6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "compiler/compileLog.hpp"
27 #include "gc/shared/collectedHeap.inline.hpp"
28 #include "gc/shared/tlab_globals.hpp"
29 #include "libadt/vectset.hpp"
30 #include "memory/universe.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/arraycopynode.hpp"
33 #include "opto/callnode.hpp"
34 #include "opto/castnode.hpp"
35 #include "opto/cfgnode.hpp"
36 #include "opto/compile.hpp"
37 #include "opto/convertnode.hpp"
38 #include "opto/graphKit.hpp"
39 #include "opto/intrinsicnode.hpp"
40 #include "opto/locknode.hpp"
41 #include "opto/loopnode.hpp"
42 #include "opto/macro.hpp"
43 #include "opto/memnode.hpp"
44 #include "opto/narrowptrnode.hpp"
45 #include "opto/node.hpp"
46 #include "opto/opaquenode.hpp"
47 #include "opto/phaseX.hpp"
48 #include "opto/rootnode.hpp"
49 #include "opto/runtime.hpp"
50 #include "opto/subnode.hpp"
51 #include "opto/subtypenode.hpp"
52 #include "opto/type.hpp"
53 #include "prims/jvmtiExport.hpp"
54 #include "runtime/continuation.hpp"
55 #include "runtime/sharedRuntime.hpp"
56 #include "utilities/macros.hpp"
57 #include "utilities/powerOfTwo.hpp"
58 #if INCLUDE_G1GC
59 #include "gc/g1/g1ThreadLocalData.hpp"
60 #endif // INCLUDE_G1GC
61
62
63 //
64 // Replace any references to "oldref" in inputs to "use" with "newref".
65 // Returns the number of replacements made.
66 //
67 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
68 int nreplacements = 0;
69 uint req = use->req();
70 for (uint j = 0; j < use->len(); j++) {
71 Node *uin = use->in(j);
72 if (uin == oldref) {
73 if (j < req)
74 use->set_req(j, newref);
75 else
76 use->set_prec(j, newref);
77 nreplacements++;
78 } else if (j >= req && uin == nullptr) {
79 break;
80 }
81 }
82 return nreplacements;
83 }
84
85 void PhaseMacroExpand::migrate_outs(Node *old, Node *target) {
86 assert(old != nullptr, "sanity");
87 for (DUIterator_Fast imax, i = old->fast_outs(imax); i < imax; i++) {
88 Node* use = old->fast_out(i);
89 _igvn.rehash_node_delayed(use);
90 imax -= replace_input(use, old, target);
91 // back up iterator
92 --i;
93 }
94 assert(old->outcnt() == 0, "all uses must be deleted");
95 }
96
97 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
98 Node* cmp;
99 if (mask != 0) {
100 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
101 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
102 } else {
103 cmp = word;
104 }
105 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
106 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
107 transform_later(iff);
108
109 // Fast path taken.
110 Node *fast_taken = transform_later(new IfFalseNode(iff));
111
112 // Fast path not-taken, i.e. slow path
113 Node *slow_taken = transform_later(new IfTrueNode(iff));
114
115 if (return_fast_path) {
116 region->init_req(edge, slow_taken); // Capture slow-control
139 // Slow-path call
140 CallNode *call = leaf_name
141 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
142 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
143
144 // Slow path call has no side-effects, uses few values
145 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
146 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0);
147 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1);
148 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2);
149 call->copy_call_debug_info(&_igvn, oldcall);
150 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
151 _igvn.replace_node(oldcall, call);
152 transform_later(call);
153
154 return call;
155 }
156
157 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
158 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
159 bs->eliminate_gc_barrier(this, p2x);
160 #ifndef PRODUCT
161 if (PrintOptoStatistics) {
162 Atomic::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
163 }
164 #endif
165 }
166
167 // Search for a memory operation for the specified memory slice.
168 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
169 Node *orig_mem = mem;
170 Node *alloc_mem = alloc->in(TypeFunc::Memory);
171 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
172 while (true) {
173 if (mem == alloc_mem || mem == start_mem ) {
174 return mem; // hit one of our sentinels
175 } else if (mem->is_MergeMem()) {
176 mem = mem->as_MergeMem()->memory_at(alias_idx);
177 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
178 Node *in = mem->in(0);
179 // we can safely skip over safepoints, calls, locks and membars because we
193 ArrayCopyNode* ac = nullptr;
194 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
195 if (ac != nullptr) {
196 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
197 return ac;
198 }
199 }
200 mem = in->in(TypeFunc::Memory);
201 } else {
202 #ifdef ASSERT
203 in->dump();
204 mem->dump();
205 assert(false, "unexpected projection");
206 #endif
207 }
208 } else if (mem->is_Store()) {
209 const TypePtr* atype = mem->as_Store()->adr_type();
210 int adr_idx = phase->C->get_alias_index(atype);
211 if (adr_idx == alias_idx) {
212 assert(atype->isa_oopptr(), "address type must be oopptr");
213 int adr_offset = atype->offset();
214 uint adr_iid = atype->is_oopptr()->instance_id();
215 // Array elements references have the same alias_idx
216 // but different offset and different instance_id.
217 if (adr_offset == offset && adr_iid == alloc->_idx) {
218 return mem;
219 }
220 } else {
221 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
222 }
223 mem = mem->in(MemNode::Memory);
224 } else if (mem->is_ClearArray()) {
225 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
226 // Can not bypass initialization of the instance
227 // we are looking.
228 debug_only(intptr_t offset;)
229 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
230 InitializeNode* init = alloc->as_Allocate()->initialization();
231 // We are looking for stored value, return Initialize node
232 // or memory edge from Allocate node.
233 if (init != nullptr) {
238 }
239 // Otherwise skip it (the call updated 'mem' value).
240 } else if (mem->Opcode() == Op_SCMemProj) {
241 mem = mem->in(0);
242 Node* adr = nullptr;
243 if (mem->is_LoadStore()) {
244 adr = mem->in(MemNode::Address);
245 } else {
246 assert(mem->Opcode() == Op_EncodeISOArray ||
247 mem->Opcode() == Op_StrCompressedCopy, "sanity");
248 adr = mem->in(3); // Destination array
249 }
250 const TypePtr* atype = adr->bottom_type()->is_ptr();
251 int adr_idx = phase->C->get_alias_index(atype);
252 if (adr_idx == alias_idx) {
253 DEBUG_ONLY(mem->dump();)
254 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
255 return nullptr;
256 }
257 mem = mem->in(MemNode::Memory);
258 } else if (mem->Opcode() == Op_StrInflatedCopy) {
259 Node* adr = mem->in(3); // Destination array
260 const TypePtr* atype = adr->bottom_type()->is_ptr();
261 int adr_idx = phase->C->get_alias_index(atype);
262 if (adr_idx == alias_idx) {
263 DEBUG_ONLY(mem->dump();)
264 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
265 return nullptr;
266 }
267 mem = mem->in(MemNode::Memory);
268 } else {
269 return mem;
270 }
271 assert(mem != orig_mem, "dead memory loop");
272 }
273 }
274
275 // Generate loads from source of the arraycopy for fields of
276 // destination needed at a deoptimization point
277 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
278 BasicType bt = ft;
283 }
284 Node* res = nullptr;
285 if (ac->is_clonebasic()) {
286 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
287 Node* base = ac->in(ArrayCopyNode::Src);
288 Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset)));
289 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
290 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
291 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
292 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
293 } else {
294 if (ac->modifies(offset, offset, &_igvn, true)) {
295 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
296 uint shift = exact_log2(type2aelembytes(bt));
297 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
298 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
299 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
300 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
301
302 Node* adr = nullptr;
303 const TypePtr* adr_type = nullptr;
304 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
305 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
306 Node* base = ac->in(ArrayCopyNode::Src);
307 adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off)));
308 adr_type = _igvn.type(base)->is_ptr()->add_offset(off);
309 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
310 // Don't emit a new load from src if src == dst but try to get the value from memory instead
311 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type->isa_oopptr(), alloc);
312 }
313 } else {
314 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
315 #ifdef _LP64
316 diff = _igvn.transform(new ConvI2LNode(diff));
317 #endif
318 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
319
320 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
321 Node* base = ac->in(ArrayCopyNode::Src);
322 adr = _igvn.transform(new AddPNode(base, base, off));
323 adr_type = _igvn.type(base)->is_ptr()->add_offset(Type::OffsetBot);
324 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
325 // Non constant offset in the array: we can't statically
326 // determine the value
327 return nullptr;
328 }
329 }
330 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
331 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
332 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
333 }
334 }
335 if (res != nullptr) {
336 if (ftype->isa_narrowoop()) {
337 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
338 res = _igvn.transform(new EncodePNode(res, ftype));
339 }
340 return res;
341 }
342 return nullptr;
343 }
344
345 //
346 // Given a Memory Phi, compute a value Phi containing the values from stores
347 // on the input paths.
348 // Note: this function is recursive, its depth is limited by the "level" argument
349 // Returns the computed Phi, or null if it cannot compute it.
350 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) {
351 assert(mem->is_Phi(), "sanity");
352 int alias_idx = C->get_alias_index(adr_t);
353 int offset = adr_t->offset();
354 int instance_id = adr_t->instance_id();
355
356 // Check if an appropriate value phi already exists.
357 Node* region = mem->in(0);
358 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
359 Node* phi = region->fast_out(k);
360 if (phi->is_Phi() && phi != mem &&
361 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
362 return phi;
363 }
364 }
365 // Check if an appropriate new value phi already exists.
366 Node* new_phi = value_phis->find(mem->_idx);
367 if (new_phi != nullptr)
368 return new_phi;
369
370 if (level <= 0) {
371 return nullptr; // Give up: phi tree too deep
372 }
373 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
374 Node *alloc_mem = alloc->in(TypeFunc::Memory);
375
376 uint length = mem->req();
377 GrowableArray <Node *> values(length, length, nullptr);
378
379 // create a new Phi for the value
380 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
381 transform_later(phi);
382 value_phis->push(phi, mem->_idx);
383
384 for (uint j = 1; j < length; j++) {
385 Node *in = mem->in(j);
386 if (in == nullptr || in->is_top()) {
387 values.at_put(j, in);
388 } else {
389 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
390 if (val == start_mem || val == alloc_mem) {
391 // hit a sentinel, return appropriate 0 value
392 values.at_put(j, _igvn.zerocon(ft));
393 continue;
394 }
395 if (val->is_Initialize()) {
396 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
397 }
398 if (val == nullptr) {
399 return nullptr; // can't find a value on this path
400 }
401 if (val == mem) {
402 values.at_put(j, mem);
403 } else if (val->is_Store()) {
404 Node* n = val->in(MemNode::ValueIn);
405 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
406 n = bs->step_over_gc_barrier(n);
407 if (is_subword_type(ft)) {
408 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
409 }
410 values.at_put(j, n);
411 } else if(val->is_Proj() && val->in(0) == alloc) {
412 values.at_put(j, _igvn.zerocon(ft));
413 } else if (val->is_Phi()) {
414 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
415 if (val == nullptr) {
416 return nullptr;
417 }
418 values.at_put(j, val);
419 } else if (val->Opcode() == Op_SCMemProj) {
420 assert(val->in(0)->is_LoadStore() ||
421 val->in(0)->Opcode() == Op_EncodeISOArray ||
422 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
423 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
424 return nullptr;
425 } else if (val->is_ArrayCopy()) {
426 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
427 if (res == nullptr) {
428 return nullptr;
429 }
430 values.at_put(j, res);
431 } else {
432 DEBUG_ONLY( val->dump(); )
436 }
437 }
438 // Set Phi's inputs
439 for (uint j = 1; j < length; j++) {
440 if (values.at(j) == mem) {
441 phi->init_req(j, phi);
442 } else {
443 phi->init_req(j, values.at(j));
444 }
445 }
446 return phi;
447 }
448
449 // Search the last value stored into the object's field.
450 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
451 assert(adr_t->is_known_instance_field(), "instance required");
452 int instance_id = adr_t->instance_id();
453 assert((uint)instance_id == alloc->_idx, "wrong allocation");
454
455 int alias_idx = C->get_alias_index(adr_t);
456 int offset = adr_t->offset();
457 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
458 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
459 Node *alloc_mem = alloc->in(TypeFunc::Memory);
460 VectorSet visited;
461
462 bool done = sfpt_mem == alloc_mem;
463 Node *mem = sfpt_mem;
464 while (!done) {
465 if (visited.test_set(mem->_idx)) {
466 return nullptr; // found a loop, give up
467 }
468 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
469 if (mem == start_mem || mem == alloc_mem) {
470 done = true; // hit a sentinel, return appropriate 0 value
471 } else if (mem->is_Initialize()) {
472 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
473 if (mem == nullptr) {
474 done = true; // Something go wrong.
475 } else if (mem->is_Store()) {
476 const TypePtr* atype = mem->as_Store()->adr_type();
477 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
478 done = true;
479 }
480 } else if (mem->is_Store()) {
481 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
482 assert(atype != nullptr, "address type must be oopptr");
483 assert(C->get_alias_index(atype) == alias_idx &&
484 atype->is_known_instance_field() && atype->offset() == offset &&
485 atype->instance_id() == instance_id, "store is correct memory slice");
486 done = true;
487 } else if (mem->is_Phi()) {
488 // try to find a phi's unique input
489 Node *unique_input = nullptr;
490 Node *top = C->top();
491 for (uint i = 1; i < mem->req(); i++) {
492 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
493 if (n == nullptr || n == top || n == mem) {
494 continue;
495 } else if (unique_input == nullptr) {
496 unique_input = n;
497 } else if (unique_input != n) {
498 unique_input = top;
499 break;
500 }
501 }
502 if (unique_input != nullptr && unique_input != top) {
503 mem = unique_input;
504 } else {
505 done = true;
506 }
507 } else if (mem->is_ArrayCopy()) {
508 done = true;
509 } else {
510 DEBUG_ONLY( mem->dump(); )
511 assert(false, "unexpected node");
512 }
513 }
514 if (mem != nullptr) {
515 if (mem == start_mem || mem == alloc_mem) {
516 // hit a sentinel, return appropriate 0 value
517 return _igvn.zerocon(ft);
518 } else if (mem->is_Store()) {
519 Node* n = mem->in(MemNode::ValueIn);
520 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
521 n = bs->step_over_gc_barrier(n);
522 return n;
523 } else if (mem->is_Phi()) {
524 // attempt to produce a Phi reflecting the values on the input paths of the Phi
525 Node_Stack value_phis(8);
526 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
527 if (phi != nullptr) {
528 return phi;
529 } else {
530 // Kill all new Phis
531 while(value_phis.is_nonempty()) {
532 Node* n = value_phis.node();
533 _igvn.replace_node(n, C->top());
534 value_phis.pop();
535 }
536 }
537 } else if (mem->is_ArrayCopy()) {
538 Node* ctl = mem->in(0);
539 Node* m = mem->in(TypeFunc::Memory);
540 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
541 // pin the loads in the uncommon trap path
542 ctl = sfpt_ctl;
543 m = sfpt_mem;
544 }
545 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
546 }
547 }
548 // Something go wrong.
549 return nullptr;
550 }
551
552 // Check the possibility of scalar replacement.
553 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) {
554 // Scan the uses of the allocation to check for anything that would
555 // prevent us from eliminating it.
556 NOT_PRODUCT( const char* fail_eliminate = nullptr; )
557 DEBUG_ONLY( Node* disq_node = nullptr; )
558 bool can_eliminate = true;
559 bool reduce_merge_precheck = (safepoints == nullptr);
560
561 Node* res = alloc->result_cast();
562 const TypeOopPtr* res_type = nullptr;
563 if (res == nullptr) {
564 // All users were eliminated.
565 } else if (!res->is_CheckCastPP()) {
566 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
567 can_eliminate = false;
568 } else {
569 res_type = igvn->type(res)->isa_oopptr();
570 if (res_type == nullptr) {
571 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
572 can_eliminate = false;
573 } else if (res_type->isa_aryptr()) {
574 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
575 if (length < 0) {
576 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
577 can_eliminate = false;
578 }
579 }
580 }
581
582 if (can_eliminate && res != nullptr) {
583 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
584 for (DUIterator_Fast jmax, j = res->fast_outs(jmax);
585 j < jmax && can_eliminate; j++) {
586 Node* use = res->fast_out(j);
587
588 if (use->is_AddP()) {
589 const TypePtr* addp_type = igvn->type(use)->is_ptr();
590 int offset = addp_type->offset();
591
592 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
593 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
594 can_eliminate = false;
595 break;
596 }
597 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
598 k < kmax && can_eliminate; k++) {
599 Node* n = use->fast_out(k);
600 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n)) {
601 DEBUG_ONLY(disq_node = n;)
602 if (n->is_Load() || n->is_LoadStore()) {
603 NOT_PRODUCT(fail_eliminate = "Field load";)
604 } else {
605 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
613 use->as_ArrayCopy()->is_copyof_validated() ||
614 use->as_ArrayCopy()->is_copyofrange_validated()) &&
615 use->in(ArrayCopyNode::Dest) == res) {
616 // ok to eliminate
617 } else if (use->is_SafePoint()) {
618 SafePointNode* sfpt = use->as_SafePoint();
619 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
620 // Object is passed as argument.
621 DEBUG_ONLY(disq_node = use;)
622 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
623 can_eliminate = false;
624 }
625 Node* sfptMem = sfpt->memory();
626 if (sfptMem == nullptr || sfptMem->is_top()) {
627 DEBUG_ONLY(disq_node = use;)
628 NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
629 can_eliminate = false;
630 } else if (!reduce_merge_precheck) {
631 safepoints->append_if_missing(sfpt);
632 }
633 } else if (reduce_merge_precheck && (use->is_Phi() || use->is_EncodeP() || use->Opcode() == Op_MemBarRelease)) {
634 // Nothing to do
635 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
636 if (use->is_Phi()) {
637 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
638 NOT_PRODUCT(fail_eliminate = "Object is return value";)
639 } else {
640 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
641 }
642 DEBUG_ONLY(disq_node = use;)
643 } else {
644 if (use->Opcode() == Op_Return) {
645 NOT_PRODUCT(fail_eliminate = "Object is return value";)
646 } else {
647 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
648 }
649 DEBUG_ONLY(disq_node = use;)
650 }
651 can_eliminate = false;
652 }
653 }
654 }
655
656 #ifndef PRODUCT
657 if (PrintEliminateAllocations && safepoints != nullptr) {
658 if (can_eliminate) {
659 tty->print("Scalar ");
660 if (res == nullptr)
661 alloc->dump();
662 else
663 res->dump();
664 } else if (alloc->_is_scalar_replaceable) {
665 tty->print("NotScalar (%s)", fail_eliminate);
666 if (res == nullptr)
667 alloc->dump();
668 else
669 res->dump();
670 #ifdef ASSERT
671 if (disq_node != nullptr) {
672 tty->print(" >>>> ");
673 disq_node->dump();
674 }
675 #endif /*ASSERT*/
676 }
677 }
678 #endif
679 return can_eliminate;
680 }
681
682 void PhaseMacroExpand::undo_previous_scalarizations(GrowableArray <SafePointNode *> safepoints_done, AllocateNode* alloc) {
683 Node* res = alloc->result_cast();
684 int nfields = 0;
709 JVMState *jvms = sfpt_done->jvms();
710 jvms->set_endoff(sfpt_done->req());
711 // Now make a pass over the debug information replacing any references
712 // to SafePointScalarObjectNode with the allocated object.
