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()) {
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) && !reduce_merge_precheck) {
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";)
611 (use->as_ArrayCopy()->is_clonebasic() ||
612 use->as_ArrayCopy()->is_arraycopy_validated() ||
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
679 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
680 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
681 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
682 }
683 #endif
684 return can_eliminate;
714 JVMState *jvms = sfpt_done->jvms();
715 jvms->set_endoff(sfpt_done->req());
716 // Now make a pass over the debug information replacing any references
717 // to SafePointScalarObjectNode with the allocated object.
718 int start = jvms->debug_start();
719 int end = jvms->debug_end();
720 for (int i = start; i < end; i++) {
721 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
722 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
723 if (scobj->first_index(jvms) == sfpt_done->req() &&
724 scobj->n_fields() == (uint)nfields) {
725 assert(scobj->alloc() == alloc, "sanity");
726 sfpt_done->set_req(i, res);
727 }
728 }
729 }
730 _igvn._worklist.push(sfpt_done);
731 }
732 }
733
734 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode *alloc, SafePointNode* sfpt) {
735 // Fields of scalar objs are referenced only at the end
736 // of regular debuginfo at the last (youngest) JVMS.
737 // Record relative start index.
738 ciInstanceKlass* iklass = nullptr;
739 BasicType basic_elem_type = T_ILLEGAL;
740 const Type* field_type = nullptr;
741 const TypeOopPtr* res_type = nullptr;
742 int nfields = 0;
743 int array_base = 0;
744 int element_size = 0;
745 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
746 Node* res = alloc->result_cast();
747
748 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
749 assert(sfpt->jvms() != nullptr, "missed JVMS");
750
751 if (res != nullptr) { // Could be null when there are no users
752 res_type = _igvn.type(res)->isa_oopptr();
753
754 if (res_type->isa_instptr()) {
755 // find the fields of the class which will be needed for safepoint debug information
756 iklass = res_type->is_instptr()->instance_klass();
757 nfields = iklass->nof_nonstatic_fields();
758 } else {
759 // find the array's elements which will be needed for safepoint debug information
760 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
761 assert(nfields >= 0, "must be an array klass.");
762 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type();
763 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
764 element_size = type2aelembytes(basic_elem_type);
765 field_type = res_type->is_aryptr()->elem();
766 }
767 }
768
769 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, nfields);
770 sobj->init_req(0, C->root());
771 transform_later(sobj);
772
773 // Scan object's fields adding an input to the safepoint for each field.
774 for (int j = 0; j < nfields; j++) {
775 intptr_t offset;
776 ciField* field = nullptr;
777 if (iklass != nullptr) {
778 field = iklass->nonstatic_field_at(j);
779 offset = field->offset_in_bytes();
780 ciType* elem_type = field->type();
781 basic_elem_type = field->layout_type();
782
783 // The next code is taken from Parse::do_get_xxx().
784 if (is_reference_type(basic_elem_type)) {
785 if (!elem_type->is_loaded()) {
786 field_type = TypeInstPtr::BOTTOM;
787 } else if (field != nullptr && field->is_static_constant()) {
788 ciObject* con = field->constant_value().as_object();
789 // Do not "join" in the previous type; it doesn't add value,
790 // and may yield a vacuous result if the field is of interface type.
791 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
792 assert(field_type != nullptr, "field singleton type must be consistent");
793 } else {
794 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
795 }
796 if (UseCompressedOops) {
797 field_type = field_type->make_narrowoop();
798 basic_elem_type = T_NARROWOOP;
799 }
800 } else {
801 field_type = Type::get_const_basic_type(basic_elem_type);
802 }
803 } else {
804 offset = array_base + j * (intptr_t)element_size;
805 }
806
807 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
808
809 Node *field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
810
811 // We weren't able to find a value for this field,
812 // give up on eliminating this allocation.
813 if (field_val == nullptr) {
814 uint last = sfpt->req() - 1;
815 for (int k = 0; k < j; k++) {
816 sfpt->del_req(last--);
817 }
818 _igvn._worklist.push(sfpt);
819
820 #ifndef PRODUCT
821 if (PrintEliminateAllocations) {
822 if (field != nullptr) {
823 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
824 field->print();
825 int field_idx = C->get_alias_index(field_addr_type);
826 tty->print(" (alias_idx=%d)", field_idx);
827 } else { // Array's element
828 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, j);
829 }
830 tty->print(", which prevents elimination of: ");
831 if (res == nullptr)
832 alloc->dump();
833 else
834 res->dump();
835 }
836 #endif
837
838 return nullptr;
839 }
840
841 if (UseCompressedOops && field_type->isa_narrowoop()) {
842 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
843 // to be able scalar replace the allocation.
844 if (field_val->is_EncodeP()) {
845 field_val = field_val->in(1);
846 } else {
847 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
848 }
849 }
850 sfpt->add_req(field_val);
851 }
852
853 sfpt->jvms()->set_endoff(sfpt->req());
854
855 return sobj;
856 }
857
858 // Do scalar replacement.
859 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
860 GrowableArray <SafePointNode *> safepoints_done;
861 Node* res = alloc->result_cast();
862 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
863
864 // Process the safepoint uses
865 while (safepoints.length() > 0) {
866 SafePointNode* sfpt = safepoints.pop();
867 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt);
868
869 if (sobj == nullptr) {
870 undo_previous_scalarizations(safepoints_done, alloc);
871 return false;
872 }
873
874 // Now make a pass over the debug information replacing any references
875 // to the allocated object with "sobj"
876 JVMState *jvms = sfpt->jvms();
877 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
878 _igvn._worklist.push(sfpt);
879
880 // keep it for rollback
881 safepoints_done.append_if_missing(sfpt);
882 }
883
884 return true;
885 }
886
887 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
888 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
889 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
890 if (ctl_proj != nullptr) {
891 igvn.replace_node(ctl_proj, n->in(0));
892 }
893 if (mem_proj != nullptr) {
894 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
895 }
896 }
897
898 // Process users of eliminated allocation.
899 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
900 Node* res = alloc->result_cast();
901 if (res != nullptr) {
902 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
903 Node *use = res->last_out(j);
904 uint oc1 = res->outcnt();
905
906 if (use->is_AddP()) {
907 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
908 Node *n = use->last_out(k);
909 uint oc2 = use->outcnt();
910 if (n->is_Store()) {
911 #ifdef ASSERT
912 // Verify that there is no dependent MemBarVolatile nodes,
913 // they should be removed during IGVN, see MemBarNode::Ideal().
914 for (DUIterator_Fast pmax, p = n->fast_outs(pmax);
915 p < pmax; p++) {
916 Node* mb = n->fast_out(p);
917 assert(mb->is_Initialize() || !mb->is_MemBar() ||
918 mb->req() <= MemBarNode::Precedent ||
919 mb->in(MemBarNode::Precedent) != n,
920 "MemBarVolatile should be eliminated for non-escaping object");
921 }
922 #endif
923 _igvn.replace_node(n, n->in(MemNode::Memory));
924 } else {
925 eliminate_gc_barrier(n);
926 }
927 k -= (oc2 - use->outcnt());
928 }
929 _igvn.remove_dead_node(use);
930 } else if (use->is_ArrayCopy()) {
931 // Disconnect ArrayCopy node
932 ArrayCopyNode* ac = use->as_ArrayCopy();
933 if (ac->is_clonebasic()) {
934 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
935 disconnect_projections(ac, _igvn);
936 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
937 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
938 disconnect_projections(membar_before->as_MemBar(), _igvn);
939 if (membar_after->is_MemBar()) {
940 disconnect_projections(membar_after->as_MemBar(), _igvn);
941 }
942 } else {
943 assert(ac->is_arraycopy_validated() ||
944 ac->is_copyof_validated() ||
945 ac->is_copyofrange_validated(), "unsupported");
946 CallProjections callprojs;
947 ac->extract_projections(&callprojs, true);
948
949 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
950 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
951 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
952
953 // Set control to top. IGVN will remove the remaining projections
954 ac->set_req(0, top());
955 ac->replace_edge(res, top(), &_igvn);
956
957 // Disconnect src right away: it can help find new
958 // opportunities for allocation elimination
959 Node* src = ac->in(ArrayCopyNode::Src);
960 ac->replace_edge(src, top(), &_igvn);
961 // src can be top at this point if src and dest of the
962 // arraycopy were the same
963 if (src->outcnt() == 0 && !src->is_top()) {
964 _igvn.remove_dead_node(src);
965 }
966 }
967 _igvn._worklist.push(ac);
968 } else {
969 eliminate_gc_barrier(use);
970 }
971 j -= (oc1 - res->outcnt());
972 }
973 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
974 _igvn.remove_dead_node(res);
975 }
976
977 //
978 // Process other users of allocation's projections
979 //
980 if (_callprojs.resproj != nullptr && _callprojs.resproj->outcnt() != 0) {
981 // First disconnect stores captured by Initialize node.
982 // If Initialize node is eliminated first in the following code,
983 // it will kill such stores and DUIterator_Last will assert.
984 for (DUIterator_Fast jmax, j = _callprojs.resproj->fast_outs(jmax); j < jmax; j++) {
985 Node* use = _callprojs.resproj->fast_out(j);
986 if (use->is_AddP()) {
987 // raw memory addresses used only by the initialization
988 _igvn.replace_node(use, C->top());
989 --j; --jmax;
990 }
991 }
992 for (DUIterator_Last jmin, j = _callprojs.resproj->last_outs(jmin); j >= jmin; ) {
993 Node* use = _callprojs.resproj->last_out(j);
994 uint oc1 = _callprojs.resproj->outcnt();
995 if (use->is_Initialize()) {
996 // Eliminate Initialize node.
997 InitializeNode *init = use->as_Initialize();
998 assert(init->outcnt() <= 2, "only a control and memory projection expected");
999 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1000 if (ctrl_proj != nullptr) {
1001 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1002 #ifdef ASSERT
1003 // If the InitializeNode has no memory out, it will die, and tmp will become null
1004 Node* tmp = init->in(TypeFunc::Control);
1005 assert(tmp == nullptr || tmp == _callprojs.fallthrough_catchproj, "allocation control projection");
1006 #endif
1007 }
1008 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1009 if (mem_proj != nullptr) {
1010 Node *mem = init->in(TypeFunc::Memory);
1011 #ifdef ASSERT
1012 if (mem->is_MergeMem()) {
1013 assert(mem->in(TypeFunc::Memory) == _callprojs.fallthrough_memproj, "allocation memory projection");
1014 } else {
1015 assert(mem == _callprojs.fallthrough_memproj, "allocation memory projection");
1016 }
1017 #endif
1018 _igvn.replace_node(mem_proj, mem);
1019 }
1020 } else {
1021 assert(false, "only Initialize or AddP expected");
1022 }
1023 j -= (oc1 - _callprojs.resproj->outcnt());
1024 }
1025 }
1026 if (_callprojs.fallthrough_catchproj != nullptr) {
1027 _igvn.replace_node(_callprojs.fallthrough_catchproj, alloc->in(TypeFunc::Control));
1028 }
1029 if (_callprojs.fallthrough_memproj != nullptr) {
1030 _igvn.replace_node(_callprojs.fallthrough_memproj, alloc->in(TypeFunc::Memory));
1031 }
1032 if (_callprojs.catchall_memproj != nullptr) {
1033 _igvn.replace_node(_callprojs.catchall_memproj, C->top());
1034 }
1035 if (_callprojs.fallthrough_ioproj != nullptr) {
1036 _igvn.replace_node(_callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1037 }
1038 if (_callprojs.catchall_ioproj != nullptr) {
1039 _igvn.replace_node(_callprojs.catchall_ioproj, C->top());
1040 }
1041 if (_callprojs.catchall_catchproj != nullptr) {
1042 _igvn.replace_node(_callprojs.catchall_catchproj, C->top());
1043 }
1044 }
1045
1046 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1047 // If reallocation fails during deoptimization we'll pop all
1048 // interpreter frames for this compiled frame and that won't play
1049 // nice with JVMTI popframe.
1050 // We avoid this issue by eager reallocation when the popframe request
1051 // is received.
1052 if (!EliminateAllocations || !alloc->_is_non_escaping) {
1053 return false;
1054 }
1055 Node* klass = alloc->in(AllocateNode::KlassNode);
1056 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1057 Node* res = alloc->result_cast();
1058 // Eliminate boxing allocations which are not used
1059 // regardless scalar replaceable status.
1060 bool boxing_alloc = C->eliminate_boxing() &&
1061 tklass->isa_instklassptr() &&
1062 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1063 if (!alloc->_is_scalar_replaceable && (!boxing_alloc || (res != nullptr))) {
1064 return false;
1065 }
1066
1067 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1068
1069 GrowableArray <SafePointNode *> safepoints;
1070 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1071 return false;
1072 }
1073
1074 if (!alloc->_is_scalar_replaceable) {
1075 assert(res == nullptr, "sanity");
1076 // We can only eliminate allocation if all debug info references
1077 // are already replaced with SafePointScalarObject because
1078 // we can't search for a fields value without instance_id.
1079 if (safepoints.length() > 0) {
1080 return false;
1081 }
1082 }
1083
1084 if (!scalar_replacement(alloc, safepoints)) {
1085 return false;
1086 }
1087
1088 CompileLog* log = C->log();
1089 if (log != nullptr) {
1090 log->head("eliminate_allocation type='%d'",
1091 log->identify(tklass->exact_klass()));
1092 JVMState* p = alloc->jvms();
1093 while (p != nullptr) {
1094 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1095 p = p->caller();
1096 }
1097 log->tail("eliminate_allocation");
1098 }
1099
1100 process_users_of_allocation(alloc);
1101
1102 #ifndef PRODUCT
1103 if (PrintEliminateAllocations) {
1104 if (alloc->is_AllocateArray())
1105 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1106 else
1107 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1108 }
1109 #endif
1110
1111 return true;
1112 }
1113
1114 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1115 // EA should remove all uses of non-escaping boxing node.
1116 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1117 return false;
1118 }
1119
1120 assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1121
1122 boxing->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1123
1124 const TypeTuple* r = boxing->tf()->range();
1125 assert(r->cnt() > TypeFunc::Parms, "sanity");
1126 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1127 assert(t != nullptr, "sanity");
1128
1129 CompileLog* log = C->log();
1130 if (log != nullptr) {
1131 log->head("eliminate_boxing type='%d'",
1132 log->identify(t->instance_klass()));
1133 JVMState* p = boxing->jvms();
1134 while (p != nullptr) {
1135 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1136 p = p->caller();
1137 }
1138 log->tail("eliminate_boxing");
1139 }
1140
1141 process_users_of_allocation(boxing);
1142
1143 #ifndef PRODUCT
1144 if (PrintEliminateAllocations) {
1288 }
1289 }
1290 #endif
1291 yank_alloc_node(alloc);
1292 return;
1293 }
1294 }
1295
1296 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1297 Node *slow_region = nullptr;
1298 Node *toobig_false = ctrl;
1299
1300 // generate the initial test if necessary
1301 if (initial_slow_test != nullptr ) {
1302 assert (expand_fast_path, "Only need test if there is a fast path");
1303 slow_region = new RegionNode(3);
1304
1305 // Now make the initial failure test. Usually a too-big test but
1306 // might be a TRUE for finalizers or a fancy class check for
1307 // newInstance0.
