[Bug tree-optimization/19038] [4.0 Regression] out-of ssa causing loops to have more than one BB

["dberlin at gcc dot gnu dot org" <gcc-bugzilla at gcc dot gnu dot org>]

------- Additional Comments From dberlin at gcc dot gnu dot org  2004-12-26 01:09 -------
Just to followup on irc discussions.
Being able to coalesce this may require moving to a different out-of-ssa
algorithm that is better at coalescing.
All of the below take different approaches to the coalescing problem, however,
they may provide some insight as to whether/how these two variables can be
coalesced.

The briggs one (which is what morgan is based on, which is what the comments
claim is used) is known not to coalesce a significant number of variables.

A comparison of the two major algorithms for out-of-ssa (briggs and sreedhar's)
can be found at
http://www.is.titech.ac.jp/~sassa/papers-written/sassa-kohama-ito-ipsi04.pdf

Sreedhar's removes roughly 80% more copies than brigg's algorithm.
Note that the authors propose their own algorithm, even though it does
significantly worse than sreedhars, so i don't consider it a serious contender here.

Sreedhar's can be found at http://tinyurl.com/5axtl
i have a copy of the paper if you have no access.

In addition, yet another different approach to out-of-ssa coalescing was
presented at cgo 2004.
http://www.cgo.org/cgo2004/papers/21_14_rastello_f_revised.pdf

I mention it because it supposedly can coalesce even more than sreedhar's
algorithm is able to.

I point these out only as references to hopefully be able coalesce these variables.
I don't believe totally rewriting out-of-ssa would be apropos for 4.0 unless the
new algorithm is significantly simpler, etc (maybe one of the above would be, i
don't know).



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=19038