713 int start = jvms->debug_start();
714 int end = jvms->debug_end();
715 for (int i = start; i < end; i++) {
716 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
717 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
718 if (scobj->first_index(jvms) == sfpt_done->req() &&
719 scobj->n_fields() == (uint)nfields) {
720 assert(scobj->alloc() == alloc, "sanity");
721 sfpt_done->set_req(i, res);
722 }
723 }
724 }
725 _igvn._worklist.push(sfpt_done);
726 }
727 }
728
729 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode *alloc, SafePointNode* sfpt) {
730 // Fields of scalar objs are referenced only at the end
731 // of regular debuginfo at the last (youngest) JVMS.
732 // Record relative start index.
733 ciInstanceKlass* iklass = nullptr;
734 BasicType basic_elem_type = T_ILLEGAL;
735 const Type* field_type = nullptr;
736 const TypeOopPtr* res_type = nullptr;
737 int nfields = 0;
738 int array_base = 0;
739 int element_size = 0;
740 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
741 Node* res = alloc->result_cast();
742
743 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
744 assert(sfpt->jvms() != nullptr, "missed JVMS");
745
746 if (res != nullptr) { // Could be null when there are no users
747 res_type = _igvn.type(res)->isa_oopptr();
748
749 if (res_type->isa_instptr()) {
750 // find the fields of the class which will be needed for safepoint debug information
751 iklass = res_type->is_instptr()->instance_klass();
752 nfields = iklass->nof_nonstatic_fields();
753 } else {
754 // find the array's elements which will be needed for safepoint debug information
755 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
756 assert(nfields >= 0, "must be an array klass.");
757 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type();
758 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
759 element_size = type2aelembytes(basic_elem_type);
760 field_type = res_type->is_aryptr()->elem();
761 }
762 }
763
764 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, nfields);
765 sobj->init_req(0, C->root());
766 transform_later(sobj);
767
768 // Scan object's fields adding an input to the safepoint for each field.
769 for (int j = 0; j < nfields; j++) {
770 intptr_t offset;
771 ciField* field = nullptr;
772 if (iklass != nullptr) {
773 field = iklass->nonstatic_field_at(j);
774 offset = field->offset_in_bytes();
775 ciType* elem_type = field->type();
776 basic_elem_type = field->layout_type();
777
778 // The next code is taken from Parse::do_get_xxx().
779 if (is_reference_type(basic_elem_type)) {
780 if (!elem_type->is_loaded()) {
781 field_type = TypeInstPtr::BOTTOM;
782 } else if (field != nullptr && field->is_static_constant()) {
783 ciObject* con = field->constant_value().as_object();
784 // Do not "join" in the previous type; it doesn't add value,
785 // and may yield a vacuous result if the field is of interface type.
786 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
787 assert(field_type != nullptr, "field singleton type must be consistent");
788 } else {
789 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
790 }
791 if (UseCompressedOops) {
792 field_type = field_type->make_narrowoop();
793 basic_elem_type = T_NARROWOOP;
794 }
795 } else {
796 field_type = Type::get_const_basic_type(basic_elem_type);
797 }
798 } else {
799 offset = array_base + j * (intptr_t)element_size;
800 }
801
802 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
803
804 Node *field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
805
806 // We weren't able to find a value for this field,
807 // give up on eliminating this allocation.
808 if (field_val == nullptr) {
809 uint last = sfpt->req() - 1;
810 for (int k = 0; k < j; k++) {
811 sfpt->del_req(last--);
812 }
813 _igvn._worklist.push(sfpt);
814
815 #ifndef PRODUCT
816 if (PrintEliminateAllocations) {
817 if (field != nullptr) {
818 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
819 field->print();
820 int field_idx = C->get_alias_index(field_addr_type);
821 tty->print(" (alias_idx=%d)", field_idx);
822 } else { // Array's element
823 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, j);
824 }
825 tty->print(", which prevents elimination of: ");
826 if (res == nullptr)
827 alloc->dump();
828 else
829 res->dump();
830 }
831 #endif
832
833 return nullptr;
834 }
835
836 if (UseCompressedOops && field_type->isa_narrowoop()) {
837 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
838 // to be able scalar replace the allocation.
839 if (field_val->is_EncodeP()) {
840 field_val = field_val->in(1);
841 } else {
842 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
843 }
844 }
845 sfpt->add_req(field_val);
846 }
847
848 sfpt->jvms()->set_endoff(sfpt->req());
849
850 return sobj;
851 }
852
853 // Do scalar replacement.
854 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
855 GrowableArray <SafePointNode *> safepoints_done;
856 Node* res = alloc->result_cast();
857 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
858
859 // Process the safepoint uses
860 while (safepoints.length() > 0) {
861 SafePointNode* sfpt = safepoints.pop();
862 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt);
863
864 if (sobj == nullptr) {
865 undo_previous_scalarizations(safepoints_done, alloc);
866 return false;
867 }
868
869 // Now make a pass over the debug information replacing any references
870 // to the allocated object with "sobj"
871 JVMState *jvms = sfpt->jvms();
872 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
873 _igvn._worklist.push(sfpt);
874
875 // keep it for rollback
876 safepoints_done.append_if_missing(sfpt);
877 }
878
879 return true;
880 }
881
882 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
883 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
884 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
885 if (ctl_proj != nullptr) {
886 igvn.replace_node(ctl_proj, n->in(0));
887 }
888 if (mem_proj != nullptr) {
889 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
890 }
891 }
892
893 // Process users of eliminated allocation.
894 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
895 Node* res = alloc->result_cast();
896 if (res != nullptr) {
897 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
898 Node *use = res->last_out(j);
899 uint oc1 = res->outcnt();
900
901 if (use->is_AddP()) {
902 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
903 Node *n = use->last_out(k);
904 uint oc2 = use->outcnt();
905 if (n->is_Store()) {
906 #ifdef ASSERT
907 // Verify that there is no dependent MemBarVolatile nodes,
908 // they should be removed during IGVN, see MemBarNode::Ideal().
909 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
910 p < pmax; p++) {
911 Node* mb = n->fast_out(p);
912 assert(mb->is_Initialize() || !mb->is_MemBar() ||
913 mb->req() <= MemBarNode::Precedent ||
914 mb->in(MemBarNode::Precedent) != n,
915 "MemBarVolatile should be eliminated for non-escaping object");
916 }
917 #endif
918 _igvn.replace_node(n, n->in(MemNode::Memory));
919 } else {
920 eliminate_gc_barrier(n);
921 }
922 k -= (oc2 - use->outcnt());
923 }
924 _igvn.remove_dead_node(use);
925 } else if (use->is_ArrayCopy()) {
926 // Disconnect ArrayCopy node
927 ArrayCopyNode* ac = use->as_ArrayCopy();
928 if (ac->is_clonebasic()) {
929 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
930 disconnect_projections(ac, _igvn);
931 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
932 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
933 disconnect_projections(membar_before->as_MemBar(), _igvn);
934 if (membar_after->is_MemBar()) {
935 disconnect_projections(membar_after->as_MemBar(), _igvn);
936 }
937 } else {
938 assert(ac->is_arraycopy_validated() ||
939 ac->is_copyof_validated() ||
940 ac->is_copyofrange_validated(), "unsupported");
941 CallProjections callprojs;
942 ac->extract_projections(&callprojs, true);
943
944 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
945 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
946 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
947
948 // Set control to top. IGVN will remove the remaining projections
949 ac->set_req(0, top());
950 ac->replace_edge(res, top(), &_igvn);
951
952 // Disconnect src right away: it can help find new
953 // opportunities for allocation elimination
954 Node* src = ac->in(ArrayCopyNode::Src);
955 ac->replace_edge(src, top(), &_igvn);
956 // src can be top at this point if src and dest of the
957 // arraycopy were the same
958 if (src->outcnt() == 0 && !src->is_top()) {
959 _igvn.remove_dead_node(src);
960 }
961 }
962 _igvn._worklist.push(ac);
963 } else {
964 eliminate_gc_barrier(use);
965 }
966 j -= (oc1 - res->outcnt());
967 }
968 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
969 _igvn.remove_dead_node(res);
970 }
971
972 //
973 // Process other users of allocation's projections
974 //
975 if (_callprojs.resproj != nullptr && _callprojs.resproj->outcnt() != 0) {
976 // First disconnect stores captured by Initialize node.
977 // If Initialize node is eliminated first in the following code,
978 // it will kill such stores and DUIterator_Last will assert.
979 for (DUIterator_Fast jmax, j = _callprojs.resproj->fast_outs(jmax); j < jmax; j++) {
980 Node* use = _callprojs.resproj->fast_out(j);
981 if (use->is_AddP()) {
982 // raw memory addresses used only by the initialization
983 _igvn.replace_node(use, C->top());
984 --j; --jmax;
985 }
986 }
987 for (DUIterator_Last jmin, j = _callprojs.resproj->last_outs(jmin); j >= jmin; ) {
988 Node* use = _callprojs.resproj->last_out(j);
989 uint oc1 = _callprojs.resproj->outcnt();
990 if (use->is_Initialize()) {
991 // Eliminate Initialize node.
992 InitializeNode *init = use->as_Initialize();
993 assert(init->outcnt() <= 2, "only a control and memory projection expected");
994 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
995 if (ctrl_proj != nullptr) {
996 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
997 #ifdef ASSERT
998 // If the InitializeNode has no memory out, it will die, and tmp will become null
999 Node* tmp = init->in(TypeFunc::Control);
1000 assert(tmp == nullptr || tmp == _callprojs.fallthrough_catchproj, "allocation control projection");
1001 #endif
1002 }
1003 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1004 if (mem_proj != nullptr) {
1005 Node *mem = init->in(TypeFunc::Memory);
1006 #ifdef ASSERT
1007 if (mem->is_MergeMem()) {
1008 assert(mem->in(TypeFunc::Memory) == _callprojs.fallthrough_memproj, "allocation memory projection");
1009 } else {
1010 assert(mem == _callprojs.fallthrough_memproj, "allocation memory projection");
1011 }
1012 #endif
1013 _igvn.replace_node(mem_proj, mem);
1014 }
1015 } else {
1016 assert(false, "only Initialize or AddP expected");
1017 }
1018 j -= (oc1 - _callprojs.resproj->outcnt());
1019 }
1020 }
1021 if (_callprojs.fallthrough_catchproj != nullptr) {
1022 _igvn.replace_node(_callprojs.fallthrough_catchproj, alloc->in(TypeFunc::Control));
1023 }
1024 if (_callprojs.fallthrough_memproj != nullptr) {
1025 _igvn.replace_node(_callprojs.fallthrough_memproj, alloc->in(TypeFunc::Memory));
1026 }
1027 if (_callprojs.catchall_memproj != nullptr) {
1028 _igvn.replace_node(_callprojs.catchall_memproj, C->top());
1029 }
1030 if (_callprojs.fallthrough_ioproj != nullptr) {
1031 _igvn.replace_node(_callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1032 }
1033 if (_callprojs.catchall_ioproj != nullptr) {
1034 _igvn.replace_node(_callprojs.catchall_ioproj, C->top());
1035 }
1036 if (_callprojs.catchall_catchproj != nullptr) {
1037 _igvn.replace_node(_callprojs.catchall_catchproj, C->top());
1038 }
1039 }
1040
1041 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1042 // If reallocation fails during deoptimization we'll pop all
1043 // interpreter frames for this compiled frame and that won't play
1044 // nice with JVMTI popframe.
1045 // We avoid this issue by eager reallocation when the popframe request
1046 // is received.
1047 if (!EliminateAllocations || !alloc->_is_non_escaping) {
1048 return false;
1049 }
1050 Node* klass = alloc->in(AllocateNode::KlassNode);
1051 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1052 Node* res = alloc->result_cast();
1053 // Eliminate boxing allocations which are not used
1054 // regardless scalar replaceable status.
1055 bool boxing_alloc = C->eliminate_boxing() &&
1056 tklass->isa_instklassptr() &&
1057 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1058 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != nullptr))) {
1059 return false;
1060 }
1061
1062 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1063
1064 GrowableArray <SafePointNode *> safepoints;
1065 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1066 return false;
1067 }
1068
1069 if (!alloc->_is_scalar_replaceable) {
1070 assert(res == nullptr, "sanity");
1071 // We can only eliminate allocation if all debug info references
1072 // are already replaced with SafePointScalarObject because
1073 // we can't search for a fields value without instance_id.
1074 if (safepoints.length() > 0) {
1075 return false;
1076 }
1077 }
1078
1079 if (!scalar_replacement(alloc, safepoints)) {
1080 return false;
1081 }
1082
1083 CompileLog* log = C->log();
1084 if (log != nullptr) {
1085 log->head("eliminate_allocation type='%d'",
1086 log->identify(tklass->exact_klass()));
1087 JVMState* p = alloc->jvms();
1088 while (p != nullptr) {
1089 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1090 p = p->caller();
1091 }
1092 log->tail("eliminate_allocation");
1093 }
1094
1095 process_users_of_allocation(alloc);
1096
1097 #ifndef PRODUCT
1098 if (PrintEliminateAllocations) {
1099 if (alloc->is_AllocateArray())
1100 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1101 else
1102 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1103 }
1104 #endif
1105
1106 return true;
1107 }
1108
1109 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1110 // EA should remove all uses of non-escaping boxing node.
1111 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1112 return false;
1113 }
1114
1115 assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1116
1117 boxing->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1118
1119 const TypeTuple* r = boxing->tf()->range();
1120 assert(r->cnt() > TypeFunc::Parms, "sanity");
1121 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1122 assert(t != nullptr, "sanity");
1123
1124 CompileLog* log = C->log();
1125 if (log != nullptr) {
1126 log->head("eliminate_boxing type='%d'",
1127 log->identify(t->instance_klass()));
1128 JVMState* p = boxing->jvms();
1129 while (p != nullptr) {
1130 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1131 p = p->caller();
1132 }
1133 log->tail("eliminate_boxing");
1134 }
1135
1136 process_users_of_allocation(boxing);
1137
1138 #ifndef PRODUCT
1139 if (PrintEliminateAllocations) {
1283 }
1284 }
1285 #endif
1286 yank_alloc_node(alloc);
1287 return;
1288 }
1289 }
1290
1291 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1292 Node *slow_region = nullptr;
1293 Node *toobig_false = ctrl;
1294
1295 // generate the initial test if necessary
1296 if (initial_slow_test != nullptr ) {
1297 assert (expand_fast_path, "Only need test if there is a fast path");
1298 slow_region = new RegionNode(3);
1299
1300 // Now make the initial failure test. Usually a too-big test but
1301 // might be a TRUE for finalizers or a fancy class check for
1302 // newInstance0.
1303 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1304 transform_later(toobig_iff);
1305 // Plug the failing-too-big test into the slow-path region
1306 Node *toobig_true = new IfTrueNode( toobig_iff );
1307 transform_later(toobig_true);
1308 slow_region ->init_req( too_big_or_final_path, toobig_true );
1309 toobig_false = new IfFalseNode( toobig_iff );
1310 transform_later(toobig_false);
1311 } else {
1312 // No initial test, just fall into next case
1313 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1314 toobig_false = ctrl;
1315 debug_only(slow_region = NodeSentinel);
1316 }
1317
1318 // If we are here there are several possibilities
1319 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1320 // no_initial_check means a constant allocation.
1321 // - If check always evaluates to false -> expand_fast_path is false (see above)
1322 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1323 // if !allocation_has_use the fast path is empty
1324 // if !allocation_has_use && no_initial_check
1325 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1326 // removed by yank_alloc_node above.
1327
1328 Node *slow_mem = mem; // save the current memory state for slow path
1329 // generate the fast allocation code unless we know that the initial test will always go slow
1330 if (expand_fast_path) {
1331 // Fast path modifies only raw memory.
1332 if (mem->is_MergeMem()) {
1333 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1334 }
1335
1336 // allocate the Region and Phi nodes for the result
1337 result_region = new RegionNode(3);
1338 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1339 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1340
1341 // Grab regular I/O before optional prefetch may change it.
1342 // Slow-path does no I/O so just set it to the original I/O.
1343 result_phi_i_o->init_req(slow_result_path, i_o);
1344
1345 // Name successful fast-path variables
1346 Node* fast_oop_ctrl;
1347 Node* fast_oop_rawmem;
1348 if (allocation_has_use) {
1349 Node* needgc_ctrl = nullptr;
1350 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1351
1352 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1353 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1354 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1355 fast_oop_ctrl, fast_oop_rawmem,
1356 prefetch_lines);
1357
1358 if (initial_slow_test != nullptr) {
1359 // This completes all paths into the slow merge point
1360 slow_region->init_req(need_gc_path, needgc_ctrl);
1361 transform_later(slow_region);
1362 } else {
1363 // No initial slow path needed!
1364 // Just fall from the need-GC path straight into the VM call.
1365 slow_region = needgc_ctrl;
1366 }
1367
1385 result_phi_i_o ->init_req(fast_result_path, i_o);
1386 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1387 } else {
1388 slow_region = ctrl;
1389 result_phi_i_o = i_o; // Rename it to use in the following code.
1390 }
1391
1392 // Generate slow-path call
1393 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1394 OptoRuntime::stub_name(slow_call_address),
1395 TypePtr::BOTTOM);
1396 call->init_req(TypeFunc::Control, slow_region);
1397 call->init_req(TypeFunc::I_O, top()); // does no i/o
1398 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1399 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1400 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1401
1402 call->init_req(TypeFunc::Parms+0, klass_node);
1403 if (length != nullptr) {
1404 call->init_req(TypeFunc::Parms+1, length);
1405 }
1406
1407 // Copy debug information and adjust JVMState information, then replace
1408 // allocate node with the call
1409 call->copy_call_debug_info(&_igvn, alloc);
1410 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1411 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1412 // path dies).
