either a tuple, reduction, or a substitution invocation).

Also, rewrite the reduction mapping code to rely on recursive calls
map_tuple(). This cleans up a lot of reduction mapping code
substantially.

dependencies, and also to treat loops more carefully.

* Dependencies are now tracked based on a list of active potential
  sources, and the active set is narrowed by filtering through
  readers/writers
* Stricter handling of dependencies within loops by not assuming
  that loops are nonempty
* Changes the barrier insertion algorithm to act in two stages: first,
  barriers are recursively inserted into loops, then they are inserted
  at the outermost level. This is mostly meant to simplify the code.
* See also #18 on gitlab

check that the index returned is a minimum, rather than the exact value from numpy (in case of duplicate min/maxs)

aliasing (closes inducer/loopy#69).

read-only.

This restriction previously applied to global temporaries.

unscheduled instructions (closes #9).

The check is to make sure that a non-global temporary gets a
definition within every subkernel that uses it.

Also make lp.save_and_reload_temporaries.

non-static loop bounds.

This is intended to help the case when one loop's bounds depend on
another loop, such as the domain

{ [i,j]: 0 <= i <= n and i <= j <= i }

In this case we want the generated code to look like

    for (i = 1; i <= n; ++i)
        for (j = i; j <= i; ++i)
            ...

rather than "for (j = 1; j <= n; ++j)" which are the static loop
bounds for j.

Additional changes that help realize this are as follows:

* Enable code generation of piecewise affine loop bounds, so that
piecewise affine constraints on the "j" loop may be expressed as
conditional expressions.

* Additionally, if "i" is marked local-parallel, we allow the "j" loop
bounds to depend on "i" as long as there are no barriers inside the
"j" loop.