Re: [PATCH] Improves __ieee754_exp() performance by greater than 5x on sparc/x86.

[Patrick McGehearty <patrick dot mcgehearty at oracle dot com>]

On 10/20/2017 6:41 AM, Szabolcs Nagy wrote:
> On 19/10/17 23:31, Patrick McGehearty wrote:
> > Table of hex float constants. I can readily adjust the formating. What
> > you see is the formating used in the original source.
> > I've been uncomfortable with hex floats approach
> > as it only works for ieee754 representations
> > that use base 2. I admit that is most current machines.
> your entire algorithm depends on ieee754 binary representation,
> that's not a good reason for avoiding hexfloats.
>
> decimal floats are not even required to be correctly rounded by
> the compiler in iso c, they are only faithfully rounded, so
> this is a portability bug in the original source too, you can
> silently get completely wrong code generation because of it.
>
> > And the prior ieee754 exp table uses hex format.
> > My second reason for resisting the change is my philosophy
> > when porting code is that every change without a good reason
> > is an opportunity to introduce errors without corresponding benefit.
> >
> > If you feel strongly about suing hex constants, I will make an effort
> > to convert these values to hex format.  This conversion seems likely
> > to require the most effort on my part.
You make good points. In the prior context of the code being built
and used for Sparc/Solaris, the Studio compiler was known to
correctly round floating point constants since the early 90's.
In the gnu context, we have to be prepared for any C standard
compliant compiler, which is a greater challenge.
I'll make the conversion to using hex constants. I'm wondering
if it would be beneficial to retain the decimal constants in comment form?
Maybe not for the big table, but for the constants used for
specific thresholds, etc.  I know when I first looked at
some of this code, I wrote a little program to display the
decimal equivalents of some of the hex constants to better understand
what the control flow was doing.

- patrick