.map(new_ctrl->_idx, mem_for_ctrl);
2427 }
2428 uint input = (uint)MemNode::Memory;
2429 _phase->igvn().replace_input_of(prev, input, new_mem);
2430 } else {
2431 uses.clear();
2432 _memory_nodes.map(new_ctrl->_idx, new_mem);
2433 uses.push(new_ctrl);
2434 for(uint next = 0; next < uses.size(); next++ ) {
2435 Node *n = uses.at(next);
2436 assert(n->is_CFG(), "");
2437 DEBUG_ONLY(if (trace) { tty->print("ZZZ ctrl"); n->dump(); });
2438 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
2439 Node* u = n->fast_out(i);
2440 if (!u->is_Root() && u->is_CFG() && u != n) {
2441 Node* m = _memory_nodes[u->_idx];
2442 if (u->is_Region() &&
2443 !has_mem_phi(u) &&
2444 u->unique_ctrl_out()->Opcode() != Op_Halt) {
2445 DEBUG_ONLY(if (trace) { tty->print("ZZZ region"); u->dump(); });
2446 DEBUG_ONLY(if (trace && m != NULL) { tty->print("ZZZ mem"); m->dump(); });
2447
2448 if (!mem_is_valid(m, u) || !m->is_Phi()) {
2449 bool push = true;
2450 bool create_phi = true;
2451 if (_phase->is_dominator(new_ctrl, u)) {
2452 create_phi = false;
2453 }
2454 if (create_phi) {
2455 Node* phi = new (_phase->C) PhiNode(u, Type::MEMORY, _phase->C->get_adr_type(_alias));
2456 _phase->register_new_node(phi, u);
2457 phis.push(phi);
2458 DEBUG_ONLY(if (trace) { tty->print("ZZZ new phi"); phi->dump(); });
2459 if (!mem_is_valid(m, u)) {
2460 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting mem"); phi->dump(); });
2461 _memory_nodes.map(u->_idx, phi);
2462 } else {
2463 DEBUG_ONLY(if (trace) { tty->print("ZZZ NOT setting mem"); m->dump(); });
2464 for (;;) {
2465 assert(m->is_Mem() || m->is_LoadStore() || m->is_Proj(), "");
2466 Node* next = NULL;
2467 if (m->is_Proj()) {
2468 next = m->in(0);
2469 } else {
2470 assert(m->is_Mem() || m->is_LoadStore(), "");
2471 assert(_alias == Compile::AliasIdxRaw, "");
2472 next = m->in(MemNode::Memory);
2473 }
2474 if (_phase->get_ctrl(next) != u) {
2475 break;
2476 }
2477 if (next->is_MergeMem()) {
2478 assert(_phase->get_ctrl(next->as_MergeMem()->memory_at(_alias)) != u, "");
2479 break;
2480 }
2481 if (next->is_Phi()) {
2482 assert(next->adr_type() == TypePtr::BOTTOM && next->in(0) == u, "");
2483 break;
2484 }
2485 m = next;
2486 }
2487
2488 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting to phi"); m->dump(); });
2489 assert(m->is_Mem() || m->is_LoadStore(), "");
2490 uint input = (uint)MemNode::Memory;
2491 _phase->igvn().replace_input_of(m, input, phi);
2492 push = false;
2493 }
2494 } else {
2495 DEBUG_ONLY(if (trace) { tty->print("ZZZ skipping region"); u->dump(); });
2496 }
2497 if (push) {
2498 uses.push(u);
2499 }
2500 }
2501 } else if (!mem_is_valid(m, u) &&
2502 !(u->Opcode() == Op_CProj && u->in(0)->Opcode() == Op_NeverBranch && u->as_Proj()->_con == 1)) {
2503 uses.push(u);
2504 }
2505 }
2506 }
2507 }
2508 for (int i = 0; i < phis.length(); i++) {
2509 Node* n = phis.at(i);
2510 Node* r = n->in(0);
2511 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi"); n->dump(); });
2512 for (uint j = 1; j < n->req(); j++) {
2513 Node* m = find_mem(r->in(j), NULL);
2514 _phase->igvn().replace_input_of(n, j, m);
2515 DEBUG_ONLY(if (trace) { tty->print("ZZZ fixing new phi: %d", j); m->dump(); });
2516 }
2517 }
2518 }
2519 uint last = _phase->C->unique();
2520 MergeMemNode* mm = NULL;
2521 int alias = _alias;
2522 DEBUG_ONLY(if (trace) { tty->print("ZZZ raw mem is"); mem->dump(); });
2523 // Process loads first to not miss an anti-dependency: if the memory
2524 // edge of a store is updated before a load is processed then an
2525 // anti-dependency may be missed.
