15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "gc/shared/barrierSet.hpp"
26 #include "gc/shared/c2/barrierSetC2.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "oops/objArrayKlass.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/cfgnode.hpp"
33 #include "opto/connode.hpp"
34 #include "opto/convertnode.hpp"
35 #include "opto/loopnode.hpp"
36 #include "opto/machnode.hpp"
37 #include "opto/movenode.hpp"
38 #include "opto/mulnode.hpp"
39 #include "opto/narrowptrnode.hpp"
40 #include "opto/phaseX.hpp"
41 #include "opto/regalloc.hpp"
42 #include "opto/regmask.hpp"
43 #include "opto/runtime.hpp"
44 #include "opto/subnode.hpp"
45 #include "opto/vectornode.hpp"
46 #include "utilities/vmError.hpp"
47
48 // Portions of code courtesy of Clifford Click
49
50 // Optimization - Graph Style
51
52 //=============================================================================
53 //------------------------------Value------------------------------------------
54 // Compute the type of the RegionNode.
503 if (left_path == nullptr || right_path == nullptr) {
504 return false;
505 }
506 Node* diamond_if = left_path->in(0);
507 if (diamond_if == nullptr || !diamond_if->is_If() || diamond_if != right_path->in(0)) {
508 // Not an IfNode merging a diamond or TOP.
509 return false;
510 }
511
512 // Check for a proper bool/cmp
513 const Node* bol = diamond_if->in(1);
514 if (!bol->is_Bool()) {
515 return false;
516 }
517 const Node* cmp = bol->in(1);
518 if (!cmp->is_Cmp()) {
519 return false;
520 }
521 return true;
522 }
523 //------------------------------Ideal------------------------------------------
524 // Return a node which is more "ideal" than the current node. Must preserve
525 // the CFG, but we can still strip out dead paths.
526 Node *RegionNode::Ideal(PhaseGVN *phase, bool can_reshape) {
527 if( !can_reshape && !in(0) ) return nullptr; // Already degraded to a Copy
528 assert(!in(0) || !in(0)->is_Root(), "not a specially hidden merge");
529
530 // Check for RegionNode with no Phi users and both inputs come from either
531 // arm of the same IF. If found, then the control-flow split is useless.
532 bool has_phis = false;
533 if (can_reshape) { // Need DU info to check for Phi users
534 try_clean_mem_phis(phase->is_IterGVN());
535 has_phis = (has_phi() != nullptr); // Cache result
536
537 if (!has_phis) { // No Phi users? Nothing merging?
538 for (uint i = 1; i < req()-1; i++) {
539 Node *if1 = in(i);
540 if( !if1 ) continue;
541 Node *iff = if1->in(0);
542 if( !iff || !iff->is_If() ) continue;
949 if (iff1 == iff2) {
950 igvn->add_users_to_worklist(iff1); // Make sure dead if is eliminated
951 igvn->replace_input_of(region, idx1, iff1->in(0));
952 igvn->replace_input_of(region, idx2, igvn->C->top());
953 return (region == this); // Remove useless if (both projections map to the same control/value)
954 }
955 BoolNode* bol1 = iff1->in(1)->isa_Bool();
956 BoolNode* bol2 = iff2->in(1)->isa_Bool();
957 if (bol1 == nullptr || bol2 == nullptr) {
958 return false; // No bool inputs found
959 }
960 Node* cmp1 = bol1->in(1);
961 Node* cmp2 = bol2->in(1);
962 bool commute = false;
963 if (!cmp1->is_Cmp() || !cmp2->is_Cmp()) {
964 return false; // No comparison
965 } else if (cmp1->Opcode() == Op_CmpF || cmp1->Opcode() == Op_CmpD ||
966 cmp2->Opcode() == Op_CmpF || cmp2->Opcode() == Op_CmpD ||
967 cmp1->Opcode() == Op_CmpP || cmp1->Opcode() == Op_CmpN ||
968 cmp2->Opcode() == Op_CmpP || cmp2->Opcode() == Op_CmpN ||
969 cmp1->is_SubTypeCheck() || cmp2->is_SubTypeCheck()) {
970 // Floats and pointers don't exactly obey trichotomy. To be on the safe side, don't transform their tests.
971 // SubTypeCheck is not commutative
972 return false;
973 } else if (cmp1 != cmp2) {
974 if (cmp1->in(1) == cmp2->in(2) &&
975 cmp1->in(2) == cmp2->in(1)) {
976 commute = true; // Same but swapped inputs, commute the test
977 } else {
978 return false; // Ifs are not comparing the same values
979 }
980 }
981 proj1 = proj1->other_if_proj();
982 proj2 = proj2->other_if_proj();
983 if (!((proj1->unique_ctrl_out_or_null() == iff2 &&
984 proj2->unique_ctrl_out_or_null() == this) ||
985 (proj2->unique_ctrl_out_or_null() == iff1 &&
986 proj1->unique_ctrl_out_or_null() == this))) {
987 return false; // Ifs are not connected through other projs
988 }
989 // Found 'iff -> proj -> iff -> proj -> this' shape where all other projs are merged
1028 st->print("#reducible ");
1029 break;
1030 case RegionNode::LoopStatus::NeverIrreducibleEntry:
1031 break; // nothing
1032 }
1033 }
1034 #endif
1035
1036 // Find the one non-null required input. RegionNode only
1037 Node *Node::nonnull_req() const {
1038 assert( is_Region(), "" );
1039 for( uint i = 1; i < _cnt; i++ )
1040 if( in(i) )
1041 return in(i);
1042 ShouldNotReachHere();
1043 return nullptr;
1044 }
1045
1046
1047 //=============================================================================
1048 // note that these functions assume that the _adr_type field is flattened
1049 uint PhiNode::hash() const {
1050 const Type* at = _adr_type;
1051 return TypeNode::hash() + (at ? at->hash() : 0);
1052 }
1053 bool PhiNode::cmp( const Node &n ) const {
1054 return TypeNode::cmp(n) && _adr_type == ((PhiNode&)n)._adr_type;
1055 }
1056 static inline
1057 const TypePtr* flatten_phi_adr_type(const TypePtr* at) {
1058 if (at == nullptr || at == TypePtr::BOTTOM) return at;
1059 return Compile::current()->alias_type(at)->adr_type();
1060 }
1061
1062 //----------------------------make---------------------------------------------
1063 // create a new phi with edges matching r and set (initially) to x
1064 PhiNode* PhiNode::make(Node* r, Node* x, const Type *t, const TypePtr* at) {
1065 uint preds = r->req(); // Number of predecessor paths
1066 assert(t != Type::MEMORY || at == flatten_phi_adr_type(at), "flatten at");
1067 PhiNode* p = new PhiNode(r, t, at);
1068 for (uint j = 1; j < preds; j++) {
1069 // Fill in all inputs, except those which the region does not yet have
1070 if (r->in(j) != nullptr)
1071 p->init_req(j, x);
1072 }
1073 return p;
1074 }
1075 PhiNode* PhiNode::make(Node* r, Node* x) {
1076 const Type* t = x->bottom_type();
1077 const TypePtr* at = nullptr;
1078 if (t == Type::MEMORY) at = flatten_phi_adr_type(x->adr_type());
1079 return make(r, x, t, at);
1080 }
1081 PhiNode* PhiNode::make_blank(Node* r, Node* x) {
1082 const Type* t = x->bottom_type();
1083 const TypePtr* at = nullptr;
1084 if (t == Type::MEMORY) at = flatten_phi_adr_type(x->adr_type());
1085 return new PhiNode(r, t, at);
1086 }
1175 np->as_Phi()->verify_adr_type(visited, at);
1176 } else if (n->bottom_type() == Type::TOP
1177 || (n->is_Mem() && n->in(MemNode::Address)->bottom_type() == Type::TOP)) {
1178 // ignore top inputs
1179 } else {
1180 const TypePtr* nat = flatten_phi_adr_type(n->adr_type());
1181 // recheck phi/non-phi consistency at leaves:
1182 assert((nat != nullptr) == (at != nullptr), "");
1183 assert(nat == at || nat == TypePtr::BOTTOM,
1184 "adr_type must be consistent at leaves of phi nest");
1185 }
1186 }
1187 }
1188
1189 // Verify a whole nest of phis rooted at this one.
