1 /*
2 * Copyright (c) 2018, 2019, Red Hat, Inc. All rights reserved.
3 *
4 * This code is free software; you can redistribute it and/or modify it
5 * under the terms of the GNU General Public License version 2 only, as
6 * published by the Free Software Foundation.
7 *
8 * This code is distributed in the hope that it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
11 * version 2 for more details (a copy is included in the LICENSE file that
12 * accompanied this code).
13 *
14 * You should have received a copy of the GNU General Public License version
15 * 2 along with this work; if not, write to the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
17 *
18 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
19 * or visit www.oracle.com if you need additional information or have any
20 * questions.
21 *
22 */
23
24 #include "precompiled.hpp"
25 #include "gc/shared/barrierSet.hpp"
26 #include "gc/shenandoah/shenandoahHeap.hpp"
27 #include "gc/shenandoah/shenandoahHeuristics.hpp"
28 #include "gc/shenandoah/shenandoahRuntime.hpp"
29 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
30 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
31 #include "gc/shenandoah/c2/shenandoahSupport.hpp"
32 #include "opto/arraycopynode.hpp"
33 #include "opto/escape.hpp"
34 #include "opto/graphKit.hpp"
35 #include "opto/idealKit.hpp"
36 #include "opto/macro.hpp"
37 #include "opto/movenode.hpp"
38 #include "opto/narrowptrnode.hpp"
39 #include "opto/rootnode.hpp"
40
41 ShenandoahBarrierSetC2* ShenandoahBarrierSetC2::bsc2() {
42 return reinterpret_cast<ShenandoahBarrierSetC2*>(BarrierSet::barrier_set()->barrier_set_c2());
43 }
44
45 ShenandoahBarrierSetC2State::ShenandoahBarrierSetC2State(Arena* comp_arena)
46 : _shenandoah_barriers(new (comp_arena) GrowableArray<ShenandoahWriteBarrierNode*>(comp_arena, 8, 0, NULL)) {
47 }
48
49 int ShenandoahBarrierSetC2State::shenandoah_barriers_count() const {
50 return _shenandoah_barriers->length();
51 }
52
53 ShenandoahWriteBarrierNode* ShenandoahBarrierSetC2State::shenandoah_barrier(int idx) const {
54 return _shenandoah_barriers->at(idx);
55 }
56
57 void ShenandoahBarrierSetC2State::add_shenandoah_barrier(ShenandoahWriteBarrierNode * n) {
58 assert(!_shenandoah_barriers->contains(n), "duplicate entry in barrier list");
59 _shenandoah_barriers->append(n);
60 }
61
62 void ShenandoahBarrierSetC2State::remove_shenandoah_barrier(ShenandoahWriteBarrierNode * n) {
63 if (_shenandoah_barriers->contains(n)) {
64 _shenandoah_barriers->remove(n);
65 }
66 }
67
68 #define __ kit->
69
70 Node* ShenandoahBarrierSetC2::shenandoah_read_barrier(GraphKit* kit, Node* obj) const {
71 if (ShenandoahReadBarrier) {
72 obj = shenandoah_read_barrier_impl(kit, obj, false, true, true);
73 }
74 return obj;
75 }
76
77 Node* ShenandoahBarrierSetC2::shenandoah_storeval_barrier(GraphKit* kit, Node* obj) const {
78 if (ShenandoahStoreValEnqueueBarrier) {
79 obj = shenandoah_write_barrier(kit, obj);
80 obj = shenandoah_enqueue_barrier(kit, obj);
81 }
82 if (ShenandoahStoreValReadBarrier) {
83 obj = shenandoah_read_barrier_impl(kit, obj, true, false, false);
84 }
85 return obj;
86 }
87
88 Node* ShenandoahBarrierSetC2::shenandoah_read_barrier_impl(GraphKit* kit, Node* obj, bool use_ctrl, bool use_mem, bool allow_fromspace) const {
89 const Type* obj_type = obj->bottom_type();
90 if (obj_type->higher_equal(TypePtr::NULL_PTR)) {
91 return obj;
92 }
93 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(obj_type);
94 Node* mem = use_mem ? __ memory(adr_type) : __ immutable_memory();
95
96 if (! ShenandoahBarrierNode::needs_barrier(&__ gvn(), NULL, obj, mem, allow_fromspace)) {
97 // We know it is null, no barrier needed.
98 return obj;
99 }
100
101 if (obj_type->meet(TypePtr::NULL_PTR) == obj_type->remove_speculative()) {
102
103 // We don't know if it's null or not. Need null-check.
104 enum { _not_null_path = 1, _null_path, PATH_LIMIT };
105 RegionNode* region = new RegionNode(PATH_LIMIT);
106 Node* phi = new PhiNode(region, obj_type);
107 Node* null_ctrl = __ top();
108 Node* not_null_obj = __ null_check_oop(obj, &null_ctrl);
109
110 region->init_req(_null_path, null_ctrl);
111 phi ->init_req(_null_path, __ zerocon(T_OBJECT));
112
113 Node* ctrl = use_ctrl ? __ control() : NULL;
114 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, not_null_obj, allow_fromspace);
115 Node* n = __ gvn().transform(rb);
116
117 region->init_req(_not_null_path, __ control());
118 phi ->init_req(_not_null_path, n);
119
120 __ set_control(__ gvn().transform(region));
121 __ record_for_igvn(region);
122 return __ gvn().transform(phi);
123
124 } else {
125 // We know it is not null. Simple barrier is sufficient.
126 Node* ctrl = use_ctrl ? __ control() : NULL;
127 ShenandoahReadBarrierNode* rb = new ShenandoahReadBarrierNode(ctrl, mem, obj, allow_fromspace);
128 Node* n = __ gvn().transform(rb);
129 __ record_for_igvn(n);
130 return n;
131 }
132 }
133
134 Node* ShenandoahBarrierSetC2::shenandoah_write_barrier_helper(GraphKit* kit, Node* obj, const TypePtr* adr_type) const {
135 ShenandoahWriteBarrierNode* wb = new ShenandoahWriteBarrierNode(kit->C, kit->control(), kit->memory(adr_type), obj);
136 Node* n = __ gvn().transform(wb);
137 if (n == wb) { // New barrier needs memory projection.
138 Node* proj = __ gvn().transform(new ShenandoahWBMemProjNode(n));
139 __ set_memory(proj, adr_type);
140 }
141 return n;
142 }
143
144 Node* ShenandoahBarrierSetC2::shenandoah_write_barrier(GraphKit* kit, Node* obj) const {
145 if (ShenandoahWriteBarrier) {
146 obj = shenandoah_write_barrier_impl(kit, obj);
147 }
148 return obj;
149 }
150
151 Node* ShenandoahBarrierSetC2::shenandoah_write_barrier_impl(GraphKit* kit, Node* obj) const {
152 if (! ShenandoahBarrierNode::needs_barrier(&__ gvn(), NULL, obj, NULL, true)) {
153 return obj;
154 }
155 const Type* obj_type = obj->bottom_type();
156 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(obj_type);
157 Node* n = shenandoah_write_barrier_helper(kit, obj, adr_type);
158 __ record_for_igvn(n);
159 return n;
160 }
161
162 bool ShenandoahBarrierSetC2::satb_can_remove_pre_barrier(GraphKit* kit, PhaseTransform* phase, Node* adr,
163 BasicType bt, uint adr_idx) const {
164 intptr_t offset = 0;
165 Node* base = AddPNode::Ideal_base_and_offset(adr, phase, offset);
166 AllocateNode* alloc = AllocateNode::Ideal_allocation(base, phase);
167
168 if (offset == Type::OffsetBot) {
169 return false; // cannot unalias unless there are precise offsets
170 }
171
172 if (alloc == NULL) {
173 return false; // No allocation found
174 }
175
176 intptr_t size_in_bytes = type2aelembytes(bt);
177
178 Node* mem = __ memory(adr_idx); // start searching here...
179
180 for (int cnt = 0; cnt < 50; cnt++) {
181
182 if (mem->is_Store()) {
183
184 Node* st_adr = mem->in(MemNode::Address);
185 intptr_t st_offset = 0;
186 Node* st_base = AddPNode::Ideal_base_and_offset(st_adr, phase, st_offset);
187
188 if (st_base == NULL) {
189 break; // inscrutable pointer
190 }
191
192 // Break we have found a store with same base and offset as ours so break
193 if (st_base == base && st_offset == offset) {
194 break;
195 }
196
197 if (st_offset != offset && st_offset != Type::OffsetBot) {
198 const int MAX_STORE = BytesPerLong;
199 if (st_offset >= offset + size_in_bytes ||
200 st_offset <= offset - MAX_STORE ||
201 st_offset <= offset - mem->as_Store()->memory_size()) {
202 // Success: The offsets are provably independent.
203 // (You may ask, why not just test st_offset != offset and be done?
204 // The answer is that stores of different sizes can co-exist
205 // in the same sequence of RawMem effects. We sometimes initialize
206 // a whole 'tile' of array elements with a single jint or jlong.)
207 mem = mem->in(MemNode::Memory);
208 continue; // advance through independent store memory
209 }
210 }
211
212 if (st_base != base
213 && MemNode::detect_ptr_independence(base, alloc, st_base,
214 AllocateNode::Ideal_allocation(st_base, phase),
215 phase)) {
216 // Success: The bases are provably independent.
217 mem = mem->in(MemNode::Memory);
218 continue; // advance through independent store memory
219 }
220 } else if (mem->is_Proj() && mem->in(0)->is_Initialize()) {
221
222 InitializeNode* st_init = mem->in(0)->as_Initialize();
223 AllocateNode* st_alloc = st_init->allocation();
224
225 // Make sure that we are looking at the same allocation site.
226 // The alloc variable is guaranteed to not be null here from earlier check.
227 if (alloc == st_alloc) {
228 // Check that the initialization is storing NULL so that no previous store
229 // has been moved up and directly write a reference
230 Node* captured_store = st_init->find_captured_store(offset,
231 type2aelembytes(T_OBJECT),
232 phase);
233 if (captured_store == NULL || captured_store == st_init->zero_memory()) {
234 return true;
235 }
236 }
237 }
238
239 // Unless there is an explicit 'continue', we must bail out here,
240 // because 'mem' is an inscrutable memory state (e.g., a call).
241 break;
242 }
243
244 return false;
245 }
246
247 #undef __
248 #define __ ideal.
249
250 void ShenandoahBarrierSetC2::satb_write_barrier_pre(GraphKit* kit,
251 bool do_load,
252 Node* obj,
253 Node* adr,
254 uint alias_idx,
255 Node* val,
256 const TypeOopPtr* val_type,
257 Node* pre_val,
258 BasicType bt) const {
259 // Some sanity checks
260 // Note: val is unused in this routine.
261
262 if (do_load) {
263 // We need to generate the load of the previous value
264 assert(obj != NULL, "must have a base");
265 assert(adr != NULL, "where are loading from?");
266 assert(pre_val == NULL, "loaded already?");
267 assert(val_type != NULL, "need a type");
268
269 if (ReduceInitialCardMarks
270 && satb_can_remove_pre_barrier(kit, &kit->gvn(), adr, bt, alias_idx)) {
271 return;
272 }
273
274 } else {
275 // In this case both val_type and alias_idx are unused.
276 assert(pre_val != NULL, "must be loaded already");
277 // Nothing to be done if pre_val is null.
278 if (pre_val->bottom_type() == TypePtr::NULL_PTR) return;
279 assert(pre_val->bottom_type()->basic_type() == T_OBJECT, "or we shouldn't be here");
280 }
281 assert(bt == T_OBJECT, "or we shouldn't be here");
282
283 IdealKit ideal(kit, true);
284
285 Node* tls = __ thread(); // ThreadLocalStorage
286
287 Node* no_base = __ top();
288 Node* zero = __ ConI(0);
289 Node* zeroX = __ ConX(0);
290
291 float likely = PROB_LIKELY(0.999);
292 float unlikely = PROB_UNLIKELY(0.999);
293
294 // Offsets into the thread
295 const int index_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset());
296 const int buffer_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
297
298 // Now the actual pointers into the thread
299 Node* buffer_adr = __ AddP(no_base, tls, __ ConX(buffer_offset));
300 Node* index_adr = __ AddP(no_base, tls, __ ConX(index_offset));
301
302 // Now some of the values
303 Node* marking;
304 Node* gc_state = __ AddP(no_base, tls, __ ConX(in_bytes(ShenandoahThreadLocalData::gc_state_offset())));
305 Node* ld = __ load(__ ctrl(), gc_state, TypeInt::BYTE, T_BYTE, Compile::AliasIdxRaw);
306 marking = __ AndI(ld, __ ConI(ShenandoahHeap::MARKING));
307 assert(ShenandoahWriteBarrierNode::is_gc_state_load(ld), "Should match the shape");
308
309 // if (!marking)
310 __ if_then(marking, BoolTest::ne, zero, unlikely); {
311 BasicType index_bt = TypeX_X->basic_type();
312 assert(sizeof(size_t) == type2aelembytes(index_bt), "Loading G1 SATBMarkQueue::_index with wrong size.");
313 Node* index = __ load(__ ctrl(), index_adr, TypeX_X, index_bt, Compile::AliasIdxRaw);
314
315 if (do_load) {
316 // load original value
317 // alias_idx correct??
318 pre_val = __ load(__ ctrl(), adr, val_type, bt, alias_idx);
319 }
320
321 // if (pre_val != NULL)
322 __ if_then(pre_val, BoolTest::ne, kit->null()); {
323 Node* buffer = __ load(__ ctrl(), buffer_adr, TypeRawPtr::NOTNULL, T_ADDRESS, Compile::AliasIdxRaw);
324
325 // is the queue for this thread full?
326 __ if_then(index, BoolTest::ne, zeroX, likely); {
327
328 // decrement the index
329 Node* next_index = kit->gvn().transform(new SubXNode(index, __ ConX(sizeof(intptr_t))));
330
331 // Now get the buffer location we will log the previous value into and store it
332 Node *log_addr = __ AddP(no_base, buffer, next_index);
333 __ store(__ ctrl(), log_addr, pre_val, T_OBJECT, Compile::AliasIdxRaw, MemNode::unordered);
334 // update the index
335 __ store(__ ctrl(), index_adr, next_index, index_bt, Compile::AliasIdxRaw, MemNode::unordered);
336
337 } __ else_(); {
338
339 // logging buffer is full, call the runtime
340 const TypeFunc *tf = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type();
341 __ make_leaf_call(tf, CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), "shenandoah_wb_pre", pre_val, tls);
342 } __ end_if(); // (!index)
343 } __ end_if(); // (pre_val != NULL)
344 } __ end_if(); // (!marking)
345
346 // Final sync IdealKit and GraphKit.
347 kit->final_sync(ideal);
348
349 if (ShenandoahSATBBarrier && adr != NULL) {
350 Node* c = kit->control();
351 Node* call = c->in(1)->in(1)->in(1)->in(0);
352 assert(is_shenandoah_wb_pre_call(call), "shenandoah_wb_pre call expected");
353 call->add_req(adr);
354 }
355 }
356
357 bool ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(Node* call) {
358 return call->is_CallLeaf() &&
359 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry);
360 }
361
362 bool ShenandoahBarrierSetC2::is_shenandoah_wb_call(Node* call) {
363 return call->is_CallLeaf() &&
364 call->as_CallLeaf()->entry_point() == CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_JRT);
365 }
366
367 bool ShenandoahBarrierSetC2::is_shenandoah_marking_if(PhaseTransform *phase, Node* n) {
368 if (n->Opcode() != Op_If) {
369 return false;
370 }
371
372 Node* bol = n->in(1);
373 assert(bol->is_Bool(), "");
374 Node* cmpx = bol->in(1);
375 if (bol->as_Bool()->_test._test == BoolTest::ne &&
376 cmpx->is_Cmp() && cmpx->in(2) == phase->intcon(0) &&
377 is_shenandoah_state_load(cmpx->in(1)->in(1)) &&
378 cmpx->in(1)->in(2)->is_Con() &&
379 cmpx->in(1)->in(2) == phase->intcon(ShenandoahHeap::MARKING)) {
380 return true;
381 }
382
383 return false;
384 }
385
386 bool ShenandoahBarrierSetC2::is_shenandoah_state_load(Node* n) {
387 if (!n->is_Load()) return false;
388 const int state_offset = in_bytes(ShenandoahThreadLocalData::gc_state_offset());
389 return n->in(2)->is_AddP() && n->in(2)->in(2)->Opcode() == Op_ThreadLocal
390 && n->in(2)->in(3)->is_Con()
391 && n->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == state_offset;
392 }
393
394 void ShenandoahBarrierSetC2::shenandoah_write_barrier_pre(GraphKit* kit,
395 bool do_load,
396 Node* obj,
397 Node* adr,
398 uint alias_idx,
399 Node* val,
400 const TypeOopPtr* val_type,
401 Node* pre_val,
402 BasicType bt) const {
403 if (ShenandoahSATBBarrier) {
404 IdealKit ideal(kit);
405 kit->sync_kit(ideal);
406
407 satb_write_barrier_pre(kit, do_load, obj, adr, alias_idx, val, val_type, pre_val, bt);
408
409 ideal.sync_kit(kit);
410 kit->final_sync(ideal);
411 }
412 }
413
414 Node* ShenandoahBarrierSetC2::shenandoah_enqueue_barrier(GraphKit* kit, Node* pre_val) const {
415 return kit->gvn().transform(new ShenandoahEnqueueBarrierNode(pre_val));
416 }
417
418 // Helper that guards and inserts a pre-barrier.