1413 if (valid_length_test != nullptr) {
1414 call->add_req(valid_length_test);
1415 }
1416 if (expand_fast_path) {
1417 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1418 } else {
1419 // Hook i_o projection to avoid its elimination during allocation
1420 // replacement (when only a slow call is generated).
1421 call->set_req(TypeFunc::I_O, result_phi_i_o);
1422 }
1423 _igvn.replace_node(alloc, call);
1424 transform_later(call);
1425
1426 // Identify the output projections from the allocate node and
1427 // adjust any references to them.
1428 // The control and io projections look like:
1429 //
1430 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1431 // Allocate Catch
1432 // ^---Proj(io) <-------+ ^---CatchProj(io)
1433 //
1434 // We are interested in the CatchProj nodes.
1435 //
1436 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1437
1438 // An allocate node has separate memory projections for the uses on
1439 // the control and i_o paths. Replace the control memory projection with
1440 // result_phi_rawmem (unless we are only generating a slow call when
1441 // both memory projections are combined)
1442 if (expand_fast_path && _callprojs.fallthrough_memproj != nullptr) {
1443 migrate_outs(_callprojs.fallthrough_memproj, result_phi_rawmem);
1444 }
1445 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1446 // catchall_memproj so we end up with a call that has only 1 memory projection.
1447 if (_callprojs.catchall_memproj != nullptr ) {
1448 if (_callprojs.fallthrough_memproj == nullptr) {
1449 _callprojs.fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1450 transform_later(_callprojs.fallthrough_memproj);
1451 }
1452 migrate_outs(_callprojs.catchall_memproj, _callprojs.fallthrough_memproj);
1453 _igvn.remove_dead_node(_callprojs.catchall_memproj);
1454 }
1455
1456 // An allocate node has separate i_o projections for the uses on the control
1457 // and i_o paths. Always replace the control i_o projection with result i_o
1458 // otherwise incoming i_o become dead when only a slow call is generated
1459 // (it is different from memory projections where both projections are
1460 // combined in such case).
1461 if (_callprojs.fallthrough_ioproj != nullptr) {
1462 migrate_outs(_callprojs.fallthrough_ioproj, result_phi_i_o);
1463 }
1464 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1465 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1466 if (_callprojs.catchall_ioproj != nullptr ) {
1467 if (_callprojs.fallthrough_ioproj == nullptr) {
1468 _callprojs.fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1469 transform_later(_callprojs.fallthrough_ioproj);
1470 }
1471 migrate_outs(_callprojs.catchall_ioproj, _callprojs.fallthrough_ioproj);
1472 _igvn.remove_dead_node(_callprojs.catchall_ioproj);
1473 }
1474
1475 // if we generated only a slow call, we are done
1476 if (!expand_fast_path) {
1477 // Now we can unhook i_o.
1478 if (result_phi_i_o->outcnt() > 1) {
1479 call->set_req(TypeFunc::I_O, top());
1480 } else {
1481 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1482 // Case of new array with negative size known during compilation.
1483 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1484 // following code since call to runtime will throw exception.
1485 // As result there will be no users of i_o after the call.
1486 // Leave i_o attached to this call to avoid problems in preceding graph.
1487 }
1488 return;
1489 }
1490
1491 if (_callprojs.fallthrough_catchproj != nullptr) {
1492 ctrl = _callprojs.fallthrough_catchproj->clone();
1493 transform_later(ctrl);
1494 _igvn.replace_node(_callprojs.fallthrough_catchproj, result_region);
1495 } else {
1496 ctrl = top();
1497 }
1498 Node *slow_result;
1499 if (_callprojs.resproj == nullptr) {
1500 // no uses of the allocation result
1501 slow_result = top();
1502 } else {
1503 slow_result = _callprojs.resproj->clone();
1504 transform_later(slow_result);
1505 _igvn.replace_node(_callprojs.resproj, result_phi_rawoop);
1506 }
1507
1508 // Plug slow-path into result merge point
1509 result_region->init_req( slow_result_path, ctrl);
1510 transform_later(result_region);
1511 if (allocation_has_use) {
1512 result_phi_rawoop->init_req(slow_result_path, slow_result);
1513 transform_later(result_phi_rawoop);
1514 }
1515 result_phi_rawmem->init_req(slow_result_path, _callprojs.fallthrough_memproj);
1516 transform_later(result_phi_rawmem);
1517 transform_later(result_phi_i_o);
1518 // This completes all paths into the result merge point
1519 }
1520
1521 // Remove alloc node that has no uses.
1522 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1523 Node* ctrl = alloc->in(TypeFunc::Control);
1524 Node* mem = alloc->in(TypeFunc::Memory);
1525 Node* i_o = alloc->in(TypeFunc::I_O);
1526
1527 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1528 if (_callprojs.resproj != nullptr) {
1529 for (DUIterator_Fast imax, i = _callprojs.resproj->fast_outs(imax); i < imax; i++) {
1530 Node* use = _callprojs.resproj->fast_out(i);
1531 use->isa_MemBar()->remove(&_igvn);
1532 --imax;
1533 --i; // back up iterator
1534 }
1535 assert(_callprojs.resproj->outcnt() == 0, "all uses must be deleted");
1536 _igvn.remove_dead_node(_callprojs.resproj);
1537 }
1538 if (_callprojs.fallthrough_catchproj != nullptr) {
1539 migrate_outs(_callprojs.fallthrough_catchproj, ctrl);
1540 _igvn.remove_dead_node(_callprojs.fallthrough_catchproj);
1541 }
1542 if (_callprojs.catchall_catchproj != nullptr) {
1543 _igvn.rehash_node_delayed(_callprojs.catchall_catchproj);
1544 _callprojs.catchall_catchproj->set_req(0, top());
1545 }
1546 if (_callprojs.fallthrough_proj != nullptr) {
1547 Node* catchnode = _callprojs.fallthrough_proj->unique_ctrl_out();
1548 _igvn.remove_dead_node(catchnode);
1549 _igvn.remove_dead_node(_callprojs.fallthrough_proj);
1550 }
1551 if (_callprojs.fallthrough_memproj != nullptr) {
1552 migrate_outs(_callprojs.fallthrough_memproj, mem);
1553 _igvn.remove_dead_node(_callprojs.fallthrough_memproj);
1554 }
1555 if (_callprojs.fallthrough_ioproj != nullptr) {
1556 migrate_outs(_callprojs.fallthrough_ioproj, i_o);
1557 _igvn.remove_dead_node(_callprojs.fallthrough_ioproj);
1558 }
1559 if (_callprojs.catchall_memproj != nullptr) {
1560 _igvn.rehash_node_delayed(_callprojs.catchall_memproj);
1561 _callprojs.catchall_memproj->set_req(0, top());
1562 }
1563 if (_callprojs.catchall_ioproj != nullptr) {
1564 _igvn.rehash_node_delayed(_callprojs.catchall_ioproj);
1565 _callprojs.catchall_ioproj->set_req(0, top());
1566 }
1567 #ifndef PRODUCT
1568 if (PrintEliminateAllocations) {
1569 if (alloc->is_AllocateArray()) {
1570 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1571 } else {
1572 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1573 }
1574 }
1575 #endif
1576 _igvn.remove_dead_node(alloc);
1577 }
1578
1579 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1580 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1581 // If initialization is performed by an array copy, any required
1582 // MemBarStoreStore was already added. If the object does not
1583 // escape no need for a MemBarStoreStore. If the object does not
1584 // escape in its initializer and memory barrier (MemBarStoreStore or
1585 // stronger) is already added at exit of initializer, also no need
1663 Node* thread = new ThreadLocalNode();
1664 transform_later(thread);
1665
1666 call->init_req(TypeFunc::Parms + 0, thread);
1667 call->init_req(TypeFunc::Parms + 1, oop);
1668 call->init_req(TypeFunc::Control, ctrl);
1669 call->init_req(TypeFunc::I_O , top()); // does no i/o
1670 call->init_req(TypeFunc::Memory , rawmem);
1671 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1672 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1673 transform_later(call);
1674 ctrl = new ProjNode(call, TypeFunc::Control);
1675 transform_later(ctrl);
1676 rawmem = new ProjNode(call, TypeFunc::Memory);
1677 transform_later(rawmem);
1678 }
1679 }
1680
1681 // Helper for PhaseMacroExpand::expand_allocate_common.
1682 // Initializes the newly-allocated storage.
1683 Node*
1684 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1685 Node* control, Node* rawmem, Node* object,
1686 Node* klass_node, Node* length,
1687 Node* size_in_bytes) {
1688 InitializeNode* init = alloc->initialization();
1689 // Store the klass & mark bits
1690 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem);
1691 if (!mark_node->is_Con()) {
1692 transform_later(mark_node);
1693 }
1694 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1695
1696 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1697 int header_size = alloc->minimum_header_size(); // conservatively small
1698
1699 // Array length
1700 if (length != nullptr) { // Arrays need length field
1701 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1702 // conservatively small header size:
1703 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1704 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1705 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1706 if (is_reference_type(elem, true)) {
1707 elem = T_OBJECT;
1708 }
1709 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1710 }
1711 }
1712
1713 // Clear the object body, if necessary.
1714 if (init == nullptr) {
1715 // The init has somehow disappeared; be cautious and clear everything.
1716 //
1717 // This can happen if a node is allocated but an uncommon trap occurs
1718 // immediately. In this case, the Initialize gets associated with the
1719 // trap, and may be placed in a different (outer) loop, if the Allocate
1720 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1721 // there can be two Allocates to one Initialize. The answer in all these
1722 // edge cases is safety first. It is always safe to clear immediately
1723 // within an Allocate, and then (maybe or maybe not) clear some more later.
1724 if (!(UseTLAB && ZeroTLAB)) {
1725 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1726 header_size, size_in_bytes,
1727 &_igvn);
1728 }
1729 } else {
1730 if (!init->is_complete()) {
1731 // Try to win by zeroing only what the init does not store.
1732 // We can also try to do some peephole optimizations,
1733 // such as combining some adjacent subword stores.
1734 rawmem = init->complete_stores(control, rawmem, object,
1735 header_size, size_in_bytes, &_igvn);
1736 }
1737 // We have no more use for this link, since the AllocateNode goes away:
1738 init->set_req(InitializeNode::RawAddress, top());
1739 // (If we keep the link, it just confuses the register allocator,
1740 // who thinks he sees a real use of the address by the membar.)
1741 }
1742
1743 return rawmem;
1744 }
1745
2075 } // EliminateNestedLocks
2076
2077 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2078 // Look for all locks of this object and mark them and
2079 // corresponding BoxLock nodes as eliminated.
2080 Node* obj = alock->obj_node();
2081 for (uint j = 0; j < obj->outcnt(); j++) {
2082 Node* o = obj->raw_out(j);
2083 if (o->is_AbstractLock() &&
2084 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2085 alock = o->as_AbstractLock();
2086 Node* box = alock->box_node();
2087 // Replace old box node with new eliminated box for all users
2088 // of the same object and mark related locks as eliminated.
2089 mark_eliminated_box(box, obj);
2090 }
2091 }
2092 }
2093 }
2094
2095 // we have determined that this lock/unlock can be eliminated, we simply
2096 // eliminate the node without expanding it.
2097 //
2098 // Note: The membar's associated with the lock/unlock are currently not
2099 // eliminated. This should be investigated as a future enhancement.
2100 //
2101 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2102
2103 if (!alock->is_eliminated()) {
2104 return false;
2105 }
2106 #ifdef ASSERT
2107 if (!alock->is_coarsened()) {
2108 // Check that new "eliminated" BoxLock node is created.
2109 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2110 assert(oldbox->is_eliminated(), "should be done already");
2111 }
2112 #endif
2113
2114 alock->log_lock_optimization(C, "eliminate_lock");
2115
2116 #ifndef PRODUCT
2117 if (PrintEliminateLocks) {
2118 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2119 }
2120 #endif
2121
2122 Node* mem = alock->in(TypeFunc::Memory);
2123 Node* ctrl = alock->in(TypeFunc::Control);
2124 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2125
2126 alock->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2127 // There are 2 projections from the lock. The lock node will
2128 // be deleted when its last use is subsumed below.
2129 assert(alock->outcnt() == 2 &&
2130 _callprojs.fallthrough_proj != nullptr &&
2131 _callprojs.fallthrough_memproj != nullptr,
2132 "Unexpected projections from Lock/Unlock");
2133
2134 Node* fallthroughproj = _callprojs.fallthrough_proj;
2135 Node* memproj_fallthrough = _callprojs.fallthrough_memproj;
2136
2137 // The memory projection from a lock/unlock is RawMem
2138 // The input to a Lock is merged memory, so extract its RawMem input
2139 // (unless the MergeMem has been optimized away.)
2140 if (alock->is_Lock()) {
2141 // Search for MemBarAcquireLock node and delete it also.
2142 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2143 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2144 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2145 Node* memproj = membar->proj_out(TypeFunc::Memory);
2146 _igvn.replace_node(ctrlproj, fallthroughproj);
2147 _igvn.replace_node(memproj, memproj_fallthrough);
2148
2149 // Delete FastLock node also if this Lock node is unique user
2150 // (a loop peeling may clone a Lock node).
2151 Node* flock = alock->as_Lock()->fastlock_node();
2152 if (flock->outcnt() == 1) {
2153 assert(flock->unique_out() == alock, "sanity");
2154 _igvn.replace_node(flock, top());
2155 }
2156 }
2157
2158 // Search for MemBarReleaseLock node and delete it also.
2159 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2160 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2181 Node* mem = lock->in(TypeFunc::Memory);
2182 Node* obj = lock->obj_node();
2183 Node* box = lock->box_node();
2184 Node* flock = lock->fastlock_node();
2185
2186 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2187
2188 // Make the merge point
2189 Node *region;
2190 Node *mem_phi;
2191 Node *slow_path;
2192
2193 region = new RegionNode(3);
2194 // create a Phi for the memory state
2195 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2196
2197 // Optimize test; set region slot 2
2198 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2199 mem_phi->init_req(2, mem);
2200
2201 // Make slow path call
2202 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2203 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2204 obj, box, nullptr);
2205
2206 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2207
2208 // Slow path can only throw asynchronous exceptions, which are always
2209 // de-opted. So the compiler thinks the slow-call can never throw an
2210 // exception. If it DOES throw an exception we would need the debug
2211 // info removed first (since if it throws there is no monitor).
2212 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr &&
2213 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock");
2214
2215 // Capture slow path
2216 // disconnect fall-through projection from call and create a new one
2217 // hook up users of fall-through projection to region
2218 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2219 transform_later(slow_ctrl);
2220 _igvn.hash_delete(_callprojs.fallthrough_proj);
2221 _callprojs.fallthrough_proj->disconnect_inputs(C);
2222 region->init_req(1, slow_ctrl);
2223 // region inputs are now complete
2224 transform_later(region);
2225 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2226
2227 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2228
2229 mem_phi->init_req(1, memproj);
2230
2231 transform_later(mem_phi);
2232
2233 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2234 }
2235
2236 //------------------------------expand_unlock_node----------------------
2237 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2238
2239 Node* ctrl = unlock->in(TypeFunc::Control);
2240 Node* mem = unlock->in(TypeFunc::Memory);
2241 Node* obj = unlock->obj_node();
2242 Node* box = unlock->box_node();
2243
2244 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2245
2246 // No need for a null check on unlock
2247
2248 // Make the merge point
2249 Node *region;
2250 Node *mem_phi;
2251
2252 region = new RegionNode(3);
2253 // create a Phi for the memory state
2254 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2255
2256 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2257 funlock = transform_later( funlock )->as_FastUnlock();
2258 // Optimize test; set region slot 2
2259 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2260 Node *thread = transform_later(new ThreadLocalNode());
2261
2262 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2263 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2264 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2265
2266 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2267 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr &&
2268 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock");
2269
2270 // No exceptions for unlocking
2271 // Capture slow path
2272 // disconnect fall-through projection from call and create a new one
2273 // hook up users of fall-through projection to region
2274 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2275 transform_later(slow_ctrl);
2276 _igvn.hash_delete(_callprojs.fallthrough_proj);
2277 _callprojs.fallthrough_proj->disconnect_inputs(C);
2278 region->init_req(1, slow_ctrl);
2279 // region inputs are now complete
2280 transform_later(region);
2281 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2282
2283 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2284 mem_phi->init_req(1, memproj );
2285 mem_phi->init_req(2, mem);
2286 transform_later(mem_phi);
2287
2288 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2289 }
2290
2291 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2292 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2293 Node* bol = check->unique_out();
2294 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2295 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2296 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2297
2298 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2299 Node* iff = bol->last_out(i);
2300 assert(iff->is_If(), "where's the if?");
2301
2302 if (iff->in(0)->is_top()) {
2303 _igvn.replace_input_of(iff, 1, C->top());
2304 continue;
2305 }
2306
2307 Node* iftrue = iff->as_If()->proj_out(1);
2308 Node* iffalse = iff->as_If()->proj_out(0);
2309 Node* ctrl = iff->in(0);
2310
2311 Node* subklass = nullptr;
2312 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2313 subklass = obj_or_subklass;
2314 } else {
2315 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2316 subklass = _igvn.transform(LoadKlassNode::make(_igvn, nullptr, C->immutable_memory(), k_adr, TypeInstPtr::KLASS));
2317 }
2318
2319 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn);
2320
2321 _igvn.replace_input_of(iff, 0, C->top());
2322 _igvn.replace_node(iftrue, not_subtype_ctrl);
2323 _igvn.replace_node(iffalse, ctrl);
2324 }
2325 _igvn.replace_node(check, C->top());
2326 }
2327
2328 //---------------------------eliminate_macro_nodes----------------------
2329 // Eliminate scalar replaced allocations and associated locks.
2330 void PhaseMacroExpand::eliminate_macro_nodes() {
2331 if (C->macro_count() == 0)
2332 return;
2333 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2334
2335 // Before elimination may re-mark (change to Nested or NonEscObj)
2336 // all associated (same box and obj) lock and unlock nodes.
2337 int cnt = C->macro_count();
2338 for (int i=0; i < cnt; i++) {
2339 Node *n = C->macro_node(i);
2340 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2341 mark_eliminated_locking_nodes(n->as_AbstractLock());
2342 }
2343 }
2344 // Re-marking may break consistency of Coarsened locks.