1190 void PhiNode::verify_adr_type(bool recursive) const {
1191 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
1192 if (Node::in_dump()) return; // muzzle asserts when printing
1193
1194 assert((_type == Type::MEMORY) == (_adr_type != nullptr), "adr_type for memory phis only");
1195
1196 if (!VerifyAliases) return; // verify thoroughly only if requested
1197
1198 assert(_adr_type == flatten_phi_adr_type(_adr_type),
1199 "Phi::adr_type must be pre-normalized");
1200
1201 if (recursive) {
1202 VectorSet visited;
1203 verify_adr_type(visited, _adr_type);
1204 }
1205 }
1206 #endif
1207
1208
1209 //------------------------------Value------------------------------------------
1210 // Compute the type of the PhiNode
1211 const Type* PhiNode::Value(PhaseGVN* phase) const {
1212 Node *r = in(0); // RegionNode
1213 if( !r ) // Copy or dead
1214 return in(1) ? phase->type(in(1)) : Type::TOP;
1458 assert(req() == 3, "same as region");
1459 RegionNode* region = in(0)->as_Region();
1460 for (uint i = 1; i < 3; i++) {
1461 Node* phi_input = in(i);
1462 if (phi_input != nullptr && phi_input->is_MergeMem() && region->in(i)->outcnt() == 1) {
1463 // Nothing is control-dependent on path #i except the region itself.
1464 MergeMemNode* merge_mem = phi_input->as_MergeMem();
1465 uint j = 3 - i;
1466 Node* other_phi_input = in(j);
1467 if (other_phi_input != nullptr && other_phi_input == merge_mem->base_memory() && !is_data_loop(region, phi_input, igvn)) {
1468 // merge_mem is a successor memory to other_phi_input, and is not pinned inside the diamond, so push it out.
1469 // Only proceed if the transformation doesn't create a data loop
1470 // This will allow the diamond to collapse completely if there are no other phis left.
1471 igvn->replace_node(this, merge_mem);
1472 return true;
1473 }
1474 }
1475 }
1476 return false;
1477 }
1478 //----------------------------check_cmove_id-----------------------------------
1479 // Check for CMove'ing a constant after comparing against the constant.
1480 // Happens all the time now, since if we compare equality vs a constant in
1481 // the parser, we "know" the variable is constant on one path and we force
1482 // it. Thus code like "if( x==0 ) {/*EMPTY*/}" ends up inserting a
1483 // conditional move: "x = (x==0)?0:x;". Yucko. This fix is slightly more
1484 // general in that we don't need constants. Since CMove's are only inserted
1485 // in very special circumstances, we do it here on generic Phi's.
1486 Node* PhiNode::is_cmove_id(PhaseTransform* phase, int true_path) {
1487 assert(true_path !=0, "only diamond shape graph expected");
1488
1489 // is_diamond_phi() has guaranteed the correctness of the nodes sequence:
1490 // phi->region->if_proj->ifnode->bool->cmp
1491 Node* region = in(0);
1492 Node* iff = region->in(1)->in(0);
1493 BoolNode* b = iff->in(1)->as_Bool();
1494 Node* cmp = b->in(1);
1495 Node* tval = in(true_path);
1496 Node* fval = in(3-true_path);
1497 Node* id = CMoveNode::is_cmove_id(phase, cmp, tval, fval, b);
1512 }
1513
1514 return id;
1515 }
1516
1517 //------------------------------Identity---------------------------------------
1518 // Check for Region being Identity.
1519 Node* PhiNode::Identity(PhaseGVN* phase) {
1520 if (must_wait_for_region_in_irreducible_loop(phase)) {
1521 return this;
1522 }
1523 // Check for no merging going on
1524 // (There used to be special-case code here when this->region->is_Loop.
1525 // It would check for a tributary phi on the backedge that the main phi
1526 // trivially, perhaps with a single cast. The unique_input method
1527 // does all this and more, by reducing such tributaries to 'this'.)
1528 Node* uin = unique_input(phase, false);
1529 if (uin != nullptr) {
1530 return uin;
1531 }
1532
1533 int true_path = is_diamond_phi();
1534 // Delay CMove'ing identity if Ideal has not had the chance to handle unsafe cases, yet.
1535 if (true_path != 0 && !(phase->is_IterGVN() && wait_for_region_igvn(phase))) {
1536 Node* id = is_cmove_id(phase, true_path);
1537 if (id != nullptr) {
1538 return id;
1539 }
1540 }
1541
1542 // Looking for phis with identical inputs. If we find one that has
1543 // type TypePtr::BOTTOM, replace the current phi with the bottom phi.
1544 if (phase->is_IterGVN() && type() == Type::MEMORY && adr_type() !=
1545 TypePtr::BOTTOM && !adr_type()->is_known_instance()) {
1546 uint phi_len = req();
1547 Node* phi_reg = region();
1548 for (DUIterator_Fast imax, i = phi_reg->fast_outs(imax); i < imax; i++) {
1549 Node* u = phi_reg->fast_out(i);
1550 assert(!u->is_Phi() || u->in(0) == phi_reg, "broken Phi/Region subgraph");
1551 if (u->is_Phi() && u->req() == phi_len && can_be_replaced_by(u->as_Phi())) {
1602 }
1603 // Check for a unique input (maybe uncasted)
1604 if (input == nullptr) {
1605 input = un;
1606 } else if (input != un) {
1607 input = NodeSentinel; // no unique input
1608 }
1609 }
1610 if (input == nullptr) {
1611 return phase->C->top(); // no inputs
1612 }
1613
1614 if (input != NodeSentinel) {
1615 return input; // one unique direct input
1616 }
1617
1618 // Nothing.
1619 return nullptr;
1620 }
1621
1622 //------------------------------is_x2logic-------------------------------------
1623 // Check for simple convert-to-boolean pattern
1624 // If:(C Bool) Region:(IfF IfT) Phi:(Region 0 1)
1625 // Convert Phi to an ConvIB.
1626 static Node *is_x2logic( PhaseGVN *phase, PhiNode *phi, int true_path ) {
1627 assert(true_path !=0, "only diamond shape graph expected");
1628
1629 // If we're late in the optimization process, we may have already expanded Conv2B nodes
1630 if (phase->C->post_loop_opts_phase() && !Matcher::match_rule_supported(Op_Conv2B)) {
1631 return nullptr;
1632 }
1633
1634 // Convert the true/false index into an expected 0/1 return.
1635 // Map 2->0 and 1->1.
1636 int flipped = 2-true_path;
1637
1638 // is_diamond_phi() has guaranteed the correctness of the nodes sequence:
1639 // phi->region->if_proj->ifnode->bool->cmp
1640 Node *region = phi->in(0);
1641 Node *iff = region->in(1)->in(0);
2069
2070 if (rc->in(0)->in(1) == nullptr || !rc->in(0)->in(1)->is_Bool()) { continue; }
2071 if (worklist.member(rc->in(0)->in(1))) {
2072 delay = true;
2073 break;
2074 }
2075
2076 if (rc->in(0)->in(1)->in(1) == nullptr || !rc->in(0)->in(1)->in(1)->is_Cmp()) { continue; }
2077 if (worklist.member(rc->in(0)->in(1)->in(1))) {
2078 delay = true;
2079 break;
2080 }
2081 }
2082
2083 if (delay) {
2084 worklist.push(this);
2085 }
2086 return delay;
2087 }
2088
2089 // If the Phi's Region is in an irreducible loop, and the Region
2090 // has had an input removed, but not yet transformed, it could be
2091 // that the Region (and this Phi) are not reachable from Root.
2092 // If we allow the Phi to collapse before the Region, this may lead
2093 // to dead-loop data. Wait for the Region to check for reachability,
2094 // and potentially remove the dead code.
2095 bool PhiNode::must_wait_for_region_in_irreducible_loop(PhaseGVN* phase) const {
2096 RegionNode* region = in(0)->as_Region();
2097 if (region->loop_status() == RegionNode::LoopStatus::MaybeIrreducibleEntry) {
2098 Node* top = phase->C->top();
2099 for (uint j = 1; j < req(); j++) {
2100 Node* rc = region->in(j); // for each control input
2101 if (rc == nullptr || phase->type(rc) == Type::TOP) {
2102 // Region is missing a control input
2103 Node* n = in(j);
2104 if (n != nullptr && n != top) {
2105 // Phi still has its input, so region just lost its input
2106 return true;
2107 }
2108 }
2505 // Phi (this) |
2506 // | |
2507 // +-----------+
2508 //
2509 // Generally, there are issues with non-termination with such circularity
2510 // (see comment further below). However, if there is a direct loop to self,
2511 // splitting the Phi through the MergeMem will result in the below.
2512 //
2513 // +---+
2514 // | |
2515 // v |
2516 // Phi |
2517 // |\ |
2518 // | +-+
2519 // (base_memory) v
2520 // MergeMem
2521 //
2522 // This split breaks the circularity and consequently does not lead to
2523 // non-termination.
2524 uint merge_width = 0;
2525 bool split_always_terminates = false; // Is splitting guaranteed to terminate?