1308 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1309 transform_later(toobig_iff);
1310 // Plug the failing-too-big test into the slow-path region
1311 Node *toobig_true = new IfTrueNode( toobig_iff );
1312 transform_later(toobig_true);
1313 slow_region ->init_req( too_big_or_final_path, toobig_true );
1314 toobig_false = new IfFalseNode( toobig_iff );
1315 transform_later(toobig_false);
1316 } else {
1317 // No initial test, just fall into next case
1318 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1319 toobig_false = ctrl;
1320 debug_only(slow_region = NodeSentinel);
1321 }
1322
1323 // If we are here there are several possibilities
1324 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1325 // no_initial_check means a constant allocation.
1326 // - If check always evaluates to false -> expand_fast_path is false (see above)
1327 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1328 // if !allocation_has_use the fast path is empty
1329 // if !allocation_has_use && no_initial_check
1330 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1331 // removed by yank_alloc_node above.
1332
1333 Node *slow_mem = mem; // save the current memory state for slow path
1334 // generate the fast allocation code unless we know that the initial test will always go slow
1335 if (expand_fast_path) {
1336 // Fast path modifies only raw memory.
1337 if (mem->is_MergeMem()) {
1338 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1339 }
1340
1341 // allocate the Region and Phi nodes for the result
1342 result_region = new RegionNode(3);
1343 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1344 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1345
1346 // Grab regular I/O before optional prefetch may change it.
1347 // Slow-path does no I/O so just set it to the original I/O.
1348 result_phi_i_o->init_req(slow_result_path, i_o);
1349
1350 // Name successful fast-path variables
1351 Node* fast_oop_ctrl;
1352 Node* fast_oop_rawmem;
1353 if (allocation_has_use) {
1354 Node* needgc_ctrl = nullptr;
1355 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1356
1357 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1358 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1359 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1360 fast_oop_ctrl, fast_oop_rawmem,
1361 prefetch_lines);
1362
1363 if (initial_slow_test != nullptr) {
1364 // This completes all paths into the slow merge point
1365 slow_region->init_req(need_gc_path, needgc_ctrl);
1366 transform_later(slow_region);
1367 } else {
1368 // No initial slow path needed!
1369 // Just fall from the need-GC path straight into the VM call.
1370 slow_region = needgc_ctrl;
1371 }
1372
1390 result_phi_i_o ->init_req(fast_result_path, i_o);
1391 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1392 } else {
1393 slow_region = ctrl;
1394 result_phi_i_o = i_o; // Rename it to use in the following code.
1395 }
1396
1397 // Generate slow-path call
1398 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1399 OptoRuntime::stub_name(slow_call_address),
1400 TypePtr::BOTTOM);
1401 call->init_req(TypeFunc::Control, slow_region);
1402 call->init_req(TypeFunc::I_O, top()); // does no i/o
1403 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1404 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1405 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1406
1407 call->init_req(TypeFunc::Parms+0, klass_node);
1408 if (length != nullptr) {
1409 call->init_req(TypeFunc::Parms+1, length);
1410 }
1411
1412 // Copy debug information and adjust JVMState information, then replace
1413 // allocate node with the call
1414 call->copy_call_debug_info(&_igvn, alloc);
1415 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1416 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1417 // path dies).
1418 if (valid_length_test != nullptr) {
1419 call->add_req(valid_length_test);
1420 }
1421 if (expand_fast_path) {
1422 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1423 } else {
1424 // Hook i_o projection to avoid its elimination during allocation
1425 // replacement (when only a slow call is generated).
1426 call->set_req(TypeFunc::I_O, result_phi_i_o);
1427 }
1428 _igvn.replace_node(alloc, call);
1429 transform_later(call);
1430
1431 // Identify the output projections from the allocate node and
1432 // adjust any references to them.
1433 // The control and io projections look like:
1434 //
1435 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1436 // Allocate Catch
1437 // ^---Proj(io) <-------+ ^---CatchProj(io)
1438 //
1439 // We are interested in the CatchProj nodes.
1440 //
1441 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1442
1443 // An allocate node has separate memory projections for the uses on
1444 // the control and i_o paths. Replace the control memory projection with
1445 // result_phi_rawmem (unless we are only generating a slow call when
1446 // both memory projections are combined)
1447 if (expand_fast_path && _callprojs.fallthrough_memproj != nullptr) {
1448 migrate_outs(_callprojs.fallthrough_memproj, result_phi_rawmem);
1449 }
1450 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1451 // catchall_memproj so we end up with a call that has only 1 memory projection.
1452 if (_callprojs.catchall_memproj != nullptr ) {
1453 if (_callprojs.fallthrough_memproj == nullptr) {
1454 _callprojs.fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1455 transform_later(_callprojs.fallthrough_memproj);
1456 }
1457 migrate_outs(_callprojs.catchall_memproj, _callprojs.fallthrough_memproj);
1458 _igvn.remove_dead_node(_callprojs.catchall_memproj);
1459 }
1460
1461 // An allocate node has separate i_o projections for the uses on the control
1462 // and i_o paths. Always replace the control i_o projection with result i_o
1463 // otherwise incoming i_o become dead when only a slow call is generated
1464 // (it is different from memory projections where both projections are
1465 // combined in such case).
1466 if (_callprojs.fallthrough_ioproj != nullptr) {
1467 migrate_outs(_callprojs.fallthrough_ioproj, result_phi_i_o);
1468 }
1469 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1470 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1471 if (_callprojs.catchall_ioproj != nullptr ) {
1472 if (_callprojs.fallthrough_ioproj == nullptr) {
1473 _callprojs.fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1474 transform_later(_callprojs.fallthrough_ioproj);
1475 }
1476 migrate_outs(_callprojs.catchall_ioproj, _callprojs.fallthrough_ioproj);
1477 _igvn.remove_dead_node(_callprojs.catchall_ioproj);
1478 }
1479
1480 // if we generated only a slow call, we are done
1481 if (!expand_fast_path) {
1482 // Now we can unhook i_o.
1483 if (result_phi_i_o->outcnt() > 1) {
1484 call->set_req(TypeFunc::I_O, top());
1485 } else {
1486 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1487 // Case of new array with negative size known during compilation.
1488 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1489 // following code since call to runtime will throw exception.
1490 // As result there will be no users of i_o after the call.
1491 // Leave i_o attached to this call to avoid problems in preceding graph.
1492 }
1493 return;
1494 }
1495
1496 if (_callprojs.fallthrough_catchproj != nullptr) {
1497 ctrl = _callprojs.fallthrough_catchproj->clone();
1498 transform_later(ctrl);
1499 _igvn.replace_node(_callprojs.fallthrough_catchproj, result_region);
1500 } else {
1501 ctrl = top();
1502 }
1503 Node *slow_result;
1504 if (_callprojs.resproj == nullptr) {
1505 // no uses of the allocation result
1506 slow_result = top();
1507 } else {
1508 slow_result = _callprojs.resproj->clone();
1509 transform_later(slow_result);
1510 _igvn.replace_node(_callprojs.resproj, result_phi_rawoop);
1511 }
1512
1513 // Plug slow-path into result merge point
1514 result_region->init_req( slow_result_path, ctrl);
1515 transform_later(result_region);
1516 if (allocation_has_use) {
1517 result_phi_rawoop->init_req(slow_result_path, slow_result);
1518 transform_later(result_phi_rawoop);
1519 }
1520 result_phi_rawmem->init_req(slow_result_path, _callprojs.fallthrough_memproj);
1521 transform_later(result_phi_rawmem);
1522 transform_later(result_phi_i_o);
1523 // This completes all paths into the result merge point
1524 }
1525
1526 // Remove alloc node that has no uses.
1527 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1528 Node* ctrl = alloc->in(TypeFunc::Control);
1529 Node* mem = alloc->in(TypeFunc::Memory);
1530 Node* i_o = alloc->in(TypeFunc::I_O);
1531
1532 alloc->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
1533 if (_callprojs.resproj != nullptr) {
1534 for (DUIterator_Fast imax, i = _callprojs.resproj->fast_outs(imax); i < imax; i++) {
1535 Node* use = _callprojs.resproj->fast_out(i);
1536 use->isa_MemBar()->remove(&_igvn);
1537 --imax;
1538 --i; // back up iterator
1539 }
1540 assert(_callprojs.resproj->outcnt() == 0, "all uses must be deleted");
1541 _igvn.remove_dead_node(_callprojs.resproj);
1542 }
1543 if (_callprojs.fallthrough_catchproj != nullptr) {
1544 migrate_outs(_callprojs.fallthrough_catchproj, ctrl);
1545 _igvn.remove_dead_node(_callprojs.fallthrough_catchproj);
1546 }
1547 if (_callprojs.catchall_catchproj != nullptr) {
1548 _igvn.rehash_node_delayed(_callprojs.catchall_catchproj);
1549 _callprojs.catchall_catchproj->set_req(0, top());
1550 }
1551 if (_callprojs.fallthrough_proj != nullptr) {
1552 Node* catchnode = _callprojs.fallthrough_proj->unique_ctrl_out();
1553 _igvn.remove_dead_node(catchnode);
1554 _igvn.remove_dead_node(_callprojs.fallthrough_proj);
1555 }
1556 if (_callprojs.fallthrough_memproj != nullptr) {
1557 migrate_outs(_callprojs.fallthrough_memproj, mem);
1558 _igvn.remove_dead_node(_callprojs.fallthrough_memproj);
1559 }
1560 if (_callprojs.fallthrough_ioproj != nullptr) {
1561 migrate_outs(_callprojs.fallthrough_ioproj, i_o);
1562 _igvn.remove_dead_node(_callprojs.fallthrough_ioproj);
1563 }
1564 if (_callprojs.catchall_memproj != nullptr) {
1565 _igvn.rehash_node_delayed(_callprojs.catchall_memproj);
1566 _callprojs.catchall_memproj->set_req(0, top());
1567 }
1568 if (_callprojs.catchall_ioproj != nullptr) {
1569 _igvn.rehash_node_delayed(_callprojs.catchall_ioproj);
1570 _callprojs.catchall_ioproj->set_req(0, top());
1571 }
1572 #ifndef PRODUCT
1573 if (PrintEliminateAllocations) {
1574 if (alloc->is_AllocateArray()) {
1575 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1576 } else {
1577 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1578 }
1579 }
1580 #endif
1581 _igvn.remove_dead_node(alloc);
1582 }
1583
1584 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1585 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1586 // If initialization is performed by an array copy, any required
1587 // MemBarStoreStore was already added. If the object does not
1588 // escape no need for a MemBarStoreStore. If the object does not
1589 // escape in its initializer and memory barrier (MemBarStoreStore or
1590 // stronger) is already added at exit of initializer, also no need
1668 Node* thread = new ThreadLocalNode();
1669 transform_later(thread);
1670
1671 call->init_req(TypeFunc::Parms + 0, thread);
1672 call->init_req(TypeFunc::Parms + 1, oop);
1673 call->init_req(TypeFunc::Control, ctrl);
1674 call->init_req(TypeFunc::I_O , top()); // does no i/o
1675 call->init_req(TypeFunc::Memory , rawmem);
1676 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1677 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1678 transform_later(call);
1679 ctrl = new ProjNode(call, TypeFunc::Control);
1680 transform_later(ctrl);
1681 rawmem = new ProjNode(call, TypeFunc::Memory);
1682 transform_later(rawmem);
1683 }
1684 }
1685
1686 // Helper for PhaseMacroExpand::expand_allocate_common.
1687 // Initializes the newly-allocated storage.
1688 Node*
1689 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1690 Node* control, Node* rawmem, Node* object,
1691 Node* klass_node, Node* length,
1692 Node* size_in_bytes) {
1693 InitializeNode* init = alloc->initialization();
1694 // Store the klass & mark bits
1695 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem);
1696 if (!mark_node->is_Con()) {
1697 transform_later(mark_node);
1698 }
1699 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1700
1701 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1702 int header_size = alloc->minimum_header_size(); // conservatively small
1703
1704 // Array length
1705 if (length != nullptr) { // Arrays need length field
1706 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1707 // conservatively small header size:
1708 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1709 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1710 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1711 if (is_reference_type(elem, true)) {
1712 elem = T_OBJECT;
1713 }
1714 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1715 }
1716 }
1717
1718 // Clear the object body, if necessary.
1719 if (init == nullptr) {
1720 // The init has somehow disappeared; be cautious and clear everything.
1721 //
1722 // This can happen if a node is allocated but an uncommon trap occurs
1723 // immediately. In this case, the Initialize gets associated with the
1724 // trap, and may be placed in a different (outer) loop, if the Allocate
1725 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1726 // there can be two Allocates to one Initialize. The answer in all these
1727 // edge cases is safety first. It is always safe to clear immediately
1728 // within an Allocate, and then (maybe or maybe not) clear some more later.
1729 if (!(UseTLAB && ZeroTLAB)) {
1730 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1731 header_size, size_in_bytes,
1732 &_igvn);
1733 }
1734 } else {
1735 if (!init->is_complete()) {
1736 // Try to win by zeroing only what the init does not store.
1737 // We can also try to do some peephole optimizations,
1738 // such as combining some adjacent subword stores.
1739 rawmem = init->complete_stores(control, rawmem, object,
1740 header_size, size_in_bytes, &_igvn);
1741 }
1742 // We have no more use for this link, since the AllocateNode goes away:
1743 init->set_req(InitializeNode::RawAddress, top());
1744 // (If we keep the link, it just confuses the register allocator,
1745 // who thinks he sees a real use of the address by the membar.)
1746 }
1747
1748 return rawmem;
1749 }
1750
2080 } // EliminateNestedLocks
2081
2082 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2083 // Look for all locks of this object and mark them and
2084 // corresponding BoxLock nodes as eliminated.
2085 Node* obj = alock->obj_node();
2086 for (uint j = 0; j < obj->outcnt(); j++) {
2087 Node* o = obj->raw_out(j);
2088 if (o->is_AbstractLock() &&
2089 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2090 alock = o->as_AbstractLock();
2091 Node* box = alock->box_node();
2092 // Replace old box node with new eliminated box for all users
2093 // of the same object and mark related locks as eliminated.
