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: [tree-ssa] Loop header copying on trees


Hello,

>  >>  >> + ftree-ch
>  >>  >> + Common
>  >>  >> + Enable loop header copying on trees
>  >>  >> + 
>  >>  >>
>  >>  >Include a small blurb describing why header copying is beneficial. 
>  >>  >Also, do we really want to control it with a flag?  Some of the -f flags
>  >>  >we have in tree-ssa should probably disappear.
>  >> And in fact, I don't think we really want to have a knob to turn off
>  >> header copying.
>  >
>  >why not? At the very least, we probably do not want it enabled at -Os.
> Funny -- we experimented with that a while back and we actually got
> more compact code copying loop headers!

I suspect this is due to the copied headers being often eliminated.  The
jump threading on tree-ssa does this kind of loop header copying
already, so I assume it might no longer be true (although I haven't
tested it).

>  >Also I do not see why having a precise control over what the compiler
>  >does should be wrong.
> The fundamental problem is you have to know far too much about the
> compiler to even begin to estimate how the gazillion knobs might
> affect the output.  I've believed for a very long time that GCC provides
> far too many knobs and switches and all it does is lead to mass
> confusion.

yes, this might be a problem. On the other hand, there are a lot of
situations where you would like to have as precise control as possible:

1) For writing testcases to testsuite, it would be very nice to be be
   able to disable all (nonesential) components except for the one you
   intend to test.
2) When debugging a misscompilation, it would often be useful to control
   the optimization options at least on per-function level (and
   preferably finer).

Perhaps we can have both, now that we have the generic way of
pass management -- we could have easy to use flags only for controlling
important features (for users), and a generic way providing the means to control
precisely what is done (for developers).

Zdenek


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