[postgis-devel] [Internet] Re: Postgis 2.4.6 regession tests fails on Solaris

[Greg Troxel]

Raúl Marín Rodríguez <rmrodriguez at carto.com> writes:

>>   something else I didn't think of
>
> There is the different floating point arithmetics that the compiler can make
> use of. IIRC correctly, I've had issues with different compilers using
> different bit sizes for floating point operations 64 or 80 bit, which
> after rounding can lead to different values. This can even happen whithin the
> same compiler: vectorized code uses 64b operators while "normal" uses whatever
> is native (usually 80b in modern x86 cpus).

Good point.  With normal IEEE754 vs extended precision, I hope that it's
a case of both behaviors being ok, and the test case being overly picky.

I think that the 80-bit extended doesn't survive memory, so it's only
while things are in registers.

> Debugging differences like this is hard even locally, so you can understand
> that nobody is too exited to do it without a machine with the "failing"
> architecture.

I certainly understand that!

Someday, I'll get a sparc64 going.  And/or build a NetBSD/sparc64 image
for qemu.