2345 if (!C->coarsened_locks_consistent()) {
2346 return; // recompile without Coarsened locks if broken
2347 }
2368 }
2369 // Next, attempt to eliminate allocations
2370 _has_locks = false;
2371 progress = true;
2372 while (progress) {
2373 progress = false;
2374 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2375 Node* n = C->macro_node(i - 1);
2376 bool success = false;
2377 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2378 switch (n->class_id()) {
2379 case Node::Class_Allocate:
2380 case Node::Class_AllocateArray:
2381 success = eliminate_allocate_node(n->as_Allocate());
2382 #ifndef PRODUCT
2383 if (success && PrintOptoStatistics) {
2384 Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
2385 }
2386 #endif
2387 break;
2388 case Node::Class_CallStaticJava:
2389 success = eliminate_boxing_node(n->as_CallStaticJava());
2390 break;
2391 case Node::Class_Lock:
2392 case Node::Class_Unlock:
2393 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2394 _has_locks = true;
2395 break;
2396 case Node::Class_ArrayCopy:
2397 break;
2398 case Node::Class_OuterStripMinedLoop:
2399 break;
2400 case Node::Class_SubTypeCheck:
2401 break;
2402 case Node::Class_Opaque1:
2403 break;
2404 default:
2405 assert(n->Opcode() == Op_LoopLimit ||
2406 n->Opcode() == Op_Opaque3 ||
2407 n->Opcode() == Op_Opaque4 ||
2408 n->Opcode() == Op_MaxL ||
2409 n->Opcode() == Op_MinL ||
2410 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2411 "unknown node type in macro list");
2412 }
2413 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2414 progress = progress || success;
2415 }
2416 }
2417 #ifndef PRODUCT
2418 if (PrintOptoStatistics) {
2419 int membar_after = count_MemBar(C);
2420 Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
2421 }
2422 #endif
2423 }
2426 // Returns true if a failure occurred.
2427 bool PhaseMacroExpand::expand_macro_nodes() {
2428 // Last attempt to eliminate macro nodes.
2429 eliminate_macro_nodes();
2430 if (C->failing()) return true;
2431
2432 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2433 bool progress = true;
2434 while (progress) {
2435 progress = false;
2436 for (int i = C->macro_count(); i > 0; i--) {
2437 Node* n = C->macro_node(i-1);
2438 bool success = false;
2439 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2440 if (n->Opcode() == Op_LoopLimit) {
2441 // Remove it from macro list and put on IGVN worklist to optimize.
2442 C->remove_macro_node(n);
2443 _igvn._worklist.push(n);
2444 success = true;
2445 } else if (n->Opcode() == Op_CallStaticJava) {
2446 // Remove it from macro list and put on IGVN worklist to optimize.
2447 C->remove_macro_node(n);
2448 _igvn._worklist.push(n);
2449 success = true;
2450 } else if (n->is_Opaque1()) {
2451 _igvn.replace_node(n, n->in(1));
2452 success = true;
2453 #if INCLUDE_RTM_OPT
2454 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2455 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2456 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2457 Node* cmp = n->unique_out();
2458 #ifdef ASSERT
2459 // Validate graph.
2460 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2461 BoolNode* bol = cmp->unique_out()->as_Bool();
2462 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2463 (bol->_test._test == BoolTest::ne), "");
2464 IfNode* ifn = bol->unique_out()->as_If();
2465 assert((ifn->outcnt() == 2) &&
2466 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != nullptr, "");
2467 #endif
2468 Node* repl = n->in(1);
2469 if (!_has_locks) {
2544 // Worst case is a macro node gets expanded into about 200 nodes.
2545 // Allow 50% more for optimization.
2546 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2547 return true;
2548 }
2549
2550 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2551 switch (n->class_id()) {
2552 case Node::Class_Lock:
2553 expand_lock_node(n->as_Lock());
2554 break;
2555 case Node::Class_Unlock:
2556 expand_unlock_node(n->as_Unlock());
2557 break;
2558 case Node::Class_ArrayCopy:
2559 expand_arraycopy_node(n->as_ArrayCopy());
2560 break;
2561 case Node::Class_SubTypeCheck:
2562 expand_subtypecheck_node(n->as_SubTypeCheck());
2563 break;
2564 default:
2565 assert(false, "unknown node type in macro list");
2566 }
2567 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2568 if (C->failing()) return true;
2569
2570 // Clean up the graph so we're less likely to hit the maximum node
2571 // limit
2572 _igvn.set_delay_transform(false);
2573 _igvn.optimize();
2574 if (C->failing()) return true;
2575 _igvn.set_delay_transform(true);
2576 }
2577
2578 // All nodes except Allocate nodes are expanded now. There could be
2579 // new optimization opportunities (such as folding newly created
2580 // load from a just allocated object). Run IGVN.
2581
2582 // expand "macro" nodes
2583 // nodes are removed from the macro list as they are processed
|
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "compiler/compileLog.hpp"
28 #include "gc/shared/collectedHeap.inline.hpp"
29 #include "gc/shared/tlab_globals.hpp"
30 #include "libadt/vectset.hpp"
31 #include "memory/universe.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/callnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/cfgnode.hpp"
37 #include "opto/compile.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/graphKit.hpp"
40 #include "opto/inlinetypenode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/locknode.hpp"
43 #include "opto/loopnode.hpp"
44 #include "opto/macro.hpp"
45 #include "opto/memnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/node.hpp"
48 #include "opto/opaquenode.hpp"
49 #include "opto/phaseX.hpp"
50 #include "opto/rootnode.hpp"
51 #include "opto/runtime.hpp"
52 #include "opto/subnode.hpp"
53 #include "opto/subtypenode.hpp"
54 #include "opto/type.hpp"
55 #include "prims/jvmtiExport.hpp"
56 #include "runtime/continuation.hpp"
57 #include "runtime/sharedRuntime.hpp"
58 #include "runtime/stubRoutines.hpp"
59 #include "utilities/macros.hpp"
60 #include "utilities/powerOfTwo.hpp"
61 #if INCLUDE_G1GC
62 #include "gc/g1/g1ThreadLocalData.hpp"
63 #endif // INCLUDE_G1GC
64
65
66 //
67 // Replace any references to "oldref" in inputs to "use" with "newref".
68 // Returns the number of replacements made.
69 //
70 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
71 int nreplacements = 0;
72 uint req = use->req();
73 for (uint j = 0; j < use->len(); j++) {
74 Node *uin = use->in(j);
75 if (uin == oldref) {
76 if (j < req)
77 use->set_req(j, newref);
78 else
79 use->set_prec(j, newref);
80 nreplacements++;
81 } else if (j >= req && uin == nullptr) {
82 break;
83 }
84 }
85 return nreplacements;
86 }
87
88 Node* PhaseMacroExpand::opt_bits_test(Node* ctrl, Node* region, int edge, Node* word, int mask, int bits, bool return_fast_path) {
89 Node* cmp;
90 if (mask != 0) {
91 Node* and_node = transform_later(new AndXNode(word, MakeConX(mask)));
92 cmp = transform_later(new CmpXNode(and_node, MakeConX(bits)));
93 } else {
94 cmp = word;
95 }
96 Node* bol = transform_later(new BoolNode(cmp, BoolTest::ne));
97 IfNode* iff = new IfNode( ctrl, bol, PROB_MIN, COUNT_UNKNOWN );
98 transform_later(iff);
99
100 // Fast path taken.
101 Node *fast_taken = transform_later(new IfFalseNode(iff));
102
103 // Fast path not-taken, i.e. slow path
104 Node *slow_taken = transform_later(new IfTrueNode(iff));
105
106 if (return_fast_path) {
107 region->init_req(edge, slow_taken); // Capture slow-control
130 // Slow-path call
131 CallNode *call = leaf_name
132 ? (CallNode*)new CallLeafNode ( slow_call_type, slow_call, leaf_name, TypeRawPtr::BOTTOM )
133 : (CallNode*)new CallStaticJavaNode( slow_call_type, slow_call, OptoRuntime::stub_name(slow_call), TypeRawPtr::BOTTOM );
134
135 // Slow path call has no side-effects, uses few values
136 copy_predefined_input_for_runtime_call(slow_path, oldcall, call );
137 if (parm0 != nullptr) call->init_req(TypeFunc::Parms+0, parm0);
138 if (parm1 != nullptr) call->init_req(TypeFunc::Parms+1, parm1);
139 if (parm2 != nullptr) call->init_req(TypeFunc::Parms+2, parm2);
140 call->copy_call_debug_info(&_igvn, oldcall);
141 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
142 _igvn.replace_node(oldcall, call);
143 transform_later(call);
144
145 return call;
146 }
147
148 void PhaseMacroExpand::eliminate_gc_barrier(Node* p2x) {
149 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
150 bs->eliminate_gc_barrier(&_igvn, p2x);
151 #ifndef PRODUCT
152 if (PrintOptoStatistics) {
153 Atomic::inc(&PhaseMacroExpand::_GC_barriers_removed_counter);
154 }
155 #endif
156 }
157
158 // Search for a memory operation for the specified memory slice.
159 static Node *scan_mem_chain(Node *mem, int alias_idx, int offset, Node *start_mem, Node *alloc, PhaseGVN *phase) {
160 Node *orig_mem = mem;
161 Node *alloc_mem = alloc->in(TypeFunc::Memory);
162 const TypeOopPtr *tinst = phase->C->get_adr_type(alias_idx)->isa_oopptr();
163 while (true) {
164 if (mem == alloc_mem || mem == start_mem ) {
165 return mem; // hit one of our sentinels
166 } else if (mem->is_MergeMem()) {
167 mem = mem->as_MergeMem()->memory_at(alias_idx);
168 } else if (mem->is_Proj() && mem->as_Proj()->_con == TypeFunc::Memory) {
169 Node *in = mem->in(0);
170 // we can safely skip over safepoints, calls, locks and membars because we
184 ArrayCopyNode* ac = nullptr;
185 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
186 if (ac != nullptr) {
187 assert(ac->is_clonebasic(), "Only basic clone is a non escaping clone");
188 return ac;
189 }
190 }
191 mem = in->in(TypeFunc::Memory);
192 } else {
193 #ifdef ASSERT
194 in->dump();
195 mem->dump();
196 assert(false, "unexpected projection");
197 #endif
198 }
199 } else if (mem->is_Store()) {
200 const TypePtr* atype = mem->as_Store()->adr_type();
201 int adr_idx = phase->C->get_alias_index(atype);
202 if (adr_idx == alias_idx) {
203 assert(atype->isa_oopptr(), "address type must be oopptr");
204 int adr_offset = atype->flat_offset();
205 uint adr_iid = atype->is_oopptr()->instance_id();
206 // Array elements references have the same alias_idx
207 // but different offset and different instance_id.
208 if (adr_offset == offset && adr_iid == alloc->_idx) {
209 return mem;
210 }
211 } else {
212 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
213 }
214 mem = mem->in(MemNode::Memory);
215 } else if (mem->is_ClearArray()) {
216 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
217 // Can not bypass initialization of the instance
218 // we are looking.
219 debug_only(intptr_t offset;)
220 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
221 InitializeNode* init = alloc->as_Allocate()->initialization();
222 // We are looking for stored value, return Initialize node
223 // or memory edge from Allocate node.
224 if (init != nullptr) {
229 }
230 // Otherwise skip it (the call updated 'mem' value).
231 } else if (mem->Opcode() == Op_SCMemProj) {
232 mem = mem->in(0);
233 Node* adr = nullptr;
234 if (mem->is_LoadStore()) {
235 adr = mem->in(MemNode::Address);
236 } else {
237 assert(mem->Opcode() == Op_EncodeISOArray ||
238 mem->Opcode() == Op_StrCompressedCopy, "sanity");
239 adr = mem->in(3); // Destination array
240 }
241 const TypePtr* atype = adr->bottom_type()->is_ptr();
242 int adr_idx = phase->C->get_alias_index(atype);
243 if (adr_idx == alias_idx) {
244 DEBUG_ONLY(mem->dump();)
245 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
246 return nullptr;
247 }
248 mem = mem->in(MemNode::Memory);
249 } else if (mem->Opcode() == Op_StrInflatedCopy) {
250 Node* adr = mem->in(3); // Destination array
251 const TypePtr* atype = adr->bottom_type()->is_ptr();
252 int adr_idx = phase->C->get_alias_index(atype);
253 if (adr_idx == alias_idx) {
254 DEBUG_ONLY(mem->dump();)
255 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
256 return nullptr;
257 }
258 mem = mem->in(MemNode::Memory);
259 } else {
260 return mem;
261 }
262 assert(mem != orig_mem, "dead memory loop");
263 }
264 }
265
266 // Generate loads from source of the arraycopy for fields of
267 // destination needed at a deoptimization point
268 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
269 BasicType bt = ft;
274 }
275 Node* res = nullptr;
276 if (ac->is_clonebasic()) {
277 assert(ac->in(ArrayCopyNode::Src) != ac->in(ArrayCopyNode::Dest), "clone source equals destination");
278 Node* base = ac->in(ArrayCopyNode::Src);
279 Node* adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(offset)));
280 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
281 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
282 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
283 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
284 } else {
285 if (ac->modifies(offset, offset, &_igvn, true)) {
286 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
287 uint shift = exact_log2(type2aelembytes(bt));
288 Node* src_pos = ac->in(ArrayCopyNode::SrcPos);
289 Node* dest_pos = ac->in(ArrayCopyNode::DestPos);
290 const TypeInt* src_pos_t = _igvn.type(src_pos)->is_int();
291 const TypeInt* dest_pos_t = _igvn.type(dest_pos)->is_int();
292
293 Node* adr = nullptr;
294 Node* base = ac->in(ArrayCopyNode::Src);
295 const TypeAryPtr* adr_type = _igvn.type(base)->is_aryptr();
296 if (adr_type->is_flat()) {
297 shift = adr_type->flat_log_elem_size();
298 }
299 if (src_pos_t->is_con() && dest_pos_t->is_con()) {
300 intptr_t off = ((src_pos_t->get_con() - dest_pos_t->get_con()) << shift) + offset;
301 adr = _igvn.transform(new AddPNode(base, base, _igvn.MakeConX(off)));
302 adr_type = _igvn.type(adr)->is_aryptr();
303 assert(adr_type == _igvn.type(base)->is_aryptr()->add_field_offset_and_offset(off), "incorrect address type");
304 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
305 // Don't emit a new load from src if src == dst but try to get the value from memory instead
306 return value_from_mem(ac->in(TypeFunc::Memory), ctl, ft, ftype, adr_type, alloc);
307 }
308 } else {
309 if (ac->in(ArrayCopyNode::Src) == ac->in(ArrayCopyNode::Dest)) {
310 // Non constant offset in the array: we can't statically
311 // determine the value
312 return nullptr;
313 }
314 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
315 #ifdef _LP64
316 diff = _igvn.transform(new ConvI2LNode(diff));
317 #endif
318 diff = _igvn.transform(new LShiftXNode(diff, _igvn.intcon(shift)));
319
320 Node* off = _igvn.transform(new AddXNode(_igvn.MakeConX(offset), diff));
321 adr = _igvn.transform(new AddPNode(base, base, off));
322 // In the case of a flat inline type array, each field has its
323 // own slice so we need to extract the field being accessed from
324 // the address computation
325 adr_type = adr_type->add_field_offset_and_offset(offset)->add_offset(Type::OffsetBot)->is_aryptr();
326 adr = _igvn.transform(new CastPPNode(adr, adr_type));
327 }
328 MergeMemNode* mergemen = _igvn.transform(MergeMemNode::make(mem))->as_MergeMem();
329 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
330 res = ArrayCopyNode::load(bs, &_igvn, ctl, mergemen, adr, adr_type, type, bt);
331 }
332 }
333 if (res != nullptr) {
334 if (ftype->isa_narrowoop()) {
335 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
336 assert(res->isa_DecodeN(), "should be narrow oop");
337 res = _igvn.transform(new EncodePNode(res, ftype));
338 }
339 return res;
340 }
341 return nullptr;
342 }
343
344 //
345 // Given a Memory Phi, compute a value Phi containing the values from stores
346 // on the input paths.
347 // Note: this function is recursive, its depth is limited by the "level" argument
348 // Returns the computed Phi, or null if it cannot compute it.
349 Node *PhaseMacroExpand::value_from_mem_phi(Node *mem, BasicType ft, const Type *phi_type, const TypeOopPtr *adr_t, AllocateNode *alloc, Node_Stack *value_phis, int level) {
350 assert(mem->is_Phi(), "sanity");
351 int alias_idx = C->get_alias_index(adr_t);
352 int offset = adr_t->flat_offset();
353 int instance_id = adr_t->instance_id();
354
355 // Check if an appropriate value phi already exists.
356 Node* region = mem->in(0);
357 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
358 Node* phi = region->fast_out(k);
359 if (phi->is_Phi() && phi != mem &&
360 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
361 return phi;
362 }
363 }
364 // Check if an appropriate new value phi already exists.
365 Node* new_phi = value_phis->find(mem->_idx);
366 if (new_phi != nullptr)
367 return new_phi;
368
369 if (level <= 0) {
370 return nullptr; // Give up: phi tree too deep
371 }
372 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
373 Node *alloc_mem = alloc->in(TypeFunc::Memory);
374
375 uint length = mem->req();
376 GrowableArray <Node *> values(length, length, nullptr);
377
378 // create a new Phi for the value
379 PhiNode *phi = new PhiNode(mem->in(0), phi_type, nullptr, mem->_idx, instance_id, alias_idx, offset);
380 transform_later(phi);
381 value_phis->push(phi, mem->_idx);
382
383 for (uint j = 1; j < length; j++) {
384 Node *in = mem->in(j);
385 if (in == nullptr || in->is_top()) {
386 values.at_put(j, in);
387 } else {
388 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
389 if (val == start_mem || val == alloc_mem) {
390 // hit a sentinel, return appropriate 0 value
391 Node* default_value = alloc->in(AllocateNode::DefaultValue);
392 if (default_value != nullptr) {
393 values.at_put(j, default_value);
394 } else {
395 assert(alloc->in(AllocateNode::RawDefaultValue) == nullptr, "default value may not be null");
396 values.at_put(j, _igvn.zerocon(ft));
397 }
398 continue;
399 }
400 if (val->is_Initialize()) {
401 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
402 }
403 if (val == nullptr) {
404 return nullptr; // can't find a value on this path
405 }
406 if (val == mem) {
407 values.at_put(j, mem);
408 } else if (val->is_Store()) {
409 Node* n = val->in(MemNode::ValueIn);
410 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
411 n = bs->step_over_gc_barrier(n);
412 if (is_subword_type(ft)) {
413 n = Compile::narrow_value(ft, n, phi_type, &_igvn, true);
414 }
415 values.at_put(j, n);
416 } else if(val->is_Proj() && val->in(0) == alloc) {
417 Node* default_value = alloc->in(AllocateNode::DefaultValue);
418 if (default_value != nullptr) {
419 values.at_put(j, default_value);
420 } else {
421 assert(alloc->in(AllocateNode::RawDefaultValue) == nullptr, "default value may not be null");
422 values.at_put(j, _igvn.zerocon(ft));
423 }
424 } else if (val->is_Phi()) {
425 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
426 if (val == nullptr) {
427 return nullptr;
428 }
429 values.at_put(j, val);
430 } else if (val->Opcode() == Op_SCMemProj) {
431 assert(val->in(0)->is_LoadStore() ||
432 val->in(0)->Opcode() == Op_EncodeISOArray ||
433 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
434 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
435 return nullptr;
436 } else if (val->is_ArrayCopy()) {
437 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
438 if (res == nullptr) {
439 return nullptr;
440 }
441 values.at_put(j, res);
442 } else {
443 DEBUG_ONLY( val->dump(); )
447 }
448 }
449 // Set Phi's inputs
450 for (uint j = 1; j < length; j++) {
451 if (values.at(j) == mem) {
452 phi->init_req(j, phi);
453 } else {
454 phi->init_req(j, values.at(j));
455 }
456 }
457 return phi;
458 }
459
460 // Search the last value stored into the object's field.
