Пример #1
0
    def visit_EUnaryOp(self, e):
        op = e.op
        if op == UOp.Distinct:
            return self.visit_iterable(e)
        elif op == UOp.The:
            return self.find_one(e.e)
        elif op == UOp.Sum:
            sum_var = fresh_var(e.type, "sum")
            loop_var = fresh_var(e.e.type.elem_type, "x")
            self.stms.append(simplify_and_optimize(seq([
                SDecl(sum_var, ENum(0).with_type(e.type)),
                SForEach(loop_var, e.e,
                    SAssign(sum_var, EBinOp(sum_var, "+", loop_var).with_type(INT)))])))
            return sum_var
        elif op == UOp.Length:
            arg = EVar("x").with_type(e.e.type.elem_type)
            return self.visit(EUnaryOp(UOp.Sum, EMap(e.e, ELambda(arg, ONE)).with_type(INT_BAG)).with_type(INT))
        elif op == UOp.All:
            arg = EVar("x").with_type(e.e.type.elem_type)
            return self.visit(EUnaryOp(UOp.Empty, EFilter(e.e, ELambda(arg, ENot(arg))).with_type(INT_BAG)).with_type(INT))
        elif op == UOp.Any:
            arg = EVar("x").with_type(e.e.type.elem_type)
            return self.visit(EUnaryOp(UOp.Exists, EFilter(e.e, ELambda(arg, arg)).with_type(INT_BAG)).with_type(INT))
        elif op == UOp.Empty:
            iterable = e.e
            v = fresh_var(BOOL, "v")
            label = fresh_name("label")
            x = fresh_var(iterable.type.elem_type, "x")
            decl = SDecl(v, ETRUE)
            find = SEscapableBlock(label,
                SForEach(x, iterable, seq([
                    SAssign(v, EFALSE),
                    SEscapeBlock(label)])))
            self.stms.append(simplify_and_optimize(seq([decl, find])))
            return v
        elif op == UOp.Exists:
            return self.visit(ENot(EUnaryOp(UOp.Empty, e.e).with_type(BOOL)))
        # elif op == UOp.AreUnique:
        #     s = fresh_var(TSet(e.e.type.elem_type), "unique_elems")
        #     u = fresh_var(BOOL, "is_unique")
        #     x = fresh_var(e.e.type.elem_type)
        #     label = fresh_name("label")
        #     self.visit(seq([
        #         SDecl(s, EEmptyList().with_type(s.type)),
        #         SDecl(u, ETRUE),
        #         SEscapableBlock(label,
        #             SForEach(x, e.e,
        #                 SIf(EEscape("{s}.find({x}) != {s}.end()", ("s", "x"), (s, x)).with_type(BOOL),
        #                     seq([SAssign(u, EFALSE), SEscapeBlock(label)]),
        #                     SEscape("{indent}{s}.insert({x});\n", ("s", "x"), (s, x)))))]))
        #     return u.id

        return self.visit_Exp(e)
Пример #2
0
 def visit_SWhile(self, w):
     self.begin_statement()
     self.write("for (;;) ")
     with self.block():
         self.visit(
             SIf(ENot(w.e), SEscape("{indent}break;\n", (), ()), SNoOp()))
         self.visit(w.body)
     self.end_statement()
Пример #3
0
    def _setup_handle_updates(self):
        """
        This method creates update code for handle objects modified by each op.
        Must be called once after all user-specified queries have been added.
        """
        for op in self.op_specs:
            print("Setting up handle updates for {}...".format(op.name))
            handles = reachable_handles_at_method(self.spec, op)
            # print("-"*60)
            for t, bag in handles.items():
                # print("  {} : {}".format(pprint(t), pprint(bag)))
                h = fresh_var(t)
                lval = EGetField(h, "val").with_type(t.value_type)
                new_val = inc.mutate(lval, op.body)

                # get set of modified handles
                modified_handles = Query(
                    fresh_name("modified_handles"), Visibility.Internal, [],
                    op.assumptions,
                    EFilter(
                        EUnaryOp(UOp.Distinct, bag).with_type(bag.type),
                        ELambda(h, ENot(EEq(lval,
                                            new_val)))).with_type(bag.type),
                    "[{}] modified handles of type {}".format(
                        op.name, pprint(t)))
                query_vars = [
                    v for v in free_vars(modified_handles)
                    if v not in self.abstract_state
                ]
                modified_handles.args = [(arg.id, arg.type)
                                         for arg in query_vars]

