< prev index next >

src/hotspot/share/opto/loopopts.cpp

Print this page

  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "gc/shared/barrierSet.hpp"
  26 #include "gc/shared/c2/barrierSetC2.hpp"
  27 #include "memory/allocation.inline.hpp"
  28 #include "memory/resourceArea.hpp"
  29 #include "opto/addnode.hpp"
  30 #include "opto/callnode.hpp"
  31 #include "opto/castnode.hpp"
  32 #include "opto/connode.hpp"
  33 #include "opto/divnode.hpp"

  34 #include "opto/loopnode.hpp"
  35 #include "opto/matcher.hpp"
  36 #include "opto/movenode.hpp"
  37 #include "opto/mulnode.hpp"
  38 #include "opto/opaquenode.hpp"
  39 #include "opto/rootnode.hpp"
  40 #include "opto/subnode.hpp"
  41 #include "opto/subtypenode.hpp"
  42 #include "opto/superword.hpp"
  43 #include "opto/vectornode.hpp"
  44 #include "utilities/checkedCast.hpp"
  45 #include "utilities/macros.hpp"
  46 
  47 //=============================================================================
  48 //------------------------------split_thru_phi---------------------------------
  49 // Split Node 'n' through merge point if there is enough win.
  50 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
  51   if ((n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) ||
  52       (n->Opcode() == Op_ConvL2I && n->bottom_type() != TypeInt::INT)) {
  53     // ConvI2L/ConvL2I may have type information on it which is unsafe to push up
  54     // so disable this for now
  55     return nullptr;
  56   }
  57 
  58   // Splitting range check CastIIs through a loop induction Phi can
  59   // cause new Phis to be created that are left unrelated to the loop
  60   // induction Phi and prevent optimizations (vectorization)
  61   if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
  62       n->in(1) == region->as_CountedLoop()->phi()) {
  63     return nullptr;
  64   }
  65 






  66   if (cannot_split_division(n, region)) {
  67     return nullptr;
  68   }
  69 
  70   SplitThruPhiWins wins(region);
  71   assert(!n->is_CFG(), "");
  72   assert(region->is_Region(), "");
  73 
  74   const Type* type = n->bottom_type();
  75   const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
  76   Node* phi;
  77   if (t_oop != nullptr && t_oop->is_known_instance_field()) {
  78     int iid    = t_oop->instance_id();
  79     int index  = C->get_alias_index(t_oop);
  80     int offset = t_oop->offset();
  81     phi = new PhiNode(region, type, nullptr, iid, index, offset);
  82   } else {
  83     phi = PhiNode::make_blank(region, n);
  84   }
  85   uint old_unique = C->unique();

 774       // CMOVE'd derived pointer?  It's a CMOVE'd derived base.  Thus
 775       // CMOVE'ing a derived pointer requires we also CMOVE the base.  If we
 776       // have a Phi for the base here that we convert to a CMOVE all is well
 777       // and good.  But if the base is dead, we'll not make a CMOVE.  Later
 778       // the allocator will have to produce a base by creating a CMOVE of the
 779       // relevant bases.  This puts the allocator in the business of
 780       // manufacturing expensive instructions, generally a bad plan.
 781       // Just Say No to Conditionally-Moved Derived Pointers.
 782       if (tp && tp->offset() != 0)
 783         return nullptr;
 784       cost++;
 785       break;
 786     }
 787     default:
 788       return nullptr;              // In particular, can't do memory or I/O
 789     }
 790     // Add in cost any speculative ops
 791     for (uint j = 1; j < region->req(); j++) {
 792       Node *proj = region->in(j);
 793       Node *inp = phi->in(j);




 794       if (get_ctrl(inp) == proj) { // Found local op
 795         cost++;
 796         // Check for a chain of dependent ops; these will all become
 797         // speculative in a CMOV.
 798         for (uint k = 1; k < inp->req(); k++)
 799           if (get_ctrl(inp->in(k)) == proj)
 800             cost += ConditionalMoveLimit; // Too much speculative goo
 801       }
 802     }
 803     // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
 804     // This will likely Split-If, a higher-payoff operation.
 805     for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
 806       Node* use = phi->fast_out(k);
 807       if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
 808         cost += ConditionalMoveLimit;
 809       // Is there a use inside the loop?
 810       // Note: check only basic types since CMoveP is pinned.
 811       if (!used_inside_loop && is_java_primitive(bt)) {
 812         IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
 813         if (r_loop == u_loop || r_loop->is_member(u_loop)) {