461 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
462 assert(adr_t->is_known_instance_field(), "instance required");
463 int instance_id = adr_t->instance_id();
464 assert((uint)instance_id == alloc->_idx, "wrong allocation");
465
466 int alias_idx = C->get_alias_index(adr_t);
467 int offset = adr_t->flat_offset();
468 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
469 Node *alloc_mem = alloc->in(TypeFunc::Memory);
470 VectorSet visited;
471
472 bool done = sfpt_mem == alloc_mem;
473 Node *mem = sfpt_mem;
474 while (!done) {
475 if (visited.test_set(mem->_idx)) {
476 return nullptr; // found a loop, give up
477 }
478 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
479 if (mem == start_mem || mem == alloc_mem) {
480 done = true; // hit a sentinel, return appropriate 0 value
481 } else if (mem->is_Initialize()) {
482 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
483 if (mem == nullptr) {
484 done = true; // Something went wrong.
485 } else if (mem->is_Store()) {
486 const TypePtr* atype = mem->as_Store()->adr_type();
487 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
488 done = true;
489 }
490 } else if (mem->is_Store()) {
491 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
492 assert(atype != nullptr, "address type must be oopptr");
493 assert(C->get_alias_index(atype) == alias_idx &&
494 atype->is_known_instance_field() && atype->flat_offset() == offset &&
495 atype->instance_id() == instance_id, "store is correct memory slice");
496 done = true;
497 } else if (mem->is_Phi()) {
498 // try to find a phi's unique input
499 Node *unique_input = nullptr;
500 Node *top = C->top();
501 for (uint i = 1; i < mem->req(); i++) {
502 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
503 if (n == nullptr || n == top || n == mem) {
504 continue;
505 } else if (unique_input == nullptr) {
506 unique_input = n;
507 } else if (unique_input != n) {
508 unique_input = top;
509 break;
510 }
511 }
512 if (unique_input != nullptr && unique_input != top) {
513 mem = unique_input;
514 } else {
515 done = true;
516 }
517 } else if (mem->is_ArrayCopy()) {
518 done = true;
519 } else {
520 DEBUG_ONLY( mem->dump(); )
521 assert(false, "unexpected node");
522 }
523 }
524 if (mem != nullptr) {
525 if (mem == start_mem || mem == alloc_mem) {
526 // hit a sentinel, return appropriate 0 value
527 Node* default_value = alloc->in(AllocateNode::DefaultValue);
528 if (default_value != nullptr) {
529 return default_value;
530 }
531 assert(alloc->in(AllocateNode::RawDefaultValue) == nullptr, "default value may not be null");
532 return _igvn.zerocon(ft);
533 } else if (mem->is_Store()) {
534 Node* n = mem->in(MemNode::ValueIn);
535 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
536 n = bs->step_over_gc_barrier(n);
537 return n;
538 } else if (mem->is_Phi()) {
539 // attempt to produce a Phi reflecting the values on the input paths of the Phi
540 Node_Stack value_phis(8);
541 Node* phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
542 if (phi != nullptr) {
543 return phi;
544 } else {
545 // Kill all new Phis
546 while(value_phis.is_nonempty()) {
547 Node* n = value_phis.node();
548 _igvn.replace_node(n, C->top());
549 value_phis.pop();
550 }
551 }
552 } else if (mem->is_ArrayCopy()) {
553 Node* ctl = mem->in(0);
554 Node* m = mem->in(TypeFunc::Memory);
555 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
556 // pin the loads in the uncommon trap path
557 ctl = sfpt_ctl;
558 m = sfpt_mem;
559 }
560 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
561 }
562 }
563 // Something went wrong.
564 return nullptr;
565 }
566
567 // Search the last value stored into the inline type's fields.
568 Node* PhaseMacroExpand::inline_type_from_mem(Node* mem, Node* ctl, ciInlineKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
569 // Subtract the offset of the first field to account for the missing oop header
570 offset -= vk->first_field_offset();
571 // Create a new InlineTypeNode and retrieve the field values from memory
572 InlineTypeNode* vt = InlineTypeNode::make_uninitialized(_igvn, vk);
573 transform_later(vt);
574 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
575 ciType* field_type = vt->field_type(i);
576 int field_offset = offset + vt->field_offset(i);
577 Node* value = nullptr;
578 if (vt->field_is_flat(i)) {
579 value = inline_type_from_mem(mem, ctl, field_type->as_inline_klass(), adr_type, field_offset, alloc);
580 } else {
581 const Type* ft = Type::get_const_type(field_type);
582 BasicType bt = type2field[field_type->basic_type()];
583 if (UseCompressedOops && !is_java_primitive(bt)) {
584 ft = ft->make_narrowoop();
585 bt = T_NARROWOOP;
586 }
587 // Each inline type field has its own memory slice
588 adr_type = adr_type->with_field_offset(field_offset);
589 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc);
590 if (value != nullptr && ft->isa_narrowoop()) {
591 assert(UseCompressedOops, "unexpected narrow oop");
592 if (value->is_EncodeP()) {
593 value = value->in(1);
594 } else {
595 value = transform_later(new DecodeNNode(value, value->get_ptr_type()));
596 }
597 }
598 }
599 if (value != nullptr) {
600 vt->set_field_value(i, value);
601 } else {
602 // We might have reached the TrackedInitializationLimit
603 return nullptr;
604 }
605 }
606 return vt;
607 }
608
609 // Check the possibility of scalar replacement.
610 bool PhaseMacroExpand::can_eliminate_allocation(PhaseIterGVN* igvn, AllocateNode *alloc, GrowableArray <SafePointNode *>* safepoints) {
611 // Scan the uses of the allocation to check for anything that would
612 // prevent us from eliminating it.
613 NOT_PRODUCT( const char* fail_eliminate = nullptr; )
614 DEBUG_ONLY( Node* disq_node = nullptr; )
615 bool can_eliminate = true;
616 bool reduce_merge_precheck = (safepoints == nullptr);
617
618 Unique_Node_List worklist;
619 Node* res = alloc->result_cast();
620 const TypeOopPtr* res_type = nullptr;
621 if (res == nullptr) {
622 // All users were eliminated.
623 } else if (!res->is_CheckCastPP()) {
624 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
625 can_eliminate = false;
626 } else {
627 worklist.push(res);
628 res_type = igvn->type(res)->isa_oopptr();
629 if (res_type == nullptr) {
630 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
631 can_eliminate = false;
632 } else if (res_type->isa_aryptr()) {
633 int length = alloc->in(AllocateNode::ALength)->find_int_con(-1);
634 if (length < 0) {
635 NOT_PRODUCT(fail_eliminate = "Array's size is not constant";)
636 can_eliminate = false;
637 }
638 }
639 }
640
641 while (can_eliminate && worklist.size() > 0) {
642 BarrierSetC2 *bs = BarrierSet::barrier_set()->barrier_set_c2();
643 res = worklist.pop();
644 for (DUIterator_Fast jmax, j = res->fast_outs(jmax); j < jmax && can_eliminate; j++) {
645 Node* use = res->fast_out(j);
646
647 if (use->is_AddP()) {
648 const TypePtr* addp_type = igvn->type(use)->is_ptr();
649 int offset = addp_type->offset();
650
651 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
652 NOT_PRODUCT(fail_eliminate = "Undefined field reference";)
653 can_eliminate = false;
654 break;
655 }
656 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
657 k < kmax && can_eliminate; k++) {
658 Node* n = use->fast_out(k);
659 if (!n->is_Store() && n->Opcode() != Op_CastP2X && !bs->is_gc_pre_barrier_node(n)) {
660 DEBUG_ONLY(disq_node = n;)
661 if (n->is_Load() || n->is_LoadStore()) {
662 NOT_PRODUCT(fail_eliminate = "Field load";)
663 } else {
664 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
672 use->as_ArrayCopy()->is_copyof_validated() ||
673 use->as_ArrayCopy()->is_copyofrange_validated()) &&
674 use->in(ArrayCopyNode::Dest) == res) {
675 // ok to eliminate
676 } else if (use->is_SafePoint()) {
677 SafePointNode* sfpt = use->as_SafePoint();
678 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
679 // Object is passed as argument.
680 DEBUG_ONLY(disq_node = use;)
681 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
682 can_eliminate = false;
683 }
684 Node* sfptMem = sfpt->memory();
685 if (sfptMem == nullptr || sfptMem->is_top()) {
686 DEBUG_ONLY(disq_node = use;)
687 NOT_PRODUCT(fail_eliminate = "null or TOP memory";)
688 can_eliminate = false;
689 } else if (!reduce_merge_precheck) {
690 safepoints->append_if_missing(sfpt);
691 }
692 } else if (use->is_InlineType() && use->as_InlineType()->get_oop() == res) {
693 // Look at uses
694 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
695 Node* u = use->fast_out(k);
696 if (u->is_InlineType()) {
697 // Use in flat field can be eliminated
698 InlineTypeNode* vt = u->as_InlineType();
699 for (uint i = 0; i < vt->field_count(); ++i) {
700 if (vt->field_value(i) == use && !vt->field_is_flat(i)) {
701 can_eliminate = false; // Use in non-flat field
702 break;
703 }
704 }
705 } else {
706 // Add other uses to the worklist to process individually
707 worklist.push(u);
708 }
709 }
710 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
711 // Store to mark word of inline type larval buffer
712 assert(res_type->is_inlinetypeptr(), "Unexpected store to mark word");
713 } else if (reduce_merge_precheck && (use->is_Phi() || use->is_EncodeP() || use->Opcode() == Op_MemBarRelease)) {
714 // Nothing to do
715 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
716 if (use->is_Phi()) {
717 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
718 NOT_PRODUCT(fail_eliminate = "Object is return value";)
719 } else {
720 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
721 }
722 DEBUG_ONLY(disq_node = use;)
723 } else {
724 if (use->Opcode() == Op_Return) {
725 NOT_PRODUCT(fail_eliminate = "Object is return value";)
726 } else {
727 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
728 }
729 DEBUG_ONLY(disq_node = use;)
730 }
731 can_eliminate = false;
732 } else {
733 assert(use->Opcode() == Op_CastP2X, "should be");
734 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
735 }
736 }
737 }
738
739 #ifndef PRODUCT
740 if (PrintEliminateAllocations && safepoints != nullptr) {
741 if (can_eliminate) {
742 tty->print("Scalar ");
743 if (res == nullptr)
744 alloc->dump();
745 else
746 res->dump();
747 } else {
748 tty->print("NotScalar (%s)", fail_eliminate);
749 if (res == nullptr)
750 alloc->dump();
751 else
752 res->dump();
753 #ifdef ASSERT
754 if (disq_node != nullptr) {
755 tty->print(" >>>> ");
756 disq_node->dump();
757 }
758 #endif /*ASSERT*/
759 }
760 }
761 #endif
762 return can_eliminate;
763 }
764
765 void PhaseMacroExpand::undo_previous_scalarizations(GrowableArray <SafePointNode *> safepoints_done, AllocateNode* alloc) {
766 Node* res = alloc->result_cast();
767 int nfields = 0;
792 JVMState *jvms = sfpt_done->jvms();
793 jvms->set_endoff(sfpt_done->req());
794 // Now make a pass over the debug information replacing any references
795 // to SafePointScalarObjectNode with the allocated object.
796 int start = jvms->debug_start();
797 int end = jvms->debug_end();
798 for (int i = start; i < end; i++) {
799 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
800 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
801 if (scobj->first_index(jvms) == sfpt_done->req() &&
802 scobj->n_fields() == (uint)nfields) {
803 assert(scobj->alloc() == alloc, "sanity");
804 sfpt_done->set_req(i, res);
805 }
806 }
807 }
808 _igvn._worklist.push(sfpt_done);
809 }
810 }
811
812 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode *alloc, SafePointNode* sfpt,
813 Unique_Node_List* value_worklist) {
814 // Fields of scalar objs are referenced only at the end
815 // of regular debuginfo at the last (youngest) JVMS.
816 // Record relative start index.
817 ciInstanceKlass* iklass = nullptr;
818 BasicType basic_elem_type = T_ILLEGAL;
819 const Type* field_type = nullptr;
820 const TypeOopPtr* res_type = nullptr;
821 int nfields = 0;
822 int array_base = 0;
823 int element_size = 0;
824 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
825 Node* res = alloc->result_cast();
826
827 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
828 assert(sfpt->jvms() != nullptr, "missed JVMS");
829
830 if (res != nullptr) { // Could be null when there are no users
831 res_type = _igvn.type(res)->isa_oopptr();
832
833 if (res_type->isa_instptr()) {
834 // find the fields of the class which will be needed for safepoint debug information
835 iklass = res_type->is_instptr()->instance_klass();
836 nfields = iklass->nof_nonstatic_fields();
837 } else {
838 // find the array's elements which will be needed for safepoint debug information
839 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
840 assert(nfields >= 0, "must be an array klass.");
841 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type();
842 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
843 element_size = type2aelembytes(basic_elem_type);
844 field_type = res_type->is_aryptr()->elem();
845 if (res_type->is_flat()) {
846 // Flat inline type array
847 element_size = res_type->is_aryptr()->flat_elem_size();
848 }
849 }
850 }
851
852 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, nfields);
853 sobj->init_req(0, C->root());
854 transform_later(sobj);
855
856 // Scan object's fields adding an input to the safepoint for each field.
857 for (int j = 0; j < nfields; j++) {
858 intptr_t offset;
859 ciField* field = nullptr;
860 if (iklass != nullptr) {
861 field = iklass->nonstatic_field_at(j);
862 offset = field->offset_in_bytes();
863 ciType* elem_type = field->type();
864 basic_elem_type = field->layout_type();
865 assert(!field->is_flat(), "flat inline type fields should not have safepoint uses");
866
867 // The next code is taken from Parse::do_get_xxx().
868 if (is_reference_type(basic_elem_type)) {
869 if (!elem_type->is_loaded()) {
870 field_type = TypeInstPtr::BOTTOM;
871 } else if (field != nullptr && field->is_static_constant()) {
872 ciObject* con = field->constant_value().as_object();
873 // Do not "join" in the previous type; it doesn't add value,
874 // and may yield a vacuous result if the field is of interface type.
875 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
876 assert(field_type != nullptr, "field singleton type must be consistent");
877 } else {
878 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
879 }
880 if (UseCompressedOops) {
881 field_type = field_type->make_narrowoop();
882 basic_elem_type = T_NARROWOOP;
883 }
884 } else {
885 field_type = Type::get_const_basic_type(basic_elem_type);
886 }
887 } else {
888 offset = array_base + j * (intptr_t)element_size;
889 }
890
891 Node* field_val = nullptr;
892 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
893 if (res_type->is_flat()) {
894 ciInlineKlass* vk = res_type->is_aryptr()->elem()->inline_klass();
895 assert(vk->flat_array(), "must be flat");
896 field_val = inline_type_from_mem(sfpt->memory(), sfpt->control(), vk, field_addr_type->isa_aryptr(), 0, alloc);
897 } else {
898 field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
899 }
900
901 // We weren't able to find a value for this field,
902 // give up on eliminating this allocation.
903 if (field_val == nullptr) {
904 uint last = sfpt->req() - 1;
905 for (int k = 0; k < j; k++) {
906 sfpt->del_req(last--);
907 }
908 _igvn._worklist.push(sfpt);
909
910 #ifndef PRODUCT
911 if (PrintEliminateAllocations) {
912 if (field != nullptr) {
913 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
914 field->print();
915 int field_idx = C->get_alias_index(field_addr_type);
916 tty->print(" (alias_idx=%d)", field_idx);
917 } else { // Array's element
918 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, j);
919 }
920 tty->print(", which prevents elimination of: ");
921 if (res == nullptr)
922 alloc->dump();
923 else
924 res->dump();
925 }
926 #endif
927
928 return nullptr;
929 }
930
931 if (UseCompressedOops && field_type->isa_narrowoop()) {
932 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
933 // to be able scalar replace the allocation.
934 if (field_val->is_EncodeP()) {
935 field_val = field_val->in(1);
936 } else if (!field_val->is_InlineType()) {
937 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
938 }
939 }
940 if (field_val->is_InlineType()) {
941 // Keep track of inline types to scalarize them later
942 value_worklist->push(field_val);
943 }
944 sfpt->add_req(field_val);
945 }
946
947 sfpt->jvms()->set_endoff(sfpt->req());
948
949 return sobj;
950 }
951
952 // Do scalar replacement.
953 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
954 GrowableArray <SafePointNode *> safepoints_done;
955 Node* res = alloc->result_cast();
956 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
957 const TypeOopPtr* res_type = nullptr;
958 if (res != nullptr) { // Could be null when there are no users
959 res_type = _igvn.type(res)->isa_oopptr();
960 }
961
962 // Process the safepoint uses
963 assert(safepoints.length() == 0 || !res_type->is_inlinetypeptr(), "Inline type allocations should not have safepoint uses");
964 Unique_Node_List value_worklist;
965 while (safepoints.length() > 0) {
966 SafePointNode* sfpt = safepoints.pop();
967 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt, &value_worklist);
968
969 if (sobj == nullptr) {
970 undo_previous_scalarizations(safepoints_done, alloc);
971 return false;
972 }
973
974 // Now make a pass over the debug information replacing any references
975 // to the allocated object with "sobj"
976 JVMState *jvms = sfpt->jvms();
977 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
978 _igvn._worklist.push(sfpt);
979
980 // keep it for rollback
981 safepoints_done.append_if_missing(sfpt);
982 }
983 // Scalarize inline types that were added to the safepoint.
984 // Don't allow linking a constant oop (if available) for flat array elements
985 // because Deoptimization::reassign_flat_array_elements needs field values.
986 bool allow_oop = (res_type != nullptr) && !res_type->is_flat();
987 for (uint i = 0; i < value_worklist.size(); ++i) {
988 InlineTypeNode* vt = value_worklist.at(i)->as_InlineType();
989 vt->make_scalar_in_safepoints(&_igvn, allow_oop);
990 }
991 return true;
992 }
993
994 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
995 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
996 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
997 if (ctl_proj != nullptr) {
998 igvn.replace_node(ctl_proj, n->in(0));
999 }
1000 if (mem_proj != nullptr) {
1001 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1002 }
1003 }
1004
1005 // Process users of eliminated allocation.
1006 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc, bool inline_alloc) {
1007 Unique_Node_List worklist;
1008 Node* res = alloc->result_cast();
1009 if (res != nullptr) {
1010 worklist.push(res);
1011 }
1012 while (worklist.size() > 0) {
1013 res = worklist.pop();
1014 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1015 Node *use = res->last_out(j);
1016 uint oc1 = res->outcnt();
1017
1018 if (use->is_AddP()) {
1019 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1020 Node *n = use->last_out(k);
1021 uint oc2 = use->outcnt();
1022 if (n->is_Store()) {
1023 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); p < pmax; p++) {
1024 MemBarNode* mb = n->fast_out(p)->isa_MemBar();
1025 if (mb != nullptr && mb->req() <= MemBarNode::Precedent && mb->in(MemBarNode::Precedent) == n) {
1026 // MemBarVolatiles should have been removed by MemBarNode::Ideal() for non-inline allocations
1027 assert(inline_alloc, "MemBarVolatile should be eliminated for non-escaping object");
1028 mb->remove(&_igvn);
1029 }
1030 }
1031 _igvn.replace_node(n, n->in(MemNode::Memory));
1032 } else {
1033 eliminate_gc_barrier(n);
1034 }
1035 k -= (oc2 - use->outcnt());
1036 }
1037 _igvn.remove_dead_node(use);
1038 } else if (use->is_ArrayCopy()) {
1039 // Disconnect ArrayCopy node
1040 ArrayCopyNode* ac = use->as_ArrayCopy();
1041 if (ac->is_clonebasic()) {
1042 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1043 disconnect_projections(ac, _igvn);
1044 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1045 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
1046 disconnect_projections(membar_before->as_MemBar(), _igvn);
1047 if (membar_after->is_MemBar()) {
1048 disconnect_projections(membar_after->as_MemBar(), _igvn);
1049 }
1050 } else {
1051 assert(ac->is_arraycopy_validated() ||
1052 ac->is_copyof_validated() ||
1053 ac->is_copyofrange_validated(), "unsupported");
1054 CallProjections* callprojs = ac->extract_projections(true);
1055
1056 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1057 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1058 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1059
1060 // Set control to top. IGVN will remove the remaining projections
1061 ac->set_req(0, top());
1062 ac->replace_edge(res, top(), &_igvn);
1063
1064 // Disconnect src right away: it can help find new
1065 // opportunities for allocation elimination
1066 Node* src = ac->in(ArrayCopyNode::Src);
1067 ac->replace_edge(src, top(), &_igvn);
1068 // src can be top at this point if src and dest of the
1069 // arraycopy were the same
1070 if (src->outcnt() == 0 && !src->is_top()) {
1071 _igvn.remove_dead_node(src);
1072 }
1073 }
1074 _igvn._worklist.push(ac);
1075 } else if (use->is_InlineType()) {
1076 assert(use->as_InlineType()->get_oop() == res, "unexpected inline type ptr use");
1077 // Cut off oop input and remove known instance id from type
1078 _igvn.rehash_node_delayed(use);
1079 use->as_InlineType()->set_oop(_igvn.zerocon(T_PRIMITIVE_OBJECT));
1080 const TypeOopPtr* toop = _igvn.type(use)->is_oopptr()->cast_to_instance_id(TypeOopPtr::InstanceBot);
1081 _igvn.set_type(use, toop);
1082 use->as_InlineType()->set_type(toop);
1083 // Process users
1084 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
1085 Node* u = use->fast_out(k);
1086 if (!u->is_InlineType()) {
1087 worklist.push(u);
1088 }
1089 }
1090 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
1091 // Store to mark word of inline type larval buffer
1092 assert(inline_alloc, "Unexpected store to mark word");
1093 _igvn.replace_node(use, use->in(MemNode::Memory));
1094 } else {
1095 eliminate_gc_barrier(use);
1096 }
1097 j -= (oc1 - res->outcnt());
1098 }
1099 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1100 _igvn.remove_dead_node(res);
1101 }
1102
1103 //
1104 // Process other users of allocation's projections
1105 //
1106 if (_callprojs->resproj[0] != nullptr && _callprojs->resproj[0]->outcnt() != 0) {
1107 // First disconnect stores captured by Initialize node.
1108 // If Initialize node is eliminated first in the following code,
1109 // it will kill such stores and DUIterator_Last will assert.
1110 for (DUIterator_Fast jmax, j = _callprojs->resproj[0]->fast_outs(jmax); j < jmax; j++) {
1111 Node* use = _callprojs->resproj[0]->fast_out(j);
1112 if (use->is_AddP()) {
1113 // raw memory addresses used only by the initialization
1114 _igvn.replace_node(use, C->top());
1115 --j; --jmax;
1116 }
1117 }
1118 for (DUIterator_Last jmin, j = _callprojs->resproj[0]->last_outs(jmin); j >= jmin; ) {
1119 Node* use = _callprojs->resproj[0]->last_out(j);
1120 uint oc1 = _callprojs->resproj[0]->outcnt();
1121 if (use->is_Initialize()) {
1122 // Eliminate Initialize node.
1123 InitializeNode *init = use->as_Initialize();
1124 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1125 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1126 if (ctrl_proj != nullptr) {
1127 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1128 #ifdef ASSERT
1129 // If the InitializeNode has no memory out, it will die, and tmp will become null
1130 Node* tmp = init->in(TypeFunc::Control);
1131 assert(tmp == nullptr || tmp == _callprojs->fallthrough_catchproj, "allocation control projection");
1132 #endif
1133 }
1134 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1135 if (mem_proj != nullptr) {
1136 Node *mem = init->in(TypeFunc::Memory);
1137 #ifdef ASSERT
1138 if (mem->is_MergeMem()) {
1139 assert(mem->in(TypeFunc::Memory) == _callprojs->fallthrough_memproj, "allocation memory projection");
1140 } else {
1141 assert(mem == _callprojs->fallthrough_memproj, "allocation memory projection");
1142 }
1143 #endif
1144 _igvn.replace_node(mem_proj, mem);
1145 }
1146 } else if (use->Opcode() == Op_MemBarStoreStore) {
1147 // Inline type buffer allocations are followed by a membar
1148 assert(inline_alloc, "Unexpected MemBarStoreStore");
1149 use->as_MemBar()->remove(&_igvn);
1150 } else {
1151 assert(false, "only Initialize or AddP expected");
1152 }
1153 j -= (oc1 - _callprojs->resproj[0]->outcnt());
1154 }
1155 }
1156 if (_callprojs->fallthrough_catchproj != nullptr) {
1157 _igvn.replace_node(_callprojs->fallthrough_catchproj, alloc->in(TypeFunc::Control));
1158 }
1159 if (_callprojs->fallthrough_memproj != nullptr) {
1160 _igvn.replace_node(_callprojs->fallthrough_memproj, alloc->in(TypeFunc::Memory));
1161 }
1162 if (_callprojs->catchall_memproj != nullptr) {
1163 _igvn.replace_node(_callprojs->catchall_memproj, C->top());
1164 }
1165 if (_callprojs->fallthrough_ioproj != nullptr) {
1166 _igvn.replace_node(_callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1167 }
1168 if (_callprojs->catchall_ioproj != nullptr) {
1169 _igvn.replace_node(_callprojs->catchall_ioproj, C->top());
1170 }
1171 if (_callprojs->catchall_catchproj != nullptr) {
1172 _igvn.replace_node(_callprojs->catchall_catchproj, C->top());
1173 }
1174 }
1175
1176 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1177 // If reallocation fails during deoptimization we'll pop all
1178 // interpreter frames for this compiled frame and that won't play
1179 // nice with JVMTI popframe.
1180 // We avoid this issue by eager reallocation when the popframe request
1181 // is received.
1182 if (!EliminateAllocations) {
1183 return false;
1184 }
1185 Node* klass = alloc->in(AllocateNode::KlassNode);
1186 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1187
1188 // Attempt to eliminate inline type buffer allocations
1189 // regardless of usage and escape/replaceable status.
1190 bool inline_alloc = tklass->isa_instklassptr() &&
1191 tklass->is_instklassptr()->instance_klass()->is_inlinetype();
1192 if (!alloc->_is_non_escaping && !inline_alloc) {
1193 return false;
1194 }
1195 // Eliminate boxing allocations which are not used
1196 // regardless scalar replaceable status.
1197 Node* res = alloc->result_cast();
1198 bool boxing_alloc = (res == nullptr) && C->eliminate_boxing() &&
1199 tklass->isa_instklassptr() &&
1200 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1201 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != nullptr))) {
1202 return false;
1203 }
1204
1205 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1206
1207 GrowableArray <SafePointNode *> safepoints;
1208 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1209 return false;
1210 }
1211
1212 if (!alloc->_is_scalar_replaceable) {
1213 assert(res == nullptr || inline_alloc, "sanity");
1214 // We can only eliminate allocation if all debug info references
1215 // are already replaced with SafePointScalarObject because
1216 // we can't search for a fields value without instance_id.
1217 if (safepoints.length() > 0) {
1218 assert(!inline_alloc, "Inline type allocations should not have safepoint uses");
1219 return false;
1220 }
1221 }
1222
1223 if (!scalar_replacement(alloc, safepoints)) {
1224 return false;
1225 }
1226
1227 CompileLog* log = C->log();
1228 if (log != nullptr) {
1229 log->head("eliminate_allocation type='%d'",
1230 log->identify(tklass->exact_klass()));
1231 JVMState* p = alloc->jvms();
1232 while (p != nullptr) {
1233 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1234 p = p->caller();
1235 }
1236 log->tail("eliminate_allocation");
1237 }
1238
1239 process_users_of_allocation(alloc, inline_alloc);
1240
1241 #ifndef PRODUCT
1242 if (PrintEliminateAllocations) {
1243 if (alloc->is_AllocateArray())
1244 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1245 else
1246 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1247 }
1248 #endif
1249
1250 return true;
1251 }
1252
1253 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1254 // EA should remove all uses of non-escaping boxing node.
1255 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1256 return false;
1257 }
1258
1259 assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1260
1261 _callprojs = boxing->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1262
1263 const TypeTuple* r = boxing->tf()->range_sig();
1264 assert(r->cnt() > TypeFunc::Parms, "sanity");
1265 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1266 assert(t != nullptr, "sanity");
1267
1268 CompileLog* log = C->log();
1269 if (log != nullptr) {
1270 log->head("eliminate_boxing type='%d'",
1271 log->identify(t->instance_klass()));
1272 JVMState* p = boxing->jvms();
1273 while (p != nullptr) {
1274 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1275 p = p->caller();
1276 }
1277 log->tail("eliminate_boxing");
1278 }
1279
1280 process_users_of_allocation(boxing);
1281
1282 #ifndef PRODUCT
1283 if (PrintEliminateAllocations) {
1427 }
1428 }
1429 #endif
1430 yank_alloc_node(alloc);
1431 return;
1432 }
1433 }
1434
1435 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1436 Node *slow_region = nullptr;
1437 Node *toobig_false = ctrl;
1438
1439 // generate the initial test if necessary
1440 if (initial_slow_test != nullptr ) {
1441 assert (expand_fast_path, "Only need test if there is a fast path");
1442 slow_region = new RegionNode(3);
1443
1444 // Now make the initial failure test. Usually a too-big test but
1445 // might be a TRUE for finalizers or a fancy class check for
1446 // newInstance0.
1447 IfNode* toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1448 transform_later(toobig_iff);
1449 // Plug the failing-too-big test into the slow-path region
1450 Node* toobig_true = new IfTrueNode(toobig_iff);
1451 transform_later(toobig_true);
1452 slow_region ->init_req( too_big_or_final_path, toobig_true );
1453 toobig_false = new IfFalseNode(toobig_iff);
1454 transform_later(toobig_false);
1455 } else {
1456 // No initial test, just fall into next case
1457 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1458 toobig_false = ctrl;
1459 debug_only(slow_region = NodeSentinel);
1460 }
1461
1462 // If we are here there are several possibilities
1463 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1464 // no_initial_check means a constant allocation.
1465 // - If check always evaluates to false -> expand_fast_path is false (see above)
1466 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1467 // if !allocation_has_use the fast path is empty
1468 // if !allocation_has_use && no_initial_check
1469 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1470 // removed by yank_alloc_node above.
1471
1472 Node *slow_mem = mem; // save the current memory state for slow path
1473 // generate the fast allocation code unless we know that the initial test will always go slow
1474 if (expand_fast_path) {
1475 // Fast path modifies only raw memory.
1476 if (mem->is_MergeMem()) {
1477 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1478 }
1479
1480 // allocate the Region and Phi nodes for the result
1481 result_region = new RegionNode(3);
1482 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1483 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1484
1485 // Grab regular I/O before optional prefetch may change it.
1486 // Slow-path does no I/O so just set it to the original I/O.
1487 result_phi_i_o->init_req(slow_result_path, i_o);
1488
1489 // Name successful fast-path variables
1490 Node* fast_oop_ctrl;
1491 Node* fast_oop_rawmem;
1492
1493 if (allocation_has_use) {
1494 Node* needgc_ctrl = nullptr;
1495 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1496
1497 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1498 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1499 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1500 fast_oop_ctrl, fast_oop_rawmem,
1501 prefetch_lines);
1502
1503 if (initial_slow_test != nullptr) {
1504 // This completes all paths into the slow merge point
1505 slow_region->init_req(need_gc_path, needgc_ctrl);
1506 transform_later(slow_region);
1507 } else {
1508 // No initial slow path needed!
1509 // Just fall from the need-GC path straight into the VM call.
1510 slow_region = needgc_ctrl;
1511 }
1512
1530 result_phi_i_o ->init_req(fast_result_path, i_o);
1531 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1532 } else {
1533 slow_region = ctrl;
1534 result_phi_i_o = i_o; // Rename it to use in the following code.
1535 }
1536
1537 // Generate slow-path call
1538 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1539 OptoRuntime::stub_name(slow_call_address),
1540 TypePtr::BOTTOM);
1541 call->init_req(TypeFunc::Control, slow_region);
1542 call->init_req(TypeFunc::I_O, top()); // does no i/o
1543 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1544 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1545 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1546
1547 call->init_req(TypeFunc::Parms+0, klass_node);
1548 if (length != nullptr) {
1549 call->init_req(TypeFunc::Parms+1, length);
1550 } else {
1551 // Let the runtime know if this is a larval allocation
1552 call->init_req(TypeFunc::Parms+1, _igvn.intcon(alloc->_larval));
1553 }
1554
1555 // Copy debug information and adjust JVMState information, then replace
1556 // allocate node with the call
1557 call->copy_call_debug_info(&_igvn, alloc);
1558 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1559 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1560 // path dies).
1561 if (valid_length_test != nullptr) {
1562 call->add_req(valid_length_test);
1563 }
1564 if (expand_fast_path) {
1565 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1566 } else {
1567 // Hook i_o projection to avoid its elimination during allocation
1568 // replacement (when only a slow call is generated).
1569 call->set_req(TypeFunc::I_O, result_phi_i_o);
1570 }
1571 _igvn.replace_node(alloc, call);
1572 transform_later(call);
1573
1574 // Identify the output projections from the allocate node and
1575 // adjust any references to them.
1576 // The control and io projections look like:
1577 //
1578 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1579 // Allocate Catch
1580 // ^---Proj(io) <-------+ ^---CatchProj(io)
1581 //
1582 // We are interested in the CatchProj nodes.
1583 //
1584 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1585
1586 // An allocate node has separate memory projections for the uses on
1587 // the control and i_o paths. Replace the control memory projection with
1588 // result_phi_rawmem (unless we are only generating a slow call when
1589 // both memory projections are combined)
1590 if (expand_fast_path && _callprojs->fallthrough_memproj != nullptr) {
1591 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, result_phi_rawmem);
1592 }
1593 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1594 // catchall_memproj so we end up with a call that has only 1 memory projection.
1595 if (_callprojs->catchall_memproj != nullptr) {
1596 if (_callprojs->fallthrough_memproj == nullptr) {
1597 _callprojs->fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1598 transform_later(_callprojs->fallthrough_memproj);
1599 }
1600 _igvn.replace_in_uses(_callprojs->catchall_memproj, _callprojs->fallthrough_memproj);
1601 _igvn.remove_dead_node(_callprojs->catchall_memproj);
1602 }
1603
1604 // An allocate node has separate i_o projections for the uses on the control
1605 // and i_o paths. Always replace the control i_o projection with result i_o
1606 // otherwise incoming i_o become dead when only a slow call is generated
1607 // (it is different from memory projections where both projections are
1608 // combined in such case).
1609 if (_callprojs->fallthrough_ioproj != nullptr) {
1610 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, result_phi_i_o);
1611 }
1612 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1613 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1614 if (_callprojs->catchall_ioproj != nullptr) {
1615 if (_callprojs->fallthrough_ioproj == nullptr) {
1616 _callprojs->fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1617 transform_later(_callprojs->fallthrough_ioproj);
1618 }
1619 _igvn.replace_in_uses(_callprojs->catchall_ioproj, _callprojs->fallthrough_ioproj);
1620 _igvn.remove_dead_node(_callprojs->catchall_ioproj);
1621 }
1622
1623 // if we generated only a slow call, we are done
1624 if (!expand_fast_path) {
1625 // Now we can unhook i_o.
1626 if (result_phi_i_o->outcnt() > 1) {
1627 call->set_req(TypeFunc::I_O, top());
1628 } else {
1629 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1630 // Case of new array with negative size known during compilation.
1631 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1632 // following code since call to runtime will throw exception.
1633 // As result there will be no users of i_o after the call.
1634 // Leave i_o attached to this call to avoid problems in preceding graph.
