[lno] we now intechange loops on a well known benchmark...
Zdenek Dvorak
rakdver@atrey.karlin.mff.cuni.cz
Thu Sep 2 14:49:00 GMT 2004
Hello,
> This patch should do the work for interchanging loops on swim, and
> probably other non optimal code. I have tested that the transform
> matrix is built and validated on a reduced testcase
> (aka. ssa-chrec-62.c)
>
> Some low level comments: the dependence on this testcase is tricky
> {{0, +, 1}_1, +, 1335}_2 vs. {0, +, 1336}_1
>
> and for getting it right one has to bound the dependence domain. This
> is done by estimating the number of iterations in loops 1 and 2 by
> looking at the size of the data and at the access functions. I will
> be grateful if somebody could comment on the way this information is
> extracted in estimate_niter_from_size_of_data. This is probably not
> the best solution, but it works on that testcase...
> + * cfgloop.h (struct loop): New fields parallel_p, and
> + estimated_nb_iterations.
we already have field for this -- loop->bounds. It would be also nice
to put the related analysis to
tree-ssa-loop-niter.c:estimate_numbers_of_iterations_loop.
Zdenek
More information about the Gcc-patches
mailing list