1099             assert(get_loop(lca)->_nest < n_loop->_nest || get_loop(lca)->_head->as_Loop()->is_in_infinite_subgraph(), "must not be moved into inner loop");
1100 
1101             // Move store out of the loop
1102             _igvn.replace_node(hook, n->in(MemNode::Memory));
1103             _igvn.replace_input_of(n, 0, lca);
1104             set_ctrl_and_loop(n, lca);
1105 
1106             // Disconnect the phi now. An empty phi can confuse other
1107             // optimizations in this pass of loop opts..
1108             if (phi->in(LoopNode::LoopBackControl) == phi) {
1109               _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
1110               n_loop->_body.yank(phi);
1111             }
1112           }
1113         }
1114       }
1115     }
1116   }
1117 }
1118 
















































1119 //------------------------------split_if_with_blocks_pre-----------------------
1120 // Do the real work in a non-recursive function.  Data nodes want to be
1121 // cloned in the pre-order so they can feed each other nicely.
1122 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1123   // Cloning these guys is unlikely to win
1124   int n_op = n->Opcode();
1125   if (n_op == Op_MergeMem) {
1126     return n;
1127   }
1128   if (n->is_Proj()) {
1129     return n;
1130   }






1131   // Do not clone-up CmpFXXX variations, as these are always
1132   // followed by a CmpI
1133   if (n->is_Cmp()) {
1134     return n;
1135   }
1136   // Attempt to use a conditional move instead of a phi/branch
1137   if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1138     Node *cmov = conditional_move( n );
1139     if (cmov) {
1140       return cmov;
1141     }
1142   }
1143   if (n->is_CFG() || n->is_LoadStore()) {
1144     return n;
1145   }
1146   if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd
1147     if (!C->major_progress()) {   // If chance of no more loop opts...
1148       _igvn._worklist.push(n);  // maybe we'll remove them
1149     }
1150     return n;

1383 
1384   return true;
1385 }
1386 
1387 // Detect if the node is the inner strip-mined loop
1388 // Return: null if it's not the case, or the exit of outer strip-mined loop
1389 static Node* is_inner_of_stripmined_loop(const Node* out) {
1390   Node* out_le = nullptr;
1391 
1392   if (out->is_CountedLoopEnd()) {
1393       const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1394 
1395       if (loop != nullptr && loop->is_strip_mined()) {
1396         out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1397       }
1398   }
1399 
1400   return out_le;
1401 }
1402 








































































































1403 //------------------------------split_if_with_blocks_post----------------------
1404 // Do the real work in a non-recursive function.  CFG hackery wants to be
1405 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1406 // info.
1407 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1408 




1409   // Cloning Cmp through Phi's involves the split-if transform.
1410   // FastLock is not used by an If
1411   if (n->is_Cmp() && !n->is_FastLock()) {
1412     Node *n_ctrl = get_ctrl(n);
1413     // Determine if the Node has inputs from some local Phi.
1414     // Returns the block to clone thru.
1415     Node *n_blk = has_local_phi_input(n);
1416     if (n_blk != n_ctrl) {
1417       return;
1418     }
1419 
1420     if (!can_split_if(n_ctrl)) {
1421       return;
1422     }
1423 
1424     if (n->outcnt() != 1) {
1425       return; // Multiple bool's from 1 compare?
1426     }
1427     Node *bol = n->unique_out();
1428     assert(bol->is_Bool(), "expect a bool here");

1532           //    accesses would start to float, since we don't pin at that point.
1533           // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned.
1534           bool pin_array_access_nodes =  n->Opcode() == Op_RangeCheck &&
1535                                          prevdom->in(0)->Opcode() != Op_RangeCheck;
1536           dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes);
1537           DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
1538           return;
1539         }
1540         prevdom = dom;
1541         dom = idom(prevdom);
1542       }
1543     }
1544   }
1545 
1546   try_sink_out_of_loop(n);
1547   if (C->failing()) {
1548     return;
1549   }
1550 
1551   try_move_store_after_loop(n);