2094 mark_eliminated_box(box, obj);
2095 }
2096 }
2097 }
2098 }
2099
2100 // we have determined that this lock/unlock can be eliminated, we simply
2101 // eliminate the node without expanding it.
2102 //
2103 // Note: The membar's associated with the lock/unlock are currently not
2104 // eliminated. This should be investigated as a future enhancement.
2105 //
2106 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2107
2108 if (!alock->is_eliminated()) {
2109 return false;
2110 }
2111 #ifdef ASSERT
2112 if (!alock->is_coarsened()) {
2113 // Check that new "eliminated" BoxLock node is created.
2114 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2115 assert(oldbox->is_eliminated(), "should be done already");
2116 }
2117 #endif
2118
2119 alock->log_lock_optimization(C, "eliminate_lock");
2120
2121 #ifndef PRODUCT
2122 if (PrintEliminateLocks) {
2123 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2124 }
2125 #endif
2126
2127 Node* mem = alock->in(TypeFunc::Memory);
2128 Node* ctrl = alock->in(TypeFunc::Control);
2129 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2130
2131 alock->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2132 // There are 2 projections from the lock. The lock node will
2133 // be deleted when its last use is subsumed below.
2134 assert(alock->outcnt() == 2 &&
2135 _callprojs.fallthrough_proj != nullptr &&
2136 _callprojs.fallthrough_memproj != nullptr,
2137 "Unexpected projections from Lock/Unlock");
2138
2139 Node* fallthroughproj = _callprojs.fallthrough_proj;
2140 Node* memproj_fallthrough = _callprojs.fallthrough_memproj;
2141
2142 // The memory projection from a lock/unlock is RawMem
2143 // The input to a Lock is merged memory, so extract its RawMem input
2144 // (unless the MergeMem has been optimized away.)
2145 if (alock->is_Lock()) {
2146 // Search for MemBarAcquireLock node and delete it also.
2147 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2148 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2149 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2150 Node* memproj = membar->proj_out(TypeFunc::Memory);
2151 _igvn.replace_node(ctrlproj, fallthroughproj);
2152 _igvn.replace_node(memproj, memproj_fallthrough);
2153
2154 // Delete FastLock node also if this Lock node is unique user
2155 // (a loop peeling may clone a Lock node).
2156 Node* flock = alock->as_Lock()->fastlock_node();
2157 if (flock->outcnt() == 1) {
2158 assert(flock->unique_out() == alock, "sanity");
2159 _igvn.replace_node(flock, top());
2160 }
2161 }
2162
2163 // Search for MemBarReleaseLock node and delete it also.
2164 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2165 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2186 Node* mem = lock->in(TypeFunc::Memory);
2187 Node* obj = lock->obj_node();
2188 Node* box = lock->box_node();
2189 Node* flock = lock->fastlock_node();
2190
2191 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2192
2193 // Make the merge point
2194 Node *region;
2195 Node *mem_phi;
2196 Node *slow_path;
2197
2198 region = new RegionNode(3);
2199 // create a Phi for the memory state
2200 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2201
2202 // Optimize test; set region slot 2
2203 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2204 mem_phi->init_req(2, mem);
2205
2206 // Make slow path call
2207 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2208 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2209 obj, box, nullptr);
2210
2211 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2212
2213 // Slow path can only throw asynchronous exceptions, which are always
2214 // de-opted. So the compiler thinks the slow-call can never throw an
2215 // exception. If it DOES throw an exception we would need the debug
2216 // info removed first (since if it throws there is no monitor).
2217 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr &&
2218 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock");
2219
2220 // Capture slow path
2221 // disconnect fall-through projection from call and create a new one
2222 // hook up users of fall-through projection to region
2223 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2224 transform_later(slow_ctrl);
2225 _igvn.hash_delete(_callprojs.fallthrough_proj);
2226 _callprojs.fallthrough_proj->disconnect_inputs(C);
2227 region->init_req(1, slow_ctrl);
2228 // region inputs are now complete
2229 transform_later(region);
2230 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2231
2232 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2233
2234 mem_phi->init_req(1, memproj);
2235
2236 transform_later(mem_phi);
2237
2238 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2239 }
2240
2241 //------------------------------expand_unlock_node----------------------
2242 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2243
2244 Node* ctrl = unlock->in(TypeFunc::Control);
2245 Node* mem = unlock->in(TypeFunc::Memory);
2246 Node* obj = unlock->obj_node();
2247 Node* box = unlock->box_node();
2248
2249 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2250
2251 // No need for a null check on unlock
2252
2253 // Make the merge point
2254 Node *region;
2255 Node *mem_phi;
2256
2257 region = new RegionNode(3);
2258 // create a Phi for the memory state
2259 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2260
2261 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2262 funlock = transform_later( funlock )->as_FastUnlock();
2263 // Optimize test; set region slot 2
2264 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2265 Node *thread = transform_later(new ThreadLocalNode());
2266
2267 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2268 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2269 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2270
2271 call->extract_projections(&_callprojs, false /*separate_io_proj*/, false /*do_asserts*/);
2272 assert(_callprojs.fallthrough_ioproj == nullptr && _callprojs.catchall_ioproj == nullptr &&
2273 _callprojs.catchall_memproj == nullptr && _callprojs.catchall_catchproj == nullptr, "Unexpected projection from Lock");
2274
2275 // No exceptions for unlocking
2276 // Capture slow path
2277 // disconnect fall-through projection from call and create a new one
2278 // hook up users of fall-through projection to region
2279 Node *slow_ctrl = _callprojs.fallthrough_proj->clone();
2280 transform_later(slow_ctrl);
2281 _igvn.hash_delete(_callprojs.fallthrough_proj);
2282 _callprojs.fallthrough_proj->disconnect_inputs(C);
2283 region->init_req(1, slow_ctrl);
2284 // region inputs are now complete
2285 transform_later(region);
2286 _igvn.replace_node(_callprojs.fallthrough_proj, region);
2287
2288 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2289 mem_phi->init_req(1, memproj );
2290 mem_phi->init_req(2, mem);
2291 transform_later(mem_phi);
2292
2293 _igvn.replace_node(_callprojs.fallthrough_memproj, mem_phi);
2294 }
2295
2296 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2297 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2298 Node* bol = check->unique_out();
2299 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2300 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2301 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2302
2303 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2304 Node* iff = bol->last_out(i);
2305 assert(iff->is_If(), "where's the if?");
2306
2307 if (iff->in(0)->is_top()) {
2308 _igvn.replace_input_of(iff, 1, C->top());
2309 continue;
2310 }
2311
2312 Node* iftrue = iff->as_If()->proj_out(1);
2313 Node* iffalse = iff->as_If()->proj_out(0);
2314 Node* ctrl = iff->in(0);
2315
2316 Node* subklass = nullptr;
2317 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2318 subklass = obj_or_subklass;
2319 } else {
2320 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2321 subklass = _igvn.transform(LoadKlassNode::make(_igvn, nullptr, C->immutable_memory(), k_adr, TypeInstPtr::KLASS));
2322 }
2323
2324 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
2325
2326 _igvn.replace_input_of(iff, 0, C->top());
2327 _igvn.replace_node(iftrue, not_subtype_ctrl);
2328 _igvn.replace_node(iffalse, ctrl);
2329 }
2330 _igvn.replace_node(check, C->top());
2331 }
2332
2333 //---------------------------eliminate_macro_nodes----------------------
2334 // Eliminate scalar replaced allocations and associated locks.
2335 void PhaseMacroExpand::eliminate_macro_nodes() {
2336 if (C->macro_count() == 0)
2337 return;
2338 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2339
2340 // Before elimination may re-mark (change to Nested or NonEscObj)
2341 // all associated (same box and obj) lock and unlock nodes.
2342 int cnt = C->macro_count();
2343 for (int i=0; i < cnt; i++) {
2344 Node *n = C->macro_node(i);
2345 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2346 mark_eliminated_locking_nodes(n->as_AbstractLock());
2347 }
2348 }
2349 // Re-marking may break consistency of Coarsened locks.
2350 if (!C->coarsened_locks_consistent()) {
2351 return; // recompile without Coarsened locks if broken
2352 }
2373 }
2374 // Next, attempt to eliminate allocations
2375 _has_locks = false;
2376 progress = true;
2377 while (progress) {
2378 progress = false;
2379 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2380 Node* n = C->macro_node(i - 1);
2381 bool success = false;
2382 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2383 switch (n->class_id()) {
2384 case Node::Class_Allocate:
2385 case Node::Class_AllocateArray:
2386 success = eliminate_allocate_node(n->as_Allocate());
2387 #ifndef PRODUCT
2388 if (success && PrintOptoStatistics) {
2389 Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
2390 }
2391 #endif
2392 break;
2393 case Node::Class_CallStaticJava:
2394 success = eliminate_boxing_node(n->as_CallStaticJava());
2395 break;
2396 case Node::Class_Lock:
2397 case Node::Class_Unlock:
2398 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2399 _has_locks = true;
2400 break;
2401 case Node::Class_ArrayCopy:
2402 break;
2403 case Node::Class_OuterStripMinedLoop:
2404 break;
2405 case Node::Class_SubTypeCheck:
2406 break;
2407 case Node::Class_Opaque1:
2408 break;
2409 default:
2410 assert(n->Opcode() == Op_LoopLimit ||
2411 n->Opcode() == Op_Opaque3 ||
2412 n->Opcode() == Op_Opaque4 ||
2413 n->Opcode() == Op_MaxL ||
2414 n->Opcode() == Op_MinL ||
2415 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2416 "unknown node type in macro list");
2417 }
2418 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2419 progress = progress || success;
2420 }
2421 }
2422 #ifndef PRODUCT
2423 if (PrintOptoStatistics) {
2424 int membar_after = count_MemBar(C);
2425 Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
2426 }
2427 #endif
2428 }
2431 // Returns true if a failure occurred.
2432 bool PhaseMacroExpand::expand_macro_nodes() {
2433 // Last attempt to eliminate macro nodes.
2434 eliminate_macro_nodes();
2435 if (C->failing()) return true;
2436
2437 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2438 bool progress = true;
2439 while (progress) {
2440 progress = false;
2441 for (int i = C->macro_count(); i > 0; i--) {
2442 Node* n = C->macro_node(i-1);
2443 bool success = false;
2444 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2445 if (n->Opcode() == Op_LoopLimit) {
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->Opcode() == Op_CallStaticJava) {
2451 // Remove it from macro list and put on IGVN worklist to optimize.
2452 C->remove_macro_node(n);
2453 _igvn._worklist.push(n);
2454 success = true;
2455 } else if (n->is_Opaque1()) {
2456 _igvn.replace_node(n, n->in(1));
2457 success = true;
2458 #if INCLUDE_RTM_OPT
2459 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2460 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2461 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2462 Node* cmp = n->unique_out();
2463 #ifdef ASSERT
2464 // Validate graph.
2465 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2466 BoolNode* bol = cmp->unique_out()->as_Bool();
2467 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2468 (bol->_test._test == BoolTest::ne), "");
2469 IfNode* ifn = bol->unique_out()->as_If();
2470 assert((ifn->outcnt() == 2) &&
2471 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != nullptr, "");
2472 #endif
2473 Node* repl = n->in(1);
2474 if (!_has_locks) {
2549 // Worst case is a macro node gets expanded into about 200 nodes.
2550 // Allow 50% more for optimization.
2551 if (C->check_node_count(300, "out of nodes before macro expansion")) {
2552 return true;
2553 }
2554
2555 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2556 switch (n->class_id()) {
2557 case Node::Class_Lock:
2558 expand_lock_node(n->as_Lock());
2559 break;
2560 case Node::Class_Unlock:
2561 expand_unlock_node(n->as_Unlock());
2562 break;
2563 case Node::Class_ArrayCopy:
2564 expand_arraycopy_node(n->as_ArrayCopy());
2565 break;
2566 case Node::Class_SubTypeCheck:
2567 expand_subtypecheck_node(n->as_SubTypeCheck());
2568 break;
2569 default:
2570 assert(false, "unknown node type in macro list");
2571 }
2572 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
2573 if (C->failing()) return true;
2574
2575 // Clean up the graph so we're less likely to hit the maximum node
2576 // limit
2577 _igvn.set_delay_transform(false);
2578 _igvn.optimize();
2579 if (C->failing()) return true;
2580 _igvn.set_delay_transform(true);
2581 }
2582
2583 // All nodes except Allocate nodes are expanded now. There could be
2584 // new optimization opportunities (such as folding newly created
2585 // load from a just allocated object). Run IGVN.
2586
2587 // expand "macro" nodes
2588 // 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()) {
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) && !reduce_merge_precheck) {
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";)
670 (use->as_ArrayCopy()->is_clonebasic() ||
671 use->as_ArrayCopy()->is_arraycopy_validated() ||
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 if (res->is_Phi() && res->as_Phi()->can_be_inline_type()) {
691 // Can only eliminate allocation if the phi had been replaced by an InlineTypeNode before which did not happen.
692 // TODO 8325106 Why wasn't it replaced by an InlineTypeNode?
693 can_eliminate = false;
694 }
695 safepoints->append_if_missing(sfpt);
696 }
697 } else if (use->is_InlineType() && use->as_InlineType()->get_oop() == res) {
698 // Look at uses
699 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
700 Node* u = use->fast_out(k);
701 if (u->is_InlineType()) {
702 // Use in flat field can be eliminated
703 InlineTypeNode* vt = u->as_InlineType();
704 for (uint i = 0; i < vt->field_count(); ++i) {
705 if (vt->field_value(i) == use && !vt->field_is_flat(i)) {
706 can_eliminate = false; // Use in non-flat field
707 break;
708 }
709 }
710 } else {
711 // Add other uses to the worklist to process individually
712 // TODO will be fixed by 8328470
713 worklist.push(use);
714 }
715 }
716 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
717 // Store to mark word of inline type larval buffer
718 assert(res_type->is_inlinetypeptr(), "Unexpected store to mark word");
719 } else if (res_type->is_inlinetypeptr() && use->Opcode() == Op_MemBarRelease) {
720 // Inline type buffer allocations are followed by a membar
721 } else if (reduce_merge_precheck && (use->is_Phi() || use->is_EncodeP() || use->Opcode() == Op_MemBarRelease)) {
722 // Nothing to do
723 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
724 if (use->is_Phi()) {
725 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
726 NOT_PRODUCT(fail_eliminate = "Object is return value";)
727 } else {
728 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
729 }
730 DEBUG_ONLY(disq_node = use;)
731 } else {
732 if (use->Opcode() == Op_Return) {
733 NOT_PRODUCT(fail_eliminate = "Object is return value";)
734 } else {
735 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
736 }
737 DEBUG_ONLY(disq_node = use;)
738 }
739 can_eliminate = false;
740 } else {
741 assert(use->Opcode() == Op_CastP2X, "should be");
742 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
743 }
744 }
745 }
746
747 #ifndef PRODUCT
748 if (PrintEliminateAllocations && safepoints != nullptr) {
749 if (can_eliminate) {
750 tty->print("Scalar ");
751 if (res == nullptr)
752 alloc->dump();
753 else
754 res->dump();
755 } else {
756 tty->print("NotScalar (%s)", fail_eliminate);
757 if (res == nullptr)
758 alloc->dump();
759 else
760 res->dump();
761 #ifdef ASSERT
762 if (disq_node != nullptr) {
763 tty->print(" >>>> ");
764 disq_node->dump();
765 }
766 #endif /*ASSERT*/
767 }
768 }
769
770 if (TraceReduceAllocationMerges && !can_eliminate && reduce_merge_precheck) {
771 tty->print_cr("\tCan't eliminate allocation because '%s': ", fail_eliminate != nullptr ? fail_eliminate : "");
772 DEBUG_ONLY(if (disq_node != nullptr) disq_node->dump();)
773 }
774 #endif
775 return can_eliminate;
805 JVMState *jvms = sfpt_done->jvms();
806 jvms->set_endoff(sfpt_done->req());
807 // Now make a pass over the debug information replacing any references
808 // to SafePointScalarObjectNode with the allocated object.