2526 for( uint i=1; i<req(); ++i ) {// For all paths in
2527 Node *ii = in(i);
2528 // TOP inputs should not be counted as safe inputs because if the
2529 // Phi references itself through all other inputs then splitting the
2530 // Phi through memory merges would create dead loop at later stage.
2531 if (ii == top) {
2532 return nullptr; // Delay optimization until graph is cleaned.
2533 }
2534 if (ii->is_MergeMem()) {
2535 MergeMemNode* n = ii->as_MergeMem();
2536 merge_width = MAX2(merge_width, n->req());
2537 if (n->base_memory() == this) {
2538 split_always_terminates = true;
2539 }
2540 }
2541 }
2542
2543 // There are cases with circular dependencies between bottom Phis
2544 // and MergeMems. Below is a minimal example.
2545 //
2546 // +------------+
2547 // | |
2548 // (base_memory) v |
2549 // MergeMem |
2550 // | |
2551 // v |
2552 // Phi (this) |
2553 // | |
2554 // v |
2555 // Phi |
2556 // | |
2557 // +----------+
2558 //
2559 // Here, we cannot break the circularity through a self-loop as there
2560 // are two Phis involved. Repeatedly splitting the Phis through the
2561 // MergeMem leads to non-termination. We check for non-termination below.
2562 // Only check for non-termination if necessary.
2563 if (!split_always_terminates && adr_type() == TypePtr::BOTTOM &&
2564 merge_width > Compile::AliasIdxRaw) {
2565 split_always_terminates = is_split_through_mergemem_terminating();
2566 }
2567
2568 if (merge_width > Compile::AliasIdxRaw) {
2569 // found at least one non-empty MergeMem
2570 const TypePtr* at = adr_type();
2571 if (at != TypePtr::BOTTOM) {
2572 // Patch the existing phi to select an input from the merge:
2573 // Phi:AT1(...MergeMem(m0, m1, m2)...) into
2574 // Phi:AT1(...m1...)
2575 int alias_idx = phase->C->get_alias_index(at);
2576 for (uint i=1; i<req(); ++i) {
2577 Node *ii = in(i);
2578 if (ii->is_MergeMem()) {
2579 MergeMemNode* n = ii->as_MergeMem();
2580 // compress paths and change unreachable cycles to TOP
2581 // If not, we can update the input infinitely along a MergeMem cycle
2582 // Equivalent code is in MemNode::Ideal_common
2583 Node *m = phase->transform(n);
2584 if (outcnt() == 0) { // Above transform() may kill us!
2585 return top;
2586 }
2587 // If transformed to a MergeMem, get the desired slice
2588 // Otherwise the returned node represents memory for every slice
2589 Node *new_mem = (m->is_MergeMem()) ?
2590 m->as_MergeMem()->memory_at(alias_idx) : m;
2591 // Update input if it is progress over what we have now
2592 if (new_mem != ii) {
2593 set_req_X(i, new_mem, phase->is_IterGVN());
2594 progress = this;
2595 }
2596 }
2597 }
2598 } else if (split_always_terminates) {
2599 // If all inputs reference this phi (directly or through data nodes) -
2600 // it is a dead loop.
2601 bool saw_safe_input = false;
2602 for (uint j = 1; j < req(); ++j) {
2603 Node* n = in(j);
2604 if (n->is_MergeMem()) {
2605 MergeMemNode* mm = n->as_MergeMem();
2606 if (mm->base_memory() == this || mm->base_memory() == mm->empty_memory()) {
2607 // Skip this input if it references back to this phi or if the memory path is dead
2608 continue;
2609 }
2610 }
2611 if (!is_unsafe_data_reference(n)) {
2612 saw_safe_input = true; // found safe input
2613 break;
2614 }
2615 }
2616 if (!saw_safe_input) {
2617 // There is a dead loop: All inputs are either dead or reference back to this phi
2618 return top;
2619 }
2620
2621 // Phi(...MergeMem(m0, m1:AT1, m2:AT2)...) into
2622 // MergeMem(Phi(...m0...), Phi:AT1(...m1...), Phi:AT2(...m2...))
2623 PhaseIterGVN* igvn = phase->is_IterGVN();
2624 assert(igvn != nullptr, "sanity check");
2625 PhiNode* new_base = (PhiNode*) clone();
2626 // Must eagerly register phis, since they participate in loops.
2627 igvn->register_new_node_with_optimizer(new_base);
2628
2629 MergeMemNode* result = MergeMemNode::make(new_base);
2630 for (uint i = 1; i < req(); ++i) {
2631 Node *ii = in(i);
2632 if (ii->is_MergeMem()) {
2633 MergeMemNode* n = ii->as_MergeMem();
2634 for (MergeMemStream mms(result, n); mms.next_non_empty2(); ) {
2635 // If we have not seen this slice yet, make a phi for it.
2636 bool made_new_phi = false;
2637 if (mms.is_empty()) {
2638 Node* new_phi = new_base->slice_memory(mms.adr_type(phase->C));
2639 made_new_phi = true;
2640 igvn->register_new_node_with_optimizer(new_phi);
2641 mms.set_memory(new_phi);
2642 }
2643 Node* phi = mms.memory();
2644 assert(made_new_phi || phi->in(i) == n, "replace the i-th merge by a slice");
2645 phi->set_req_X(i, mms.memory2(), phase);
2646 }
2647 }
2648 }
2649 // Distribute all self-loops.
2650 { // (Extra braces to hide mms.)
2651 for (MergeMemStream mms(result); mms.next_non_empty(); ) {
2652 Node* phi = mms.memory();
2653 for (uint i = 1; i < req(); ++i) {
2658 }
2659 }
2660
2661 // We could immediately transform the new Phi nodes here, but that can
2662 // result in creating an excessive number of new nodes within a single
2663 // IGVN iteration. We have put the Phi nodes on the IGVN worklist, so
2664 // they are transformed later on in any case.
2665
2666 // Replace self with the result.
2667 return result;
2668 }
2669 }
2670 //
2671 // Other optimizations on the memory chain
2672 //
2673 const TypePtr* at = adr_type();
2674 for( uint i=1; i<req(); ++i ) {// For all paths in
2675 Node *ii = in(i);
2676 Node *new_in = MemNode::optimize_memory_chain(ii, at, nullptr, phase);
2677 if (ii != new_in ) {
2678 set_req_X(i, new_in, phase);
2679 progress = this;
2680 }
2681 }
2682 }
2683
2684 #ifdef _LP64
2685 // Push DecodeN/DecodeNKlass down through phi.
2686 // The rest of phi graph will transform by split EncodeP node though phis up.
2687 if ((UseCompressedOops || UseCompressedClassPointers) && can_reshape && progress == nullptr) {
2688 bool may_push = true;
2689 bool has_decodeN = false;
2690 bool is_decodeN = false;
2691 for (uint i=1; i<req(); ++i) {// For all paths in
2692 Node *ii = in(i);
2693 if (ii->is_DecodeNarrowPtr() && ii->bottom_type() == bottom_type()) {
2694 // Do optimization if a non dead path exist.
2695 if (ii->in(1)->bottom_type() != Type::TOP) {
2696 has_decodeN = true;
2697 is_decodeN = ii->is_DecodeN();
2698 }
2726 if (is_decodeN) {
2727 new_ii = new EncodePNode(ii, narrow_t);
2728 } else {
2729 new_ii = new EncodePKlassNode(ii, narrow_t);
2730 }
2731 igvn->register_new_node_with_optimizer(new_ii);
2732 }
2733 }
2734 new_phi->set_req(i, new_ii);
2735 }
2736 igvn->register_new_node_with_optimizer(new_phi, this);
2737 if (is_decodeN) {
2738 progress = new DecodeNNode(new_phi, bottom_type());
2739 } else {
2740 progress = new DecodeNKlassNode(new_phi, bottom_type());
2741 }
2742 }
2743 }
2744 #endif
2745
2746 // Try to convert a Phi with two duplicated convert nodes into a phi of the pre-conversion type and the convert node
2747 // proceeding the phi, to de-duplicate the convert node and compact the IR.
2748 if (can_reshape && progress == nullptr) {
2749 ConvertNode* convert = in(1)->isa_Convert();
2750 if (convert != nullptr) {
2751 int conv_op = convert->Opcode();
2752 bool ok = true;
2753
2754 // Check the rest of the inputs
2755 for (uint i = 2; i < req(); i++) {
2756 // Make sure that all inputs are of the same type of convert node
2757 if (in(i)->Opcode() != conv_op) {
2758 ok = false;
2759 break;
2760 }
2761 }
2762
2763 if (ok) {
2764 // Find the local bottom type to set as the type of the phi
2765 const Type* source_type = Type::get_const_basic_type(convert->in_type()->basic_type());
2788 // only after the non-bottom memory phi is processed by igvn, PhiNode::Identity doesn't run and the transformation
2789 // doesn't happen.
