Re: [PATCH] Value profile based optimizations, part 1

[Daniel Jacobowitz <drow at mvista dot com>]

On Mon, Jun 09, 2003 at 05:25:05PM +0100, Nathan Sidwell wrote:
> Daniel Jacobowitz wrote:
> 
> >>turn on those optimizations, if not found don't. This might be acheived
> >>by turning them on at -O2 and then silently turn them off if the .da data
> >>lookup fails (i,e, make the flags tri-valued, off/on/weakly-on).
> >
> >
> >Don't we normally require an option before searching for a .da file? 
> >With that option and if a .da file is found, then this seems like a
> >good idea.
> yes, I think Honza's point is that we want to look at -O2 anyway, and
> *silently* not do those optimizations if no .da file is found. After
> all, the user had to give some special option to generate the .da
> file, why should they *have* to do something special to use them?
> 
> The current behaviour is -fbranch-probabilities moans when a .da file
> is not found (which is the right thing).

No, I think his point was that these optimizations should be performed
at -O3 (without profile data) or -O2 (with profile data).

Strongly object to using .da files without an explicit option.  That's
a reproducibility nightmare.  "Why doesn't this testcase trigger the
bug?" "Oh, the user had dd if=/dev/random bs=1024 of=test.da lying
around, and it's named test.o..."

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer