17 Jan
2023
17 Jan
'23
12:12 p.m.
Peter Corlett wrote:
On Tue, Jan 17, 2023 at 10:16:18AM +0000, Peter
Coghlan via cctalk wrote:
[...]
In the context of the original question posed, I think it would be interesting
to know of if todays technology artifical intelligence can perform these sort
of tasks by "learning how" or "figuring out how" ie without being
programmed
specifically to carry out these tasks. Even better, could it deal with more
complex difficulties such as the address space already being full and the
translated code being larger than the original code?
How about translating code from Z80 which has
several registers to 6502
with rather fewer? That would seem to need some more intelligent thinking
on how to simulate the unavailable registers without causing additional
difficulties.
It is often said that the 6502 has 256 registers, i.e. zero page.
So e.g. LD (HL), A could be mechanically transformed into the sequence LDX
#0, LDA (h, X), and STA a, with h and a being zero page locations. On the
65C02 the first two operations can be replaced with a simple LDA (h),
although it may still be useful to index via X to simulate EXX without
performing an expensive copy.
As it stands, that replaces a one byte instruction with five byte sequence
which is obviously not great, but a relatively simple peephole optimiser can
eliminate many of the redundant loads and stores so it wouldn't be quite so
bad. After all, one important source of stores is the flags register, which
I ignored in the code fragment. A _good_ optimiser can do a lot of clever
analysis and transformation, and would probably be needed to handle all of
the edge cases well, but would be too large and CPU-intensive to run on a
Z80 or 6502 system.
It'd be easier to bodge a Z80 into a 6502 machine than try and translate the
code. That's what often happened back in the day, after all.
If the artificial intelligence machine came up with the answer:
"It'd be too much trouble to translate the code, stick in a Z80 instead"
would that be a useful advance on having a human being come up with the same
conclusion? Would it be able to justify it's conclusion to a beancounter?
Regards,
Peter Coghlan.