By just adding the next predicate to avoid generation of immediate 16-bits stores
I got +5% jbb2005 on woodcrest:

  // Store Short/Char Immediate
  instruct storeImmI16(memory mem, immI16 src) %{
+   predicate(UseImm16ForStore);
    match(Set mem (StoreC mem src));

    ins_cost(150);
    format %{ "MOV16  $mem,$src" %}
    opcode(0xC7);     /* C7 /0 Same as 32 store immediate with prefix */
    ins_encode( SizePrefix, OpcP, RMopc_Mem(0x00,mem),  Con16( src ));
    ins_pipe( ialu_mem_imm );
  %}


hsdev-3% sh compare ref.jbb.n2 ref.jbb.n2_n16s
============================================================================
ref.jbb.n2
  Benchmark         Samples        Mean     Stdev             Geomean Weight
  specjbb2000            20   120380.43    591.43
  specjbb2005            20    51259.08    741.56
============================================================================
ref.jbb.n2_n16s
  Benchmark         Samples        Mean     Stdev   %Diff    P   Significant
  specjbb2000            20   122106.15    856.48    1.43 0.000          Yes
  specjbb2005            20    54137.34    989.22    5.62 0.000          Yes
============================================================================
Posted Date : 2007-03-06 22:42:15.0

N/A

don't use immediate 16-bits value store to memory on Intel cpus
Posted Date : 2007-03-07 00:48:22.0