Pavel Panchekha

Results for this were quite bad on the nightlies, surprisingly so. Clearly I'm not going to merge this, but let me leave the PR open for a bit so I...

Great issue. By default, Herbie turns off support for the `hypot` function—if that's turned on it does recommend using the library `hypot` function. So without access to `hypot` what it's...

It would be nice if Herbie did something better—perhaps `|x| + sqrt(1 + y^2)`—but that at least explains the result you're seeing.

Right, yes. I meant to write that above but messed it up! So when Herbie says `[x, y] = sort([x, y])` it means you should sort the `x` and `y`...

That's pretty funny, @dalev. Here's what's going on here: - First, Herbie proposes sorting x, y, and z. That means that y is closer to zero than either x or...

Thanks @dalev and all for the bug report. For now, a quick workaround you can try is passing `--disable generate:taylor`. That's the pass that generates the buggy output and disabling...

The latest unstable version should work, but easier yet is --enable rules:numerics —Pavel Panchekha > On Aug 17, 2021, at 2:48 PM, Christoph Lassner ***@***.***> wrote: > >  >...

Good suggestion. We don't have `ldexp` right now because it takes an integer argument—that doesn't work for complicated and uninteresting reasons—but we should be able to do it soon.

Thanks for the bug report! Let me what I can answer... First, bits of error is absolutely measuring ULPs of error and taking the logarithm, not strictly counting bits. More...

Herbie now regularly achieves the correct suggested result, so marking this fixed. On the subtler issue of people using Herbie’s results directly, uhh, can’t really fix that but I think...