On Wed, 13 May 2026 at 19:02, Dragon Archer <[email protected]> wrote:
>
> On Wed, 13 May 2026 at 16:46, Jonathan Wakely <[email protected]> wrote:
> > On Wed, 13 May 2026 at 16:14, Dragon Archer <[email protected]> 
> > wrote:
> > >
> > > Thank you for your review and thoughtful questions.
> > >
> > > > Seems reasonable to me (though pragma push/pop_macro don't seem strictly
> > > > necessary as Jonathan pointed out).
> > > >
> > > > Though in theory I guess there's a risk of a performance regression by
> > > > removing an assert in Ryu, if the compiler relies on an assert condition
> > > > to be true in order to safely apply an important optimization.
> > >
> > > I understand the concern. In my view:
> > >
> > > 1. Replacing an unconditional `assert` (which always evaluates the 
> > > condition at runtime) with `__glibcxx_assert` (which is completely 
> > > removed when `_GLIBCXX_ASSERTIONS` is not defined) actually **removes** 
> > > runtime checks in non‑debug builds. This performance benefit likely 
> > > outweighs any hypothetical optimization that the compiler might have 
> > > derived from the assertion.
> > >
> > > 2. If we are worried about missing optimization hints, we could consider 
> > > using `__builtin_assume` – that would be more semantically precise. 
> > > However, that would require guarding it with `#ifdef _GLIBCXX_ASSERTIONS` 
> > > and defining a fallback. I personally prefer keeping the current approach 
> > > with `__glibcxx_assert` because it is already used throughout libstdc++ 
> > > and matches the debugging philosophy.
> >
> > I don't want such changes in the Ryu code, it will just make it
> > difficult to rebase on upstream. However, we might be able to replace
> > it with Teju Jagua anyway.
>
> As discussed in PR/125228, this patch only changes
> libstdc++-v3/src/c++11/debug.cc and 
> libstdc++-v3/src/c++17/floating_to_chars.cc,
> both are pure GCC files that are not shared with upstream.
> So there should be no rebase issues.

Yes, I know this patch doesn't change the ryu code. But to add uses of
__builtin_assume into the ryu code would require changing the ryu
code. Just replacing all uses of assert with __builtin_assert is fine,
but replacing them with __builtin_assume would not be fine, and so
more specific changes to the ryu code would be needed.

And I'm saying I wouldn't want such changes. You suggested it as
something to consider, I'm saying it's not worth considering.

>
> > >
> > > 3. If the community later decides that `__glibcxx_assert` should also 
> > > serve as an optimization hint in release mode, they could define it to 
> > > `__builtin_assume` instead of empty when `_GLIBCXX_ASSERTIONS` is **not** 
> > > defined. That would be a separate change, and will automatically be 
> > > effective for this patch as well.
> >
> > That won't happen while I have any say in the matter.
> >
>
> I didn't want to propose that change either.
>
> > >
> > > So for this patch, I believe using `__glibcxx_assert` is the right and 
> > > simple solution.
> > >
> > > > Just wondering, is there a legit problem/concern with __FILE__ appearing
> > > > in the final library or is it a matter of QoI (which I agree with)?
> > >
> > > It is primarily a QoI improvement. The presence of `__FILE__` strings 
> > > increases binary size and may leak internal build paths (which is a minor 
> > > privacy/security concern for some users).
> >
> > Users who care about that can easily compile in /tmp or another
> > non-secret path. You can also compile with --enable-cxx-flags=-DNDEBUG
> > to disable those assertions in your own builds.
> >
>
> Yes, that's true, but I think this patch is clean enough and can be
> made available in upstream without causing any issues. It is a small
> change that improves the quality of the library without affecting
> functionality (and may even improve performance in non‑debug builds).
>
> > > Eliminating them makes the library cleaner, but there is no functional 
> > > correctness issue. BTW, I think the removal of runtime checks will 
> > > improve perfomance slightly.
> > >
> > > Again, thank you for your time and feedback.
>

Reply via email to