Re: [Sbcl-devel] slow printing of floating point numbers in SBCL

[Nathan F. <fr...@gm...>]

On 6/5/07, Nikodemus Siivola <nik...@ra...> wrote:
> I wonder is if using a small memoization cache would help -- obviously
> it would help with the "print PI in a loop" test

FWIW, a profile of "print a million random single-floats" with current SBCL
looks like:

           Self        Total        Cumul
  Nr  Count     %  Count     %  Count     %    Calls  Function
------------------------------------------------------------------------
   1     33  13.5     33  13.5     33  13.5        -  SB-VM::GENERIC-+
   2     31  12.7     84  34.4     64  26.2        -  SB-IMPL::FLONUM-TO-DIGITS
   3     28  11.5     44  18.0     92  37.7        -
SB-IMPL::OUTPUT-BYTES/UTF-8
   4     22   9.0     65  26.6    114  46.7        -  SB-IMPL::FD-SOUT
   5     21   8.6     21   8.6    135  55.3        -  TRUNCATE

Everything else is < 4%.

Eliminating the first two will likely involve some careful analysis and
inserting declarations in FLONUM-TO-DIGITS.

Eliminating the second two may be simpler: there is a FIXME in FD-SOUT:

                 (and (find #\newline thing :start start :end end)
                      ;; FIXME why do we need both calls?
                      ;; Is find faster forwards than
                      ;; position is backwards?
                      (position #\newline thing
                                :from-end t
                                :start start
                                :end end)))))

Surely we don't need to have both calls, do we?  Just calling POSITION
should work fine, right?

There is also this clever optimization:

          (if (and (typep thing 'base-string)
                   (eq (fd-stream-external-format stream) :latin-1))
              (ecase (fd-stream-buffering stream)
                (:full
                 (output-raw-bytes stream thing start end))
                (:line
                 (output-raw-bytes stream thing start end)
                 (when last-newline
                   (flush-output-buffer stream)))
                (:none
                 (frob-output stream thing start end nil)))
              ;;; ...actually do external-format conversion

It seems that another win would be to check for :utf-8 or whatever
and try to send base-strings when printing floats (FLONUM-TO-DIGITS
does this, but FLONUM-TO-STRING may not).  Doing this extra check
might also speed up other code that is careful about outputting
base-strings.  (It is also yet another reason to implement real
base-char string-streams...)

Do these sound reasonable?

-Nathan