419 void ShenandoahBarrierSetC2::insert_pre_barrier(GraphKit* kit, Node* base_oop, Node* offset,
420 Node* pre_val, bool need_mem_bar) const {
421 // We could be accessing the referent field of a reference object. If so, when G1
422 // is enabled, we need to log the value in the referent field in an SATB buffer.
423 // This routine performs some compile time filters and generates suitable
424 // runtime filters that guard the pre-barrier code.
425 // Also add memory barrier for non volatile load from the referent field
426 // to prevent commoning of loads across safepoint.
427
428 // Some compile time checks.
429
430 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
431 const TypeX* otype = offset->find_intptr_t_type();
432 if (otype != NULL && otype->is_con() &&
433 otype->get_con() != java_lang_ref_Reference::referent_offset) {
434 // Constant offset but not the reference_offset so just return
435 return;
436 }
437
438 // We only need to generate the runtime guards for instances.
439 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
440 if (btype != NULL) {
441 if (btype->isa_aryptr()) {
442 // Array type so nothing to do
443 return;
444 }
445
446 const TypeInstPtr* itype = btype->isa_instptr();
447 if (itype != NULL) {
448 // Can the klass of base_oop be statically determined to be
449 // _not_ a sub-class of Reference and _not_ Object?
450 ciKlass* klass = itype->klass();
451 if ( klass->is_loaded() &&
452 !klass->is_subtype_of(kit->env()->Reference_klass()) &&
453 !kit->env()->Object_klass()->is_subtype_of(klass)) {
454 return;
455 }
456 }
457 }
458
459 // The compile time filters did not reject base_oop/offset so
460 // we need to generate the following runtime filters
461 //
462 // if (offset == java_lang_ref_Reference::_reference_offset) {
463 // if (instance_of(base, java.lang.ref.Reference)) {
464 // pre_barrier(_, pre_val, ...);
465 // }
466 // }
467
468 float likely = PROB_LIKELY( 0.999);
469 float unlikely = PROB_UNLIKELY(0.999);
470
471 IdealKit ideal(kit);
472
473 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
474
475 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
476 // Update graphKit memory and control from IdealKit.
477 kit->sync_kit(ideal);
478
479 Node* ref_klass_con = kit->makecon(TypeKlassPtr::make(kit->env()->Reference_klass()));
480 Node* is_instof = kit->gen_instanceof(base_oop, ref_klass_con);
481
482 // Update IdealKit memory and control from graphKit.
483 __ sync_kit(kit);
484
485 Node* one = __ ConI(1);
486 // is_instof == 0 if base_oop == NULL
487 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
488
489 // Update graphKit from IdeakKit.
490 kit->sync_kit(ideal);
491
492 // Use the pre-barrier to record the value in the referent field
493 satb_write_barrier_pre(kit, false /* do_load */,
494 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
495 pre_val /* pre_val */,
496 T_OBJECT);
497 if (need_mem_bar) {
498 // Add memory barrier to prevent commoning reads from this field
499 // across safepoint since GC can change its value.
500 kit->insert_mem_bar(Op_MemBarCPUOrder);
501 }
502 // Update IdealKit from graphKit.
503 __ sync_kit(kit);
504
505 } __ end_if(); // _ref_type != ref_none
506 } __ end_if(); // offset == referent_offset
507
508 // Final sync IdealKit and GraphKit.
509 kit->final_sync(ideal);
510 }
511
512 #undef __
513
514 const TypeFunc* ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type() {
515 const Type **fields = TypeTuple::fields(2);
516 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
517 fields[TypeFunc::Parms+1] = TypeRawPtr::NOTNULL; // thread
518 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+2, fields);
519
520 // create result type (range)
521 fields = TypeTuple::fields(0);
522 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
523
524 return TypeFunc::make(domain, range);
525 }
526
527 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type() {
528 const Type **fields = TypeTuple::fields(1);
529 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
530 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
531
532 // create result type (range)
533 fields = TypeTuple::fields(0);
534 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+0, fields);
535
536 return TypeFunc::make(domain, range);
537 }
538
539 const TypeFunc* ShenandoahBarrierSetC2::shenandoah_write_barrier_Type() {
540 const Type **fields = TypeTuple::fields(1);
541 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL; // original field value
542 const TypeTuple *domain = TypeTuple::make(TypeFunc::Parms+1, fields);
543
544 // create result type (range)
545 fields = TypeTuple::fields(1);
546 fields[TypeFunc::Parms+0] = TypeInstPtr::NOTNULL;
547 const TypeTuple *range = TypeTuple::make(TypeFunc::Parms+1, fields);
548
549 return TypeFunc::make(domain, range);
550 }
551
552 void ShenandoahBarrierSetC2::resolve_address(C2Access& access) const {
553 const TypePtr* adr_type = access.addr().type();
554
555 if ((access.decorators() & IN_NATIVE) == 0 && (adr_type->isa_instptr() || adr_type->isa_aryptr())) {
556 int off = adr_type->is_ptr()->offset();
557 int base_off = adr_type->isa_instptr() ? instanceOopDesc::base_offset_in_bytes() :
558 arrayOopDesc::base_offset_in_bytes(adr_type->is_aryptr()->elem()->array_element_basic_type());
559 assert(off != Type::OffsetTop, "unexpected offset");
560 if (off == Type::OffsetBot || off >= base_off) {
561 DecoratorSet decorators = access.decorators();
562 bool is_write = (decorators & C2_WRITE_ACCESS) != 0;
563 GraphKit* kit = NULL;
564 if (access.is_parse_access()) {
565 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
566 kit = parse_access.kit();
567 }
568 Node* adr = access.addr().node();
569 assert(adr->is_AddP(), "unexpected address shape");
570 Node* base = adr->in(AddPNode::Base);
571
572 if (is_write) {
573 if (kit != NULL) {
574 base = shenandoah_write_barrier(kit, base);
575 } else {
576 assert(access.is_opt_access(), "either parse or opt access");
577 assert((access.decorators() & C2_ARRAY_COPY) != 0, "can be skipped for clone");
578 }
579 } else {
580 if (adr_type->isa_instptr()) {
581 Compile* C = access.gvn().C;
582 ciField* field = C->alias_type(adr_type)->field();
583
584 // Insert read barrier for Shenandoah.
