[PATCH] D27152: Merge strings using sharded hash tables.

[Rui Ueyama via llvm-commits]

Our guarantee to our users can be different from our internal goal (for
ease of development) as long as the former is more relaxed than the latter.
We shouldn't guarantee to users no more than "if you give the exact same
inputs, you'll get the same binary". But we want to keep command line
options from changing output as long as possible so that they are easy to
debug. They don't contradict.

On Sun, Dec 11, 2016 at 7:55 PM, Rafael Avila de Espindola <
rafael.espindola at gmail.com> wrote:

> Sean Silva <chisophugis at gmail.com> writes:
> > That is the rule that I am in favor of, but Rafael seems to prefer that
> > --threads/--no-threads should not affect the output. Hence we need to
> come
> > to agreement and document the conclusion.
>
> I would appreciated if we tried to avoid the difference. Doing a final
> sort on the string number (i.e, when it was first seen) would also allow
> dynamically varing how many shards we use, so it might be a good thing
> for performance too.
>
> > Also, the wording should be clear that the output *binary* is guaranteed
> > the same. We are already nondeterministic w.r.t. the order of user
> > diagnostics.
>
> Correct.
>
> Cheers,
> Rafael
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20161212/6c3cf610/attachment.html>