30 #include "opto/addnode.hpp"
31 #include "opto/arraycopynode.hpp"
32 #include "opto/callnode.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/compile.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/locknode.hpp"
39 #include "opto/loopnode.hpp"
40 #include "opto/macro.hpp"
41 #include "opto/memnode.hpp"
42 #include "opto/narrowptrnode.hpp"
43 #include "opto/node.hpp"
44 #include "opto/opaquenode.hpp"
45 #include "opto/phaseX.hpp"
46 #include "opto/rootnode.hpp"
47 #include "opto/runtime.hpp"
48 #include "opto/subnode.hpp"
49 #include "opto/type.hpp"
50 #include "runtime/sharedRuntime.hpp"
51 #include "utilities/macros.hpp"
52 #if INCLUDE_G1GC
53 #include "gc/g1/g1ThreadLocalData.hpp"
54 #endif // INCLUDE_G1GC
55 #if INCLUDE_SHENANDOAHGC
56 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
57 #endif
58
59
60 //
61 // Replace any references to "oldref" in inputs to "use" with "newref".
62 // Returns the number of replacements made.
63 //
64 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
65 int nreplacements = 0;
66 uint req = use->req();
67 for (uint j = 0; j < use->len(); j++) {
68 Node *uin = use->in(j);
69 if (uin == oldref) {
70 if (j < req)
71 use->set_req(j, newref);
72 else
73 use->set_prec(j, newref);
74 nreplacements++;
75 } else if (j >= req && uin == NULL) {
76 break;
77 }
78 }
79 return nreplacements;
80 }
81
82 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
83 // Copy debug information and adjust JVMState information
84 uint old_dbg_start = oldcall->tf()->domain()->cnt();
85 uint new_dbg_start = newcall->tf()->domain()->cnt();
86 int jvms_adj = new_dbg_start - old_dbg_start;
87 assert (new_dbg_start == newcall->req(), "argument count mismatch");
88
89 // SafePointScalarObject node could be referenced several times in debug info.
90 // Use Dict to record cloned nodes.
91 Dict* sosn_map = new Dict(cmpkey,hashkey);
92 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
93 Node* old_in = oldcall->in(i);
94 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
95 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
96 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
97 uint old_unique = C->unique();
98 Node* new_in = old_sosn->clone(sosn_map);
99 if (old_unique != C->unique()) { // New node?
100 new_in->set_req(0, C->root()); // reset control edge
101 new_in = transform_later(new_in); // Register new node.
102 }
103 old_in = new_in;
104 }
105 newcall->add_req(old_in);
260 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
261 return in;
262 }
263 }
264 mem = in->in(TypeFunc::Memory);
265 } else if (in->is_MemBar()) {
266 ArrayCopyNode* ac = NULL;
267 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
268 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
269 return ac;
270 }
271 mem = in->in(TypeFunc::Memory);
272 } else {
273 assert(false, "unexpected projection");
274 }
275 } else if (mem->is_Store()) {
276 const TypePtr* atype = mem->as_Store()->adr_type();
277 int adr_idx = phase->C->get_alias_index(atype);
278 if (adr_idx == alias_idx) {
279 assert(atype->isa_oopptr(), "address type must be oopptr");
280 int adr_offset = atype->offset();
281 uint adr_iid = atype->is_oopptr()->instance_id();
282 // Array elements references have the same alias_idx
283 // but different offset and different instance_id.
284 if (adr_offset == offset && adr_iid == alloc->_idx)
285 return mem;
286 } else {
287 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
288 }
289 mem = mem->in(MemNode::Memory);
290 } else if (mem->is_ClearArray()) {
291 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
292 // Can not bypass initialization of the instance
293 // we are looking.
294 debug_only(intptr_t offset;)
295 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
296 InitializeNode* init = alloc->as_Allocate()->initialization();
297 // We are looking for stored value, return Initialize node
298 // or memory edge from Allocate node.
299 if (init != NULL)
300 return init;
303 }
304 // Otherwise skip it (the call updated 'mem' value).
305 } else if (mem->Opcode() == Op_SCMemProj) {
306 mem = mem->in(0);
307 Node* adr = NULL;
308 if (mem->is_LoadStore()) {
309 adr = mem->in(MemNode::Address);
310 } else {
311 assert(mem->Opcode() == Op_EncodeISOArray ||
312 mem->Opcode() == Op_StrCompressedCopy, "sanity");
313 adr = mem->in(3); // Destination array
314 }
315 const TypePtr* atype = adr->bottom_type()->is_ptr();
316 int adr_idx = phase->C->get_alias_index(atype);
317 if (adr_idx == alias_idx) {
318 DEBUG_ONLY(mem->dump();)
319 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
320 return NULL;
321 }
322 mem = mem->in(MemNode::Memory);
323 } else if (mem->Opcode() == Op_StrInflatedCopy) {
324 Node* adr = mem->in(3); // Destination array
325 const TypePtr* atype = adr->bottom_type()->is_ptr();
326 int adr_idx = phase->C->get_alias_index(atype);
327 if (adr_idx == alias_idx) {
328 DEBUG_ONLY(mem->dump();)
329 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
330 return NULL;
331 }
332 mem = mem->in(MemNode::Memory);
333 } else {
334 return mem;
335 }
336 assert(mem != orig_mem, "dead memory loop");
337 }
338 }
339
340 // Generate loads from source of the arraycopy for fields of
341 // destination needed at a deoptimization point
342 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
343 BasicType bt = ft;
344 const Type *type = ftype;
345 if (ft == T_NARROWOOP) {
346 bt = T_OBJECT;
347 type = ftype->make_oopptr();
348 }
349 Node* res = NULL;
350 if (ac->is_clonebasic()) {
351 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
352 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
353 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
354 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
355 } else {
356 if (ac->modifies(offset, offset, &_igvn, true)) {
357 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
358 uint shift = exact_log2(type2aelembytes(bt));
359 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
360 #ifdef _LP64
361 diff = _igvn.transform(new ConvI2LNode(diff));
362 #endif
363 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
364
365 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
366 Node* base = ac->in(ArrayCopyNode::Src);
367 Node* adr = _igvn.transform(new AddPNode(base, base, off));
368 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
369 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
370 }
371 }
372 if (res != NULL) {
373 res = _igvn.transform(res);
374 if (ftype->isa_narrowoop()) {
375 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
376 res = _igvn.transform(new EncodePNode(res, ftype));
377 }
378 return res;
379 }
380 return NULL;
381 }
382
383 //
384 // Given a Memory Phi, compute a value Phi containing the values from stores
385 // on the input paths.
386 // Note: this function is recursive, its depth is limited by the "level" argument
387 // Returns the computed Phi, or NULL if it cannot compute it.
388 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) {
389 assert(mem->is_Phi(), "sanity");
390 int alias_idx = C->get_alias_index(adr_t);
391 int offset = adr_t->offset();
392 int instance_id = adr_t->instance_id();
393
394 // Check if an appropriate value phi already exists.
395 Node* region = mem->in(0);
396 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
397 Node* phi = region->fast_out(k);
398 if (phi->is_Phi() && phi != mem &&
399 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
400 return phi;
401 }
402 }
403 // Check if an appropriate new value phi already exists.
404 Node* new_phi = value_phis->find(mem->_idx);
405 if (new_phi != NULL)
406 return new_phi;
407
408 if (level <= 0) {
409 return NULL; // Give up: phi tree too deep
410 }
411 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
412 Node *alloc_mem = alloc->in(TypeFunc::Memory);
413
414 uint length = mem->req();
415 GrowableArray <Node *> values(length, length, NULL, false);
416
417 // create a new Phi for the value
418 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
419 transform_later(phi);
420 value_phis->push(phi, mem->_idx);
421
422 for (uint j = 1; j < length; j++) {
423 Node *in = mem->in(j);
424 if (in == NULL || in->is_top()) {
425 values.at_put(j, in);
426 } else {
427 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
428 if (val == start_mem || val == alloc_mem) {
429 // hit a sentinel, return appropriate 0 value
430 values.at_put(j, _igvn.zerocon(ft));
431 continue;
432 }
433 if (val->is_Initialize()) {
434 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
435 }
436 if (val == NULL) {
437 return NULL; // can't find a value on this path
438 }
439 if (val == mem) {
440 values.at_put(j, mem);
441 } else if (val->is_Store()) {
442 Node* n = val->in(MemNode::ValueIn);
443 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
444 n = bs->step_over_gc_barrier(n);
445 values.at_put(j, n);
446 } else if(val->is_Proj() && val->in(0) == alloc) {
447 values.at_put(j, _igvn.zerocon(ft));
448 } else if (val->is_Phi()) {
449 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
450 if (val == NULL) {
451 return NULL;
452 }
453 values.at_put(j, val);
454 } else if (val->Opcode() == Op_SCMemProj) {
455 assert(val->in(0)->is_LoadStore() ||
456 val->in(0)->Opcode() == Op_EncodeISOArray ||
457 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
458 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
459 return NULL;
460 } else if (val->is_ArrayCopy()) {
461 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
462 if (res == NULL) {
463 return NULL;
464 }
465 values.at_put(j, res);
466 } else {
467 #ifdef ASSERT
473 }
474 }
475 // Set Phi's inputs
476 for (uint j = 1; j < length; j++) {
477 if (values.at(j) == mem) {
478 phi->init_req(j, phi);
479 } else {
480 phi->init_req(j, values.at(j));
481 }
482 }
483 return phi;
484 }
485
486 // Search the last value stored into the object's field.
487 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
488 assert(adr_t->is_known_instance_field(), "instance required");
489 int instance_id = adr_t->instance_id();
490 assert((uint)instance_id == alloc->_idx, "wrong allocation");
491
492 int alias_idx = C->get_alias_index(adr_t);
493 int offset = adr_t->offset();
494 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
495 Node *alloc_ctrl = alloc->in(TypeFunc::Control);
496 Node *alloc_mem = alloc->in(TypeFunc::Memory);
497 Arena *a = Thread::current()->resource_area();
498 VectorSet visited(a);
499
500
501 bool done = sfpt_mem == alloc_mem;
502 Node *mem = sfpt_mem;
503 while (!done) {
504 if (visited.test_set(mem->_idx)) {
505 return NULL; // found a loop, give up
506 }
507 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
508 if (mem == start_mem || mem == alloc_mem) {
509 done = true; // hit a sentinel, return appropriate 0 value
510 } else if (mem->is_Initialize()) {
511 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
512 if (mem == NULL) {
513 done = true; // Something go wrong.
514 } else if (mem->is_Store()) {
515 const TypePtr* atype = mem->as_Store()->adr_type();
516 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
517 done = true;
518 }
519 } else if (mem->is_Store()) {
520 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
521 assert(atype != NULL, "address type must be oopptr");
522 assert(C->get_alias_index(atype) == alias_idx &&
523 atype->is_known_instance_field() && atype->offset() == offset &&
524 atype->instance_id() == instance_id, "store is correct memory slice");
525 done = true;
526 } else if (mem->is_Phi()) {
527 // try to find a phi's unique input
528 Node *unique_input = NULL;
529 Node *top = C->top();
530 for (uint i = 1; i < mem->req(); i++) {
531 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
532 if (n == NULL || n == top || n == mem) {
533 continue;
534 } else if (unique_input == NULL) {
535 unique_input = n;
536 } else if (unique_input != n) {
537 unique_input = top;
538 break;
539 }
540 }
541 if (unique_input != NULL && unique_input != top) {
542 mem = unique_input;
543 } else {
544 done = true;
545 }
546 } else if (mem->is_ArrayCopy()) {
547 done = true;
548 } else {
549 assert(false, "unexpected node");
550 }
551 }
552 if (mem != NULL) {
553 if (mem == start_mem || mem == alloc_mem) {
554 // hit a sentinel, return appropriate 0 value
555 return _igvn.zerocon(ft);
556 } else if (mem->is_Store()) {
557 Node* n = mem->in(MemNode::ValueIn);
558 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
559 n = bs->step_over_gc_barrier(n);
560 return n;
561 } else if (mem->is_Phi()) {
562 // attempt to produce a Phi reflecting the values on the input paths of the Phi
563 Node_Stack value_phis(a, 8);
564 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
565 if (phi != NULL) {
566 return phi;
567 } else {
568 // Kill all new Phis
569 while(value_phis.is_nonempty()) {
570 Node* n = value_phis.node();
571 _igvn.replace_node(n, C->top());
572 value_phis.pop();
573 }
574 }
575 } else if (mem->is_ArrayCopy()) {
576 Node* ctl = mem->in(0);
577 Node* m = mem->in(TypeFunc::Memory);
578 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
579 // pin the loads in the uncommon trap path
580 ctl = sfpt_ctl;
581 m = sfpt_mem;
582 }
583 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
584 }
585 }
586 // Something go wrong.
587 return NULL;
588 }
589
590 // Check the possibility of scalar replacement.
591 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
592 // Scan the uses of the allocation to check for anything that would
593 // prevent us from eliminating it.
594 NOT_PRODUCT( const char* fail_eliminate = NULL; )
595 DEBUG_ONLY( Node* disq_node = NULL; )
596 bool can_eliminate = true;
597
598 Node* res = alloc->result_cast();
599 const TypeOopPtr* res_type = NULL;
600 if (res == NULL) {
601 // All users were eliminated.
602 } else if (!res->is_CheckCastPP()) {
603 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
604 can_eliminate = false;
605 } else {
606 res_type = _igvn.type(res)->isa_oopptr();
607 if (res_type == NULL) {
608 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
609 can_eliminate = false;
625 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
626 int offset = addp_type->offset();
627
628 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
629 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
630 can_eliminate = false;
631 break;
632 }
633 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
634 k < kmax && can_eliminate; k++) {
635 Node* n = use->fast_out(k);
636 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
637 SHENANDOAHGC_ONLY((!UseShenandoahGC || !ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(n)) &&)
638 !(n->is_ArrayCopy() &&
639 n->as_ArrayCopy()->is_clonebasic() &&
640 n->in(ArrayCopyNode::Dest) == use)) {
641 DEBUG_ONLY(disq_node = n;)
642 if (n->is_Load() || n->is_LoadStore()) {
643 NOT_PRODUCT(fail_eliminate = "Field load";)
644 } else {
645 NOT_PRODUCT(fail_eliminate = "Not store field referrence";)
646 }
647 can_eliminate = false;
648 }
649 }
650 } else if (use->is_ArrayCopy() &&
651 (use->as_ArrayCopy()->is_arraycopy_validated() ||
652 use->as_ArrayCopy()->is_copyof_validated() ||
653 use->as_ArrayCopy()->is_copyofrange_validated()) &&
654 use->in(ArrayCopyNode::Dest) == res) {
655 // ok to eliminate
656 } else if (use->is_SafePoint()) {
657 SafePointNode* sfpt = use->as_SafePoint();
658 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
659 // Object is passed as argument.
