This is the mail archive of the 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] Fix PR34263: Cleaning up latch blocks


Thanks for your reply!

> Effectively stashing the value of a_2 to load back in the case where we
> exit. Is this right? (except it is actually working on the coalesced
> root variables, not the actual SSA_NAMES)

Yes, that's correct.

> Seems to me there is an older PR that involves a similar situation too ,
> but I dont recall it.

It might be PR19038, which this patch was originally taken from.

> Should there be some kind of limit to the number of copies being
> inserted? Ie, if there are 10 copies being inserted on the edge, I would
> think 10 extra copies in the loop would tend to start adding up and
> offset whatever gain you might get.

That sounds reasonable, I can add a default number to consider only small
latches. (and maybe add a new --param option to tune it?)

> As long as there isn't disagreement on whether the extra copies in and
> out of the loop are a problem, the patch istself is OK.

I will address your comments and re-submit the patch.


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