                # modify each one
                subqueries = []
                state_update_stm = inc.mutate_in_place(
                    lval,
                    lval,
                    op.body,
                    abstract_state=self.abstract_state,
                    assumptions=list(op.assumptions) +
                    [EDeepIn(h, bag),
                     EIn(h, modified_handles.ret)],
                    invariants=self.abstract_invariants,
                    subgoals_out=subqueries)
                for sub_q in subqueries:
                    sub_q.docstring = "[{}] {}".format(op.name,
                                                       sub_q.docstring)
                    state_update_stm = self._add_subquery(
                        sub_q=sub_q, used_by=state_update_stm)
                if state_update_stm != SNoOp():
                    state_update_stm = SForEach(
                        h,
                        ECall(modified_handles.name,
                              query_vars).with_type(bag.type),
                        state_update_stm)
                    state_update_stm = self._add_subquery(
                        sub_q=modified_handles, used_by=state_update_stm)
                self.handle_updates[(t, op.name)] = state_update_stm
Пример #4
0
    def array_resize_for_index(self, elem_type, a, i):
        """Resize the array until `i` is a legal index.

        When i < 0, it will do nothing instead.
        """
        new_a = fresh_name(hint="new_array")
        if elem_type == BOOL:
            t = "long"
        else:
            t = self.strip_generics(self.visit(elem_type, name=""))
        len = EEscape("{a}.length", ["a"], [a]).with_type(INT)
        double_and_incr_size = SEscape(
            "{{indent}}{t}[] {new_a} = new {t}[({{len}} << 1) + 1];\n{{indent}}System.arraycopy({{a}}, 0, {new_a}, 0, {{len}});\n{{indent}}{{a}} = {new_a};\n".format(t=t, new_a=new_a),
            ["a", "len"], [a, len])
        self.visit(SWhile(
            EAll([EBinOp(i, ">=", ZERO).with_type(BOOL), ENot(self.array_in_bounds(elem_type, a, i))]),
            double_and_incr_size))
Пример #5
0
 def visit_EBinOp(self, e):
     op = e.op
     if op == "==":
         return self._eq(e.e1, e.e2)
     elif op == "===":
         # rewrite deep-equality test into regular equality
         op = "=="
     elif op == "!=":
         return self.visit(ENot(EEq(e.e1, e.e2)))
     elif op == BOp.Or:
         return self.visit(ECond(e.e1, ETRUE, e.e2).with_type(BOOL))
     elif op == BOp.And:
         return self.visit(ECond(e.e1, e.e2, EFALSE).with_type(BOOL))
     elif op == BOp.In:
         if isinstance(e.e2.type, TSet):
             return self.test_set_containment_native(e.e2, e.e1)
         else:
             raise Exception(
                 "{!r} operator is supposed to be handled by simplify_and_optimize"
                 .format(op))
     return "({e1} {op} {e2})".format(e1=self.visit(e.e1),
                                      op=op,
                                      e2=self.visit(e.e2))
Пример #6
0
def stream(iterable : Exp, loop_var : EVar, body : Stm) -> Stm:
    """Convert an iterable expression to a streaming operation.

    Input:
      iterable - an expression with an iterable type (Bag, Set, or List), not
        yet optimized
      loop_var - a variable to use as the loop variable
      body - a statement to run on that variable, not yet optimized

    Output:
      A statement equivalent to
        for (loop_var in iterable) { body }
      that eliminates as many intermediate collections and objects as possible.

    NOTE: The output of function will not be correct if the body modifies any
    free variable in the iterable expression or writes to any pointers that
    are read by the iterable expression.

    Generating code for the expression

        Map {func} (Filter {predicate} big_collection)

    might create two new collections as large as `big_collection`: one to hold
    the result of the filter and one to hold the result of the map.  If all the
    code needs to do is to iterate over the result, then there is no reason to
    make the two new collections.

