On 15/08/2026 08:38, Collin Funk wrote:
Pádraig Brady <[email protected]> writes:
On 15/08/2026 06:11, Collin Funk wrote:
I pushed this mostly because I prefer marking default cases with
unreachable so the reader knows that not all representable values
can/should occur there. However, I did notice it from an interesting
-fanalyzer bug/limitation.
-- 8< --
Analyzer mistakenly reports that the CSWTCH lookup table can
overflow
here since it cannot infer the value of C from the previously called
switch statement in main. See:
<https://gcc.gnu.org/bugzilla/PR126885>.
* src/stdbuf.c (optc_to_fileno): Add a default label with a call to
unreachable.
---
src/stdbuf.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/src/stdbuf.c b/src/stdbuf.c
index ff19c6056..fb7de79b3 100644
--- a/src/stdbuf.c
+++ b/src/stdbuf.c
@@ -184,6 +184,8 @@ optc_to_fileno (int c)
case 'o':
ret = STDOUT_FILENO;
break;
+ default:
+ unreachable ();
}
return ret;
That doesn't introduce a bug but does introduce coupling.
I.e. we only call optc_to_fileno() with one of the handled cases,
but that function in isolation can be called with anything,
so the code is now brittle if adjusted in future.
Also unreachable() is dangerous IMHO. See:
https://github.com/coreutils/coreutils/commit/e661c7a52
There with clang at least I saw code just run off the end
of a function into arbitrary code.
I need to audit/think more about its use in coreutils.
Ouch. I forgot about that commit.
I never really minded just calling abort, to be honest. I guess that
affirm (false) is nicer since you can leave it up to the person
configuring what behavior they prefer.
Yes we've more liberties here than a library for e.g.
affirm(false) is better as by default we'll abort() unless built with
--disable-assert
A more appropriate adjustment might be to:
default:
ret = -1;
break;
Yes, that makes sense. Sadly, though, analyzer is still unhappy with it.
But I guess analyzer is always a bit noisy, so one shouldn't lose sleep
over it.
That's surprising.
TBH I've not gained _anything_ useful from -fanalyzer yet.
cheers,
Padraig