1635 }
1636 return;
1637 }
1638
1639 if (_callprojs->fallthrough_catchproj != nullptr) {
1640 ctrl = _callprojs->fallthrough_catchproj->clone();
1641 transform_later(ctrl);
1642 _igvn.replace_node(_callprojs->fallthrough_catchproj, result_region);
1643 } else {
1644 ctrl = top();
1645 }
1646 Node *slow_result;
1647 if (_callprojs->resproj[0] == nullptr) {
1648 // no uses of the allocation result
1649 slow_result = top();
1650 } else {
1651 slow_result = _callprojs->resproj[0]->clone();
1652 transform_later(slow_result);
1653 _igvn.replace_node(_callprojs->resproj[0], result_phi_rawoop);
1654 }
1655
1656 // Plug slow-path into result merge point
1657 result_region->init_req( slow_result_path, ctrl);
1658 transform_later(result_region);
1659 if (allocation_has_use) {
1660 result_phi_rawoop->init_req(slow_result_path, slow_result);
1661 transform_later(result_phi_rawoop);
1662 }
1663 result_phi_rawmem->init_req(slow_result_path, _callprojs->fallthrough_memproj);
1664 transform_later(result_phi_rawmem);
1665 transform_later(result_phi_i_o);
1666 // This completes all paths into the result merge point
1667 }
1668
1669 // Remove alloc node that has no uses.
1670 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1671 Node* ctrl = alloc->in(TypeFunc::Control);
1672 Node* mem = alloc->in(TypeFunc::Memory);
1673 Node* i_o = alloc->in(TypeFunc::I_O);
1674
1675 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1676 if (_callprojs->resproj[0] != nullptr) {
1677 for (DUIterator_Fast imax, i = _callprojs->resproj[0]->fast_outs(imax); i < imax; i++) {
1678 Node* use = _callprojs->resproj[0]->fast_out(i);
1679 use->isa_MemBar()->remove(&_igvn);
1680 --imax;
1681 --i; // back up iterator
1682 }
1683 assert(_callprojs->resproj[0]->outcnt() == 0, "all uses must be deleted");
1684 _igvn.remove_dead_node(_callprojs->resproj[0]);
1685 }
1686 if (_callprojs->fallthrough_catchproj != nullptr) {
1687 _igvn.replace_in_uses(_callprojs->fallthrough_catchproj, ctrl);
1688 _igvn.remove_dead_node(_callprojs->fallthrough_catchproj);
1689 }
1690 if (_callprojs->catchall_catchproj != nullptr) {
1691 _igvn.rehash_node_delayed(_callprojs->catchall_catchproj);
1692 _callprojs->catchall_catchproj->set_req(0, top());
1693 }
1694 if (_callprojs->fallthrough_proj != nullptr) {
1695 Node* catchnode = _callprojs->fallthrough_proj->unique_ctrl_out();
1696 _igvn.remove_dead_node(catchnode);
1697 _igvn.remove_dead_node(_callprojs->fallthrough_proj);
1698 }
1699 if (_callprojs->fallthrough_memproj != nullptr) {
1700 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, mem);
1701 _igvn.remove_dead_node(_callprojs->fallthrough_memproj);
1702 }
1703 if (_callprojs->fallthrough_ioproj != nullptr) {
1704 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, i_o);
1705 _igvn.remove_dead_node(_callprojs->fallthrough_ioproj);
1706 }
1707 if (_callprojs->catchall_memproj != nullptr) {
1708 _igvn.rehash_node_delayed(_callprojs->catchall_memproj);
1709 _callprojs->catchall_memproj->set_req(0, top());
1710 }
1711 if (_callprojs->catchall_ioproj != nullptr) {
1712 _igvn.rehash_node_delayed(_callprojs->catchall_ioproj);
1713 _callprojs->catchall_ioproj->set_req(0, top());
1714 }
1715 #ifndef PRODUCT
1716 if (PrintEliminateAllocations) {
1717 if (alloc->is_AllocateArray()) {
1718 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1719 } else {
1720 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1721 }
1722 }
1723 #endif
1724 _igvn.remove_dead_node(alloc);
1725 }
1726
1727 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1728 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1729 // If initialization is performed by an array copy, any required
1730 // MemBarStoreStore was already added. If the object does not
1731 // escape no need for a MemBarStoreStore. If the object does not
1732 // escape in its initializer and memory barrier (MemBarStoreStore or
1733 // stronger) is already added at exit of initializer, also no need
1811 Node* thread = new ThreadLocalNode();
1812 transform_later(thread);
1813
1814 call->init_req(TypeFunc::Parms + 0, thread);
1815 call->init_req(TypeFunc::Parms + 1, oop);
1816 call->init_req(TypeFunc::Control, ctrl);
1817 call->init_req(TypeFunc::I_O , top()); // does no i/o
1818 call->init_req(TypeFunc::Memory , rawmem);
1819 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1820 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1821 transform_later(call);
1822 ctrl = new ProjNode(call, TypeFunc::Control);
1823 transform_later(ctrl);
1824 rawmem = new ProjNode(call, TypeFunc::Memory);
1825 transform_later(rawmem);
1826 }
1827 }
1828
1829 // Helper for PhaseMacroExpand::expand_allocate_common.
1830 // Initializes the newly-allocated storage.
1831 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1832 Node* control, Node* rawmem, Node* object,
1833 Node* klass_node, Node* length,
1834 Node* size_in_bytes) {
1835 InitializeNode* init = alloc->initialization();
1836 // Store the klass & mark bits
1837 Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
1838 if (!mark_node->is_Con()) {
1839 transform_later(mark_node);
1840 }
1841 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1842
1843 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1844 int header_size = alloc->minimum_header_size(); // conservatively small
1845
1846 // Array length
1847 if (length != nullptr) { // Arrays need length field
1848 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1849 // conservatively small header size:
1850 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1851 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1852 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1853 if (is_reference_type(elem, true)) {
1854 elem = T_OBJECT;
1855 }
1856 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1857 }
1858 }
1859
1860 // Clear the object body, if necessary.
1861 if (init == nullptr) {
1862 // The init has somehow disappeared; be cautious and clear everything.
1863 //
1864 // This can happen if a node is allocated but an uncommon trap occurs
1865 // immediately. In this case, the Initialize gets associated with the
1866 // trap, and may be placed in a different (outer) loop, if the Allocate
1867 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1868 // there can be two Allocates to one Initialize. The answer in all these
1869 // edge cases is safety first. It is always safe to clear immediately
1870 // within an Allocate, and then (maybe or maybe not) clear some more later.
1871 if (!(UseTLAB && ZeroTLAB)) {
1872 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1873 alloc->in(AllocateNode::DefaultValue),
1874 alloc->in(AllocateNode::RawDefaultValue),
1875 header_size, size_in_bytes,
1876 &_igvn);
1877 }
1878 } else {
1879 if (!init->is_complete()) {
1880 // Try to win by zeroing only what the init does not store.
1881 // We can also try to do some peephole optimizations,
1882 // such as combining some adjacent subword stores.
1883 rawmem = init->complete_stores(control, rawmem, object,
1884 header_size, size_in_bytes, &_igvn);
1885 }
1886 // We have no more use for this link, since the AllocateNode goes away:
1887 init->set_req(InitializeNode::RawAddress, top());
1888 // (If we keep the link, it just confuses the register allocator,
1889 // who thinks he sees a real use of the address by the membar.)
1890 }
1891
1892 return rawmem;
1893 }
1894
2224 } // EliminateNestedLocks
2225
2226 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2227 // Look for all locks of this object and mark them and
2228 // corresponding BoxLock nodes as eliminated.
2229 Node* obj = alock->obj_node();
2230 for (uint j = 0; j < obj->outcnt(); j++) {
2231 Node* o = obj->raw_out(j);
2232 if (o->is_AbstractLock() &&
2233 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2234 alock = o->as_AbstractLock();
2235 Node* box = alock->box_node();
2236 // Replace old box node with new eliminated box for all users
2237 // of the same object and mark related locks as eliminated.
2238 mark_eliminated_box(box, obj);
2239 }
2240 }
2241 }
2242 }
2243
2244 void PhaseMacroExpand::inline_type_guard(Node** ctrl, LockNode* lock) {
2245 Node* obj = lock->obj_node();
2246 const TypePtr* obj_type = _igvn.type(obj)->make_ptr();
2247 if (!obj_type->can_be_inline_type()) {
2248 return;
2249 }
2250 Node* mark = make_load(*ctrl, lock->memory(), obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2251 Node* value_mask = _igvn.MakeConX(markWord::inline_type_pattern);
2252 Node* is_value = _igvn.transform(new AndXNode(mark, value_mask));
2253 Node* cmp = _igvn.transform(new CmpXNode(is_value, value_mask));
2254 Node* bol = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2255 Node* unc_ctrl = generate_slow_guard(ctrl, bol, nullptr);
2256
2257 int trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_class_check, Deoptimization::Action_none);
2258 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2259 const TypePtr* no_memory_effects = nullptr;
2260 CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap",
2261 no_memory_effects);
2262 unc->init_req(TypeFunc::Control, unc_ctrl);
2263 unc->init_req(TypeFunc::I_O, lock->i_o());
2264 unc->init_req(TypeFunc::Memory, lock->memory());
2265 unc->init_req(TypeFunc::FramePtr, lock->in(TypeFunc::FramePtr));
2266 unc->init_req(TypeFunc::ReturnAdr, lock->in(TypeFunc::ReturnAdr));
2267 unc->init_req(TypeFunc::Parms+0, _igvn.intcon(trap_request));
2268 unc->set_cnt(PROB_UNLIKELY_MAG(4));
2269 unc->copy_call_debug_info(&_igvn, lock);
2270
2271 assert(unc->peek_monitor_box() == lock->box_node(), "wrong monitor");
2272 assert((obj_type->is_inlinetypeptr() && unc->peek_monitor_obj()->is_SafePointScalarObject()) ||
2273 (obj->is_InlineType() && obj->in(1) == unc->peek_monitor_obj()) ||
2274 (obj == unc->peek_monitor_obj()), "wrong monitor");
2275
2276 // pop monitor and push obj back on stack: we trap before the monitorenter
2277 unc->pop_monitor();
2278 unc->grow_stack(unc->jvms(), 1);
2279 unc->set_stack(unc->jvms(), unc->jvms()->stk_size()-1, obj);
2280 _igvn.register_new_node_with_optimizer(unc);
2281
2282 unc_ctrl = _igvn.transform(new ProjNode(unc, TypeFunc::Control));
2283 Node* halt = _igvn.transform(new HaltNode(unc_ctrl, lock->in(TypeFunc::FramePtr), "monitor enter on inline type"));
2284 _igvn.add_input_to(C->root(), halt);
2285 }
2286
2287 // we have determined that this lock/unlock can be eliminated, we simply
2288 // eliminate the node without expanding it.
2289 //
2290 // Note: The membar's associated with the lock/unlock are currently not
2291 // eliminated. This should be investigated as a future enhancement.
2292 //
2293 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2294
2295 if (!alock->is_eliminated()) {
2296 return false;
2297 }
2298 #ifdef ASSERT
2299 if (!alock->is_coarsened()) {
2300 // Check that new "eliminated" BoxLock node is created.
2301 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2302 assert(oldbox->is_eliminated(), "should be done already");
2303 }
2304 #endif
2305
2306 alock->log_lock_optimization(C, "eliminate_lock");
2307
2308 #ifndef PRODUCT
2309 if (PrintEliminateLocks) {
2310 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2311 }
2312 #endif
2313
2314 Node* mem = alock->in(TypeFunc::Memory);
2315 Node* ctrl = alock->in(TypeFunc::Control);
2316 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2317
2318 _callprojs = alock->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2319 // There are 2 projections from the lock. The lock node will
2320 // be deleted when its last use is subsumed below.
2321 assert(alock->outcnt() == 2 &&
2322 _callprojs->fallthrough_proj != nullptr &&
2323 _callprojs->fallthrough_memproj != nullptr,
2324 "Unexpected projections from Lock/Unlock");
2325
2326 Node* fallthroughproj = _callprojs->fallthrough_proj;
2327 Node* memproj_fallthrough = _callprojs->fallthrough_memproj;
2328
2329 // The memory projection from a lock/unlock is RawMem
2330 // The input to a Lock is merged memory, so extract its RawMem input
2331 // (unless the MergeMem has been optimized away.)
2332 if (alock->is_Lock()) {
2333 // Deoptimize and re-execute if object is an inline type
2334 inline_type_guard(&ctrl, alock->as_Lock());
2335
2336 // Search for MemBarAcquireLock node and delete it also.
2337 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2338 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2339 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2340 Node* memproj = membar->proj_out(TypeFunc::Memory);
2341 _igvn.replace_node(ctrlproj, fallthroughproj);
2342 _igvn.replace_node(memproj, memproj_fallthrough);
2343
2344 // Delete FastLock node also if this Lock node is unique user
2345 // (a loop peeling may clone a Lock node).
2346 Node* flock = alock->as_Lock()->fastlock_node();
2347 if (flock->outcnt() == 1) {
2348 assert(flock->unique_out() == alock, "sanity");
2349 _igvn.replace_node(flock, top());
2350 }
2351 }
2352
2353 // Search for MemBarReleaseLock node and delete it also.
2354 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2355 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2376 Node* mem = lock->in(TypeFunc::Memory);
2377 Node* obj = lock->obj_node();
2378 Node* box = lock->box_node();
2379 Node* flock = lock->fastlock_node();
2380
2381 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2382
2383 // Make the merge point
2384 Node *region;
2385 Node *mem_phi;
2386 Node *slow_path;
2387
2388 region = new RegionNode(3);
2389 // create a Phi for the memory state
2390 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2391
2392 // Optimize test; set region slot 2
2393 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2394 mem_phi->init_req(2, mem);
2395
2396 // Deoptimize and re-execute if object is an inline type
2397 inline_type_guard(&slow_path, lock);
2398
2399 // Make slow path call
2400 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2401 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2402 obj, box, nullptr);
2403
2404 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2405
2406 // Slow path can only throw asynchronous exceptions, which are always
2407 // de-opted. So the compiler thinks the slow-call can never throw an
2408 // exception. If it DOES throw an exception we would need the debug
2409 // info removed first (since if it throws there is no monitor).
2410 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2411 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2412
2413 // Capture slow path
2414 // disconnect fall-through projection from call and create a new one
2415 // hook up users of fall-through projection to region
2416 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2417 transform_later(slow_ctrl);
2418 _igvn.hash_delete(_callprojs->fallthrough_proj);
2419 _callprojs->fallthrough_proj->disconnect_inputs(C);
2420 region->init_req(1, slow_ctrl);
2421 // region inputs are now complete
2422 transform_later(region);
2423 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2424
2425 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2426
2427 mem_phi->init_req(1, memproj);
2428
2429 transform_later(mem_phi);
2430
2431 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2432 }
2433
2434 //------------------------------expand_unlock_node----------------------
2435 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2436
2437 Node* ctrl = unlock->in(TypeFunc::Control);
2438 Node* mem = unlock->in(TypeFunc::Memory);
2439 Node* obj = unlock->obj_node();
2440 Node* box = unlock->box_node();
2441
2442 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2443
2444 // No need for a null check on unlock
2445
2446 // Make the merge point
2447 Node *region;
2448 Node *mem_phi;
2449
2450 region = new RegionNode(3);
2451 // create a Phi for the memory state
2452 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2453
2454 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2455 funlock = transform_later( funlock )->as_FastUnlock();
2456 // Optimize test; set region slot 2
2457 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2458 Node *thread = transform_later(new ThreadLocalNode());
2459
2460 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2461 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2462 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2463
2464 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2465 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2466 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2467
2468 // No exceptions for unlocking
2469 // Capture slow path
2470 // disconnect fall-through projection from call and create a new one
2471 // hook up users of fall-through projection to region
2472 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2473 transform_later(slow_ctrl);
2474 _igvn.hash_delete(_callprojs->fallthrough_proj);
2475 _callprojs->fallthrough_proj->disconnect_inputs(C);
2476 region->init_req(1, slow_ctrl);
2477 // region inputs are now complete
2478 transform_later(region);
2479 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2480
2481 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2482 mem_phi->init_req(1, memproj );
2483 mem_phi->init_req(2, mem);
2484 transform_later(mem_phi);
2485
2486 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2487 }
2488
2489 // An inline type might be returned from the call but we don't know its
2490 // type. Either we get a buffered inline type (and nothing needs to be done)
2491 // or one of the values being returned is the klass of the inline type
2492 // and we need to allocate an inline type instance of that type and
2493 // initialize it with other values being returned. In that case, we
2494 // first try a fast path allocation and initialize the value with the
2495 // inline klass's pack handler or we fall back to a runtime call.
2496 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2497 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2498 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2499 if (ret == nullptr) {
2500 return;
2501 }
2502 const TypeFunc* tf = call->_tf;
2503 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2504 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2505 call->_tf = new_tf;
2506 // Make sure the change of type is applied before projections are processed by igvn
2507 _igvn.set_type(call, call->Value(&_igvn));
2508 _igvn.set_type(ret, ret->Value(&_igvn));
2509
2510 // Before any new projection is added:
2511 CallProjections* projs = call->extract_projections(true, true);
2512
2513 // Create temporary hook nodes that will be replaced below.
2514 // Add an input to prevent hook nodes from being dead.
