[libc-commits] [libc] [libc][math] Fix double-rounding errors with denormal inputs for hypot function. (PR #208247)

[via libc-commits]

lntue wrote:

> the rule for underflow after rounding is somewhat complex. I tried to explain it in slides 40-41 of https://members.loria.fr/PZimmermann/talks/module1.pdf

Thanks for the slides and the clear explanation!  The plot in the slide 41 is exactly why I'm leaning toward raising underflow before rounding whenever we are in control of the rounding or with not much cost, since that can be done without probing the current rounding mode.

I'm switching this back to underflow with pre-rounding.

https://github.com/llvm/llvm-project/pull/208247