660 DEBUG_ONLY(disq_node = use;)
661 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
662 can_eliminate = false;
663 }
664 Node* sfptMem = sfpt->memory();
665 if (sfptMem == NULL || sfptMem->is_top()) {
666 DEBUG_ONLY(disq_node = use;)
667 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
668 can_eliminate = false;
669 } else {
670 safepoints.append_if_missing(sfpt);
671 }
672 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
673 if (use->is_Phi()) {
674 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
675 NOT_PRODUCT(fail_eliminate = "Object is return value";)
676 } else {
677 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
678 }
679 DEBUG_ONLY(disq_node = use;)
680 } else {
681 if (use->Opcode() == Op_Return) {
682 NOT_PRODUCT(fail_eliminate = "Object is return value";)
683 }else {
684 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
685 }
686 DEBUG_ONLY(disq_node = use;)
687 }
688 can_eliminate = false;
689 }
690 }
691 }
692
693 #ifndef PRODUCT
694 if (PrintEliminateAllocations) {
695 if (can_eliminate) {
696 tty->print("Scalar ");
697 if (res == NULL)
698 alloc->dump();
699 else
700 res->dump();
701 } else if (alloc->_is_scalar_replaceable) {
702 tty->print("NotScalar (%s)", fail_eliminate);
703 if (res == NULL)
704 alloc->dump();
705 else
706 res->dump();
707 #ifdef ASSERT
708 if (disq_node != NULL) {
731 Node* res = alloc->result_cast();
732 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
733 const TypeOopPtr* res_type = NULL;
734 if (res != NULL) { // Could be NULL when there are no users
735 res_type = _igvn.type(res)->isa_oopptr();
736 }
737
738 if (res != NULL) {
739 klass = res_type->klass();
740 if (res_type->isa_instptr()) {
741 // find the fields of the class which will be needed for safepoint debug information
742 assert(klass->is_instance_klass(), "must be an instance klass.");
743 iklass = klass->as_instance_klass();
744 nfields = iklass->nof_nonstatic_fields();
745 } else {
746 // find the array's elements which will be needed for safepoint debug information
747 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
748 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
749 elem_type = klass->as_array_klass()->element_type();
750 basic_elem_type = elem_type->basic_type();
751 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
752 element_size = type2aelembytes(basic_elem_type);
753 }
754 }
755 //
756 // Process the safepoint uses
757 //
758 while (safepoints.length() > 0) {
759 SafePointNode* sfpt = safepoints.pop();
760 Node* mem = sfpt->memory();
761 Node* ctl = sfpt->control();
762 assert(sfpt->jvms() != NULL, "missed JVMS");
763 // Fields of scalar objs are referenced only at the end
764 // of regular debuginfo at the last (youngest) JVMS.
765 // Record relative start index.
766 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
767 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
768 #ifdef ASSERT
769 alloc,
770 #endif
771 first_ind, nfields);
772 sobj->init_req(0, C->root());
773 transform_later(sobj);
774
775 // Scan object's fields adding an input to the safepoint for each field.
776 for (int j = 0; j < nfields; j++) {
777 intptr_t offset;
778 ciField* field = NULL;
779 if (iklass != NULL) {
780 field = iklass->nonstatic_field_at(j);
781 offset = field->offset();
782 elem_type = field->type();
783 basic_elem_type = field->layout_type();
784 } else {
785 offset = array_base + j * (intptr_t)element_size;
786 }
787
788 const Type *field_type;
789 // The next code is taken from Parse::do_get_xxx().
790 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
791 if (!elem_type->is_loaded()) {
792 field_type = TypeInstPtr::BOTTOM;
793 } else if (field != NULL && field->is_static_constant()) {
794 // This can happen if the constant oop is non-perm.
795 ciObject* con = field->constant_value().as_object();
796 // Do not "join" in the previous type; it doesn't add value,
797 // and may yield a vacuous result if the field is of interface type.
798 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
799 assert(field_type != NULL, "field singleton type must be consistent");
800 } else {
801 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
802 }
803 if (UseCompressedOops) {
804 field_type = field_type->make_narrowoop();
805 basic_elem_type = T_NARROWOOP;
806 }
807 } else {
808 field_type = Type::get_const_basic_type(basic_elem_type);
809 }
810
811 const TypeOopPtr *field_addr_type = res_type->add_offset(offset)->isa_oopptr();
812
813 Node *field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
814 if (field_val == NULL) {
815 // We weren't able to find a value for this field,
816 // give up on eliminating this allocation.
817
818 // Remove any extra entries we added to the safepoint.
819 uint last = sfpt->req() - 1;
820 for (int k = 0; k < j; k++) {
821 sfpt->del_req(last--);
822 }
823 _igvn._worklist.push(sfpt);
824 // rollback processed safepoints
825 while (safepoints_done.length() > 0) {
826 SafePointNode* sfpt_done = safepoints_done.pop();
827 // remove any extra entries we added to the safepoint
828 last = sfpt_done->req() - 1;
829 for (int k = 0; k < nfields; k++) {
830 sfpt_done->del_req(last--);
831 }
832 JVMState *jvms = sfpt_done->jvms();
833 jvms->set_endoff(sfpt_done->req());
859 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
860 sfpt->_idx, j);
861 }
862 tty->print(", which prevents elimination of: ");
863 if (res == NULL)
864 alloc->dump();
865 else
866 res->dump();
867 }
868 #endif
869 return false;
870 }
871 if (UseCompressedOops && field_type->isa_narrowoop()) {
872 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
873 // to be able scalar replace the allocation.
874 if (field_val->is_EncodeP()) {
875 field_val = field_val->in(1);
876 } else {
877 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
878 }
879 }
880 sfpt->add_req(field_val);
881 }
882 JVMState *jvms = sfpt->jvms();
883 jvms->set_endoff(sfpt->req());
884 // Now make a pass over the debug information replacing any references
885 // to the allocated object with "sobj"
886 int start = jvms->debug_start();
887 int end = jvms->debug_end();
888 sfpt->replace_edges_in_range(res, sobj, start, end);
889 _igvn._worklist.push(sfpt);
890 safepoints_done.append_if_missing(sfpt); // keep it for rollback
891 }
892 return true;
893 }
894
895 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
896 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
897 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
898 if (ctl_proj != NULL) {
899 igvn.replace_node(ctl_proj, n->in(0));
900 }
901 if (mem_proj != NULL) {
902 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
903 }
904 }
905
906 // Process users of eliminated allocation.
907 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
908 Node* res = alloc->result_cast();
909 if (res != NULL) {
910 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
911 Node *use = res->last_out(j);
936 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
937 disconnect_projections(ac, _igvn);
938 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
939 Node* membar_before = alloc->in(0)->in(0);
940 disconnect_projections(membar_before->as_MemBar(), _igvn);
941 if (membar_after->is_MemBar()) {
942 disconnect_projections(membar_after->as_MemBar(), _igvn);
943 }
944 } else {
945 eliminate_gc_barrier(n);
946 }
947 k -= (oc2 - use->outcnt());
948 }
949 _igvn.remove_dead_node(use);
950 } else if (use->is_ArrayCopy()) {
951 // Disconnect ArrayCopy node
952 ArrayCopyNode* ac = use->as_ArrayCopy();
953 assert(ac->is_arraycopy_validated() ||
954 ac->is_copyof_validated() ||
955 ac->is_copyofrange_validated(), "unsupported");
956 CallProjections callprojs;
957 ac->extract_projections(&callprojs, true);
958
959 _igvn.replace_node(callprojs.fallthrough_ioproj, ac->in(TypeFunc::I_O));
960 _igvn.replace_node(callprojs.fallthrough_memproj, ac->in(TypeFunc::Memory));
961 _igvn.replace_node(callprojs.fallthrough_catchproj, ac->in(TypeFunc::Control));
962
963 // Set control to top. IGVN will remove the remaining projections
964 ac->set_req(0, top());
965 ac->replace_edge(res, top());
966
967 // Disconnect src right away: it can help find new
968 // opportunities for allocation elimination
969 Node* src = ac->in(ArrayCopyNode::Src);
970 ac->replace_edge(src, top());
971 // src can be top at this point if src and dest of the
972 // arraycopy were the same
973 if (src->outcnt() == 0 && !src->is_top()) {
974 _igvn.remove_dead_node(src);
975 }
976
977 _igvn._worklist.push(ac);
978 } else {
979 eliminate_gc_barrier(use);
980 }
981 j -= (oc1 - res->outcnt());
982 }
983 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
984 _igvn.remove_dead_node(res);
985 }
986
987 //
988 // Process other users of allocation's projections
989 //
990 if (_resproj != NULL && _resproj->outcnt() != 0) {
991 // First disconnect stores captured by Initialize node.
992 // If Initialize node is eliminated first in the following code,
993 // it will kill such stores and DUIterator_Last will assert.
994 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
995 Node *use = _resproj->fast_out(j);
996 if (use->is_AddP()) {
997 // raw memory addresses used only by the initialization
1109 if (alloc->is_AllocateArray())
1110 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1111 else
1112 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1113 }
1114 #endif
1115
1116 return true;
1117 }
1118
1119 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1120 // EA should remove all uses of non-escaping boxing node.
1121 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1122 return false;
1123 }
1124
1125 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1126
1127 extract_call_projections(boxing);
1128
1129 const TypeTuple* r = boxing->tf()->range();
1130 assert(r->cnt() > TypeFunc::Parms, "sanity");
1131 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1132 assert(t != NULL, "sanity");
1133
1134 CompileLog* log = C->log();
1135 if (log != NULL) {
1136 log->head("eliminate_boxing type='%d'",
1137 log->identify(t->klass()));
1138 JVMState* p = boxing->jvms();
1139 while (p != NULL) {
1140 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1141 p = p->caller();
1142 }
1143 log->tail("eliminate_boxing");
1144 }
1145
1146 process_users_of_allocation(boxing);
1147
1148 #ifndef PRODUCT
1149 if (PrintEliminateAllocations) {
1269 Node *result_phi_i_o = NULL;
1270
1271 // The initial slow comparison is a size check, the comparison
1272 // we want to do is a BoolTest::gt
1273 bool always_slow = false;
1274 int tv = _igvn.find_int_con(initial_slow_test, -1);
1275 if (tv >= 0) {
1276 always_slow = (tv == 1);
1277 initial_slow_test = NULL;
1278 } else {
1279 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1280 }
1281
1282 if (C->env()->dtrace_alloc_probes() ||
1283 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1284 // Force slow-path allocation
1285 always_slow = true;
1286 initial_slow_test = NULL;
1287 }
1288
1289
1290 enum { too_big_or_final_path = 1, need_gc_path = 2 };
1291 Node *slow_region = NULL;
1292 Node *toobig_false = ctrl;
1293
1294 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1295 // generate the initial test if necessary
1296 if (initial_slow_test != NULL ) {
1297 slow_region = new RegionNode(3);
1298
1299 // Now make the initial failure test. Usually a too-big test but
1300 // might be a TRUE for finalizers or a fancy class check for
1301 // newInstance0.
1302 IfNode *toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1303 transform_later(toobig_iff);
1304 // Plug the failing-too-big test into the slow-path region
1305 Node *toobig_true = new IfTrueNode( toobig_iff );
1306 transform_later(toobig_true);
1307 slow_region ->init_req( too_big_or_final_path, toobig_true );
1308 toobig_false = new IfFalseNode( toobig_iff );
1309 transform_later(toobig_false);
1310 } else { // No initial test, just fall into next case
1311 toobig_false = ctrl;
1312 debug_only(slow_region = NodeSentinel);
1313 }
1314
1315 Node *slow_mem = mem; // save the current memory state for slow path
1316 // generate the fast allocation code unless we know that the initial test will always go slow
1317 if (!always_slow) {
1318 // Fast path modifies only raw memory.
1319 if (mem->is_MergeMem()) {
1320 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1321 }
1322
1323 // allocate the Region and Phi nodes for the result
1324 result_region = new RegionNode(3);
1325 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1326 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1327 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1328
1329 // Grab regular I/O before optional prefetch may change it.
1330 // Slow-path does no I/O so just set it to the original I/O.
1331 result_phi_i_o->init_req(slow_result_path, i_o);
1332
1333 Node* needgc_ctrl = NULL;
1334 // Name successful fast-path variables
1335 Node* fast_oop_ctrl;
1336 Node* fast_oop_rawmem;
1337
1338 intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1339
1340 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1341 Node* fast_oop = bs->obj_allocate(this, ctrl, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1342 fast_oop_ctrl, fast_oop_rawmem,
1343 prefetch_lines);
1344
1345 if (initial_slow_test) {
1346 slow_region->init_req(need_gc_path, needgc_ctrl);
1347 // This completes all paths into the slow merge point
1348 transform_later(slow_region);
1349 } else { // No initial slow path needed!
1350 // Just fall from the need-GC path straight into the VM call.
1351 slow_region = needgc_ctrl;
1352 }
1353
1354 InitializeNode* init = alloc->initialization();
1355 fast_oop_rawmem = initialize_object(alloc,
1356 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1357 klass_node, length, size_in_bytes);
1358
1359 // If initialization is performed by an array copy, any required
1360 // MemBarStoreStore was already added. If the object does not
1361 // escape no need for a MemBarStoreStore. If the object does not
1362 // escape in its initializer and memory barrier (MemBarStoreStore or
1363 // stronger) is already added at exit of initializer, also no need
1364 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1365 // so that stores that initialize this object can't be reordered
1366 // with a subsequent store that makes this object accessible by
1367 // other threads.
1368 // Other threads include java threads and JVM internal threads
1369 // (for example concurrent GC threads). Current concurrent GC
1564 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1565 _igvn.remove_dead_node(_ioproj_catchall);
1566 }
1567
1568 // if we generated only a slow call, we are done
1569 if (always_slow) {
1570 // Now we can unhook i_o.