2515 Node* ctl = new Node(call);
2516 Node* mem = new Node(ctl);
2517 Node* io = new Node(ctl);
2518 Node* ex_ctl = new Node(ctl);
2519 Node* ex_mem = new Node(ctl);
2520 Node* ex_io = new Node(ctl);
2521 Node* res = new Node(ctl);
2522
2523 // Allocate a new buffered inline type only if a new one is not returned
2524 Node* cast = transform_later(new CastP2XNode(ctl, res));
2525 Node* mask = MakeConX(0x1);
2526 Node* masked = transform_later(new AndXNode(cast, mask));
2527 Node* cmp = transform_later(new CmpXNode(masked, mask));
2528 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2529 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2530 transform_later(allocation_iff);
2531 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2532 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2533 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2534
2535 // Try to allocate a new buffered inline instance either from TLAB or eden space
2536 Node* needgc_ctrl = nullptr; // needgc means slowcase, i.e. allocation failed
2537 CallLeafNoFPNode* handler_call;
2538 const bool alloc_in_place = UseTLAB;
2539 if (alloc_in_place) {
2540 Node* fast_oop_ctrl = nullptr;
2541 Node* fast_oop_rawmem = nullptr;
2542 Node* mask2 = MakeConX(-2);
2543 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2544 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2545 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeInstKlassPtr::OBJECT_OR_NULL));
2546 Node* layout_val = make_load(nullptr, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2547 Node* size_in_bytes = ConvI2X(layout_val);
2548 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2549 Node* fast_oop = bs->obj_allocate(this, mem, allocation_ctl, size_in_bytes, io, needgc_ctrl,
2550 fast_oop_ctrl, fast_oop_rawmem,
2551 AllocateInstancePrefetchLines);
2552 // Allocation succeed, initialize buffered inline instance header firstly,
2553 // and then initialize its fields with an inline class specific handler
2554 Node* mark_node = makecon(TypeRawPtr::make((address)markWord::inline_type_prototype().value()));
2555 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2556 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2557 if (UseCompressedClassPointers) {
2558 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2559 }
2560 Node* fixed_block = make_load(fast_oop_ctrl, fast_oop_rawmem, klass_node, in_bytes(InstanceKlass::adr_inlineklass_fixed_block_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2561 Node* pack_handler = make_load(fast_oop_ctrl, fast_oop_rawmem, fixed_block, in_bytes(InlineKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2562 handler_call = new CallLeafNoFPNode(OptoRuntime::pack_inline_type_Type(),
2563 nullptr,
2564 "pack handler",
2565 TypeRawPtr::BOTTOM);
2566 handler_call->init_req(TypeFunc::Control, fast_oop_ctrl);
2567 handler_call->init_req(TypeFunc::Memory, fast_oop_rawmem);
2568 handler_call->init_req(TypeFunc::I_O, top());
2569 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2570 handler_call->init_req(TypeFunc::ReturnAdr, top());
2571 handler_call->init_req(TypeFunc::Parms, pack_handler);
2572 handler_call->init_req(TypeFunc::Parms+1, fast_oop);
2573 } else {
2574 needgc_ctrl = allocation_ctl;
2575 }
2576
2577 // Allocation failed, fall back to a runtime call
2578 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_inline_type_fields_Type(),
2579 StubRoutines::store_inline_type_fields_to_buf(),
2580 "store_inline_type_fields",
2581 TypePtr::BOTTOM);
2582 slow_call->init_req(TypeFunc::Control, needgc_ctrl);
2583 slow_call->init_req(TypeFunc::Memory, mem);
2584 slow_call->init_req(TypeFunc::I_O, io);
2585 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2586 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2587 slow_call->init_req(TypeFunc::Parms, res);
2588
2589 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2590 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2591 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2592 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2593 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2594 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2595 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2596
2597 Node* ex_r = new RegionNode(3);
2598 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2599 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2600 ex_r->init_req(1, slow_excp);
2601 ex_mem_phi->init_req(1, slow_mem);
2602 ex_io_phi->init_req(1, slow_io);
2603 ex_r->init_req(2, ex_ctl);
2604 ex_mem_phi->init_req(2, ex_mem);
2605 ex_io_phi->init_req(2, ex_io);
2606 transform_later(ex_r);
2607 transform_later(ex_mem_phi);
2608 transform_later(ex_io_phi);
2609
2610 // We don't know how many values are returned. This assumes the
2611 // worst case, that all available registers are used.
2612 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2613 if (domain->field_at(i) == Type::HALF) {
2614 slow_call->init_req(i, top());
2615 if (alloc_in_place) {
2616 handler_call->init_req(i+1, top());
2617 }
2618 continue;
2619 }
2620 Node* proj = transform_later(new ProjNode(call, i));
2621 slow_call->init_req(i, proj);
2622 if (alloc_in_place) {
2623 handler_call->init_req(i+1, proj);
2624 }
2625 }
2626 // We can safepoint at that new call
2627 slow_call->copy_call_debug_info(&_igvn, call);
2628 transform_later(slow_call);
2629 if (alloc_in_place) {
2630 transform_later(handler_call);
2631 }
2632
2633 Node* fast_ctl = nullptr;
2634 Node* fast_res = nullptr;
2635 MergeMemNode* fast_mem = nullptr;
2636 if (alloc_in_place) {
2637 fast_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2638 Node* rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2639 fast_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2640 fast_mem = MergeMemNode::make(mem);
2641 fast_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2642 transform_later(fast_mem);
2643 }
2644
2645 Node* r = new RegionNode(alloc_in_place ? 4 : 3);
2646 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2647 Node* io_phi = new PhiNode(r, Type::ABIO);
2648 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2649 r->init_req(1, no_allocation_ctl);
2650 mem_phi->init_req(1, mem);
2651 io_phi->init_req(1, io);
2652 res_phi->init_req(1, no_allocation_res);
2653 r->init_req(2, slow_norm);
2654 mem_phi->init_req(2, slow_mem);
2655 io_phi->init_req(2, slow_io);
2656 res_phi->init_req(2, slow_res);
2657 if (alloc_in_place) {
2658 r->init_req(3, fast_ctl);
2659 mem_phi->init_req(3, fast_mem);
2660 io_phi->init_req(3, io);
2661 res_phi->init_req(3, fast_res);
2662 }
2663 transform_later(r);
2664 transform_later(mem_phi);
2665 transform_later(io_phi);
2666 transform_later(res_phi);
2667
2668 // Do not let stores that initialize this buffer be reordered with a subsequent
2669 // store that would make this buffer accessible by other threads.
2670 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
2671 transform_later(mb);
2672 mb->init_req(TypeFunc::Memory, mem_phi);
2673 mb->init_req(TypeFunc::Control, r);
2674 r = new ProjNode(mb, TypeFunc::Control);
2675 transform_later(r);
2676 mem_phi = new ProjNode(mb, TypeFunc::Memory);
2677 transform_later(mem_phi);
2678
2679 assert(projs->nb_resproj == 1, "unexpected number of results");
2680 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2681 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2682 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2683 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2684 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2685 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2686 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2687 // The CatchNode should not use the ex_io_phi. Re-connect it to the catchall_ioproj.
2688 Node* cn = projs->fallthrough_catchproj->in(0);
2689 _igvn.replace_input_of(cn, 1, projs->catchall_ioproj);
2690
2691 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2692 _igvn.replace_node(mem, projs->fallthrough_memproj);
2693 _igvn.replace_node(io, projs->fallthrough_ioproj);
2694 _igvn.replace_node(res, projs->resproj[0]);
2695 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2696 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2697 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2698 }
2699
2700 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2701 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2702 Node* bol = check->unique_out();
2703 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2704 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2705 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2706
2707 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2708 Node* iff = bol->last_out(i);
2709 assert(iff->is_If(), "where's the if?");
2710
2711 if (iff->in(0)->is_top()) {
2712 _igvn.replace_input_of(iff, 1, C->top());
2713 continue;
2714 }
2715
2716 Node* iftrue = iff->as_If()->proj_out(1);
2717 Node* iffalse = iff->as_If()->proj_out(0);
2718 Node* ctrl = iff->in(0);
2719
2720 Node* subklass = nullptr;
2721 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2722 subklass = obj_or_subklass;
2723 } else {
2724 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2725 subklass = _igvn.transform(LoadKlassNode::make(_igvn, nullptr, C->immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2726 }
2727
2728 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn);
2729
2730 _igvn.replace_input_of(iff, 0, C->top());
2731 _igvn.replace_node(iftrue, not_subtype_ctrl);
2732 _igvn.replace_node(iffalse, ctrl);
2733 }
2734 _igvn.replace_node(check, C->top());
2735 }
2736
2737 // FlatArrayCheckNode (array1 array2 ...) is expanded into:
2738 //
2739 // long mark = array1.mark | array2.mark | ...;
2740 // long locked_bit = markWord::unlocked_value & array1.mark & array2.mark & ...;
2741 // if (locked_bit == 0) {
2742 // // One array is locked, load prototype header from the klass
2743 // mark = array1.klass.proto | array2.klass.proto | ...
2744 // }
2745 // if ((mark & markWord::flat_array_bit_in_place) == 0) {
2746 // ...
2747 // }
2748 void PhaseMacroExpand::expand_flatarraycheck_node(FlatArrayCheckNode* check) {
2749 bool array_inputs = _igvn.type(check->in(FlatArrayCheckNode::ArrayOrKlass))->isa_oopptr() != nullptr;
2750 if (UseArrayMarkWordCheck && array_inputs) {
2751 Node* mark = MakeConX(0);
2752 Node* locked_bit = MakeConX(markWord::unlocked_value);
2753 Node* mem = check->in(FlatArrayCheckNode::Memory);
2754 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2755 Node* ary = check->in(i);
2756 const TypeOopPtr* t = _igvn.type(ary)->isa_oopptr();
2757 assert(t != nullptr, "Mixing array and klass inputs");
2758 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2759 Node* mark_adr = basic_plus_adr(ary, oopDesc::mark_offset_in_bytes());
2760 Node* mark_load = _igvn.transform(LoadNode::make(_igvn, nullptr, mem, mark_adr, mark_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
2761 mark = _igvn.transform(new OrXNode(mark, mark_load));
2762 locked_bit = _igvn.transform(new AndXNode(locked_bit, mark_load));
2763 }
2764 assert(!mark->is_Con(), "Should have been optimized out");
2765 Node* cmp = _igvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
2766 Node* is_unlocked = _igvn.transform(new BoolNode(cmp, BoolTest::ne));
2767
2768 // BoolNode might be shared, replace each if user
2769 Node* old_bol = check->unique_out();
2770 assert(old_bol->is_Bool() && old_bol->as_Bool()->_test._test == BoolTest::ne, "unexpected condition");
2771 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2772 IfNode* old_iff = old_bol->last_out(i)->as_If();
2773 Node* ctrl = old_iff->in(0);
2774 RegionNode* region = new RegionNode(3);
2775 Node* mark_phi = new PhiNode(region, TypeX_X);
2776
2777 // Check if array is unlocked
2778 IfNode* iff = _igvn.transform(new IfNode(ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
2779
2780 // Unlocked: Use bits from mark word
2781 region->init_req(1, _igvn.transform(new IfTrueNode(iff)));
2782 mark_phi->init_req(1, mark);
2783
2784 // Locked: Load prototype header from klass
2785 ctrl = _igvn.transform(new IfFalseNode(iff));
2786 Node* proto = MakeConX(0);
2787 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2788 Node* ary = check->in(i);
2789 // Make loads control dependent to make sure they are only executed if array is locked
2790 Node* klass_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
2791 Node* klass = _igvn.transform(LoadKlassNode::make(_igvn, ctrl, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2792 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
2793 Node* proto_load = _igvn.transform(LoadNode::make(_igvn, ctrl, C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
2794 proto = _igvn.transform(new OrXNode(proto, proto_load));
2795 }
2796 region->init_req(2, ctrl);
2797 mark_phi->init_req(2, proto);
2798
2799 // Check if flat array bits are set
2800 Node* mask = MakeConX(markWord::flat_array_bit_in_place);
2801 Node* masked = _igvn.transform(new AndXNode(_igvn.transform(mark_phi), mask));
2802 cmp = _igvn.transform(new CmpXNode(masked, MakeConX(0)));
2803 Node* is_not_flat = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2804
2805 ctrl = _igvn.transform(region);
2806 iff = _igvn.transform(new IfNode(ctrl, is_not_flat, PROB_MAX, COUNT_UNKNOWN))->as_If();
2807 _igvn.replace_node(old_iff, iff);
2808 }
2809 _igvn.replace_node(check, C->top());
2810 } else {
2811 // Fall back to layout helper check
2812 Node* lhs = intcon(0);
2813 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2814 Node* array_or_klass = check->in(i);
2815 Node* klass = nullptr;
2816 const TypePtr* t = _igvn.type(array_or_klass)->is_ptr();
2817 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2818 if (t->isa_oopptr() != nullptr) {
2819 Node* klass_adr = basic_plus_adr(array_or_klass, oopDesc::klass_offset_in_bytes());
2820 klass = transform_later(LoadKlassNode::make(_igvn, nullptr, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2821 } else {
2822 assert(t->isa_aryklassptr(), "Unexpected input type");
2823 klass = array_or_klass;
2824 }
2825 Node* lh_addr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
2826 Node* lh_val = _igvn.transform(LoadNode::make(_igvn, nullptr, C->immutable_memory(), lh_addr, lh_addr->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
2827 lhs = _igvn.transform(new OrINode(lhs, lh_val));
2828 }
2829 Node* masked = transform_later(new AndINode(lhs, intcon(Klass::_lh_array_tag_flat_value_bit_inplace)));
2830 Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
2831 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2832 Node* old_bol = check->unique_out();
2833 _igvn.replace_node(old_bol, bol);
2834 _igvn.replace_node(check, C->top());
2835 }
2836 }
2837
2838 //---------------------------eliminate_macro_nodes----------------------
2839 // Eliminate scalar replaced allocations and associated locks.
2840 void PhaseMacroExpand::eliminate_macro_nodes() {
2841 if (C->macro_count() == 0)
2842 return;
2843 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2844
2845 // Before elimination may re-mark (change to Nested or NonEscObj)
2846 // all associated (same box and obj) lock and unlock nodes.
2847 int cnt = C->macro_count();
2848 for (int i=0; i < cnt; i++) {
2849 Node *n = C->macro_node(i);
2850 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2851 mark_eliminated_locking_nodes(n->as_AbstractLock());
2852 }
2853 }
2854 // Re-marking may break consistency of Coarsened locks.
2855 if (!C->coarsened_locks_consistent()) {
2856 return; // recompile without Coarsened locks if broken
2857 }
2878 }
2879 // Next, attempt to eliminate allocations
2880 _has_locks = false;
2881 progress = true;
2882 while (progress) {
2883 progress = false;
2884 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2885 Node* n = C->macro_node(i - 1);
2886 bool success = false;
2887 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2888 switch (n->class_id()) {
2889 case Node::Class_Allocate:
2890 case Node::Class_AllocateArray:
2891 success = eliminate_allocate_node(n->as_Allocate());
2892 #ifndef PRODUCT
2893 if (success && PrintOptoStatistics) {
2894 Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
2895 }
2896 #endif
2897 break;
2898 case Node::Class_CallStaticJava: {
2899 CallStaticJavaNode* call = n->as_CallStaticJava();
2900 if (!call->method()->is_method_handle_intrinsic()) {
2901 success = eliminate_boxing_node(n->as_CallStaticJava());
2902 }
2903 break;
2904 }
2905 case Node::Class_Lock:
2906 case Node::Class_Unlock:
2907 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2908 _has_locks = true;
2909 break;
2910 case Node::Class_ArrayCopy:
2911 break;
2912 case Node::Class_OuterStripMinedLoop:
2913 break;
2914 case Node::Class_SubTypeCheck:
2915 break;
2916 case Node::Class_Opaque1:
2917 break;
2918 case Node::Class_FlatArrayCheck:
2919 break;
2920 default:
2921 assert(n->Opcode() == Op_LoopLimit ||
2922 n->Opcode() == Op_Opaque3 ||
2923 n->Opcode() == Op_Opaque4 ||
2924 n->Opcode() == Op_MaxL ||
2925 n->Opcode() == Op_MinL ||
2926 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2927 "unknown node type in macro list");
2928 }
2929 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2930 progress = progress || success;
2931 }
2932 }
2933 #ifndef PRODUCT
2934 if (PrintOptoStatistics) {
2935 int membar_after = count_MemBar(C);
2936 Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
2937 }
2938 #endif
2939 }
2942 // Returns true if a failure occurred.
2943 bool PhaseMacroExpand::expand_macro_nodes() {
2944 // Last attempt to eliminate macro nodes.
2945 eliminate_macro_nodes();
2946 if (C->failing()) return true;
2947
2948 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2949 bool progress = true;
2950 while (progress) {
2951 progress = false;
2952 for (int i = C->macro_count(); i > 0; i--) {
2953 Node* n = C->macro_node(i-1);
2954 bool success = false;
2955 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2956 if (n->Opcode() == Op_LoopLimit) {
2957 // Remove it from macro list and put on IGVN worklist to optimize.
2958 C->remove_macro_node(n);
2959 _igvn._worklist.push(n);
2960 success = true;
2961 } else if (n->Opcode() == Op_CallStaticJava) {
2962 CallStaticJavaNode* call = n->as_CallStaticJava();
2963 if (!call->method()->is_method_handle_intrinsic()) {
2964 // Remove it from macro list and put on IGVN worklist to optimize.
2965 C->remove_macro_node(n);
2966 _igvn._worklist.push(n);
2967 success = true;
2968 }
2969 } else if (n->is_Opaque1()) {
2970 _igvn.replace_node(n, n->in(1));
2971 success = true;
2972 #if INCLUDE_RTM_OPT
2973 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2974 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2975 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2976 Node* cmp = n->unique_out();
2977 #ifdef ASSERT
2978 // Validate graph.
2979 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2980 BoolNode* bol = cmp->unique_out()->as_Bool();
2981 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2982 (bol->_test._test == BoolTest::ne), "");
2983 IfNode* ifn = bol->unique_out()->as_If();
2984 assert((ifn->outcnt() == 2) &&
2985 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != nullptr, "");
2986 #endif
2987 Node* repl = n->in(1);
2988 if (!_has_locks) {
3063 // Worst case is a macro node gets expanded into about 200 nodes.
3064 // Allow 50% more for optimization.
3065 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3066 return true;
3067 }
3068
3069 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3070 switch (n->class_id()) {
3071 case Node::Class_Lock:
3072 expand_lock_node(n->as_Lock());
3073 break;
3074 case Node::Class_Unlock:
3075 expand_unlock_node(n->as_Unlock());
3076 break;
3077 case Node::Class_ArrayCopy:
3078 expand_arraycopy_node(n->as_ArrayCopy());
3079 break;
3080 case Node::Class_SubTypeCheck:
3081 expand_subtypecheck_node(n->as_SubTypeCheck());
3082 break;
3083 case Node::Class_CallStaticJava:
3084 expand_mh_intrinsic_return(n->as_CallStaticJava());
3085 C->remove_macro_node(n);
3086 break;
3087 case Node::Class_FlatArrayCheck:
3088 expand_flatarraycheck_node(n->as_FlatArrayCheck());
3089 break;
3090 default:
3091 assert(false, "unknown node type in macro list");
3092 }
3093 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
3094 if (C->failing()) return true;
3095
3096 // Clean up the graph so we're less likely to hit the maximum node
3097 // limit
3098 _igvn.set_delay_transform(false);
3099 _igvn.optimize();
3100 if (C->failing()) return true;
3101 _igvn.set_delay_transform(true);
3102 }
3103
3104 // All nodes except Allocate nodes are expanded now. There could be
3105 // new optimization opportunities (such as folding newly created
3106 // load from a just allocated object). Run IGVN.
3107
3108 // expand "macro" nodes
3109 // nodes are removed from the macro list as they are processed
|