Copy stable/10@r272459 to releng/10.1 as part of
the 10.1-RELEASE process.

Approved by: re (implicit)
Sponsored by: The FreeBSD Foundation

Copy head (r256279) to stable/10 as part of the 10.0-RELEASE cycle.

Approved by: re (implicit)
Sponsored by: The FreeBSD Foundation

Add regression tests of the emulation of the long branch conditional form.
These tests can obviously only be run on Merced processors, as they don't
implement the long branch.

Wrap calls to memcpy(3) in a function called block_copy(). This way,
and as long as we're not compiling with IPA, gcc(1) won't optimize
the call away. The whole purpose of using memcpy(3) is to avoid
misaligned loads and stores when we need to read or write the value
in the unaligned memory location. But if gcc(1) optimizes the call
to memcpy(3) away, it will typically introduce misaligned loads and
stores. In this context that's not a good idea.

Skip the register based postinc stores here and don't count them
in the number of tests. This way we avoid skipped tests for non-
existent memory access combinations. The number of tests dropped
from 60 to 48.

Make the local variables global so that the compiler cannot assume
too much about them. This prevents certain peephole optimizations
at -O that invalidate the tests.

Fix the unaligned store with post increment test: The misaligned pointer
stopped pointing to the value the moment we wrote it due to the post
increment. So, grab the value for comparison out of the data structure
directly.

Add tests for post increment. This bumps the number of tests up to
60. The postinc store tests currently fail (value mismatch). Hence
the score as of this commit is 48 out of 60. Either the kernel or
the tests need to be fixed.

Regression test for unaligned loads and stores for short, int, long,
float, double and long double types. No post-increment tests yet.
All tests are skipped if the debug.unaligned_test sysctl variable
cannot be set to 1.