1552 }
1553 
1554 // Transform:
1555 //
1556 // if (some_condition) {
1557 //   // body 1
1558 // } else {
1559 //   // body 2
1560 // }
1561 // if (some_condition) {
1562 //   // body 3
1563 // } else {
1564 //   // body 4
1565 // }
1566 //
1567 // into:
1568 //
1569 //
1570 // if (some_condition) {
1571 //   // body 1

2041   uint i;
2042   for (i = 1; i < phi->req(); i++) {
2043     Node* b = phi->in(i);
2044     if (b->is_Phi()) {
2045       _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
2046     } else {
2047       assert(b->is_Bool() || b->is_OpaqueConstantBool() || b->is_OpaqueInitializedAssertionPredicate(),
2048              "bool, non-null check with OpaqueConstantBool or Initialized Assertion Predicate with its Opaque node");
2049     }
2050   }
2051   Node* n = phi->in(1);
2052   Node* sample_opaque = nullptr;
2053   Node *sample_bool = nullptr;
2054   if (n->is_OpaqueConstantBool() || n->is_OpaqueInitializedAssertionPredicate()) {
2055     sample_opaque = n;
2056     sample_bool = n->in(1);
2057     assert(sample_bool->is_Bool(), "wrong type");
2058   } else {
2059     sample_bool = n;
2060   }
2061   Node *sample_cmp = sample_bool->in(1);








2062 
2063   // Make Phis to merge the Cmp's inputs.
2064   PhiNode *phi1 = new PhiNode(phi->in(0), Type::TOP);
2065   PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
2066   for (i = 1; i < phi->req(); i++) {
2067     Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
2068     Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
2069     phi1->set_req(i, n1);
2070     phi2->set_req(i, n2);
2071     phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2072     phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2073   }
2074   // See if these Phis have been made before.
2075   // Register with optimizer
2076   Node *hit1 = _igvn.hash_find_insert(phi1);
2077   if (hit1) {                   // Hit, toss just made Phi
2078     _igvn.remove_dead_node(phi1); // Remove new phi
2079     assert(hit1->is_Phi(), "" );
2080     phi1 = (PhiNode*)hit1;      // Use existing phi
2081   } else {                      // Miss
2082     _igvn.register_new_node_with_optimizer(phi1);
2083   }
2084   Node *hit2 = _igvn.hash_find_insert(phi2);

  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "gc/shared/barrierSet.hpp"
  26 #include "gc/shared/c2/barrierSetC2.hpp"
  27 #include "memory/allocation.inline.hpp"
  28 #include "memory/resourceArea.hpp"
  29 #include "opto/addnode.hpp"
  30 #include "opto/callnode.hpp"
  31 #include "opto/castnode.hpp"
  32 #include "opto/connode.hpp"
  33 #include "opto/divnode.hpp"
  34 #include "opto/inlinetypenode.hpp"
  35 #include "opto/loopnode.hpp"
  36 #include "opto/matcher.hpp"
  37 #include "opto/movenode.hpp"
  38 #include "opto/mulnode.hpp"
  39 #include "opto/opaquenode.hpp"
  40 #include "opto/rootnode.hpp"
  41 #include "opto/subnode.hpp"
  42 #include "opto/subtypenode.hpp"
  43 #include "opto/superword.hpp"
  44 #include "opto/vectornode.hpp"
  45 #include "utilities/checkedCast.hpp"
  46 #include "utilities/macros.hpp"
  47 
  48 //=============================================================================
  49 //------------------------------split_thru_phi---------------------------------
  50 // Split Node 'n' through merge point if there is enough win.
  51 Node* PhaseIdealLoop::split_thru_phi(Node* n, Node* region, int policy) {
  52   if ((n->Opcode() == Op_ConvI2L && n->bottom_type() != TypeLong::LONG) ||
  53       (n->Opcode() == Op_ConvL2I && n->bottom_type() != TypeInt::INT)) {
  54     // ConvI2L/ConvL2I may have type information on it which is unsafe to push up
  55     // so disable this for now
  56     return nullptr;
  57   }
  58 
  59   // Splitting range check CastIIs through a loop induction Phi can
  60   // cause new Phis to be created that are left unrelated to the loop
  61   // induction Phi and prevent optimizations (vectorization)
  62   if (n->Opcode() == Op_CastII && region->is_CountedLoop() &&
  63       n->in(1) == region->as_CountedLoop()->phi()) {
  64     return nullptr;
  65   }
  66 
  67   // Inline types should not be split through Phis because they cannot be merged
  68   // through Phi nodes but each value input needs to be merged individually.
  69   if (n->is_InlineType()) {
  70     return nullptr;
  71   }
  72 
  73   if (cannot_split_division(n, region)) {
  74     return nullptr;
  75   }
  76 
  77   SplitThruPhiWins wins(region);
  78   assert(!n->is_CFG(), "");
  79   assert(region->is_Region(), "");
  80 
  81   const Type* type = n->bottom_type();
  82   const TypeOopPtr* t_oop = _igvn.type(n)->isa_oopptr();
  83   Node* phi;
  84   if (t_oop != nullptr && t_oop->is_known_instance_field()) {
  85     int iid    = t_oop->instance_id();
  86     int index  = C->get_alias_index(t_oop);
  87     int offset = t_oop->offset();
  88     phi = new PhiNode(region, type, nullptr, iid, index, offset);
  89   } else {
  90     phi = PhiNode::make_blank(region, n);
  91   }
  92   uint old_unique = C->unique();