809 int start = jvms->debug_start();
810 int end = jvms->debug_end();
811 for (int i = start; i < end; i++) {
812 if (sfpt_done->in(i)->is_SafePointScalarObject()) {
813 SafePointScalarObjectNode* scobj = sfpt_done->in(i)->as_SafePointScalarObject();
814 if (scobj->first_index(jvms) == sfpt_done->req() &&
815 scobj->n_fields() == (uint)nfields) {
816 assert(scobj->alloc() == alloc, "sanity");
817 sfpt_done->set_req(i, res);
818 }
819 }
820 }
821 _igvn._worklist.push(sfpt_done);
822 }
823 }
824
825 SafePointScalarObjectNode* PhaseMacroExpand::create_scalarized_object_description(AllocateNode *alloc, SafePointNode* sfpt,
826 Unique_Node_List* value_worklist) {
827 // Fields of scalar objs are referenced only at the end
828 // of regular debuginfo at the last (youngest) JVMS.
829 // Record relative start index.
830 ciInstanceKlass* iklass = nullptr;
831 BasicType basic_elem_type = T_ILLEGAL;
832 const Type* field_type = nullptr;
833 const TypeOopPtr* res_type = nullptr;
834 int nfields = 0;
835 int array_base = 0;
836 int element_size = 0;
837 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
838 Node* res = alloc->result_cast();
839
840 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
841 assert(sfpt->jvms() != nullptr, "missed JVMS");
842
843 if (res != nullptr) { // Could be null when there are no users
844 res_type = _igvn.type(res)->isa_oopptr();
845
846 if (res_type->isa_instptr()) {
847 // find the fields of the class which will be needed for safepoint debug information
848 iklass = res_type->is_instptr()->instance_klass();
849 nfields = iklass->nof_nonstatic_fields();
850 } else {
851 // find the array's elements which will be needed for safepoint debug information
852 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
853 assert(nfields >= 0, "must be an array klass.");
854 basic_elem_type = res_type->is_aryptr()->elem()->array_element_basic_type();
855 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
856 element_size = type2aelembytes(basic_elem_type);
857 field_type = res_type->is_aryptr()->elem();
858 if (res_type->is_flat()) {
859 // Flat inline type array
860 element_size = res_type->is_aryptr()->flat_elem_size();
861 }
862 }
863 }
864
865 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type, alloc, first_ind, nfields);
866 sobj->init_req(0, C->root());
867 transform_later(sobj);
868
869 // Scan object's fields adding an input to the safepoint for each field.
870 for (int j = 0; j < nfields; j++) {
871 intptr_t offset;
872 ciField* field = nullptr;
873 if (iklass != nullptr) {
874 field = iklass->nonstatic_field_at(j);
875 offset = field->offset_in_bytes();
876 ciType* elem_type = field->type();
877 basic_elem_type = field->layout_type();
878 assert(!field->is_flat(), "flat inline type fields should not have safepoint uses");
879
880 // The next code is taken from Parse::do_get_xxx().
881 if (is_reference_type(basic_elem_type)) {
882 if (!elem_type->is_loaded()) {
883 field_type = TypeInstPtr::BOTTOM;
884 } else if (field != nullptr && field->is_static_constant()) {
885 ciObject* con = field->constant_value().as_object();
886 // Do not "join" in the previous type; it doesn't add value,
887 // and may yield a vacuous result if the field is of interface type.
888 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
889 assert(field_type != nullptr, "field singleton type must be consistent");
890 } else {
891 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
892 }
893 if (UseCompressedOops) {
894 field_type = field_type->make_narrowoop();
895 basic_elem_type = T_NARROWOOP;
896 }
897 } else {
898 field_type = Type::get_const_basic_type(basic_elem_type);
899 }
900 } else {
901 offset = array_base + j * (intptr_t)element_size;
902 }
903
904 Node* field_val = nullptr;
905 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
906 if (res_type->is_flat()) {
907 ciInlineKlass* inline_klass = res_type->is_aryptr()->elem()->inline_klass();
908 assert(inline_klass->flat_in_array(), "must be flat in array");
909 field_val = inline_type_from_mem(sfpt->memory(), sfpt->control(), inline_klass, field_addr_type->isa_aryptr(), 0, alloc);
910 } else {
911 field_val = value_from_mem(sfpt->memory(), sfpt->control(), basic_elem_type, field_type, field_addr_type, alloc);
912 }
913
914 // We weren't able to find a value for this field,
915 // give up on eliminating this allocation.
916 if (field_val == nullptr) {
917 uint last = sfpt->req() - 1;
918 for (int k = 0; k < j; k++) {
919 sfpt->del_req(last--);
920 }
921 _igvn._worklist.push(sfpt);
922
923 #ifndef PRODUCT
924 if (PrintEliminateAllocations) {
925 if (field != nullptr) {
926 tty->print("=== At SafePoint node %d can't find value of field: ", sfpt->_idx);
927 field->print();
928 int field_idx = C->get_alias_index(field_addr_type);
929 tty->print(" (alias_idx=%d)", field_idx);
930 } else { // Array's element
931 tty->print("=== At SafePoint node %d can't find value of array element [%d]", sfpt->_idx, j);
932 }
933 tty->print(", which prevents elimination of: ");
934 if (res == nullptr)
935 alloc->dump();
936 else
937 res->dump();
938 }
939 #endif
940
941 return nullptr;
942 }
943
944 if (UseCompressedOops && field_type->isa_narrowoop()) {
945 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
946 // to be able scalar replace the allocation.
947 if (field_val->is_EncodeP()) {
948 field_val = field_val->in(1);
949 } else if (!field_val->is_InlineType()) {
950 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
951 }
952 }
953
954 // Keep track of inline types to scalarize them later
955 if (field_val->is_InlineType()) {
956 value_worklist->push(field_val);
957 } else if (field_val->is_Phi()) {
958 PhiNode* phi = field_val->as_Phi();
959 // Eagerly replace inline type phis now since we could be removing an inline type allocation where we must
960 // scalarize all its fields in safepoints.
961 field_val = phi->try_push_inline_types_down(&_igvn, true);
962 if (field_val->is_InlineType()) {
963 value_worklist->push(field_val);
964 }
965 }
966 sfpt->add_req(field_val);
967 }
968
969 sfpt->jvms()->set_endoff(sfpt->req());
970
971 return sobj;
972 }
973
974 // Do scalar replacement.
975 bool PhaseMacroExpand::scalar_replacement(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
976 GrowableArray <SafePointNode *> safepoints_done;
977 Node* res = alloc->result_cast();
978 assert(res == nullptr || res->is_CheckCastPP(), "unexpected AllocateNode result");
979 const TypeOopPtr* res_type = nullptr;
980 if (res != nullptr) { // Could be null when there are no users
981 res_type = _igvn.type(res)->isa_oopptr();
982 }
983
984 // Process the safepoint uses
985 assert(safepoints.length() == 0 || !res_type->is_inlinetypeptr(), "Inline type allocations should not have safepoint uses");
986 Unique_Node_List value_worklist;
987 while (safepoints.length() > 0) {
988 SafePointNode* sfpt = safepoints.pop();
989 SafePointScalarObjectNode* sobj = create_scalarized_object_description(alloc, sfpt, &value_worklist);
990
991 if (sobj == nullptr) {
992 undo_previous_scalarizations(safepoints_done, alloc);
993 return false;
994 }
995
996 // Now make a pass over the debug information replacing any references
997 // to the allocated object with "sobj"
998 JVMState *jvms = sfpt->jvms();
999 sfpt->replace_edges_in_range(res, sobj, jvms->debug_start(), jvms->debug_end(), &_igvn);
1000 _igvn._worklist.push(sfpt);
1001
1002 // keep it for rollback
1003 safepoints_done.append_if_missing(sfpt);
1004 }
1005 // Scalarize inline types that were added to the safepoint.
1006 // Don't allow linking a constant oop (if available) for flat array elements
1007 // because Deoptimization::reassign_flat_array_elements needs field values.
1008 bool allow_oop = (res_type != nullptr) && !res_type->is_flat();
1009 for (uint i = 0; i < value_worklist.size(); ++i) {
1010 InlineTypeNode* vt = value_worklist.at(i)->as_InlineType();
1011 vt->make_scalar_in_safepoints(&_igvn, allow_oop);
1012 }
1013 return true;
1014 }
1015
1016 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
1017 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
1018 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
1019 if (ctl_proj != nullptr) {
1020 igvn.replace_node(ctl_proj, n->in(0));
1021 }
1022 if (mem_proj != nullptr) {
1023 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
1024 }
1025 }
1026
1027 // Process users of eliminated allocation.
1028 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc, bool inline_alloc) {
1029 Unique_Node_List worklist;
1030 Node* res = alloc->result_cast();
1031 if (res != nullptr) {
1032 worklist.push(res);
1033 }
1034 while (worklist.size() > 0) {
1035 res = worklist.pop();
1036 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1037 Node *use = res->last_out(j);
1038 uint oc1 = res->outcnt();
1039
1040 if (use->is_AddP()) {
1041 for (DUIterator_Last kmin, k = use->last_outs(kmin); k >= kmin; ) {
1042 Node *n = use->last_out(k);
1043 uint oc2 = use->outcnt();
1044 if (n->is_Store()) {
1045 for (DUIterator_Fast pmax, p = n->fast_outs(pmax); p < pmax; p++) {
1046 MemBarNode* mb = n->fast_out(p)->isa_MemBar();
1047 if (mb != nullptr && mb->req() <= MemBarNode::Precedent && mb->in(MemBarNode::Precedent) == n) {
1048 // MemBarVolatiles should have been removed by MemBarNode::Ideal() for non-inline allocations
1049 assert(inline_alloc, "MemBarVolatile should be eliminated for non-escaping object");
1050 mb->remove(&_igvn);
1051 }
1052 }
1053 _igvn.replace_node(n, n->in(MemNode::Memory));
1054 } else {
1055 eliminate_gc_barrier(n);
1056 }
1057 k -= (oc2 - use->outcnt());
1058 }
1059 _igvn.remove_dead_node(use);
1060 } else if (use->is_ArrayCopy()) {
1061 // Disconnect ArrayCopy node
1062 ArrayCopyNode* ac = use->as_ArrayCopy();
1063 if (ac->is_clonebasic()) {
1064 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1065 disconnect_projections(ac, _igvn);
1066 assert(alloc->in(TypeFunc::Memory)->is_Proj() && alloc->in(TypeFunc::Memory)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1067 Node* membar_before = alloc->in(TypeFunc::Memory)->in(0);
1068 disconnect_projections(membar_before->as_MemBar(), _igvn);
1069 if (membar_after->is_MemBar()) {
1070 disconnect_projections(membar_after->as_MemBar(), _igvn);
1071 }
1072 } else {
1073 assert(ac->is_arraycopy_validated() ||
1074 ac->is_copyof_validated() ||
1075 ac->is_copyofrange_validated(), "unsupported");
1076 CallProjections* callprojs = ac->extract_projections(true);
1077
1078 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1079 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1080 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1081
1082 // Set control to top. IGVN will remove the remaining projections
1083 ac->set_req(0, top());
1084 ac->replace_edge(res, top(), &_igvn);
1085
1086 // Disconnect src right away: it can help find new
1087 // opportunities for allocation elimination
1088 Node* src = ac->in(ArrayCopyNode::Src);
1089 ac->replace_edge(src, top(), &_igvn);
1090 // src can be top at this point if src and dest of the
1091 // arraycopy were the same
1092 if (src->outcnt() == 0 && !src->is_top()) {
1093 _igvn.remove_dead_node(src);
1094 }
1095 }
1096 _igvn._worklist.push(ac);
1097 } else if (use->is_InlineType()) {
1098 assert(use->as_InlineType()->get_oop() == res, "unexpected inline type ptr use");
1099 // Cut off oop input and remove known instance id from type
1100 _igvn.rehash_node_delayed(use);
1101 use->as_InlineType()->set_oop(_igvn, _igvn.zerocon(T_OBJECT));
1102 const TypeOopPtr* toop = _igvn.type(use)->is_oopptr()->cast_to_instance_id(TypeOopPtr::InstanceBot);
1103 _igvn.set_type(use, toop);
1104 use->as_InlineType()->set_type(toop);
1105 // Process users
1106 for (DUIterator_Fast kmax, k = use->fast_outs(kmax); k < kmax; k++) {
1107 Node* u = use->fast_out(k);
1108 if (!u->is_InlineType()) {
1109 worklist.push(u);
1110 }
1111 }
1112 } else if (use->Opcode() == Op_StoreX && use->in(MemNode::Address) == res) {
1113 // Store to mark word of inline type larval buffer
1114 assert(inline_alloc, "Unexpected store to mark word");
1115 _igvn.replace_node(use, use->in(MemNode::Memory));
1116 } else if (use->Opcode() == Op_MemBarRelease) {
1117 // Inline type buffer allocations are followed by a membar
1118 assert(inline_alloc, "Unexpected MemBarRelease");
1119 use->as_MemBar()->remove(&_igvn);
1120 } else {
1121 eliminate_gc_barrier(use);
1122 }
1123 j -= (oc1 - res->outcnt());
1124 }
1125 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1126 _igvn.remove_dead_node(res);
1127 }
1128
1129 //
1130 // Process other users of allocation's projections
1131 //
1132 if (_callprojs->resproj[0] != nullptr && _callprojs->resproj[0]->outcnt() != 0) {
1133 // First disconnect stores captured by Initialize node.
1134 // If Initialize node is eliminated first in the following code,
1135 // it will kill such stores and DUIterator_Last will assert.
1136 for (DUIterator_Fast jmax, j = _callprojs->resproj[0]->fast_outs(jmax); j < jmax; j++) {
1137 Node* use = _callprojs->resproj[0]->fast_out(j);
1138 if (use->is_AddP()) {
1139 // raw memory addresses used only by the initialization
1140 _igvn.replace_node(use, C->top());
1141 --j; --jmax;
1142 }
1143 }
1144 for (DUIterator_Last jmin, j = _callprojs->resproj[0]->last_outs(jmin); j >= jmin; ) {
1145 Node* use = _callprojs->resproj[0]->last_out(j);
1146 uint oc1 = _callprojs->resproj[0]->outcnt();
1147 if (use->is_Initialize()) {
1148 // Eliminate Initialize node.
1149 InitializeNode *init = use->as_Initialize();
1150 assert(init->outcnt() <= 2, "only a control and memory projection expected");
1151 Node *ctrl_proj = init->proj_out_or_null(TypeFunc::Control);
1152 if (ctrl_proj != nullptr) {
1153 _igvn.replace_node(ctrl_proj, init->in(TypeFunc::Control));
1154 #ifdef ASSERT
1155 // If the InitializeNode has no memory out, it will die, and tmp will become null
1156 Node* tmp = init->in(TypeFunc::Control);
1157 assert(tmp == nullptr || tmp == _callprojs->fallthrough_catchproj, "allocation control projection");
1158 #endif
1159 }
1160 Node *mem_proj = init->proj_out_or_null(TypeFunc::Memory);
1161 if (mem_proj != nullptr) {
1162 Node *mem = init->in(TypeFunc::Memory);
1163 #ifdef ASSERT
1164 if (mem->is_MergeMem()) {
1165 assert(mem->in(TypeFunc::Memory) == _callprojs->fallthrough_memproj, "allocation memory projection");
1166 } else {
1167 assert(mem == _callprojs->fallthrough_memproj, "allocation memory projection");
1168 }
1169 #endif
1170 _igvn.replace_node(mem_proj, mem);
1171 }
1172 } else if (use->Opcode() == Op_MemBarStoreStore) {
1173 // Inline type buffer allocations are followed by a membar
1174 assert(inline_alloc, "Unexpected MemBarStoreStore");
1175 use->as_MemBar()->remove(&_igvn);
1176 } else {
1177 assert(false, "only Initialize or AddP expected");
1178 }
1179 j -= (oc1 - _callprojs->resproj[0]->outcnt());
1180 }
1181 }
1182 if (_callprojs->fallthrough_catchproj != nullptr) {
1183 _igvn.replace_node(_callprojs->fallthrough_catchproj, alloc->in(TypeFunc::Control));
1184 }
1185 if (_callprojs->fallthrough_memproj != nullptr) {
1186 _igvn.replace_node(_callprojs->fallthrough_memproj, alloc->in(TypeFunc::Memory));
1187 }
1188 if (_callprojs->catchall_memproj != nullptr) {
1189 _igvn.replace_node(_callprojs->catchall_memproj, C->top());
1190 }
1191 if (_callprojs->fallthrough_ioproj != nullptr) {
1192 _igvn.replace_node(_callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O));
1193 }
1194 if (_callprojs->catchall_ioproj != nullptr) {
1195 _igvn.replace_node(_callprojs->catchall_ioproj, C->top());
1196 }
1197 if (_callprojs->catchall_catchproj != nullptr) {
1198 _igvn.replace_node(_callprojs->catchall_catchproj, C->top());
1199 }
1200 }
1201
1202 bool PhaseMacroExpand::eliminate_allocate_node(AllocateNode *alloc) {
1203 // If reallocation fails during deoptimization we'll pop all
1204 // interpreter frames for this compiled frame and that won't play
1205 // nice with JVMTI popframe.
1206 // We avoid this issue by eager reallocation when the popframe request
1207 // is received.
1208 if (!EliminateAllocations) {
1209 return false;
1210 }
1211 Node* klass = alloc->in(AllocateNode::KlassNode);
1212 const TypeKlassPtr* tklass = _igvn.type(klass)->is_klassptr();
1213
1214 // Attempt to eliminate inline type buffer allocations
1215 // regardless of usage and escape/replaceable status.
1216 bool inline_alloc = tklass->isa_instklassptr() &&
1217 tklass->is_instklassptr()->instance_klass()->is_inlinetype();
1218 if (!alloc->_is_non_escaping && !inline_alloc) {
1219 return false;
1220 }
1221 // Eliminate boxing allocations which are not used
1222 // regardless scalar replaceable status.
1223 Node* res = alloc->result_cast();
1224 bool boxing_alloc = (res == nullptr) && C->eliminate_boxing() &&
1225 tklass->isa_instklassptr() &&
1226 tklass->is_instklassptr()->instance_klass()->is_box_klass();
1227 if (!alloc->_is_scalar_replaceable && !boxing_alloc && !inline_alloc) {
1228 return false;
1229 }
1230
1231 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1232
1233 GrowableArray <SafePointNode *> safepoints;
1234 if (!can_eliminate_allocation(&_igvn, alloc, &safepoints)) {
1235 return false;
1236 }
1237
1238 if (!alloc->_is_scalar_replaceable) {
1239 assert(res == nullptr || inline_alloc, "sanity");
1240 // We can only eliminate allocation if all debug info references
1241 // are already replaced with SafePointScalarObject because
1242 // we can't search for a fields value without instance_id.
1243 if (safepoints.length() > 0) {
1244 assert(!inline_alloc, "Inline type allocations should not have safepoint uses");
1245 return false;
1246 }
1247 }
1248
1249 if (!scalar_replacement(alloc, safepoints)) {
1250 return false;
1251 }
1252
1253 CompileLog* log = C->log();
1254 if (log != nullptr) {
1255 log->head("eliminate_allocation type='%d'",
1256 log->identify(tklass->exact_klass()));
1257 JVMState* p = alloc->jvms();
1258 while (p != nullptr) {
1259 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1260 p = p->caller();
1261 }
1262 log->tail("eliminate_allocation");
1263 }
1264
1265 process_users_of_allocation(alloc, inline_alloc);
1266
1267 #ifndef PRODUCT
1268 if (PrintEliminateAllocations) {
1269 if (alloc->is_AllocateArray())
1270 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1271 else
1272 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1273 }
1274 #endif
1275
1276 return true;
1277 }
1278
1279 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1280 // EA should remove all uses of non-escaping boxing node.
1281 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != nullptr) {
1282 return false;
1283 }
1284
1285 assert(boxing->result_cast() == nullptr, "unexpected boxing node result");
1286
1287 _callprojs = boxing->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1288
1289 const TypeTuple* r = boxing->tf()->range_sig();
1290 assert(r->cnt() > TypeFunc::Parms, "sanity");
1291 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1292 assert(t != nullptr, "sanity");
1293
1294 CompileLog* log = C->log();
1295 if (log != nullptr) {
1296 log->head("eliminate_boxing type='%d'",
1297 log->identify(t->instance_klass()));
1298 JVMState* p = boxing->jvms();
1299 while (p != nullptr) {
1300 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1301 p = p->caller();
1302 }
1303 log->tail("eliminate_boxing");
1304 }
1305
1306 process_users_of_allocation(boxing);
1307
1308 #ifndef PRODUCT
1309 if (PrintEliminateAllocations) {
1453 }
1454 }
1455 #endif
1456 yank_alloc_node(alloc);
1457 return;
1458 }
1459 }
1460
1461 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1462 Node *slow_region = nullptr;
1463 Node *toobig_false = ctrl;
1464
1465 // generate the initial test if necessary
1466 if (initial_slow_test != nullptr ) {
1467 assert (expand_fast_path, "Only need test if there is a fast path");
1468 slow_region = new RegionNode(3);
1469
1470 // Now make the initial failure test. Usually a too-big test but
1471 // might be a TRUE for finalizers or a fancy class check for
1472 // newInstance0.
1473 IfNode* toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1474 transform_later(toobig_iff);
1475 // Plug the failing-too-big test into the slow-path region
1476 Node* toobig_true = new IfTrueNode(toobig_iff);
1477 transform_later(toobig_true);
1478 slow_region ->init_req( too_big_or_final_path, toobig_true );
1479 toobig_false = new IfFalseNode(toobig_iff);
1480 transform_later(toobig_false);
1481 } else {
1482 // No initial test, just fall into next case
1483 assert(allocation_has_use || !expand_fast_path, "Should already have been handled");
1484 toobig_false = ctrl;
1485 debug_only(slow_region = NodeSentinel);
1486 }
1487
1488 // If we are here there are several possibilities
1489 // - expand_fast_path is false - then only a slow path is expanded. That's it.
1490 // no_initial_check means a constant allocation.
1491 // - If check always evaluates to false -> expand_fast_path is false (see above)
1492 // - If check always evaluates to true -> directly into fast path (but may bailout to slowpath)
1493 // if !allocation_has_use the fast path is empty
1494 // if !allocation_has_use && no_initial_check
1495 // - Then there are no fastpath that can fall out to slowpath -> no allocation code at all.
1496 // removed by yank_alloc_node above.
1497
1498 Node *slow_mem = mem; // save the current memory state for slow path
1499 // generate the fast allocation code unless we know that the initial test will always go slow
1500 if (expand_fast_path) {
1501 // Fast path modifies only raw memory.
1502 if (mem->is_MergeMem()) {
1503 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1504 }
1505
1506 // allocate the Region and Phi nodes for the result
1507 result_region = new RegionNode(3);
1508 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1509 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1510
1511 // Grab regular I/O before optional prefetch may change it.
1512 // Slow-path does no I/O so just set it to the original I/O.
1513 result_phi_i_o->init_req(slow_result_path, i_o);
1514
1515 // Name successful fast-path variables
1516 Node* fast_oop_ctrl;
1517 Node* fast_oop_rawmem;
1518
1519 if (allocation_has_use) {
1520 Node* needgc_ctrl = nullptr;
1521 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1522
1523 intx prefetch_lines = length != nullptr ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1524 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1525 Node* fast_oop = bs->obj_allocate(this, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1526 fast_oop_ctrl, fast_oop_rawmem,
1527 prefetch_lines);
1528
1529 if (initial_slow_test != nullptr) {
1530 // This completes all paths into the slow merge point
1531 slow_region->init_req(need_gc_path, needgc_ctrl);
1532 transform_later(slow_region);
1533 } else {
1534 // No initial slow path needed!
1535 // Just fall from the need-GC path straight into the VM call.
1536 slow_region = needgc_ctrl;
1537 }
1538
1556 result_phi_i_o ->init_req(fast_result_path, i_o);
1557 result_phi_rawmem->init_req(fast_result_path, fast_oop_rawmem);
1558 } else {
1559 slow_region = ctrl;
1560 result_phi_i_o = i_o; // Rename it to use in the following code.
1561 }
1562
1563 // Generate slow-path call
1564 CallNode *call = new CallStaticJavaNode(slow_call_type, slow_call_address,
1565 OptoRuntime::stub_name(slow_call_address),
1566 TypePtr::BOTTOM);
1567 call->init_req(TypeFunc::Control, slow_region);
1568 call->init_req(TypeFunc::I_O, top()); // does no i/o
1569 call->init_req(TypeFunc::Memory, slow_mem); // may gc ptrs
1570 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1571 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1572
1573 call->init_req(TypeFunc::Parms+0, klass_node);
1574 if (length != nullptr) {
1575 call->init_req(TypeFunc::Parms+1, length);
1576 } else {
1577 // Let the runtime know if this is a larval allocation
1578 call->init_req(TypeFunc::Parms+1, _igvn.intcon(alloc->_larval));
1579 }
1580
1581 // Copy debug information and adjust JVMState information, then replace
1582 // allocate node with the call
1583 call->copy_call_debug_info(&_igvn, alloc);
1584 // For array allocations, copy the valid length check to the call node so Compile::final_graph_reshaping() can verify
1585 // that the call has the expected number of CatchProj nodes (in case the allocation always fails and the fallthrough
1586 // path dies).
1587 if (valid_length_test != nullptr) {
1588 call->add_req(valid_length_test);
1589 }
1590 if (expand_fast_path) {
1591 call->set_cnt(PROB_UNLIKELY_MAG(4)); // Same effect as RC_UNCOMMON.
1592 } else {
1593 // Hook i_o projection to avoid its elimination during allocation
1594 // replacement (when only a slow call is generated).
1595 call->set_req(TypeFunc::I_O, result_phi_i_o);
1596 }
1597 _igvn.replace_node(alloc, call);
1598 transform_later(call);
1599
1600 // Identify the output projections from the allocate node and
1601 // adjust any references to them.
1602 // The control and io projections look like:
1603 //
1604 // v---Proj(ctrl) <-----+ v---CatchProj(ctrl)
1605 // Allocate Catch
1606 // ^---Proj(io) <-------+ ^---CatchProj(io)
1607 //
1608 // We are interested in the CatchProj nodes.
1609 //
1610 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1611
1612 // An allocate node has separate memory projections for the uses on
1613 // the control and i_o paths. Replace the control memory projection with
1614 // result_phi_rawmem (unless we are only generating a slow call when
1615 // both memory projections are combined)
1616 if (expand_fast_path && _callprojs->fallthrough_memproj != nullptr) {
1617 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, result_phi_rawmem);
1618 }
1619 // Now change uses of catchall_memproj to use fallthrough_memproj and delete
1620 // catchall_memproj so we end up with a call that has only 1 memory projection.
1621 if (_callprojs->catchall_memproj != nullptr) {
1622 if (_callprojs->fallthrough_memproj == nullptr) {
1623 _callprojs->fallthrough_memproj = new ProjNode(call, TypeFunc::Memory);
1624 transform_later(_callprojs->fallthrough_memproj);
1625 }
1626 _igvn.replace_in_uses(_callprojs->catchall_memproj, _callprojs->fallthrough_memproj);
1627 _igvn.remove_dead_node(_callprojs->catchall_memproj);
1628 }
1629
1630 // An allocate node has separate i_o projections for the uses on the control
1631 // and i_o paths. Always replace the control i_o projection with result i_o
1632 // otherwise incoming i_o become dead when only a slow call is generated
1633 // (it is different from memory projections where both projections are
1634 // combined in such case).
