parse(Float64,::String) doesn't respect rounding mode

[dpsanders]

In

https://travis-ci.org/dpsanders/ValidatedNumerics.jl/builds/84036519

the tests for the ValidatedNumerics.jl package, which include lots of rounding mode changes,
run correctly on Mac OS, but give some incorrect results on Linux.

Could this be due to extended precision mode? How can this be turned off?
Or is there another explanation?

Cc @simonbyrne

[simonbyrne]

I'm not really familiar with the package: can you provide some specific examples of the incorrect behaviour?

[dpsanders]

I'll try and reduce to some minimal cases. cc @lbenet

[vtjnash]

Are you calling libm functions? I seem to recall that those may modify the rounding mode on some systems, or give incorrect answers if the rounding mode is changed.

[dpsanders]

We should just be doing rounding mode changes like

julia> a = with_rounding(Float64, RoundDown) do
           parse(Float64, "0.1")
       end
0.09999999999999999

julia> 

julia> b = with_rounding(Float64, RoundUp) do
           parse(Float64, "0.1")
       end
0.1

[JeffreySarnoff]

with your example using rc4 on Win7, a==b==0.1
Julia Version 0.4.0-rc4 Commit e9c9c92* (2015-10-04 03:14 UTC)
System: Windows (x86_64-w64-mingw32)
CPU: Intel(R) Core(TM) i7-4960X CPU @ 3.60GHz WORD_SIZE: 64

with your example using rc2 on Linux (same box), a!=b, b==0.1, a=b-eps(b)
Julia Version 0.4.0-rc2 Commit fa52609* (2015-09-18 17:51 UTC)
System: Linux (x86_64-unknown-linux-gnu)
CPU: Intel(R) Core(TM) i7-4960X CPU @ 3.60GHz WORD_SIZE: 64

[JeffreySarnoff]

and using rc4 on same Linux also worked correctly

[simonbyrne]

From what I understand, we use the system strtod for this, so I'm guessing this is a problem with glibc. They're actually pretty careful about this stuff, so if that's the problem it may be worth reporting upstream. What version are you using? (you can find out by running /lib/libc.so.6).

[lbenet]

The expected results for the tests (which fail in travis!) correspond to your results in Linux; we get exactly the same in our Macs and also in linux with Julia Version 0.4.0-rc4+4, Commit 7048c40* (2015-10-07 00:33 UTC). The system is Linux (x86_64-suse-linux)
CPU: AMD Opteron(tm) Processor 6376.

As for Simon's question, on that linux machine we have:
GNU C Library (GNU libc) stable release version 2.17 (git c758a6861537).

Do you know this info for travis?

[JeffreySarnoff]

GNU C Library (Ubuntu EGLIBC 2.19-0ubuntu6.6) stable release version 2.19,
Copyright (C) 2014 Free Software Foundation, Inc.
Compiled by GNU CC version 4.8.2.
Compiled on a Linux 3.13.11 system on 2015-02-25.

[simonbyrne]

This looks like it might have been fixed in glibc 2.17:
https://sourceware.org/bugzilla/show_bug.cgi?id=3479

From what I can tell, travis is running eglibc 2.15, so that would explain it. I don't know if there is any way to update it via apt (though I would suspect not, since it is so tightly linked to the OS).

At some point in future I think we should try to implement our own version, this would also fix our issues with MSVC (#6349).

[lbenet]

This looks like it might have been fixed in glibc 2.17:
https://sourceware.org/bugzilla/show_bug.cgi?id=3479

From what I can tell, travis is running eglibc 2.15, so that would explain it. I don't know if there is any way to update it via apt (though I would suspect not, since it is so tightly linked to the OS).

@tkelman Is it possible to upgrade travis to incorporate some bug fixes in glibc?

[tkelman]

They may have an opt-in way to use ubuntu 14.04 by now, not sure. If it is only available via a docker worker then we would have to replace all the apt-get steps in our travis setup.

[tkelman]

If you're asking about a package, then try putting the following in your yml:

sudo: required
services:
  - docker

Actually reading http://blog.travis-ci.com/2015-08-19-using-docker-on-travis-ci/, it looks like we could maybe transition our main Linux Travis setup to caching things via docker.

[lbenet]

The problem is indeed related with our package (see JuliaIntervals/ValidatedNumerics.jl#56); while tests pass locally (in Linux and Mac), they fail in travis.

I followed your suggestion and finally the tests pass also in travis! Thanks a lot !!!

[dpsanders]

These seems to be fixed.