 781       // CMOVE'd derived pointer?  It's a CMOVE'd derived base.  Thus
 782       // CMOVE'ing a derived pointer requires we also CMOVE the base.  If we
 783       // have a Phi for the base here that we convert to a CMOVE all is well
 784       // and good.  But if the base is dead, we'll not make a CMOVE.  Later
 785       // the allocator will have to produce a base by creating a CMOVE of the
 786       // relevant bases.  This puts the allocator in the business of
 787       // manufacturing expensive instructions, generally a bad plan.
 788       // Just Say No to Conditionally-Moved Derived Pointers.
 789       if (tp && tp->offset() != 0)
 790         return nullptr;
 791       cost++;
 792       break;
 793     }
 794     default:
 795       return nullptr;              // In particular, can't do memory or I/O
 796     }
 797     // Add in cost any speculative ops
 798     for (uint j = 1; j < region->req(); j++) {
 799       Node *proj = region->in(j);
 800       Node *inp = phi->in(j);
 801       if (inp->isa_InlineType()) {
 802         // TODO 8302217 This prevents PhiNode::push_inline_types_through
 803         return nullptr;
 804       }
 805       if (get_ctrl(inp) == proj) { // Found local op
 806         cost++;
 807         // Check for a chain of dependent ops; these will all become
 808         // speculative in a CMOV.
 809         for (uint k = 1; k < inp->req(); k++)
 810           if (get_ctrl(inp->in(k)) == proj)
 811             cost += ConditionalMoveLimit; // Too much speculative goo
 812       }
 813     }
 814     // See if the Phi is used by a Cmp or Narrow oop Decode/Encode.
 815     // This will likely Split-If, a higher-payoff operation.
 816     for (DUIterator_Fast kmax, k = phi->fast_outs(kmax); k < kmax; k++) {
 817       Node* use = phi->fast_out(k);
 818       if (use->is_Cmp() || use->is_DecodeNarrowPtr() || use->is_EncodeNarrowPtr())
 819         cost += ConditionalMoveLimit;
 820       // Is there a use inside the loop?
 821       // Note: check only basic types since CMoveP is pinned.
 822       if (!used_inside_loop && is_java_primitive(bt)) {
 823         IdealLoopTree* u_loop = get_loop(has_ctrl(use) ? get_ctrl(use) : use);
 824         if (r_loop == u_loop || r_loop->is_member(u_loop)) {