1635 if (_callprojs->fallthrough_ioproj != nullptr) {
1636 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, result_phi_i_o);
1637 }
1638 // Now change uses of catchall_ioproj to use fallthrough_ioproj and delete
1639 // catchall_ioproj so we end up with a call that has only 1 i_o projection.
1640 if (_callprojs->catchall_ioproj != nullptr) {
1641 if (_callprojs->fallthrough_ioproj == nullptr) {
1642 _callprojs->fallthrough_ioproj = new ProjNode(call, TypeFunc::I_O);
1643 transform_later(_callprojs->fallthrough_ioproj);
1644 }
1645 _igvn.replace_in_uses(_callprojs->catchall_ioproj, _callprojs->fallthrough_ioproj);
1646 _igvn.remove_dead_node(_callprojs->catchall_ioproj);
1647 }
1648
1649 // if we generated only a slow call, we are done
1650 if (!expand_fast_path) {
1651 // Now we can unhook i_o.
1652 if (result_phi_i_o->outcnt() > 1) {
1653 call->set_req(TypeFunc::I_O, top());
1654 } else {
1655 assert(result_phi_i_o->unique_ctrl_out() == call, "sanity");
1656 // Case of new array with negative size known during compilation.
1657 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1658 // following code since call to runtime will throw exception.
1659 // As result there will be no users of i_o after the call.
1660 // Leave i_o attached to this call to avoid problems in preceding graph.
1661 }
1662 return;
1663 }
1664
1665 if (_callprojs->fallthrough_catchproj != nullptr) {
1666 ctrl = _callprojs->fallthrough_catchproj->clone();
1667 transform_later(ctrl);
1668 _igvn.replace_node(_callprojs->fallthrough_catchproj, result_region);
1669 } else {
1670 ctrl = top();
1671 }
1672 Node *slow_result;
1673 if (_callprojs->resproj[0] == nullptr) {
1674 // no uses of the allocation result
1675 slow_result = top();
1676 } else {
1677 slow_result = _callprojs->resproj[0]->clone();
1678 transform_later(slow_result);
1679 _igvn.replace_node(_callprojs->resproj[0], result_phi_rawoop);
1680 }
1681
1682 // Plug slow-path into result merge point
1683 result_region->init_req( slow_result_path, ctrl);
1684 transform_later(result_region);
1685 if (allocation_has_use) {
1686 result_phi_rawoop->init_req(slow_result_path, slow_result);
1687 transform_later(result_phi_rawoop);
1688 }
1689 result_phi_rawmem->init_req(slow_result_path, _callprojs->fallthrough_memproj);
1690 transform_later(result_phi_rawmem);
1691 transform_later(result_phi_i_o);
1692 // This completes all paths into the result merge point
1693 }
1694
1695 // Remove alloc node that has no uses.
1696 void PhaseMacroExpand::yank_alloc_node(AllocateNode* alloc) {
1697 Node* ctrl = alloc->in(TypeFunc::Control);
1698 Node* mem = alloc->in(TypeFunc::Memory);
1699 Node* i_o = alloc->in(TypeFunc::I_O);
1700
1701 _callprojs = alloc->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
1702 if (_callprojs->resproj[0] != nullptr) {
1703 for (DUIterator_Fast imax, i = _callprojs->resproj[0]->fast_outs(imax); i < imax; i++) {
1704 Node* use = _callprojs->resproj[0]->fast_out(i);
1705 use->isa_MemBar()->remove(&_igvn);
1706 --imax;
1707 --i; // back up iterator
1708 }
1709 assert(_callprojs->resproj[0]->outcnt() == 0, "all uses must be deleted");
1710 _igvn.remove_dead_node(_callprojs->resproj[0]);
1711 }
1712 if (_callprojs->fallthrough_catchproj != nullptr) {
1713 _igvn.replace_in_uses(_callprojs->fallthrough_catchproj, ctrl);
1714 _igvn.remove_dead_node(_callprojs->fallthrough_catchproj);
1715 }
1716 if (_callprojs->catchall_catchproj != nullptr) {
1717 _igvn.rehash_node_delayed(_callprojs->catchall_catchproj);
1718 _callprojs->catchall_catchproj->set_req(0, top());
1719 }
1720 if (_callprojs->fallthrough_proj != nullptr) {
1721 Node* catchnode = _callprojs->fallthrough_proj->unique_ctrl_out();
1722 _igvn.remove_dead_node(catchnode);
1723 _igvn.remove_dead_node(_callprojs->fallthrough_proj);
1724 }
1725 if (_callprojs->fallthrough_memproj != nullptr) {
1726 _igvn.replace_in_uses(_callprojs->fallthrough_memproj, mem);
1727 _igvn.remove_dead_node(_callprojs->fallthrough_memproj);
1728 }
1729 if (_callprojs->fallthrough_ioproj != nullptr) {
1730 _igvn.replace_in_uses(_callprojs->fallthrough_ioproj, i_o);
1731 _igvn.remove_dead_node(_callprojs->fallthrough_ioproj);
1732 }
1733 if (_callprojs->catchall_memproj != nullptr) {
1734 _igvn.rehash_node_delayed(_callprojs->catchall_memproj);
1735 _callprojs->catchall_memproj->set_req(0, top());
1736 }
1737 if (_callprojs->catchall_ioproj != nullptr) {
1738 _igvn.rehash_node_delayed(_callprojs->catchall_ioproj);
1739 _callprojs->catchall_ioproj->set_req(0, top());
1740 }
1741 #ifndef PRODUCT
1742 if (PrintEliminateAllocations) {
1743 if (alloc->is_AllocateArray()) {
1744 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1745 } else {
1746 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1747 }
1748 }
1749 #endif
1750 _igvn.remove_dead_node(alloc);
1751 }
1752
1753 void PhaseMacroExpand::expand_initialize_membar(AllocateNode* alloc, InitializeNode* init,
1754 Node*& fast_oop_ctrl, Node*& fast_oop_rawmem) {
1755 // If initialization is performed by an array copy, any required
1756 // MemBarStoreStore was already added. If the object does not
1757 // escape no need for a MemBarStoreStore. If the object does not
1758 // escape in its initializer and memory barrier (MemBarStoreStore or
1759 // stronger) is already added at exit of initializer, also no need
1837 Node* thread = new ThreadLocalNode();
1838 transform_later(thread);
1839
1840 call->init_req(TypeFunc::Parms + 0, thread);
1841 call->init_req(TypeFunc::Parms + 1, oop);
1842 call->init_req(TypeFunc::Control, ctrl);
1843 call->init_req(TypeFunc::I_O , top()); // does no i/o
1844 call->init_req(TypeFunc::Memory , rawmem);
1845 call->init_req(TypeFunc::ReturnAdr, alloc->in(TypeFunc::ReturnAdr));
1846 call->init_req(TypeFunc::FramePtr, alloc->in(TypeFunc::FramePtr));
1847 transform_later(call);
1848 ctrl = new ProjNode(call, TypeFunc::Control);
1849 transform_later(ctrl);
1850 rawmem = new ProjNode(call, TypeFunc::Memory);
1851 transform_later(rawmem);
1852 }
1853 }
1854
1855 // Helper for PhaseMacroExpand::expand_allocate_common.
1856 // Initializes the newly-allocated storage.
1857 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1858 Node* control, Node* rawmem, Node* object,
1859 Node* klass_node, Node* length,
1860 Node* size_in_bytes) {
1861 InitializeNode* init = alloc->initialization();
1862 // Store the klass & mark bits
1863 Node* mark_node = alloc->make_ideal_mark(&_igvn, control, rawmem);
1864 if (!mark_node->is_Con()) {
1865 transform_later(mark_node);
1866 }
1867 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1868
1869 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1870 int header_size = alloc->minimum_header_size(); // conservatively small
1871
1872 // Array length
1873 if (length != nullptr) { // Arrays need length field
1874 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1875 // conservatively small header size:
1876 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1877 if (_igvn.type(klass_node)->isa_aryklassptr()) { // we know the exact header size in most cases:
1878 BasicType elem = _igvn.type(klass_node)->is_klassptr()->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
1879 if (is_reference_type(elem, true)) {
1880 elem = T_OBJECT;
1881 }
1882 header_size = Klass::layout_helper_header_size(Klass::array_layout_helper(elem));
1883 }
1884 }
1885
1886 // Clear the object body, if necessary.
1887 if (init == nullptr) {
1888 // The init has somehow disappeared; be cautious and clear everything.
1889 //
1890 // This can happen if a node is allocated but an uncommon trap occurs
1891 // immediately. In this case, the Initialize gets associated with the
1892 // trap, and may be placed in a different (outer) loop, if the Allocate
1893 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1894 // there can be two Allocates to one Initialize. The answer in all these
1895 // edge cases is safety first. It is always safe to clear immediately
1896 // within an Allocate, and then (maybe or maybe not) clear some more later.
1897 if (!(UseTLAB && ZeroTLAB)) {
1898 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1899 alloc->in(AllocateNode::DefaultValue),
1900 alloc->in(AllocateNode::RawDefaultValue),
1901 header_size, size_in_bytes,
1902 &_igvn);
1903 }
1904 } else {
1905 if (!init->is_complete()) {
1906 // Try to win by zeroing only what the init does not store.
1907 // We can also try to do some peephole optimizations,
1908 // such as combining some adjacent subword stores.
1909 rawmem = init->complete_stores(control, rawmem, object,
1910 header_size, size_in_bytes, &_igvn);
1911 }
1912 // We have no more use for this link, since the AllocateNode goes away:
1913 init->set_req(InitializeNode::RawAddress, top());
1914 // (If we keep the link, it just confuses the register allocator,
1915 // who thinks he sees a real use of the address by the membar.)
1916 }
1917
1918 return rawmem;
1919 }
1920
2250 } // EliminateNestedLocks
2251
2252 if (alock->is_non_esc_obj()) { // Lock is used for non escaping object
2253 // Look for all locks of this object and mark them and
2254 // corresponding BoxLock nodes as eliminated.
2255 Node* obj = alock->obj_node();
2256 for (uint j = 0; j < obj->outcnt(); j++) {
2257 Node* o = obj->raw_out(j);
2258 if (o->is_AbstractLock() &&
2259 o->as_AbstractLock()->obj_node()->eqv_uncast(obj)) {
2260 alock = o->as_AbstractLock();
2261 Node* box = alock->box_node();
2262 // Replace old box node with new eliminated box for all users
2263 // of the same object and mark related locks as eliminated.
2264 mark_eliminated_box(box, obj);
2265 }
2266 }
2267 }
2268 }
2269
2270 void PhaseMacroExpand::inline_type_guard(Node** ctrl, LockNode* lock) {
2271 Node* obj = lock->obj_node();
2272 const TypePtr* obj_type = _igvn.type(obj)->make_ptr();
2273 if (!obj_type->can_be_inline_type()) {
2274 return;
2275 }
2276 Node* mark = make_load(*ctrl, lock->memory(), obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2277 Node* value_mask = _igvn.MakeConX(markWord::inline_type_pattern);
2278 Node* is_value = _igvn.transform(new AndXNode(mark, value_mask));
2279 Node* cmp = _igvn.transform(new CmpXNode(is_value, value_mask));
2280 Node* bol = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2281 Node* unc_ctrl = generate_slow_guard(ctrl, bol, nullptr);
2282
2283 int trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_class_check, Deoptimization::Action_none);
2284 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2285 const TypePtr* no_memory_effects = nullptr;
2286 CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap",
2287 no_memory_effects);
2288 unc->init_req(TypeFunc::Control, unc_ctrl);
2289 unc->init_req(TypeFunc::I_O, lock->i_o());
2290 unc->init_req(TypeFunc::Memory, lock->memory());
2291 unc->init_req(TypeFunc::FramePtr, lock->in(TypeFunc::FramePtr));
2292 unc->init_req(TypeFunc::ReturnAdr, lock->in(TypeFunc::ReturnAdr));
2293 unc->init_req(TypeFunc::Parms+0, _igvn.intcon(trap_request));
2294 unc->set_cnt(PROB_UNLIKELY_MAG(4));
2295 unc->copy_call_debug_info(&_igvn, lock);
2296
2297 assert(unc->peek_monitor_box() == lock->box_node(), "wrong monitor");
2298 assert((obj_type->is_inlinetypeptr() && unc->peek_monitor_obj()->is_SafePointScalarObject()) ||
2299 (obj->is_InlineType() && obj->in(1) == unc->peek_monitor_obj()) ||
2300 (obj == unc->peek_monitor_obj()), "wrong monitor");
2301
2302 // pop monitor and push obj back on stack: we trap before the monitorenter
2303 unc->pop_monitor();
2304 unc->grow_stack(unc->jvms(), 1);
2305 unc->set_stack(unc->jvms(), unc->jvms()->stk_size()-1, obj);
2306 _igvn.register_new_node_with_optimizer(unc);
2307
2308 unc_ctrl = _igvn.transform(new ProjNode(unc, TypeFunc::Control));
2309 Node* halt = _igvn.transform(new HaltNode(unc_ctrl, lock->in(TypeFunc::FramePtr), "monitor enter on inline type"));
2310 _igvn.add_input_to(C->root(), halt);
2311 }
2312
2313 // we have determined that this lock/unlock can be eliminated, we simply
2314 // eliminate the node without expanding it.
2315 //
2316 // Note: The membar's associated with the lock/unlock are currently not
2317 // eliminated. This should be investigated as a future enhancement.
2318 //
2319 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2320
2321 if (!alock->is_eliminated()) {
2322 return false;
2323 }
2324 #ifdef ASSERT
2325 if (!alock->is_coarsened()) {
2326 // Check that new "eliminated" BoxLock node is created.
2327 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2328 assert(oldbox->is_eliminated(), "should be done already");
2329 }
2330 #endif
2331
2332 alock->log_lock_optimization(C, "eliminate_lock");
2333
2334 #ifndef PRODUCT
2335 if (PrintEliminateLocks) {
2336 tty->print_cr("++++ Eliminated: %d %s '%s'", alock->_idx, (alock->is_Lock() ? "Lock" : "Unlock"), alock->kind_as_string());
2337 }
2338 #endif
2339
2340 Node* mem = alock->in(TypeFunc::Memory);
2341 Node* ctrl = alock->in(TypeFunc::Control);
2342 guarantee(ctrl != nullptr, "missing control projection, cannot replace_node() with null");
2343
2344 _callprojs = alock->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2345 // There are 2 projections from the lock. The lock node will
2346 // be deleted when its last use is subsumed below.
2347 assert(alock->outcnt() == 2 &&
2348 _callprojs->fallthrough_proj != nullptr &&
2349 _callprojs->fallthrough_memproj != nullptr,
2350 "Unexpected projections from Lock/Unlock");
2351
2352 Node* fallthroughproj = _callprojs->fallthrough_proj;
2353 Node* memproj_fallthrough = _callprojs->fallthrough_memproj;
2354
2355 // The memory projection from a lock/unlock is RawMem
2356 // The input to a Lock is merged memory, so extract its RawMem input
2357 // (unless the MergeMem has been optimized away.)
