https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71177
Tavian Barnes <tavianator at gmail dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
Attachment #38516|0 |1
is obsolete| |
--- Comment #2 from Tavian Barnes <tavianator at gmail dot com> ---
Created attachment 38520
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=38520&action=edit
Further reduced testcase
> Please don't set regression markers if you don't know what you are doing (but
> feel free to suggest that something should have regression markers).
>
> Happens also in 5.3.0, 4.9.3 and 7.0, so not a regression.
Oops, sorry! This only showed up in my actual build on upgrading to GCC 6, but
I didn't check the reduced version against older GCCs.
> If you could reduce the testcase further and point out at a missing
> optimization, the chances of getting this fixed would increase. Otherwise,
> someone needs to look at the dump files and figure out why the ranges are not
> affected by the check and whether they should.
I've attached a further reduced testcase.
> The 'else' block is not unreachable, but loop optimizations look at the
> possible
> ranges within the loop and those may not be taking into account any range
> information derived from the caller.
The else block is unreachable *from within normalize()*. And since the warning
only shows up due to constant propagation from normalize(), it would be nice to
avoid it.
Specifically, normalize() calls resize(length - 2) which calls (in the
unreached else block) append(n - length). (n - length) gets constant-folded to
(size_t)-2 which then overflows a ptrdiff_t later.