2790 // Look for non-bottom Phis that should be transformed and enqueue them for igvn so that PhiNode::Identity executes for
2791 // them.
2792 if (can_reshape && type() == Type::MEMORY && adr_type() == TypePtr::BOTTOM) {
2793 PhaseIterGVN* igvn = phase->is_IterGVN();
2794 uint phi_len = req();
2795 Node* phi_reg = region();
2796 for (DUIterator_Fast imax, i = phi_reg->fast_outs(imax); i < imax; i++) {
2797 Node* u = phi_reg->fast_out(i);
2798 assert(!u->is_Phi() || (u->in(0) == phi_reg && u->req() == phi_len), "broken Phi/Region subgraph");
2799 if (u->is_Phi() && u->as_Phi()->can_be_replaced_by(this)) {
2800 igvn->_worklist.push(u);
2801 }
2802 }
2803 }
2804
2805 return progress; // Return any progress
2806 }
2807
2808 static int compare_types(const Type* const& e1, const Type* const& e2) {
2809 return (intptr_t)e1 - (intptr_t)e2;
2810 }
2811
2812 // Collect types at casts that are going to be eliminated at that Phi and store them in a TypeTuple.
2813 // Sort the types using an arbitrary order so a list of some types always hashes to the same TypeTuple (and TypeTuple
2814 // pointer comparison is enough to tell if 2 list of types are the same or not)
2815 const TypeTuple* PhiNode::collect_types(PhaseGVN* phase) const {
2816 const Node* region = in(0);
2817 const Type* phi_type = bottom_type();
2818 ResourceMark rm;
2819 GrowableArray<const Type*> types;
2820 for (uint i = 1; i < req(); i++) {
2821 if (region->in(i) == nullptr || phase->type(region->in(i)) == Type::TOP) {
2822 continue;
2823 }
2824 Node* in = Node::in(i);
2825 const Type* t = phase->type(in);
2826 if (in == nullptr || in == this || t == Type::TOP) {
2827 continue;
3177 #ifndef PRODUCT
3178 void CatchProjNode::dump_spec(outputStream *st) const {
3179 ProjNode::dump_spec(st);
3180 st->print("@bci %d ",_handler_bci);
3181 }
3182 #endif
3183
3184 //=============================================================================
3185 //------------------------------Identity---------------------------------------
3186 // Check for CreateEx being Identity.
3187 Node* CreateExNode::Identity(PhaseGVN* phase) {
3188 if( phase->type(in(1)) == Type::TOP ) return in(1);
3189 if( phase->type(in(0)) == Type::TOP ) return in(0);
3190 if (phase->type(in(0)->in(0)) == Type::TOP) {
3191 assert(in(0)->is_CatchProj(), "control is CatchProj");
3192 return phase->C->top(); // dead code
3193 }
3194 // We only come from CatchProj, unless the CatchProj goes away.
3195 // If the CatchProj is optimized away, then we just carry the
3196 // exception oop through.
3197 CallNode *call = in(1)->in(0)->as_Call();
3198
3199 return (in(0)->is_CatchProj() && in(0)->in(0)->is_Catch() &&
3200 in(0)->in(0)->in(1) == in(1)) ? this : call->in(TypeFunc::Parms);
3201 }
3202
3203 //=============================================================================
3204 //------------------------------Value------------------------------------------
3205 // Check for being unreachable.
3206 const Type* NeverBranchNode::Value(PhaseGVN* phase) const {
3207 if (!in(0) || in(0)->is_top()) return Type::TOP;
3208 return bottom_type();
3209 }
3210
3211 //------------------------------Ideal------------------------------------------
3212 // Check for no longer being part of a loop
3213 Node *NeverBranchNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3214 if (can_reshape && !in(0)->is_Region()) {
3215 // Dead code elimination can sometimes delete this projection so
3216 // if it's not there, there's nothing to do.
|
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "gc/shared/barrierSet.hpp"
26 #include "gc/shared/c2/barrierSetC2.hpp"
27 #include "memory/allocation.inline.hpp"
28 #include "memory/resourceArea.hpp"
29 #include "oops/objArrayKlass.hpp"
30 #include "opto/addnode.hpp"
31 #include "opto/castnode.hpp"
32 #include "opto/cfgnode.hpp"
33 #include "opto/connode.hpp"
34 #include "opto/convertnode.hpp"
35 #include "opto/inlinetypenode.hpp"
36 #include "opto/loopnode.hpp"
37 #include "opto/machnode.hpp"
38 #include "opto/movenode.hpp"
39 #include "opto/mulnode.hpp"
40 #include "opto/narrowptrnode.hpp"
41 #include "opto/phaseX.hpp"
42 #include "opto/regalloc.hpp"
43 #include "opto/regmask.hpp"
44 #include "opto/runtime.hpp"
45 #include "opto/subnode.hpp"
46 #include "opto/vectornode.hpp"
47 #include "utilities/vmError.hpp"
48
49 // Portions of code courtesy of Clifford Click
50
51 // Optimization - Graph Style
52
53 //=============================================================================
54 //------------------------------Value------------------------------------------
55 // Compute the type of the RegionNode.
504 if (left_path == nullptr || right_path == nullptr) {
505 return false;
506 }
507 Node* diamond_if = left_path->in(0);
508 if (diamond_if == nullptr || !diamond_if->is_If() || diamond_if != right_path->in(0)) {
509 // Not an IfNode merging a diamond or TOP.
510 return false;
511 }
512
513 // Check for a proper bool/cmp
514 const Node* bol = diamond_if->in(1);
515 if (!bol->is_Bool()) {
516 return false;
517 }
518 const Node* cmp = bol->in(1);
519 if (!cmp->is_Cmp()) {
520 return false;
521 }
522 return true;
523 }
524
525 //------------------------------Ideal------------------------------------------
526 // Return a node which is more "ideal" than the current node. Must preserve
527 // the CFG, but we can still strip out dead paths.
528 Node *RegionNode::Ideal(PhaseGVN *phase, bool can_reshape) {
529 if( !can_reshape && !in(0) ) return nullptr; // Already degraded to a Copy
530 assert(!in(0) || !in(0)->is_Root(), "not a specially hidden merge");
531
532 // Check for RegionNode with no Phi users and both inputs come from either
533 // arm of the same IF. If found, then the control-flow split is useless.
534 bool has_phis = false;
535 if (can_reshape) { // Need DU info to check for Phi users
536 try_clean_mem_phis(phase->is_IterGVN());
537 has_phis = (has_phi() != nullptr); // Cache result
538
539 if (!has_phis) { // No Phi users? Nothing merging?
540 for (uint i = 1; i < req()-1; i++) {
541 Node *if1 = in(i);
542 if( !if1 ) continue;
543 Node *iff = if1->in(0);
544 if( !iff || !iff->is_If() ) continue;
951 if (iff1 == iff2) {
952 igvn->add_users_to_worklist(iff1); // Make sure dead if is eliminated
953 igvn->replace_input_of(region, idx1, iff1->in(0));
954 igvn->replace_input_of(region, idx2, igvn->C->top());
955 return (region == this); // Remove useless if (both projections map to the same control/value)
956 }
957 BoolNode* bol1 = iff1->in(1)->isa_Bool();
958 BoolNode* bol2 = iff2->in(1)->isa_Bool();
959 if (bol1 == nullptr || bol2 == nullptr) {
960 return false; // No bool inputs found
961 }
962 Node* cmp1 = bol1->in(1);
963 Node* cmp2 = bol2->in(1);
964 bool commute = false;
965 if (!cmp1->is_Cmp() || !cmp2->is_Cmp()) {
966 return false; // No comparison
967 } else if (cmp1->Opcode() == Op_CmpF || cmp1->Opcode() == Op_CmpD ||
968 cmp2->Opcode() == Op_CmpF || cmp2->Opcode() == Op_CmpD ||
969 cmp1->Opcode() == Op_CmpP || cmp1->Opcode() == Op_CmpN ||
970 cmp2->Opcode() == Op_CmpP || cmp2->Opcode() == Op_CmpN ||
971 cmp1->is_SubTypeCheck() || cmp2->is_SubTypeCheck() ||
972 cmp1->is_FlatArrayCheck() || cmp2->is_FlatArrayCheck()) {
973 // Floats and pointers don't exactly obey trichotomy. To be on the safe side, don't transform their tests.