585 if (field != NULL &&
586 ((ShenandoahOptimizeStaticFinals && field->is_static() && field->is_final()) ||
587 (ShenandoahOptimizeInstanceFinals && !field->is_static() && field->is_final()) ||
588 (ShenandoahOptimizeStableFinals && field->is_stable()))) {
589 // Skip the barrier for special fields
590 } else {
591 if (kit != NULL) {
592 base = shenandoah_read_barrier(kit, base);
593 } else {
594 assert(access.is_opt_access(), "either parse or opt access");
595 assert((access.decorators() & C2_ARRAY_COPY) != 0, "can be skipped for arraycopy");
596 }
597 }
598 } else {
599 if (kit != NULL) {
600 base = shenandoah_read_barrier(kit, base);
601 } else {
602 assert(access.is_opt_access(), "either parse or opt access");
603 assert((access.decorators() & C2_ARRAY_COPY) != 0, "can be skipped for arraycopy");
604 }
605 }
606 }
607 if (base != adr->in(AddPNode::Base)) {
608 assert(kit != NULL, "no barrier should have been added");
609
610 Node* address = adr->in(AddPNode::Address);
611
612 if (address->is_AddP()) {
613 assert(address->in(AddPNode::Base) == adr->in(AddPNode::Base), "unexpected address shape");
614 assert(!address->in(AddPNode::Address)->is_AddP(), "unexpected address shape");
615 assert(address->in(AddPNode::Address) == adr->in(AddPNode::Base), "unexpected address shape");
616 address = address->clone();
617 address->set_req(AddPNode::Base, base);
618 address->set_req(AddPNode::Address, base);
619 address = kit->gvn().transform(address);
620 } else {
621 assert(address == adr->in(AddPNode::Base), "unexpected address shape");
622 address = base;
623 }
624 adr = adr->clone();
625 adr->set_req(AddPNode::Base, base);
626 adr->set_req(AddPNode::Address, address);
627 adr = kit->gvn().transform(adr);
628 access.addr().set_node(adr);
629 }
630 }
631 }
632 }
633
634 Node* ShenandoahBarrierSetC2::store_at_resolved(C2Access& access, C2AccessValue& val) const {
635 DecoratorSet decorators = access.decorators();
636
637 const TypePtr* adr_type = access.addr().type();
638 Node* adr = access.addr().node();
639
640 bool anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
641 bool on_heap = (decorators & IN_HEAP) != 0;
642
643 if (!access.is_oop() || (!on_heap && !anonymous)) {
644 return BarrierSetC2::store_at_resolved(access, val);
645 }
646
647 if (access.is_parse_access()) {
648 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
649 GraphKit* kit = parse_access.kit();
650
651 uint adr_idx = kit->C->get_alias_index(adr_type);
652 assert(adr_idx != Compile::AliasIdxTop, "use other store_to_memory factory" );
653 Node* value = val.node();
654 value = shenandoah_storeval_barrier(kit, value);
655 val.set_node(value);
656 shenandoah_write_barrier_pre(kit, true /* do_load */, /*kit->control(),*/ access.base(), adr, adr_idx, val.node(),
657 static_cast<const TypeOopPtr*>(val.type()), NULL /* pre_val */, access.type());
658 } else {
659 assert(access.is_opt_access(), "only for optimization passes");
660 assert(((decorators & C2_TIGHLY_COUPLED_ALLOC) != 0 || !ShenandoahSATBBarrier) && (decorators & C2_ARRAY_COPY) != 0, "unexpected caller of this code");
661 C2OptAccess& opt_access = static_cast<C2OptAccess&>(access);
662 PhaseGVN& gvn = opt_access.gvn();
663 MergeMemNode* mm = opt_access.mem();
664
665 if (ShenandoahStoreValReadBarrier) {
666 RegionNode* region = new RegionNode(3);
667 const Type* v_t = gvn.type(val.node());
668 Node* phi = new PhiNode(region, v_t->isa_oopptr() ? v_t->is_oopptr()->cast_to_nonconst() : v_t);
669 Node* cmp = gvn.transform(new CmpPNode(val.node(), gvn.zerocon(T_OBJECT)));
670 Node* bol = gvn.transform(new BoolNode(cmp, BoolTest::ne));
671 IfNode* iff = new IfNode(opt_access.ctl(), bol, PROB_LIKELY_MAG(3), COUNT_UNKNOWN);
672
673 gvn.transform(iff);
674 if (gvn.is_IterGVN()) {
675 gvn.is_IterGVN()->_worklist.push(iff);
676 } else {
677 gvn.record_for_igvn(iff);
678 }
679
680 Node* null_true = gvn.transform(new IfFalseNode(iff));
681 Node* null_false = gvn.transform(new IfTrueNode(iff));
682 region->init_req(1, null_true);
683 region->init_req(2, null_false);
684 phi->init_req(1, gvn.zerocon(T_OBJECT));
685 Node* cast = new CastPPNode(val.node(), gvn.type(val.node())->join_speculative(TypePtr::NOTNULL));
686 cast->set_req(0, null_false);
687 cast = gvn.transform(cast);
688 Node* rb = gvn.transform(new ShenandoahReadBarrierNode(null_false, gvn.C->immutable_memory(), cast, false));
689 phi->init_req(2, rb);
690 opt_access.set_ctl(gvn.transform(region));
691 val.set_node(gvn.transform(phi));
692 }
693 if (ShenandoahStoreValEnqueueBarrier) {
694 const TypePtr* adr_type = ShenandoahBarrierNode::brooks_pointer_type(gvn.type(val.node()));
695 int alias = gvn.C->get_alias_index(adr_type);
696 Node* wb = new ShenandoahWriteBarrierNode(gvn.C, opt_access.ctl(), mm->memory_at(alias), val.node());
697 Node* wb_transformed = gvn.transform(wb);
698 Node* enqueue = gvn.transform(new ShenandoahEnqueueBarrierNode(wb_transformed));
699 if (wb_transformed == wb) {
700 Node* proj = gvn.transform(new ShenandoahWBMemProjNode(wb));
701 mm->set_memory_at(alias, proj);
702 }
703 val.set_node(enqueue);
704 }
705 }
706 return BarrierSetC2::store_at_resolved(access, val);
707 }
708
709 Node* ShenandoahBarrierSetC2::load_at_resolved(C2Access& access, const Type* val_type) const {
710 DecoratorSet decorators = access.decorators();
711
712 Node* adr = access.addr().node();
713 Node* obj = access.base();
714
715 bool mismatched = (decorators & C2_MISMATCHED) != 0;
716 bool unknown = (decorators & ON_UNKNOWN_OOP_REF) != 0;
717 bool on_heap = (decorators & IN_HEAP) != 0;
718 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
719 bool is_unordered = (decorators & MO_UNORDERED) != 0;
720 bool need_cpu_mem_bar = !is_unordered || mismatched || !on_heap;
721
722 Node* top = Compile::current()->top();
723
724 Node* offset = adr->is_AddP() ? adr->in(AddPNode::Offset) : top;
725 Node* load = BarrierSetC2::load_at_resolved(access, val_type);
726
727 // If we are reading the value of the referent field of a Reference
728 // object (either by using Unsafe directly or through reflection)
729 // then, if SATB is enabled, we need to record the referent in an
730 // SATB log buffer using the pre-barrier mechanism.
731 // Also we need to add memory barrier to prevent commoning reads
732 // from this field across safepoint since GC can change its value.
733 bool need_read_barrier = ShenandoahKeepAliveBarrier &&
734 (on_heap && (on_weak || (unknown && offset != top && obj != top)));
735
736 if (!access.is_oop() || !need_read_barrier) {
737 return load;
738 }
739
740 assert(access.is_parse_access(), "entry not supported at optimization time");
741 C2ParseAccess& parse_access = static_cast<C2ParseAccess&>(access);
742 GraphKit* kit = parse_access.kit();
743
744 if (on_weak) {
745 // Use the pre-barrier to record the value in the referent field
746 satb_write_barrier_pre(kit, false /* do_load */,
747 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
748 load /* pre_val */, T_OBJECT);
749 // Add memory barrier to prevent commoning reads from this field
750 // across safepoint since GC can change its value.
751 kit->insert_mem_bar(Op_MemBarCPUOrder);
752 } else if (unknown) {
753 // We do not require a mem bar inside pre_barrier if need_mem_bar
754 // is set: the barriers would be emitted by us.
