Re: Help needed: r-base fails with its Bessel functions
- To: Dirk Eddelbuettel <edd@debian.org>
- Cc: Phil Blundell <pb@nexus.co.uk>, debian-arm@lists.debian.org
- Subject: Re: Help needed: r-base fails with its Bessel functions
- From: Nicholas Clark <nick@ccl4.org>
- Date: Tue, 9 Oct 2001 11:30:13 +0100
- Message-id: <20011009113013.A32194@plum.flirble.org>
- In-reply-to: <20010921072942.A24685@debian.org>; from edd@debian.org on Fri, Sep 21, 2001 at 07:29:43AM -0500
- References: <20010920161152.A12694@debian.org> <edd@debian.org> <E15kL4n-0005xf-00@mill.nexus.co.uk> <20010921072942.A24685@debian.org>
On Fri, Sep 21, 2001 at 07:29:43AM -0500, Dirk Eddelbuettel wrote:
> On Fri, Sep 21, 2001 at 08:50:21AM +0100, Phil Blundell wrote:
> > >Could you Arm experts look into this? Is there a genuine Arm floating point
> > >issue hiding?
> >
> > Maybe. ARM has a wacky format for storing doubles, so if R wants to poke
> > around with the actual binary representation of floating point numbers it will
> > probably run into trouble.
>
> Note that a) r-base 0.90 compiled fine (since then we used to require Atlas
> to build, so that never happened on too many archs, now re-designed) abd b)
> R has a gazillion other regression tests that would fail.
>
> I suspect it is either the src/nmath/bessel* code, or its interaction with
> libc. Recall that the segment which failed passed NaN, NaN around as
> arguments.
What kernel is on the machine attempting to build? IIRC there were NaN
bugs in NWFPE 0.94 (or whatever is with 2.2.18) that are fixed in the next
NWFPE (whatever is in 2.4.0). Russell backported the later NWFPE to 2.2.19
[bugs were specifically that some comparison operators, (including LE, IIRC)
were not behaving correctly when NaN was one of the arguments.]
Nicholas Clark
Reply to: