[Rd] Issues of R_pretty in src/appl/pretty.c

Martin Maechler maechler at stat.math.ethz.ch
Tue Aug 15 10:55:10 CEST 2017


>>>>> Martin Maechler <maechler at stat.math.ethz.ch>
>>>>>     on Mon, 14 Aug 2017 11:46:07 +0200 writes:

>>>>> Suharto Anggono Suharto Anggono via R-devel <r-devel at r-project.org>
>>>>>     on Fri, 11 Aug 2017 17:11:06 +0000 writes:
>>>>> Suharto Anggono Suharto Anggono via R-devel <r-devel at r-project.org>
>>>>>     on Fri, 11 Aug 2017 17:11:06 +0000 writes:

    >> See https://stat.ethz.ch/pipermail/r-devel/2017-August/074746.html for the origin of the example here.

    >> That
    >> pretty(c(-1,1)*1e300, n = 1e9, min.n = 1) gave 20 intervals, far from 1e9, but
    >> pretty(c(-1,1)*1e300, n = 1e6, min.n = 1) gave 1000000 intervals
    >> (on a machine), made me trace through the code to function 'R_pretty' in https://svn.r-project.org/R/trunk/src/appl/pretty.c .

    > thank you.

    >> *lo is -1e300, *up is 1e300.
    >> cell = fmax2(fabs(*lo),fabs(*up));
    >> 'cell' is 1e300.
    >> i_small = dx < cell * U * imax2(1,*ndiv) * DBL_EPSILON *3;
    >> When *ndiv is (int) 1e9, apparently cell * U * imax2(1,*ndiv) overflows to infinity and 'i_small' is 1 (true). It doesn't happen when *ndiv is (int) 1e6.

    > well spotted!

    >> Putting parentheses may avoid the floating point overflow. For example,
    >> i_small = dx < cell * (U * imax2(1,*ndiv) * DBL_EPSILON) *3;

    > yes... but only if the compiler optimization steps  "keep the parentheses".
    > AFAIK, there is no guarantee for that.
    > To make sure, I'd replace the above by

    > U *= imax2(1,*ndiv) * DBL_EPSILON;
    > i_small = dx < cell * U * 3;


    >> The part
    >> U = (1 + (h5 >= 1.5*h+.5)) ? 1/(1+h) : 1.5/(1+h5);
    >> is strange. Because (h5 >= 1.5*h+.5) is 1 or 0, (1 + (h5 >= 1.5*h+.5)) is never zero and 1/(1+h) will always be chosen.

    > Yes, strange indeed!
    > here was as a change (not by me!) adding wrong parentheses
    > there (or maybe adding what the previously "missing" parens
    > implied, but not what they intended!).
    > The original code had been
     
    > U = 1 + (h5 >= 1.5*h+.5) ? 1/(1+h) : 1.5/(1+h5);

    > and "of course" was intended to mean

    > U = 1 + ((h5 >= 1.5*h+.5) ? 1/(1+h) : 1.5/(1+h5));

    > and this what I'll change it to, now.


    >> The comment for 'rounding_eps' says "1e-7 is consistent with seq.default()". Currently, seq.default() uses 1e-10 as fuzz.

    > Hmm, yes, thank you; this was correct when written,
    > but seq.default had been changed in the mean time,
    > namely in  svn r51095 | 2010-02-03

    > Usually we are cautious / reluctant to change such things w/o
    > any bug that we see to fix.
    > OTOH, we did have  bug cases we'd wanted to amend for seq() /
    > seq.int();
    > and I'll look into updating the "pretty - epsilon" also to
    > 1e-10.

    > Thank you for your analysis and suggestions!

I've committed now what I think has been suggested
above ... to R-devel only :
------------------------------------------------------------------------
r73094 | maechler | 2017-08-15 09:10:27 +0200 (Tue, 15. Aug 2017) | 1 Zeile
Geänderte Pfade:
   M doc/NEWS.Rd
   M src/appl/pretty.c
   M src/main/engine.c
   M tests/reg-large.R
   M tests/reg-tests-2.Rout.save

pretty(x, n) fix overflow for large n suggested by Suhartu Aggano, R-devel, 2017-08-11



More information about the R-devel mailing list