When our argv[0] is not gcc, samba will not optimize the project.....
Nevertheless, the default ist -O2.

Since we have problem we heavily merging in build histories (cpython),
We now try a more robust variant for each commit:

1. Build the parent
   1.a) Last Build was the parent, and that was good -> Do nothing
   1.b) Last Build not the parent. git reset and build
      A) was ok? fine
      B) rebuild failed -> git clean, reset, try again
2. Build the commit itself

We always try to build children first to avoid heavy jumping between
different banches and commits. But that is only an optimization. Lets
give it a try.

Some build systems need a reconfigure. However, we do not want to
measure that, since it is single threaded and way too slow.

I suspect the mbedtls programs to cause a lot of failed commits.
Currently, we see 66/500 commits fail in mbedtls.

During the historical build of cPython, I discovered that hashing of
vector operations was not implemented. Again this had no influence on
the incremental builds, since there was no chance to detect a change.

In C you can have extern definition and declaration in one place:

extern void foo() {
 ...
}

These must not be ignored.

Sometimes (I didn't search for the reason), the objectfile cache
handling was broken.