755 insert_pre_barrier(kit, obj, offset, load, !need_cpu_mem_bar);
756 }
757
758 return load;
759 }
760
761 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_val_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
762 Node* new_val, const Type* value_type) const {
763 GraphKit* kit = access.kit();
764 if (access.is_oop()) {
765 new_val = shenandoah_storeval_barrier(kit, new_val);
766 shenandoah_write_barrier_pre(kit, false /* do_load */,
767 NULL, NULL, max_juint, NULL, NULL,
768 expected_val /* pre_val */, T_OBJECT);
769
770 MemNode::MemOrd mo = access.mem_node_mo();
771 Node* mem = access.memory();
772 Node* adr = access.addr().node();
773 const TypePtr* adr_type = access.addr().type();
774 Node* load_store = NULL;
775
776 #ifdef _LP64
777 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
778 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
779 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
780 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangeNNode(kit->control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
781 } else
782 #endif
783 {
784 load_store = kit->gvn().transform(new ShenandoahCompareAndExchangePNode(kit->control(), mem, adr, new_val, expected_val, adr_type, value_type->is_oopptr(), mo));
785 }
786
787 access.set_raw_access(load_store);
788 pin_atomic_op(access);
789
790 #ifdef _LP64
791 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
792 return kit->gvn().transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
793 }
794 #endif
795 return load_store;
796 }
797 return BarrierSetC2::atomic_cmpxchg_val_at_resolved(access, expected_val, new_val, value_type);
798 }
799
800 Node* ShenandoahBarrierSetC2::atomic_cmpxchg_bool_at_resolved(C2AtomicParseAccess& access, Node* expected_val,
801 Node* new_val, const Type* value_type) const {
802 GraphKit* kit = access.kit();
803 if (access.is_oop()) {
804 new_val = shenandoah_storeval_barrier(kit, new_val);
805 shenandoah_write_barrier_pre(kit, false /* do_load */,
806 NULL, NULL, max_juint, NULL, NULL,
807 expected_val /* pre_val */, T_OBJECT);
808 DecoratorSet decorators = access.decorators();
809 MemNode::MemOrd mo = access.mem_node_mo();
810 Node* mem = access.memory();
811 bool is_weak_cas = (decorators & C2_WEAK_CMPXCHG) != 0;
812 Node* load_store = NULL;
813 Node* adr = access.addr().node();
814 #ifdef _LP64
815 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
816 Node *newval_enc = kit->gvn().transform(new EncodePNode(new_val, new_val->bottom_type()->make_narrowoop()));
817 Node *oldval_enc = kit->gvn().transform(new EncodePNode(expected_val, expected_val->bottom_type()->make_narrowoop()));
818 if (is_weak_cas) {
819 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
820 } else {
821 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapNNode(kit->control(), mem, adr, newval_enc, oldval_enc, mo));
822 }
823 } else
824 #endif
825 {
826 if (is_weak_cas) {
827 load_store = kit->gvn().transform(new ShenandoahWeakCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
828 } else {
829 load_store = kit->gvn().transform(new ShenandoahCompareAndSwapPNode(kit->control(), mem, adr, new_val, expected_val, mo));
830 }
831 }
832 access.set_raw_access(load_store);
833 pin_atomic_op(access);
834 return load_store;
835 }
836 return BarrierSetC2::atomic_cmpxchg_bool_at_resolved(access, expected_val, new_val, value_type);
837 }
838
839 Node* ShenandoahBarrierSetC2::atomic_xchg_at_resolved(C2AtomicParseAccess& access, Node* val, const Type* value_type) const {
840 GraphKit* kit = access.kit();
841 if (access.is_oop()) {
842 val = shenandoah_storeval_barrier(kit, val);
843 }
844 Node* result = BarrierSetC2::atomic_xchg_at_resolved(access, val, value_type);
845 if (access.is_oop()) {
846 shenandoah_write_barrier_pre(kit, false /* do_load */,
847 NULL, NULL, max_juint, NULL, NULL,
848 result /* pre_val */, T_OBJECT);
849 }
850 return result;
851 }
852
853 void ShenandoahBarrierSetC2::clone(GraphKit* kit, Node* src, Node* dst, Node* size, bool is_array) const {
854 assert(!src->is_AddP(), "unexpected input");
855 src = shenandoah_read_barrier(kit, src);
856 BarrierSetC2::clone(kit, src, dst, size, is_array);
857 }
858
859 Node* ShenandoahBarrierSetC2::resolve(GraphKit* kit, Node* n, DecoratorSet decorators) const {
860 bool is_write = decorators & ACCESS_WRITE;
861 if (is_write) {
862 return shenandoah_write_barrier(kit, n);
863 } else {
864 return shenandoah_read_barrier(kit, n);
865 }
866 }
867
868 Node* ShenandoahBarrierSetC2::obj_allocate(PhaseMacroExpand* macro, Node* ctrl, Node* mem, Node* toobig_false, Node* size_in_bytes,
869 Node*& i_o, Node*& needgc_ctrl,
870 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem,
871 intx prefetch_lines) const {
872 PhaseIterGVN& igvn = macro->igvn();
873
874 // Allocate several words more for the Shenandoah brooks pointer.
875 size_in_bytes = new AddXNode(size_in_bytes, igvn.MakeConX(ShenandoahBrooksPointer::byte_size()));
876 macro->transform_later(size_in_bytes);
877
878 Node* fast_oop = BarrierSetC2::obj_allocate(macro, ctrl, mem, toobig_false, size_in_bytes,
879 i_o, needgc_ctrl, fast_oop_ctrl, fast_oop_rawmem,
880 prefetch_lines);
881
882 // Bump up object for Shenandoah brooks pointer.
883 fast_oop = new AddPNode(macro->top(), fast_oop, igvn.MakeConX(ShenandoahBrooksPointer::byte_size()));
884 macro->transform_later(fast_oop);
885
886 // Initialize Shenandoah brooks pointer to point to the object itself.
887 fast_oop_rawmem = macro->make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, ShenandoahBrooksPointer::byte_offset(), fast_oop, T_OBJECT);
888
889 return fast_oop;
890 }
891
892 // Support for GC barriers emitted during parsing
893 bool ShenandoahBarrierSetC2::is_gc_barrier_node(Node* node) const {
894 if (node->Opcode() != Op_CallLeaf && node->Opcode() != Op_CallLeafNoFP) {
895 return false;
896 }
897 CallLeafNode *call = node->as_CallLeaf();
898 if (call->_name == NULL) {
899 return false;
900 }
901
902 return strcmp(call->_name, "shenandoah_clone_barrier") == 0 ||
903 strcmp(call->_name, "shenandoah_cas_obj") == 0 ||
904 strcmp(call->_name, "shenandoah_wb_pre") == 0;
905 }
906
907 Node* ShenandoahBarrierSetC2::step_over_gc_barrier(Node* c) const {
908 return ShenandoahBarrierNode::skip_through_barrier(c);
909 }
910
911 bool ShenandoahBarrierSetC2::expand_barriers(Compile* C, PhaseIterGVN& igvn) const {
912 return !ShenandoahWriteBarrierNode::expand(C, igvn);
913 }
914
915 bool ShenandoahBarrierSetC2::optimize_loops(PhaseIdealLoop* phase, LoopOptsMode mode, VectorSet& visited, Node_Stack& nstack, Node_List& worklist) const {
916 if (mode == LoopOptsShenandoahExpand) {
917 assert(UseShenandoahGC, "only for shenandoah");
918 ShenandoahWriteBarrierNode::pin_and_expand(phase);
919 return true;
920 } else if (mode == LoopOptsShenandoahPostExpand) {
921 assert(UseShenandoahGC, "only for shenandoah");
922 visited.Clear();
923 ShenandoahWriteBarrierNode::optimize_after_expansion(visited, nstack, worklist, phase);
924 return true;
925 }
926 GrowableArray<MemoryGraphFixer*> memory_graph_fixers;
927 ShenandoahWriteBarrierNode::optimize_before_expansion(phase, memory_graph_fixers, false);
928 return false;
929 }
930
931 bool ShenandoahBarrierSetC2::array_copy_requires_gc_barriers(bool tightly_coupled_alloc, BasicType type, bool is_clone, ArrayCopyPhase phase) const {
932 bool is_oop = type == T_OBJECT || type == T_ARRAY;
933 if (!is_oop) {
934 return false;
935 }
936
937 if (tightly_coupled_alloc) {
938 if (phase == Optimization) {
939 return false;
940 }
941 return !is_clone;
942 }
943 if (phase == Optimization) {
944 return !ShenandoahStoreValEnqueueBarrier;
945 }
946 return true;
947 }
948
949 bool ShenandoahBarrierSetC2::clone_needs_postbarrier(ArrayCopyNode *ac, PhaseIterGVN& igvn) {
950 Node* src = ac->in(ArrayCopyNode::Src);
951 const TypeOopPtr* src_type = igvn.type(src)->is_oopptr();
952 if (src_type->isa_instptr() != NULL) {
953 ciInstanceKlass* ik = src_type->klass()->as_instance_klass();
954 if ((src_type->klass_is_exact() || (!ik->is_interface() && !ik->has_subklass())) && !ik->has_injected_fields()) {
955 if (ik->has_object_fields()) {
956 return true;
957 } else {
958 if (!src_type->klass_is_exact()) {
959 igvn.C->dependencies()->assert_leaf_type(ik);
960 }
961 }
962 } else {
963 return true;
964 }
965 } else if (src_type->isa_aryptr()) {
966 BasicType src_elem = src_type->klass()->as_array_klass()->element_type()->basic_type();
967 if (src_elem == T_OBJECT || src_elem == T_ARRAY) {
968 return true;
969 }
970 } else {
971 return true;
972 }
973 return false;
974 }
975
976 void ShenandoahBarrierSetC2::clone_barrier_at_expansion(ArrayCopyNode* ac, Node* call, PhaseIterGVN& igvn) const {
977 assert(ac->is_clonebasic(), "no other kind of arraycopy here");
978
979 if (!clone_needs_postbarrier(ac, igvn)) {
980 BarrierSetC2::clone_barrier_at_expansion(ac, call, igvn);
981 return;
982 }
983
984 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
985 Node* c = new ProjNode(call,TypeFunc::Control);
986 c = igvn.transform(c);
987 Node* m = new ProjNode(call, TypeFunc::Memory);
988 m = igvn.transform(m);
989
990 Node* dest = ac->in(ArrayCopyNode::Dest);
991 assert(dest->is_AddP(), "bad input");
992 Node* barrier_call = new CallLeafNode(ShenandoahBarrierSetC2::shenandoah_clone_barrier_Type(),
993 CAST_FROM_FN_PTR(address, ShenandoahRuntime::shenandoah_clone_barrier),
994 "shenandoah_clone_barrier", raw_adr_type);
995 barrier_call->init_req(TypeFunc::Control, c);
996 barrier_call->init_req(TypeFunc::I_O , igvn.C->top());
997 barrier_call->init_req(TypeFunc::Memory , m);
998 barrier_call->init_req(TypeFunc::ReturnAdr, igvn.C->top());
999 barrier_call->init_req(TypeFunc::FramePtr, igvn.C->top());
1000 barrier_call->init_req(TypeFunc::Parms+0, dest->in(AddPNode::Base));
1001
1002 barrier_call = igvn.transform(barrier_call);
1003 c = new ProjNode(barrier_call,TypeFunc::Control);
1004 c = igvn.transform(c);
1005 m = new ProjNode(barrier_call, TypeFunc::Memory);
1006 m = igvn.transform(m);
1007
1008 Node* out_c = ac->proj_out(TypeFunc::Control);
1009 Node* out_m = ac->proj_out(TypeFunc::Memory);
1010 igvn.replace_node(out_c, c);
1011 igvn.replace_node(out_m, m);
1012 }
1013
1014
1015 // Support for macro expanded GC barriers
1016 void ShenandoahBarrierSetC2::register_potential_barrier_node(Node* node) const {
1017 if (node->Opcode() == Op_ShenandoahWriteBarrier) {
1018 state()->add_shenandoah_barrier((ShenandoahWriteBarrierNode*) node);
1019 }
1020 }
1021
1022 void ShenandoahBarrierSetC2::unregister_potential_barrier_node(Node* node) const {
1023 if (node->Opcode() == Op_ShenandoahWriteBarrier) {
1024 state()->remove_shenandoah_barrier((ShenandoahWriteBarrierNode*) node);
1025 }
1026 }
1027
1028 void ShenandoahBarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* n) const {
1029 if (is_shenandoah_wb_pre_call(n)) {
1030 shenandoah_eliminate_wb_pre(n, ¯o->igvn());
1031 }
1032 }
1033
1034 void ShenandoahBarrierSetC2::shenandoah_eliminate_wb_pre(Node* call, PhaseIterGVN* igvn) const {
1035 assert(UseShenandoahGC && is_shenandoah_wb_pre_call(call), "");
1036 Node* c = call->as_Call()->proj_out(TypeFunc::Control);
1037 c = c->unique_ctrl_out();
1038 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
1039 c = c->unique_ctrl_out();
1040 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
1041 Node* iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
1042 assert(iff->is_If(), "expect test");
1043 if (!is_shenandoah_marking_if(igvn, iff)) {
1044 c = c->unique_ctrl_out();
1045 assert(c->is_Region() && c->req() == 3, "where's the pre barrier control flow?");
1046 iff = c->in(1)->is_IfProj() ? c->in(1)->in(0) : c->in(2)->in(0);
1047 assert(is_shenandoah_marking_if(igvn, iff), "expect marking test");
1048 }
1049 Node* cmpx = iff->in(1)->in(1);
1050 igvn->replace_node(cmpx, igvn->makecon(TypeInt::CC_EQ));
1051 igvn->rehash_node_delayed(call);
1052 call->del_req(call->req()-1);
1053 }
1054
1055 void ShenandoahBarrierSetC2::enqueue_useful_gc_barrier(PhaseIterGVN* igvn, Node* node) const {
1056 if (node->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(node)) {
1057 igvn->add_users_to_worklist(node);
1058 }
1059 }
1060
1061 void ShenandoahBarrierSetC2::eliminate_useless_gc_barriers(Unique_Node_List &useful, Compile* C) const {
1062 for (uint i = 0; i < useful.size(); i++) {
1063 Node* n = useful.at(i);
1064 if (n->Opcode() == Op_AddP && ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(n)) {
1065 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1066 C->record_for_igvn(n->fast_out(i));
1067 }
1068 }
1069 }
1070 for (int i = state()->shenandoah_barriers_count()-1; i >= 0; i--) {
1071 ShenandoahWriteBarrierNode* n = state()->shenandoah_barrier(i);
1072 if (!useful.member(n)) {
1073 state()->remove_shenandoah_barrier(n);
1074 }
1075 }
1076
1077 }
1078
1079 bool ShenandoahBarrierSetC2::has_special_unique_user(const Node* node) const {
1080 assert(node->outcnt() == 1, "match only for unique out");
1081 Node* n = node->unique_out();
1082 return node->Opcode() == Op_ShenandoahWriteBarrier && n->Opcode() == Op_ShenandoahWBMemProj;
1083 }
1084
1085 void ShenandoahBarrierSetC2::add_users_to_worklist(Unique_Node_List* worklist) const {}
1086
1087 void* ShenandoahBarrierSetC2::create_barrier_state(Arena* comp_arena) const {
1088 return new(comp_arena) ShenandoahBarrierSetC2State(comp_arena);
1089 }
1090
1091 ShenandoahBarrierSetC2State* ShenandoahBarrierSetC2::state() const {
1092 return reinterpret_cast<ShenandoahBarrierSetC2State*>(Compile::current()->barrier_set_state());
1093 }
1094
1095 // If the BarrierSetC2 state has kept macro nodes in its compilation unit state to be
1096 // expanded later, then now is the time to do so.
1097 bool ShenandoahBarrierSetC2::expand_macro_nodes(PhaseMacroExpand* macro) const { return false; }
1098
1099 #ifdef ASSERT
1100 void ShenandoahBarrierSetC2::verify_gc_barriers(Compile* compile, CompilePhase phase) const {
1101 if (ShenandoahVerifyOptoBarriers && phase == BarrierSetC2::BeforeExpand) {
1102 ShenandoahBarrierNode::verify(Compile::current()->root());
1103 } else if (phase == BarrierSetC2::BeforeCodeGen) {
1104 // Verify G1 pre-barriers
1105 const int marking_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset());
1106
1107 ResourceArea *area = Thread::current()->resource_area();
1108 Unique_Node_List visited(area);
1109 Node_List worklist(area);
1110 // We're going to walk control flow backwards starting from the Root
1111 worklist.push(compile->root());
1112 while (worklist.size() > 0) {
1113 Node *x = worklist.pop();
1114 if (x == NULL || x == compile->top()) continue;
1115 if (visited.member(x)) {
1116 continue;
1117 } else {
1118 visited.push(x);
1119 }
1120
1121 if (x->is_Region()) {
1122 for (uint i = 1; i < x->req(); i++) {
1123 worklist.push(x->in(i));
1124 }
1125 } else {
1126 worklist.push(x->in(0));
1127 // We are looking for the pattern:
1128 // /->ThreadLocal
1129 // If->Bool->CmpI->LoadB->AddP->ConL(marking_offset)
1130 // \->ConI(0)
1131 // We want to verify that the If and the LoadB have the same control
1132 // See GraphKit::g1_write_barrier_pre()
1133 if (x->is_If()) {
1134 IfNode *iff = x->as_If();
1135 if (iff->in(1)->is_Bool() && iff->in(1)->in(1)->is_Cmp()) {
1136 CmpNode *cmp = iff->in(1)->in(1)->as_Cmp();
1137 if (cmp->Opcode() == Op_CmpI && cmp->in(2)->is_Con() && cmp->in(2)->bottom_type()->is_int()->get_con() == 0
1138 && cmp->in(1)->is_Load()) {
1139 LoadNode *load = cmp->in(1)->as_Load();
1140 if (load->Opcode() == Op_LoadB && load->in(2)->is_AddP() && load->in(2)->in(2)->Opcode() == Op_ThreadLocal
1141 && load->in(2)->in(3)->is_Con()
1142 && load->in(2)->in(3)->bottom_type()->is_intptr_t()->get_con() == marking_offset) {
1143
1144 Node *if_ctrl = iff->in(0);
1145 Node *load_ctrl = load->in(0);
1146
1147 if (if_ctrl != load_ctrl) {
1148 // Skip possible CProj->NeverBranch in infinite loops
1149 if ((if_ctrl->is_Proj() && if_ctrl->Opcode() == Op_CProj)
1150 && (if_ctrl->in(0)->is_MultiBranch() && if_ctrl->in(0)->Opcode() == Op_NeverBranch)) {
1151 if_ctrl = if_ctrl->in(0)->in(0);
1152 }
1153 }
1154 assert(load_ctrl != NULL && if_ctrl == load_ctrl, "controls must match");
1155 }
1156 }
1157 }
1158 }
1159 }
1160 }
1161 }
1162 }
1163 #endif
1164
1165 Node* ShenandoahBarrierSetC2::ideal_node(PhaseGVN* phase, Node* n, bool can_reshape) const {
1166 if (is_shenandoah_wb_pre_call(n)) {
1167 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1168 if (n->req() > cnt) {
1169 Node* addp = n->in(cnt);
1170 if (has_only_shenandoah_wb_pre_uses(addp)) {
1171 n->del_req(cnt);
1172 if (can_reshape) {
1173 phase->is_IterGVN()->_worklist.push(addp);
1174 }
1175 return n;
1176 }
1177 }
1178 }
1179 if (n->Opcode() == Op_CmpP) {
1180 Node* in1 = n->in(1);
1181 Node* in2 = n->in(2);
1182 if (in1->bottom_type() == TypePtr::NULL_PTR) {
1183 in2 = step_over_gc_barrier(in2);
1184 }
1185 if (in2->bottom_type() == TypePtr::NULL_PTR) {
1186 in1 = step_over_gc_barrier(in1);
1187 }
1188 PhaseIterGVN* igvn = phase->is_IterGVN();
1189 if (in1 != n->in(1)) {
1190 if (igvn != NULL) {
1191 n->set_req_X(1, in1, igvn);
1192 } else {
1193 n->set_req(1, in1);
1194 }
1195 assert(in2 == n->in(2), "only one change");
1196 return n;
1197 }
1198 if (in2 != n->in(2)) {
1199 if (igvn != NULL) {
1200 n->set_req_X(2, in2, igvn);
1201 } else {
1202 n->set_req(2, in2);
1203 }
1204 return n;
1205 }
1206 } else if (can_reshape &&
1207 n->Opcode() == Op_If &&
1208 ShenandoahWriteBarrierNode::is_heap_stable_test(n) &&
1209 n->in(0) != NULL) {
1210 Node* dom = n->in(0);
1211 Node* prev_dom = n;
1212 int op = n->Opcode();
1213 int dist = 16;
1214 // Search up the dominator tree for another heap stable test
1215 while (dom->Opcode() != op || // Not same opcode?