1110             assert(get_loop(lca)->_nest < n_loop->_nest || get_loop(lca)->_head->as_Loop()->is_in_infinite_subgraph(), "must not be moved into inner loop");
1111 
1112             // Move store out of the loop
1113             _igvn.replace_node(hook, n->in(MemNode::Memory));
1114             _igvn.replace_input_of(n, 0, lca);
1115             set_ctrl_and_loop(n, lca);
1116 
1117             // Disconnect the phi now. An empty phi can confuse other
1118             // optimizations in this pass of loop opts..
1119             if (phi->in(LoopNode::LoopBackControl) == phi) {
1120               _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
1121               n_loop->_body.yank(phi);
1122             }
1123           }
1124         }
1125       }
1126     }
1127   }
1128 }
1129 
1130 // We can't use immutable memory for the flat array check because we are loading the mark word which is
1131 // mutable. Although the bits we are interested in are immutable (we check for markWord::unlocked_value),
1132 // we need to use raw memory to not break anti dependency analysis. Below code will attempt to still move
1133 // flat array checks out of loops, mainly to enable loop unswitching.
1134 void PhaseIdealLoop::move_flat_array_check_out_of_loop(Node* n) {
1135   // Skip checks for more than one array
1136   if (n->req() > 3) {
1137     return;
1138   }
1139   Node* mem = n->in(FlatArrayCheckNode::Memory);
1140   Node* array = n->in(FlatArrayCheckNode::ArrayOrKlass)->uncast();
1141   IdealLoopTree* check_loop = get_loop(get_ctrl(n));
1142   IdealLoopTree* ary_loop = get_loop(get_ctrl(array));
1143 
1144   // Check if array is loop invariant
1145   if (!check_loop->is_member(ary_loop)) {
1146     // Walk up memory graph from the check until we leave the loop
1147     VectorSet wq;
1148     wq.set(mem->_idx);
1149     while (check_loop->is_member(get_loop(ctrl_or_self(mem)))) {
1150       if (mem->is_Phi()) {
1151         mem = mem->in(1);
1152       } else if (mem->is_MergeMem()) {
1153         mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1154       } else if (mem->is_Proj()) {
1155         mem = mem->in(0);
1156       } else if (mem->is_MemBar() || mem->is_SafePoint()) {
1157         mem = mem->in(TypeFunc::Memory);
1158       } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
1159         mem = mem->in(MemNode::Memory);
1160       } else {
1161 #ifdef ASSERT
1162         mem->dump();
1163 #endif
1164         ShouldNotReachHere();
1165       }
1166       if (wq.test_set(mem->_idx)) {
1167         return;
1168       }
1169     }
1170     // Replace memory input and re-compute ctrl to move the check out of the loop
1171     _igvn.replace_input_of(n, 1, mem);
1172     set_ctrl_and_loop(n, get_early_ctrl(n));
1173     Node* bol = n->unique_out();
1174     set_ctrl_and_loop(bol, get_early_ctrl(bol));
1175   }
1176 }
1177 
1178 //------------------------------split_if_with_blocks_pre-----------------------
1179 // Do the real work in a non-recursive function.  Data nodes want to be
1180 // cloned in the pre-order so they can feed each other nicely.
1181 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1182   // Cloning these guys is unlikely to win
1183   int n_op = n->Opcode();
1184   if (n_op == Op_MergeMem) {
1185     return n;
1186   }
1187   if (n->is_Proj()) {
1188     return n;
1189   }
1190 
1191   if (n->isa_FlatArrayCheck()) {
1192     move_flat_array_check_out_of_loop(n);
1193     return n;
1194   }
1195 
1196   // Do not clone-up CmpFXXX variations, as these are always
1197   // followed by a CmpI
1198   if (n->is_Cmp()) {
1199     return n;
1200   }
1201   // Attempt to use a conditional move instead of a phi/branch
1202   if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1203     Node *cmov = conditional_move( n );
1204     if (cmov) {
1205       return cmov;
1206     }
1207   }
1208   if (n->is_CFG() || n->is_LoadStore()) {
1209     return n;
1210   }
1211   if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd
1212     if (!C->major_progress()) {   // If chance of no more loop opts...
1213       _igvn._worklist.push(n);  // maybe we'll remove them
1214     }
1215     return n;

