[llvm-dev] [poison] is select-of-select to logic+select allowed?

[Davide Italiano via llvm-dev]

On Tue, May 23, 2017 at 9:27 AM, John Regehr via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
> Would it make sense to first commit to a semantics for select, document
> this, and attempt to comment out all now-invalid optimizations?  We don't
> want to remove them entirely since, as Nuno points out, we can bring some of
> them back later using appropriate freezes.
>
> What I don't know is who would be affected by the performance regressions in
> the meantime, but I suspect that they wouldn't be too big of a deal for most
> of us.
>
> John
>

As somebody who has been bitten by `undef` funny behaviour several
times (most recently in NewGVN, see, e.g.
https://bugs.llvm.org/show_bug.cgi?id=33129), I'm in favour of this.
My anecdotal evidence/past experience is that it causes too much pain
for relatively little benefit.
In fact, we're (likely) going to disable aggressive `undef`
optimizations, at least for PHIs, in NewGVN, at least until `undef` is
fixed.
For `select` in particular, it doesn't seem to be a big problem, but I
think some benchmarking is needed to make sure things don't regress
badly.

--
Davide