2358 if (alock->is_Lock()) {
2359 // Deoptimize and re-execute if object is an inline type
2360 inline_type_guard(&ctrl, alock->as_Lock());
2361
2362 // Search for MemBarAcquireLock node and delete it also.
2363 MemBarNode* membar = fallthroughproj->unique_ctrl_out()->as_MemBar();
2364 assert(membar != nullptr && membar->Opcode() == Op_MemBarAcquireLock, "");
2365 Node* ctrlproj = membar->proj_out(TypeFunc::Control);
2366 Node* memproj = membar->proj_out(TypeFunc::Memory);
2367 _igvn.replace_node(ctrlproj, fallthroughproj);
2368 _igvn.replace_node(memproj, memproj_fallthrough);
2369
2370 // Delete FastLock node also if this Lock node is unique user
2371 // (a loop peeling may clone a Lock node).
2372 Node* flock = alock->as_Lock()->fastlock_node();
2373 if (flock->outcnt() == 1) {
2374 assert(flock->unique_out() == alock, "sanity");
2375 _igvn.replace_node(flock, top());
2376 }
2377 }
2378
2379 // Search for MemBarReleaseLock node and delete it also.
2380 if (alock->is_Unlock() && ctrl->is_Proj() && ctrl->in(0)->is_MemBar()) {
2381 MemBarNode* membar = ctrl->in(0)->as_MemBar();
2402 Node* mem = lock->in(TypeFunc::Memory);
2403 Node* obj = lock->obj_node();
2404 Node* box = lock->box_node();
2405 Node* flock = lock->fastlock_node();
2406
2407 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2408
2409 // Make the merge point
2410 Node *region;
2411 Node *mem_phi;
2412 Node *slow_path;
2413
2414 region = new RegionNode(3);
2415 // create a Phi for the memory state
2416 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2417
2418 // Optimize test; set region slot 2
2419 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2420 mem_phi->init_req(2, mem);
2421
2422 // Deoptimize and re-execute if object is an inline type
2423 inline_type_guard(&slow_path, lock);
2424
2425 // Make slow path call
2426 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2427 OptoRuntime::complete_monitor_locking_Java(), nullptr, slow_path,
2428 obj, box, nullptr);
2429
2430 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2431
2432 // Slow path can only throw asynchronous exceptions, which are always
2433 // de-opted. So the compiler thinks the slow-call can never throw an
2434 // exception. If it DOES throw an exception we would need the debug
2435 // info removed first (since if it throws there is no monitor).
2436 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2437 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2438
2439 // Capture slow path
2440 // disconnect fall-through projection from call and create a new one
2441 // hook up users of fall-through projection to region
2442 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2443 transform_later(slow_ctrl);
2444 _igvn.hash_delete(_callprojs->fallthrough_proj);
2445 _callprojs->fallthrough_proj->disconnect_inputs(C);
2446 region->init_req(1, slow_ctrl);
2447 // region inputs are now complete
2448 transform_later(region);
2449 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2450
2451 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory));
2452
2453 mem_phi->init_req(1, memproj);
2454
2455 transform_later(mem_phi);
2456
2457 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2458 }
2459
2460 //------------------------------expand_unlock_node----------------------
2461 void PhaseMacroExpand::expand_unlock_node(UnlockNode *unlock) {
2462
2463 Node* ctrl = unlock->in(TypeFunc::Control);
2464 Node* mem = unlock->in(TypeFunc::Memory);
2465 Node* obj = unlock->obj_node();
2466 Node* box = unlock->box_node();
2467
2468 assert(!box->as_BoxLock()->is_eliminated(), "sanity");
2469
2470 // No need for a null check on unlock
2471
2472 // Make the merge point
2473 Node *region;
2474 Node *mem_phi;
2475
2476 region = new RegionNode(3);
2477 // create a Phi for the memory state
2478 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2479
2480 FastUnlockNode *funlock = new FastUnlockNode( ctrl, obj, box );
2481 funlock = transform_later( funlock )->as_FastUnlock();
2482 // Optimize test; set region slot 2
2483 Node *slow_path = opt_bits_test(ctrl, region, 2, funlock, 0, 0);
2484 Node *thread = transform_later(new ThreadLocalNode());
2485
2486 CallNode *call = make_slow_call((CallNode *) unlock, OptoRuntime::complete_monitor_exit_Type(),
2487 CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C),
2488 "complete_monitor_unlocking_C", slow_path, obj, box, thread);
2489
2490 _callprojs = call->extract_projections(false /*separate_io_proj*/, false /*do_asserts*/);
2491 assert(_callprojs->fallthrough_ioproj == nullptr && _callprojs->catchall_ioproj == nullptr &&
2492 _callprojs->catchall_memproj == nullptr && _callprojs->catchall_catchproj == nullptr, "Unexpected projection from Lock");
2493
2494 // No exceptions for unlocking
2495 // Capture slow path
2496 // disconnect fall-through projection from call and create a new one
2497 // hook up users of fall-through projection to region
2498 Node *slow_ctrl = _callprojs->fallthrough_proj->clone();
2499 transform_later(slow_ctrl);
2500 _igvn.hash_delete(_callprojs->fallthrough_proj);
2501 _callprojs->fallthrough_proj->disconnect_inputs(C);
2502 region->init_req(1, slow_ctrl);
2503 // region inputs are now complete
2504 transform_later(region);
2505 _igvn.replace_node(_callprojs->fallthrough_proj, region);
2506
2507 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2508 mem_phi->init_req(1, memproj );
2509 mem_phi->init_req(2, mem);
2510 transform_later(mem_phi);
2511
2512 _igvn.replace_node(_callprojs->fallthrough_memproj, mem_phi);
2513 }
2514
2515 // An inline type might be returned from the call but we don't know its
2516 // type. Either we get a buffered inline type (and nothing needs to be done)
2517 // or one of the values being returned is the klass of the inline type
2518 // and we need to allocate an inline type instance of that type and
2519 // initialize it with other values being returned. In that case, we
2520 // first try a fast path allocation and initialize the value with the
2521 // inline klass's pack handler or we fall back to a runtime call.
2522 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2523 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2524 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2525 if (ret == nullptr) {
2526 return;
2527 }
2528 const TypeFunc* tf = call->_tf;
2529 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2530 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2531 call->_tf = new_tf;
2532 // Make sure the change of type is applied before projections are processed by igvn
2533 _igvn.set_type(call, call->Value(&_igvn));
2534 _igvn.set_type(ret, ret->Value(&_igvn));
2535
2536 // Before any new projection is added:
2537 CallProjections* projs = call->extract_projections(true, true);
2538
2539 // Create temporary hook nodes that will be replaced below.
2540 // Add an input to prevent hook nodes from being dead.
2541 Node* ctl = new Node(call);
2542 Node* mem = new Node(ctl);
2543 Node* io = new Node(ctl);
2544 Node* ex_ctl = new Node(ctl);
2545 Node* ex_mem = new Node(ctl);
2546 Node* ex_io = new Node(ctl);
2547 Node* res = new Node(ctl);
2548
2549 // Allocate a new buffered inline type only if a new one is not returned
2550 Node* cast = transform_later(new CastP2XNode(ctl, res));
2551 Node* mask = MakeConX(0x1);
2552 Node* masked = transform_later(new AndXNode(cast, mask));
2553 Node* cmp = transform_later(new CmpXNode(masked, mask));
2554 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2555 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2556 transform_later(allocation_iff);
2557 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2558 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2559 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2560
2561 // Try to allocate a new buffered inline instance either from TLAB or eden space
2562 Node* needgc_ctrl = nullptr; // needgc means slowcase, i.e. allocation failed
2563 CallLeafNoFPNode* handler_call;
2564 const bool alloc_in_place = UseTLAB;
2565 if (alloc_in_place) {
2566 Node* fast_oop_ctrl = nullptr;
2567 Node* fast_oop_rawmem = nullptr;
2568 Node* mask2 = MakeConX(-2);
2569 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2570 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2571 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeInstKlassPtr::OBJECT_OR_NULL));
2572 Node* layout_val = make_load(nullptr, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2573 Node* size_in_bytes = ConvI2X(layout_val);
2574 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
2575 Node* fast_oop = bs->obj_allocate(this, mem, allocation_ctl, size_in_bytes, io, needgc_ctrl,
2576 fast_oop_ctrl, fast_oop_rawmem,
2577 AllocateInstancePrefetchLines);
2578 // Allocation succeed, initialize buffered inline instance header firstly,
2579 // and then initialize its fields with an inline class specific handler
2580 Node* mark_node = makecon(TypeRawPtr::make((address)markWord::inline_type_prototype().value()));
2581 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2582 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2583 if (UseCompressedClassPointers) {
2584 fast_oop_rawmem = make_store(fast_oop_ctrl, fast_oop_rawmem, fast_oop, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2585 }
2586 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);
2587 Node* pack_handler = make_load(fast_oop_ctrl, fast_oop_rawmem, fixed_block, in_bytes(InlineKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2588 handler_call = new CallLeafNoFPNode(OptoRuntime::pack_inline_type_Type(),
2589 nullptr,
2590 "pack handler",
2591 TypeRawPtr::BOTTOM);
2592 handler_call->init_req(TypeFunc::Control, fast_oop_ctrl);
2593 handler_call->init_req(TypeFunc::Memory, fast_oop_rawmem);
2594 handler_call->init_req(TypeFunc::I_O, top());
2595 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2596 handler_call->init_req(TypeFunc::ReturnAdr, top());
2597 handler_call->init_req(TypeFunc::Parms, pack_handler);
2598 handler_call->init_req(TypeFunc::Parms+1, fast_oop);
2599 } else {
2600 needgc_ctrl = allocation_ctl;
2601 }
2602
2603 // Allocation failed, fall back to a runtime call
2604 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_inline_type_fields_Type(),
2605 StubRoutines::store_inline_type_fields_to_buf(),
2606 "store_inline_type_fields",
2607 TypePtr::BOTTOM);
2608 slow_call->init_req(TypeFunc::Control, needgc_ctrl);
2609 slow_call->init_req(TypeFunc::Memory, mem);
2610 slow_call->init_req(TypeFunc::I_O, io);
2611 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2612 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2613 slow_call->init_req(TypeFunc::Parms, res);
2614
2615 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2616 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2617 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2618 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2619 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2620 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2621 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2622
2623 Node* ex_r = new RegionNode(3);
2624 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2625 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2626 ex_r->init_req(1, slow_excp);
2627 ex_mem_phi->init_req(1, slow_mem);
2628 ex_io_phi->init_req(1, slow_io);
2629 ex_r->init_req(2, ex_ctl);
2630 ex_mem_phi->init_req(2, ex_mem);
2631 ex_io_phi->init_req(2, ex_io);
2632 transform_later(ex_r);
2633 transform_later(ex_mem_phi);
2634 transform_later(ex_io_phi);
2635
2636 // We don't know how many values are returned. This assumes the
2637 // worst case, that all available registers are used.
2638 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2639 if (domain->field_at(i) == Type::HALF) {
2640 slow_call->init_req(i, top());
2641 if (alloc_in_place) {
2642 handler_call->init_req(i+1, top());
2643 }
2644 continue;
2645 }
2646 Node* proj = transform_later(new ProjNode(call, i));
2647 slow_call->init_req(i, proj);
2648 if (alloc_in_place) {
2649 handler_call->init_req(i+1, proj);
2650 }
2651 }
2652 // We can safepoint at that new call
2653 slow_call->copy_call_debug_info(&_igvn, call);
2654 transform_later(slow_call);
2655 if (alloc_in_place) {
2656 transform_later(handler_call);
2657 }
2658
2659 Node* fast_ctl = nullptr;
2660 Node* fast_res = nullptr;
2661 MergeMemNode* fast_mem = nullptr;
2662 if (alloc_in_place) {
2663 fast_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2664 Node* rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2665 fast_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2666 fast_mem = MergeMemNode::make(mem);
2667 fast_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2668 transform_later(fast_mem);
2669 }
2670
2671 Node* r = new RegionNode(alloc_in_place ? 4 : 3);
2672 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2673 Node* io_phi = new PhiNode(r, Type::ABIO);
2674 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2675 r->init_req(1, no_allocation_ctl);
2676 mem_phi->init_req(1, mem);
2677 io_phi->init_req(1, io);
2678 res_phi->init_req(1, no_allocation_res);
2679 r->init_req(2, slow_norm);
2680 mem_phi->init_req(2, slow_mem);
2681 io_phi->init_req(2, slow_io);
2682 res_phi->init_req(2, slow_res);
2683 if (alloc_in_place) {
2684 r->init_req(3, fast_ctl);
2685 mem_phi->init_req(3, fast_mem);
2686 io_phi->init_req(3, io);
2687 res_phi->init_req(3, fast_res);
2688 }
2689 transform_later(r);
2690 transform_later(mem_phi);
2691 transform_later(io_phi);
2692 transform_later(res_phi);
2693
2694 // Do not let stores that initialize this buffer be reordered with a subsequent
2695 // store that would make this buffer accessible by other threads.
2696 MemBarNode* mb = MemBarNode::make(C, Op_MemBarStoreStore, Compile::AliasIdxBot);
2697 transform_later(mb);
2698 mb->init_req(TypeFunc::Memory, mem_phi);
2699 mb->init_req(TypeFunc::Control, r);
2700 r = new ProjNode(mb, TypeFunc::Control);
2701 transform_later(r);
2702 mem_phi = new ProjNode(mb, TypeFunc::Memory);
2703 transform_later(mem_phi);
2704
2705 assert(projs->nb_resproj == 1, "unexpected number of results");
2706 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2707 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2708 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2709 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2710 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2711 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2712 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2713 // The CatchNode should not use the ex_io_phi. Re-connect it to the catchall_ioproj.