1448 
1449   return true;
1450 }
1451 
1452 // Detect if the node is the inner strip-mined loop
1453 // Return: null if it's not the case, or the exit of outer strip-mined loop
1454 static Node* is_inner_of_stripmined_loop(const Node* out) {
1455   Node* out_le = nullptr;
1456 
1457   if (out->is_CountedLoopEnd()) {
1458       const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1459 
1460       if (loop != nullptr && loop->is_strip_mined()) {
1461         out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1462       }
1463   }
1464 
1465   return out_le;
1466 }
1467 
1468 bool PhaseIdealLoop::flat_array_element_type_check(Node *n) {
1469   // If the CmpP is a subtype check for a value that has just been
1470   // loaded from an array, the subtype check guarantees the value
1471   // can't be stored in a flat array and the load of the value
1472   // happens with a flat array check then: push the type check
1473   // through the phi of the flat array check. This needs special
1474   // logic because the subtype check's input is not a phi but a
1475   // LoadKlass that must first be cloned through the phi.
1476   if (n->Opcode() != Op_CmpP) {
1477     return false;
1478   }
1479 
1480   Node* klassptr = n->in(1);
1481   Node* klasscon = n->in(2);
1482 
1483   if (klassptr->is_DecodeNarrowPtr()) {
1484     klassptr = klassptr->in(1);
1485   }
1486 
1487   if (klassptr->Opcode() != Op_LoadKlass && klassptr->Opcode() != Op_LoadNKlass) {
1488     return false;
1489   }
1490 
1491   if (!klasscon->is_Con()) {
1492     return false;
1493   }
1494 
1495   Node* addr = klassptr->in(MemNode::Address);
1496 
1497   if (!addr->is_AddP()) {
1498     return false;
1499   }
1500 
1501   intptr_t offset;
1502   Node* obj = AddPNode::Ideal_base_and_offset(addr, &_igvn, offset);
1503 
1504   if (obj == nullptr) {
1505     return false;
1506   }
1507 
1508   // TODO 8378077: The code below does not work anymore with off-heap accesses which set their bases to top with
1509   // JDK-8373343. Also: flat_array_element_type_check() was introduced with JDK-8228622 for a specific check to enable
1510   // split-if but JDK-8245729 changed how that check looks like. Is it still relevant? This should be revisited.
1511   if (addr->in(AddPNode::Base)->is_top()) {
1512     return false;
1513   }
1514 
1515   if (obj->Opcode() == Op_CastPP) {
1516     obj = obj->in(1);
1517   }
1518 
1519   if (!obj->is_Phi()) {
1520     return false;
1521   }
1522 
1523   Node* region = obj->in(0);
1524 
1525   Node* phi = PhiNode::make_blank(region, n->in(1));
1526   for (uint i = 1; i < region->req(); i++) {
1527     Node* in = obj->in(i);
1528     Node* ctrl = region->in(i);
1529     if (addr->in(AddPNode::Base) != obj) {
1530       Node* cast = addr->in(AddPNode::Base);
1531       assert(cast->Opcode() == Op_CastPP && cast->in(0) != nullptr, "inconsistent subgraph");
1532       Node* cast_clone = cast->clone();
1533       cast_clone->set_req(0, ctrl);
1534       cast_clone->set_req(1, in);
1535       register_new_node(cast_clone, ctrl);
1536       const Type* tcast = cast_clone->Value(&_igvn);
1537       _igvn.set_type(cast_clone, tcast);
1538       cast_clone->as_Type()->set_type(tcast);
1539       in = cast_clone;
1540     }
1541     Node* addr_clone = addr->clone();
1542     addr_clone->set_req(AddPNode::Base, in);
1543     addr_clone->set_req(AddPNode::Address, in);
1544     register_new_node(addr_clone, ctrl);
1545     _igvn.set_type(addr_clone, addr_clone->Value(&_igvn));
1546     Node* klassptr_clone = klassptr->clone();
1547     klassptr_clone->set_req(2, addr_clone);
1548     register_new_node(klassptr_clone, ctrl);
1549     _igvn.set_type(klassptr_clone, klassptr_clone->Value(&_igvn));
1550     if (klassptr != n->in(1)) {
1551       Node* decode = n->in(1);
1552       assert(decode->is_DecodeNarrowPtr(), "inconsistent subgraph");
1553       Node* decode_clone = decode->clone();
1554       decode_clone->set_req(1, klassptr_clone);
1555       register_new_node(decode_clone, ctrl);
1556       _igvn.set_type(decode_clone, decode_clone->Value(&_igvn));
1557       klassptr_clone = decode_clone;
1558     }
1559     phi->set_req(i, klassptr_clone);
1560   }
1561   register_new_node(phi, region);
1562   Node* orig = n->in(1);
1563   _igvn.replace_input_of(n, 1, phi);
1564   split_if_with_blocks_post(n);
1565   if (n->outcnt() != 0) {
1566     _igvn.replace_input_of(n, 1, orig);
1567     _igvn.remove_dead_node(phi);
1568   }
1569   return true;
1570 }
1571 
1572 //------------------------------split_if_with_blocks_post----------------------
1573 // Do the real work in a non-recursive function.  CFG hackery wants to be
1574 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1575 // info.
1576 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1577 
1578   if (flat_array_element_type_check(n)) {
1579     return;
1580   }
1581 
1582   // Cloning Cmp through Phi's involves the split-if transform.
1583   // FastLock is not used by an If
1584   if (n->is_Cmp() && !n->is_FastLock()) {
1585     Node *n_ctrl = get_ctrl(n);
1586     // Determine if the Node has inputs from some local Phi.
1587     // Returns the block to clone thru.
1588     Node *n_blk = has_local_phi_input(n);
1589     if (n_blk != n_ctrl) {
1590       return;
1591     }
1592 
1593     if (!can_split_if(n_ctrl)) {
1594       return;
1595     }
1596 
1597     if (n->outcnt() != 1) {
1598       return; // Multiple bool's from 1 compare?
1599     }
1600     Node *bol = n->unique_out();
1601     assert(bol->is_Bool(), "expect a bool here");