    This function is mutually recursive with `simplify_and_optimize`, so any
    transformations performed by that method are also applied to the output of
    this one.
    """

    if isinstance(iterable, EEmptyList):
        return SNoOp()
    elif isinstance(iterable, ESingleton):
        setup, value = simplify_and_optimize_expression(iterable.e)
        # SScoped because if the iterable is e.g. [x] + [y], then the body
        # might be appear in the same block twice.  If the body declares any
        # variables, that will cause problems in languages like Java or C++.
        return seq([setup, SScoped(re_use(value, loop_var, simplify_and_optimize(body)))])
    elif isinstance(iterable, ECond):
        cond_setup, cond = simplify_and_optimize_expression(iterable.cond)
        return seq([
            cond_setup,
            SIf(cond,
                stream(iterable.then_branch, loop_var, body),
                stream(iterable.else_branch, loop_var, body))])
    elif isinstance(iterable, EUnaryOp) and iterable.op == UOp.Distinct:
        tmp = fresh_var(TSet(iterable.type.elem_type), "distinct_elems")
        return seq([
            SDecl(tmp, EEmptyList().with_type(tmp.type)),
            stream(iterable.e, loop_var, SIf(
                ENot(EBinOp(loop_var, BOp.In, tmp).with_type(BOOL)),
                seq([body, SCall(tmp, "add", [loop_var])]),
                SNoOp()))])
    elif isinstance(iterable, EBinOp) and iterable.op == "+":
        return seq([
            stream(iterable.e1, loop_var, body),
            stream(iterable.e2, loop_var, body)])
    elif isinstance(iterable, EBinOp) and iterable.op == "-":
        if is_hashable(iterable.type.elem_type):
            h_setup, h = histogram(iterable.e2)
            val_ref = fresh_var(INT, "count")
            return seq([
                simplify_and_optimize(h_setup),
                stream(
                    iterable.e1,
                    loop_var,
                    SIf(EGt(EMapGet(h, loop_var).with_type(INT), ZERO),
                        SMapUpdate(h, loop_var, val_ref, SAssign(val_ref, EBinOp(val_ref, "-", ONE).with_type(INT))),
                        body))])
        else:
            rhs = fresh_var(iterable.e2.type, "bag_subtraction_right")
            return seq([
                simplify_and_optimize(SDecl(rhs, iterable.e2)),
                stream(
                    iterable.e1,
                    loop_var,
                    SIf(EIn(loop_var, rhs),
                        SCall(rhs, "remove", (loop_var,)),
                        body))])
    elif isinstance(iterable, EFilter):
        return stream(
            EFlatMap(iterable.e, ELambda(iterable.predicate.arg,
                ECond(iterable.predicate.body,
                    ESingleton(iterable.predicate.arg).with_type(iterable.type),
                    EEmptyList().with_type(iterable.type)).with_type(iterable.type))).with_type(iterable.type),
            loop_var,
            body)
    elif isinstance(iterable, EMap):
        return stream(
            EFlatMap(iterable.e, ELambda(iterable.transform_function.arg,
                ESingleton(iterable.transform_function.body).with_type(iterable.type))).with_type(iterable.type),
            loop_var,
            body)
    elif isinstance(iterable, EFlatMap):
        inner_loop_var = fresh_var(
            iterable.transform_function.arg.type,
            iterable.transform_function.arg.id)
        return stream(
            iterable.e,
            inner_loop_var,
            stream(iterable.transform_function.apply_to(inner_loop_var), loop_var, body))
    elif isinstance(iterable, EListSlice):
        elem_type = iterable.type.elem_type
        l = fresh_var(iterable.e.type, "list")
        s = fresh_var(INT, "start")
        e = fresh_var(INT, "end")
        return simplify_and_optimize(seq([
            SDecl(l, iterable.e),
            SDecl(s, max_of(iterable.start, ZERO)),
            SDecl(e, min_of(iterable.end, ELen(l))),
            SWhile(ELt(s, e), seq([
                SDecl(loop_var, EListGet(l, s).with_type(elem_type)),
                body,
                SAssign(s, EBinOp(s, "+", ONE).with_type(INT))]))]))
    elif isinstance(iterable, ELet):
        v = fresh_var(
            iterable.body_function.arg.type,
            iterable.body_function.arg.id)
        return seq([
            simplify_and_optimize(SDecl(v, iterable.e)),
            stream(iterable.body_function.apply_to(v), loop_var, body)])
    elif isinstance(iterable, EMove):
        return stream(iterable.e, loop_var, body)
    else:
        assert is_collection(iterable.type), repr(iterable)
        setup, e = simplify_and_optimize_expression(iterable)
        return seq([setup, SForEach(loop_var, e, simplify_and_optimize(body))])
Пример #7
0
def possibly_useful_nonrecursive(
    solver,
    e: Exp,
    context: Context,
    pool=RUNTIME_POOL,
    assumptions: Exp = ETRUE,
    ops: [Op] = ()) -> bool:
    """Heuristic filter to ignore expressions that are almost certainly useless."""