974 // SubTypeCheck is not commutative
975 return false;
976 } else if (cmp1 != cmp2) {
977 if (cmp1->in(1) == cmp2->in(2) &&
978 cmp1->in(2) == cmp2->in(1)) {
979 commute = true; // Same but swapped inputs, commute the test
980 } else {
981 return false; // Ifs are not comparing the same values
982 }
983 }
984 proj1 = proj1->other_if_proj();
985 proj2 = proj2->other_if_proj();
986 if (!((proj1->unique_ctrl_out_or_null() == iff2 &&
987 proj2->unique_ctrl_out_or_null() == this) ||
988 (proj2->unique_ctrl_out_or_null() == iff1 &&
989 proj1->unique_ctrl_out_or_null() == this))) {
990 return false; // Ifs are not connected through other projs
991 }
992 // Found 'iff -> proj -> iff -> proj -> this' shape where all other projs are merged
1031 st->print("#reducible ");
1032 break;
1033 case RegionNode::LoopStatus::NeverIrreducibleEntry:
1034 break; // nothing
1035 }
1036 }
1037 #endif
1038
1039 // Find the one non-null required input. RegionNode only
1040 Node *Node::nonnull_req() const {
1041 assert( is_Region(), "" );
1042 for( uint i = 1; i < _cnt; i++ )
1043 if( in(i) )
1044 return in(i);
1045 ShouldNotReachHere();
1046 return nullptr;
1047 }
1048
1049
1050 //=============================================================================
1051 // note that these functions assume that the _adr_type field is flat
1052 uint PhiNode::hash() const {
1053 const Type* at = _adr_type;
1054 return TypeNode::hash() + (at ? at->hash() : 0);
1055 }
1056 bool PhiNode::cmp( const Node &n ) const {
1057 return TypeNode::cmp(n) && _adr_type == ((PhiNode&)n)._adr_type;
1058 }
1059 static inline
1060 const TypePtr* flatten_phi_adr_type(const TypePtr* at) {
1061 if (at == nullptr || at == TypePtr::BOTTOM) return at;
1062 return Compile::current()->alias_type(at)->adr_type();
1063 }
1064
1065 //----------------------------make---------------------------------------------
1066 // create a new phi with edges matching r and set (initially) to x
1067 PhiNode* PhiNode::make(Node* r, Node* x, const Type *t, const TypePtr* at) {
1068 uint preds = r->req(); // Number of predecessor paths
1069 assert(t != Type::MEMORY || at == flatten_phi_adr_type(at) || (flatten_phi_adr_type(at) == TypeAryPtr::INLINES && Compile::current()->flat_accesses_share_alias()), "flatten at");
1070 PhiNode* p = new PhiNode(r, t, at);
1071 for (uint j = 1; j < preds; j++) {
1072 // Fill in all inputs, except those which the region does not yet have
1073 if (r->in(j) != nullptr)
1074 p->init_req(j, x);
1075 }
1076 return p;
1077 }
1078 PhiNode* PhiNode::make(Node* r, Node* x) {
1079 const Type* t = x->bottom_type();
1080 const TypePtr* at = nullptr;
1081 if (t == Type::MEMORY) at = flatten_phi_adr_type(x->adr_type());
1082 return make(r, x, t, at);
1083 }
1084 PhiNode* PhiNode::make_blank(Node* r, Node* x) {
1085 const Type* t = x->bottom_type();
1086 const TypePtr* at = nullptr;
1087 if (t == Type::MEMORY) at = flatten_phi_adr_type(x->adr_type());
1088 return new PhiNode(r, t, at);
1089 }
1178 np->as_Phi()->verify_adr_type(visited, at);
1179 } else if (n->bottom_type() == Type::TOP
1180 || (n->is_Mem() && n->in(MemNode::Address)->bottom_type() == Type::TOP)) {
1181 // ignore top inputs
1182 } else {
1183 const TypePtr* nat = flatten_phi_adr_type(n->adr_type());
1184 // recheck phi/non-phi consistency at leaves:
1185 assert((nat != nullptr) == (at != nullptr), "");
1186 assert(nat == at || nat == TypePtr::BOTTOM,
1187 "adr_type must be consistent at leaves of phi nest");
1188 }
1189 }
1190 }
1191
1192 // Verify a whole nest of phis rooted at this one.
1193 void PhiNode::verify_adr_type(bool recursive) const {
1194 if (VMError::is_error_reported()) return; // muzzle asserts when debugging an error
1195 if (Node::in_dump()) return; // muzzle asserts when printing
1196
1197 assert((_type == Type::MEMORY) == (_adr_type != nullptr), "adr_type for memory phis only");
1198 // Flat array element shouldn't get their own memory slice until flat_accesses_share_alias is cleared.
1199 // It could be the graph has no loads/stores and flat_accesses_share_alias is never cleared. EA could still
1200 // creates per element Phis but that wouldn't be a problem as there are no memory accesses for that array.
1201 assert(_adr_type == nullptr || _adr_type->isa_aryptr() == nullptr ||
1202 _adr_type->is_aryptr()->is_known_instance() ||
1203 !_adr_type->is_aryptr()->is_flat() ||
1204 !Compile::current()->flat_accesses_share_alias() ||
1205 _adr_type == TypeAryPtr::INLINES, "flat array element shouldn't get its own slice yet");
1206
1207 if (!VerifyAliases) return; // verify thoroughly only if requested
1208
1209 assert(_adr_type == flatten_phi_adr_type(_adr_type),
1210 "Phi::adr_type must be pre-normalized");
1211
1212 if (recursive) {
1213 VectorSet visited;
1214 verify_adr_type(visited, _adr_type);
1215 }
1216 }
1217 #endif
1218
1219
1220 //------------------------------Value------------------------------------------
1221 // Compute the type of the PhiNode
1222 const Type* PhiNode::Value(PhaseGVN* phase) const {
1223 Node *r = in(0); // RegionNode
1224 if( !r ) // Copy or dead
1225 return in(1) ? phase->type(in(1)) : Type::TOP;
1469 assert(req() == 3, "same as region");
1470 RegionNode* region = in(0)->as_Region();
1471 for (uint i = 1; i < 3; i++) {
1472 Node* phi_input = in(i);
1473 if (phi_input != nullptr && phi_input->is_MergeMem() && region->in(i)->outcnt() == 1) {
1474 // Nothing is control-dependent on path #i except the region itself.
1475 MergeMemNode* merge_mem = phi_input->as_MergeMem();
1476 uint j = 3 - i;
1477 Node* other_phi_input = in(j);
1478 if (other_phi_input != nullptr && other_phi_input == merge_mem->base_memory() && !is_data_loop(region, phi_input, igvn)) {
1479 // merge_mem is a successor memory to other_phi_input, and is not pinned inside the diamond, so push it out.
1480 // Only proceed if the transformation doesn't create a data loop
1481 // This will allow the diamond to collapse completely if there are no other phis left.
1482 igvn->replace_node(this, merge_mem);
1483 return true;
1484 }
1485 }
1486 }
1487 return false;
1488 }
1489
1490 //----------------------------check_cmove_id-----------------------------------
1491 // Check for CMove'ing a constant after comparing against the constant.
1492 // Happens all the time now, since if we compare equality vs a constant in
1493 // the parser, we "know" the variable is constant on one path and we force
1494 // it. Thus code like "if( x==0 ) {/*EMPTY*/}" ends up inserting a
1495 // conditional move: "x = (x==0)?0:x;". Yucko. This fix is slightly more
1496 // general in that we don't need constants. Since CMove's are only inserted
1497 // in very special circumstances, we do it here on generic Phi's.
1498 Node* PhiNode::is_cmove_id(PhaseTransform* phase, int true_path) {
1499 assert(true_path !=0, "only diamond shape graph expected");
1500
1501 // is_diamond_phi() has guaranteed the correctness of the nodes sequence:
1502 // phi->region->if_proj->ifnode->bool->cmp
1503 Node* region = in(0);
1504 Node* iff = region->in(1)->in(0);
1505 BoolNode* b = iff->in(1)->as_Bool();
1506 Node* cmp = b->in(1);
1507 Node* tval = in(true_path);
1508 Node* fval = in(3-true_path);
1509 Node* id = CMoveNode::is_cmove_id(phase, cmp, tval, fval, b);
1524 }
1525
1526 return id;
1527 }
1528
1529 //------------------------------Identity---------------------------------------
1530 // Check for Region being Identity.
1531 Node* PhiNode::Identity(PhaseGVN* phase) {
1532 if (must_wait_for_region_in_irreducible_loop(phase)) {
1533 return this;
1534 }
1535 // Check for no merging going on
1536 // (There used to be special-case code here when this->region->is_Loop.
1537 // It would check for a tributary phi on the backedge that the main phi
1538 // trivially, perhaps with a single cast. The unique_input method
1539 // does all this and more, by reducing such tributaries to 'this'.)
