Back and forth, again... (sigh). With base 3.15 this seems to terminate the
build process more reliably if anything goes wrong.

When the value of a parameter is the unquoted empty string as in "var=",
this would not be recognized and the default value was used instead. Just a
stupid unnecessary explicit test I had to remove. Thanks to Hovanes Egiyan
for reporting this problem.

Pointer + pointer is illegal, but pointer - pointer is fine.

This makes it fit better when included as part of the pdf manual.

This is yet another attempt at fixing the multi-target rule that generates
the SNL parser. we create a directory, run lemon with it as its output dir,
then move the generated files to the current dir.

Review is now complete up to (not including) section 'Asynchronous Use of
pvGet'.

This is so that test programs that are also listed under TESTPROD_HOST can
still be built after running compiler tests that fail.

This failed when the event flag was part of the structure to be initialized.
We now call build_channel_tree on the event flag expression if it has no
channel tree yet.

We now allow them if they are inside a loop inside the wait statement.

This is because we always use the channel now to access the variable, except
when initializing, and for pointer there is no value to initialize.

This is a stopgap measure to avoid generation of semantically wrong code.
The problem is that for wait statements we generate a loop and user code
must not break out of this loop.

At least that is what I think ATM.