On Thu, Feb 06, 2003 at 01:37:42PM +0100, Leopold Toetsch wrote:
> This is one thing I allways wanted to try ;-)
>
> fast_core MOps: 11
> Prederef: 17.5
> CGoto MOps: 19.4
> CGP MOps: 27.5
> CGP -O3 MOps: 65 !!!1
>
> This runloop combines the faster dispatch of opcodes via computed goto
> and the clever register addressing due to predereferencing registers and
> constants.
This generates some pretty good code (for instance, for sub_i_i_i) :
(*(INTVAL *)cur_opcode[1]) =
(*(INTVAL *)cur_opcode[2]) - (*(INTVAL *)cur_opcode[3]);
goto *ops_addr[*(cur_opcode += 4)];
mov 0x4(%esi),%ecx
mov 0x8(%esi),%edx
mov 0xc(%esi),%eax
mov (%eax),%eax
mov (%edx),%edx
sub %eax,%edx
mov %edx,(%ecx)
add $0x10,%esi
mov (%esi),%eax
shl $0x2,%eax
jmp *0x80f75a0(%eax)
Why not also predereference the operation address? This would save a
memory read. The goto would become:
goto **(cur_opcode += 4);
It seems hard to improve upon this. One possibility is to have some
new instructions that uses an accumulator:
accu = accu - (*(INTVAL *)cur_opcode[1]);
goto **(cur_opcode += 2);
mov 0x4(%esi),%eax
mov (%eax),%eax
sub %ebx,%eax
add $0x8,%esi
jmp *(%esi)
-- Jérôme
