RFC: removing slow paths in various dbl-64 libm functions

[Florian Weimer]

On 12/03/2015 04:01 AM, Carlos O'Donell wrote:
> On 12/01/2015 07:41 PM, Joseph Myers wrote:
>> I don't think timing oracle avoidance is a goal for libm functions (or for 
>> glibc functions in general, e.g. memcmp).
> 
> I agree that at present there are no goals regarding timing oracle avoidance.
> 
> Any discussions about timing oracle issues would be had in a new thread and
> with clear guidelines discussed in order to attain community consensus.

I don't think it's reasonable to avoid timing oracles for math
functions.  I would expect data-dependent timing for quite a few
floating point instructions (notably division, we don't even have to go
into NaN territory).

Non-polynomial functions (such as trigonometric functions) are used in
cryptography, but only to generate constant tables which are not
data-dependent at all, and those tables usually are precomputed, to
avoid issues with platforms where these floating-point operations have
errors so large that the computed tables are not correct.

Here's an example which uses |sin n|:

  <https://tools.ietf.org/html/rfc1321#section-3.4>

> The particular case of memcmp is certainly the most interesting, and Florian
> and I have discussed it privately. A public discussion is welcome if someone
> wants to start it.

There has been some public discussion as well:

  <https://sourceware.org/ml/libc-alpha/2014-01/msg00763.html>

Florian