Re: [PATCH?] Implement exp and log builtins as x86 intrinsics

[Richard Henderson <rth at redhat dot com>]

On Sun, Aug 04, 2002 at 02:55:52PM -0600, Roger Sayle wrote:
> At the very least, the fact that these index spaces currently
> struggle to avoid values used by the other is perhaps misleading.

At present, most of the numbers are as were found before
the symbolic constants were added.  I.e. largely at random.

> I take it you believe that defining splitters for exp?f2 and log?f2
> is the correct approach,

Well, not splitters, per se.  Rather just generate them
expanded via define_expand.

> and that there isn't a simple "one line" fix to expose the stack
> temporary that could be used with the "monolith" implementation above?

You might have success with (clobber (match_scratch:XF 1 "=u")).

Also note that the expansion is larger than a call for -Os, and
that P4 optimization guides recommend using software methods for
these insns when less than 80 bits of result is needed.


r~