1705           //    accesses would start to float, since we don't pin at that point.
1706           // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned.
1707           bool pin_array_access_nodes =  n->Opcode() == Op_RangeCheck &&
1708                                          prevdom->in(0)->Opcode() != Op_RangeCheck;
1709           dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes);
1710           DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
1711           return;
1712         }
1713         prevdom = dom;
1714         dom = idom(prevdom);
1715       }
1716     }
1717   }
1718 
1719   try_sink_out_of_loop(n);
1720   if (C->failing()) {
1721     return;
1722   }
1723 
1724   try_move_store_after_loop(n);
1725 
1726   // Remove multiple allocations of the same inline type
1727   if (n->is_InlineType()) {
1728     n->as_InlineType()->remove_redundant_allocations(this);
1729   }
1730 }
1731 
1732 // Transform:
1733 //
1734 // if (some_condition) {
1735 //   // body 1
1736 // } else {
1737 //   // body 2
1738 // }
1739 // if (some_condition) {
1740 //   // body 3
1741 // } else {
1742 //   // body 4
1743 // }
1744 //
1745 // into:
1746 //
1747 //
1748 // if (some_condition) {
1749 //   // body 1

2219   uint i;
2220   for (i = 1; i < phi->req(); i++) {
2221     Node* b = phi->in(i);
2222     if (b->is_Phi()) {
2223       _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
2224     } else {
2225       assert(b->is_Bool() || b->is_OpaqueConstantBool() || b->is_OpaqueInitializedAssertionPredicate(),
2226              "bool, non-null check with OpaqueConstantBool or Initialized Assertion Predicate with its Opaque node");
2227     }
2228   }
2229   Node* n = phi->in(1);
2230   Node* sample_opaque = nullptr;
2231   Node *sample_bool = nullptr;
2232   if (n->is_OpaqueConstantBool() || n->is_OpaqueInitializedAssertionPredicate()) {
2233     sample_opaque = n;
2234     sample_bool = n->in(1);
2235     assert(sample_bool->is_Bool(), "wrong type");
2236   } else {
2237     sample_bool = n;
2238   }
2239   Node* sample_cmp = sample_bool->in(1);
2240   const Type* t = Type::TOP;
2241   const TypePtr* at = nullptr;
2242   if (sample_cmp->is_FlatArrayCheck()) {
2243     // Left input of a FlatArrayCheckNode is memory, set the (adr) type of the phi accordingly
2244     assert(sample_cmp->in(1)->bottom_type() == Type::MEMORY, "unexpected input type");
2245     t = Type::MEMORY;
2246     at = TypeRawPtr::BOTTOM;
2247   }
2248 
2249   // Make Phis to merge the Cmp's inputs.
2250   PhiNode *phi1 = new PhiNode(phi->in(0), t, at);
2251   PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
2252   for (i = 1; i < phi->req(); i++) {
2253     Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
2254     Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
2255     phi1->set_req(i, n1);
2256     phi2->set_req(i, n2);
2257     phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2258     phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2259   }
2260   // See if these Phis have been made before.
2261   // Register with optimizer
2262   Node *hit1 = _igvn.hash_find_insert(phi1);
2263   if (hit1) {                   // Hit, toss just made Phi
2264     _igvn.remove_dead_node(phi1); // Remove new phi
2265     assert(hit1->is_Phi(), "" );
2266     phi1 = (PhiNode*)hit1;      // Use existing phi
2267   } else {                      // Miss
2268     _igvn.register_new_node_with_optimizer(phi1);
2269   }
2270   Node *hit2 = _igvn.hash_find_insert(phi2);
< prev index next >