1216 !ShenandoahWriteBarrierNode::is_heap_stable_test(dom) || // Not same input 1?
1217 prev_dom->in(0) != dom) { // One path of test does not dominate?
1218 if (dist < 0) return NULL;
1219
1220 dist--;
1221 prev_dom = dom;
1222 dom = IfNode::up_one_dom(dom);
1223 if (!dom) return NULL;
1224 }
1225
1226 // Check that we did not follow a loop back to ourselves
1227 if (n == dom) {
1228 return NULL;
1229 }
1230
1231 return n->as_If()->dominated_by(prev_dom, phase->is_IterGVN());
1232 }
1233
1234 return NULL;
1235 }
1236
1237 Node* ShenandoahBarrierSetC2::identity_node(PhaseGVN* phase, Node* n) const {
1238 if (n->is_Load()) {
1239 Node *mem = n->in(MemNode::Memory);
1240 Node *value = n->as_Load()->can_see_stored_value(mem, phase);
1241 if (value) {
1242 PhaseIterGVN *igvn = phase->is_IterGVN();
1243 if (igvn != NULL &&
1244 value->is_Phi() &&
1245 value->req() > 2 &&
1246 value->in(1) != NULL &&
1247 value->in(1)->is_ShenandoahBarrier()) {
1248 if (igvn->_worklist.member(value) ||
1249 igvn->_worklist.member(value->in(0)) ||
1250 (value->in(0)->in(1) != NULL &&
1251 value->in(0)->in(1)->is_IfProj() &&
1252 (igvn->_worklist.member(value->in(0)->in(1)) ||
1253 (value->in(0)->in(1)->in(0) != NULL &&
1254 igvn->_worklist.member(value->in(0)->in(1)->in(0)))))) {
1255 igvn->_worklist.push(n);
1256 return n;
1257 }
1258 }
1259 // (This works even when value is a Con, but LoadNode::Value
1260 // usually runs first, producing the singleton type of the Con.)
1261 Node *value_no_barrier = step_over_gc_barrier(value->Opcode() == Op_EncodeP ? value->in(1) : value);
1262 if (value->Opcode() == Op_EncodeP) {
1263 if (value_no_barrier != value->in(1)) {
1264 Node *encode = value->clone();
1265 encode->set_req(1, value_no_barrier);
1266 encode = phase->transform(encode);
1267 return encode;
1268 }
1269 } else {
1270 return value_no_barrier;
1271 }
1272 }
1273 }
1274 return n;
1275 }
1276
1277 bool ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(Node* n) {
1278 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) {
1279 Node* u = n->fast_out(i);
1280 if (!is_shenandoah_wb_pre_call(u)) {
1281 return false;
1282 }
1283 }
1284 return n->outcnt() > 0;
1285 }
1286
1287 bool ShenandoahBarrierSetC2::flatten_gc_alias_type(const TypePtr*& adr_type) const {
1288 int offset = adr_type->offset();
1289 if (offset == ShenandoahBrooksPointer::byte_offset()) {
1290 if (adr_type->isa_aryptr()) {
1291 adr_type = TypeAryPtr::make(adr_type->ptr(), adr_type->isa_aryptr()->ary(), adr_type->isa_aryptr()->klass(), false, offset);
1292 } else if (adr_type->isa_instptr()) {
1293 adr_type = TypeInstPtr::make(adr_type->ptr(), ciEnv::current()->Object_klass(), false, NULL, offset);
1294 }
1295 return true;
1296 } else {
1297 return false;
1298 }
1299 }
1300
1301 bool ShenandoahBarrierSetC2::final_graph_reshaping(Compile* compile, Node* n, uint opcode) const {
1302 switch (opcode) {
1303 case Op_CallLeaf:
1304 case Op_CallLeafNoFP: {
1305 assert (n->is_Call(), "");
1306 CallNode *call = n->as_Call();
1307 if (ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(call)) {
1308 uint cnt = ShenandoahBarrierSetC2::write_ref_field_pre_entry_Type()->domain()->cnt();
1309 if (call->req() > cnt) {
1310 assert(call->req() == cnt + 1, "only one extra input");
1311 Node *addp = call->in(cnt);
1312 assert(!ShenandoahBarrierSetC2::has_only_shenandoah_wb_pre_uses(addp), "useless address computation?");
1313 call->del_req(cnt);
1314 }
1315 }
1316 return false;
1317 }
1318 case Op_ShenandoahCompareAndSwapP:
1319 case Op_ShenandoahCompareAndSwapN:
1320 case Op_ShenandoahWeakCompareAndSwapN:
1321 case Op_ShenandoahWeakCompareAndSwapP:
1322 case Op_ShenandoahCompareAndExchangeP:
1323 case Op_ShenandoahCompareAndExchangeN:
1324 #ifdef ASSERT
1325 if( VerifyOptoOopOffsets ) {
1326 MemNode* mem = n->as_Mem();
1327 // Check to see if address types have grounded out somehow.
1328 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
1329 ciInstanceKlass *k = tp->klass()->as_instance_klass();
1330 bool oop_offset_is_sane = k->contains_field_offset(tp->offset());
1331 assert( !tp || oop_offset_is_sane, "" );
1332 }
1333 #endif
1334 return true;
1335 case Op_ShenandoahReadBarrier:
1336 return true;
1337 case Op_ShenandoahWriteBarrier:
1338 assert(false, "should have been expanded already");
1339 return true;
1340 default:
1341 return false;
1342 }
1343 }
1344
1345 #ifdef ASSERT
1346 bool ShenandoahBarrierSetC2::verify_gc_alias_type(const TypePtr* adr_type, int offset) const {
1347 if (offset == ShenandoahBrooksPointer::byte_offset() &&
1348 (adr_type->base() == Type::AryPtr || adr_type->base() == Type::OopPtr)) {
1349 return true;
1350 } else {
1351 return false;
1352 }
1353 }
1354 #endif
1355
1356 bool ShenandoahBarrierSetC2::escape_add_to_con_graph(ConnectionGraph* conn_graph, PhaseGVN* gvn, Unique_Node_List* delayed_worklist, Node* n, uint opcode) const {
1357 switch (opcode) {
1358 case Op_ShenandoahCompareAndExchangeP:
1359 case Op_ShenandoahCompareAndExchangeN:
1360 conn_graph->add_objload_to_connection_graph(n, delayed_worklist);
1361 // fallthrough
1362 case Op_ShenandoahWeakCompareAndSwapP:
1363 case Op_ShenandoahWeakCompareAndSwapN:
1364 case Op_ShenandoahCompareAndSwapP:
1365 case Op_ShenandoahCompareAndSwapN:
1366 conn_graph->add_to_congraph_unsafe_access(n, opcode, delayed_worklist);
1367 return true;
1368 case Op_StoreP: {
1369 Node* adr = n->in(MemNode::Address);
1370 const Type* adr_type = gvn->type(adr);
1371 // Pointer stores in G1 barriers looks like unsafe access.
