------- Comment #8 from christoph dot mallon at gmx dot de 2007-02-19 16:28
-------
(In reply to comment #7)
> Constant propagation propagates through the PHI in this case:
I basically wrote this in comment #4 and #6 (though I thought it was more
general, see the sidenote below)
> [...]
>
> so this would be the place where a warning could trigger. But it's likely
> this will introduce more false positives (so the "may" form is only done
> very late).
An optimisation - which is not stricly necessary and could just be left out -
removes the cause for the warning, therefore i doubt this generates false
positives. Also this warning was correctly reported before GCC 4.1, why should
it suddenly trigger false positives?
> it's like -Wstrict-overflow -- warn about valid optimizations of undefined
> code (that is here, make it "dead"). In that sense it's a dup of the
> mentioned PRs (it's treated as dead code).
>
> So, let's make this a request for a warning for an optimization based on
> the undefinedness of a value instead ;)
Why should this be a request? I still think it is a regression, because all
GCCs before 4.1 correctly reported this warning (which i stated in the original
posting).
Further I think this warning is highly valuable. I often write code where a
local variable gets declared and later gets values assigned in different code
paths and I want that my compiler tells me if i forgot to do so on one path.
I also expanded this test case a little:
int f(int x, int z)
{
int y;
if (x == 42) y = z;
return y;
}
The only difference is that I don't assign a constant, but another local
variable to 'y'. Now the optimisation does not trigger (because it stricly
operates on constants). So this is very confusing: If the code operates on
constants the warning is missing (if the Phi is simple enough). On the other
hand it gets shown if non-constants or Phis with multiple non-undefined values
are involved.
Sidenote: In the non-constant case this could also be counted as missed
optimisation. If the value gets calculated in a dominating block, the Phi could
also be replaced by this value.
IAACC - I Am A Compiler Constructor. My student research project was about
if-conversion on SSA, so I think I know a bit about this stuff in general. (;
--
christoph dot mallon at gmx dot de changed:
What |Removed |Added
----------------------------------------------------------------------------
Severity|enhancement |minor
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30856