2526 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2527 Node* u = mem->out(i);
2528 if (u->_idx < last && u->is_Load() && _phase->C->get_alias_index(u->adr_type()) == alias) {
2529 Node* m = find_mem(_phase->get_ctrl(u), u);
2530 if (m != mem) {
2531 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2532 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2533 --i;
2534 }
2535 }
2536 }
2537 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2538 Node* u = mem->out(i);
2539 if (u->_idx < last) {
2540 if (u->is_Mem()) {
2541 if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2542 Node* m = find_mem(_phase->get_ctrl(u), u);
2543 if (m != mem) {
2544 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2545 _phase->igvn().replace_input_of(u, MemNode::Memory, m);
2546 --i;
2547 }
2548 }
2549 } else if (u->is_MergeMem()) {
2550 MergeMemNode* u_mm = u->as_MergeMem();
2551 if (u_mm->memory_at(alias) == mem) {
2552 MergeMemNode* newmm = NULL;
2553 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2554 Node* uu = u->fast_out(j);
2555 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2556 if (uu->is_Phi()) {
2557 assert(uu->adr_type() == TypePtr::BOTTOM, "");
2558 Node* region = uu->in(0);
2559 int nb = 0;
2560 for (uint k = 1; k < uu->req(); k++) {
2561 if (uu->in(k) == u) {
2562 Node* m = find_mem(region->in(k), NULL);
2563 if (m != mem) {
2564 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", k); uu->dump(); });
2565 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2566 if (newmm != u) {
2567 _phase->igvn().replace_input_of(uu, k, newmm);
2568 nb++;
2569 --jmax;
2570 }
2571 }
2572 }
2573 }
2574 if (nb > 0) {
2575 --j;
2576 }
2577 } else {
2578 Node* m = find_mem(_phase->ctrl_or_self(uu), uu);
2579 if (m != mem) {
2580 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); uu->dump(); });
2581 newmm = clone_merge_mem(u, mem, m, _phase->ctrl_or_self(m), i);
2582 if (newmm != u) {
2583 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2584 --j, --jmax;
2585 }
2586 }
2587 }
2588 }
2589 }
2590 } else if (u->is_Phi()) {
2591 assert(u->bottom_type() == Type::MEMORY, "what else?");
2592 if (_phase->C->get_alias_index(u->adr_type()) == alias || u->adr_type() == TypePtr::BOTTOM) {
2593 Node* region = u->in(0);
2594 bool replaced = false;
2595 for (uint j = 1; j < u->req(); j++) {
2596 if (u->in(j) == mem) {
2597 Node* m = find_mem(region->in(j), NULL);
2598 Node* nnew = m;
2599 if (m != mem) {
2600 if (u->adr_type() == TypePtr::BOTTOM) {
2601 mm = allocate_merge_mem(mem, m, _phase->ctrl_or_self(m));
2602 nnew = mm;
2603 }
2604 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of phi %d", j); u->dump(); });
2605 _phase->igvn().replace_input_of(u, j, nnew);
2606 replaced = true;
2607 }
2608 }
2609 }
2610 if (replaced) {
2611 --i;
2612 }
2613 }
2614 } else if ((u->adr_type() == TypePtr::BOTTOM) ||
2615 u->adr_type() == NULL) {
2616 assert(u->adr_type() != NULL ||
2617 u->Opcode() == Op_Rethrow ||
2618 u->Opcode() == Op_Return ||
2619 u->Opcode() == Op_SafePoint ||
2620 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2621 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2622 u->Opcode() == Op_CallLeaf, "");
2623 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2624 if (m != mem) {
2625 mm = allocate_merge_mem(mem, m, _phase->get_ctrl(m));
2626 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2627 --i;
2628 }
2629 } else if (_phase->C->get_alias_index(u->adr_type()) == alias) {
2630 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2631 if (m != mem) {
2632 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2633 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2634 --i;
2635 }
2636 } else if (u->adr_type() != TypePtr::BOTTOM &&
2637 _memory_nodes[_phase->ctrl_or_self(u)->_idx] == u) {
2638 Node* m = find_mem(_phase->ctrl_or_self(u), u);
2639 assert(m != mem, "");
2640 // u is on the wrong slice...