1372 // Ignore such stores to be able scalar replace non-escaping
1373 // allocations.
1374 if (adr_type->isa_rawptr() && adr->is_AddP()) {
1375 Node* base = conn_graph->get_addp_base(adr);
1376 if (base->Opcode() == Op_LoadP &&
1377 base->in(MemNode::Address)->is_AddP()) {
1378 adr = base->in(MemNode::Address);
1379 Node* tls = conn_graph->get_addp_base(adr);
1380 if (tls->Opcode() == Op_ThreadLocal) {
1381 int offs = (int) gvn->find_intptr_t_con(adr->in(AddPNode::Offset), Type::OffsetBot);
1382 const int buf_offset = in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset());
1383 if (offs == buf_offset) {
1384 return true; // Pre barrier previous oop value store.
1385 }
1386 }
1387 }
1388 }
1389 return false;
1390 }
1391 case Op_ShenandoahReadBarrier:
1392 case Op_ShenandoahWriteBarrier:
1393 // Barriers 'pass through' its arguments. I.e. what goes in, comes out.
1394 // It doesn't escape.
1395 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahBarrierNode::ValueIn), delayed_worklist);
1396 break;
1397 case Op_ShenandoahEnqueueBarrier:
1398 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), delayed_worklist);
1399 break;
1400 default:
1401 // Nothing
1402 break;
1403 }
1404 return false;
1405 }
1406
1407 bool ShenandoahBarrierSetC2::escape_add_final_edges(ConnectionGraph* conn_graph, PhaseGVN* gvn, Node* n, uint opcode) const {
1408 switch (opcode) {
1409 case Op_ShenandoahCompareAndExchangeP:
1410 case Op_ShenandoahCompareAndExchangeN: {
1411 Node *adr = n->in(MemNode::Address);
1412 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, adr, NULL);
1413 // fallthrough
1414 }
1415 case Op_ShenandoahCompareAndSwapP:
1416 case Op_ShenandoahCompareAndSwapN:
1417 case Op_ShenandoahWeakCompareAndSwapP:
1418 case Op_ShenandoahWeakCompareAndSwapN:
1419 return conn_graph->add_final_edges_unsafe_access(n, opcode);
1420 case Op_ShenandoahReadBarrier:
1421 case Op_ShenandoahWriteBarrier:
1422 // Barriers 'pass through' its arguments. I.e. what goes in, comes out.
1423 // It doesn't escape.
1424 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(ShenandoahBarrierNode::ValueIn), NULL);
1425 return true;
1426 case Op_ShenandoahEnqueueBarrier:
1427 conn_graph->add_local_var_and_edge(n, PointsToNode::NoEscape, n->in(1), NULL);
1428 return true;
1429 default:
1430 // Nothing
1431 break;
1432 }
1433 return false;
1434 }
1435
1436 bool ShenandoahBarrierSetC2::escape_has_out_with_unsafe_object(Node* n) const {
1437 return n->has_out_with(Op_ShenandoahCompareAndExchangeP) || n->has_out_with(Op_ShenandoahCompareAndExchangeN) ||
1438 n->has_out_with(Op_ShenandoahCompareAndSwapP, Op_ShenandoahCompareAndSwapN, Op_ShenandoahWeakCompareAndSwapP, Op_ShenandoahWeakCompareAndSwapN);
1439
1440 }
1441
1442 bool ShenandoahBarrierSetC2::escape_is_barrier_node(Node* n) const {
1443 return n->is_ShenandoahBarrier();
1444 }
1445
1446 bool ShenandoahBarrierSetC2::matcher_find_shared_visit(Matcher* matcher, Matcher::MStack& mstack, Node* n, uint opcode, bool& mem_op, int& mem_addr_idx) const {
1447 switch (opcode) {
1448 case Op_ShenandoahReadBarrier:
1449 if (n->in(ShenandoahBarrierNode::ValueIn)->is_DecodeNarrowPtr()) {
1450 matcher->set_shared(n->in(ShenandoahBarrierNode::ValueIn)->in(1));
1451 }
1452 matcher->set_shared(n);
1453 return true;
1454 default:
1455 break;
1456 }
1457 return false;
1458 }
1459
1460 bool ShenandoahBarrierSetC2::matcher_find_shared_post_visit(Matcher* matcher, Node* n, uint opcode) const {
1461 switch (opcode) {
1462 case Op_ShenandoahCompareAndExchangeP:
1463 case Op_ShenandoahCompareAndExchangeN:
1464 case Op_ShenandoahWeakCompareAndSwapP:
1465 case Op_ShenandoahWeakCompareAndSwapN:
1466 case Op_ShenandoahCompareAndSwapP:
1467 case Op_ShenandoahCompareAndSwapN: { // Convert trinary to binary-tree
1468 Node* newval = n->in(MemNode::ValueIn);
1469 Node* oldval = n->in(LoadStoreConditionalNode::ExpectedIn);
1470 Node* pair = new BinaryNode(oldval, newval);
1471 n->set_req(MemNode::ValueIn,pair);
1472 n->del_req(LoadStoreConditionalNode::ExpectedIn);
1473 return true;
1474 }
1475 default:
1476 break;
1477 }
1478 return false;
1479 }
1480
1481 bool ShenandoahBarrierSetC2::matcher_is_store_load_barrier(Node* x, uint xop) const {
1482 return xop == Op_ShenandoahCompareAndExchangeP ||
1483 xop == Op_ShenandoahCompareAndExchangeN ||
1484 xop == Op_ShenandoahWeakCompareAndSwapP ||
1485 xop == Op_ShenandoahWeakCompareAndSwapN ||
1486 xop == Op_ShenandoahCompareAndSwapN ||
1487 xop == Op_ShenandoahCompareAndSwapP;
1488 }
1489
1490 void ShenandoahBarrierSetC2::igvn_add_users_to_worklist(PhaseIterGVN* igvn, Node* use) const {
1491 if (use->is_ShenandoahBarrier()) {
1492 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1493 Node* u = use->fast_out(i2);
1494 Node* cmp = use->find_out_with(Op_CmpP);
1495 if (u->Opcode() == Op_CmpP) {
1496 igvn->_worklist.push(cmp);
1497 }
1498 }
1499 }
1500 }
1501
1502 void ShenandoahBarrierSetC2::ccp_analyze(PhaseCCP* ccp, Unique_Node_List& worklist, Node* use) const {
1503 if (use->is_ShenandoahBarrier()) {
1504 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) {
1505 Node* p = use->fast_out(i2);
1506 if (p->Opcode() == Op_AddP) {
1507 for (DUIterator_Fast i3max, i3 = p->fast_outs(i3max); i3 < i3max; i3++) {
1508 Node* q = p->fast_out(i3);
1509 if (q->is_Load()) {
1510 if(q->bottom_type() != ccp->type(q)) {
1511 worklist.push(q);
1512 }
1513 }
1514 }
1515 }
1516 }
1517 }
1518 }
1519
1520 Node* ShenandoahBarrierSetC2::split_if_pre(PhaseIdealLoop* phase, Node* n) const {
1521 if (n->Opcode() == Op_ShenandoahReadBarrier) {
1522 ((ShenandoahReadBarrierNode*)n)->try_move(phase);
1523 } else if (n->Opcode() == Op_ShenandoahWriteBarrier) {
1524 return ((ShenandoahWriteBarrierNode*)n)->try_split_thru_phi(phase);
1525 }
1526
1527 return NULL;
1528 }
1529
1530 bool ShenandoahBarrierSetC2::build_loop_late_post(PhaseIdealLoop* phase, Node* n) const {
1531 return ShenandoahBarrierNode::build_loop_late_post(phase, n);
1532 }
1533
1534 bool ShenandoahBarrierSetC2::sink_node(PhaseIdealLoop* phase, Node* n, Node* x, Node* x_ctrl, Node* n_ctrl) const {
1535 if (n->is_ShenandoahBarrier()) {
1536 return x->as_ShenandoahBarrier()->sink_node(phase, x_ctrl, n_ctrl);
1537 }
1538 if (n->is_MergeMem()) {
1539 // PhaseIdealLoop::split_if_with_blocks_post() would:
1540 // _igvn._worklist.yank(x);
1541 // which sometimes causes chains of MergeMem which some of
1542 // shenandoah specific code doesn't support
1543 phase->register_new_node(x, x_ctrl);
1544 return true;
1545 }
1546 return false;
1547 }