Simpler than phys_cast sometimes.

I've also added an osv/ prefix to make sure we don't clash with system headers.

We probably should either adopt this globally or drop it, but I fear we'll run into
problems with the latter approach.

Various glibc tests need it.

The payload is likely to play games with the stack area (like mmap()ing guard
pages), so give it virtual memory to play with instead of physical memory.

Fixes crash with mode=release.

Exceptions are challenging on x64 because the ABI reserves 128 bytes below
%rsp for local variables; this will be overwritten by an exception or
interrupt.  We can't use -mno-red-zone because the payload may not have been
compiled for it.  Instead, use the interrupt stack stack table mechanism
and a separate stack for exceptions.  May fix random corruptions when timers
are used.

Look up the build directory from the running object.

glibc headers define various standard functions as macros, which breaks
the build.  #undef them in the source files.

It is illegal to put a field after a variable-length array, what's its offset?
Breaks mode=release.

Silences 'Subproject commit -dirty' messages

Not completely silent unfortunately.

Benchmark suite

Example:

  make cmdline='java.so -jar bench.jar'

One test so far: java.lang.System.nanoTime()

We're only about 500% slower than native.

Note that the glue code to set up the class path is written in Java.

Needed by Java formatter.

Signed-off-by: Avi Kivity <avi@cloudius-systems.com>