1540 Node* uin = unique_input(phase, false);
1541 if (uin != nullptr) {
1542 return uin;
1543 }
1544 uin = unique_constant_input_recursive(phase);
1545 if (uin != nullptr) {
1546 return uin;
1547 }
1548
1549 int true_path = is_diamond_phi();
1550 // Delay CMove'ing identity if Ideal has not had the chance to handle unsafe cases, yet.
1551 if (true_path != 0 && !(phase->is_IterGVN() && wait_for_region_igvn(phase))) {
1552 Node* id = is_cmove_id(phase, true_path);
1553 if (id != nullptr) {
1554 return id;
1555 }
1556 }
1557
1558 // Looking for phis with identical inputs. If we find one that has
1559 // type TypePtr::BOTTOM, replace the current phi with the bottom phi.
1560 if (phase->is_IterGVN() && type() == Type::MEMORY && adr_type() !=
1561 TypePtr::BOTTOM && !adr_type()->is_known_instance()) {
1562 uint phi_len = req();
1563 Node* phi_reg = region();
1564 for (DUIterator_Fast imax, i = phi_reg->fast_outs(imax); i < imax; i++) {
1565 Node* u = phi_reg->fast_out(i);
1566 assert(!u->is_Phi() || u->in(0) == phi_reg, "broken Phi/Region subgraph");
1567 if (u->is_Phi() && u->req() == phi_len && can_be_replaced_by(u->as_Phi())) {
1618 }
1619 // Check for a unique input (maybe uncasted)
1620 if (input == nullptr) {
1621 input = un;
1622 } else if (input != un) {
1623 input = NodeSentinel; // no unique input
1624 }
1625 }
1626 if (input == nullptr) {
1627 return phase->C->top(); // no inputs
1628 }
1629
1630 if (input != NodeSentinel) {
1631 return input; // one unique direct input
1632 }
1633
1634 // Nothing.
1635 return nullptr;
1636 }
1637
1638 // Find the unique input, try to look recursively through input Phis
1639 Node* PhiNode::unique_constant_input_recursive(PhaseGVN* phase) {
1640 if (!phase->is_IterGVN()) {
1641 return nullptr;
1642 }
1643
1644 ResourceMark rm;
1645 Node* unique = nullptr;
1646 Unique_Node_List visited;
1647 visited.push(this);
1648
1649 for (uint visited_idx = 0; visited_idx < visited.size(); visited_idx++) {
1650 Node* current = visited.at(visited_idx);
1651 for (uint i = 1; i < current->req(); i++) {
1652 Node* phi_in = current->in(i);
1653 if (phi_in == nullptr) {
1654 continue;
1655 }
1656
1657 if (phi_in->is_Phi()) {
1658 visited.push(phi_in);
1659 } else {
1660 if (unique == nullptr) {
1661 if (!phi_in->is_Con()) {
1662 return nullptr;
1663 }
1664 unique = phi_in;
1665 } else if (unique != phi_in) {
1666 return nullptr;
1667 }
1668 }
1669 }
1670 }
1671 return unique;
1672 }
1673
1674 //------------------------------is_x2logic-------------------------------------
1675 // Check for simple convert-to-boolean pattern
1676 // If:(C Bool) Region:(IfF IfT) Phi:(Region 0 1)
1677 // Convert Phi to an ConvIB.
1678 static Node *is_x2logic( PhaseGVN *phase, PhiNode *phi, int true_path ) {
1679 assert(true_path !=0, "only diamond shape graph expected");
1680
1681 // If we're late in the optimization process, we may have already expanded Conv2B nodes
1682 if (phase->C->post_loop_opts_phase() && !Matcher::match_rule_supported(Op_Conv2B)) {
1683 return nullptr;
1684 }
1685
1686 // Convert the true/false index into an expected 0/1 return.
1687 // Map 2->0 and 1->1.
1688 int flipped = 2-true_path;
1689
1690 // is_diamond_phi() has guaranteed the correctness of the nodes sequence:
1691 // phi->region->if_proj->ifnode->bool->cmp
1692 Node *region = phi->in(0);
1693 Node *iff = region->in(1)->in(0);
2121
2122 if (rc->in(0)->in(1) == nullptr || !rc->in(0)->in(1)->is_Bool()) { continue; }
2123 if (worklist.member(rc->in(0)->in(1))) {
2124 delay = true;
2125 break;
2126 }
2127
2128 if (rc->in(0)->in(1)->in(1) == nullptr || !rc->in(0)->in(1)->in(1)->is_Cmp()) { continue; }
2129 if (worklist.member(rc->in(0)->in(1)->in(1))) {
2130 delay = true;
2131 break;
2132 }
2133 }
2134
2135 if (delay) {
2136 worklist.push(this);
2137 }
2138 return delay;
2139 }
2140
2141 // Push inline type input nodes (and null) down through the phi recursively (can handle data loops).
2142 InlineTypeNode* PhiNode::push_inline_types_down(PhaseGVN* phase, bool can_reshape, ciInlineKlass* inline_klass) {
2143 assert(inline_klass != nullptr, "must be");
2144 InlineTypeNode* vt = InlineTypeNode::make_null(*phase, inline_klass, /* transform = */ false)->clone_with_phis(phase, in(0), nullptr, !_type->maybe_null(), true);
2145 if (can_reshape) {
2146 // Replace phi right away to be able to use the inline
2147 // type node when reaching the phi again through data loops.
2148 PhaseIterGVN* igvn = phase->is_IterGVN();
2149 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2150 Node* u = fast_out(i);
2151 igvn->rehash_node_delayed(u);
2152 imax -= u->replace_edge(this, vt);
2153 --i;
2154 }
2155 igvn->rehash_node_delayed(this);
2156 assert(outcnt() == 0, "should be dead now");
2157 }
2158 ResourceMark rm;
2159 Node_List casts;
2160 for (uint i = 1; i < req(); ++i) {
2161 Node* n = in(i);
2162 while (n->is_ConstraintCast()) {
2163 casts.push(n);
2164 n = n->in(1);
2165 }
2166 if (phase->type(n)->is_zero_type()) {
2167 n = InlineTypeNode::make_null(*phase, inline_klass);
2168 } else if (n->is_Phi()) {
2169 assert(can_reshape, "can only handle phis during IGVN");
2170 n = phase->transform(n->as_Phi()->push_inline_types_down(phase, can_reshape, inline_klass));
2171 }
2172 while (casts.size() != 0) {
2173 // Push the cast(s) through the InlineTypeNode
2174 // TODO 8302217 Can we avoid cloning? See InlineTypeNode::clone_if_required
2175 Node* cast = casts.pop()->clone();
2176 cast->set_req_X(1, n->as_InlineType()->get_oop(), phase);
2177 n = n->clone();
2178 n->as_InlineType()->set_oop(*phase, phase->transform(cast));
2179 n = phase->transform(n);
2180 if (n->is_top()) {
2181 break;
2182 }
2183 }
2184 bool transform = !can_reshape && (i == (req()-1)); // Transform phis on last merge
2185 assert(n->is_top() || n->is_InlineType(), "Only InlineType or top at this point.");
2186 if (n->is_InlineType()) {
2187 vt->merge_with(phase, n->as_InlineType(), i, transform);
2188 } // else nothing to do: phis above vt created by clone_with_phis are initialized to top already.
2189 }
2190 return vt;
2191 }
2192
2193 // If the Phi's Region is in an irreducible loop, and the Region
2194 // has had an input removed, but not yet transformed, it could be
2195 // that the Region (and this Phi) are not reachable from Root.
2196 // If we allow the Phi to collapse before the Region, this may lead
2197 // to dead-loop data. Wait for the Region to check for reachability,
2198 // and potentially remove the dead code.
2199 bool PhiNode::must_wait_for_region_in_irreducible_loop(PhaseGVN* phase) const {
2200 RegionNode* region = in(0)->as_Region();
2201 if (region->loop_status() == RegionNode::LoopStatus::MaybeIrreducibleEntry) {
2202 Node* top = phase->C->top();
2203 for (uint j = 1; j < req(); j++) {
2204 Node* rc = region->in(j); // for each control input
2205 if (rc == nullptr || phase->type(rc) == Type::TOP) {
2206 // Region is missing a control input
2207 Node* n = in(j);
2208 if (n != nullptr && n != top) {
2209 // Phi still has its input, so region just lost its input
2210 return true;
2211 }
2212 }
2609 // Phi (this) |
2610 // | |
2611 // +-----------+
2612 //
2613 // Generally, there are issues with non-termination with such circularity
2614 // (see comment further below). However, if there is a direct loop to self,
2615 // splitting the Phi through the MergeMem will result in the below.