    state_vars = OrderedSet(v for v, p in context.vars() if p == STATE_POOL)
    args = OrderedSet(v for v, p in context.vars() if p == RUNTIME_POOL)
    assumptions = EAll([assumptions, context.path_condition()])
    at_runtime = pool == RUNTIME_POOL

    h = extension_handler(type(e))
    if h is not None:
        res = h.possibly_useful(e, context, pool, assumptions, ops, solver)
        if not res:
            return res

    if isinstance(e, EStateVar) and not free_vars(e.e):
        return No("constant value in state position")
    if (isinstance(e, EDropFront)
            or isinstance(e, EDropBack)) and not at_runtime:
        return No("EDrop* in state position")
    if not allow_big_sets.value and isinstance(e, EFlatMap) and not at_runtime:
        return No("EFlatMap in state position")
    if not allow_int_arithmetic_state.value and not at_runtime and isinstance(
            e, EBinOp) and e.type == INT:
        return No("integer arithmetic in state position")
    if is_collection(e.type) and not is_scalar(e.type.elem_type):
        return No("collection of nonscalar: e {}\n elem_type: {}\n".format(
            e, e.type.elem_type))
    if isinstance(e.type, TMap) and not is_scalar(e.type.k):
        return No("bad key type {}".format(pprint(e.type.k)))
    if isinstance(e.type, TMap) and isinstance(e.type.v, TMap):
        return No("map to map")
    # This check is probably a bad idea: whether `the` is legal may depend on
    # the contex that the expression is embedded within, so we can't skip it
    # during synthesis just because it looks invalid now.
    # if isinstance(e, EUnaryOp) and e.op == UOp.The:
    #     len = EUnaryOp(UOp.Length, e.e).with_type(INT)
    #     if not valid(EImplies(assumptions, EBinOp(len, "<=", ENum(1).with_type(INT)).with_type(BOOL))):
    #         return No("illegal application of 'the': could have >1 elems")
    if not at_runtime and isinstance(
            e, EBinOp) and e.op == "-" and is_collection(e.type):
        return No("collection subtraction in state position")
    # if not at_runtime and isinstance(e, ESingleton):
    #     return No("singleton in state position")
    if not allow_nonzero_state_constants.value and not at_runtime and isinstance(
            e, ENum) and e.val != 0:
        return No("nonzero integer constant in state position")
    if not allow_binop_state.value and at_runtime and isinstance(
            e, EStateVar) and isinstance(e.e, EBinOp) and is_scalar(
                e.e.e1.type) and is_scalar(e.e.e2.type):
        return No(
            "constant-time binary operator {!r} in state position".format(
                e.e.op))
    if not allow_conditional_state.value and not at_runtime and isinstance(
            e, ECond):
        return No("conditional in state position")
    if isinstance(e, EMakeMap2) and isinstance(e.e, EEmptyList):
        return No("trivially empty map")
    if isinstance(e, EMakeMap2) and isinstance(e.e, ESingleton):
        return No("really tiny map")
    if not at_runtime and (isinstance(e, EArgMin) or isinstance(e, EArgMax)):
        # Cozy has no way to efficiently implement mins/maxes when more than
        # one element may leave the collection.
        from cozy.state_maintenance import mutate
        for op in ops:
            elems = e.e
            elems_prime = mutate(elems, op.body)
            formula = EAll([assumptions] + list(op.assumptions) + [
                EGt(
                    ELen(
                        EBinOp(elems, "-", elems_prime).with_type(elems.type)),
                    ONE)
            ])
            if solver.satisfiable(formula):
                return No(
                    "more than one element might be removed during {}".format(
                        op.name))
    if not allow_peels.value and not at_runtime and isinstance(e, EFilter):
        # catch "peels": removal of zero or one elements
        if solver.valid(
                EImplies(
                    assumptions,
                    ELe(
                        ELen(
                            EFilter(
                                e.e,
                                ELambda(e.predicate.arg, ENot(
                                    e.predicate.body))).with_type(e.type)),
                        ONE))):
            return No("filter is a peel")
    if not allow_big_maps.value and not at_runtime and isinstance(
            e, EMakeMap2) and is_collection(e.type.v):
        all_collections = [sv for sv in state_vars if is_collection(sv.type)]
        total_size = ENum(0).with_type(INT)
        for c in all_collections:
            total_size = EBinOp(total_size, "+",
                                EUnaryOp(UOp.Length,
                                         c).with_type(INT)).with_type(INT)
        my_size = EUnaryOp(
            UOp.Length,
            EFlatMap(
                EUnaryOp(UOp.Distinct, e.e).with_type(e.e.type),
                e.value_function).with_type(e.type.v)).with_type(INT)
        s = EImplies(assumptions,
                     EBinOp(total_size, ">=", my_size).with_type(BOOL))
        if not solver.valid(s):
            return No("non-polynomial-sized map")

