std.internal.math.gammfunction assumes 80-bit real

[dlangBugzillaToGithub]

johannespfau (@jpf91) reported this on 2014-07-03T15:21:19Z

Transfered from https://issues.dlang.org/show_bug.cgi?id=13032

Description

std.internal.math.gammfunction assumes that reals are in x86 Extended Precision Format. This is only true for x86 and gammafunction is actually the only module with failing tests on ARM.

A simple way to test for systems where real==double is to use the -mlong-double flag with gdc:

----
#Install gdc via linux distribution or from gdcproject.org/downloads/
wget http://gdcproject.org/downloads/binaries/x86_64-linux-gnu/native_2.065_gcc4.9.0_a8ad6a6678_20140615.tar.xz
tar xf native_2.065_gcc4.9.0_a8ad6a6678_20140615.tar.xz
wget https://raw.githubusercontent.com/D-Programming-GDC/GDC/master/libphobos/src/std/internal/math/gammafunction.d
echo "void main(){}" > main.d
# With default precision, tests should pass on x86 systems
./x86_64-gdcproject-linux-gnu/bin/x86_64-unknown-linux-gnu-gdc gammafunction.d main.d -funittest -g
./a.out
# With real==double precision, tests fail
./x86_64-gdcproject-linux-gnu/bin/x86_64-unknown-linux-gnu-gdc gammafunction.d main.d -funittest  -g -mlong-double-64
./a.out
----

[dlangBugzillaToGithub]

johannespfau (@jpf91) commented on 2014-07-03T17:07:02Z

Sorry, testing is not as simple as I initially thought. Mixing code compiled with -mlong-double-64 and code compiled without is not possible.

So in order to test this on x86 gdc needs to be built from source and phobos/drutime have to be compiled with "-mlong-double-64" as well.

Use
DFLAGS="-mlong-double-64" ./configure
DFLAGS="-mlong-double-64" make
for this.


It might be simpler to just use one of the existing ARM (cross)compilers and test on an ARM system.