axis() produces junk on DEC alpha (PR#274)
ripley@stats.ox.ac.uk
ripley@stats.ox.ac.uk
Mon, 13 Sep 1999 19:08:06 +0200 (MET DST)
On Mon, 13 Sep 1999, Martin Maechler wrote:
> >>>>> On Mon, 13 Sep 1999 17:41:07 +0200 (MET DST), ripley@stats.ox.ac.uk (Brian D. Ripley) said:
>
> >> Date: Mon, 13 Sep 1999 10:57:23 +0200 (MET DST)
> >> From: Albrecht Gebhardt <albrecht.gebhardt@uni-klu.ac.at>
> >>
> >> On Mon, 13 Sep 1999, Prof Brian D Ripley wrote:
>
> BDR> [...]
>
> BDR> I'm puzzled. Are you saying -Inf passes R_FINITE and passes
> BDR> ISNAN? (Wow!, if so.) In any case, let us try and fix the macros
> BDR> (which crop up in many places) not two instances.
> BDR>
> >> NaN passed R_FINITE (see above). Then it gets multiplied by some
> >> values and serves as tickmark length.
> >>
> >> >
> >> > There are too many options in Arith.h to guess: can you please tell us
> >> > what path is being taken (HAVE_FINITE, IEEE_754, FINITE_BROKEN, ...).
>
> >> in src/include/Rconfig.h:
> >> /* Floating Point Arithmetic */
> >> /* #undef HAVE_MATHERR */ /* System V */
> >> #define HAVE_ISNAN 1 /* IEEE Arith indicator */
> >> #define HAVE_FINITE 1
> >>
> >> /* #undef HAVE_FLOATINGPOINT_H */
> >> /* #undef HAVE_FPU_CONTROL_H */
> >> /* #undef HAVE_IEEEFP_H */ /* "-Wall" */
> >> /* #undef HAVE_IEEE754_H */ /* Linux */
> >>
> >> #ifdef HAVE_ISNAN
> >> #ifdef HAVE_FINITE
> >> #define IEEE_754
> >> #endif
> >> #endif
> >>
> >> and some lines below:
> >> /* Bug Workarounds */
> >> /* #undef HAVE_OSF_SPRINTF_BUG */
> >> #define CALLOC_BROKEN 1
> >> #define FINITE_BROKEN 1
> >> #define LOG_BROKEN 1
> >>
> >>
> >> >
> >> > I believe this can all be tested from interpreted R code. I believe you
> >> > should get (is.na(c(2.0, Inf, -Inf, NaN, NA)) etc)
> >> > 2.0 Inf -Inf NaN NA
> >> > is.na F F F T T
> >> > is.finite T F F F F
> >> > is.infinite F T T F F
> >> > is.nan F F F T F
> >>
> >>
> >> > is.na(c(2.0, Inf, -Inf, NaN, NA))
> >> [1] FALSE FALSE FALSE TRUE TRUE
> >> > is.finite(c(2.0, Inf, -Inf, NaN, NA))
> >> [1] TRUE TRUE TRUE TRUE TRUE
> >> > is.infinite(c(2.0, Inf, -Inf, NaN, NA))
> >> [1] FALSE FALSE FALSE FALSE FALSE
> >> > is.nan(c(2.0, Inf, -Inf, NaN, NA))
> >> [1] FALSE FALSE FALSE TRUE FALSE
> >>
> >> Seems that everything is R_FINITE!
>
> BDR> [...]
>
> BDR> OK, so I think you are at the line
>
> BDR> # define R_FINITE(x) ((x) != R_NaReal)
>
> BDR> in Arith.h, and that is clearly inadequate (although it should have
> BDR> caught the NA, I think). Can you please try replacing that line by
>
> BDR> static int R_FINITE(double x) {
> BDR> return !isnan(x) & (x != R_PosInf) & (x != R_NegInf);
> BDR> }
>
> BDR> and try those tests again.
>
> BDR> (Kurt: looks like finite can be broken in more ways than one. I suggest
> BDR> return(finite(1./0.) & finite(0./0.) & finite(-1./0.)) in aclocal.m4,
>
> shouldn't this be "|" instead of "&" [broken if either is finite] ?
Um. Yes, I think so: the man page is almost impossible to decipher, and
I had read it reversed.
> {weak voice in the background:
> it may be worth tyring to be both right and efficient ..
> but I agree that "right" is much more important!
Not, I think, on platforms where the inbuilt ways are broken.
--
Brian D. Ripley, ripley@stats.ox.ac.uk
Professor of Applied Statistics, http://www.stats.ox.ac.uk/~ripley/
University of Oxford, Tel: +44 1865 272861 (self)
1 South Parks Road, +44 1865 272860 (secr)
Oxford OX1 3TG, UK Fax: +44 1865 272595
-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-.-
r-devel mailing list -- Read http://www.ci.tuwien.ac.at/~hornik/R/R-FAQ.html
Send "info", "help", or "[un]subscribe"
(in the "body", not the subject !) To: r-devel-request@stat.math.ethz.ch
_._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._._