    return True
Пример #8
0
def improve(target: Exp,
            context: Context,
            assumptions: Exp = ETRUE,
            stop_callback: Callable[[], bool] = never_stop,
            hints: [Exp] = (),
            examples: [{
                str: object
            }] = (),
            cost_model: CostModel = None,
            ops: [Op] = (),
            improve_count: Value = None):
    """Improve the target expression using enumerative synthesis.

    This function is a generator that yields increasingly better and better
    versions of the input expression `target` in the given `context`.  The
    `cost_model` defines "better".

    It periodically calls `stop_callback` and exits gracefully when
    `stop_callback` returns True.

    Other parameters:
        - assumptions: a precondition.  The yielded improvements will only be
          correct when the assumptions are true.
        - hints: expressions that might be useful.  These will be explored
          first when looking for improvements.
        - examples: inputs that will be used internally to differentiate
          semantically distinct expressions.  This procedure discovers more
          examples as it runs, so there usually isn't a reason to provide any.
        - ops: update operations.  This function may make different choices
          about what expressions are state expressions based on what changes
          can happen to that state.

    Key differences from "regular" enumerative synthesis:
        - Expressions are either "state" expressions or "runtime" expressions,
          allowing this algorithm to choose what things to store on the data
          structure and what things to compute at query execution time. (The
          cost model is ultimately responsible for this choice.)
        - If a better version of *any subexpression* for the target is found,
          it is immediately substituted in and the overall expression is
          returned. This "smooths out" the search space a little, allowing us
          find kinda-good solutions very quickly, even if the best possible
          solution is out of reach.  This is more desireable than running for
          an indeterminate amount of time doing nothing.
    """

    print("call to improve:")
    print("""improve(
        target={target!r},
        context={context!r},
        assumptions={assumptions!r},
        stop_callback={stop_callback!r},
        hints={hints!r},
        examples={examples!r},
        cost_model={cost_model!r},
        ops={ops!r})""".format(target=target,
                               context=context,
                               assumptions=assumptions,
                               stop_callback=stop_callback,
                               hints=hints,
                               examples=examples,
                               cost_model=cost_model,
                               ops=ops))

    target = inline_lets(target)
    target = freshen_binders(target, context)
    assumptions = freshen_binders(assumptions, context)

    if heuristic_done(target):
        print("The target already looks great!")
        return

    print()
    print("improving: {}".format(pprint(target)))
    print("subject to: {}".format(pprint(assumptions)))
    print()

    is_wf = exp_wf(target, context=context, assumptions=assumptions)
    assert is_wf, "initial target is not well-formed: {}".format(is_wf)

    state_vars = [v for (v, p) in context.vars() if p == STATE_POOL]
    if eliminate_vars.value and can_elim_vars(target, assumptions, state_vars):
        print("This job does not depend on state_vars.")
        # TODO: what can we do about it?

