< 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();

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




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

1100             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");
1101 
1102             // Move store out of the loop
1103             _igvn.replace_node(hook, n->in(MemNode::Memory));
1104             _igvn.replace_input_of(n, 0, lca);
1105             set_ctrl_and_loop(n, lca);
1106 
1107             // Disconnect the phi now. An empty phi can confuse other
1108             // optimizations in this pass of loop opts..
1109             if (phi->in(LoopNode::LoopBackControl) == phi) {
1110               _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
1111               n_loop->_body.yank(phi);
1112             }
1113           }
1114         }
1115       }
1116     }
1117   }
1118 }
1119 
















































1120 //------------------------------split_if_with_blocks_pre-----------------------
1121 // Do the real work in a non-recursive function.  Data nodes want to be
1122 // cloned in the pre-order so they can feed each other nicely.
1123 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1124   // Cloning these guys is unlikely to win
1125   int n_op = n->Opcode();
1126   if (n_op == Op_MergeMem) {
1127     return n;
1128   }
1129   if (n->is_Proj()) {
1130     return n;
1131   }






1132   // Do not clone-up CmpFXXX variations, as these are always
1133   // followed by a CmpI
1134   if (n->is_Cmp()) {
1135     return n;
1136   }
1137   // Attempt to use a conditional move instead of a phi/branch
1138   if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1139     Node *cmov = conditional_move( n );
1140     if (cmov) {
1141       return cmov;
1142     }
1143   }
1144   if (n->is_CFG() || n->is_LoadStore()) {
1145     return n;
1146   }
1147   if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd
1148     if (!C->major_progress()) {   // If chance of no more loop opts...
1149       _igvn._worklist.push(n);  // maybe we'll remove them
1150     }
1151     return n;

1384 
1385   return true;
1386 }
1387 
1388 // Detect if the node is the inner strip-mined loop
1389 // Return: null if it's not the case, or the exit of outer strip-mined loop
1390 static Node* is_inner_of_stripmined_loop(const Node* out) {
1391   Node* out_le = nullptr;
1392 
1393   if (out->is_CountedLoopEnd()) {
1394       const CountedLoopNode* loop = out->as_CountedLoopEnd()->loopnode();
1395 
1396       if (loop != nullptr && loop->is_strip_mined()) {
1397         out_le = loop->in(LoopNode::EntryControl)->as_OuterStripMinedLoop()->outer_loop_exit();
1398       }
1399   }
1400 
1401   return out_le;
1402 }
1403 


































































































1404 //------------------------------split_if_with_blocks_post----------------------
1405 // Do the real work in a non-recursive function.  CFG hackery wants to be
1406 // in the post-order, so it can dirty the I-DOM info and not use the dirtied
1407 // info.
1408 void PhaseIdealLoop::split_if_with_blocks_post(Node *n) {
1409 




1410   // Cloning Cmp through Phi's involves the split-if transform.
1411   // FastLock is not used by an If
1412   if (n->is_Cmp() && !n->is_FastLock()) {
1413     Node *n_ctrl = get_ctrl(n);
1414     // Determine if the Node has inputs from some local Phi.
1415     // Returns the block to clone thru.
1416     Node *n_blk = has_local_phi_input(n);
1417     if (n_blk != n_ctrl) {
1418       return;
1419     }
1420 
1421     if (!can_split_if(n_ctrl)) {
1422       return;
1423     }
1424 
1425     if (n->outcnt() != 1) {
1426       return; // Multiple bool's from 1 compare?
1427     }
1428     Node *bol = n->unique_out();
1429     assert(bol->is_Bool(), "expect a bool here");

1533           //    accesses would start to float, since we don't pin at that point.
1534           // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned.
1535           bool pin_array_access_nodes =  n->Opcode() == Op_RangeCheck &&
1536                                          prevdom->in(0)->Opcode() != Op_RangeCheck;
1537           dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes);
1538           DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
1539           return;
1540         }
1541         prevdom = dom;
1542         dom = idom(prevdom);
1543       }
1544     }
1545   }
1546 
1547   try_sink_out_of_loop(n);
1548   if (C->failing()) {
1549     return;
1550   }
1551 
1552   try_move_store_after_loop(n);





1553 }
1554 
1555 // Transform:
1556 //
1557 // if (some_condition) {
1558 //   // body 1
1559 // } else {
1560 //   // body 2
1561 // }
1562 // if (some_condition) {
1563 //   // body 3
1564 // } else {
1565 //   // body 4
1566 // }
1567 //
1568 // into:
1569 //
1570 //
1571 // if (some_condition) {
1572 //   // 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_OpaqueNotNull() || b->is_OpaqueInitializedAssertionPredicate(),
2048              "bool, non-null check with OpaqueNotNull 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_OpaqueNotNull() || 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();

