src/hotspot/share/opto/lcm.cpp

*** 224,10 ***
--- 224,11 ---
      case Op_StoreC:
      case Op_StoreD:
      case Op_StoreF:
      case Op_StoreI:
      case Op_StoreL:
+     case Op_StoreLSpecial:
      case Op_StoreP:
      case Op_StoreN:
      case Op_StoreNKlass:
        was_store = true;         // Memory op is a store op
        // Stores will have their address in slot 2 (memory in slot 1).

*** 313,13 ***
          } else {
            // only regular oops are expected here
            tptr = base->bottom_type()->is_ptr();
          }
          // Give up if offset is not a compile-time constant.
!         if (offset == Type::OffsetBot || tptr->_offset == Type::OffsetBot)
            continue;
!         offset += tptr->_offset; // correct if base is offsetted
          // Give up if reference is beyond page size.
          if (MacroAssembler::needs_explicit_null_check(offset))
            continue;
          // Give up if base is a decode node and the heap base is not protected.
          if (base->is_Mach() && base->as_Mach()->ideal_Opcode() == Op_DecodeN &&
--- 314,13 ---
          } else {
            // only regular oops are expected here
            tptr = base->bottom_type()->is_ptr();
          }
          // Give up if offset is not a compile-time constant.
!         if (offset == Type::OffsetBot || tptr->offset() == Type::OffsetBot)
            continue;
!         offset += tptr->offset(); // correct if base is offsetted
          // Give up if reference is beyond page size.
          if (MacroAssembler::needs_explicit_null_check(offset))
            continue;
          // Give up if base is a decode node and the heap base is not protected.
          if (base->is_Mach() && base->as_Mach()->ideal_Opcode() == Op_DecodeN &&

*** 360,11 ***
          // a scratch register for the implementation of the node (e.g. in
          // late-expanded GC barriers).
          continue;
        }
        // Block of memory-op input
!       Block *inb = get_block_for_node(mach->in(j));
        Block *b = block;          // Start from nul check
        while( b != inb && b->_dom_depth > inb->_dom_depth )
          b = b->_idom;           // search upwards for input
        // See if input dominates null check
        if( b != inb )
--- 361,15 ---
          // a scratch register for the implementation of the node (e.g. in
          // late-expanded GC barriers).
          continue;
        }
        // Block of memory-op input
!       Block* inb = get_block_for_node(mach->in(j));
+       if (mach->in(j)->is_Con() && mach->in(j)->req() == 1 && inb == get_block_for_node(mach)) {
+         // Ignore constant loads scheduled in the same block (we can simply hoist them as well)
+         continue;
+       }
        Block *b = block;          // Start from nul check
        while( b != inb && b->_dom_depth > inb->_dom_depth )
          b = b->_idom;           // search upwards for input
        // See if input dominates null check
        if( b != inb )

*** 433,10 ***
--- 438,31 ---
        }
        move_node_and_its_projections_to_block(val, block);
      }
    }
  
+   // Hoist constant load inputs as well.
+   for (uint i = 1; i < best->req(); ++i) {
+     Node* n = best->in(i);
+     if (n->is_Con() && get_block_for_node(n) == get_block_for_node(best)) {
+       get_block_for_node(n)->find_remove(n);
+       block->add_inst(n);
+       map_node_to_block(n, block);
+       // Constant loads may kill flags (for example, when XORing a register).
+       // Check for flag-killing projections that also need to be hoisted.
+       for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++) {
+         Node* proj = n->fast_out(j);
+         if (proj->is_MachProj()) {
+           get_block_for_node(proj)->find_remove(proj);
+           block->add_inst(proj);
+           map_node_to_block(proj, block);
+         }
+       }
+     }
+   }
+ 
+ 
    // Move any MachTemp inputs to the end of the test block.
    for (uint i = 0; i < best->req(); i++) {
      Node* n = best->in(i);
      if (n == nullptr || !n->is_MachTemp()) {
        continue;

*** 741,10 ***
--- 767,11 ---
          case Op_StoreC:
          case Op_StoreD:
          case Op_StoreF:
          case Op_StoreI:
          case Op_StoreL:
+         case Op_StoreLSpecial:
          case Op_StoreP:
          case Op_StoreN:
          case Op_StoreVector:
          case Op_StoreVectorMasked:
          case Op_StoreVectorScatter:

*** 907,11 ***
    // Act as if the call defines the Frame Pointer.
    // Certainly the FP is alive and well after the call.
    regs.Insert(_matcher.c_frame_pointer());
  
    // Set all registers killed and not already defined by the call.
!   uint r_cnt = mcall->tf()->range()->cnt();
    int op = mcall->ideal_Opcode();
    MachProjNode *proj = new MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
    map_node_to_block(proj, block);
    block->insert_node(proj, node_cnt++);
  
--- 934,11 ---
    // Act as if the call defines the Frame Pointer.
    // Certainly the FP is alive and well after the call.
    regs.Insert(_matcher.c_frame_pointer());
  
    // Set all registers killed and not already defined by the call.
!   uint r_cnt = mcall->tf()->range_cc()->cnt();
    int op = mcall->ideal_Opcode();
    MachProjNode *proj = new MachProjNode( mcall, r_cnt+1, RegMask::Empty, MachProjNode::fat_proj );
    map_node_to_block(proj, block);
    block->insert_node(proj, node_cnt++);