1571 if (result_phi_i_o->outcnt() > 1) {
1572 call->set_req(TypeFunc::I_O, top());
1573 } else {
1574 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1575 // Case of new array with negative size known during compilation.
1576 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1577 // following code since call to runtime will throw exception.
1578 // As result there will be no users of i_o after the call.
1579 // Leave i_o attached to this call to avoid problems in preceding graph.
1580 }
1581 return;
1582 }
1583
1584
1585 if (_fallthroughcatchproj != NULL) {
1586 ctrl = _fallthroughcatchproj->clone();
1587 transform_later(ctrl);
1588 _igvn.replace_node(_fallthroughcatchproj, result_region);
1589 } else {
1590 ctrl = top();
1591 }
1592 Node *slow_result;
1593 if (_resproj == NULL) {
1594 // no uses of the allocation result
1595 slow_result = top();
1596 } else {
1597 slow_result = _resproj->clone();
1598 transform_later(slow_result);
1599 _igvn.replace_node(_resproj, result_phi_rawoop);
1600 }
1601
1602 // Plug slow-path into result merge point
1603 result_region ->init_req( slow_result_path, ctrl );
1604 result_phi_rawoop->init_req( slow_result_path, slow_result);
1605 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1606 transform_later(result_region);
1607 transform_later(result_phi_rawoop);
1608 transform_later(result_phi_rawmem);
1609 transform_later(result_phi_i_o);
1610 // This completes all paths into the result merge point
1611 }
1612
1613
1614 // Helper for PhaseMacroExpand::expand_allocate_common.
1615 // Initializes the newly-allocated storage.
1616 Node*
1617 PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1618 Node* control, Node* rawmem, Node* object,
1619 Node* klass_node, Node* length,
1620 Node* size_in_bytes) {
1621 InitializeNode* init = alloc->initialization();
1622 // Store the klass & mark bits
1623 Node* mark_node = NULL;
1624 // For now only enable fast locking for non-array types
1625 if (UseBiasedLocking && (length == NULL)) {
1626 mark_node = make_load(control, rawmem, klass_node, in_bytes(Klass::prototype_header_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
1627 } else {
1628 mark_node = makecon(TypeRawPtr::make((address)markOopDesc::prototype()));
1629 }
1630 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
1631
1632 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
1633 int header_size = alloc->minimum_header_size(); // conservatively small
1634
1635 // Array length
1636 if (length != NULL) { // Arrays need length field
1637 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1638 // conservatively small header size:
1639 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1640 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1641 if (k->is_array_klass()) // we know the exact header size in most cases:
1642 header_size = Klass::layout_helper_header_size(k->layout_helper());
1643 }
1644
1645 // Clear the object body, if necessary.
1646 if (init == NULL) {
1647 // The init has somehow disappeared; be cautious and clear everything.
1648 //
1649 // This can happen if a node is allocated but an uncommon trap occurs
1650 // immediately. In this case, the Initialize gets associated with the
1651 // trap, and may be placed in a different (outer) loop, if the Allocate
1652 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1653 // there can be two Allocates to one Initialize. The answer in all these
1654 // edge cases is safety first. It is always safe to clear immediately
1655 // within an Allocate, and then (maybe or maybe not) clear some more later.
1656 if (!(UseTLAB && ZeroTLAB)) {
1657 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1658 header_size, size_in_bytes,
1659 &_igvn);
1660 }
1661 } else {
1662 if (!init->is_complete()) {
1663 // Try to win by zeroing only what the init does not store.
1664 // We can also try to do some peephole optimizations,
1665 // such as combining some adjacent subword stores.
1666 rawmem = init->complete_stores(control, rawmem, object,
1667 header_size, size_in_bytes, &_igvn);
1668 }
1669 // We have no more use for this link, since the AllocateNode goes away:
1670 init->set_req(InitializeNode::RawAddress, top());
1671 // (If we keep the link, it just confuses the register allocator,
1672 // who thinks he sees a real use of the address by the membar.)
1673 }
1674
1675 return rawmem;
1676 }
1677
2018 // Replace old box node with new eliminated box for all users
2019 // of the same object and mark related locks as eliminated.
2020 mark_eliminated_box(box, obj);
2021 }
2022 }
2023 }
2024 }
2025
2026 // we have determined that this lock/unlock can be eliminated, we simply
2027 // eliminate the node without expanding it.
2028 //
2029 // Note: The membar's associated with the lock/unlock are currently not
2030 // eliminated. This should be investigated as a future enhancement.
2031 //
2032 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2033
2034 if (!alock->is_eliminated()) {
2035 return false;
2036 }
2037 #ifdef ASSERT
2038 if (!alock->is_coarsened()) {
2039 // Check that new "eliminated" BoxLock node is created.
2040 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2041 assert(oldbox->is_eliminated(), "should be done already");
2042 }
2043 #endif
2044
2045 alock->log_lock_optimization(C, "eliminate_lock");
2046
2047 #ifndef PRODUCT
2048 if (PrintEliminateLocks) {
2049 if (alock->is_Lock()) {
2050 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2051 } else {
2052 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2053 }
2054 }
2055 #endif
2056
2057 Node* mem = alock->in(TypeFunc::Memory);
2298 // region->in(2) is set to fast path - the object is locked to the current thread.
2299
2300 slow_path->init_req(2, ctrl); // Capture slow-control
2301 slow_mem->init_req(2, fast_lock_mem_phi);
2302
2303 transform_later(slow_path);
2304 transform_later(slow_mem);
2305 // Reset lock's memory edge.
2306 lock->set_req(TypeFunc::Memory, slow_mem);
2307
2308 } else {
2309 region = new RegionNode(3);
2310 // create a Phi for the memory state
2311 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2312
2313 // Optimize test; set region slot 2
2314 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2315 mem_phi->init_req(2, mem);
2316 }
2317
2318 // Make slow path call
2319 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2320 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2321 obj, box, NULL);
2322
2323 extract_call_projections(call);
2324
2325 // Slow path can only throw asynchronous exceptions, which are always
2326 // de-opted. So the compiler thinks the slow-call can never throw an
2327 // exception. If it DOES throw an exception we would need the debug
2328 // info removed first (since if it throws there is no monitor).
2329 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2330 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2331
2332 // Capture slow path
2333 // disconnect fall-through projection from call and create a new one
2334 // hook up users of fall-through projection to region
2335 Node *slow_ctrl = _fallthroughproj->clone();
2336 transform_later(slow_ctrl);
2337 _igvn.hash_delete(_fallthroughproj);
2399 // No exceptions for unlocking
2400 // Capture slow path
2401 // disconnect fall-through projection from call and create a new one
2402 // hook up users of fall-through projection to region
2403 Node *slow_ctrl = _fallthroughproj->clone();
2404 transform_later(slow_ctrl);
2405 _igvn.hash_delete(_fallthroughproj);
2406 _fallthroughproj->disconnect_inputs(NULL, C);
2407 region->init_req(1, slow_ctrl);
2408 // region inputs are now complete
2409 transform_later(region);
2410 _igvn.replace_node(_fallthroughproj, region);
2411
2412 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2413 mem_phi->init_req(1, memproj );
2414 mem_phi->init_req(2, mem);
2415 transform_later(mem_phi);
2416 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2417 }
2418
2419 //---------------------------eliminate_macro_nodes----------------------
2420 // Eliminate scalar replaced allocations and associated locks.
2421 void PhaseMacroExpand::eliminate_macro_nodes() {
2422 if (C->macro_count() == 0)
2423 return;
2424
2425 // First, attempt to eliminate locks
2426 int cnt = C->macro_count();
2427 for (int i=0; i < cnt; i++) {
2428 Node *n = C->macro_node(i);
2429 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2430 // Before elimination mark all associated (same box and obj)
2431 // lock and unlock nodes.
2432 mark_eliminated_locking_nodes(n->as_AbstractLock());
2433 }
2434 }
2435 bool progress = true;
2436 while (progress) {
2437 progress = false;
2438 for (int i = C->macro_count(); i > 0; i--) {
2443 success = eliminate_locking_node(n->as_AbstractLock());
2444 }
2445 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2446 progress = progress || success;
2447 }
2448 }
2449 // Next, attempt to eliminate allocations
2450 _has_locks = false;
2451 progress = true;
2452 while (progress) {
2453 progress = false;
2454 for (int i = C->macro_count(); i > 0; i--) {
2455 Node * n = C->macro_node(i-1);
2456 bool success = false;
2457 debug_only(int old_macro_count = C->macro_count(););
2458 switch (n->class_id()) {
2459 case Node::Class_Allocate:
2460 case Node::Class_AllocateArray:
2461 success = eliminate_allocate_node(n->as_Allocate());
2462 break;
2463 case Node::Class_CallStaticJava:
2464 success = eliminate_boxing_node(n->as_CallStaticJava());
2465 break;
2466 case Node::Class_Lock:
2467 case Node::Class_Unlock:
2468 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2469 _has_locks = true;
2470 break;
2471 case Node::Class_ArrayCopy:
2472 break;
2473 case Node::Class_OuterStripMinedLoop:
2474 break;
2475 default:
2476 assert(n->Opcode() == Op_LoopLimit ||
2477 n->Opcode() == Op_Opaque1 ||
2478 n->Opcode() == Op_Opaque2 ||
2479 n->Opcode() == Op_Opaque3 ||
2480 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2481 "unknown node type in macro list");
2482 }
2483 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2484 progress = progress || success;
2485 }
2495 // Make sure expansion will not cause node limit to be exceeded.
2496 // Worst case is a macro node gets expanded into about 200 nodes.
2497 // Allow 50% more for optimization.
2498 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2499 return true;
2500
2501 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2502 bool progress = true;
2503 while (progress) {
2504 progress = false;
2505 for (int i = C->macro_count(); i > 0; i--) {
2506 Node * n = C->macro_node(i-1);
2507 bool success = false;
2508 debug_only(int old_macro_count = C->macro_count(););
2509 if (n->Opcode() == Op_LoopLimit) {
2510 // Remove it from macro list and put on IGVN worklist to optimize.
2511 C->remove_macro_node(n);
2512 _igvn._worklist.push(n);
2513 success = true;
2514 } else if (n->Opcode() == Op_CallStaticJava) {
2515 // Remove it from macro list and put on IGVN worklist to optimize.
2516 C->remove_macro_node(n);
2517 _igvn._worklist.push(n);
2518 success = true;
2519 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2520 _igvn.replace_node(n, n->in(1));
2521 success = true;
2522 #if INCLUDE_RTM_OPT
2523 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2524 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2525 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2526 Node* cmp = n->unique_out();
2527 #ifdef ASSERT
2528 // Validate graph.
2529 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2530 BoolNode* bol = cmp->unique_out()->as_Bool();
2531 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2532 (bol->_test._test == BoolTest::ne), "");
2533 IfNode* ifn = bol->unique_out()->as_If();
2534 assert((ifn->outcnt() == 2) &&
2535 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2536 #endif
2537 Node* repl = n->in(1);
2538 if (!_has_locks) {
2578 int macro_count = C->macro_count();
2579 Node * n = C->macro_node(macro_count-1);
2580 assert(n->is_macro(), "only macro nodes expected here");
2581 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2582 // node is unreachable, so don't try to expand it
2583 C->remove_macro_node(n);
2584 continue;
2585 }
2586 switch (n->class_id()) {
2587 case Node::Class_Allocate:
2588 expand_allocate(n->as_Allocate());
2589 break;
2590 case Node::Class_AllocateArray:
2591 expand_allocate_array(n->as_AllocateArray());
2592 break;
2593 case Node::Class_Lock:
2594 expand_lock_node(n->as_Lock());
2595 break;
2596 case Node::Class_Unlock:
2597 expand_unlock_node(n->as_Unlock());
2598 break;
2599 default:
2600 assert(false, "unknown node type in macro list");
2601 }
2602 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2603 if (C->failing()) return true;
2604 }
2605
2606 _igvn.set_delay_transform(false);
2607 _igvn.optimize();
2608 if (C->failing()) return true;
2609 return false;
2610 }
|
30 #include "opto/addnode.hpp"
31 #include "opto/arraycopynode.hpp"
32 #include "opto/callnode.hpp"
33 #include "opto/castnode.hpp"
34 #include "opto/cfgnode.hpp"
35 #include "opto/compile.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/locknode.hpp"
39 #include "opto/loopnode.hpp"
40 #include "opto/macro.hpp"
41 #include "opto/memnode.hpp"
42 #include "opto/narrowptrnode.hpp"
43 #include "opto/node.hpp"
44 #include "opto/opaquenode.hpp"
45 #include "opto/phaseX.hpp"
46 #include "opto/rootnode.hpp"
47 #include "opto/runtime.hpp"
48 #include "opto/subnode.hpp"
49 #include "opto/type.hpp"
50 #include "opto/valuetypenode.hpp"
51 #include "runtime/sharedRuntime.hpp"
52 #include "utilities/macros.hpp"
53 #if INCLUDE_G1GC
54 #include "gc/g1/g1ThreadLocalData.hpp"
55 #endif // INCLUDE_G1GC
56 #if INCLUDE_SHENANDOAHGC
57 #include "gc/shenandoah/c2/shenandoahBarrierSetC2.hpp"
58 #endif
59
60
61 //
62 // Replace any references to "oldref" in inputs to "use" with "newref".
63 // Returns the number of replacements made.
64 //
65 int PhaseMacroExpand::replace_input(Node *use, Node *oldref, Node *newref) {
66 int nreplacements = 0;
67 uint req = use->req();
68 for (uint j = 0; j < use->len(); j++) {
69 Node *uin = use->in(j);
70 if (uin == oldref) {
71 if (j < req)
72 use->set_req(j, newref);
73 else
74 use->set_prec(j, newref);
75 nreplacements++;
76 } else if (j >= req && uin == NULL) {
77 break;
78 }
79 }
80 return nreplacements;
81 }
82
83 void PhaseMacroExpand::copy_call_debug_info(CallNode *oldcall, CallNode * newcall) {
84 // Copy debug information and adjust JVMState information
85 uint old_dbg_start = oldcall->tf()->domain_sig()->cnt();
86 uint new_dbg_start = newcall->tf()->domain_sig()->cnt();
87 int jvms_adj = new_dbg_start - old_dbg_start;
88 assert (new_dbg_start == newcall->req(), "argument count mismatch");
89
90 // SafePointScalarObject node could be referenced several times in debug info.
91 // Use Dict to record cloned nodes.