    hints = ([freshen_binders(h, context) for h in hints] + [
        freshen_binders(wrap_naked_statevars(a, state_vars), context)
        for a in break_conj(assumptions)
    ] + [target])
    print("{} hints".format(len(hints)))
    for h in hints:
        print(" - {}".format(pprint(h)))
    vars = list(v for (v, p) in context.vars())
    funcs = context.funcs()

    solver = solver_for_context(context, assumptions=assumptions)

    if not solver.satisfiable(ETRUE):
        print("assumptions are unsat; this query will never be called")
        yield construct_value(target.type)
        return

    is_good = possibly_useful(solver, target, context)
    assert is_good, "WARNING: this target is already a bad idea\n is_good = {}, target = {}".format(
        is_good, target)

    examples = list(examples)

    if cost_model is None:
        cost_model = CostModel(funcs=funcs, assumptions=assumptions)

    watched_targets = [target]
    blacklist = {}

    while True:
        # 1. find any potential improvement to any sub-exp of target
        for new_target in search_for_improvements(targets=watched_targets,
                                                  wf_solver=solver,
                                                  context=context,
                                                  examples=examples,
                                                  cost_model=cost_model,
                                                  stop_callback=stop_callback,
                                                  hints=hints,
                                                  ops=ops,
                                                  blacklist=blacklist):
            print("Found candidate improvement: {}".format(pprint(new_target)))

            # 2. check
            with task("verifying candidate"):
                counterexample = solver.satisfy(ENot(EEq(target, new_target)))

            if counterexample is not None:
                if counterexample in examples:
                    print("assumptions = {!r}".format(assumptions))
                    print("duplicate example: {!r}".format(counterexample))
                    print("old target = {!r}".format(target))
                    print("new target = {!r}".format(new_target))
                    raise Exception("got a duplicate example")
                # a. if incorrect: add example, restart
                examples.append(counterexample)
                print("new example: {!r}".format(counterexample))
                print("wrong; restarting with {} examples".format(
                    len(examples)))
                break
            else:
                # b. if correct: yield it, watch the new target, goto 1
                print("The candidate is valid!")
                print(repr(new_target))
                print("Determining whether to yield it...")
                with task("updating frontier"):
                    to_evict = []
                    keep = True
                    old_better = None
                    for old_target in watched_targets:
                        evc = retention_policy(new_target, context, old_target,
                                               context, RUNTIME_POOL,
                                               cost_model)
                        if old_target not in evc:
                            to_evict.append(old_target)
                        if new_target not in evc:
                            old_better = old_target
                            keep = False
                            break
                    for t in to_evict:
                        watched_targets.remove(t)
                    if not keep:
                        print(
                            "Whoops! Looks like we already found something better."
                        )
                        print(" --> {}".format(pprint(old_better)))
                        continue
                    if target in to_evict:
                        print("Yep, it's an improvement!")
                        yield new_target
                        if heuristic_done(new_target):
                            print("target now matches doneness heuristic")
                            return
                        target = new_target
                    else:
                        print("Nope, it isn't substantially better!")

                watched_targets.append(new_target)
                print("Now watching {} targets".format(len(watched_targets)))
                break

        if improve_count is not None:
            with improve_count.get_lock():
                improve_count.value += 1
Пример #9
0
def improve(target: Exp,
            context: Context,
            assumptions: Exp = T,
            stop_callback=never_stop,
            hints: [Exp] = (),
            examples: [{
                str: object
            }] = (),
            cost_model: CostModel = None,
            ops: [Op] = ()):
    """
    Improve the target expression using enumerative synthesis.
    This function is a generator that yields increasingly better and better
    versions of the input expression `target`.

    Notes on internals of this algorithm follow.

