[dataflow] Don't rescan everything after CSE
Tue Mar 6 16:38:00 GMT 2007
Paolo Bonzini wrote:
> This saves ~1% of compilation time.
> More ideas that I have, but did not implement are:
> 1) delay computation of LR/UR until where we did it in the trunk. fwprop and GCSE do not need it, for example.
I do not think that this will buy you that much. Nothing is actually
done with these unless you call df_analyze and if the passes do not call
it, there is no work done. DF lets all of the dataflow problems rot
until you do a df_analyze. there is no tweeking of the solutions as
changes are made. The only thing that is done, is that dirty bits get
set as structural changes are made.
> 2) don't update REG_DEAD/REG_UNUSED notes for patches that don't need it. Updating them could/should be a changeable flag.
I had thought about this and then rejected it because it is too complex
(not from the point of view of the clients but just the amount of code
that had to be changed in df-scan). I also decided that many of the
optimizations that ignore the notes are just buggy and will eventually
get changed so i feared that the amount of time to build them and tear
them down would eventually rise.
> Bootstrapped/regtested i686-pc-linux-gnu.
More information about the Gcc-patches