Fixes #319

Gregory Malecha authored 5 years ago and Robbert Krebbers committed 5 years ago

- remove "odd" comment
- move atomic triples to bi_scope

This also gets rid of [val_for_compare]-normalization; instead we introduce a
[LitErased] literal that is suited for use by erasure theorems.

This is the name used by x86 (CMPXCHG) and LLVM.  Also reorder the result (value
first, boolean second) for consistency with LLVM, because why not.

This is a change enabling backward compatibility.

This closes issue #220.

We add a specific constructor to the type of expressions for injecting
values in expressions.

The advantage are :
- Values can be assumed to be always closed when performing
  substitutions (even though they could contain free variables, but it
  turns out it does not cause any problem in the proofs in
  practice). This means that we no longer need the `Closed` typeclass
  and everything that comes with it (all the reflection-based machinery
  contained in tactics.v is no longer necessary). I have not measured
  anything, but I guess this would have a significant performance
  impact.

- There is only one constructor for values. As a result, the AsVal and
  IntoVal typeclasses are no longer necessary: an expression which is
  a value will always unify with `Val _`, and therefore lemmas can be
  stated using this constructor.

Of course, this means that there are two ways of writing such a thing
as "The pair of integers 1 and 2": Either by using the value
constructor applied to the pair represented as a value, or by using
the expression pair constructor. So we add reduction rules that
transform reduced pair, injection and closure expressions into values.
At first, this seems weird, because of the redundancy. But in fact,
this has some meaning, since the machine migth actually be doing
something to e.g., allocate the pair or the closure.

These additional steps of computation show up in the proofs, and some
additional wp_* tactics need to be called.