    Key differences from "regular" enumerative synthesis:
        - Expressions are either "state" expressions or "runtime" expressions,
          allowing this algorithm to choose what things to store on the data
          structure and what things to compute at query execution time. (The
          cost model is ultimately responsible for this choice.)
        - If a better version of *any subexpression* for the target is found,
          it is immediately substituted in and the overall expression is
          returned. This "smooths out" the search space a little, and lets us
          find kinda-good solutions very quickly, even if the best possible
          solution is out of reach.
    """

    print("call to improve:")
    print("""improve(
        target={target!r},
        context={context!r},
        assumptions={assumptions!r},
        stop_callback={stop_callback!r},
        hints={hints!r},
        examples={examples!r},
        cost_model={cost_model!r},
        ops={ops!r})""".format(target=target,
                               context=context,
                               assumptions=assumptions,
                               stop_callback=stop_callback,
                               hints=hints,
                               examples=examples,
                               cost_model=cost_model,
                               ops=ops))

    target = freshen_binders(target, context)
    assumptions = freshen_binders(assumptions, context)

    print()
    print("improving: {}".format(pprint(target)))
    print("subject to: {}".format(pprint(assumptions)))
    print()

    is_wf = exp_wf(target, context=context, assumptions=assumptions)
    if not is_wf:
        print(
            "WARNING: initial target is not well-formed [{}]; this might go poorly..."
            .format(is_wf))
        print(pprint(is_wf.offending_subexpression))
        print(pprint(is_wf.offending_subexpression.type))

    state_vars = [v for (v, p) in context.vars() if p == STATE_POOL]
    if eliminate_vars.value and can_elim_vars(target, assumptions, state_vars):
        print("This job does not depend on state_vars.")
        # TODO: what can we do about it?

    hints = ([freshen_binders(h, context) for h in hints] + [
        freshen_binders(wrap_naked_statevars(a, state_vars), context)
        for a in break_conj(assumptions)
    ] + [target])
    print("{} hints".format(len(hints)))
    for h in hints:
        print(" - {}".format(pprint(h)))
    vars = list(v for (v, p) in context.vars())
    funcs = context.funcs()

    solver = solver_for_context(context, assumptions=assumptions)

    if not solver.satisfiable(T):
        print("assumptions are unsat; this query will never be called")
        yield construct_value(target.type)
        return

    examples = list(examples)

    if cost_model is None:
        cost_model = CostModel(funcs=funcs, assumptions=assumptions)

    watched_targets = [target]
    learner = Learner(watched_targets,
                      solver,
                      context,
                      examples,
                      cost_model,
                      stop_callback,
                      hints,
                      ops=ops)

    while True:
        # 1. find any potential improvement to any sub-exp of target
        for new_target in learner.search():
            print("Found candidate improvement: {}".format(pprint(new_target)))

            # 2. check
            with task("verifying candidate"):
                counterexample = solver.satisfy(ENot(EEq(target, new_target)))

            if counterexample is not None:
                if counterexample in examples:
                    print("assumptions = {!r}".format(assumptions))
                    print("duplicate example: {!r}".format(counterexample))
                    print("old target = {!r}".format(target))
                    print("new target = {!r}".format(new_target))
                    raise Exception("got a duplicate example")
                # a. if incorrect: add example, reset the learner
                examples.append(counterexample)
                event("new example: {!r}".format(counterexample))
                print("wrong; restarting with {} examples".format(
                    len(examples)))
                learner.reset(examples)
                break
            else:
                # b. if correct: yield it, watch the new target, goto 1
                print("The candidate is valid!")
                print(repr(new_target))
                print("Determining whether to yield it...")
                with task("updating frontier"):
                    to_evict = []
                    keep = True
                    old_better = None
                    for old_target in watched_targets:
                        evc = eviction_policy(new_target, context, old_target,
                                              context, RUNTIME_POOL,
                                              cost_model)
                        if old_target not in evc:
                            to_evict.append(old_target)
                        if new_target not in evc:
                            old_better = old_target
                            keep = False
                            break
                    for t in to_evict:
                        watched_targets.remove(t)
                    if not keep:
                        print(
                            "Whoops! Looks like we already found something better."
                        )
                        print(" --> {}".format(pprint(old_better)))
                        continue
                    if target in to_evict:
                        print("Yep, it's an improvement!")
                        yield new_target
                        if heuristic_done(new_target):
                            print("target now matches doneness heuristic")
                            return
                        target = new_target
                    else:
                        print("Nope, it isn't substantially better!")

                watched_targets.append(new_target)
                print("Now watching {} targets".format(len(watched_targets)))
                learner.watch(watched_targets)
                break