signature.test fails on big endian

[librsync]

signature.test fails on big endian

[wRAR]

Seems to work on s390x.

[sourcefrog]

Can you have a look whether WORDS_BIGENDIAN and NATIVE_LITTLE_ENDIAN are getting set correctly?

[wRAR]

WORDS_BIGENDIAN is set, NATIVE_LITTLE_ENDIAN is unset (checked by a program including config.h and blake2-impl.h).

[sourcefrog]

I think the next step is to add a test that BLAKE2 produces the right hashes for specified input.

[sourcefrog]

This might have been fixed by the change to CMake - please re-test if you have a bigendian machine.

[wRAR]

Still fails in 01215e2

[dbaarda]

This is either a bug in the CMake config for blake2, or its a bug in blake2 itself.

The blake2 we use is a copy of upstream taken some time ago, so if this is a blake2 bug, it may be fixed upstream already. I also note that Debian sid now has a libb2 package, so this is starting to be available as a library on some platforms, though I don't think it's widely enough available that we don't need to provide it ourselves. IMHO we should;

1. Check the endian related config settings needed by blake2 vs what we have in our config.h and CMakeLists.txt to make sure we are setting things right.

2. Get a newer copy from upstream to see if it's an upstream bug that's been fixed.

3. Change our cmake config to use the platform provided library if it exists.

[dbaarda]

I had a look and there is nothing endian related in our CMakeLists.txt. Our config.h.cmake includes a bunch of #ifdef tests that look to me like it would default to little-endian unless on apple with AC_APPLE_UNIVERSAL_BUILD and BIG_ENDIAN defined, which is not exactly robust. IMHO we should use cmake's TestBigEndian.

However, your tests seem to indicate that the existing limited endian check is correctly detecting your platform as big-endian. This suggests it must be a blake2 bug in the version we have a copy of.

[dbaarda]

I just created pull request #113 to update to the latest blake2 reference implementation. Can someone check/confirm if this fixes this problem?

[dbaarda]

For the record, pull request #114 adds support for using libb2 if it is found.