92 Dict* sosn_map = new Dict(cmpkey,hashkey);
93 for (uint i = old_dbg_start; i < oldcall->req(); i++) {
94 Node* old_in = oldcall->in(i);
95 // Clone old SafePointScalarObjectNodes, adjusting their field contents.
96 if (old_in != NULL && old_in->is_SafePointScalarObject()) {
97 SafePointScalarObjectNode* old_sosn = old_in->as_SafePointScalarObject();
98 uint old_unique = C->unique();
99 Node* new_in = old_sosn->clone(sosn_map);
100 if (old_unique != C->unique()) { // New node?
101 new_in->set_req(0, C->root()); // reset control edge
102 new_in = transform_later(new_in); // Register new node.
103 }
104 old_in = new_in;
105 }
106 newcall->add_req(old_in);
261 if (call->as_ArrayCopy()->modifies(offset, offset, phase, false)) {
262 return in;
263 }
264 }
265 mem = in->in(TypeFunc::Memory);
266 } else if (in->is_MemBar()) {
267 ArrayCopyNode* ac = NULL;
268 if (ArrayCopyNode::may_modify(tinst, in->as_MemBar(), phase, ac)) {
269 assert(ac != NULL && ac->is_clonebasic(), "Only basic clone is a non escaping clone");
270 return ac;
271 }
272 mem = in->in(TypeFunc::Memory);
273 } else {
274 assert(false, "unexpected projection");
275 }
276 } else if (mem->is_Store()) {
277 const TypePtr* atype = mem->as_Store()->adr_type();
278 int adr_idx = phase->C->get_alias_index(atype);
279 if (adr_idx == alias_idx) {
280 assert(atype->isa_oopptr(), "address type must be oopptr");
281 int adr_offset = atype->flattened_offset();
282 uint adr_iid = atype->is_oopptr()->instance_id();
283 // Array elements references have the same alias_idx
284 // but different offset and different instance_id.
285 if (adr_offset == offset && adr_iid == alloc->_idx)
286 return mem;
287 } else {
288 assert(adr_idx == Compile::AliasIdxRaw, "address must match or be raw");
289 }
290 mem = mem->in(MemNode::Memory);
291 } else if (mem->is_ClearArray()) {
292 if (!ClearArrayNode::step_through(&mem, alloc->_idx, phase)) {
293 // Can not bypass initialization of the instance
294 // we are looking.
295 debug_only(intptr_t offset;)
296 assert(alloc == AllocateNode::Ideal_allocation(mem->in(3), phase, offset), "sanity");
297 InitializeNode* init = alloc->as_Allocate()->initialization();
298 // We are looking for stored value, return Initialize node
299 // or memory edge from Allocate node.
300 if (init != NULL)
301 return init;
304 }
305 // Otherwise skip it (the call updated 'mem' value).
306 } else if (mem->Opcode() == Op_SCMemProj) {
307 mem = mem->in(0);
308 Node* adr = NULL;
309 if (mem->is_LoadStore()) {
310 adr = mem->in(MemNode::Address);
311 } else {
312 assert(mem->Opcode() == Op_EncodeISOArray ||
313 mem->Opcode() == Op_StrCompressedCopy, "sanity");
314 adr = mem->in(3); // Destination array
315 }
316 const TypePtr* atype = adr->bottom_type()->is_ptr();
317 int adr_idx = phase->C->get_alias_index(atype);
318 if (adr_idx == alias_idx) {
319 DEBUG_ONLY(mem->dump();)
320 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
321 return NULL;
322 }
323 mem = mem->in(MemNode::Memory);
324 } else if (mem->Opcode() == Op_StrInflatedCopy) {
325 Node* adr = mem->in(3); // Destination array
326 const TypePtr* atype = adr->bottom_type()->is_ptr();
327 int adr_idx = phase->C->get_alias_index(atype);
328 if (adr_idx == alias_idx) {
329 DEBUG_ONLY(mem->dump();)
330 assert(false, "Object is not scalar replaceable if a StrInflatedCopy node accesses its field");
331 return NULL;
332 }
333 mem = mem->in(MemNode::Memory);
334 } else {
335 return mem;
336 }
337 assert(mem != orig_mem, "dead memory loop");
338 }
339 }
340
341 // Generate loads from source of the arraycopy for fields of
342 // destination needed at a deoptimization point
343 Node* PhaseMacroExpand::make_arraycopy_load(ArrayCopyNode* ac, intptr_t offset, Node* ctl, Node* mem, BasicType ft, const Type *ftype, AllocateNode *alloc) {
344 BasicType bt = ft;
345 const Type *type = ftype;
346 if (ft == T_NARROWOOP) {
347 bt = T_OBJECT;
348 type = ftype->make_oopptr();
349 }
350 Node* res = NULL;
351 if (ac->is_clonebasic()) {
352 Node* base = ac->in(ArrayCopyNode::Src)->in(AddPNode::Base);
353 Node* adr = _igvn.transform(new AddPNode(base, base, MakeConX(offset)));
354 const TypePtr* adr_type = _igvn.type(base)->is_ptr()->add_offset(offset);
355 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
356 } else {
357 if (ac->modifies(offset, offset, &_igvn, true)) {
358 assert(ac->in(ArrayCopyNode::Dest) == alloc->result_cast(), "arraycopy destination should be allocation's result");
359 uint shift = exact_log2(type2aelembytes(bt));
360 Node* diff = _igvn.transform(new SubINode(ac->in(ArrayCopyNode::SrcPos), ac->in(ArrayCopyNode::DestPos)));
361 #ifdef _LP64
362 diff = _igvn.transform(new ConvI2LNode(diff));
363 #endif
364 diff = _igvn.transform(new LShiftXNode(diff, intcon(shift)));
365
366 Node* off = _igvn.transform(new AddXNode(MakeConX(offset), diff));
367 Node* base = ac->in(ArrayCopyNode::Src);
368 Node* adr = _igvn.transform(new AddPNode(base, base, off));
369 const TypePtr* adr_type = _igvn.type(base)->is_ptr();
370 if (adr_type->isa_aryptr()) {
371 // In the case of a flattened value type array, each field has its
372 // own slice so we need to extract the field being accessed from
373 // the address computation
374 adr_type = adr_type->is_aryptr()->add_field_offset_and_offset(offset);
375 adr = _igvn.transform(new CastPPNode(adr, adr_type));
376 } else {
377 adr_type = adr_type->add_offset(offset);
378 }
379 res = LoadNode::make(_igvn, ctl, mem, adr, adr_type, type, bt, MemNode::unordered, LoadNode::Pinned);
380 }
381 }
382 if (res != NULL) {
383 res = _igvn.transform(res);
384 if (ftype->isa_narrowoop()) {
385 // PhaseMacroExpand::scalar_replacement adds DecodeN nodes
386 assert(res->isa_DecodeN(), "should be narrow oop");
387 res = _igvn.transform(new EncodePNode(res, ftype));
388 }
389 return res;
390 }
391 return NULL;
392 }
393
394 //
395 // Given a Memory Phi, compute a value Phi containing the values from stores
396 // on the input paths.
397 // Note: this function is recursive, its depth is limited by the "level" argument
398 // Returns the computed Phi, or NULL if it cannot compute it.
399 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) {
400 assert(mem->is_Phi(), "sanity");
401 int alias_idx = C->get_alias_index(adr_t);
402 int offset = adr_t->flattened_offset();
403 int instance_id = adr_t->instance_id();
404
405 // Check if an appropriate value phi already exists.
406 Node* region = mem->in(0);
407 for (DUIterator_Fast kmax, k = region->fast_outs(kmax); k < kmax; k++) {
408 Node* phi = region->fast_out(k);
409 if (phi->is_Phi() && phi != mem &&
410 phi->as_Phi()->is_same_inst_field(phi_type, (int)mem->_idx, instance_id, alias_idx, offset)) {
411 return phi;
412 }
413 }
414 // Check if an appropriate new value phi already exists.
415 Node* new_phi = value_phis->find(mem->_idx);
416 if (new_phi != NULL)
417 return new_phi;
418
419 if (level <= 0) {
420 return NULL; // Give up: phi tree too deep
421 }
422 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
423 Node *alloc_mem = alloc->in(TypeFunc::Memory);
424
425 uint length = mem->req();
426 GrowableArray <Node *> values(length, length, NULL, false);
427
428 // create a new Phi for the value
429 PhiNode *phi = new PhiNode(mem->in(0), phi_type, NULL, mem->_idx, instance_id, alias_idx, offset);
430 transform_later(phi);
431 value_phis->push(phi, mem->_idx);
432
433 for (uint j = 1; j < length; j++) {
434 Node *in = mem->in(j);
435 if (in == NULL || in->is_top()) {
436 values.at_put(j, in);
437 } else {
438 Node *val = scan_mem_chain(in, alias_idx, offset, start_mem, alloc, &_igvn);
439 if (val == start_mem || val == alloc_mem) {
440 // hit a sentinel, return appropriate 0 value
441 Node* default_value = alloc->in(AllocateNode::DefaultValue);
442 if (default_value != NULL) {
443 values.at_put(j, default_value);
444 } else {
445 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
446 values.at_put(j, _igvn.zerocon(ft));
447 }
448 continue;
449 }
450 if (val->is_Initialize()) {
451 val = val->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
452 }
453 if (val == NULL) {
454 return NULL; // can't find a value on this path
455 }
456 if (val == mem) {
457 values.at_put(j, mem);
458 } else if (val->is_Store()) {
459 Node* n = val->in(MemNode::ValueIn);
460 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
461 n = bs->step_over_gc_barrier(n);
462 values.at_put(j, n);
463 } else if(val->is_Proj() && val->in(0) == alloc) {
464 Node* default_value = alloc->in(AllocateNode::DefaultValue);
465 if (default_value != NULL) {
466 values.at_put(j, default_value);
467 } else {
468 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
469 values.at_put(j, _igvn.zerocon(ft));
470 }
471 } else if (val->is_Phi()) {
472 val = value_from_mem_phi(val, ft, phi_type, adr_t, alloc, value_phis, level-1);
473 if (val == NULL) {
474 return NULL;
475 }
476 values.at_put(j, val);
477 } else if (val->Opcode() == Op_SCMemProj) {
478 assert(val->in(0)->is_LoadStore() ||
479 val->in(0)->Opcode() == Op_EncodeISOArray ||
480 val->in(0)->Opcode() == Op_StrCompressedCopy, "sanity");
481 assert(false, "Object is not scalar replaceable if a LoadStore node accesses its field");
482 return NULL;
483 } else if (val->is_ArrayCopy()) {
484 Node* res = make_arraycopy_load(val->as_ArrayCopy(), offset, val->in(0), val->in(TypeFunc::Memory), ft, phi_type, alloc);
485 if (res == NULL) {
486 return NULL;
487 }
488 values.at_put(j, res);
489 } else {
490 #ifdef ASSERT
496 }
497 }
498 // Set Phi's inputs
499 for (uint j = 1; j < length; j++) {
500 if (values.at(j) == mem) {
501 phi->init_req(j, phi);
502 } else {
503 phi->init_req(j, values.at(j));
504 }
505 }
506 return phi;
507 }
508
509 // Search the last value stored into the object's field.
510 Node *PhaseMacroExpand::value_from_mem(Node *sfpt_mem, Node *sfpt_ctl, BasicType ft, const Type *ftype, const TypeOopPtr *adr_t, AllocateNode *alloc) {
511 assert(adr_t->is_known_instance_field(), "instance required");
512 int instance_id = adr_t->instance_id();
513 assert((uint)instance_id == alloc->_idx, "wrong allocation");
514
515 int alias_idx = C->get_alias_index(adr_t);
516 int offset = adr_t->flattened_offset();
517 Node *start_mem = C->start()->proj_out_or_null(TypeFunc::Memory);
518 Node *alloc_mem = alloc->in(TypeFunc::Memory);
519 Arena *a = Thread::current()->resource_area();
520 VectorSet visited(a);
521
522 bool done = sfpt_mem == alloc_mem;
523 Node *mem = sfpt_mem;
524 while (!done) {
525 if (visited.test_set(mem->_idx)) {
526 return NULL; // found a loop, give up
527 }
528 mem = scan_mem_chain(mem, alias_idx, offset, start_mem, alloc, &_igvn);
529 if (mem == start_mem || mem == alloc_mem) {
530 done = true; // hit a sentinel, return appropriate 0 value
531 } else if (mem->is_Initialize()) {
532 mem = mem->as_Initialize()->find_captured_store(offset, type2aelembytes(ft), &_igvn);
533 if (mem == NULL) {
534 done = true; // Something went wrong.
535 } else if (mem->is_Store()) {
536 const TypePtr* atype = mem->as_Store()->adr_type();
537 assert(C->get_alias_index(atype) == Compile::AliasIdxRaw, "store is correct memory slice");
538 done = true;
539 }
540 } else if (mem->is_Store()) {
541 const TypeOopPtr* atype = mem->as_Store()->adr_type()->isa_oopptr();
542 assert(atype != NULL, "address type must be oopptr");
543 assert(C->get_alias_index(atype) == alias_idx &&
544 atype->is_known_instance_field() && atype->flattened_offset() == offset &&
545 atype->instance_id() == instance_id, "store is correct memory slice");
546 done = true;
547 } else if (mem->is_Phi()) {
548 // try to find a phi's unique input
549 Node *unique_input = NULL;
550 Node *top = C->top();
551 for (uint i = 1; i < mem->req(); i++) {
552 Node *n = scan_mem_chain(mem->in(i), alias_idx, offset, start_mem, alloc, &_igvn);
553 if (n == NULL || n == top || n == mem) {
554 continue;
555 } else if (unique_input == NULL) {
556 unique_input = n;
557 } else if (unique_input != n) {
558 unique_input = top;
559 break;
560 }
561 }
562 if (unique_input != NULL && unique_input != top) {
563 mem = unique_input;
564 } else {
565 done = true;
566 }
567 } else if (mem->is_ArrayCopy()) {
568 done = true;
569 } else {
570 assert(false, "unexpected node");
571 }
572 }
573 if (mem != NULL) {
574 if (mem == start_mem || mem == alloc_mem) {
575 // hit a sentinel, return appropriate 0 value
576 Node* default_value = alloc->in(AllocateNode::DefaultValue);
577 if (default_value != NULL) {
578 return default_value;
579 }
580 assert(alloc->in(AllocateNode::RawDefaultValue) == NULL, "default value may not be null");
581 return _igvn.zerocon(ft);
582 } else if (mem->is_Store()) {
583 Node* n = mem->in(MemNode::ValueIn);
584 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
585 n = bs->step_over_gc_barrier(n);
586 return n;
587 } else if (mem->is_Phi()) {
588 // attempt to produce a Phi reflecting the values on the input paths of the Phi
589 Node_Stack value_phis(a, 8);
590 Node * phi = value_from_mem_phi(mem, ft, ftype, adr_t, alloc, &value_phis, ValueSearchLimit);
591 if (phi != NULL) {
592 return phi;
593 } else {
594 // Kill all new Phis
595 while(value_phis.is_nonempty()) {
596 Node* n = value_phis.node();
597 _igvn.replace_node(n, C->top());
598 value_phis.pop();
599 }
600 }
601 } else if (mem->is_ArrayCopy()) {
602 Node* ctl = mem->in(0);
603 Node* m = mem->in(TypeFunc::Memory);
604 if (sfpt_ctl->is_Proj() && sfpt_ctl->as_Proj()->is_uncommon_trap_proj(Deoptimization::Reason_none)) {
605 // pin the loads in the uncommon trap path
606 ctl = sfpt_ctl;
607 m = sfpt_mem;
608 }
609 return make_arraycopy_load(mem->as_ArrayCopy(), offset, ctl, m, ft, ftype, alloc);
610 }
611 }
612 // Something went wrong.