2616 //
2617 // +---+
2618 // | |
2619 // v |
2620 // Phi |
2621 // |\ |
2622 // | +-+
2623 // (base_memory) v
2624 // MergeMem
2625 //
2626 // This split breaks the circularity and consequently does not lead to
2627 // non-termination.
2628 uint merge_width = 0;
2629 // TODO revisit this with JDK-8247216
2630 bool mergemem_only = true;
2631 bool split_always_terminates = false; // Is splitting guaranteed to terminate?
2632 for( uint i=1; i<req(); ++i ) {// For all paths in
2633 Node *ii = in(i);
2634 // TOP inputs should not be counted as safe inputs because if the
2635 // Phi references itself through all other inputs then splitting the
2636 // Phi through memory merges would create dead loop at later stage.
2637 if (ii == top) {
2638 return nullptr; // Delay optimization until graph is cleaned.
2639 }
2640 if (ii->is_MergeMem()) {
2641 MergeMemNode* n = ii->as_MergeMem();
2642 merge_width = MAX2(merge_width, n->req());
2643 if (n->base_memory() == this) {
2644 split_always_terminates = true;
2645 }
2646 } else {
2647 mergemem_only = false;
2648 }
2649 }
2650
2651 // There are cases with circular dependencies between bottom Phis
2652 // and MergeMems. Below is a minimal example.
2653 //
2654 // +------------+
2655 // | |
2656 // (base_memory) v |
2657 // MergeMem |
2658 // | |
2659 // v |
2660 // Phi (this) |
2661 // | |
2662 // v |
2663 // Phi |
2664 // | |
2665 // +----------+
2666 //
2667 // Here, we cannot break the circularity through a self-loop as there
2668 // are two Phis involved. Repeatedly splitting the Phis through the
2669 // MergeMem leads to non-termination. We check for non-termination below.
2670 // Only check for non-termination if necessary.
2671 if (!mergemem_only && !split_always_terminates && adr_type() == TypePtr::BOTTOM &&
2672 merge_width > Compile::AliasIdxRaw) {
2673 split_always_terminates = is_split_through_mergemem_terminating();
2674 }
2675
2676 if (merge_width > Compile::AliasIdxRaw) {
2677 // found at least one non-empty MergeMem
2678 const TypePtr* at = adr_type();
2679 if (at != TypePtr::BOTTOM) {
2680 // Patch the existing phi to select an input from the merge:
2681 // Phi:AT1(...MergeMem(m0, m1, m2)...) into
2682 // Phi:AT1(...m1...)
2683 int alias_idx = phase->C->get_alias_index(at);
2684 for (uint i=1; i<req(); ++i) {
2685 Node *ii = in(i);
2686 if (ii->is_MergeMem()) {
2687 MergeMemNode* n = ii->as_MergeMem();
2688 // compress paths and change unreachable cycles to TOP
2689 // If not, we can update the input infinitely along a MergeMem cycle
2690 // Equivalent code is in MemNode::Ideal_common
2691 Node *m = phase->transform(n);
2692 if (outcnt() == 0) { // Above transform() may kill us!
2693 return top;
2694 }
2695 // If transformed to a MergeMem, get the desired slice
2696 // Otherwise the returned node represents memory for every slice
2697 Node *new_mem = (m->is_MergeMem()) ?
2698 m->as_MergeMem()->memory_at(alias_idx) : m;
2699 // Update input if it is progress over what we have now
2700 if (new_mem != ii) {
2701 set_req_X(i, new_mem, phase->is_IterGVN());
2702 progress = this;
2703 }
2704 }
2705 }
2706 } else if (mergemem_only || split_always_terminates) {
2707 // If all inputs reference this phi (directly or through data nodes) -
2708 // it is a dead loop.
2709 bool saw_safe_input = false;
2710 for (uint j = 1; j < req(); ++j) {
2711 Node* n = in(j);
2712 if (n->is_MergeMem()) {
2713 MergeMemNode* mm = n->as_MergeMem();
2714 if (mm->base_memory() == this || mm->base_memory() == mm->empty_memory()) {
2715 // Skip this input if it references back to this phi or if the memory path is dead
2716 continue;
2717 }
2718 }
2719 if (!is_unsafe_data_reference(n)) {
2720 saw_safe_input = true; // found safe input
2721 break;
2722 }
2723 }
2724 if (!saw_safe_input) {
2725 // There is a dead loop: All inputs are either dead or reference back to this phi
2726 return top;
2727 }
2728
2729 // Phi(...MergeMem(m0, m1:AT1, m2:AT2)...) into
2730 // MergeMem(Phi(...m0...), Phi:AT1(...m1...), Phi:AT2(...m2...))
2731 PhaseIterGVN* igvn = phase->is_IterGVN();
2732 assert(igvn != nullptr, "sanity check");
2733 PhiNode* new_base = (PhiNode*) clone();
2734 // Must eagerly register phis, since they participate in loops.
2735 igvn->register_new_node_with_optimizer(new_base);
2736
2737 MergeMemNode* result = MergeMemNode::make(new_base);
2738 for (uint i = 1; i < req(); ++i) {
2739 Node *ii = in(i);
2740 if (ii->is_MergeMem()) {
2741 MergeMemNode* n = ii->as_MergeMem();
2742 if (igvn) {
2743 // TODO revisit this with JDK-8247216
2744 // Put 'n' on the worklist because it might be modified by MergeMemStream::iteration_setup
2745 igvn->_worklist.push(n);
2746 }
2747 for (MergeMemStream mms(result, n); mms.next_non_empty2(); ) {
2748 // If we have not seen this slice yet, make a phi for it.
2749 bool made_new_phi = false;
2750 if (mms.is_empty()) {
2751 Node* new_phi = new_base->slice_memory(mms.adr_type(phase->C));
2752 made_new_phi = true;
2753 igvn->register_new_node_with_optimizer(new_phi);
2754 mms.set_memory(new_phi);
2755 }
2756 Node* phi = mms.memory();
2757 assert(made_new_phi || phi->in(i) == n, "replace the i-th merge by a slice");
2758 phi->set_req_X(i, mms.memory2(), phase);
2759 }
2760 }
2761 }
2762 // Distribute all self-loops.
2763 { // (Extra braces to hide mms.)
2764 for (MergeMemStream mms(result); mms.next_non_empty(); ) {
2765 Node* phi = mms.memory();
2766 for (uint i = 1; i < req(); ++i) {
2771 }
2772 }
2773
2774 // We could immediately transform the new Phi nodes here, but that can
2775 // result in creating an excessive number of new nodes within a single
2776 // IGVN iteration. We have put the Phi nodes on the IGVN worklist, so
2777 // they are transformed later on in any case.
2778
2779 // Replace self with the result.
2780 return result;
2781 }
2782 }
2783 //
2784 // Other optimizations on the memory chain
2785 //
2786 const TypePtr* at = adr_type();
2787 for( uint i=1; i<req(); ++i ) {// For all paths in
2788 Node *ii = in(i);
2789 Node *new_in = MemNode::optimize_memory_chain(ii, at, nullptr, phase);
2790 if (ii != new_in ) {
2791 set_req_X(i, new_in, phase->is_IterGVN());
2792 progress = this;
2793 }
2794 }
2795 }
2796
2797 #ifdef _LP64
2798 // Push DecodeN/DecodeNKlass down through phi.
2799 // The rest of phi graph will transform by split EncodeP node though phis up.
2800 if ((UseCompressedOops || UseCompressedClassPointers) && can_reshape && progress == nullptr) {
2801 bool may_push = true;
2802 bool has_decodeN = false;
2803 bool is_decodeN = false;
2804 for (uint i=1; i<req(); ++i) {// For all paths in
2805 Node *ii = in(i);
2806 if (ii->is_DecodeNarrowPtr() && ii->bottom_type() == bottom_type()) {
2807 // Do optimization if a non dead path exist.
2808 if (ii->in(1)->bottom_type() != Type::TOP) {
2809 has_decodeN = true;
2810 is_decodeN = ii->is_DecodeN();
2811 }
2839 if (is_decodeN) {
2840 new_ii = new EncodePNode(ii, narrow_t);
2841 } else {
2842 new_ii = new EncodePKlassNode(ii, narrow_t);
2843 }
2844 igvn->register_new_node_with_optimizer(new_ii);
2845 }
2846 }
2847 new_phi->set_req(i, new_ii);
2848 }
2849 igvn->register_new_node_with_optimizer(new_phi, this);
2850 if (is_decodeN) {
2851 progress = new DecodeNNode(new_phi, bottom_type());
2852 } else {
2853 progress = new DecodeNKlassNode(new_phi, bottom_type());
2854 }
2855 }
2856 }
2857 #endif
2858
2859 Node* inline_type = try_push_inline_types_down(phase, can_reshape);
2860 if (inline_type != this) {
2861 return inline_type;
2862 }
2863
2864 // Try to convert a Phi with two duplicated convert nodes into a phi of the pre-conversion type and the convert node
2865 // proceeding the phi, to de-duplicate the convert node and compact the IR.
