This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [patch] Store motion rewrite



On Monday, February 24, 2003, at 06:13 PM, Zdenek Dvorak wrote:


Hello,

Zdenek> Perhaps it would be better to just fix Daniel's patch and base further
Zdenek> work on it?


Part of the problem is that Daniel's patch is out of date. It
will take time just to mrege Daniel's patch into the current sources, let
alone fix whatever bug is causing the bootstrap problem on some
architectures.


I am eager and willing to help with store motion in the rtlopt
branch. My personal suggestion is that we stay with your new
implementation, whose internals you understand. We debug whatever Athlon
bootstrap failures and forward-port Daniel's speed improvements to your
infrastructure. You now know of Daniel's previous effort and can use it
as a base for ideas. If we use Daniel's code base, you would need to
learn and understand it.

[sorry if you get this twice; I have forgotten to post it to the list last
time :-( ]


here is the patch updated for mainline. It enables load and store
motion of non-symbol only mems. I haven't integrated Daniel's patch into
it, as I haven't observed any measurable slowdown (but perhaps I was
just using bad testcases).
Just in case someone wants to know about this part of the message, as we discussed later in the thread, Zdenek's way of tracking where the reg was set effectively incorporates what i was doing, he just did it in a slightly different place (I did it by memoizing the results of checking a given reg at the end, he spread it out over the functions.).

His way might be slightly slower when you have a billion stores, regs, and bb's.

--Dan


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]