613 return NULL;
614 }
615
616 // Search the last value stored into the value type's fields.
617 Node* PhaseMacroExpand::value_type_from_mem(Node* mem, Node* ctl, ciValueKlass* vk, const TypeAryPtr* adr_type, int offset, AllocateNode* alloc) {
618 // Subtract the offset of the first field to account for the missing oop header
619 offset -= vk->first_field_offset();
620 // Create a new ValueTypeNode and retrieve the field values from memory
621 ValueTypeNode* vt = ValueTypeNode::make_uninitialized(_igvn, vk)->as_ValueType();
622 for (int i = 0; i < vk->nof_declared_nonstatic_fields(); ++i) {
623 ciType* field_type = vt->field_type(i);
624 int field_offset = offset + vt->field_offset(i);
625 // Each value type field has its own memory slice
626 adr_type = adr_type->with_field_offset(field_offset);
627 Node* value = NULL;
628 if (vt->field_is_flattened(i)) {
629 value = value_type_from_mem(mem, ctl, field_type->as_value_klass(), adr_type, field_offset, alloc);
630 } else {
631 const Type* ft = Type::get_const_type(field_type);
632 BasicType bt = field_type->basic_type();
633 if (UseCompressedOops && !is_java_primitive(bt)) {
634 ft = ft->make_narrowoop();
635 bt = T_NARROWOOP;
636 }
637 value = value_from_mem(mem, ctl, bt, ft, adr_type, alloc);
638 if (value != NULL && ft->isa_narrowoop()) {
639 assert(UseCompressedOops, "unexpected narrow oop");
640 value = transform_later(new DecodeNNode(value, value->get_ptr_type()));
641 }
642 }
643 if (value != NULL) {
644 vt->set_field_value(i, value);
645 } else {
646 // We might have reached the TrackedInitializationLimit
647 return NULL;
648 }
649 }
650 return transform_later(vt);
651 }
652
653 // Check the possibility of scalar replacement.
654 bool PhaseMacroExpand::can_eliminate_allocation(AllocateNode *alloc, GrowableArray <SafePointNode *>& safepoints) {
655 // Scan the uses of the allocation to check for anything that would
656 // prevent us from eliminating it.
657 NOT_PRODUCT( const char* fail_eliminate = NULL; )
658 DEBUG_ONLY( Node* disq_node = NULL; )
659 bool can_eliminate = true;
660
661 Node* res = alloc->result_cast();
662 const TypeOopPtr* res_type = NULL;
663 if (res == NULL) {
664 // All users were eliminated.
665 } else if (!res->is_CheckCastPP()) {
666 NOT_PRODUCT(fail_eliminate = "Allocation does not have unique CheckCastPP";)
667 can_eliminate = false;
668 } else {
669 res_type = _igvn.type(res)->isa_oopptr();
670 if (res_type == NULL) {
671 NOT_PRODUCT(fail_eliminate = "Neither instance or array allocation";)
672 can_eliminate = false;
688 const TypePtr* addp_type = _igvn.type(use)->is_ptr();
689 int offset = addp_type->offset();
690
691 if (offset == Type::OffsetTop || offset == Type::OffsetBot) {
692 NOT_PRODUCT(fail_eliminate = "Undefined field referrence";)
693 can_eliminate = false;
694 break;
695 }
696 for (DUIterator_Fast kmax, k = use->fast_outs(kmax);
697 k < kmax && can_eliminate; k++) {
698 Node* n = use->fast_out(k);
699 if (!n->is_Store() && n->Opcode() != Op_CastP2X &&
700 SHENANDOAHGC_ONLY((!UseShenandoahGC || !ShenandoahBarrierSetC2::is_shenandoah_wb_pre_call(n)) &&)
701 !(n->is_ArrayCopy() &&
702 n->as_ArrayCopy()->is_clonebasic() &&
703 n->in(ArrayCopyNode::Dest) == use)) {
704 DEBUG_ONLY(disq_node = n;)
705 if (n->is_Load() || n->is_LoadStore()) {
706 NOT_PRODUCT(fail_eliminate = "Field load";)
707 } else {
708 NOT_PRODUCT(fail_eliminate = "Not store field reference";)
709 }
710 can_eliminate = false;
711 }
712 }
713 } else if (use->is_ArrayCopy() &&
714 (use->as_ArrayCopy()->is_arraycopy_validated() ||
715 use->as_ArrayCopy()->is_copyof_validated() ||
716 use->as_ArrayCopy()->is_copyofrange_validated()) &&
717 use->in(ArrayCopyNode::Dest) == res) {
718 // ok to eliminate
719 } else if (use->is_SafePoint()) {
720 SafePointNode* sfpt = use->as_SafePoint();
721 if (sfpt->is_Call() && sfpt->as_Call()->has_non_debug_use(res)) {
722 // Object is passed as argument.
723 DEBUG_ONLY(disq_node = use;)
724 NOT_PRODUCT(fail_eliminate = "Object is passed as argument";)
725 can_eliminate = false;
726 }
727 Node* sfptMem = sfpt->memory();
728 if (sfptMem == NULL || sfptMem->is_top()) {
729 DEBUG_ONLY(disq_node = use;)
730 NOT_PRODUCT(fail_eliminate = "NULL or TOP memory";)
731 can_eliminate = false;
732 } else {
733 safepoints.append_if_missing(sfpt);
734 }
735 } else if (use->is_ValueType() && use->isa_ValueType()->get_oop() == res) {
736 // ok to eliminate
737 } else if (use->is_Store()) {
738 // store to mark work
739 } else if (use->Opcode() != Op_CastP2X) { // CastP2X is used by card mark
740 if (use->is_Phi()) {
741 if (use->outcnt() == 1 && use->unique_out()->Opcode() == Op_Return) {
742 NOT_PRODUCT(fail_eliminate = "Object is return value";)
743 } else {
744 NOT_PRODUCT(fail_eliminate = "Object is referenced by Phi";)
745 }
746 DEBUG_ONLY(disq_node = use;)
747 } else {
748 if (use->Opcode() == Op_Return) {
749 NOT_PRODUCT(fail_eliminate = "Object is return value";)
750 } else {
751 NOT_PRODUCT(fail_eliminate = "Object is referenced by node";)
752 }
753 DEBUG_ONLY(disq_node = use;)
754 }
755 can_eliminate = false;
756 } else {
757 assert(use->Opcode() == Op_CastP2X, "should be");
758 assert(!use->has_out_with(Op_OrL), "should have been removed because oop is never null");
759 }
760 }
761 }
762
763 #ifndef PRODUCT
764 if (PrintEliminateAllocations) {
765 if (can_eliminate) {
766 tty->print("Scalar ");
767 if (res == NULL)
768 alloc->dump();
769 else
770 res->dump();
771 } else if (alloc->_is_scalar_replaceable) {
772 tty->print("NotScalar (%s)", fail_eliminate);
773 if (res == NULL)
774 alloc->dump();
775 else
776 res->dump();
777 #ifdef ASSERT
778 if (disq_node != NULL) {
801 Node* res = alloc->result_cast();
802 assert(res == NULL || res->is_CheckCastPP(), "unexpected AllocateNode result");
803 const TypeOopPtr* res_type = NULL;
804 if (res != NULL) { // Could be NULL when there are no users
805 res_type = _igvn.type(res)->isa_oopptr();
806 }
807
808 if (res != NULL) {
809 klass = res_type->klass();
810 if (res_type->isa_instptr()) {
811 // find the fields of the class which will be needed for safepoint debug information
812 assert(klass->is_instance_klass(), "must be an instance klass.");
813 iklass = klass->as_instance_klass();
814 nfields = iklass->nof_nonstatic_fields();
815 } else {
816 // find the array's elements which will be needed for safepoint debug information
817 nfields = alloc->in(AllocateNode::ALength)->find_int_con(-1);
818 assert(klass->is_array_klass() && nfields >= 0, "must be an array klass.");
819 elem_type = klass->as_array_klass()->element_type();
820 basic_elem_type = elem_type->basic_type();
821 if (elem_type->is_valuetype() && !klass->is_value_array_klass()) {
822 assert(basic_elem_type == T_VALUETYPE, "unexpected element basic type");
823 basic_elem_type = T_OBJECT;
824 }
825 array_base = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
826 element_size = type2aelembytes(basic_elem_type);
827 if (klass->is_value_array_klass()) {
828 // Flattened value type array
829 element_size = klass->as_value_array_klass()->element_byte_size();
830 }
831 }
832 }
833 //
834 // Process the safepoint uses
835 //
836 Unique_Node_List value_worklist;
837 while (safepoints.length() > 0) {
838 SafePointNode* sfpt = safepoints.pop();
839 Node* mem = sfpt->memory();
840 Node* ctl = sfpt->control();
841 assert(sfpt->jvms() != NULL, "missed JVMS");
842 // Fields of scalar objs are referenced only at the end
843 // of regular debuginfo at the last (youngest) JVMS.
844 // Record relative start index.
845 uint first_ind = (sfpt->req() - sfpt->jvms()->scloff());
846 SafePointScalarObjectNode* sobj = new SafePointScalarObjectNode(res_type,
847 #ifdef ASSERT
848 alloc,
849 #endif
850 first_ind, nfields);
851 sobj->init_req(0, C->root());
852 transform_later(sobj);
853
854 // Scan object's fields adding an input to the safepoint for each field.
855 for (int j = 0; j < nfields; j++) {
856 intptr_t offset;
857 ciField* field = NULL;
858 if (iklass != NULL) {
859 field = iklass->nonstatic_field_at(j);
860 offset = field->offset();
861 elem_type = field->type();
862 basic_elem_type = field->layout_type();
863 assert(!field->is_flattened(), "flattened value type fields should not have safepoint uses");
864 } else {
865 offset = array_base + j * (intptr_t)element_size;
866 }
867
868 const Type *field_type;
869 // The next code is taken from Parse::do_get_xxx().
870 if (basic_elem_type == T_OBJECT || basic_elem_type == T_ARRAY) {
871 if (!elem_type->is_loaded()) {
872 field_type = TypeInstPtr::BOTTOM;
873 } else if (field != NULL && field->is_static_constant()) {
874 // This can happen if the constant oop is non-perm.
875 ciObject* con = field->constant_value().as_object();
876 // Do not "join" in the previous type; it doesn't add value,
877 // and may yield a vacuous result if the field is of interface type.
878 field_type = TypeOopPtr::make_from_constant(con)->isa_oopptr();
879 assert(field_type != NULL, "field singleton type must be consistent");
880 } else {
881 field_type = TypeOopPtr::make_from_klass(elem_type->as_klass());
882 }
883 if (UseCompressedOops) {
884 field_type = field_type->make_narrowoop();
885 basic_elem_type = T_NARROWOOP;
886 }
887 } else {
888 field_type = Type::get_const_basic_type(basic_elem_type);
889 }
890
891 Node* field_val = NULL;
892 const TypeOopPtr* field_addr_type = res_type->add_offset(offset)->isa_oopptr();
893 if (klass->is_value_array_klass()) {
894 ciValueKlass* vk = elem_type->as_value_klass();
895 assert(vk->flatten_array(), "must be flattened");
896 field_val = value_type_from_mem(mem, ctl, vk, field_addr_type->isa_aryptr(), 0, alloc);
897 } else {
898 field_val = value_from_mem(mem, ctl, basic_elem_type, field_type, field_addr_type, alloc);
899 }
900 if (field_val == NULL) {
901 // We weren't able to find a value for this field,
902 // give up on eliminating this allocation.
903
904 // Remove any extra entries we added to the safepoint.
905 uint last = sfpt->req() - 1;
906 for (int k = 0; k < j; k++) {
907 sfpt->del_req(last--);
908 }
909 _igvn._worklist.push(sfpt);
910 // rollback processed safepoints
911 while (safepoints_done.length() > 0) {
912 SafePointNode* sfpt_done = safepoints_done.pop();
913 // remove any extra entries we added to the safepoint
914 last = sfpt_done->req() - 1;
915 for (int k = 0; k < nfields; k++) {
916 sfpt_done->del_req(last--);
917 }
918 JVMState *jvms = sfpt_done->jvms();
919 jvms->set_endoff(sfpt_done->req());
945 tty->print("=== At SafePoint node %d can't find value of array element [%d]",
946 sfpt->_idx, j);
947 }
948 tty->print(", which prevents elimination of: ");
949 if (res == NULL)
950 alloc->dump();
951 else
952 res->dump();
953 }
954 #endif
955 return false;
956 }
957 if (UseCompressedOops && field_type->isa_narrowoop()) {
958 // Enable "DecodeN(EncodeP(Allocate)) --> Allocate" transformation
959 // to be able scalar replace the allocation.
