pgcc-1.0.1 with and without haifa

[Ronald Wahl]

Hi,

I just nbenched pgcc-1.0.1 with and without the haifa scheduler. I noticed
that with the haifa scheduler I get faster integer but slower floating
point code than without it ( on a K6 ). The following optimization(s) will
increase fp-performance on a K6:

	* floating point instructions are limited to the first decoder 

	* floating point instructions can be executed in parallel with
          short decodable integer instructions if the floating point
          instructions is the first one of a pair and both (floating
          point and integer instruction) are short decodable and do not
	  depend on the same ressources (load/store-unit).

	* using of load-execute instructions instead of doing this by two
          seperate instructions. (Maybe this is already done.)

	* Replacing FDIV with FMUL where possible.

	* Doing fp- and integer-multiply in parallel.

	* using of integer instructions to move floating point data.

ron

-- 
\ Ronald Wahl --- rwa@informatik.tu-chemnitz.de   \
 \ WWW: http://www.tu-chemnitz.de/~row             \
  \ Talk: rwa@goliath.csn.tu-chemnitz.de            \
   \ PGP key available by finger to my email address \