Dan Sugalski <[EMAIL PROTECTED]> wrote:
> Or 3) Toss the prederef stuff entirely.
And here is, why I want to keep the CGP core:
sub_i_i_i
0x81bbef0 <cgp_core+33488>: mov 0x4(%esi),%ecx
0x81bbef3 <cgp_core+33491>: mov 0x8(%esi),%edx
0x81bbef6 <cgp_core+33494>: mov 0xc(%esi),%eax
0x81bbef9 <cgp_core+33497>: add $0x10,%esi
0x81bbefc <cgp_core+33500>: mov (%eax,%edi,1),%eax
0x81bbeff <cgp_core+33503>: mov (%edx,%edi,1),%edx
0x81bbf02 <cgp_core+33506>: sub %eax,%edx
0x81bbf04 <cgp_core+33508>: mov %edx,(%ecx,%edi,1)
0x81bbf07 <cgp_core+33511>: jmp *(%esi)
if_i_ic
0x81b4152 <cgp_core+1330>: mov 0x4(%esi),%eax
0x81b4155 <cgp_core+1333>: cmpl $0x0,(%eax,%edi,1)
0x81b4159 <cgp_core+1337>: je 0x81b4167 <cgp_core+1351>
0x81b415b <cgp_core+1339>: mov 0x8(%esi),%eax
0x81b415e <cgp_core+1342>: mov (%eax),%eax
0x81b4160 <cgp_core+1344>: shl $0x2,%eax
0x81b4163 <cgp_core+1347>: add %eax,%esi
0x81b4165 <cgp_core+1349>: jmp *(%esi)
0x81b4167 <cgp_core+1351>: add $0xc,%esi
0x81b416a <cgp_core+1354>: jmp *(%esi)
%esi ... cur_opcode
%edi ... register frame pointer
A register access is 2 CPU instructions only:
mov 8(%esi), %edx # cur_opcode[2], i.e. offset of REG_INT(x)
mov (%edx, %edi, 1), %edx # get *(base + offset)
That's all.
$ ./parrot -C mops.pasm
Iterations: 100000000
Estimated ops: 200000000
Elapsed time: 2.156002
M op/s: 92.764291
That's an Athlon 800 - 8.5 CPU instructions per Parrot instruction.
leo
