[Bug middle-end/78977] [7 Regression] g++7 snprintf() of double produces wrong code with -O3

[h2+bugs at fsfe dot org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78977

--- Comment #3 from h2+bugs at fsfe dot org ---
(In reply to Jakub Jelinek from comment #1)
> You haven't provided a self-contained test, so it is hard to know exactly,
> but I bet it is the -fprintf-return-value optimization.  Try
> -fno-printf-return-value if it helps, 

Yes, that "fixes" it.

> if it does, then we really need self-contained small test.

Like I said, it doesn't happen on all the time and I have had difficulties
extracting it. I could make something smaller than our whole repo, but it would
likely still be quite a lot of code. Or is there some way I can make GCC dump
an intermediate version of all the code that it actually uses for building the
binary?

> Perhaps FreeBSD libc snprintf prints %g differently from what GCC expects,
> rounds it differently etc.

Hm, could be. Note that FreeBSD still has some brokenness in the lib that
mandates using -D_GLIBCXX_USE_C99=1 if you want C99 (which is required for
c++11). But I thought this was only related to lacking precision in the math
library.

(In reply to Martin Sebor from comment #2)
> It's possible that this bug will be fixed by my patch for bug 78696.  The
> patch was approved just before the holidays but I am yet to commit it.  Let
> me retest it and check it in today.

Great, let me know if it's merged, then I will try a newer snapshot!