Ok the python 3.5 optimization process could not be done while cross compiling. So I needed to compile on the device in a chroot.

After patching 3.5.6 for compiling warnings I realized that there is 3.5.9 which already has the patches. So I started compiling 3.5.9.
I did many compilings maybe in total 20 different settings (not all on device ) and there was no 100% clean build. Tested 4 differents gcc version but in the end I did come back to stock gcc. For that I need to disable the optimization flag -fprofile-correction, because gcc4.2.1 does not supporting it.

I will attach the build log from my last build. There are some assembler messages but it builds. With other gcc's I was also getting Error: offset too big which breaks compiling.
It could be that one of the linking libraries is not clean or that something was changed in the floating point instructions handler between gcc version or it is binutils (in reality I really don't have a clue).

In principle it should run faster than the previous 3.5.6 build. To gain more speed it could be worth to try a thumb build.

The building took some time only running the test to generate profile data took 4 hours. On device development is fun, but the build time is frustrating. Build log as attachment.

The package needs testing I don't have an use for it, so it is not tested.
https://openrepos.net/content/halftux/python35