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