To that end, the attached patch adds the checker under its own new
pass. The pass runs very early without optimization, and relatively
late with it to benefit from the passes above. With optimization
the pass also can (but doesn't yet) fold the return value from these
functions into a constant. In the Linux kernel, it folds 88 snprintf
calls (I didn't find any instances where the whole call could be
folded into a string constant).
So I haven't dug into the patch yet, but what (if any) interaction is
there between the early and late passes? Do they both emit warnings, or
is it more that the early passe builds a candidate set which is refined
by the later pass, or something totally different?
Thanks for the feedback.
The pass only runs once. Without optimization, it's invoked early,
does its thing, and is never invoked again. With optimization,
it's also invoked early but returns without doing anything, only
to to invoked again much later on to do real work. I wasn't sure
where among all the passes might be the best spot to insert it so
I put in the same place as Aldy's -Walloca patch because it relies
on the same prior passes.
I tested the patch with LTO but due to bug 71907 no warnings are
issued. Once the bug is resolved I'll re-test it and see about
adding test cases to verify it.
Note Richi checked in a patch a couple days ago to stream the abstract
origins, which is supposed to fix 71907.
Thanks for the heads up. I'll give it a whirl.
Onward to looking at the patch :-)
Re-reading it again myself I see a few typos, commented out chunks
of code left over from debugging, and missing comments. I'll clean
this up in the next update along with other requested changes.
Martin