2641 assert(u->is_ClearArray(), "");
2642 DEBUG_ONLY(if (trace) { tty->print("ZZZ setting memory of use"); u->dump(); });
2643 _phase->igvn().replace_input_of(u, u->find_edge(mem), m);
2644 --i;
2645 }
2646 }
2647 }
2648 #ifdef ASSERT
2649 assert(new_mem->outcnt() > 0, "");
2650 for (int i = 0; i < phis.length(); i++) {
2651 Node* n = phis.at(i);
2652 assert(n->outcnt() > 0, "new phi must have uses now");
2653 }
2654 #endif
2655 }
2656
2657 MergeMemNode* MemoryGraphFixer::allocate_merge_mem(Node* mem, Node* rep_proj, Node* rep_ctrl) const {
2658 MergeMemNode* mm = MergeMemNode::make(_phase->C, mem);
2659 mm->set_memory_at(_alias, rep_proj);
2660 _phase->register_new_node(mm, rep_ctrl);
2661 return mm;
2662 }
2663
2664 MergeMemNode* MemoryGraphFixer::clone_merge_mem(Node* u, Node* mem, Node* rep_proj, Node* rep_ctrl, DUIterator& i) const {
2665 MergeMemNode* newmm = NULL;
2666 MergeMemNode* u_mm = u->as_MergeMem();
2667 Node* c = _phase->get_ctrl(u);
2668 if (_phase->is_dominator(c, rep_ctrl)) {
2669 c = rep_ctrl;
2670 } else {
2671 assert(_phase->is_dominator(rep_ctrl, c), "one must dominate the other");
2672 }
2673 if (u->outcnt() == 1) {
2674 if (u->req() > (uint)_alias && u->in(_alias) == mem) {
2675 _phase->igvn().replace_input_of(u, _alias, rep_proj);
2676 --i;
2677 } else {
2678 _phase->igvn().rehash_node_delayed(u);
2679 u_mm->set_memory_at(_alias, rep_proj);
2680 }
2681 newmm = u_mm;
2682 _phase->set_ctrl_and_loop(u, c);
2683 } else {
2684 // can't simply clone u and then change one of its input because
2685 // it adds and then removes an edge which messes with the
2686 // DUIterator
2687 newmm = MergeMemNode::make(_phase->C, u_mm->base_memory());
2688 for (uint j = 0; j < u->req(); j++) {
2689 if (j < newmm->req()) {
2690 if (j == (uint)_alias) {
2691 newmm->set_req(j, rep_proj);
2692 } else if (newmm->in(j) != u->in(j)) {
2693 newmm->set_req(j, u->in(j));
2694 }
2695 } else if (j == (uint)_alias) {
2696 newmm->add_req(rep_proj);
2697 } else {
2698 newmm->add_req(u->in(j));
2699 }
2700 }
2701 if ((uint)_alias >= u->req()) {
2702 newmm->set_memory_at(_alias, rep_proj);
2703 }
2704 _phase->register_new_node(newmm, c);
2705 }
2706 return newmm;
2707 }
2708
2709 bool MemoryGraphFixer::should_process_phi(Node* phi) const {
2710 if (phi->adr_type() == TypePtr::BOTTOM) {
2711 Node* region = phi->in(0);
2712 for (DUIterator_Fast jmax, j = region->fast_outs(jmax); j < jmax; j++) {
2713 Node* uu = region->fast_out(j);
2714 if (uu->is_Phi() && uu != phi && uu->bottom_type() == Type::MEMORY && _phase->C->get_alias_index(uu->adr_type()) == _alias) {
2715 return false;
2716 }
2717 }
2718 return true;
2719 }
2720 return _phase->C->get_alias_index(phi->adr_type()) == _alias;
2721 }
2722
2723 void MemoryGraphFixer::fix_memory_uses(Node* mem, Node* replacement, Node* rep_proj, Node* rep_ctrl) const {
2724 uint last = _phase-> C->unique();
2725 MergeMemNode* mm = NULL;
2726 assert(mem->bottom_type() == Type::MEMORY, "");
2727 for (DUIterator i = mem->outs(); mem->has_out(i); i++) {
2728 Node* u = mem->out(i);
2729 if (u != replacement && u->_idx < last) {
2730 if (u->is_MergeMem()) {
2731 MergeMemNode* u_mm = u->as_MergeMem();
2732 if (u_mm->memory_at(_alias) == mem) {
2733 MergeMemNode* newmm = NULL;
2734 for (DUIterator_Fast jmax, j = u->fast_outs(jmax); j < jmax; j++) {
2735 Node* uu = u->fast_out(j);
2736 assert(!uu->is_MergeMem(), "chain of MergeMems?");
2737 if (uu->is_Phi()) {
2738 if (should_process_phi(uu)) {
2739 Node* region = uu->in(0);
2740 int nb = 0;
2741 for (uint k = 1; k < uu->req(); k++) {
2742 if (uu->in(k) == u && _phase->is_dominator(rep_ctrl, region->in(k))) {
2743 if (newmm == NULL) {
2744 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2745 }
2746 if (newmm != u) {
2747 _phase->igvn().replace_input_of(uu, k, newmm);
2748 nb++;
2749 --jmax;
2750 }
2751 }
2752 }
2753 if (nb > 0) {
2754 --j;
2755 }
2756 }
2757 } else {
2758 if (rep_ctrl != uu && ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(uu), replacement, uu, _phase)) {
2759 if (newmm == NULL) {
2760 newmm = clone_merge_mem(u, mem, rep_proj, rep_ctrl, i);
2761 }
2762 if (newmm != u) {
2763 _phase->igvn().replace_input_of(uu, uu->find_edge(u), newmm);
2764 --j, --jmax;
2765 }
2766 }
2767 }
2768 }
2769 }
2770 } else if (u->is_Phi()) {
2771 assert(u->bottom_type() == Type::MEMORY, "what else?");
2772 Node* region = u->in(0);
2773 if (should_process_phi(u)) {
2774 bool replaced = false;
2775 for (uint j = 1; j < u->req(); j++) {
2776 if (u->in(j) == mem && _phase->is_dominator(rep_ctrl, region->in(j))) {
2777 Node* nnew = rep_proj;
2778 if (u->adr_type() == TypePtr::BOTTOM) {
2779 if (mm == NULL) {
2780 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2781 }
2782 nnew = mm;
2783 }
2784 _phase->igvn().replace_input_of(u, j, nnew);
2785 replaced = true;
2786 }
2787 }
2788 if (replaced) {
2789 --i;
2790 }
2791
2792 }
2793 } else if ((u->adr_type() == TypePtr::BOTTOM) ||
2794 u->adr_type() == NULL) {
2795 assert(u->adr_type() != NULL ||
2796 u->Opcode() == Op_Rethrow ||
2797 u->Opcode() == Op_Return ||
2798 u->Opcode() == Op_SafePoint ||
2799 u->Opcode() == Op_StoreIConditional ||
2800 u->Opcode() == Op_StoreLConditional ||
2801 (u->is_CallStaticJava() && u->as_CallStaticJava()->uncommon_trap_request() != 0) ||
2802 (u->is_CallStaticJava() && u->as_CallStaticJava()->_entry_point == OptoRuntime::rethrow_stub()) ||
2803 u->Opcode() == Op_CallLeaf, err_msg("%s", u->Name()));
