[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[no subject]



MRG, JM, and myself discussed the problem of COMPLEX and DOUBLE precision
this morning, and came to the conclusion that
1) COMPLEX with arbitrary-valued components are what is required, and the
   operations on two complices  will have to be done with the generic
   rational operations (since BIGNUMs or FLONUM could occur).  This seems
   to be acceptible especially since we are not now planning to open-compile
   the COMPLEX arithmetic.  For a discussion why, see me.
2) On all new machines, there will be only one FLONUM type, namely what the
   machine will reasonably support.  Generally this means 50 to 100 bits.
   Whatever is less is generally unsatisfactory, and all new winning machines
   do floating-point operations on quantities of this size just as fast as
   on smaller-sized operands (modulo a little epsilon).  The retrofit to
   the KL10 for this would probalby take about 4-6 man months since it
   would mean completely changeing the FLONUM format; and even though this
   could be done in a way that only wasted space (the second word would merely
   be a waste on the KA10s since they dont really have Doubleprecision
   instructions - the loader having converted from DFBSR to FBSR)
   we still dont think the retrofit is worth it.  So for the PDP10 series
   we will admit the new data type DOUBLE, which will be closed-compiled.
   At least some of the benefit of the KL10 floating-point hardware will
   be realized.  Since standard flonums will compile as at present, then
   most of the IMSL stuff will still compile adequately.