Forbid in-place conversions of global objects.

[stevelinton]

After doing some work on cleaning up the library around ConvertToVectorRep and CopyToVectorRep,
I am experimenting with having the kernel check for, and ban, in-place conversion of lists and matrices in the global/atomic region (region zero). This is basically immutable lists and matrices.

This is easy to do, but currently breaks quite a few tests. Mostly these are easy to fix and I'm working through them. If I can make this work it might be worth doing the same for string conversion, blist conversion, etc.

The nice thing is it then legal to make a new mutable (and so thread-local) vector, convert it and then optionally make it immutable, which can never cause a race condition, but illegal to convert an object that could in principle be accessible by another thread.

[stevelinton]

Questions arise like:

Should the error trigger if the conversion doesn't actually have to do anything (ie the list is already converted)?

[rbehrends]

This is necessary, anyway, because in-place mutation of public objects is a source of race conditions, anyway (see also IsStringConv). I think it may be possible to at least get sane behavior with macros in lieu of functions in some places. For example, IsStringConv(obj) would be something like:

#define IsStringConv(obj) (obj = IsString(obj) ? obj : AsString(obj))

(Obviously, there are potential issues with side effects in the obj expression, but at least in the kernel every use of IsStringConv() refers to a variable.)

[stevelinton]

It's easy enough to write the code to pass conversion of atomic objects that are already in the right format. The question is, is it a good idea. Either the caller knows the object will be in the right format, in which case, why are they calling, or they don't in which case this is just a bug waiting to happen.

[stevelinton]

I now have this working again to the extent that testinstall in the main thread runs OK, although I did have to make one small change to the autpgrp package to get rid of a diff in morpheus.tst. The code is in stevelinton/copy-matrep, but be careful merging this, I think I may have broken something in my repo.

The essential difference is that the kernel compressed vector code tries quite hard not to do in place
conversions on any object unless it (a) is writable (b) is not in the global region and (c) does not have IsLockedRepresentationObject. The makes

MakeImmutable(mat);
ConvertToMatrixRep(mat);

fail. The lines work in the reverse order (assuming mat starts our thread-local and mutable).

ImmutableMatrix(mat)

works fine.