2804 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2805 if (mm == NULL) {
2806 mm = allocate_merge_mem(mem, rep_proj, rep_ctrl);
2807 }
2808 _phase->igvn().replace_input_of(u, u->find_edge(mem), mm);
2809 --i;
2810 }
2811 } else if (_phase->C->get_alias_index(u->adr_type()) == _alias) {
2812 if (ShenandoahBarrierC2Support::is_dominator(rep_ctrl, _phase->ctrl_or_self(u), replacement, u, _phase)) {
2813 _phase->igvn().replace_input_of(u, u->find_edge(mem), rep_proj);
2814 --i;
2815 }
2816 }
2817 }
2818 }
2819 }
2820
2821 ShenandoahLoadReferenceBarrierNode::ShenandoahLoadReferenceBarrierNode(Node* ctrl, Node* obj)
2822 : Node(ctrl, obj) {
2823 Compile::current()->add_shenandoah_barrier(this);
2824 }
2825
2826 const Type* ShenandoahLoadReferenceBarrierNode::bottom_type() const {
2827 if (in(ValueIn) == NULL || in(ValueIn)->is_top()) {
2828 return Type::TOP;
2829 }
2830 const Type* t = in(ValueIn)->bottom_type();
2831 if (t == TypePtr::NULL_PTR) {
2832 return t;
2833 }
2834 return t->is_oopptr();
2835 }
2836
2837 const Type* ShenandoahLoadReferenceBarrierNode::Value(PhaseTransform *phase) const {
2838 // Either input is TOP ==> the result is TOP
2839 const Type *t2 = phase->type(in(ValueIn));
2840 if( t2 == Type::TOP ) return Type::TOP;
2841
2842 if (t2 == TypePtr::NULL_PTR) {
2843 return t2;
2844 }
2845
2846 const Type* type = t2->is_oopptr();
2847 return type;
2848 }
2849
2850 Node* ShenandoahLoadReferenceBarrierNode::Identity(PhaseTransform *phase) {
2851 Node* value = in(ValueIn);
2852 if (!needs_barrier(phase, value)) {
2853 return value;
2854 }
2855 return this;
2856 }
2857
2858 bool ShenandoahLoadReferenceBarrierNode::needs_barrier(PhaseTransform* phase, Node* n) {
2859 Unique_Node_List visited;
2860 return needs_barrier_impl(phase, n, visited);
2861 }
2862
2863 bool ShenandoahLoadReferenceBarrierNode::needs_barrier_impl(PhaseTransform* phase, Node* n, Unique_Node_List &visited) {
2864 if (n == NULL) return false;
2865 if (visited.member(n)) {
2866 return false; // Been there.
2867 }
2868 visited.push(n);
2869
2870 if (n->is_Allocate()) {
2871 // tty->print_cr("optimize barrier on alloc");
2872 return false;
2873 }
2874 if (n->is_Call()) {
2875 // tty->print_cr("optimize barrier on call");
2876 return false;
2877 }
2878
2879 const Type* type = phase->type(n);
2880 if (type == Type::TOP) {
2881 return false;
2882 }
2883 if (type->make_ptr()->higher_equal(TypePtr::NULL_PTR)) {
2884 // tty->print_cr("optimize barrier on null");
2885 return false;
2886 }
2887 // Impl detail: Need to check isa_(narrow)oop before calling to make_oopptr on potentially non-oop types
2888 // in 8u, otherwise make_oopptr would assert. make_oopptr is fixed later during JDK-8078629.