2866 if (can_reshape && progress == nullptr) {
2867 ConvertNode* convert = in(1)->isa_Convert();
2868 if (convert != nullptr) {
2869 int conv_op = convert->Opcode();
2870 bool ok = true;
2871
2872 // Check the rest of the inputs
2873 for (uint i = 2; i < req(); i++) {
2874 // Make sure that all inputs are of the same type of convert node
2875 if (in(i)->Opcode() != conv_op) {
2876 ok = false;
2877 break;
2878 }
2879 }
2880
2881 if (ok) {
2882 // Find the local bottom type to set as the type of the phi
2883 const Type* source_type = Type::get_const_basic_type(convert->in_type()->basic_type());
2906 // only after the non-bottom memory phi is processed by igvn, PhiNode::Identity doesn't run and the transformation
2907 // doesn't happen.
2908 // Look for non-bottom Phis that should be transformed and enqueue them for igvn so that PhiNode::Identity executes for
2909 // them.
2910 if (can_reshape && type() == Type::MEMORY && adr_type() == TypePtr::BOTTOM) {
2911 PhaseIterGVN* igvn = phase->is_IterGVN();
2912 uint phi_len = req();
2913 Node* phi_reg = region();
2914 for (DUIterator_Fast imax, i = phi_reg->fast_outs(imax); i < imax; i++) {
2915 Node* u = phi_reg->fast_out(i);
2916 assert(!u->is_Phi() || (u->in(0) == phi_reg && u->req() == phi_len), "broken Phi/Region subgraph");
2917 if (u->is_Phi() && u->as_Phi()->can_be_replaced_by(this)) {
2918 igvn->_worklist.push(u);
2919 }
2920 }
2921 }
2922
2923 return progress; // Return any progress
2924 }
2925
2926 // Check recursively if inputs are either an inline type, constant null
2927 // or another Phi (including self references through data loops). If so,
2928 // push the inline types down through the phis to enable folding of loads.
2929 Node* PhiNode::try_push_inline_types_down(PhaseGVN* phase, const bool can_reshape) {
2930 if (!can_be_inline_type()) {
2931 return this;
2932 }
2933
2934 ciInlineKlass* inline_klass;
2935 if (can_push_inline_types_down(phase, can_reshape, inline_klass)) {
2936 assert(inline_klass != nullptr, "must be");
2937 return push_inline_types_down(phase, can_reshape, inline_klass);
2938 }
2939 return this;
2940 }
2941
2942 bool PhiNode::can_push_inline_types_down(PhaseGVN* phase, const bool can_reshape, ciInlineKlass*& inline_klass) {
2943 if (req() <= 2) {
2944 // Dead phi.
2945 return false;
2946 }
2947 inline_klass = nullptr;
2948
2949 // TODO 8302217 We need to prevent endless pushing through
2950 bool only_phi = (outcnt() != 0);
2951 for (DUIterator_Fast imax, i = fast_outs(imax); i < imax; i++) {
2952 Node* n = fast_out(i);
2953 if (n->is_InlineType() && n->in(1) == this) {
2954 return false;
2955 }
2956 if (!n->is_Phi()) {
2957 only_phi = false;
2958 }
2959 }
2960 if (only_phi) {
2961 return false;
2962 }
2963
2964 ResourceMark rm;
2965 Unique_Node_List worklist;
2966 worklist.push(this);
2967 Node_List casts;
2968
2969 for (uint next = 0; next < worklist.size(); next++) {
2970 Node* phi = worklist.at(next);
2971 for (uint i = 1; i < phi->req(); i++) {
2972 Node* n = phi->in(i);
2973 if (n == nullptr) {
2974 return false;
2975 }
2976 while (n->is_ConstraintCast()) {
2977 if (n->in(0) != nullptr && n->in(0)->is_top()) {
2978 // Will die, don't optimize
2979 return false;
2980 }
2981 casts.push(n);
2982 n = n->in(1);
2983 }
2984 const Type* type = phase->type(n);
2985 if (n->is_InlineType() && (inline_klass == nullptr || inline_klass == type->inline_klass())) {
2986 inline_klass = type->inline_klass();
2987 } else if (n->is_Phi() && can_reshape && n->bottom_type()->isa_ptr()) {
2988 worklist.push(n);
2989 } else if (!type->is_zero_type()) {
2990 return false;
2991 }
2992 }
2993 }
2994 if (inline_klass == nullptr) {
2995 return false;
2996 }
2997
2998 // Check if cast nodes can be pushed through
2999 const Type* t = Type::get_const_type(inline_klass);
3000 while (casts.size() != 0 && t != nullptr) {
3001 Node* cast = casts.pop();
3002 if (t->filter(cast->bottom_type()) == Type::TOP) {
3003 return false;
3004 }
3005 }
3006
3007 return true;
3008 }
3009
3010 #ifdef ASSERT
3011 bool PhiNode::can_push_inline_types_down(PhaseGVN* phase) {
3012 if (!can_be_inline_type()) {
3013 return false;
3014 }
3015
3016 ciInlineKlass* inline_klass;
3017 return can_push_inline_types_down(phase, true, inline_klass);
3018 }
3019 #endif // ASSERT
3020
3021 static int compare_types(const Type* const& e1, const Type* const& e2) {
3022 return (intptr_t)e1 - (intptr_t)e2;
3023 }
3024
3025 // Collect types at casts that are going to be eliminated at that Phi and store them in a TypeTuple.
3026 // Sort the types using an arbitrary order so a list of some types always hashes to the same TypeTuple (and TypeTuple
3027 // pointer comparison is enough to tell if 2 list of types are the same or not)
3028 const TypeTuple* PhiNode::collect_types(PhaseGVN* phase) const {
3029 const Node* region = in(0);
3030 const Type* phi_type = bottom_type();
3031 ResourceMark rm;
3032 GrowableArray<const Type*> types;
3033 for (uint i = 1; i < req(); i++) {
3034 if (region->in(i) == nullptr || phase->type(region->in(i)) == Type::TOP) {
3035 continue;
3036 }
3037 Node* in = Node::in(i);
3038 const Type* t = phase->type(in);
3039 if (in == nullptr || in == this || t == Type::TOP) {
3040 continue;
3390 #ifndef PRODUCT
3391 void CatchProjNode::dump_spec(outputStream *st) const {
3392 ProjNode::dump_spec(st);
3393 st->print("@bci %d ",_handler_bci);
3394 }
3395 #endif
3396
3397 //=============================================================================
3398 //------------------------------Identity---------------------------------------
3399 // Check for CreateEx being Identity.
3400 Node* CreateExNode::Identity(PhaseGVN* phase) {
3401 if( phase->type(in(1)) == Type::TOP ) return in(1);
3402 if( phase->type(in(0)) == Type::TOP ) return in(0);
3403 if (phase->type(in(0)->in(0)) == Type::TOP) {
3404 assert(in(0)->is_CatchProj(), "control is CatchProj");
3405 return phase->C->top(); // dead code
3406 }
3407 // We only come from CatchProj, unless the CatchProj goes away.
3408 // If the CatchProj is optimized away, then we just carry the
3409 // exception oop through.
3410
3411 // CheckCastPPNode::Ideal() for inline types reuses the exception
3412 // paths of a call to perform an allocation: we can see a Phi here.
3413 if (in(1)->is_Phi()) {
3414 return this;
3415 }
3416 CallNode *call = in(1)->in(0)->as_Call();
3417
3418 return (in(0)->is_CatchProj() && in(0)->in(0)->is_Catch() &&
3419 in(0)->in(0)->in(1) == in(1)) ? this : call->in(TypeFunc::Parms);
3420 }
3421
3422 //=============================================================================
3423 //------------------------------Value------------------------------------------
3424 // Check for being unreachable.
3425 const Type* NeverBranchNode::Value(PhaseGVN* phase) const {
3426 if (!in(0) || in(0)->is_top()) return Type::TOP;
3427 return bottom_type();
3428 }
3429
3430 //------------------------------Ideal------------------------------------------
3431 // Check for no longer being part of a loop
3432 Node *NeverBranchNode::Ideal(PhaseGVN *phase, bool can_reshape) {
3433 if (can_reshape && !in(0)->is_Region()) {
3434 // Dead code elimination can sometimes delete this projection so
3435 // if it's not there, there's nothing to do.
|