But it was recently made thread local. If the value left
in it at the end of one request was smart allocated, the
next request would try to free it - but after its already
been swept.
We could make this RequestLocal, but that seems to add even
more overhead. We could probably make it a __thread TypedValue,
and just set it to uninit at the start of each request (and then
tvAsVariant() where its used). This just puts it back to how
it was, since global_variables() is already being dealt with
appropriately.
TestExtCurl was failing in the fbmake build, because for some
reason the other TestServer was being used. AFAICT it's within the
compiler's rights to do whatever it wants here, though.
TestServer finds a free port to use as the server, but it
used the standard admin and rpc ports. If you had a
webserver running, it would own the admin port, causing
TestServer's tests to time out (because it uses the admin
port to stop them). Similarly rpc tests would fail because
they would be talking to the wrong server.
Find free ports for the admin and rpc servers, and use them.
This change is mostly for FB internal organizational reasons.
Building is not effected beyond the fact that the target now
lands in hphp/hhvm/hhvm rather than src/hhvm/hhvm.
mwilliams