2889 if ((type->isa_oopptr() || type->isa_narrowoop()) && type->make_oopptr()->const_oop() != NULL) {
2890 // tty->print_cr("optimize barrier on constant");
2891 return false;
2892 }
2893
2894 switch (n->Opcode()) {
2895 case Op_AddP:
2896 return true; // TODO: Can refine?
2897 case Op_LoadP:
2898 case Op_GetAndSetN:
2899 case Op_GetAndSetP:
2900 return true;
2901 case Op_Phi: {
2902 for (uint i = 1; i < n->req(); i++) {
2903 if (needs_barrier_impl(phase, n->in(i), visited)) return true;
2904 }
2905 return false;
2906 }
2907 case Op_CheckCastPP:
2908 case Op_CastPP:
2909 return needs_barrier_impl(phase, n->in(1), visited);
2910 case Op_Proj:
2911 return needs_barrier_impl(phase, n->in(0), visited);
2912 case Op_ShenandoahLoadReferenceBarrier:
2913 // tty->print_cr("optimize barrier on barrier");
2914 return false;
2915 case Op_Parm:
2916 // tty->print_cr("optimize barrier on input arg");
2917 return false;
2918 case Op_DecodeN:
2919 case Op_EncodeP:
2920 return needs_barrier_impl(phase, n->in(1), visited);
2921 case Op_LoadN:
2922 return true;
2923 case Op_CMoveN:
2924 case Op_CMoveP:
2925 return needs_barrier_impl(phase, n->in(2), visited) ||
2926 needs_barrier_impl(phase, n->in(3), visited);
2927 case Op_CreateEx:
2928 return false;
2929 default:
2930 break;
2931 }
2932 #ifdef ASSERT
2933 tty->print("need barrier on?: ");
2934 tty->print_cr("ins:");
2935 n->dump(2);
2936 tty->print_cr("outs:");
2937 n->dump(-2);
2938 ShouldNotReachHere();
2939 #endif
2940 return true;
2941 }
2942
2943 bool ShenandoahLoadReferenceBarrierNode::is_redundant() {
2944 Unique_Node_List visited;
2945 Node_Stack stack(0);
2946 stack.push(this, 0);
2947
2948 // Check if the barrier is actually useful: go over nodes looking for useful uses
2949 // (e.g. memory accesses). Stop once we detected a required use. Otherwise, walk
2950 // until we ran out of nodes, and then declare the barrier redundant.
2951 while (stack.size() > 0) {
2952 Node* n = stack.node();
2953 if (visited.member(n)) {
2954 stack.pop();
2955 continue;
2956 }
2957 visited.push(n);
2958 bool visit_users = false;
2959 switch (n->Opcode()) {
2960 case Op_CallStaticJava:
2961 case Op_CallDynamicJava:
2962 case Op_CallLeaf:
2963 case Op_CallLeafNoFP:
2964 case Op_CompareAndSwapL:
2965 case Op_CompareAndSwapI:
2966 case Op_CompareAndSwapN:
2967 case Op_CompareAndSwapP:
2968 case Op_ShenandoahCompareAndSwapN:
2969 case Op_ShenandoahCompareAndSwapP:
2970 case Op_GetAndSetL:
2971 case Op_GetAndSetI:
2972 case Op_GetAndSetP:
2973 case Op_GetAndSetN:
2974 case Op_GetAndAddL:
2975 case Op_GetAndAddI:
2976 case Op_FastLock:
2977 case Op_FastUnlock:
2978 case Op_Rethrow:
2979 case Op_Return:
2980 case Op_StoreB:
2981 case Op_StoreC:
2982 case Op_StoreD:
2983 case Op_StoreF:
2984 case Op_StoreL:
2985 case Op_StoreLConditional:
2986 case Op_StoreI:
2987 case Op_StoreIConditional:
2988 case Op_StoreN:
2989 case Op_StoreP:
2990 case Op_StoreVector:
2991 case Op_EncodeP:
2992 case Op_CastP2X:
2993 case Op_SafePoint:
2994 case Op_EncodeISOArray:
2995 case Op_AryEq:
2996 case Op_StrEquals:
2997 case Op_StrComp:
2998 case Op_StrIndexOf:
2999 // Known to require barriers
3000 return false;
3001 case Op_CmpP: {
3002 if (n->in(1)->bottom_type()->higher_equal(TypePtr::NULL_PTR) ||
3003 n->in(2)->bottom_type()->higher_equal(TypePtr::NULL_PTR)) {
3004 // One of the sides is known null, no need for barrier.