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

1111             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");
1112 
1113             // Move store out of the loop
1114             _igvn.replace_node(hook, n->in(MemNode::Memory));
1115             _igvn.replace_input_of(n, 0, lca);
1116             set_ctrl_and_loop(n, lca);
1117 
1118             // Disconnect the phi now. An empty phi can confuse other
1119             // optimizations in this pass of loop opts..
1120             if (phi->in(LoopNode::LoopBackControl) == phi) {
1121               _igvn.replace_node(phi, phi->in(LoopNode::EntryControl));
1122               n_loop->_body.yank(phi);
1123             }
1124           }
1125         }
1126       }
1127     }
1128   }
1129 }
1130 
1131 // We can't use immutable memory for the flat array check because we are loading the mark word which is
1132 // mutable. Although the bits we are interested in are immutable (we check for markWord::unlocked_value),
1133 // we need to use raw memory to not break anti dependency analysis. Below code will attempt to still move
1134 // flat array checks out of loops, mainly to enable loop unswitching.
1135 void PhaseIdealLoop::move_flat_array_check_out_of_loop(Node* n) {
1136   // Skip checks for more than one array
1137   if (n->req() > 3) {
1138     return;
1139   }
1140   Node* mem = n->in(FlatArrayCheckNode::Memory);
1141   Node* array = n->in(FlatArrayCheckNode::ArrayOrKlass)->uncast();
1142   IdealLoopTree* check_loop = get_loop(get_ctrl(n));
1143   IdealLoopTree* ary_loop = get_loop(get_ctrl(array));
1144 
1145   // Check if array is loop invariant
1146   if (!check_loop->is_member(ary_loop)) {
1147     // Walk up memory graph from the check until we leave the loop
1148     VectorSet wq;
1149     wq.set(mem->_idx);
1150     while (check_loop->is_member(get_loop(ctrl_or_self(mem)))) {
1151       if (mem->is_Phi()) {
1152         mem = mem->in(1);
1153       } else if (mem->is_MergeMem()) {
1154         mem = mem->as_MergeMem()->memory_at(Compile::AliasIdxRaw);
1155       } else if (mem->is_Proj()) {
1156         mem = mem->in(0);
1157       } else if (mem->is_MemBar() || mem->is_SafePoint()) {
1158         mem = mem->in(TypeFunc::Memory);
1159       } else if (mem->is_Store() || mem->is_LoadStore() || mem->is_ClearArray()) {
1160         mem = mem->in(MemNode::Memory);
1161       } else {
1162 #ifdef ASSERT
1163         mem->dump();
1164 #endif
1165         ShouldNotReachHere();
1166       }
1167       if (wq.test_set(mem->_idx)) {
1168         return;
1169       }
1170     }
1171     // Replace memory input and re-compute ctrl to move the check out of the loop
1172     _igvn.replace_input_of(n, 1, mem);
1173     set_ctrl_and_loop(n, get_early_ctrl(n));
1174     Node* bol = n->unique_out();
1175     set_ctrl_and_loop(bol, get_early_ctrl(bol));
1176   }
1177 }
1178 
1179 //------------------------------split_if_with_blocks_pre-----------------------
1180 // Do the real work in a non-recursive function.  Data nodes want to be
1181 // cloned in the pre-order so they can feed each other nicely.
1182 Node *PhaseIdealLoop::split_if_with_blocks_pre( Node *n ) {
1183   // Cloning these guys is unlikely to win
1184   int n_op = n->Opcode();
1185   if (n_op == Op_MergeMem) {
1186     return n;
1187   }
1188   if (n->is_Proj()) {
1189     return n;
1190   }
1191 
1192   if (n->isa_FlatArrayCheck()) {
1193     move_flat_array_check_out_of_loop(n);
1194     return n;
1195   }
1196 
1197   // Do not clone-up CmpFXXX variations, as these are always
1198   // followed by a CmpI
1199   if (n->is_Cmp()) {
1200     return n;
1201   }
1202   // Attempt to use a conditional move instead of a phi/branch
1203   if (ConditionalMoveLimit > 0 && n_op == Op_Region) {
1204     Node *cmov = conditional_move( n );
1205     if (cmov) {
1206       return cmov;
1207     }
1208   }
1209   if (n->is_CFG() || n->is_LoadStore()) {
1210     return n;
1211   }
1212   if (n->is_Opaque1()) { // Opaque nodes cannot be mod'd
1213     if (!C->major_progress()) {   // If chance of no more loop opts...
1214       _igvn._worklist.push(n);  // maybe we'll remove them
1215     }
1216     return n;