960 if (field_val->is_EncodeP()) {
961 field_val = field_val->in(1);
962 } else {
963 field_val = transform_later(new DecodeNNode(field_val, field_val->get_ptr_type()));
964 }
965 } else if (field_val->is_ValueType()) {
966 // Keep track of value types to scalarize them later
967 value_worklist.push(field_val);
968 }
969 sfpt->add_req(field_val);
970 }
971 JVMState *jvms = sfpt->jvms();
972 jvms->set_endoff(sfpt->req());
973 // Now make a pass over the debug information replacing any references
974 // to the allocated object with "sobj"
975 int start = jvms->debug_start();
976 int end = jvms->debug_end();
977 sfpt->replace_edges_in_range(res, sobj, start, end);
978 _igvn._worklist.push(sfpt);
979 safepoints_done.append_if_missing(sfpt); // keep it for rollback
980 }
981 // Scalarize value types that were added to the safepoint
982 for (uint i = 0; i < value_worklist.size(); ++i) {
983 Node* vt = value_worklist.at(i);
984 vt->as_ValueType()->make_scalar_in_safepoints(&_igvn);
985 }
986 return true;
987 }
988
989 static void disconnect_projections(MultiNode* n, PhaseIterGVN& igvn) {
990 Node* ctl_proj = n->proj_out_or_null(TypeFunc::Control);
991 Node* mem_proj = n->proj_out_or_null(TypeFunc::Memory);
992 if (ctl_proj != NULL) {
993 igvn.replace_node(ctl_proj, n->in(0));
994 }
995 if (mem_proj != NULL) {
996 igvn.replace_node(mem_proj, n->in(TypeFunc::Memory));
997 }
998 }
999
1000 // Process users of eliminated allocation.
1001 void PhaseMacroExpand::process_users_of_allocation(CallNode *alloc) {
1002 Node* res = alloc->result_cast();
1003 if (res != NULL) {
1004 for (DUIterator_Last jmin, j = res->last_outs(jmin); j >= jmin; ) {
1005 Node *use = res->last_out(j);
1030 Node* membar_after = ac->proj_out(TypeFunc::Control)->unique_ctrl_out();
1031 disconnect_projections(ac, _igvn);
1032 assert(alloc->in(0)->is_Proj() && alloc->in(0)->in(0)->Opcode() == Op_MemBarCPUOrder, "mem barrier expected before allocation");
1033 Node* membar_before = alloc->in(0)->in(0);
1034 disconnect_projections(membar_before->as_MemBar(), _igvn);
1035 if (membar_after->is_MemBar()) {
1036 disconnect_projections(membar_after->as_MemBar(), _igvn);
1037 }
1038 } else {
1039 eliminate_gc_barrier(n);
1040 }
1041 k -= (oc2 - use->outcnt());
1042 }
1043 _igvn.remove_dead_node(use);
1044 } else if (use->is_ArrayCopy()) {
1045 // Disconnect ArrayCopy node
1046 ArrayCopyNode* ac = use->as_ArrayCopy();
1047 assert(ac->is_arraycopy_validated() ||
1048 ac->is_copyof_validated() ||
1049 ac->is_copyofrange_validated(), "unsupported");
1050 CallProjections* callprojs = ac->extract_projections(true);
1051
1052 _igvn.replace_node(callprojs->fallthrough_ioproj, ac->in(TypeFunc::I_O));
1053 _igvn.replace_node(callprojs->fallthrough_memproj, ac->in(TypeFunc::Memory));
1054 _igvn.replace_node(callprojs->fallthrough_catchproj, ac->in(TypeFunc::Control));
1055
1056 // Set control to top. IGVN will remove the remaining projections
1057 ac->set_req(0, top());
1058 ac->replace_edge(res, top());
1059
1060 // Disconnect src right away: it can help find new
1061 // opportunities for allocation elimination
1062 Node* src = ac->in(ArrayCopyNode::Src);
1063 ac->replace_edge(src, top());
1064 // src can be top at this point if src and dest of the
1065 // arraycopy were the same
1066 if (src->outcnt() == 0 && !src->is_top()) {
1067 _igvn.remove_dead_node(src);
1068 }
1069
1070 _igvn._worklist.push(ac);
1071 } else if (use->is_ValueType()) {
1072 assert(use->isa_ValueType()->get_oop() == res, "unexpected value type use");
1073 _igvn.rehash_node_delayed(use);
1074 use->isa_ValueType()->set_oop(_igvn.zerocon(T_VALUETYPE));
1075 } else if (use->is_Store()) {
1076 _igvn.replace_node(use, use->in(MemNode::Memory));
1077 } else {
1078 eliminate_gc_barrier(use);
1079 }
1080 j -= (oc1 - res->outcnt());
1081 }
1082 assert(res->outcnt() == 0, "all uses of allocated objects must be deleted");
1083 _igvn.remove_dead_node(res);
1084 }
1085
1086 //
1087 // Process other users of allocation's projections
1088 //
1089 if (_resproj != NULL && _resproj->outcnt() != 0) {
1090 // First disconnect stores captured by Initialize node.
1091 // If Initialize node is eliminated first in the following code,
1092 // it will kill such stores and DUIterator_Last will assert.
1093 for (DUIterator_Fast jmax, j = _resproj->fast_outs(jmax); j < jmax; j++) {
1094 Node *use = _resproj->fast_out(j);
1095 if (use->is_AddP()) {
1096 // raw memory addresses used only by the initialization
1208 if (alloc->is_AllocateArray())
1209 tty->print_cr("++++ Eliminated: %d AllocateArray", alloc->_idx);
1210 else
1211 tty->print_cr("++++ Eliminated: %d Allocate", alloc->_idx);
1212 }
1213 #endif
1214
1215 return true;
1216 }
1217
1218 bool PhaseMacroExpand::eliminate_boxing_node(CallStaticJavaNode *boxing) {
1219 // EA should remove all uses of non-escaping boxing node.
1220 if (!C->eliminate_boxing() || boxing->proj_out_or_null(TypeFunc::Parms) != NULL) {
1221 return false;
1222 }
1223
1224 assert(boxing->result_cast() == NULL, "unexpected boxing node result");
1225
1226 extract_call_projections(boxing);
1227
1228 const TypeTuple* r = boxing->tf()->range_sig();
1229 assert(r->cnt() > TypeFunc::Parms, "sanity");
1230 const TypeInstPtr* t = r->field_at(TypeFunc::Parms)->isa_instptr();
1231 assert(t != NULL, "sanity");
1232
1233 CompileLog* log = C->log();
1234 if (log != NULL) {
1235 log->head("eliminate_boxing type='%d'",
1236 log->identify(t->klass()));
1237 JVMState* p = boxing->jvms();
1238 while (p != NULL) {
1239 log->elem("jvms bci='%d' method='%d'", p->bci(), log->identify(p->method()));
1240 p = p->caller();
1241 }
1242 log->tail("eliminate_boxing");
1243 }
1244
1245 process_users_of_allocation(boxing);
1246
1247 #ifndef PRODUCT
1248 if (PrintEliminateAllocations) {
1368 Node *result_phi_i_o = NULL;
1369
1370 // The initial slow comparison is a size check, the comparison
1371 // we want to do is a BoolTest::gt
1372 bool always_slow = false;
1373 int tv = _igvn.find_int_con(initial_slow_test, -1);
1374 if (tv >= 0) {
1375 always_slow = (tv == 1);
1376 initial_slow_test = NULL;
1377 } else {
1378 initial_slow_test = BoolNode::make_predicate(initial_slow_test, &_igvn);
1379 }
1380
1381 if (C->env()->dtrace_alloc_probes() ||
1382 (!UseTLAB && !Universe::heap()->supports_inline_contig_alloc())) {
1383 // Force slow-path allocation
1384 always_slow = true;
1385 initial_slow_test = NULL;
1386 }
1387
1388 Node *slow_region = NULL;
1389 Node *toobig_false = ctrl;
1390
1391 assert (initial_slow_test == NULL || !always_slow, "arguments must be consistent");
1392 // generate the initial test if necessary
1393 if (initial_slow_test != NULL ) {
1394 if (slow_region == NULL) {
1395 slow_region = new RegionNode(1);
1396 }
1397 // Now make the initial failure test. Usually a too-big test but
1398 // might be a TRUE for finalizers or a fancy class check for
1399 // newInstance0.
1400 IfNode* toobig_iff = new IfNode(ctrl, initial_slow_test, PROB_MIN, COUNT_UNKNOWN);
1401 transform_later(toobig_iff);
1402 // Plug the failing-too-big test into the slow-path region
1403 Node* toobig_true = new IfTrueNode(toobig_iff);
1404 transform_later(toobig_true);
1405 slow_region ->add_req(toobig_true);
1406 toobig_false = new IfFalseNode(toobig_iff);
1407 transform_later(toobig_false);
1408 } else { // No initial test, just fall into next case
1409 toobig_false = ctrl;
1410 }
1411
1412 Node *slow_mem = mem; // save the current memory state for slow path
1413 // generate the fast allocation code unless we know that the initial test will always go slow
1414 if (!always_slow) {
1415 // Fast path modifies only raw memory.
1416 if (mem->is_MergeMem()) {
1417 mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1418 }
1419
1420 // allocate the Region and Phi nodes for the result
1421 result_region = new RegionNode(3);
1422 result_phi_rawmem = new PhiNode(result_region, Type::MEMORY, TypeRawPtr::BOTTOM);
1423 result_phi_rawoop = new PhiNode(result_region, TypeRawPtr::BOTTOM);
1424 result_phi_i_o = new PhiNode(result_region, Type::ABIO); // I/O is used for Prefetch
1425
1426 // Grab regular I/O before optional prefetch may change it.
1427 // Slow-path does no I/O so just set it to the original I/O.
1428 result_phi_i_o->init_req(slow_result_path, i_o);
1429
1430 Node* needgc_ctrl = NULL;
1431 // Name successful fast-path variables
1432 Node* fast_oop_ctrl;
1433 Node* fast_oop_rawmem;
1434
1435 intx prefetch_lines = length != NULL ? AllocatePrefetchLines : AllocateInstancePrefetchLines;
1436
1437 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
1438 Node* fast_oop = bs->obj_allocate(this, ctrl, mem, toobig_false, size_in_bytes, i_o, needgc_ctrl,
1439 fast_oop_ctrl, fast_oop_rawmem,
1440 prefetch_lines);
1441
1442 if (slow_region != NULL) {
1443 slow_region->add_req(needgc_ctrl);
1444 // This completes all paths into the slow merge point
1445 transform_later(slow_region);
1446 } else {
1447 // Just fall from the need-GC path straight into the VM call.
1448 slow_region = needgc_ctrl;
1449 }
1450
1451 InitializeNode* init = alloc->initialization();
1452 fast_oop_rawmem = initialize_object(alloc,
1453 fast_oop_ctrl, fast_oop_rawmem, fast_oop,
1454 klass_node, length, size_in_bytes);
1455
1456 // If initialization is performed by an array copy, any required
1457 // MemBarStoreStore was already added. If the object does not
1458 // escape no need for a MemBarStoreStore. If the object does not
1459 // escape in its initializer and memory barrier (MemBarStoreStore or
1460 // stronger) is already added at exit of initializer, also no need
1461 // for a MemBarStoreStore. Otherwise we need a MemBarStoreStore
1462 // so that stores that initialize this object can't be reordered
1463 // with a subsequent store that makes this object accessible by
1464 // other threads.
1465 // Other threads include java threads and JVM internal threads
1466 // (for example concurrent GC threads). Current concurrent GC
1661 assert(_ioproj_catchall->outcnt() == 0, "all uses must be deleted");
1662 _igvn.remove_dead_node(_ioproj_catchall);
1663 }
1664
1665 // if we generated only a slow call, we are done
1666 if (always_slow) {
1667 // Now we can unhook i_o.
1668 if (result_phi_i_o->outcnt() > 1) {
1669 call->set_req(TypeFunc::I_O, top());
1670 } else {
1671 assert(result_phi_i_o->unique_ctrl_out() == call, "");
1672 // Case of new array with negative size known during compilation.
1673 // AllocateArrayNode::Ideal() optimization disconnect unreachable
1674 // following code since call to runtime will throw exception.
1675 // As result there will be no users of i_o after the call.
1676 // Leave i_o attached to this call to avoid problems in preceding graph.
1677 }
1678 return;
1679 }
1680
1681 if (_fallthroughcatchproj != NULL) {
1682 ctrl = _fallthroughcatchproj->clone();
1683 transform_later(ctrl);
1684 _igvn.replace_node(_fallthroughcatchproj, result_region);
1685 } else {
1686 ctrl = top();
1687 }
1688 Node *slow_result;
1689 if (_resproj == NULL) {
1690 // no uses of the allocation result
1691 slow_result = top();
1692 } else {
1693 slow_result = _resproj->clone();
1694 transform_later(slow_result);
1695 _igvn.replace_node(_resproj, result_phi_rawoop);
1696 }
1697
1698 // Plug slow-path into result merge point
1699 result_region ->init_req( slow_result_path, ctrl );
1700 result_phi_rawoop->init_req( slow_result_path, slow_result);
1701 result_phi_rawmem->init_req( slow_result_path, _memproj_fallthrough );
1702 transform_later(result_region);
1703 transform_later(result_phi_rawoop);
1704 transform_later(result_phi_rawmem);
1705 transform_later(result_phi_i_o);
1706 // This completes all paths into the result merge point
1707 }
1708
1709
1710 // Helper for PhaseMacroExpand::expand_allocate_common.
1711 // Initializes the newly-allocated storage.
