AMD K6 support

[Kurt Garloff]

On Thu, Aug 20, 1998 at 08:02:21AM -0400, John Carr wrote:
> 
> The parameters in your patch conflict with the latest optimization
> guide from AMD (for example, AMD says fpmul and imul take 2 cycles).
> 
> I am also working on AMD K6 optimizations for egcs.  I'll send my
> patch for discussion this weekend.

I did not yet look at your patch, sorry.
I updated mine though and fixed a few typos (thanks for pointing it out, 
Kamil) and experimented with my fp code to get better results. Also, the
numbers used should be closer to the AMD  specs, now. fpmul is 3 cycles and
cost_imul is 4. You can argue about these, but at least for imul 4 isn't
unrealistic: AMD docs say 2-3 cycles latency and 2-3 cycles throughput ...
My results went worse when I further decreased those values.

I think your patch is much more complicated and will be included into egcs
later, but for now, I'm happy with my lightweight patch and I hope others
will be, too.

Have a look at the appended diff.

-- 
+--------------------------------------------------------------+
|  cand.phys. Kurt Garloff        C1-O4-101 (Chemiegeb"aude)   |
|Lehrstuhl Hochfrequenztechnik  (Fakult"at f"ur Elektrotechnik)|
|   Universit"at Dortmund             D - 44221 Dortmund       |
|   Tel.: (0231) 755-3947            Fax: (0231) 755-4631      |
|            garloff@hft.e-technik.uni-dortmund.de             |
|                 Public PGP key available !                   | 
+--------------------------------------------------------------+