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

1700           //    accesses would start to float, since we don't pin at that point.
1701           // 3. If we move from regular if: don't pin. All array accesses are already assumed to be pinned.
1702           bool pin_array_access_nodes =  n->Opcode() == Op_RangeCheck &&
1703                                          prevdom->in(0)->Opcode() != Op_RangeCheck;
1704           dominated_by(prevdom->as_IfProj(), n->as_If(), false, pin_array_access_nodes);
1705           DEBUG_ONLY( if (VerifyLoopOptimizations) { verify(); } );
1706           return;
1707         }
1708         prevdom = dom;
1709         dom = idom(prevdom);
1710       }
1711     }
1712   }
1713 
1714   try_sink_out_of_loop(n);
1715   if (C->failing()) {
1716     return;
1717   }
1718 
1719   try_move_store_after_loop(n);
1720 
1721   // Remove multiple allocations of the same inline type
1722   if (n->is_InlineType()) {
1723     n->as_InlineType()->remove_redundant_allocations(this);
1724   }
1725 }
1726 
1727 // Transform:
1728 //
1729 // if (some_condition) {
1730 //   // body 1
1731 // } else {
1732 //   // body 2
1733 // }
1734 // if (some_condition) {
1735 //   // body 3
1736 // } else {
1737 //   // body 4
1738 // }
1739 //
1740 // into:
1741 //
1742 //
1743 // if (some_condition) {
1744 //   // body 1

2213   uint i;
2214   for (i = 1; i < phi->req(); i++) {
2215     Node* b = phi->in(i);
2216     if (b->is_Phi()) {
2217       _igvn.replace_input_of(phi, i, clone_iff(b->as_Phi()));
2218     } else {
2219       assert(b->is_Bool() || b->is_OpaqueNotNull() || b->is_OpaqueInitializedAssertionPredicate(),
2220              "bool, non-null check with OpaqueNotNull or Initialized Assertion Predicate with its Opaque node");
2221     }
2222   }
2223   Node* n = phi->in(1);
2224   Node* sample_opaque = nullptr;
2225   Node *sample_bool = nullptr;
2226   if (n->is_OpaqueNotNull() || n->is_OpaqueInitializedAssertionPredicate()) {
2227     sample_opaque = n;
2228     sample_bool = n->in(1);
2229     assert(sample_bool->is_Bool(), "wrong type");
2230   } else {
2231     sample_bool = n;
2232   }
2233   Node* sample_cmp = sample_bool->in(1);
2234   const Type* t = Type::TOP;
2235   const TypePtr* at = nullptr;
2236   if (sample_cmp->is_FlatArrayCheck()) {
2237     // Left input of a FlatArrayCheckNode is memory, set the (adr) type of the phi accordingly
2238     assert(sample_cmp->in(1)->bottom_type() == Type::MEMORY, "unexpected input type");
2239     t = Type::MEMORY;
2240     at = TypeRawPtr::BOTTOM;
2241   }
2242 
2243   // Make Phis to merge the Cmp's inputs.
2244   PhiNode *phi1 = new PhiNode(phi->in(0), t, at);
2245   PhiNode *phi2 = new PhiNode(phi->in(0), Type::TOP);
2246   for (i = 1; i < phi->req(); i++) {
2247     Node *n1 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(1) : phi->in(i)->in(1)->in(1)->in(1);
2248     Node *n2 = sample_opaque == nullptr ? phi->in(i)->in(1)->in(2) : phi->in(i)->in(1)->in(1)->in(2);
2249     phi1->set_req(i, n1);
2250     phi2->set_req(i, n2);
2251     phi1->set_type(phi1->type()->meet_speculative(n1->bottom_type()));
2252     phi2->set_type(phi2->type()->meet_speculative(n2->bottom_type()));
2253   }
2254   // See if these Phis have been made before.
2255   // Register with optimizer
2256   Node *hit1 = _igvn.hash_find_insert(phi1);
2257   if (hit1) {                   // Hit, toss just made Phi
2258     _igvn.remove_dead_node(phi1); // Remove new phi
2259     assert(hit1->is_Phi(), "" );
2260     phi1 = (PhiNode*)hit1;      // Use existing phi
2261   } else {                      // Miss
2262     _igvn.register_new_node_with_optimizer(phi1);
2263   }
2264   Node *hit2 = _igvn.hash_find_insert(phi2);
< prev index next >