FP: Fix wrong model due to partial assignment (#2910)

For a simple query `(not (= (fp.isSubnormal x) false))`, we were getting
a wrong model. The issue was that `(sign x)` was not assigned a value
and did not appear in the shared terms. In
`TheoryFp::collectModelInfo()`, however, we generate an expression that
connects the components of `x` to `x`, which contains `(sign x)`. As a
result, the normalization while building a model did not result in a
constant. This commit fixes the issue by marking `(sign x)` (and
`(significand x)`) as assignable. Assignable terms can take any value
while building a model if there is no existing value.

1 files changed, 8 insertions, 0 deletions

diff --git a/src/theory/theory_model_builder.cpp b/src/theory/theory_model_builder.cpp
index 77cdfb79b..102979056 100644
--- a/src/theory/theory_model_builder.cpp
+++ b/src/theory/theory_model_builder.cpp
@@ -47,6 +47,14 @@ bool TheoryEngineModelBuilder::isAssignable(TNode n)
       return !n.getType().isFunction();
     }
   }
+  else if (n.getKind() == kind::FLOATINGPOINT_COMPONENT_SIGN)
+  {
+    // Extracting the sign of a floating-point number acts similar to a
+    // selector on a datatype, i.e. if `(sign x)` wasn't assigned a value, we
+    // can pick an arbitrary one. Note that the other components of a
+    // floating-point number should always be assigned a value.
+    return true;
+  }
   else
   {
     // non-function variables, and fully applied functions