Test failure on powerpc and ppc64el

[avalentino]

We are trying to update debian packages for numexpr to v2.6.1 and we have test failures on powerpc and ppc64el [1], [2]
I feels very strange because v2.6.0 (currently in testing) seems to work correctly, and changes between v2.6.0 ans v2.6.1 are really minimal [3].

The main change between the two builds seems to be the updated gcc version (5.3.1 vs 6.1.1).

Any idea?

[1] https://buildd.debian.org/status/fetch.php?pkg=numexpr&arch=powerpc&ver=2.6.1-1&stamp=1472197317
[2] https://buildd.debian.org/status/fetch.php?pkg=numexpr&arch=ppc64el&ver=2.6.1-1&stamp=1472197359
[3] v2.6.0...v2.6.1

[FrancescAlted]

Hmm, this is really strange. I have no experimented with gcc 6 yet, but is there a way to build numexpr 2.6.1 with gcc-5? I am using gcc 5.4.0 (before was gcc 5.3.something) on an Ubuntu 16.04 box and all the tests passes.

[avalentino]

Thanks @FrancescAlted, Indeed it is very strange.
The default gcc version on the next Debian stable will be gcc 6 so I think it is not possible to use old versions of building packages.
I will ask for a guest access to powerpc and ppc64el platforms to investigate.

[leitao]

This seems to be an issue on converting double to long, which is truncating other than rounding. For example, powl(3,3) is returning 26,99999999996, which is being trucated to 26 instead of 27.

@ferseiti created a initial patch that fixes this version. I just reviewed and we should send a final patch in a PR next week.

[ferseiti]

@leitao That is correct. That also happens with other values.
The fact is that simply casting will not guarantee what we expect to be a default behaviour (rounding).
Just please disregard the patch linked, since I removed it.
I will commit it with llround which is the correct type-generic argument macro for returning long long type.

[ferseiti]

Hello.
I have created a PR for this.
I have tested it on ppc64el and x86_64.
Regards.

[avalentino]

@FrancescAlted probably this one can be closed since PR #235 has been merged now

[FrancescAlted]

Thanks @avalentino for the reminder. Closing.