Re: [tcb] Incremental SSA updates

[Diego Novillo <dnovillo at redhat dot com>]

Zdenek Dvorak wrote:

> I did not try to measure.  But if the patch is really used in an
> incremental fashion, scanning long basic blocks most definitely is not
> cheap.
You need numbers to back that up.  We do very many IL scans in the 
course of the optimizers.  I've never seen IL scans too high in the list.

> No it is not; update_ssa features for example
> rewrite_blocks (start_bb, REWRITE_UPDATE);
Certainly, but blocks that have no uses nor defs of affected names are 
ignored (local patch for now).  So, it doesn't really matter.

> It is supposed to be used for example during updates of loop closed ssa
> form, where you know that only some of the uses are affected.
OK.

> Your approach scans whole subgraph dominated by start_bb, which
> may be significantly more than necessary area.  In the use in vrp,
> it will usually be almost whole program.
Not at all, only the affected blocks and in those affected blocks, the 
affected statements are scanned.

In any case, give me a couple of days to do some timings and take a good 
look at your patch.  I still haven't had the time to examine it closely.


Diego.