3005 } else {
3006 return false;
3007 }
3008 break;
3009 }
3010 case Op_LoadB:
3011 case Op_LoadUB:
3012 case Op_LoadUS:
3013 case Op_LoadD:
3014 case Op_LoadF:
3015 case Op_LoadL:
3016 case Op_LoadI:
3017 case Op_LoadS:
3018 case Op_LoadN:
3019 case Op_LoadP:
3020 case Op_LoadVector: {
3021 const TypePtr* adr_type = n->adr_type();
3022 int alias_idx = Compile::current()->get_alias_index(adr_type);
3023 Compile::AliasType* alias_type = Compile::current()->alias_type(alias_idx);
3024 ciField* field = alias_type->field();
3025 bool is_static = field != NULL && field->is_static();
3026 bool is_final = field != NULL && field->is_final();
3027
3028 if (ShenandoahOptimizeStaticFinals && is_static && is_final) {
3029 // Loading the constant does not require barriers: it should be handled
3030 // as part of GC roots already.
3031 } else {
3032 return false;
3033 }
3034 break;
3035 }
3036 case Op_Conv2B:
3037 case Op_LoadRange:
3038 case Op_LoadKlass:
3039 case Op_LoadNKlass:
3040 // Do not require barriers
3041 break;
3042 case Op_AddP:
3043 case Op_CheckCastPP:
3044 case Op_CastPP:
3045 case Op_CMoveP:
3046 case Op_Phi:
3047 case Op_ShenandoahLoadReferenceBarrier:
3048 // Whether or not these need the barriers depends on their users
3049 visit_users = true;
3050 break;
3051 default: {
3052 #ifdef ASSERT
3053 fatal(err_msg("Unknown node in is_redundant: %s", NodeClassNames[n->Opcode()]));
3054 #else
3055 // Default to have excess barriers, rather than miss some.
3056 return false;
3057 #endif
3058 }
3059 }
3060
3061 stack.pop();
3062 if (visit_users) {
3063 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
3064 Node* user = n->fast_out(i);
3065 if (user != NULL) {
3066 stack.push(user, 0);
3067 }
3068 }
3069 }
3070 }
3071
3072 // No need for barrier found.
3073 return true;
3074 }
3075
3076 CallStaticJavaNode* ShenandoahLoadReferenceBarrierNode::pin_and_expand_null_check(PhaseIterGVN& igvn) {
3077 Node* val = in(ValueIn);
3078
3079 const Type* val_t = igvn.type(val);
3080
3081 if (val_t->meet(TypePtr::NULL_PTR) != val_t &&
3082 val->Opcode() == Op_CastPP &&
3083 val->in(0) != NULL &&
3084 val->in(0)->Opcode() == Op_IfTrue &&
3085 val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none) &&
3086 val->in(0)->in(0)->is_If() &&
3087 val->in(0)->in(0)->in(1)->Opcode() == Op_Bool &&
3088 val->in(0)->in(0)->in(1)->as_Bool()->_test._test == BoolTest::ne &&
3089 val->in(0)->in(0)->in(1)->in(1)->Opcode() == Op_CmpP &&
3090 val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1) &&
3091 val->in(0)->in(0)->in(1)->in(1)->in(2)->bottom_type() == TypePtr::NULL_PTR) {
3092 assert(val->in(0)->in(0)->in(1)->in(1)->in(1) == val->in(1), "");
3093 CallStaticJavaNode* unc = val->in(0)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none);
3094 return unc;
3095 }
3096 return NULL;
3097 }