Mercurial > jdk8-mips64-public > hotspot / changeset

     1.1 --- a/src/cpu/mips/vm/mips_64.ad	Sun Jun 12 09:32:41 2016 +0800
     1.2 +++ b/src/cpu/mips/vm/mips_64.ad	Tue Jun 14 20:04:50 2016 +0800
     1.3 @@ -3055,6 +3055,7 @@
     1.4        __ ld(tmpReg, Address(tmpReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
     1.5        __ sltiu(AT, tmpReg, 1);  /* Jin: AT = !tmpReg; */
     1.6        __ bne(tmpReg, R0, DONE_LABEL);
     1.7 +      __ delayed()->nop();
     1.8  
     1.9        __ cmpxchg(TREG, Address(boxReg, ObjectMonitor::owner_offset_in_bytes()-2), tmpReg) ;
    1.10          // Intentional fall-through into DONE_LABEL ...
    1.11 @@ -3155,11 +3156,12 @@
    1.12  
    1.13        __ ld(tmpReg, Address(objReg, 0)) ;       // Examine the object's markword
    1.14        __ ld(AT, Address(boxReg, 0)) ;            // Examine the displaced header
    1.15 -      //masm.jccb  (Assembler::zero, DONE_LABEL) ;      // 0 indicates recursive stack-lock
    1.16        __ beq(AT, R0, DONE_LABEL) ;      // 0 indicates recursive stack-lock
    1.17 -      __ move(AT, 0x1); /* delay slot */
    1.18 -
    1.19 -      __ andi(AT, tmpReg, 0x02) ;                     // Inflated?
    1.20 +      //__ move(AT, 0x1);
    1.21 +      //__ delayed()->nop();
    1.22 +      __ delayed()->daddiu(AT, R0, 0x1);
    1.23 +
    1.24 +      __ andi(AT, tmpReg, markOopDesc::monitor_value) ;                     // Inflated?
    1.25        __ beq(AT, R0, Stacked) ;                     // Inflated?
    1.26        __ delayed()->nop();
    1.27  
    1.28 @@ -3187,13 +3189,14 @@
    1.29        // IA32's memory-model is SPO, so STs are ordered with respect to
    1.30        // each other and there's no need for an explicit barrier (fence).
    1.31        // See also http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
    1.32 -
    1.33  #ifdef OPT_THREAD
    1.34        __ move(boxReg, TREG);
    1.35  #else
    1.36        __ get_thread (boxReg) ;
    1.37  #endif
    1.38  
    1.39 +#ifndef _LP64
    1.40 +
    1.41        // Note that we could employ various encoding schemes to reduce
    1.42        // the number of loads below (currently 4) to just 2 or 3.
    1.43        // Refer to the comments in synchronizer.cpp.
