linspace construction failures

[StefanKarpinski]

Specifically this line fails on 32-bit systems. The failure is an InexactError and goes away when linspace construction in that loop is commented out, suggesting that there's some code somewhere that does a trunc, floor or ceil to Int with a value that is too large to represent as a 32-bit integer.

[timholy]

Hmm, tricky. I just searched through twiceprecision, and all the Int conversions that look dangerous seem to be protected already via a maxintfloat. There are some unprotected computations on len, but I don't think that's relevant for any of your test cases and probably isn't a bug anyway.

[StefanKarpinski]

I think this will require a 32-bit VM to debug. Fortunately, it fails on both Win32 and Linux32.

[tkelman]

you can use multilib compilers to build 32 bit julia on 64 bit linux. set ARCH = i686 in Make.user.

[JeffBezanson]

I just saw what looks like a range length test failure on 64-bit, apparently due to unlucky random numbers:
https://travis-ci.org/JuliaLang/julia/jobs/199738457

[StefanKarpinski]

Hmm. Guess I didn't test thoroughly enough. Hi ho, hi ho, it's off to work we go...

[StefanKarpinski]

I've actually managed to reproduce this on 64-bit with enough iterations. Some cases:

(start, step, stop, n) = (-1.7123204f0, 3.2072393f-8, -1.7123204f0, 2)
(start, step, stop, n) = (1.3173739f0, -2.3759904f-8, 1.3173739f0, 3)
(start, step, stop, n) = (1.2467973f0, 4.1729544f-9, 1.2467973f0, 2)