2714 Node* cn = projs->fallthrough_catchproj->in(0);
2715 _igvn.replace_input_of(cn, 1, projs->catchall_ioproj);
2716
2717 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2718 _igvn.replace_node(mem, projs->fallthrough_memproj);
2719 _igvn.replace_node(io, projs->fallthrough_ioproj);
2720 _igvn.replace_node(res, projs->resproj[0]);
2721 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2722 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2723 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2724 }
2725
2726 void PhaseMacroExpand::expand_subtypecheck_node(SubTypeCheckNode *check) {
2727 assert(check->in(SubTypeCheckNode::Control) == nullptr, "should be pinned");
2728 Node* bol = check->unique_out();
2729 Node* obj_or_subklass = check->in(SubTypeCheckNode::ObjOrSubKlass);
2730 Node* superklass = check->in(SubTypeCheckNode::SuperKlass);
2731 assert(bol->is_Bool() && bol->as_Bool()->_test._test == BoolTest::ne, "unexpected bool node");
2732
2733 for (DUIterator_Last imin, i = bol->last_outs(imin); i >= imin; --i) {
2734 Node* iff = bol->last_out(i);
2735 assert(iff->is_If(), "where's the if?");
2736
2737 if (iff->in(0)->is_top()) {
2738 _igvn.replace_input_of(iff, 1, C->top());
2739 continue;
2740 }
2741
2742 Node* iftrue = iff->as_If()->proj_out(1);
2743 Node* iffalse = iff->as_If()->proj_out(0);
2744 Node* ctrl = iff->in(0);
2745
2746 Node* subklass = nullptr;
2747 if (_igvn.type(obj_or_subklass)->isa_klassptr()) {
2748 subklass = obj_or_subklass;
2749 } else {
2750 Node* k_adr = basic_plus_adr(obj_or_subklass, oopDesc::klass_offset_in_bytes());
2751 subklass = _igvn.transform(LoadKlassNode::make(_igvn, nullptr, C->immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2752 }
2753
2754 Node* not_subtype_ctrl = Phase::gen_subtype_check(subklass, superklass, &ctrl, nullptr, _igvn, check->method(), check->bci());
2755
2756 _igvn.replace_input_of(iff, 0, C->top());
2757 _igvn.replace_node(iftrue, not_subtype_ctrl);
2758 _igvn.replace_node(iffalse, ctrl);
2759 }
2760 _igvn.replace_node(check, C->top());
2761 }
2762
2763 // FlatArrayCheckNode (array1 array2 ...) is expanded into:
2764 //
2765 // long mark = array1.mark | array2.mark | ...;
2766 // long locked_bit = markWord::unlocked_value & array1.mark & array2.mark & ...;
2767 // if (locked_bit == 0) {
2768 // // One array is locked, load prototype header from the klass
2769 // mark = array1.klass.proto | array2.klass.proto | ...
2770 // }
2771 // if ((mark & markWord::flat_array_bit_in_place) == 0) {
2772 // ...
2773 // }
2774 void PhaseMacroExpand::expand_flatarraycheck_node(FlatArrayCheckNode* check) {
2775 bool array_inputs = _igvn.type(check->in(FlatArrayCheckNode::ArrayOrKlass))->isa_oopptr() != nullptr;
2776 if (UseArrayMarkWordCheck && array_inputs) {
2777 Node* mark = MakeConX(0);
2778 Node* locked_bit = MakeConX(markWord::unlocked_value);
2779 Node* mem = check->in(FlatArrayCheckNode::Memory);
2780 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2781 Node* ary = check->in(i);
2782 const TypeOopPtr* t = _igvn.type(ary)->isa_oopptr();
2783 assert(t != nullptr, "Mixing array and klass inputs");
2784 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2785 Node* mark_adr = basic_plus_adr(ary, oopDesc::mark_offset_in_bytes());
2786 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));
2787 mark = _igvn.transform(new OrXNode(mark, mark_load));
2788 locked_bit = _igvn.transform(new AndXNode(locked_bit, mark_load));
2789 }
2790 assert(!mark->is_Con(), "Should have been optimized out");
2791 Node* cmp = _igvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
2792 Node* is_unlocked = _igvn.transform(new BoolNode(cmp, BoolTest::ne));
2793
2794 // BoolNode might be shared, replace each if user
2795 Node* old_bol = check->unique_out();
2796 assert(old_bol->is_Bool() && old_bol->as_Bool()->_test._test == BoolTest::ne, "unexpected condition");
2797 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2798 IfNode* old_iff = old_bol->last_out(i)->as_If();
2799 Node* ctrl = old_iff->in(0);
2800 RegionNode* region = new RegionNode(3);
2801 Node* mark_phi = new PhiNode(region, TypeX_X);
2802
2803 // Check if array is unlocked
2804 IfNode* iff = _igvn.transform(new IfNode(ctrl, is_unlocked, PROB_MAX, COUNT_UNKNOWN))->as_If();
2805
2806 // Unlocked: Use bits from mark word
2807 region->init_req(1, _igvn.transform(new IfTrueNode(iff)));
2808 mark_phi->init_req(1, mark);
2809
2810 // Locked: Load prototype header from klass
2811 ctrl = _igvn.transform(new IfFalseNode(iff));
2812 Node* proto = MakeConX(0);
2813 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2814 Node* ary = check->in(i);
2815 // Make loads control dependent to make sure they are only executed if array is locked
2816 Node* klass_adr = basic_plus_adr(ary, oopDesc::klass_offset_in_bytes());
2817 Node* klass = _igvn.transform(LoadKlassNode::make(_igvn, ctrl, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2818 Node* proto_adr = basic_plus_adr(klass, in_bytes(Klass::prototype_header_offset()));
2819 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));
2820 proto = _igvn.transform(new OrXNode(proto, proto_load));
2821 }
2822 region->init_req(2, ctrl);
2823 mark_phi->init_req(2, proto);
2824
2825 // Check if flat array bits are set
2826 Node* mask = MakeConX(markWord::flat_array_bit_in_place);
2827 Node* masked = _igvn.transform(new AndXNode(_igvn.transform(mark_phi), mask));
2828 cmp = _igvn.transform(new CmpXNode(masked, MakeConX(0)));
2829 Node* is_not_flat = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2830
2831 ctrl = _igvn.transform(region);
2832 iff = _igvn.transform(new IfNode(ctrl, is_not_flat, PROB_MAX, COUNT_UNKNOWN))->as_If();
2833 _igvn.replace_node(old_iff, iff);
2834 }
2835 _igvn.replace_node(check, C->top());
2836 } else {
2837 // Fall back to layout helper check
2838 Node* lhs = intcon(0);
2839 for (uint i = FlatArrayCheckNode::ArrayOrKlass; i < check->req(); ++i) {
2840 Node* array_or_klass = check->in(i);
2841 Node* klass = nullptr;
2842 const TypePtr* t = _igvn.type(array_or_klass)->is_ptr();
2843 assert(!t->is_flat() && !t->is_not_flat(), "Should have been optimized out");
2844 if (t->isa_oopptr() != nullptr) {
2845 Node* klass_adr = basic_plus_adr(array_or_klass, oopDesc::klass_offset_in_bytes());
2846 klass = transform_later(LoadKlassNode::make(_igvn, nullptr, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
2847 } else {
2848 assert(t->isa_klassptr(), "Unexpected input type");
2849 klass = array_or_klass;
2850 }
2851 Node* lh_addr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
2852 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));
2853 lhs = _igvn.transform(new OrINode(lhs, lh_val));
2854 }
2855 Node* masked = transform_later(new AndINode(lhs, intcon(Klass::_lh_array_tag_flat_value_bit_inplace)));
2856 Node* cmp = transform_later(new CmpINode(masked, intcon(0)));
2857 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2858 Node* m2b = transform_later(new Conv2BNode(masked));
2859 // The matcher expects the input to If nodes to be produced by a Bool(CmpI..)
2860 // pattern, but the input to other potential users (e.g. Phi) to be some
2861 // other pattern (e.g. a Conv2B node, possibly idealized as a CMoveI).
2862 Node* old_bol = check->unique_out();
2863 for (DUIterator_Last imin, i = old_bol->last_outs(imin); i >= imin; --i) {
2864 Node* user = old_bol->last_out(i);
2865 for (uint j = 0; j < user->req(); j++) {
2866 Node* n = user->in(j);
2867 if (n == old_bol) {
2868 _igvn.replace_input_of(user, j, user->is_If() ? bol : m2b);
2869 }
2870 }
2871 }
2872 _igvn.replace_node(check, C->top());
2873 }
2874 }
2875
2876 //---------------------------eliminate_macro_nodes----------------------
2877 // Eliminate scalar replaced allocations and associated locks.
2878 void PhaseMacroExpand::eliminate_macro_nodes() {
2879 if (C->macro_count() == 0)
2880 return;
2881 NOT_PRODUCT(int membar_before = count_MemBar(C);)
2882
2883 // Before elimination may re-mark (change to Nested or NonEscObj)
2884 // all associated (same box and obj) lock and unlock nodes.
2885 int cnt = C->macro_count();
2886 for (int i=0; i < cnt; i++) {
2887 Node *n = C->macro_node(i);
2888 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2889 mark_eliminated_locking_nodes(n->as_AbstractLock());
2890 }
2891 }
2892 // Re-marking may break consistency of Coarsened locks.
2893 if (!C->coarsened_locks_consistent()) {
2894 return; // recompile without Coarsened locks if broken
2895 }
2916 }
2917 // Next, attempt to eliminate allocations
2918 _has_locks = false;
2919 progress = true;
2920 while (progress) {
2921 progress = false;
2922 for (int i = C->macro_count(); i > 0; i = MIN2(i - 1, C->macro_count())) { // more than 1 element can be eliminated at once
2923 Node* n = C->macro_node(i - 1);
2924 bool success = false;
2925 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2926 switch (n->class_id()) {
2927 case Node::Class_Allocate:
2928 case Node::Class_AllocateArray:
2929 success = eliminate_allocate_node(n->as_Allocate());
2930 #ifndef PRODUCT
2931 if (success && PrintOptoStatistics) {
2932 Atomic::inc(&PhaseMacroExpand::_objs_scalar_replaced_counter);
2933 }
2934 #endif
2935 break;
2936 case Node::Class_CallStaticJava: {
2937 CallStaticJavaNode* call = n->as_CallStaticJava();
2938 if (!call->method()->is_method_handle_intrinsic()) {
2939 success = eliminate_boxing_node(n->as_CallStaticJava());
2940 }
2941 break;
2942 }
2943 case Node::Class_Lock:
2944 case Node::Class_Unlock:
2945 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2946 _has_locks = true;
2947 break;
2948 case Node::Class_ArrayCopy:
2949 break;
2950 case Node::Class_OuterStripMinedLoop:
2951 break;
2952 case Node::Class_SubTypeCheck:
2953 break;
2954 case Node::Class_Opaque1:
2955 break;
2956 case Node::Class_FlatArrayCheck:
2957 break;
2958 default:
2959 assert(n->Opcode() == Op_LoopLimit ||
2960 n->Opcode() == Op_Opaque3 ||
2961 n->Opcode() == Op_Opaque4 ||
2962 n->Opcode() == Op_MaxL ||
2963 n->Opcode() == Op_MinL ||
2964 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2965 "unknown node type in macro list");
2966 }
2967 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2968 progress = progress || success;
2969 }
2970 }
2971 #ifndef PRODUCT
2972 if (PrintOptoStatistics) {
2973 int membar_after = count_MemBar(C);
2974 Atomic::add(&PhaseMacroExpand::_memory_barriers_removed_counter, membar_before - membar_after);
2975 }
2976 #endif
2977 }
2980 // Returns true if a failure occurred.
2981 bool PhaseMacroExpand::expand_macro_nodes() {
2982 // Last attempt to eliminate macro nodes.
2983 eliminate_macro_nodes();
2984 if (C->failing()) return true;
2985
2986 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2987 bool progress = true;
2988 while (progress) {
2989 progress = false;
2990 for (int i = C->macro_count(); i > 0; i--) {
2991 Node* n = C->macro_node(i-1);
2992 bool success = false;
2993 DEBUG_ONLY(int old_macro_count = C->macro_count();)
2994 if (n->Opcode() == Op_LoopLimit) {
2995 // Remove it from macro list and put on IGVN worklist to optimize.
2996 C->remove_macro_node(n);
2997 _igvn._worklist.push(n);
2998 success = true;
2999 } else if (n->Opcode() == Op_CallStaticJava) {
3000 CallStaticJavaNode* call = n->as_CallStaticJava();
3001 if (!call->method()->is_method_handle_intrinsic()) {
3002 // Remove it from macro list and put on IGVN worklist to optimize.
3003 C->remove_macro_node(n);
3004 _igvn._worklist.push(n);
3005 success = true;
3006 }
3007 } else if (n->is_Opaque1()) {
3008 _igvn.replace_node(n, n->in(1));
3009 success = true;
3010 #if INCLUDE_RTM_OPT
3011 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
3012 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
3013 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
3014 Node* cmp = n->unique_out();
3015 #ifdef ASSERT
3016 // Validate graph.
3017 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
3018 BoolNode* bol = cmp->unique_out()->as_Bool();
3019 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
3020 (bol->_test._test == BoolTest::ne), "");
3021 IfNode* ifn = bol->unique_out()->as_If();
3022 assert((ifn->outcnt() == 2) &&
3023 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != nullptr, "");
3024 #endif
3025 Node* repl = n->in(1);
3026 if (!_has_locks) {
3101 // Worst case is a macro node gets expanded into about 200 nodes.
3102 // Allow 50% more for optimization.
3103 if (C->check_node_count(300, "out of nodes before macro expansion")) {
3104 return true;
3105 }
3106
3107 DEBUG_ONLY(int old_macro_count = C->macro_count();)
3108 switch (n->class_id()) {
3109 case Node::Class_Lock:
3110 expand_lock_node(n->as_Lock());
3111 break;
3112 case Node::Class_Unlock:
3113 expand_unlock_node(n->as_Unlock());
3114 break;
3115 case Node::Class_ArrayCopy:
3116 expand_arraycopy_node(n->as_ArrayCopy());
3117 break;
3118 case Node::Class_SubTypeCheck:
3119 expand_subtypecheck_node(n->as_SubTypeCheck());
3120 break;
3121 case Node::Class_CallStaticJava:
3122 expand_mh_intrinsic_return(n->as_CallStaticJava());
3123 C->remove_macro_node(n);
3124 break;
3125 case Node::Class_FlatArrayCheck:
3126 expand_flatarraycheck_node(n->as_FlatArrayCheck());
3127 break;
3128 default:
3129 assert(false, "unknown node type in macro list");
3130 }
3131 assert(C->macro_count() == (old_macro_count - 1), "expansion must have deleted one node from macro list");
3132 if (C->failing()) return true;
3133
3134 // Clean up the graph so we're less likely to hit the maximum node
3135 // limit
3136 _igvn.set_delay_transform(false);
3137 _igvn.optimize();
3138 if (C->failing()) return true;
3139 _igvn.set_delay_transform(true);
3140 }
3141
3142 // All nodes except Allocate nodes are expanded now. There could be
3143 // new optimization opportunities (such as folding newly created
3144 // load from a just allocated object). Run IGVN.
3145
3146 // expand "macro" nodes
3147 // nodes are removed from the macro list as they are processed
|