    1.44 @@ -3340,6 +3343,82 @@
    1.45        if ((EmitSync & 65536) == 0) {
    1.46           __ bind (CheckSucc) ;
    1.47        }
    1.48 +#else // _LP64
    1.49 +      // It's inflated
    1.50 +      __ ld(AT, Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
    1.51 +      __ xorr(boxReg, boxReg, AT);
    1.52 +
    1.53 +      __ ld(AT, Address (tmpReg, ObjectMonitor::recursions_offset_in_bytes()-2)) ;
    1.54 +      __ orr(boxReg, boxReg, AT);
    1.55 +
    1.56 +      __ move(AT, R0);
    1.57 +      __ bne(boxReg, R0, DONE_LABEL);
    1.58 +      __ delayed()->nop();
    1.59 +
    1.60 +      __ ld(boxReg, Address (tmpReg, ObjectMonitor::EntryList_offset_in_bytes()-2)) ;
    1.61 +      __ ld(AT, Address (tmpReg, ObjectMonitor::cxq_offset_in_bytes()-2)) ;
    1.62 +      __ orr(boxReg, boxReg, AT);
    1.63 +
    1.64 +      __ move(AT, R0);
    1.65 +      __ bne(boxReg, R0, CheckSucc);
    1.66 +      __ delayed()->nop();
    1.67 +
    1.68 +      __ sd(R0, Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
    1.69 +      __ move(AT, 0x1);
    1.70 +      __ b(DONE_LABEL);
    1.71 +      __ delayed()->nop();
    1.72 +
    1.73 +
    1.74 +    if ((EmitSync & 65536) == 0) {
    1.75 +      Label LSuccess, LGoSlowPath ;
    1.76 +      __ bind (CheckSucc);
    1.77 +      __ ld(AT, Address (tmpReg, ObjectMonitor::succ_offset_in_bytes()-2)) ;
    1.78 +      __ beq(AT, R0, LGoSlowPath);
    1.79 +      __ delayed()->nop();
    1.80 +
    1.81 +      // I'd much rather use lock:andl m->_owner, 0 as it's faster than the
    1.82 +      // the explicit ST;MEMBAR combination, but masm doesn't currently support
    1.83 +      // "ANDQ M,IMM".  Don't use MFENCE here.  lock:add to TOS, xchg, etc
    1.84 +      // are all faster when the write buffer is populated.
    1.85 +      __ sd(R0, Address (tmpReg, ObjectMonitor::owner_offset_in_bytes()-2)) ;
    1.86 +      if (os::is_MP()) {
    1.87 +         // lock (); 
    1.88 +         //addl (Address(rsp, 0), 0); //?
    1.89 +      }
    1.90 +      __ ld(AT, Address (tmpReg, ObjectMonitor::succ_offset_in_bytes()-2)) ;
    1.91 +      __ bne(AT, R0, LSuccess);
    1.92 +      __ delayed()->nop();
    1.93 +
    1.94 +      __ move(boxReg, R0) ;                  // box is really EAX
    1.95 +      //if (os::is_MP()) { lock(); }
    1.96 +      __ cmpxchg(SP, Address(tmpReg, ObjectMonitor::owner_offset_in_bytes()-2), boxReg);
    1.97 +      __ beq(AT, R0, LSuccess);
    1.98 +      __ delayed()->nop();
    1.99 +      // Intentional fall-through into slow-path
   1.100 +
   1.101 +      __ bind  (LGoSlowPath);
   1.102 +      __ ori(boxReg, boxReg, 1) ;                      // set ICC.ZF=0 to indicate failure
   1.103 +      __ move(AT, R0);
   1.104 +      __ b(DONE_LABEL) ;
   1.105 +      __ delayed()->nop();
   1.106 +
   1.107 +
   1.108 +      __ bind  (LSuccess);
   1.109 +      __ move(boxReg, R0) ;                 // set ICC.ZF=1 to indicate success
   1.110 +      __ move(AT, 0x1) ;
   1.111 +      __ b(DONE_LABEL) ;
   1.112 +      __ delayed()->nop();
   1.113 +    }
   1.114 +
   1.115 +    __ bind  (Stacked);
   1.116 +    __ ld(tmpReg, Address(boxReg, 0)) ;
   1.117 +    //if (os::is_MP()) { lock(); }
   1.118 +    __ cmpxchg(tmpReg, Address(objReg, 0), boxReg); // Uses EAX which is box
   1.119 +
   1.120 +    if (EmitSync & 65536) {
   1.121 +       __ bind (CheckSucc);
   1.122 +    }
   1.123 +#endif
   1.124  
   1.125        __ bind(DONE_LABEL);
   1.126  
   1.127 @@ -3357,7 +3436,7 @@
   1.128  
   1.129      __ addu(flags, R0, R0);    
   1.130      __ beq(AT, R0, L);
   1.131 -    __ nop();
   1.132 +    __ delayed()->nop();
   1.133      __ move(flags, 0xFFFFFFFF);
   1.134      __ bind(L);
   1.135    %}