This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Why does the athlon_fp DFA have so many states?
- From: Jan Hubicka <hubicka at ucw dot cz>
- To: "Naveen Sharma, Noida" <naveens at noida dot hcltech dot com>
- Cc: Ranjit Mathew <rmathew at hotmail dot com>, gcc at gcc dot gnu dot org
- Date: Fri, 7 Nov 2003 12:24:41 +0100
- Subject: Re: Why does the athlon_fp DFA have so many states?
- References: <1B3885BC15C7024C845AAC78314766C59EC064@EXCH-01>
> > While compiling mainline/3.3 on i686-pc-linux-gnu
> > I have always noted that the DFA for athlon_fp has the
> > maximum number of asterisks ('*').
>
> I have sometimes noticed system going out of virtual
> memory while processing athlon DFA description.
> Complex pipelines lead to a large number of DFA states.
> Athlon's pipeline seems complex at first glance. (though
> I am not very familiar with athlon internals.)
> It might help, if it is possible to do some more automaton
> factoring for athlon. But thats just a guess.
There is no way to factor the automaton more with current design.
The Athlon FP pipes has both high pararelism and occasionally high
latencies resulting in huge state space. The current description is
very inexact to fit in memory.
I was thinking about two sollutions to this. Either we can decompose
the automatons serially or use different representation (in the Athlon
space an bitmask of states with 3 alternatives per insn would suffice
resulting in tiny table).
Honza
>
> Best Regards,
> Naveen Sharma.