1712 Node* PhaseMacroExpand::initialize_object(AllocateNode* alloc,
1713 Node* control, Node* rawmem, Node* object,
1714 Node* klass_node, Node* length,
1715 Node* size_in_bytes) {
1716 InitializeNode* init = alloc->initialization();
1717 // Store the klass & mark bits
1718 Node* mark_node = alloc->make_ideal_mark(&_igvn, object, control, rawmem, klass_node);
1719 if (!mark_node->is_Con()) {
1720 transform_later(mark_node);
1721 }
1722 rawmem = make_store(control, rawmem, object, oopDesc::mark_offset_in_bytes(), mark_node, TypeX_X->basic_type());
1723
1724 BasicType bt = T_METADATA;
1725 Node* metadata = klass_node;
1726 Node* properties = alloc->in(AllocateNode::StorageProperties);
1727 if (properties != NULL) {
1728 // Encode array storage properties into klass pointer
1729 assert(EnableValhalla, "array storage properties not supported");
1730 if (UseCompressedClassPointers) {
1731 // Compress the klass pointer before inserting the storage properties value
1732 metadata = transform_later(new EncodePKlassNode(metadata, metadata->bottom_type()->make_narrowklass()));
1733 metadata = transform_later(new CastN2INode(metadata));
1734 metadata = transform_later(new OrINode(metadata, transform_later(new ConvL2INode(properties))));
1735 bt = T_INT;
1736 } else {
1737 metadata = transform_later(new CastP2XNode(NULL, metadata));
1738 metadata = transform_later(new OrXNode(metadata, properties));
1739 bt = T_LONG;
1740 }
1741 }
1742 rawmem = make_store(control, rawmem, object, oopDesc::klass_offset_in_bytes(), metadata, bt);
1743
1744 int header_size = alloc->minimum_header_size(); // conservatively small
1745
1746 // Array length
1747 if (length != NULL) { // Arrays need length field
1748 rawmem = make_store(control, rawmem, object, arrayOopDesc::length_offset_in_bytes(), length, T_INT);
1749 // conservatively small header size:
1750 header_size = arrayOopDesc::base_offset_in_bytes(T_BYTE);
1751 ciKlass* k = _igvn.type(klass_node)->is_klassptr()->klass();
1752 if (k->is_array_klass()) // we know the exact header size in most cases:
1753 header_size = Klass::layout_helper_header_size(k->layout_helper());
1754 }
1755
1756 // Clear the object body, if necessary.
1757 if (init == NULL) {
1758 // The init has somehow disappeared; be cautious and clear everything.
1759 //
1760 // This can happen if a node is allocated but an uncommon trap occurs
1761 // immediately. In this case, the Initialize gets associated with the
1762 // trap, and may be placed in a different (outer) loop, if the Allocate
1763 // is in a loop. If (this is rare) the inner loop gets unrolled, then
1764 // there can be two Allocates to one Initialize. The answer in all these
1765 // edge cases is safety first. It is always safe to clear immediately
1766 // within an Allocate, and then (maybe or maybe not) clear some more later.
1767 if (!(UseTLAB && ZeroTLAB)) {
1768 rawmem = ClearArrayNode::clear_memory(control, rawmem, object,
1769 alloc->in(AllocateNode::DefaultValue),
1770 alloc->in(AllocateNode::RawDefaultValue),
1771 header_size, size_in_bytes,
1772 &_igvn);
1773 }
1774 } else {
1775 if (!init->is_complete()) {
1776 // Try to win by zeroing only what the init does not store.
1777 // We can also try to do some peephole optimizations,
1778 // such as combining some adjacent subword stores.
1779 rawmem = init->complete_stores(control, rawmem, object,
1780 header_size, size_in_bytes, &_igvn);
1781 }
1782 // We have no more use for this link, since the AllocateNode goes away:
1783 init->set_req(InitializeNode::RawAddress, top());
1784 // (If we keep the link, it just confuses the register allocator,
1785 // who thinks he sees a real use of the address by the membar.)
1786 }
1787
1788 return rawmem;
1789 }
1790
2131 // Replace old box node with new eliminated box for all users
2132 // of the same object and mark related locks as eliminated.
2133 mark_eliminated_box(box, obj);
2134 }
2135 }
2136 }
2137 }
2138
2139 // we have determined that this lock/unlock can be eliminated, we simply
2140 // eliminate the node without expanding it.
2141 //
2142 // Note: The membar's associated with the lock/unlock are currently not
2143 // eliminated. This should be investigated as a future enhancement.
2144 //
2145 bool PhaseMacroExpand::eliminate_locking_node(AbstractLockNode *alock) {
2146
2147 if (!alock->is_eliminated()) {
2148 return false;
2149 }
2150 #ifdef ASSERT
2151 const Type* obj_type = _igvn.type(alock->obj_node());
2152 assert(!obj_type->isa_valuetype() && !obj_type->is_valuetypeptr(), "Eliminating lock on value type");
2153 if (!alock->is_coarsened()) {
2154 // Check that new "eliminated" BoxLock node is created.
2155 BoxLockNode* oldbox = alock->box_node()->as_BoxLock();
2156 assert(oldbox->is_eliminated(), "should be done already");
2157 }
2158 #endif
2159
2160 alock->log_lock_optimization(C, "eliminate_lock");
2161
2162 #ifndef PRODUCT
2163 if (PrintEliminateLocks) {
2164 if (alock->is_Lock()) {
2165 tty->print_cr("++++ Eliminated: %d Lock", alock->_idx);
2166 } else {
2167 tty->print_cr("++++ Eliminated: %d Unlock", alock->_idx);
2168 }
2169 }
2170 #endif
2171
2172 Node* mem = alock->in(TypeFunc::Memory);
2413 // region->in(2) is set to fast path - the object is locked to the current thread.
2414
2415 slow_path->init_req(2, ctrl); // Capture slow-control
2416 slow_mem->init_req(2, fast_lock_mem_phi);
2417
2418 transform_later(slow_path);
2419 transform_later(slow_mem);
2420 // Reset lock's memory edge.
2421 lock->set_req(TypeFunc::Memory, slow_mem);
2422
2423 } else {
2424 region = new RegionNode(3);
2425 // create a Phi for the memory state
2426 mem_phi = new PhiNode( region, Type::MEMORY, TypeRawPtr::BOTTOM);
2427
2428 // Optimize test; set region slot 2
2429 slow_path = opt_bits_test(ctrl, region, 2, flock, 0, 0);
2430 mem_phi->init_req(2, mem);
2431 }
2432
2433 const TypeOopPtr* objptr = _igvn.type(obj)->make_oopptr();
2434 if (objptr->can_be_value_type()) {
2435 // Deoptimize and re-execute if a value
2436 assert(EnableValhalla, "should only be used if value types are enabled");
2437 Node* mark = make_load(slow_path, mem, obj, oopDesc::mark_offset_in_bytes(), TypeX_X, TypeX_X->basic_type());
2438 Node* value_mask = _igvn.MakeConX(markOopDesc::always_locked_pattern);
2439 Node* is_value = _igvn.transform(new AndXNode(mark, value_mask));
2440 Node* cmp = _igvn.transform(new CmpXNode(is_value, value_mask));
2441 Node* bol = _igvn.transform(new BoolNode(cmp, BoolTest::eq));
2442 Node* unc_ctrl = generate_slow_guard(&slow_path, bol, NULL);
2443
2444 int trap_request = Deoptimization::make_trap_request(Deoptimization::Reason_class_check, Deoptimization::Action_none);
2445 address call_addr = SharedRuntime::uncommon_trap_blob()->entry_point();
2446 const TypePtr* no_memory_effects = NULL;
2447 JVMState* jvms = lock->jvms();
2448 CallNode* unc = new CallStaticJavaNode(OptoRuntime::uncommon_trap_Type(), call_addr, "uncommon_trap",
2449 jvms->bci(), no_memory_effects);
2450
2451 unc->init_req(TypeFunc::Control, unc_ctrl);
2452 unc->init_req(TypeFunc::I_O, lock->i_o());
2453 unc->init_req(TypeFunc::Memory, mem); // may gc ptrs
2454 unc->init_req(TypeFunc::FramePtr, lock->in(TypeFunc::FramePtr));
2455 unc->init_req(TypeFunc::ReturnAdr, lock->in(TypeFunc::ReturnAdr));
2456 unc->init_req(TypeFunc::Parms+0, _igvn.intcon(trap_request));
2457 unc->set_cnt(PROB_UNLIKELY_MAG(4));
2458 copy_call_debug_info(lock, unc);
2459
2460 assert(unc->peek_monitor_box() == box, "wrong monitor");
2461 assert(unc->peek_monitor_obj() == obj, "wrong monitor");
2462
2463 // pop monitor and push obj back on stack: we trap before the monitorenter
2464 unc->pop_monitor();
2465 unc->grow_stack(unc->jvms(), 1);
2466 unc->set_stack(unc->jvms(), unc->jvms()->stk_size()-1, obj);
2467
2468 _igvn.register_new_node_with_optimizer(unc);
2469
2470 Node* ctrl = _igvn.transform(new ProjNode(unc, TypeFunc::Control));
2471 Node* halt = _igvn.transform(new HaltNode(ctrl, lock->in(TypeFunc::FramePtr)));
2472 C->root()->add_req(halt);
2473 }
2474
2475 // Make slow path call
2476 CallNode *call = make_slow_call((CallNode *) lock, OptoRuntime::complete_monitor_enter_Type(),
2477 OptoRuntime::complete_monitor_locking_Java(), NULL, slow_path,
2478 obj, box, NULL);
2479
2480 extract_call_projections(call);
2481
2482 // Slow path can only throw asynchronous exceptions, which are always
2483 // de-opted. So the compiler thinks the slow-call can never throw an
2484 // exception. If it DOES throw an exception we would need the debug
2485 // info removed first (since if it throws there is no monitor).
2486 assert ( _ioproj_fallthrough == NULL && _ioproj_catchall == NULL &&
2487 _memproj_catchall == NULL && _catchallcatchproj == NULL, "Unexpected projection from Lock");
2488
2489 // Capture slow path
2490 // disconnect fall-through projection from call and create a new one
2491 // hook up users of fall-through projection to region
2492 Node *slow_ctrl = _fallthroughproj->clone();
2493 transform_later(slow_ctrl);
2494 _igvn.hash_delete(_fallthroughproj);
2556 // No exceptions for unlocking
2557 // Capture slow path
2558 // disconnect fall-through projection from call and create a new one
2559 // hook up users of fall-through projection to region
2560 Node *slow_ctrl = _fallthroughproj->clone();
2561 transform_later(slow_ctrl);
2562 _igvn.hash_delete(_fallthroughproj);
2563 _fallthroughproj->disconnect_inputs(NULL, C);
2564 region->init_req(1, slow_ctrl);
2565 // region inputs are now complete
2566 transform_later(region);
2567 _igvn.replace_node(_fallthroughproj, region);
2568
2569 Node *memproj = transform_later(new ProjNode(call, TypeFunc::Memory) );
2570 mem_phi->init_req(1, memproj );
2571 mem_phi->init_req(2, mem);
2572 transform_later(mem_phi);
2573 _igvn.replace_node(_memproj_fallthrough, mem_phi);
2574 }
2575
2576 // A value type might be returned from the call but we don't know its
2577 // type. Either we get a buffered value (and nothing needs to be done)
2578 // or one of the values being returned is the klass of the value type
2579 // and we need to allocate a value type instance of that type and
2580 // initialize it with other values being returned. In that case, we
2581 // first try a fast path allocation and initialize the value with the
2582 // value klass's pack handler or we fall back to a runtime call.
