Optimize VM

[bit-hack]

The follow opcode frequencies were gathers from the 1k fuzz tests.
The VM should be optimized to prefer the more frequently called instructions.

Also we should look at common instruction pairs, and optimize for them.

  20 584890 INS_CONST
  21 240505 INS_LOCALS
  16 239147 INS_CALL
  26 153893 INS_GETG
  8  37696  INS_NOT
  15 30480  INS_CJMP
  22 24626  INS_GETV
  23 24321  INS_SETV
  14 5629   INS_JMP
  28 3962   INS_GETGI
  29 3524   INS_SETGI
  27 2882   INS_SETG
  0  2877   INS_ADD
  12 2581   INS_GEQ
  11 2418   INS_LEQ
  7  2265   INS_OR
  3  2179   INS_MUL
  9  2121   INS_LT
  2  1764   INS_SUB
  10 1734   INS_GT
  13 1581   INS_EQ
  17 1546   INS_RET
  4  1380   INS_DIV
  24 1283   INS_GETVI
  6  939    INS_AND
  25 826    INS_SETVI
  19 214    INS_POP
  1  0      INS_INC
  5  0      INS_MOD
  18 0      INS_SCALL

[bit-hack]

I added the += operator so that loop carried dependencies such as i = i + 1 are more compact. I still need to add this to the fuzz tester. But the current ordering is now:

     INS_CONST 140302
      INS_CJMP 65145
       INS_JMP 60748
      INS_GETG 41666
    INS_LOCALS 23992
      INS_CALL 22631
      INS_SETV 9727
       INS_NOT 7318
      INS_GETV 5596
     INS_GETGI 4525
        INS_LT 4455
        INS_OR 3831
     INS_SETGI 3168
       INS_DIV 2756
       INS_LEQ 2281
       INS_ADD 2240
       INS_AND 1829
       INS_RET 1724
        INS_EQ 1136
     INS_GETVI 897
       INS_SUB 527
      INS_SETG 521
        INS_GT 460
       INS_GEQ 344
       INS_MUL 275
       INS_POP 202
     INS_SETVI 72
       INS_MOD 0
       INS_NEG 0
     INS_SCALL 0
      INS_ACCV 0

[bit-hack]

Recheck these numbers, but a little bit of work has gone into improving the speed of the VM. Currently the bottle neck is the collection phase of the garbage collector. Marking nodes could be made more efficient using marking bitmaps.