Summary:
LLVM and Llair use a form of records, in particular for values of
constant structs and arrays. In Llair, these use standard `select` and
`update` operations a la McCarthy's theory of functional arrays, with
a compact `record` operation for constructing complete records. This
is fine and logically well-understood. The issue is that once
constructed, these values are accessed using instructions that (may)
operate over byte-ranges, rather than struct member indices. The
backend uses a theory of sequences to represent such values (the
contents of memory). So some code depends on high fidelity
interoperation between these two views.
This diff resolves this by removing the record theory from the backend
and instead encoding them using the sequence theory. The approach
taken keeps records in Llair and translates them to sequences in
Llair_to_Fol. This choice is made since the encoding into the sequence
theory involves terms that do not have types that are expressible in
terms of the source types. In particular, `(update r i e)`, is encoded
as the concatenation of the prefix of `r` up to the offset of index
`i`, followed by `e` (possibly with padding), and then the suffix of
`r` from index `i+1` on. The prefix and suffix sequences do not
necessarily have source-expressible types.
Reviewed By: jvillard
Differential Revision: D24800866
fbshipit-source-id: e7238c558
Summary:
The support for recursive references to globals from within their
initializers is enough to handle all the cases of recursive structs
that have been encountered so far. Therefore this diff removes the
complication of recursive records entirely.
Reviewed By: jvillard
Differential Revision: D24772955
fbshipit-source-id: f59f06257
Summary:
Change Arith.map to not descend through non-interpreted arithmetic
operators. For example, in `2×(x × y) + 3×z + 4`, `map ~f` will apply
`f` to the subterms `x × y` and `z`, but not `x` or `y`.
The logical notion of "subterm" that is needed by the solver does not
coincide with the representation. This is essentially due to not
"flattening" or "purifying" terms. That is, traditionally `x × y`
would not be permitted as an indeterminate of a polynomial. Instead, a
new variable would need to be introduced: `v = x × y` and then the
polynomial would be expressed as `2×v + 3×z + 4`. Taking maximal
non-interpreted subterms as the definition of "subterm" results in
subterms in the non-flattened representation that are equivalent to
those that would result from flattening the representation.
Reviewed By: jvillard
Differential Revision: D24746235
fbshipit-source-id: d8fcf46a1
Summary:
The implementation of Arithmetic relies on the partial projection from
terms to polynomials. This diff enables it to also embed polynomials
back into terms.
Reviewed By: jvillard
Differential Revision: D24746223
fbshipit-source-id: b6010e7b7
Summary:
Add a distinction between interpreted and uninterpreted arithmetic
terms, and use it in Context.classify. This enables correctly
classifying non-linear terms such as `x × y` as uninterpreted.
Reviewed By: ngorogiannis
Differential Revision: D24746228
fbshipit-source-id: 1a4b0e3bd
Summary:
In the process of computing `Context.solve`, fresh variables can be
generated. Not all of these end up in the final solution
substitution. Currently all of the freshly generated variables are
returned to the client, which leads to extraneous existentials. This
diff trims the returned fresh variables to only those that appear in
the final solution.
Reviewed By: ngorogiannis
Differential Revision: D24746241
fbshipit-source-id: 59a2f221b
Summary:
When exceptions are used due to the lack of goto, use `raise_notrace`
instead of `raise` to avoid the overhead of populating the backtrace.
Reviewed By: ngorogiannis
Differential Revision: D24630525
fbshipit-source-id: c5051d9c4
Josh Berdine