2583 void PhaseMacroExpand::expand_mh_intrinsic_return(CallStaticJavaNode* call) {
2584 assert(call->method()->is_method_handle_intrinsic(), "must be a method handle intrinsic call");
2585 Node* ret = call->proj_out_or_null(TypeFunc::Parms);
2586 if (ret == NULL) {
2587 return;
2588 }
2589 const TypeFunc* tf = call->_tf;
2590 const TypeTuple* domain = OptoRuntime::store_value_type_fields_Type()->domain_cc();
2591 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain);
2592 call->_tf = new_tf;
2593 // Make sure the change of type is applied before projections are processed by igvn
2594 _igvn.set_type(call, call->Value(&_igvn));
2595 _igvn.set_type(ret, ret->Value(&_igvn));
2596
2597 // Before any new projection is added:
2598 CallProjections* projs = call->extract_projections(true, true);
2599
2600 Node* ctl = new Node(1);
2601 Node* mem = new Node(1);
2602 Node* io = new Node(1);
2603 Node* ex_ctl = new Node(1);
2604 Node* ex_mem = new Node(1);
2605 Node* ex_io = new Node(1);
2606 Node* res = new Node(1);
2607
2608 Node* cast = transform_later(new CastP2XNode(ctl, res));
2609 Node* mask = MakeConX(0x1);
2610 Node* masked = transform_later(new AndXNode(cast, mask));
2611 Node* cmp = transform_later(new CmpXNode(masked, mask));
2612 Node* bol = transform_later(new BoolNode(cmp, BoolTest::eq));
2613 IfNode* allocation_iff = new IfNode(ctl, bol, PROB_MAX, COUNT_UNKNOWN);
2614 transform_later(allocation_iff);
2615 Node* allocation_ctl = transform_later(new IfTrueNode(allocation_iff));
2616 Node* no_allocation_ctl = transform_later(new IfFalseNode(allocation_iff));
2617
2618 Node* no_allocation_res = transform_later(new CheckCastPPNode(no_allocation_ctl, res, TypeInstPtr::BOTTOM));
2619
2620 Node* mask2 = MakeConX(-2);
2621 Node* masked2 = transform_later(new AndXNode(cast, mask2));
2622 Node* rawklassptr = transform_later(new CastX2PNode(masked2));
2623 Node* klass_node = transform_later(new CheckCastPPNode(allocation_ctl, rawklassptr, TypeKlassPtr::OBJECT_OR_NULL));
2624
2625 Node* slowpath_bol = NULL;
2626 Node* top_adr = NULL;
2627 Node* old_top = NULL;
2628 Node* new_top = NULL;
2629 if (UseTLAB) {
2630 Node* end_adr = NULL;
2631 set_eden_pointers(top_adr, end_adr);
2632 Node* end = make_load(ctl, mem, end_adr, 0, TypeRawPtr::BOTTOM, T_ADDRESS);
2633 old_top = new LoadPNode(ctl, mem, top_adr, TypeRawPtr::BOTTOM, TypeRawPtr::BOTTOM, MemNode::unordered);
2634 transform_later(old_top);
2635 Node* layout_val = make_load(NULL, mem, klass_node, in_bytes(Klass::layout_helper_offset()), TypeInt::INT, T_INT);
2636 Node* size_in_bytes = ConvI2X(layout_val);
2637 new_top = new AddPNode(top(), old_top, size_in_bytes);
2638 transform_later(new_top);
2639 Node* slowpath_cmp = new CmpPNode(new_top, end);
2640 transform_later(slowpath_cmp);
2641 slowpath_bol = new BoolNode(slowpath_cmp, BoolTest::ge);
2642 transform_later(slowpath_bol);
2643 } else {
2644 slowpath_bol = intcon(1);
2645 top_adr = top();
2646 old_top = top();
2647 new_top = top();
2648 }
2649 IfNode* slowpath_iff = new IfNode(allocation_ctl, slowpath_bol, PROB_UNLIKELY_MAG(4), COUNT_UNKNOWN);
2650 transform_later(slowpath_iff);
2651
2652 Node* slowpath_true = new IfTrueNode(slowpath_iff);
2653 transform_later(slowpath_true);
2654
2655 CallStaticJavaNode* slow_call = new CallStaticJavaNode(OptoRuntime::store_value_type_fields_Type(),
2656 StubRoutines::store_value_type_fields_to_buf(),
2657 "store_value_type_fields",
2658 call->jvms()->bci(),
2659 TypePtr::BOTTOM);
2660 slow_call->init_req(TypeFunc::Control, slowpath_true);
2661 slow_call->init_req(TypeFunc::Memory, mem);
2662 slow_call->init_req(TypeFunc::I_O, io);
2663 slow_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2664 slow_call->init_req(TypeFunc::ReturnAdr, call->in(TypeFunc::ReturnAdr));
2665 slow_call->init_req(TypeFunc::Parms, res);
2666
2667 Node* slow_ctl = transform_later(new ProjNode(slow_call, TypeFunc::Control));
2668 Node* slow_mem = transform_later(new ProjNode(slow_call, TypeFunc::Memory));
2669 Node* slow_io = transform_later(new ProjNode(slow_call, TypeFunc::I_O));
2670 Node* slow_res = transform_later(new ProjNode(slow_call, TypeFunc::Parms));
2671 Node* slow_catc = transform_later(new CatchNode(slow_ctl, slow_io, 2));
2672 Node* slow_norm = transform_later(new CatchProjNode(slow_catc, CatchProjNode::fall_through_index, CatchProjNode::no_handler_bci));
2673 Node* slow_excp = transform_later(new CatchProjNode(slow_catc, CatchProjNode::catch_all_index, CatchProjNode::no_handler_bci));
2674
2675 Node* ex_r = new RegionNode(3);
2676 Node* ex_mem_phi = new PhiNode(ex_r, Type::MEMORY, TypePtr::BOTTOM);
2677 Node* ex_io_phi = new PhiNode(ex_r, Type::ABIO);
2678 ex_r->init_req(1, slow_excp);
2679 ex_mem_phi->init_req(1, slow_mem);
2680 ex_io_phi->init_req(1, slow_io);
2681 ex_r->init_req(2, ex_ctl);
2682 ex_mem_phi->init_req(2, ex_mem);
2683 ex_io_phi->init_req(2, ex_io);
2684
2685 transform_later(ex_r);
2686 transform_later(ex_mem_phi);
2687 transform_later(ex_io_phi);
2688
2689 Node* slowpath_false = new IfFalseNode(slowpath_iff);
2690 transform_later(slowpath_false);
2691 Node* rawmem = new StorePNode(slowpath_false, mem, top_adr, TypeRawPtr::BOTTOM, new_top, MemNode::unordered);
2692 transform_later(rawmem);
2693 Node* mark_node = makecon(TypeRawPtr::make((address)markOopDesc::always_locked_prototype()));
2694 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::mark_offset_in_bytes(), mark_node, T_ADDRESS);
2695 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_offset_in_bytes(), klass_node, T_METADATA);
2696 if (UseCompressedClassPointers) {
2697 rawmem = make_store(slowpath_false, rawmem, old_top, oopDesc::klass_gap_offset_in_bytes(), intcon(0), T_INT);
2698 }
2699 Node* fixed_block = make_load(slowpath_false, rawmem, klass_node, in_bytes(InstanceKlass::adr_valueklass_fixed_block_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2700 Node* pack_handler = make_load(slowpath_false, rawmem, fixed_block, in_bytes(ValueKlass::pack_handler_offset()), TypeRawPtr::BOTTOM, T_ADDRESS);
2701
2702 CallLeafNoFPNode* handler_call = new CallLeafNoFPNode(OptoRuntime::pack_value_type_Type(),
2703 NULL,
2704 "pack handler",
2705 TypeRawPtr::BOTTOM);
2706 handler_call->init_req(TypeFunc::Control, slowpath_false);
2707 handler_call->init_req(TypeFunc::Memory, rawmem);
2708 handler_call->init_req(TypeFunc::I_O, top());
2709 handler_call->init_req(TypeFunc::FramePtr, call->in(TypeFunc::FramePtr));
2710 handler_call->init_req(TypeFunc::ReturnAdr, top());
2711 handler_call->init_req(TypeFunc::Parms, pack_handler);
2712 handler_call->init_req(TypeFunc::Parms+1, old_top);
2713
2714 // We don't know how many values are returned. This assumes the
2715 // worst case, that all available registers are used.
2716 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2717 if (domain->field_at(i) == Type::HALF) {
2718 slow_call->init_req(i, top());
2719 handler_call->init_req(i+1, top());
2720 continue;
2721 }
2722 Node* proj = transform_later(new ProjNode(call, i));
2723 slow_call->init_req(i, proj);
2724 handler_call->init_req(i+1, proj);
2725 }
2726
2727 // We can safepoint at that new call
2728 copy_call_debug_info(call, slow_call);
2729 transform_later(slow_call);
2730 transform_later(handler_call);
2731
2732 Node* handler_ctl = transform_later(new ProjNode(handler_call, TypeFunc::Control));
2733 rawmem = transform_later(new ProjNode(handler_call, TypeFunc::Memory));
2734 Node* slowpath_false_res = transform_later(new ProjNode(handler_call, TypeFunc::Parms));
2735
2736 MergeMemNode* slowpath_false_mem = MergeMemNode::make(mem);
2737 slowpath_false_mem->set_memory_at(Compile::AliasIdxRaw, rawmem);
2738 transform_later(slowpath_false_mem);
2739
2740 Node* r = new RegionNode(4);
2741 Node* mem_phi = new PhiNode(r, Type::MEMORY, TypePtr::BOTTOM);
2742 Node* io_phi = new PhiNode(r, Type::ABIO);
2743 Node* res_phi = new PhiNode(r, TypeInstPtr::BOTTOM);
2744
2745 r->init_req(1, no_allocation_ctl);
2746 mem_phi->init_req(1, mem);
2747 io_phi->init_req(1, io);
2748 res_phi->init_req(1, no_allocation_res);
2749 r->init_req(2, slow_norm);
2750 mem_phi->init_req(2, slow_mem);
2751 io_phi->init_req(2, slow_io);
2752 res_phi->init_req(2, slow_res);
2753 r->init_req(3, handler_ctl);
2754 mem_phi->init_req(3, slowpath_false_mem);
2755 io_phi->init_req(3, io);
2756 res_phi->init_req(3, slowpath_false_res);
2757
2758 transform_later(r);
2759 transform_later(mem_phi);
2760 transform_later(io_phi);
2761 transform_later(res_phi);
2762
2763 assert(projs->nb_resproj == 1, "unexpected number of results");
2764 _igvn.replace_in_uses(projs->fallthrough_catchproj, r);
2765 _igvn.replace_in_uses(projs->fallthrough_memproj, mem_phi);
2766 _igvn.replace_in_uses(projs->fallthrough_ioproj, io_phi);
2767 _igvn.replace_in_uses(projs->resproj[0], res_phi);
2768 _igvn.replace_in_uses(projs->catchall_catchproj, ex_r);
2769 _igvn.replace_in_uses(projs->catchall_memproj, ex_mem_phi);
2770 _igvn.replace_in_uses(projs->catchall_ioproj, ex_io_phi);
2771
2772 _igvn.replace_node(ctl, projs->fallthrough_catchproj);
2773 _igvn.replace_node(mem, projs->fallthrough_memproj);
2774 _igvn.replace_node(io, projs->fallthrough_ioproj);
2775 _igvn.replace_node(res, projs->resproj[0]);
2776 _igvn.replace_node(ex_ctl, projs->catchall_catchproj);
2777 _igvn.replace_node(ex_mem, projs->catchall_memproj);
2778 _igvn.replace_node(ex_io, projs->catchall_ioproj);
2779 }
2780
2781 //---------------------------eliminate_macro_nodes----------------------
2782 // Eliminate scalar replaced allocations and associated locks.
2783 void PhaseMacroExpand::eliminate_macro_nodes() {
2784 if (C->macro_count() == 0)
2785 return;
2786
2787 // First, attempt to eliminate locks
2788 int cnt = C->macro_count();
2789 for (int i=0; i < cnt; i++) {
2790 Node *n = C->macro_node(i);
2791 if (n->is_AbstractLock()) { // Lock and Unlock nodes
2792 // Before elimination mark all associated (same box and obj)
2793 // lock and unlock nodes.
2794 mark_eliminated_locking_nodes(n->as_AbstractLock());
2795 }
2796 }
2797 bool progress = true;
2798 while (progress) {
2799 progress = false;
2800 for (int i = C->macro_count(); i > 0; i--) {
2805 success = eliminate_locking_node(n->as_AbstractLock());
2806 }
2807 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2808 progress = progress || success;
2809 }
2810 }
2811 // Next, attempt to eliminate allocations
2812 _has_locks = false;
2813 progress = true;
2814 while (progress) {
2815 progress = false;
2816 for (int i = C->macro_count(); i > 0; i--) {
2817 Node * n = C->macro_node(i-1);
2818 bool success = false;
2819 debug_only(int old_macro_count = C->macro_count(););
2820 switch (n->class_id()) {
2821 case Node::Class_Allocate:
2822 case Node::Class_AllocateArray:
2823 success = eliminate_allocate_node(n->as_Allocate());
2824 break;
2825 case Node::Class_CallStaticJava: {
2826 CallStaticJavaNode* call = n->as_CallStaticJava();
2827 if (!call->method()->is_method_handle_intrinsic()) {
2828 success = eliminate_boxing_node(n->as_CallStaticJava());
2829 }
2830 break;
2831 }
2832 case Node::Class_Lock:
2833 case Node::Class_Unlock:
2834 assert(!n->as_AbstractLock()->is_eliminated(), "sanity");
2835 _has_locks = true;
2836 break;
2837 case Node::Class_ArrayCopy:
2838 break;
2839 case Node::Class_OuterStripMinedLoop:
2840 break;
2841 default:
2842 assert(n->Opcode() == Op_LoopLimit ||
2843 n->Opcode() == Op_Opaque1 ||
2844 n->Opcode() == Op_Opaque2 ||
2845 n->Opcode() == Op_Opaque3 ||
2846 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(n),
2847 "unknown node type in macro list");
2848 }
2849 assert(success == (C->macro_count() < old_macro_count), "elimination reduces macro count");
2850 progress = progress || success;
2851 }
2861 // Make sure expansion will not cause node limit to be exceeded.
2862 // Worst case is a macro node gets expanded into about 200 nodes.
2863 // Allow 50% more for optimization.
2864 if (C->check_node_count(C->macro_count() * 300, "out of nodes before macro expansion" ) )
2865 return true;
2866
2867 // Eliminate Opaque and LoopLimit nodes. Do it after all loop optimizations.
2868 bool progress = true;
2869 while (progress) {
2870 progress = false;
2871 for (int i = C->macro_count(); i > 0; i--) {
2872 Node * n = C->macro_node(i-1);
2873 bool success = false;
2874 debug_only(int old_macro_count = C->macro_count(););
2875 if (n->Opcode() == Op_LoopLimit) {
2876 // Remove it from macro list and put on IGVN worklist to optimize.
2877 C->remove_macro_node(n);
2878 _igvn._worklist.push(n);
2879 success = true;
2880 } else if (n->Opcode() == Op_CallStaticJava) {
2881 CallStaticJavaNode* call = n->as_CallStaticJava();
2882 if (!call->method()->is_method_handle_intrinsic()) {
2883 // Remove it from macro list and put on IGVN worklist to optimize.
2884 C->remove_macro_node(n);
2885 _igvn._worklist.push(n);
2886 success = true;
2887 }
2888 } else if (n->Opcode() == Op_Opaque1 || n->Opcode() == Op_Opaque2) {
2889 _igvn.replace_node(n, n->in(1));
2890 success = true;
2891 #if INCLUDE_RTM_OPT
2892 } else if ((n->Opcode() == Op_Opaque3) && ((Opaque3Node*)n)->rtm_opt()) {
2893 assert(C->profile_rtm(), "should be used only in rtm deoptimization code");
2894 assert((n->outcnt() == 1) && n->unique_out()->is_Cmp(), "");
2895 Node* cmp = n->unique_out();
2896 #ifdef ASSERT
2897 // Validate graph.
2898 assert((cmp->outcnt() == 1) && cmp->unique_out()->is_Bool(), "");
2899 BoolNode* bol = cmp->unique_out()->as_Bool();
2900 assert((bol->outcnt() == 1) && bol->unique_out()->is_If() &&
2901 (bol->_test._test == BoolTest::ne), "");
2902 IfNode* ifn = bol->unique_out()->as_If();
2903 assert((ifn->outcnt() == 2) &&
2904 ifn->proj_out(1)->is_uncommon_trap_proj(Deoptimization::Reason_rtm_state_change) != NULL, "");
2905 #endif
2906 Node* repl = n->in(1);
2907 if (!_has_locks) {
2947 int macro_count = C->macro_count();
2948 Node * n = C->macro_node(macro_count-1);
2949 assert(n->is_macro(), "only macro nodes expected here");
2950 if (_igvn.type(n) == Type::TOP || (n->in(0) != NULL && n->in(0)->is_top())) {
2951 // node is unreachable, so don't try to expand it
2952 C->remove_macro_node(n);
2953 continue;
2954 }
2955 switch (n->class_id()) {
2956 case Node::Class_Allocate:
2957 expand_allocate(n->as_Allocate());
2958 break;
2959 case Node::Class_AllocateArray:
2960 expand_allocate_array(n->as_AllocateArray());
2961 break;
2962 case Node::Class_Lock:
2963 expand_lock_node(n->as_Lock());
2964 break;
2965 case Node::Class_Unlock:
2966 expand_unlock_node(n->as_Unlock());
2967 break;
2968 case Node::Class_CallStaticJava:
2969 expand_mh_intrinsic_return(n->as_CallStaticJava());
2970 C->remove_macro_node(n);
2971 break;
2972 default:
2973 assert(false, "unknown node type in macro list");
2974 }
2975 assert(C->macro_count() < macro_count, "must have deleted a node from macro list");
2976 if (C->failing()) return true;
2977 }
2978
2979 _igvn.set_delay_transform(false);
2980 _igvn.optimize();
2981 if (C->failing()) return true;
2982 return false;
2983 }
|